CN118157132B

CN118157132B - Data mining method and device for voltage monitoring system based on neural network

Info

Publication number: CN118157132B
Application number: CN202410574344.7A
Authority: CN
Inventors: 张德金
Original assignee: Nanjing Yishun Hong Information Technology Co ltd
Current assignee: Nanjing Yishun Hong Information Technology Co ltd
Priority date: 2024-05-10
Filing date: 2024-05-10
Publication date: 2024-08-02
Anticipated expiration: 2044-05-10
Also published as: CN118157132A

Abstract

The invention discloses a voltage monitoring system data mining method and device based on a neural network, which are used for acquiring and preprocessing power distribution network data in a preset area and a time range to obtain preprocessed power distribution network data; acquiring power distribution network data from the power distribution network data, and constructing a power distribution network topology; constructing and training a graph convolutional neural network based on the network topology of the power distribution network; analyzing distribution conditions of voltage out-of-limit data of the power distribution network according to the power distribution network data and the output result of the graph convolution neural network, and clustering all nodes in the power distribution network; based on the clustered power distribution network, performing time causal discovery and space causal discovery, and calculating the average causal effect value of each causal edge. The voltage out-of-limit analysis and tracing speed is improved, and response capability is improved.

Description

Data mining method and device for voltage monitoring system based on neural network

Technical Field

The invention relates to a voltage detection technology, in particular to a data mining method and device of a voltage monitoring system based on a neural network.

Background

The safe and stable operation of the power system is the basis for guaranteeing the normal operation of the modern society. The voltage is one of the key indexes for measuring the quality of the electric energy, and abnormal fluctuation of the voltage can influence the normal operation of the user equipment, and can induce cascading failure and even cause power failure accidents. Therefore, the voltage running state of the power distribution network is monitored in real time, and abnormal conditions such as voltage out-of-limit and the like are found and diagnosed in time, so that the method has important significance in guaranteeing the safe and economic running of the power grid.

Traditional voltage monitoring mainly relies on manual periodic meter reading and threshold alarm setting. This approach is difficult to cope with increasingly complex power distribution environments, and has problems of poor real-time performance, low reliability, and the like. In recent years, under the drive of new technologies such as the Internet of things and big data, a voltage online monitoring system is continuously developed, and a large number of measuring points collect voltage data in real time and transmit the voltage data to cloud centralized management, so that abundant data support is provided for power distribution operation and maintenance. How to mine valuable information from massive, high-dimensional, unstructured raw voltage data in space-time dimensions, and serve anomaly diagnosis and decision optimization of a power distribution network, still faces a plurality of challenges.

Currently, data-driven intelligent voltage anomaly detection methods are receiving extensive attention in academia and industry. Some scholars adopt the traditional machine learning models such as a support vector machine, a random forest and the like, and the classification and identification of typical abnormal modes such as voltage sag, sudden rise, harmonic distortion and the like are realized through manually constructing features. However, the method has the problems of complex characteristic engineering, low calculation efficiency, poor generalization capability and the like. In recent years, deep learning techniques represented by convolutional neural networks and long-term and short-term memory networks have been developed in the fields of computer vision, natural language processing, and the like. Some researchers try to apply the models such as CNN, LSTM and the like to time sequence voltage data modeling, directly learn high-order features in original data, and avoid the limitation of manual feature extraction. However, most of the existing methods are used for modeling aiming at the voltage time sequence of a single measuring point, so that the spatial correlation caused by the topology structure of the power distribution network is ignored, and the voltage running state of the network is difficult to be mastered globally.

In addition, the reasons for the occurrence of the voltage abnormality are complicated, and the voltage abnormality has external factors such as equipment faults, environmental interference and the like, and has internal mechanisms such as load fluctuation, tide transfer and the like. Only the voltage data is considered to be insufficient, and the association analysis is also needed to be carried out by combining multi-source heterogeneous data such as load, tide, equipment state and the like. At present, students explore the correlation between voltage abnormality and other factors by adopting methods such as multivariate time series classification, abnormal point detection and the like, but deep mining from correlation to causality is not yet available. Most research focuses on detection and identification of voltage anomalies, but not enough power assistance is provided on how to support power distribution scheduling decisions. Practical systems should be able to locate the fault point, infer the cause, evaluate the impact, and propose a solution, as well as to find the problem. Knowledge driving and data driving are combined, and an end-to-end intelligent flow from monitoring, analysis to decision is built, so that the problems to be solved are urgently.

