CN118365139A

CN118365139A - Risk situation awareness method, system, equipment and medium

Info

Publication number: CN118365139A
Application number: CN202410773136.XA
Authority: CN
Inventors: 魏新迟; 宋平; 王杰; 李存斌; 张琪祁; 周健; 时珊珊
Original assignee: State Grid Shanghai Electric Power Co Ltd
Current assignee: State Grid Shanghai Electric Power Co Ltd
Priority date: 2024-06-17
Filing date: 2024-06-17
Publication date: 2024-07-19
Anticipated expiration: 2044-06-17
Also published as: CN118365139B

Abstract

The application relates to the technical field of smart grids, in particular to a risk situation awareness method, a risk situation awareness system, computer equipment and a storage medium. The risk situation awareness method comprises the following steps: acquiring historical operation data of situation indexes of each power node in the target power distribution network; performing cluster recognition according to the historical operation data aiming at each power node to obtain a cluster result data set containing typical operation states of each node; taking node real-time index data of each node as input, and determining a prediction result data set of each power node based on a node state prediction model obtained by pre-training; and classifying and judging the prediction result data set and the clustering result data set based on a classification algorithm model so as to determine the risk situation of each power node. The method for situation awareness of the urban power distribution network can be used for realizing multi-energy access.

Description

Risk situation awareness method, system, equipment and medium

Technical Field

The application relates to the technical field of smart grids, in particular to a risk situation awareness method, a risk situation awareness system, computer equipment and a storage medium.

Background

Situation awareness is a means for acquiring, understanding and displaying security elements capable of causing system situation change and forward prediction of recent development trend based on security big data. Under the background of access of a large amount of distributed renewable energy sources, risk situation awareness is carried out on an information physical system (CPS, cyber PHYSICAL SYSTEMS) of the urban flexible power distribution network, and real-time and fine risk early warning can be realized, so that the important task of safety construction of the flexible urban power grid is realized.

At present, risk situation awareness of an urban power grid only aims at two situations, namely, the risk of occurrence of a power grid physical system and the risk of occurrence of a power grid information physical system under network attack. For the urban power distribution network with multi-energy access, risks are generated in a physical layer, a network layer (monitoring equipment, a dispatching system and the like), a coupling layer (transmission equipment and the like) and an environment layer (temperature, climate and the like) of the urban power distribution network. Meanwhile, the urban flexible power distribution network is required to have the capability of comprehensively sensing physical, network, environment and other multi-layer information, so that effective response is made on disturbance of each layer, and the existing situation sensing index cannot be suitable for the urban power distribution network with multi-energy access.

Therefore, a method for situation awareness of the urban power distribution network with multiple energy access is needed.

Disclosure of Invention

The embodiment of the application provides a risk situation awareness method, a risk situation awareness system, computer equipment and a storage medium.

In a first aspect of the embodiment of the present application, there is provided a risk situation awareness method, including:

Acquiring historical operation data of situation indexes of each power node in a target power distribution network; wherein, the situation indexes refer to indexes in a pre-constructed risk situation index system, and the risk situation index system at least comprises: a physical layer, a network layer, a coupling layer, and an environmental layer;

Performing cluster recognition on the historical operation data aiming at each power node to obtain a cluster result data set containing typical operation states of each power node; the typical operation state refers to a state that the operation parameters of all the power nodes are in a preset normal operation parameter range;

Taking node real-time index data of each node as input, and determining the node prediction state of each power node based on a node state prediction model obtained by pre-training to obtain a prediction result data set;

and classifying and judging the prediction result data set and the clustering result data set based on a classification algorithm model so as to determine the risk situation of each power node.

In an optional embodiment of the present application, the performing cluster recognition on the historical operation data for each power node to obtain a cluster result data set including typical operation states of each power node includes:

sampling the historical operation data by adopting a variable grid division density deviation sampling method to obtain a sampling data set;

Generating a clustering decision graph for each power node based on the sampled data set;

And determining a neighborhood distance threshold parameter of the clustering decision graph, and carrying out clustering recognition on each historical operation data in the sampling data set according to the minimum neighborhood point parameter to obtain the clustering result data set containing the typical operation state of each power node.

In an optional embodiment of the present application, the sampling the historical operating data by using a variable grid-partitioned density deviation sampling method to obtain a sampled data sample set includes:

Dividing each historical operation data in each situation index into a plurality of intervals according to each situation index;

calculating the density similarity of the historical operation data among all the intervals;

merging all intervals with the density similarity larger than a preset threshold value into an interval grid;

And sampling according to a density deviation sampling method aiming at each interval grid to obtain the sampling data set.

In an alternative embodiment of the application, the density similarity is calculated according to the following formula:

Wherein, Representing the density similarity of the two intervals,Is the firstThe first of the dimension indexesHistorical operating data for each interval.

In an alternative embodiment of the application, the probability f that each interval grid is sampled is:

Wherein, Represent the firstThe grid density of the individual inter-zone grids,As a total number of grids,Is a sampling index.

In an optional embodiment of the application, the generating a cluster decision graph for each power node based on the sampled data set comprises:

calculating the reachable distance of each sample point aiming at the sample point in the sampling data set of each power node;

sequencing all the sample points according to the reachable distance to obtain an ordered result queue;

And drawing an reachable distance graph of each sample point according to the arrangement sequence of each sample point in the result queue to obtain the clustering decision graph aiming at each power node.

In an optional embodiment of the present application, the node state prediction model is a transducer model, and correspondingly, the determining, by using node real-time index data of each node as input, a node prediction state of each power node based on a node state prediction model obtained by training in advance, to obtain a prediction result data set includes:

In the transducer model encoder, a multi-head self-attention layer converts each real-time index data in an input sequence into a query, a key and a value, calculates attention distribution among the real-time index data in each power node, and obtains fusion information representing each sample point.

And carrying out forward propagation on the fusion information based on a full connection layer in the encoder to obtain encoder output.

In an alternative embodiment of the present application, the calculation formula of the multi-head self-attention layer is as follows:

Wherein, The vector dimensions of the keys are represented,、、Representing different vectors of real-time index data.

In an alternative embodiment of the present application, the calculation formula of the encoder output is:

Wherein, Represent the firstThe group multi-headed attention encoder output,、、Are all the vectors of the two-dimensional vector,Is thatDimension of the vector.

In an alternative embodiment of the present application, the decoder in the transducer model further comprises a masked multi-headed self-attention layer, and the input data of the masked multi-headed self-attention layer is node real-time index data before the current processing sample point.

In an alternative embodiment of the application, the transducer model uses a piecewise loss function in training.

In an alternative embodiment of the application, the segment loss function is as follows:

Wherein: In order to predict the value of the loss, Is the firstThe first sample ofThe true value of the individual index(s),In order to be able to predict the value,In order to measure the volume of the sample to be tested,For the segmentation threshold value,Is the error coefficient under the different segments.

In an optional embodiment of the application, before the classifying algorithm model is used for classifying and judging the prediction result data set and the clustering result data set to determine the risk situation of each power node, the method further comprises:

carrying out standardized processing on each data in the clustering result data set;

And carrying out class label coding on the clustering result data set to be used as input of pre-constructed classification algorithm model training.

In an alternative embodiment of the present application, the node real-time index data of each node further comprises season index data.

In an optional embodiment of the present application, the node real-time index data includes an illumination intensity index and a wind power index; the method further comprises the steps of: and calculating the real-time numerical value and the change rate of the node real-time index data.

In an alternative embodiment of the present application, the rate of change of the node real-time index data is calculated according to the following formula:

Wherein, Representing the change rate of the real-time index data of the node, which is used for representing the power node in the process ofThe index stability of the moment of time,For the real-time radiation intensity per minute within the sampling interval,Is the mean value of the real-time radiation intensity over the sampling interval.

In an optional embodiment of the present application, the classifying the prediction result data set and the clustering result data set based on the classification algorithm model to determine a risk situation of each power node includes:

if the predicted result data set is in the clustered result data set, determining that the risk situation of the current node is a health state;

And if the predicted result data set is not in the clustered result data set, determining that the risk situation of the current node is a risk state.

In a second aspect of the embodiment of the present application, there is provided a risk situation awareness system, including:

The acquisition module is used for acquiring historical operation data of situation indexes of each power node in the target power distribution network; wherein, the situation indexes refer to indexes in a pre-constructed risk situation index system, and the risk situation index system at least comprises: a physical layer, a network layer, a coupling layer, and an environmental layer;

The identification module is used for carrying out cluster identification on the historical operation data aiming at each power node to obtain a cluster result data set containing the typical operation state of each power node; the typical operation state refers to a state that the operation parameters of all the power nodes are in a preset normal operation parameter range;

the first determining module is used for taking node real-time index data of each node as input, determining the node prediction state of each power node based on a node state prediction model obtained by training in advance, and obtaining a prediction result data set;

And the second determining module is used for carrying out classification judgment on the prediction result data set and the clustering result data set based on a classification algorithm model so as to determine the risk situation of each power node.

In a third aspect of the embodiment of the present application, there is provided a computer apparatus including: comprising a memory and a processor, said memory storing a computer program, characterized in that the processor implements the steps of the method according to any of the preceding claims when executing said computer program.

In a fourth aspect of embodiments of the present application, there is provided a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method as described in any of the above.

The embodiment of the application provides a risk situation awareness method, which takes CPS nodes of a power distribution network as objects, realizes node typical running state identification through clustering based on data density, realizes node future running characteristic prediction and risk situation awareness through deep learning on the basis, and judges whether future time is in a risk state or not. In the first aspect, indexes (node voltage, node power, node network side frequency, equipment temperature, log alarm information, vulnerability information, data transmission delay, data loss, wind power (only for wind power nodes), illumination intensity (only for photoelectric nodes) and environmental temperature) of four physical, information, coupling and environmental layers of the CPS node of the power distribution network are monitored and early-warned in real time. The unit-level power CPS of each wind power plant, photovoltaic power station, user unit building, transformer substation and the like is used as a node. The current situation that a large amount of renewable energy sources are connected into the urban flexible power distribution network and the construction requirement of the flexible power grid are considered, when the risk situation awareness index is constructed, the environment layer index is added on the basis of the traditional physical layer, network layer and coupling layer, so that the situation awareness process considers the influence of environment factors on power grid equipment, the accuracy of risk state identification is improved, and the defect of the prior art in the situation awareness index dimension is overcome; in the second aspect, the embodiment of the application realizes the prediction of the future operation characteristics of the power nodes in the urban flexible power distribution network area through a neural network algorithm based on a self-attention mechanism, and carries out the future situation awareness of the nodes on the basis of the future characteristic prediction value; according to the third aspect, the running state of each power node in the urban flexible power distribution network area is accurately captured, the risk source can be positioned during risk situation awareness, and the situation awareness accuracy is improved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application and do not constitute a limitation on the application. In the drawings:

FIG. 1 is an overall flow chart of a risk situation awareness method according to an embodiment of the present application;

Fig. 2 is a schematic diagram of a risk situation index system in a risk situation awareness method according to an embodiment of the present application;

FIG. 3 is a process of encoding and decoding a transducer model according to an embodiment of the present application;

FIG. 4 is a graph of reachable distance of photovoltaic power plant node data OPTICS (Ordering points to IDENTIFY THE clustering structure, method of determining cluster structure by point-to-point ordering) in accordance with one embodiment of the application;

FIG. 5 is an illustration of index features of different exemplary states of a photovoltaic power plant node in accordance with one embodiment of the present application;

FIG. 6 is a graph showing the predicted generated power of a photovoltaic power plant according to an embodiment of the present application;

FIG. 7 is a graph showing the total error of index value predictions for a photovoltaic power plant node in accordance with one embodiment of the present application;

FIG. 8 is a graph showing the evaluation index of the classification result of the photovoltaic power station node in one embodiment of the present application;

FIG. 9 is a graph showing a photovoltaic power plant node risk situation discrimination result in an embodiment of the present application;

FIG. 10 is a graph showing the predicted results of environmental temperature indicators according to one embodiment of the present application;

FIG. 11 is a graph showing the predicted result of data loss values according to an embodiment of the present application;

FIG. 12 is a diagram of a user element node classification result evaluation index in accordance with an embodiment of the present application;

FIG. 13 is a graph showing a risk situation determination result of a subscriber unit node in an embodiment of the present application;

FIG. 14 shows the accuracy of identifying risk situations by the risk situation awareness method according to the embodiment of the present application under different indexes in an embodiment of the present application;

FIG. 15 is a graph showing the predicted generated power of a photovoltaic power plant using a Informer model according to one embodiment of the present application;

FIG. 16 is a diagram showing total error in index value prediction for a photovoltaic power plant node by using Informer model in accordance with one embodiment of the present application.

Detailed Description

In the process of implementing the application, the inventor finds that a method for situation awareness of the urban power distribution network with multi-energy access is needed.

In view of the above problems, the embodiments of the present application provide a risk situation awareness method, a risk situation awareness system, a computer device, and a storage medium.

The scheme in the embodiment of the application can be realized by adopting various computer languages, such as object-oriented programming language Java, an transliteration script language JavaScript and the like.

In order to make the technical solutions and advantages of the embodiments of the present application more apparent, the following detailed description of exemplary embodiments of the present application is provided in conjunction with the accompanying drawings, and it is apparent that the described embodiments are only some embodiments of the present application and not exhaustive of all embodiments. It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other.

Referring to fig. 1, the risk situation awareness method provided by the embodiment of the present application may be divided into three levels, including S1 situation awareness, S2 situation awareness, and S3 situation prediction, where the specific implementation contents of each level are elaborated with reference to the following implementation:

the risk situation awareness method provided by the embodiment of the application comprises the following steps 201 to 204:

Step 201, acquiring historical operation data of situation indexes of each power node in a target power distribution network;

The situation indexes refer to indexes in a pre-constructed risk situation index system, and different from the traditional method, the whole regional power grid is taken as a research object. Fig. 2 is an exemplary risk situation indicator system according to an embodiment of the present application, where the risk situation indicator system at least includes: the physical layer (refers to a specific entity device layer), the network layer (refers to a software system, a module and the like of an information layer), the coupling layer (hardware or software of an interaction layer between the physical layer and the network layer) and the environment layer (refers to an environment layer where each entity in the physical layer and/or the network layer is located); the situation index in the physical layer may be, for example: node voltage, node power, node network side frequency, equipment temperature, log alarm information, vulnerability information, data transmission delay, data loss, environmental temperature and the like; the situation index in the network layer may be, for example: log alarm information, vulnerability information and the like; the situation index in the coupling layer may be, for example: data transmission delay, data loss, etc.; the situation index in the environment layer may be, for example: wind power (wind power node), illumination intensity (light Fu Jiedian), ambient temperature, etc.