Accordingly, research and innovation are needed in order to solve the above-mentioned problems of the prior art.

Disclosure of Invention

The invention aims to provide a voltage monitoring system data mining method and device based on a neural network, so as to solve the problems in the prior art.

According to an aspect of the present application, there is provided a voltage monitoring system data mining method based on a neural network, including the steps of:

s1, collecting and preprocessing power distribution network data in a preset area and a time range to obtain preprocessed power distribution network data;

S2, acquiring power distribution network data from the power distribution network data, and constructing a power distribution network topology; constructing and training a graph convolutional neural network based on the network topology of the power distribution network;

S3, analyzing distribution conditions of voltage out-of-limit data of the power distribution network according to the power distribution network data and the output result of the graph convolution neural network, and clustering all nodes in the power distribution network;

And S4, based on the clustered power distribution network, performing time causal discovery and space causal discovery, and calculating an average causal effect value of each causal edge.

According to another aspect of the present application, there is also provided a voltage monitoring system data mining apparatus based on a neural network, including:

at least one processor; and

A memory communicatively coupled to at least one of the processors; wherein,

The memory stores instructions executable by the processor for execution by the processor to implement the neural network-based voltage monitoring system data mining method of any of the above-described aspects.

The method has the beneficial effects that the speed of voltage out-of-limit analysis and tracing is improved, and the response capability is improved. The related art effects will be described below.

Drawings

Fig. 1 is a flow chart of the present invention.

Fig. 2 is a flowchart of step S3 of the present invention.

Fig. 3 is a flowchart of step S4 of the present invention.

Detailed Description

As shown in fig. 1, a data mining method of a voltage monitoring system based on a neural network is provided, which includes the following steps:

In this embodiment, a network topology diagram is constructed with power distribution stations and switching stations as nodes and lines as edges. On this basis, a multi-layer GCN is designed to learn the spatial characteristics of the voltage data. The GCN can effectively capture the topology information of the node neighborhood by defining convolution operation on the graph structure, so that the spatial dependency relationship among different measuring points is characterized. By considering the topology structure of the distribution network and learning the association mode among the measuring points through the GCN, the running state of the network can be grasped from the global angle through anomaly detection and root cause analysis, and the limitation of the existing method is overcome.

By utilizing the spatial features extracted by GCN and the extracted temporal features, the space-time comprehensive feature vector of each measuring point is obtained, and the time and space fluctuation modes of the voltage are described. On the basis, a clustering algorithm is adopted to divide the measuring point groups with similar modes. Based on the clustering result, further carrying out causal analysis of two dimensions of time and space. In the time dimension, a time sequence causal graph among different physical quantities is constructed by methods such as condition independence test and the like, and the front cause and the back cause of voltage abnormality are revealed. In the space dimension, a dependent structure among the measuring points can be constructed through a PC algorithm, and the direction of the edge is corrected by using priori knowledge, so that a causal topological graph is obtained. Based on the two causal graphs, the average causal effect of each causal edge can be calculated, and the influence of different factors on the voltage abnormality can be quantitatively evaluated. Therefore, on the basis of multi-view feature learning of heterogeneous data, the causative mechanism of voltage abnormality is explored from the two angles of time and space, and deep mining from correlation to causality is realized.

Based on the characteristics extracted by the GCN, the scheme realizes distributed detection and positioning of the voltage out-of-limit of the power distribution network. The method can not only find out voltage abnormality in time, but also identify the area possibly having problems, and strive for time for problem isolation and treatment. The causal analysis can locate the root cause of voltage abnormality, such as regional voltage collapse caused by overload of a certain line, etc. The method provides basis for the dispatcher to diagnose the problem and make countermeasures. Based on the causal graph, the influence of different factor changes on the system can be further simulated, and the severity degree and the influence range of voltage abnormality are evaluated. Furthermore, the quantitative analysis of causality helps to determine the most effective intervention. For example, by comparing the average causal effects of different lines, it can be determined which line's power flow is controlled to improve the local voltage quality problem to the greatest extent. The method realizes end-to-end closed loop from anomaly detection, root cause analysis to scheduling decision, and opens up the whole flow of monitoring, analysis and control.

In a word, through space-time feature fusion and heterogeneous data association analysis, the scheme can furthest excavate the intrinsic value of multi-source data, explore complex influence factors and evolution rules of voltage abnormality, and the analysis of space-time distribution features and influence factors of the voltage abnormality can reflect weak links of a power distribution network in actual operation, so that the fine management level is improved.