Through the setting of the environment layer, the consideration of environmental factors is increased when the risk situation index is selected, and the toughness characteristics such as the perception force, the cooperative force and the like of the toughness power grid are considered, so that real-time data of each node of the target power distribution network, including power data in a physical layer, network information transmission data in a network layer and the like, are obtained; the detection of the environmental layer on wind power and illumination intensity reflects the receiving and transmitting capacity of distributed power supply nodes in the power distribution network, so that the strain of the power grid is perceived, and meanwhile, the prediction precision is improved, and therefore, the index system can effectively reflect risk factors threatening the toughness of the power system. Wherein a part of the index interpretation is shown in table 1.

Table 1 partial index quantization method

The data types in the historical operation data include, but are not limited to, the index data, and may also include other various influencing parameters, and the embodiment of the present application is not limited specifically. This historical operating data belongs to S1 in fig. 1: and the situation awareness level is used for awareness of the operation situation of each power node through the grid big data.

Step 202, for each power node, performing cluster recognition on the historical operation data to obtain a cluster result data set containing typical operation states of each node;

The typical operation state refers to a state that the operation parameters of each power node are in a preset normal operation parameter range, and can be colloquially understood as a preset normal operation state. Opposite to this typical operating state, i.e. atypical operating state, may also be referred to as abnormal operating state, i.e. operating state which exceeds the normal operating range.

Step 203, taking node real-time index data of each node as input, and determining a node prediction state of each power node based on a node state prediction model obtained by pre-training to obtain a prediction result data set;

The node state prediction model is obtained by training the operation data of each node as input quantity based on the state of each node as supervision, and a pre-training model of the node state prediction model and the training process of the pre-training model are not described herein, and can be flexibly adjusted according to actual conditions, so that only the function of predicting the state of the future node according to the operation data of each node is required. The node real-time index data is the operation data of each node, and the operation data belongs to the power node situation index in the step 201. The data in the prediction result data set are the prediction data or the prediction node state of each power node, and are not real data or real states.

And 204, classifying and judging the prediction result data set and the clustering result data set based on a classification algorithm model so as to determine the risk situation of each power node.

For example, it may be determined whether each data or each node state in the prediction result data set belongs to the clustering result data set in step 202, if so, it means that the prediction result is one of the typical operation states of each node, and the future node situation is in a normal situation, otherwise, it belongs to an abnormal situation.

In an alternative embodiment of the present application, the typical operation state of each node may be determined based on an OPTICS clustering algorithm (Ordering points to IDENTIFY THE clustering structure, a density-based clustering algorithm), and the specific process may be as follows:

defining a data set for each power node in the target distribution network (I.e., clustering result data set) storing node sample data (i.e., historical operating data)WhereinIs an index measurement value (namely, historical operation data of each situation index). The OPTICS algorithm selects a neighborhood distance threshold parameter by generating a decision graphAnd combining the minimum neighborhood point parametersDBSCAN (Density-Based Spatial Clustering of Applications with Noise, a Density-based clustering algorithm) clustering is performed, and the problem that a distance-based clustering method cannot process high-dimensional irregular data is solved. The DBSCAN algorithm needs to define six concepts first, and can be specifically as follows:

Core object: for a pair of Which is provided withNeighborhood (radius is)Neighborhood of (f) number of samples in a sampleDefinition ofIs a core object.

The density is direct: if the power node sampleTo core objectsIs the Euclidean distance of (2)ThenDensity up to。

The density can be achieved: if a certain node sample setAny of (3)Density up toThenAndThe density can be achieved.

Core distance: for a certain core objectCore distance:

（1）

in the formula (1), Representation ofAt a distance ofThe number of sample points in the neighborhood of (a),Representation ofTime of daySample points for which there is no core distance as non-core points,Representation and rendering ofThe outermost sample point within the smallest neighborhood that becomes the core object,Representing the distance between the two sample points.

The distance can be reached: for a core objectThe reachable distance is as follows:

（2）

in the formula (2), Is thatA kind of electronic deviceAt any point in the vicinity of the point,Representation ofTime of dayThere are no sample points of reachable distance as non-core points.

The embodiment of the application provides a risk situation awareness method, which takes each power node in a CPS of a target power distribution network as an object, realizes node typical running state identification through clustering based on data density, realizes node future running characteristic prediction and risk situation awareness through deep learning on the basis, and judges whether the future moment is in a risk state or not.

In a first aspect, a risk situation indicator system in an embodiment of the present application at least includes: and the physical layer, the network layer, the coupling layer and the environment layer monitor and early warn each power node in the CPS of the target power distribution network in real time according to the historical operation data of the 4 layers, and determine the risk situation in the CPS of the target power distribution network. The running state of each power node in the target power distribution network is accurately captured and determined, and the risk source can be positioned during risk situation awareness, so that the situation awareness accuracy is improved.

In the second aspect, when the risk situation awareness index system is constructed, the environment layer index is added on the basis of the traditional physical layer, the network layer and the coupling layer, so that the situation awareness process considers the influence of environment factors on power grid equipment, the accuracy of risk state identification is improved, and the defect of the prior art in the situation awareness index dimension is overcome;

In the third aspect, the embodiment of the application determines the node prediction state of each power node based on the node state prediction model obtained by pre-training, predicts the future operation characteristics of each power node in the target power distribution network, and perceives the future situation of the node based on the future characteristic prediction value, thereby realizing the perception of the future risk situation of each node of the target power distribution network;

In a fourth aspect, the embodiment of the present application performs classification judgment on the prediction result data set and the clustering result data set based on a classification algorithm model to determine risk situations of each power node, where the prediction result data set and the clustering result data set are both performed based on each node, and may accurately position the power node with potential risk while performing risk situation prediction, determine a risk source, and improve situation awareness accuracy.

In an optional embodiment of the present application, step 202, performing cluster recognition on the historical operation data for each power node to obtain a cluster result data set including typical operation states of each power node, includes the following steps 301 to 303:

Step 301, sampling the historical operation data by adopting a variable grid division density deviation sampling method to obtain a sampling data set;

step 302, generating a clustering decision graph for each power node based on the sampling data set;

Step 303, determining a neighborhood distance threshold parameter of the clustering decision graph, and performing clustering recognition on each historical operation data in the sampling data set according to the minimum neighborhood point parameter to obtain the clustering result data set containing the typical operation state of each power node.

With continued reference to fig. 1, in S2 situation understanding, the operation characteristics of the CPS node of the target power distribution network are obtained through analysis of the operation situation big data. Firstly, the embodiment of the application samples the historical real-time power grid, network, coupling and environment layer data of each node, so that the running time of the model can be greatly reduced; secondly, the embodiment of the application clusters the sampled data through an OPTICS algorithm based on data density, and identifies the typical running state of the nodes in a data driving mode, thereby adaptively capturing the running characteristics of the power nodes in different complex scenes, and overcoming the defect that the running state of subjective definition cannot adapt to the current power grid polygon and complexity.

The embodiment of the application adopts a Variable grid partition density deviation sampling (VG_GBS, variable GRID DENSITY Biased Sampling) method, the sampling points of the method can cover the whole distribution interval as much as possible, the sampling probability is inversely proportional to the distribution density, the too sparse sampling points of the operation state with less data can be avoided, and the reliability of the data is improved. Meanwhile, the clustering data set is determined based on the clustering decision diagram, so that the risk situation awareness efficiency of the embodiment of the application can be greatly improved.

In an optional embodiment of the present application, the step 301 of sampling the historical operating data by using a variable grid division density deviation sampling method to obtain a sampled data set may include the following steps 401 to 404:

Step 401, dividing each historical operation data in each situation index into a plurality of intervals according to each situation index;

step 402, calculating the density similarity of the historical operation data among all intervals;

step 403, merging all intervals with the density similarity larger than a preset threshold value into an interval grid;

the variable meshing process in the embodiment of the application can be, for example: setting the historical operation data of each power node (namely the values of the situation awareness indexes of the nodes at each moment) as a model ) WhereinIs a feature dimension (i.e. a situational awareness index dimension),For the number of time points, for the firstThe dimension data are sorted, the equal depth of the dimension data is divided into k sections, and the density similarity can be calculated according to the following formula (3):

（3）

In the formula (3), epsilon represents the density similarity of two intervals, Is the firstThe first dimension indexHistorical operating data for each interval.

Let the ordered data set be represented asThenIf (3)The two intervals are combined. After the merging work is completed, the whole node historical operation data sample set is divided into a plurality of interval grids with different sizes.

And 404, sampling according to a density deviation sampling method for each interval grid to obtain the sampling set.

In an alternative embodiment of the present application, the probability of each interval grid being sampled is as described aboveThe method comprises the following steps:

（4）

in the formula (4) of the present invention, Represent the firstThe grid density of the inter-grid, which may be represented by the total number of historical operating data points for the power nodes within the inter-grid,Representing the total number of grids,Representing the sampling index.

The embodiment of the application is based on merging each interval with the density similarity larger than a preset threshold value into one interval grid, and sampling according to a density deviation sampling method for each interval grid to obtain the sampling set. The interval grids are divided according to the density similarity, then the divided interval grids are sampled according to a density deviation sampling method, sampling is more reasonable, too sparse operation state sampling points with less data can be avoided, and the reliability of the data is improved.

In an optional embodiment of the present application, the generating a cluster decision map for each power node based on the sampled data set in step 302 includes the following steps 501-503:

Step 501, calculating the reachable distance of each sample point aiming at the sample point in the sampling data set of each power node;

Step 502, sorting all sample points according to the reachable distance to obtain an ordered result queue;

and step 503, drawing an reachable distance graph of each sample point according to the arrangement sequence of each sample point in the result queue, and obtaining the clustering decision graph aiming at each power node.

For example: knowing the power node sample dataset D, two queues are created: an ordered queue O and a result queue R; first case: if all points in the power node sample data set D are processed or core points do not exist, ending the algorithm; in the second case, a node history running data sample point p which is not processed (i.e. is not in the result queue R) and is a core object is selected, p is firstly put in the result queue R, p is deleted from the sample data set D, then all densities of p in the sample data set D are found to reach the sample point x, and the reachable distance from x to p is calculated: if x is not in the ordered queue O, putting x and the reachable distance into the ordered queue O; if x is in the ordered queue O and the new reachable distance of x is smaller, updating the reachable distance of x, and finally reordering the data in the ordered queue O according to the reachable distance from small to large; and in the third case, if the ordered queue O is empty, returning to the step, otherwise, taking out the first sample point y (i.e. the sample point with the smallest distance) in the O, putting the first sample point y into the result queue R, and deleting y from the sample data set D and the ordered queue O. If y is not a core object, repeatedly searching a sample point with the smallest reachable distance of the residual data in the ordered queue O; fourth case: y is a core object, finding all density direct sample points of y in a sample data set D, calculating the reachable distance to y, and updating all density direct sample points of y into an ordered queue O according to the above;

And repeating the steps until the algorithm is finished, and finally obtaining an orderly output result and a corresponding reachable distance. After the algorithm is finished, the obtained ordered result queue R stores corresponding reachable distances, and the reachable distances are sequentially drawn into a reachable distance graph, namely an OPTICS decision graph.

Trough setting presented by reachable distance in OPTICS decision diagramThe parameters are as follows:

（5）

in the formula (5) of the present invention, And (5) carrying out DBSCAN clustering treatment on the maximum value of all the wave valley points. The DBSCAN clustering step may be: selecting unprocessed core objects, and collecting all density reachable samples into a cluster; and continuing to select other core objects, and repeating the clustering step until the clustering is completed.

The embodiment of the application realizes the identification of the typical running state of each power node in the target power distribution network by means of the sampling algorithm and the clustering algorithm based on the data density, and compared with the traditional subjective definition method of the typical running state, the method can more comprehensively find different running states of each node; meanwhile, the advantage of the data density clustering method in processing high-dimensional irregular data is adopted, the recognition rate of the normal state and the abnormal state is improved, namely, the accuracy of situation recognition results is improved on the basis of reducing labor cost, and the method is more suitable for recognizing the typical running state of the power node.

In an optional embodiment of the present application, the node state prediction model is a transducer model, and correspondingly, in the step 203, the node real-time index data of each node is used as input, and the node prediction state of each power node is determined based on the node state prediction model obtained by training in advance, so as to obtain a prediction result data set, which includes the following steps 601-602:

in step 601, in the transducer model encoder, the multi-head self-attention layer converts the real-time running data sample points of the nodes at each time of the real-time index data in the input sequence into query, key and value, and calculates the attention distribution among the real-time index data in each power node to obtain fusion information representing each sample point.

And 602, forward propagating the fusion information based on a full connection layer in the encoder to obtain encoder output.

The embodiment of the application predicts the future operation characteristics or the future operation states of all power nodes in the target power distribution network based on an improved converter deep learning model. Referring to fig. 3, an exemplary architecture of a transducer model is shown, and the transducer model is constructed based on a self-attention mechanism, which enables the transducer model to calculate the relation between different feature dimensions in an input sequence, so that the transducer model is more suitable for real-time operation data of high-dimensional power nodes with complex association relations between indexes. In the transducer model, the historical operation data of each node is taken as input, is divided into a plurality of subsequences after passing through an embedding layer, each subsequence passes through a plurality of self-attention layers and a full-connection layer, and finally is output through linear transformation, namely the real-time operation data of the node at a future time point.