As shown in fig. 2, according to an aspect of the present application, step S3 is further:

s31, decomposing an original voltage time sequence into subsequences with different frequency scales by adopting a multi-scale time sequence decomposition method, and acquiring and analyzing fluctuation characteristics of the voltage at different time scales;

Step S32, extracting statistical characteristics of each sub-sequence, wherein the statistical characteristics are taken as time characteristics and comprise a time ratio exceeding a threshold value, a maximum out-of-limit amplitude and an out-of-limit duration;

S33, adopting a graph convolution neural network to extract the spatial characteristics of the voltage on the network topology of the power distribution network;

Step S34, the extracted time features and the spatial features are spliced according to the bits to obtain time-space comprehensive features of the measuring points, a comprehensive feature matrix is formed, and the time fluctuation characteristics and the spatial correlation modes of the voltages at the nodes are analyzed according to the comprehensive feature matrix;

And S35, performing cluster analysis on the comprehensive feature matrix of each measuring point by adopting a DBSCAN algorithm to find out measuring point clusters with similar voltage abnormal space-time features.

The step S31 further includes:

Step S311, reading an original voltage time sequence, and filling zero vectors with preset lengths in the head and the tail of the original voltage time sequence respectively;

Step S312, constructing and adopting a causal convolution layer to extract local features of different time scales, wherein convolution operation of an ith scale is as follows: v ^t,i=σ(∑_j=1 ^Kw_i,jv^t-j+1+b_i), i=1, 2, K;

Where w _i,j is the convolution kernel parameter, b _i is the bias term, σ is the activation function; the convolution kernel parameters satisfy causal constraints: when j > i, w _i,j =0, so that the characteristic of the ith scale depends on only the scale smaller than or equal to i to accord with the time-dependent sequence;

Step S313, a pooling layer realizes dynamic scale aggregation, and calculates attention weight ：α_t,i=（exp（w_a ^Ttanh（W_av^t,i+b_a））/∑_j=1 ^Kexp（w_a ^Ttanh（W_av^t,i+b_a））; of an ith scale at a t moment, wherein w _a,W_a and b _a are parameters of an attention network, and alpha _t,i represents importance degree of the ith scale at the t moment; weighting and summing the K subsequences according to the attention weights to obtain an aggregated sequence;

Step S314, training a neural network with a causal convolution layer, utilizing a trained model to forward propagate an original sequence, and outputting the original sequence at a final causal convolution layer to obtain a desired multi-scale decomposition subsequence.

In this embodiment, causal convolution can adaptively adjust the receptive field of the convolution kernel according to the dynamics of the time series, capturing the fluctuation modes of different time spans. So that the extracted features can better reflect the multi-scale features of the voltage time series. Unlike modeling each scale component independently, this approach introduces causal constraints between scales in causal convolution, requiring that coarser scale features only rely on finer scale features. This causal constraint conforms to the physical law of voltage fluctuation from high frequency to low frequency, so that the extracted features have consistent physical interpretation across scales.

By adaptively extracting the multi-scale fluctuation characteristics of the voltage sequence, the method can accurately describe transient disturbance such as voltage sag and the like, and provides finer-granularity characteristic support for subsequent abnormality detection. By mining causal dependencies among time scales, the method can be used for describing slow time scale change trends of the voltage, such as voltage long-term reduction caused by equipment aging, and the like. This helps to investigate the long-acting influencing factors of voltage anomalies. The time characteristics extracted by the method are input into a graph convolution neural network, so that multi-scale space-time characteristic learning of voltage data can be realized. Not only can the spatial correlation modes on different time scales be revealed, but also the dynamic evolution rule of the spatial correlation can be studied. In the diagnosis of voltage abnormality, it is necessary to judge the duration, severity, etc. of the abnormality in order to formulate a corresponding control strategy. The extraction of the multi-scale temporal features allows for anomaly analysis to be performed at different time spans, including both transient and steady state features. Through the association characteristics of causal convolution learning voltage under different time lags, relevant modes such as periodicity, trend and the like in a voltage time sequence can be discovered, and basis is provided for applications such as anomaly prediction, load adjustment and the like. Compared to pure autocorrelation analysis, multi-scale causal convolution can take into account longer time-distance dependencies, revealing more complex temporal patterns.