The embodiment of the application realizes the prediction of the future real-time operation characteristics of the power nodes in the urban flexible power distribution network area based on a transducer model by means of a self-attention mechanism, and carries out the future situation awareness of the nodes on the basis of the future characteristic predicted value. The accuracy of future feature prediction is improved through the learning capacity setting of the self-attention mechanism on the relation among the situation indexes, so that the model operation effect added into the prediction step is not inferior to that of the traditional situation awareness method based on the true value, and the timeliness of risk early warning is improved on the premise of ensuring the situation identification accuracy.

Furthermore, it is to be further explained that: the transducer model comprises two parts: an encoder and a decoder. The encoder is mainly responsible for converting the input sequence into a fixed-length vector representation, which the decoder decodes into the output sequence.

In an encoder, each layer includes two sublayers: a multi-headed self-focusing layer and a fully-connected layer. The multi-head self-attention layer converts the node real-time operation data sample points of each time in the input sequence into a query (Q), a key (K) and a value (V), calculates attention distribution between each time point and other operation data sample points of all time points, and obtains a weighted sum to represent fusion information of the sample points. And the full connection layer carries out forward propagation on the fusion information to obtain the output of the layer.

（6）

in the formula (6) of the present invention, The vector dimensions of the keys are represented,、、Representing different vectors of real-time index data. The calculation formula shows that for each queryThe multi-head self-attention layer will be based on the multi-head self-attention layer and all keysFor all valuesWeighted summation is performed.

In an optional embodiment of the present application, the above-mentioned calculation formula of the encoder output is:

（7）

in the formula (7) of the present invention, Represent the firstThe group multi-headed attention encoder output,、、Are vectors, and are all the same as the vector,Is thatDimension of the vector.

In an optional embodiment of the present application, the decoder in the foregoing transform model further includes a masking multi-head self-attention layer, and input data of the masking multi-head self-attention layer is node real-time index data before the current processing sample point.

Can be used in the encoding processRespectively withGroup independent weight vector、、Multiplying to obtain vector、、Then (1)Group self-attention settlement results:

（8）

in the formula (8), the expression "a", Is thatDimension of the vector. Will beSplicing with an optimizable weight matrixMultiplying to obtain multi-head attention output result. As shown in fig. 4-16, willAnd obtaining the output of the coding layer through residual connection and FNN (Feedforword Neural Network, feedforward neural network).

According to the embodiment of the application, the prediction of the future operation characteristics of each power node in the target power distribution network is realized through a neural network algorithm based on a self-attention mechanism, and the future situation of the node is perceived on the basis of the future characteristic prediction value, so that the risk situation perception of the target power distribution network is realized.

In the decoding process, the target sequence is adopted as input by the decoding layer, and the h+1th input sequence is that by masking the multi-head self-attention layerUsing output of coding layers、Instead of in the encoding-decoding attention layer、And then the output is obtained after the same steps as the coding layer.

In the decoder, a masking multi-head self-attention layer is added in addition to the multi-head self-attention layer and full-connection layer in the encoder. This layer is similar to the multi-headed self-attention layer in the encoder, but only the position before this position is considered when calculating the attention profile, i.e. only the nodes before this point in time are considered to run the data in real time, thus avoiding the problem of using future point in time information in the decoder.

The training process of the transducer model is typically done using Maximum Likelihood Estimation (MLE). I.e. the model predicts the probability of the output sequence for a given input sequence and maximizes its probability value. At the same time, to avoid overfitting, regularization terms, such as L2 regularization, are also typically added.

The method comprises the steps of acquiring historical real-time sequence data of each power node, and predicting the future running state of the node by means of a transducer, wherein the specific process is as follows:

1) Input device

Defining window length(Number of time steps representing historical real-time data of power node) index numberInput vector：

（9）

（10）

In the formula (9) and the formula (10),A position code representing a point in time is presented,Represent the firstUnder the time nodeNormalized measurement of each index.

In an optional embodiment of the present application, before the step 204 of classifying the prediction result data set and the clustering result data set based on the classification algorithm model to determine risk situations of each power node, the method further includes the following steps 701-702:

step 701, performing standardization processing on each data in the clustering result data set;

And 702, performing class label coding on the clustering result data set, and using the class label coding as input of pre-constructed classification algorithm model training.

The standardized processing can be unified, normalized, dimensionless, data class conversion or class division of the data format, and the like, and the embodiment of the application is not particularly limited and can be flexibly adjusted according to actual conditions. According to the embodiment of the application, through carrying out standardized processing on each data in the clustering result data set and carrying out class label coding on the clustering result data set, the complexity of subsequent data processing can be greatly reduced, the data processing efficiency is improved, and the risk situation sensing efficiency of the embodiment of the application is further improved as the input of the pre-constructed classification algorithm model training.

In an optional embodiment of the present application, the node real-time index data of each node further includes seasonal index data, that is, a seasonal situation index is introduced into the environmental layer of the situation index system of fig. 2, so that the situation awareness process considers the influence of seasonal variation on the power grid equipment, thereby improving the accuracy of risk status identification.

In one embodiment of the application, the judgment can be performed based on the future running risk situation of the nodes of the nerve random forest:

And combining the clustering result and the node future operation characteristic prediction result data set, identifying the category of the future operation state through a classification algorithm, and identifying the moment of the atypical operation state as a risk situation. The classification algorithm is a neural random forest (Neural Random Forest, NRF), and the clustered data set D# is subjected to class label coding and is used as model input:

（11）

in the formula (11), the color of the sample is, Is the firstThe first group of the personal groupAnd (3) taking a convolutional neural network (Convolutional Neural Networks, CNN) as a random network fused in the NRF, receiving a prediction result of each decision tree as input, calculating through a convolutional layer, a pooling layer and a full-connection layer, and outputting a final prediction category to improve the accuracy of the model. And taking the future time node data as input of a classifier for completing training to obtain a risk situation discrimination result.

In an optional embodiment of the present application, the node real-time index data includes an illumination intensity index and a wind power index; the method further comprises the following steps:

and calculating the real-time numerical value and the change rate of the node real-time index data.

In an alternative embodiment of the application, the data preprocessing step employs a min-max normalization process for a large number of irregular, non-uniform dimensions of data prior to inputting the model. The characteristics of equipment temperature, environmental temperature and the like are greatly influenced by seasons, and one-dimensional seasonal characteristics are added to the original data from a typical running state identification link in order to enable the model to identify abnormal states in different seasons: map {0,0.5,1} to { winter, spring/autumn, summer }, respectively. And simultaneously calculating real-time numerical values and change rates of the illumination intensity and the wind power index, thereby considering the influence of environmental factors of the renewable energy nodes on the stability of the power grid. The change rate in the node real-time index data is calculated as follows:

（12）

in the formula (12) of the present invention, Representing the change rate of the real-time index data of the node, which is used for representing the power node in the process ofThe illumination intensity (or wind power) index stability at the moment,For real-time illumination radiation intensity (or wind speed) per minute during the sampling interval,Is the mean value of the real-time illumination radiation intensity (or wind speed) in the sampling interval.