As shown in fig. 3, according to an aspect of the present application, the process of performing time causal discovery in step S4 is further:

S41, calling network data of the power distribution network, and constructing a multi-element time sequence data set;

step S42, performing ADF stability test on each row of the multi-element time sequence data, and performing stabilization treatment on the non-stable sequence by adopting differential operation; determining the optimal hysteresis order of any two variables through BIC criteria;

step S43, respectively constructing vector autoregressive models according to variables in the multi-element time series data in sequence, and calculating F statistics;

and S44, constructing a local causal directed graph based on the calculation result, and combining the local causal directed graph into a global multivariate time sequence causal graph.

According to one aspect of the present application, the process of performing spatial causal discovery in step S4 is further:

S45, extracting physical characteristics and data characteristics from the power distribution network data, and selecting representative measuring points from the measuring points to construct a causal discovered node set; wherein each node corresponds to a multidimensional time series; the physical characteristics include electrical distance, and the data characteristics include mutual information;

step S46, invoking a preconfigured algorithm to learn the edge structure of the causal graph, namely initializing a completely undirected graph, connecting edges between any two nodes, and then deleting the edges step by step to obtain a skeleton graph;

step S47, deducing the direction of each undirected edge by utilizing set theory and graph theory knowledge based on the skeleton diagram;

and S48, manually setting the direction of the undirected edge in the pair by using priori knowledge, and then applying a depth-first search algorithm to ensure the non-circularity of the skeleton map, so as to obtain a final space causal map.

According to an aspect of the present application, the step S46 is further:

Step S461, selecting edges with different node depths for each edge in the causal graph, and constructing a node subset;

step S462, for each node subset, testing the condition independence of any two nodes by adopting condition mutual information and bias correlation coefficients; if the two nodes are independent, deleting the edges formed by the two nodes;

Step S463, repeat steps S461 to S462 until all sides have been subjected to the independence test.

According to an aspect of the application, the step S47 is further:

Step S471, reading a skeleton diagram, and searching all unshielded collision structures in the skeleton diagram, wherein the collision structures refer to any three nodes A, B, C, and if A is connected with B and B is connected with C, but A is not directly connected with C, the node B is called as a collision node;

step S472, orienting the edge direction by applying an edge direction orientation rule, the rule including:

If A points in the direction of B and B is connected to C, but A is not connected to C, B should also point to C;

If A points to B, C also points to B, and A has a connection with C itself, then A points to C;

if there is a directed path from node A to node B and there is also an undirected edge between A and B, then the undirected edge can be oriented to point from A to B;

if node A points to node B and there is a third node C such that A is not connected to C, but there is a directional edge between B and C, then B, C's directional edge may be oriented such that C points to B;

step S473, repeat steps S471 and S472 until all orientable edges in the skeleton diagram are oriented.

According to one aspect of the present application, step S47 further comprises:

And step S474, evaluating the stability of each causal edge by repeated sampling, and eliminating edges with the occurrence frequency lower than a threshold value.

In this embodiment, by constructing the vector autoregressive model, the method can characterize the time-dependent relationship between the voltage and other physical quantities. The causal test not only considers the history information of the variable, but also considers the influence of other variables, and can distinguish real time sequence causal and false correlation. ADF inspection and BIC criteria ensure statistical validity of causal analysis. Based on physical and data characteristics, a space causal graph reflecting causal topology of the power distribution network is constructed through condition independence test and edge direction rule. By iteratively deleting edges and orientations, computational overhead can be reduced while the accuracy of the causal structure is ensured. The introduction of a priori knowledge enhances the interpretability of the causal relationship. This enables the learned spatial dependency structure to reveal the propagation mechanism of the electrical quantity in the distribution network, characterizing the spatial impact range of the fault disturbance. And fusing the time causal graph and the space causal graph to form a combined space-time causal model. The analysis of the voltage anomaly causes can be simultaneously developed in two dimensions of time and space, and the analysis comprises evolution rules in different time scales and propagation paths in different space positions. And (3) adopting resampling to evaluate the stability of each causal edge, and eliminating the edge with lower confidence. The stability analysis enhances the robustness of the causal graph to sample disturbance and overcomes the pseudo causal risk caused by limited samples.