In an optional embodiment of the present application, the classifying algorithm model is used to classify and judge the prediction result data set and the clustering result data set to determine a risk situation of each power node, where the classifying algorithm model includes the following two cases:

first case: if the predicted result data set is in the clustered result data set, determining that the risk situation of the current node is a health state;

Second case: and if the predicted result data set is not in the clustered result data set, determining that the risk situation of the current node is a risk state.

The embodiment of the application determines whether the risk situation of the current node belongs to the health state or the analysis state by judging whether the predicted result data set is in the clustered result data set, and the determination method is simple, quick and high in accuracy.

In an alternative embodiment of the present application, because the internal range of each index data of the power node is large, a single loss function is adopted during model training to cause uneven error distribution, and the converter model calculates errors by adopting a piecewise loss function during training. Therefore, the embodiment of the application adopts the segmentation loss function to calculate the error, divides the interval according to the size of the real value standardized by each index, and multiplies different intervals by different coefficients and square root error results to calculate the loss function so as to perform model tuning, thereby improving the accuracy and reliability of the transducer model.

In an alternative embodiment of the present application, the segment loss function is as follows:

（13）

（14）

In the formulas (13) and (14): the predicted loss value is indicated as such, Represent the firstThe first sample ofThe true value of the individual index(s),The predicted value is represented by a value of the prediction,Which is indicative of the volume of the sample being measured,Representing the segmentation threshold value,Representing the error coefficients under the different segments.

According to the embodiment of the application, the accuracy of future feature prediction can be improved through the arrangement of the piecewise loss function, so that the model operation effect added into the prediction step is not inferior to that of the traditional situation awareness method based on the true value, and the timeliness of risk early warning is improved on the premise of ensuring the situation identification accuracy.

In order to further explain the risk situation awareness method provided by the embodiment of the application, different embodiments are provided below by taking the target power distribution network as an urban flexible power distribution network as an example.

The distributed energy is accessed in a large scale and high permeability, so that the CPS complexity of the urban distribution network is increased, and meanwhile, various operation risks are brought. The embodiment of the application provides a city toughness power distribution network CPS risk situation sensing method based on deep learning, which comprises the following steps: urban flexible distribution network CPS node typical running state identification model based on data density; node future operation feature prediction model based on improved transducer: NRF-based node future risk situation discrimination model. Real-time risk monitoring and early warning can be provided for CPS operation and maintenance of the urban flexible power distribution network, and self-healing capacity and emergency response capacity of the CPS operation and maintenance are improved.

Selecting CPS operation parameters of a distribution network in a certain area for performing calculation analysis (including nodes such as a plurality of wind power stations, photovoltaic power stations, power distribution stations, user unit buildings and the like), wherein the data span of a model training set is 2020/01/01-2020/12/31, and the data step length is 30min; the prediction set chooses 2021 data.

Example 1:

With continued reference to fig. 1, in one embodiment, a photovoltaic power plant node is the subject.

(1) The situation awareness index used in the step S1 comprises the following steps: node voltage, node power, node network side frequency, equipment temperature, log alarm information, vulnerability information, data transmission delay, data loss, illumination intensity and environmental temperature.

(2) Step S2

1) Sampling

For each node, the number of samples is set at 20% (3504) of the total data, let t=0.75, k=20, e=0.5 for vg_dbs sampling.

2) Clustering

Clustering the sampled results by using an OPTICS algorithm, wherein the OPTICS algorithm needs to determine two parameters: order the=30, And assuming(Inf is an infinite sign) and the reachable distance is shown in fig. 4.

By means of profile coefficientsThe compactness of each cluster and the separation degree of different clusters in the clustering result can be judged, the value range is [ -1,1], and the closer to 1, the higher the clustering quality is represented. Using grid search, inDetermining the optimum in rangeAnd (5) performing DBSCAN clustering on the result.

Table 2 shows comparison of clustering results of the OPTICS algorithm, and compared with other methods, the method has the advantages that the profile coefficient of a sample after the OPTICS clustering is higher, so that the defects that a traditional distance-based clustering algorithm cannot accurately reflect the distribution rule of high-dimensional data and cannot find non-spherical clusters are overcome, and a better clustering effect can be achieved when power nodes with high feature dimensions and uneven distribution run data in real time. In addition, density clustering algorithms can filter outliers (i.e., noise) through the setting of density thresholds, whereas distance-based clustering algorithms cannot identify noise and therefore are subject to outliers when identifying typical operating conditions.

And calculating the average value of the sample points of each index value in each cluster to obtain different typical state index characteristics. And (5) removing indexes with insignificant mean differences among different clusters, and screening out features influencing the clustering result, as shown in fig. 5. It can be seen that OPTICS classifies photovoltaic power plant nodes into a high temperature season low load state (class 1), a low temperature season low load state (class 2), a low temperature season high load state (class 3), a high temperature season high load state (class 5), and successfully identifies the state when the change in the morning and evening illumination radiation is significant (class 4), and the state when it is subject to a small network attack but does not pose a threat (class 6). The clustering result can successfully reflect complex typical states.

Comparison of clustering results of Table 2 OPTICS with other methods

(3) Step S3

1) Prediction

Setting input parameters of a transducer model training: the number of coding layers is 6, the number of decoding layers is 6, the input sequence length is 48, the output sequence length is 6, the vector dimension is 128, and the number of multi-attention heads is 8. The specific piecewise function parameter settings are shown in table 3. By means of root mean square errorAverage absolute percentage errorThe accuracy of the improved transducer predictions was verified and the error pairs before and after model tuning are shown in table 4. In addition, the traditional BP network and Bi-LSTM network prediction method are selected for comparison, and the effectiveness of the used deep learning method is verified.

TABLE 3 segment loss function parameter settings

TABLE 4 error of the results of the transform predictions before and after improvement

The root mean square error before and after model tuning has no obvious change, and the average absolute percentage error after tuning is obviously reduced, which shows that the proposed loss function adjustment method plays an effective role in reducing the percentage error of the prediction result.

And selecting the data of the measured nodes in the 0 th day of 2021, 5 months, 23 th day, 0 th day, 5 months, 26 th day and 0 th day for analysis, wherein the abnormality of power generation fluctuation occurs in the photovoltaic power station. Fig. 6 shows predicted values of the actual illumination radiation, the actual electricity-using-side required power, the actual generated power, and the generated power of the three networks in this period. It can be seen that about 10 days of 5 months 24, the actual illumination radiation is sufficient, the load demand of the electricity utilization side is kept at a higher level, but the actual power generation of the photovoltaic power station suddenly drops to an abnormal value, and the node is recovered at 13 days of 5 months 24, and is in an abnormal situation. From the predicted values of the three networks, the improved transducer well predicts the fluctuation of the power generated by the power station, and after the power station is recovered to the normal state, the predicted value is again close to the true value within 2 time steps. And the predicted results of Bi-LSTM and BP networks are greatly different from the true values.

Fig. 7 shows the total error values of all index predictions for the nodes for the three networks during this time period, and it can be seen that under normal conditions, the prediction error value of the improved transducer model is slightly lower than that of the other two networks, and remains at a lower level. Under the risk situation, the prediction errors of the three networks are greatly increased, but the errors of the improved transducer are obviously lower.

2) Classification

The method I and the method II are used for measured node data simultaneously, the accuracy rate and recall rate of each state of the node are calculated by the three methods, and the result is shown in fig. 8. It can be seen that the effect of the method of the model provided by the embodiment of the application is no more than that of the other two models for the identification of the normal state. The accuracy rate and recall rate of the abnormal state identification are obviously higher than those of other models, so that more abnormal states are correctly captured, and fewer normal states are misdetected as abnormal.

The recognition results of the time periods before and after the node presents the abnormal situation are shown in fig. 9, and in the abnormal state, the judgment result has delay of 1 time step, so that the abnormal state can be basically recognized; in other time periods, the judgment result has 3 errors, and the recognition accuracy rate in the time period in the graph is 96.53%.

Example 2:

In one embodiment, a subscriber unit building node is targeted.

(1) The situation awareness index used in the step S1 comprises the following steps: node voltage, node power, node network side frequency, equipment temperature, log alarm information, vulnerability information, data transmission delay, data loss and environmental temperature.

(2) Step S2

1) Sampling

For each node, VG_DBS sampling is performed with the same parameters as in example 1.

2) Clustering

The result of clustering the selected user unit nodes shows that the number of clusters is 4, the contour coefficient is 0.47, and the noise ratio is 3.32%.

(3) Step S3

1) Prediction

Data in a high-temperature period (7 months, 28 days, 0 time to 8 months, 1 day, 0 time) in 2021 summer is selected for analysis, and when the sensor of a unit building node is damaged due to the influence of high temperature in the period of 7 months, 29 days and 12 days, a large amount of data is lost in the duration time, and the sensor is in an abnormal situation. The prediction results of the model on the node data are shown in fig. 10 and 11.

As can be seen from FIG. 10, in the normal state, the improved transducer model has higher accuracy in predicting the environmental temperature index with little fluctuation of error. When the abnormal state is reached, the real values of the data loss indexes shown in fig. 12 are greatly changed, the prediction errors of the three methods are suddenly increased and decreased, the prediction result of the improved transducer model is closer to the real values, and after the abnormal state is ended, the prediction error is quickly reduced to a normal level, so that the situation that the abnormal state of other indexes is caused by the fact that the related information among the indexes is successfully learned by means of a self-attention mechanism can be predicted.

2) Classification

The accuracy of the proposed method for identifying future operational states of the subscriber unit node is shown in fig. 12, and it can be seen that, for the subscriber unit node, the proposed model has better identification effects on several states than the other two methods, and is more obvious at the moment of the abnormal state.

In the time period before and after the node is abnormal, the situation recognition result is shown in fig. 13, so that the model can accurately and timely recognize the abnormal state when the risk occurs, and the predicted value is close to the true value in 1 time step after the risk is over. The recognition accuracy in the time period in the graph is 97.2%.

It can be found that after the prediction step is added in the method provided by the embodiment of the application, the accuracy of situation identification is not reduced, but the result is better than other models, and the prediction accuracy of the improved transducer algorithm is explained again. In addition, the method provided by the embodiment of the application identifies more normal states for each node, and the discrimination accuracy of each running state is in a higher level, especially in a part related to risks, so that the typical state automatic identification strategy driven by the data can more comprehensively identify the running state of the node, the method is suitable for the CPS of the current complex and changeable power grid, and the accuracy of the method provided by the embodiment of the application in distinguishing the normal state from the abnormal state is higher than that of the traditional method by means of the improvement of the data density on the clustering effect of the high-dimensional data and the processing effect on the noise point.

Example 3:

in one implementation manner, the indexes in the method provided by the embodiment of the application are replaced by the indexes used in the literature a and the literature b, the risk situation awareness is carried out by taking a plurality of nodes as objects, and the accuracy of the result of identifying the risk states of the nodes under different indexes is counted, as shown in fig. 14. It can be seen that, since the network layer factor is considered in the document a, the application effect of the index provided by the document a is better than that of the document b which only considers the physical layer factor, and the risk state identification result reaches the highest value under the index provided by the embodiment of the application, which indicates that the addition of the environmental layer index successfully leads the model to learn the influence of the environmental factor on the power node equipment, and further improves the accuracy of situation awareness.

Example 4:

In one embodiment, the clustering model is replaced with a density peak clustering (DENSITY PEAKS clustering, DPC) algorithm, the predictive model is replaced with a Informer model also with the aid of a self-attention mechanism, targeting the measured photovoltaic power plant node.

In step S2, on the basis of sampling, a density-based DPC model is adopted to conduct typical running state identification, the clustering result is shown in table 5, the contour coefficient is larger than 0.5, and a good clustering level is achieved.

TABLE 5 DPC clustering results

The result of predicting the generated power of the photovoltaic power station in the period before and after the abnormal situation by the Informer model is shown in fig. 15, and the total error of the predicted values of the indexes is shown in fig. 16. It can be seen that the accuracy of the model for prediction in the normal state is maintained at 95% or more, and fluctuation of the generated electric power is successfully distinguished in the abnormal state.

It should be understood that, although the steps in the flowchart are shown in sequence as indicated by the arrows, the steps are not necessarily performed in sequence as indicated by the arrows. The steps are not strictly limited to the order of execution unless explicitly recited herein, and the steps may be executed in other orders. Moreover, at least some of the steps in the figures may include multiple sub-steps or stages that are not necessarily performed at the same time, but may be performed at different times, nor does the order in which the sub-steps or stages are performed necessarily performed in sequence, but may be performed alternately or alternately with at least a portion of other steps or other steps.

One embodiment of the present application provides a risk situation awareness system, comprising:

In an optional embodiment of the present application, the identification module is specifically configured to sample the historical operating data by using a variable grid-divided density deviation sampling method to obtain a sampled data set; generating a clustering decision graph for each power node based on the sampled data set; and determining a neighborhood distance threshold parameter of the clustering decision graph, and carrying out clustering recognition on each historical operation data in the sampling data set according to the minimum neighborhood point parameter to obtain the clustering result data set containing the typical operation state of each power node.

In an optional embodiment of the present application, the identification module is specifically configured to divide, for each situation indicator, each historical operation data in the situation indicator into a plurality of intervals; calculating the density similarity of the historical operation data among all the intervals; merging all intervals with the density similarity larger than a preset threshold value into an interval grid; and sampling according to a density deviation sampling method aiming at each interval grid to obtain the sampling data set.

wherein epsilon represents the density similarity of the two intervals, Is the firstUpper middle of dimension indexHistorical operating data for each interval.

Wherein, Represent the firstThe grid density of the individual interval grid, expressed in terms of said historical running data point total of power nodes within the grid,As a total number of grids,Is a sampling index.

In an optional embodiment of the present application, the identification module is specifically configured to calculate, for sample points in the sampled data set of each power node, an reachable distance of each sample point; sequencing all the sample points according to the reachable distance to obtain an ordered result queue; and drawing an reachable distance graph of each sample point according to the arrangement sequence of each sample point in the result queue to obtain the clustering decision graph aiming at each power node.