By learning causal dependency relations among different lines and devices, the method can predict the affected degree of other lines and loads when one device fails and quantify potential negative effects. The method provides basis for identifying weak links and improving the grid structure, and is beneficial to comprehensively improving the power supply reliability of the power distribution network. Through the spatiotemporal causal graph, the method can infer the time course and the spatial path of the voltage anomaly propagation. By combining the state data acquired in real time, the fault type can be rapidly diagnosed, the fault position can be positioned, the fault severity degree can be estimated, and decision support can be provided for rush repair. By dynamically updating the causal graph and evaluating the average effect of each causal link, the high risk mode of the distribution operation can be identified in a rolling manner and the evolution trend of the high risk mode can be evaluated. By causal tracing, the method can locate key influencing factors of quality problems such as voltage out-of-limit, harmonic pollution and the like and quantify the relative contribution degree of the key influencing factors.

In another embodiment of the present application, the data acquisition and preprocessing process specifically includes:

The distribution automation terminal (such as FTU, DTU and the like) is used for collecting distribution operation data in a period of time (such as 1 month), and the distribution operation data mainly comprises the following steps: voltage data: real-time voltage of each monitoring point, A/B/C three-phase voltage and the like. Current data: current of bus and branch, three-phase current, etc. Power data: active power, reactive power, etc. Topology data: physical topological connection information of the power distribution network. Status data: a startup and shutdown state, a device commissioning state, etc.

Preprocessing the collected original data: duplicate or illegal data records are removed. The missing values are filled in, and methods such as linear interpolation, K nearest neighbor and the like can be used. Smoothing noise data, such as using moving average, wavelet transform, etc. The offset value and the abnormal value are detected and corrected.

The multi-source heterogeneous data are aligned according to time stamps and integrated into a unified data view. It is necessary to ensure that data of different acquisition frequencies (e.g. seconds, minutes) can be correctly aggregated.

And carrying out operations such as normalization and standardization on the data according to the input requirements of the model. Normalization may map data to the [0,1] interval using maximum-minimum normalization; normalization data can be converted to a distribution with a mean of 0 and standard deviation of 1 using Z-score. Not all acquired data is effective for modeling and it is necessary to screen out key features with differentiation. Statistical indicators (e.g., variance, correlation coefficients) or machine learning methods (e.g., lasso regression, decision trees) can be used to evaluate the importance of features, top-N features are selected for modeling.

In another embodiment of the present application, the process of clustering measurement points using the DBSCAN algorithm is specifically as follows:

After the space-time comprehensive characteristics of each measuring point are obtained, the characteristic matrix is required to be normalized according to columns, and the characteristics of different dimensions are scaled to similar dimensions so as not to influence the clustering result.

DBSCAN is based on density clustering, and similarity among samples needs to be calculated. For extracted high-dimensional features, cosine similarity can be used;

the performance of DBSCAN is very sensitive to the parameters epsilon and minPts. Epsilon controls the neighborhood radius of the samples, minPts controls the minimum neighborhood number of samples that become the core object. And (3) traversing a plurality of groups of parameter combinations with optimal indexes such as profile coefficients on a 2D parameter plane by adopting grid search.

In another embodiment of the present application, further comprising:

Constructing a network topological graph: and abstracting a topological structure undirected graph G= (V, E) of the power distribution network by taking the power distribution transformer substation and the switch station as nodes and the line as edges. V= { V1, V2,..vn } is a node set, E is an edge set.

S22: constructing a graph convolution neural network: designing an L-layer graph convolution neural network. The forward propagation rule of the first layer is:

H^l+1=σ(D*^（-1/2）A* D*^（-1/2）H^lW^l；

Where a=a+i is a adjacency matrix a plus self-connection, D is a degree matrix of a, H ^l is a feature matrix of the layer I node, W ^l is a learnable weight matrix, and σ is an activation function, for example.

S23: constructing a space-time attention layer: a self-attention layer is introduced after the GCN to adaptively learn the spatio-temporal correlation weights between nodes. Z=attention (Q, K, V) =softmax (QK ^T/sqrt（d_k)) V;

q, K, V are query matrix, key matrix, value matrix, respectively, which are obtained by GCN output H { (L) } linear transformation. d_k is the dimension of K. The Attention output Z is a weighted node characteristic representation.

S24: predicting voltage anomalies: and transmitting the attention layer output Z into a feedforward neural network, and outputting a voltage predicted value v _i ^t+1 at the next moment of each measuring point. The voltage residual r _i ^t+1=|v_i ^t+1-v*_i ^t+1 is used for measuring the deviation degree of the actual value and the predicted value, and if the deviation degree exceeds a preset threshold value, the abnormality is judged.

In another embodiment of the application, the implementation of the causal convolution layer is as follows:

Input layer: the raw one-dimensional voltage time series data is shaped as (seq _len, 1), where seq _len is the time step. To ensure time-shift invariance, zero vectors of length (kernel _size -1) are filled in the beginning and end of the sequence respectively.

Causal convolution layer: this layer is the core of the causal convolutional neural network, whose convolution operation formula is shown above,

The receptive field of the network is K, i.e. one output feature is related to the inputs of the last K time steps. By superimposing multiple causal convolution layers, the receptive field can be enlarged and longer time dependence modeled.

Attention pooling layer: this layer applies a focus mechanism on the causal convolution feature map of multiple scales, adaptively aggregating features of different time scales.

Specifically, assuming the first layer causal convolution outputs Kl feature graphs v ^t,i _i=1 ^Kl, the attention pooling layer first calculates the attention weights for each feature graph (as shown above)

Then, the attention pooling layer sums the Kl feature maps by weight to obtain an aggregate feature u ^t:

u^t=∑_i=1 ^Klα^t,iv^t,i；

In another embodiment of the application, spectral clustering is employed. Spectral clustering utilizes eigenvalues and eigenvectors of the laplace matrix of samples to perform clustering, and can identify non-convex clusters. The method mainly comprises the following steps:

And calculating a similarity matrix W of the sample to obtain a corresponding degree matrix D and a Laplacian matrix L=D-W.

And carrying out feature decomposition on the L, and taking feature vectors corresponding to the first k minimum non-zero feature values to form a feature matrix U.

And carrying out K-means clustering on each row of U to obtain final cluster division.

In another embodiment of the present application, the causal inference and root cause analysis process is specifically:

A causal test, such as a Granger causal test, is applied to the multivariate time series (e.g., voltage, current, power factor, etc.) at each measurement point to infer the chronological dependencies between the variables. If the past value of variable X significantly contributes to predicting the current value of Y, then X is said to be the Granger cause of Y.

And constructing a Bayesian causal graph by taking the measuring points as nodes and the causal relationship as directed edges. The conditional probability distribution of the directed edges is learned from the data, and the causal learning method, such as a causal structure learning method of a PC algorithm, is used for judging the edge direction, so as to determine the topological structure of the causal graph.

For each causal edge xi→xj, the Average Causal Effect (ACE) of Xi on Xj is measured:

ACE_Xi→Xj=E[Xj|do(Xi=xi¹)]- E[Xj|do(Xi=xi⁰)]]；

do (xi=xi ^k) represents intervention on variable Xi, set to Xi ^k. ACE measures the variation of Xj caused by one unit of variation of Xi under the condition of controlling other variables, so as to evaluate the influence of the reason node on the result node.

S34: root cause localization and risk assessment: when detecting that a voltage abnormality occurs at a certain measuring point, carrying out probability reasoning in the causal graph. Considering all possible combinations of causes Pak, the posterior probability for each combination given the voltage anomaly evidence Evid is calculated based on the bayesian theorem:

P (pak| Evid) = (P (Evid |pak) P (Pak) }/(Σ _i P (Evid |pai) P (Pai))); and selecting the reason combination with the maximum posterior probability as a diagnosis result. And meanwhile, evaluating risk values (such as influences on power supply reliability) under the combined conditions of all reasons, and making an emergency plan.

In another embodiment of the application, one example is given as follows: for example, a power distribution company in a certain city monitors that partial area voltages of A, B, C10 kV power distribution lines in the district are in a low state for a long time, so that the power supply quality of a user is affected, and a reason needs to be found as soon as possible and a solution is formulated. The area is provided with 50 ring main units, each ring main unit is provided with an intelligent electric power instrument, and operating parameters such as voltage, current, power and the like are collected in a period of 1 minute. Meanwhile, the distribution network topology, line parameters, user information and the like are recorded into the distribution geographic information system.

Step 1: data acquisition and preprocessing

The power distribution automation master station is used for collecting power distribution operation data of 2022, 1 month and 1 day to 2023, 1 month and 1 day for one year, and the power distribution operation data mainly comprise voltage, current and power data of 50 measuring points on A, B, C lines, on-off states and the like. The original data is cleaned to remove duplicate or illegal recordings and aligned according to the time stamps into a unified time series data set. Interpolation complement is carried out on the missing data, and obvious abnormal data are corrected. And normalizing each physical quantity to the [0,1] interval, and eliminating the dimension influence.

Step 2: graph convolution neural network construction

And constructing a network structure diagram taking 50 measuring points as nodes and a line as an edge according to topology information provided by the GIS system.