In an optional embodiment of the present application, the node state prediction model is a transducer model, and the first determining module is specifically configured to, in the transducer model encoder, convert each real-time index data in an input sequence into a query, a key, and a value by using a multi-head self-attention layer, and calculate attention distribution among the real-time index data in each power node, so as to obtain fusion information representing each sample point; and carrying out forward propagation on the fusion information based on a full connection layer in the encoder to obtain encoder output.

Wherein, The vector dimensions of the keys are represented,、、The differences representing the real-time index data are vectors.

Wherein: the predicted loss value is indicated as such, Represent the firstThe first sample ofThe true value of the individual index(s),The predicted value is represented by a value of the prediction,Which is indicative of the volume of the sample being measured,Representing the segmentation threshold value,Representing the error coefficients under the different segments.

In an optional embodiment of the present application, the second determining module is further configured to perform a normalization process on each data in the clustering result dataset; and carrying out class label coding on the clustering result data set to be used as input of pre-constructed classification algorithm model training.

In an optional embodiment of the present application, the node real-time index data includes an illumination intensity index and a wind power index; the first determining module is further configured to calculate a real-time value and a change rate of the node real-time index data.

Wherein, Representing the change rate of the real-time index data of the node, which is used for representing the power node in the process ofThe index stability of the moment of time, _{Representation of} The real-time radiation intensity per minute during the sampling interval,Is the mean value of the real-time radiation intensity over the sampling interval.

In an optional embodiment of the present application, the second determining module is further configured to determine that a risk situation of the current node is a health state if the prediction result dataset is in the clustering result dataset; and if the predicted result data set is not in the clustered result data set, determining that the risk situation of the current node is a risk state.

For specific limitations on the risk situation awareness system, reference may be made to the above limitation on the risk situation awareness method, and no further description is given here. The various modules in the risk situation awareness system described above may be implemented in whole or in part by software, hardware, and combinations thereof. The above modules may be embedded in hardware or may be independent of a processor in the computer device, or may be stored in software in a memory in the computer device, so that the processor may call and execute operations corresponding to the above modules.

In one embodiment, a computer device is provided that includes a processor, a memory, a network interface, and a database connected by a system bus. Wherein the processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device is for storing data. The network interface of the computer device is used for communicating with an external terminal through a network connection. The computer program is executed by a processor to implement a risk situation awareness method as described above. Comprising the following steps: the risk situation awareness method comprises a memory and a processor, wherein the memory stores a computer program, and the processor realizes any step in the risk situation awareness method when executing the computer program.

In one embodiment, a computer readable storage medium is provided, having stored thereon a computer program which, when executed by a processor, may implement any of the steps of the risk situation awareness method as described above.

It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

While preferred embodiments of the present application have been described, additional variations and modifications in those embodiments may occur to those skilled in the art once they learn of the basic inventive concepts. It is therefore intended that the following claims be interpreted as including the preferred embodiments and all such alterations and modifications as fall within the scope of the application.

It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.

Claims

1. A risk situation awareness method, comprising:

Performing cluster recognition on the historical operation data aiming at each power node to obtain a cluster result data set containing typical operation states of each power node; the typical operation state refers to a state that the operation parameters of all the power nodes are in a preset normal operation parameter range; performing cluster recognition on the historical operation data aiming at each power node to obtain a cluster result data set containing typical operation states of each power node, wherein the cluster result data set comprises: sampling the historical operation data by adopting a variable grid division density deviation sampling method to obtain a sampling data set; generating a clustering decision graph for each power node based on the sampled data set; determining a neighborhood distance threshold parameter of the clustering decision graph, and carrying out clustering recognition on each historical operation data in the sampling data set according to a minimum neighborhood point parameter to obtain a clustering result data set containing typical operation states of each power node;

2. The risk situation awareness method of claim 1 wherein the sampling the historical operating data using a variable meshing density bias sampling method to obtain a set of sampled data samples comprises:

3. The risk situation awareness method according to claim 2, wherein the density similarity is calculated according to the formula:

；

4. The risk situation awareness method according to claim 2, wherein the probability f that each inter-zone grid is sampled is:

；

5. The risk situation awareness method of claim 1 wherein the generating a cluster decision graph for each power node based on the sampled data set comprises:

6. The risk situation awareness method according to claim 1, wherein the node state prediction model is a transducer model, and the corresponding method takes node real-time index data of each node as input, determines a node prediction state of each power node based on a node state prediction model obtained by training in advance, and obtains a prediction result data set, and includes:

In the converter model encoder, a multi-head self-attention layer converts each real-time index data in an input sequence into a query, a key and a value, calculates attention distribution among the real-time index data in each power node, and obtains fusion information representing each sample point;

7. The risk situation awareness method of claim 6 wherein the multi-headed self-attention layer is calculated as:

；

8. The risk situation awareness method of claim 6 wherein the encoder output has a calculation formula:

；

9. The risk situation awareness method of claim 6 wherein the decoder in the fransformer model further comprises a masked multi-headed self-attention layer whose input data is node real-time index data prior to the current processing sample point.

10. The risk situation awareness method of claim 6 in which the Transformer model employs a piecewise loss function during training.

11. The risk situation awareness method of claim 10 wherein the piecewise loss function is as follows:

；

12. The risk situation awareness method of claim 1, further comprising, prior to the classifying the prediction result dataset and the clustering result dataset based on a classification algorithm model to determine risk situations for each power node:

13. The risk situation awareness method of claim 1 wherein the node real-time metric data for each node further comprises seasonal metric data.

14. The risk situation awareness method according to claim 1, wherein the node real-time index data includes an illumination intensity index and a wind power index; the method further comprises the steps of:

15. The risk situation awareness method of claim 14 wherein the rate of change of the node real-time indicator data is calculated according to the formula:

；

Wherein, Representing the change rate of the real-time index data of the node, which is used for representing the power node in the process ofThe index stability of the moment of time,Representing the real-time radiation intensity per minute during the sampling interval,Is the mean value of the real-time radiation intensity over the sampling interval.

16. The risk situation awareness method of claim 1 wherein the classifying the prediction result dataset and the clustering result dataset based on a classification algorithm model to determine risk situations for each power node comprises:

17. A risk situation awareness system, comprising:

the identification module is used for carrying out cluster identification on the historical operation data aiming at each power node to obtain a cluster result data set containing the typical operation state of each power node; the typical operation state refers to a state that the operation parameters of all the power nodes are in a preset normal operation parameter range; performing cluster recognition on the historical operation data aiming at each power node to obtain a cluster result data set containing typical operation states of each power node, wherein the cluster result data set comprises: sampling the historical operation data by adopting a variable grid division density deviation sampling method to obtain a sampling data set; generating a clustering decision graph for each power node based on the sampled data set; determining a neighborhood distance threshold parameter of the clustering decision graph, and carrying out clustering recognition on each historical operation data in the sampling data set according to a minimum neighborhood point parameter to obtain a clustering result data set containing typical operation states of each power node;

18. A computer device, comprising: comprising a memory and a processor, said memory storing a computer program, characterized in that the processor implements the steps of the method of any of claims 1 to 16 when said computer program is executed.

19. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the method of any of claims 1 to 16.