Designing a 3-layer graph convolution neural network, taking normalized voltage, current and power data as node characteristics, and training through a semi-supervised classification task. The semi-supervised is to use the historical voltage anomaly labels (marked by the backtracking property of an operation and maintenance expert) of a small number of measuring points as part of training samples, and simultaneously to use a large number of unlabeled samples to perform self-supervised learning, and extract the voltage operation state of the whole network in an end-to-end mode for embedding.

The model uses Adam optimizer, initial learning rate 0.001, 10% lower learning rate per 50 rounds, training 500 rounds.

Step 3: voltage anomaly spatiotemporal analysis

Spatial dimension: and (3) using the trained GCN model to perform time-by-time reasoning on all the measuring points to obtain the abnormal probability distribution of the measuring points. And clustering the measuring points with high abnormality probability (such as top 20%) by using a DBSCAN algorithm, wherein each cluster represents an abnormal region. And carrying out visual display on the abnormal region in the GIS system, analyzing the spatial distribution characteristics of the abnormal region, and primarily judging the relevance between the abnormal region and the geographic environment and the peripheral equipment.

Time dimension: and extracting voltage data of each measuring point for one year to form a time sequence for the clustered abnormal areas.

The original voltage sequence is decomposed into a pattern of fluctuations on different time scales using an adaptive time series decomposition method, mainly into long-period components (reflecting the trend of the slow time scale), short-period components (reflecting the fluctuation of the fast time scale) and random noise. Analyzing the evolution trend of the long period component, focusing on the inflection point of the long-term voltage drop trend, and deducing possible policy factors or equipment aging factors. The fluctuation intensity of the short period component is analyzed, the occurrence time and severity of the instantaneous impact are identified, and possible fault factors or abnormal loads are deduced.

Step 4: voltage anomaly spatiotemporal causal analysis

Time causal analysis: and (3) constructing a multi-element time sequence data set comprising a plurality of physical quantities such as voltage, current, power, temperature and the like aiming at the measuring points of the abnormal region in the step (3). And constructing a time causal graph among the physical quantities by adopting a stationarity test and a Granger causal test, and revealing an advanced early warning factor of voltage abnormality. And calculating the causal strength, and evaluating the contribution degree of different physical quantities to the voltage abnormality.

Spatial causal analysis: and (3) further refining the space causal topology of the abnormal region on the basis of the space clustering in the step (3). And learning a dependent structure among abnormal measuring points from the data by adopting a PC stability algorithm and a minimum description length criterion, and carrying out directional correction on edges according to the power distribution professional knowledge to finally obtain a Directed Acyclic Graph (DAG). On the basis of DAG, calculating average causal effect, and quantitatively analyzing the influence mechanism of key measuring points in an abnormal area.

Step 5: analysis result output

And outputting a visual display of the space-time abnormal region, wherein the visual display comprises a spatial distribution diagram and a multi-scale decomposition diagram of a time sequence in a GIS system. And outputting a space-time causal relation graph, wherein the space-time causal relation graph comprises a time causal graph of the key physical quantity and a space DAG graph of the key measuring point, and labeling the quantitative contribution degree. An anomaly diagnostic report is formed listing key influencing factors of voltage anomalies, including problematic lines/equipment, time of occurrence, severity, etc., and giving a preliminary correction scheme.

Analysis of results: through the analysis of the voltage abnormality, the power distribution network is found to have the following problems:

Spatial dimension: the voltage abnormality is mainly concentrated in the southeast corner of the line A and the middle part of the line C, and obvious space aggregation is presented. In combination with GIS information, the two abnormal areas are just in an industrial park and a newly built district, the load density is high, and the abnormal areas are presumed to be related to overload operation. Time dimension: the voltage data of the abnormal area shows a slow descending trend for a long time, and two obvious jumps appear in 6 months and 11 months of 2022, which are consistent with the two-stage construction and production time of the industrial park. At the same time, the voltage fluctuation is obvious in the morning and evening peak time of each day, and the voltage fluctuation should be related to the load curve of a park and a district. Causal analysis: the voltage in the abnormal region is most correlated with the advance of the three-phase imbalance and the current fluctuation rate in time, and is a precursor signal of the voltage abnormality. In space, the end measuring point of the line with the longest power supply radius is a key influence node, and has remarkable conduction effect on the voltage abnormality of the peripheral area.

Based on the above analysis, it is suggested that: and carrying out load assessment on the abnormal region, and if necessary, increasing capacity and upgrading the transformer and the circuit to match the load increase requirement. And (3) throwing a power quality control device such as SVG and the like into the abnormal region, balancing the three-phase load in real time, and inhibiting current fluctuation. And the power supply topology of a park and a community is optimized, capacitance compensation and line upgrading are properly added at the terminal measuring point, and the power supply reliability is improved. And an abnormality early warning mechanism of indexes such as three-phase unbalance degree, current fluctuation rate and the like is established, and an abnormality treatment time window is advanced. And carrying out space-time causal analysis in a normalized mode, dynamically sensing the running state of the power grid, and guiding the precise planning.

In another embodiment of the present application, the LSTM model may also be used for processing, specifically as follows:

Determining the number and the hierarchy of LSTM model units, wherein each LSTM model unit comprises an input gate, a forgetting gate and an output gate, and the gates control the inflow, the reservation and the outflow of information so as to determine the basic structure of a network and initialize the network structure;

adjusting the dimension and length of the preprocessed data, and performing feature scaling to convert the preprocessed data into a sequence which can be processed by the LSTM model, and acquiring input data to ensure that the input data is suitable for model processing;

defining an LSTM layer in an LSTM model, and setting the weights of an activation function and each gate to determine the long-term dependence of the LSTM model on data processing and learning;

Introducing an attention mechanism layer behind the LSTM layer, and assigning weights for the output of the time steps so as to improve the attention capability of the LSTM model to the key information;

setting a loss function for evaluating the difference between the predicted and actual values of the LSTM model, and setting an optimizer for updating the weight of the LSTM model to improve the prediction performance;

training the LSTM model using the training set data and learning patterns and relationships in the data by a back propagation algorithm;

Evaluating LSTM model performance by using the verification set, and adjusting model parameters according to the evaluation result to perform model evaluation and tuning;

After the LSTM model is trained, the structure and the weight of the LSTM model are saved so as to be deployed to a voltage detection system for real-time data mining and analysis.

The method comprises the steps of selecting a long-short-term memory network LSTM, establishing an LSTM model, processing high-dimensional data in a voltage detection system through hierarchical feature extraction, enabling the model to extract meaningful features from original data, and further processing and analyzing through a subsequent hierarchical structure, so that the data processing efficiency and accuracy of the system are improved; the long-term memory network LSTM has strong robustness and generalization capability, can accurately predict and identify unseen data samples, and can accurately predict and respond to abnormal operation or abnormal conditions of the system; the LSTM model can effectively process large-scale data and improve data processing efficiency.

at least one processor; and

A memory communicatively coupled to at least one of the processors; wherein,

The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to the specific details of the above embodiments, and various equivalent changes can be made to the technical solution of the present invention within the scope of the technical concept of the present invention, and all the equivalent changes belong to the protection scope of the present invention.

Claims

1. The data mining method of the voltage monitoring system based on the neural network is characterized by comprising the following steps of:

S4, based on the clustered power distribution network, performing time causal discovery and space causal discovery, and calculating an average causal effect value of each causal edge;

the process of performing the time causal discovery in step S4 is further:

Step S44, constructing a local causal directed graph based on a calculation result, and combining the local causal directed graph into a global multivariate time sequence causal graph;

the process of performing the spatial causal discovery in step S4 is further:

step S48, manually setting the direction of the undirected edge in the pair by using priori knowledge, and then applying a depth-first search algorithm to ensure the non-circularity of the skeleton map to obtain a final space causal map;

The step S46 is further:

Step S463, repeating steps S461 to S462 until all edges are subjected to the independence test;

the step S47 is further as follows:

If a directed path exists between node A and node B, and an undirected edge exists between A and B, the undirected edge is oriented to point from A to B;

If node A points to node B and there is a third node C such that A is not connected to C, but there is a directional edge between B and C, then the directional edge of B, C is oriented to C point to B;

step S473, repeating steps S471 and S472 until all orientable edges in the skeleton diagram are oriented;

2. The neural network-based voltage monitoring system data mining method according to claim 1, wherein step S3 is further:

3. The method for mining voltage monitoring system data based on neural network according to claim 2, wherein the step S31 further comprises:

4. A neural network-based voltage monitoring system data mining apparatus, comprising:

at least one processor; and

A memory communicatively coupled to at least one of the processors; wherein,

The memory stores instructions executable by the processor for execution by the processor to implement the neural network-based voltage monitoring system data mining method of any one of claims 1 to 3.