CN117761802A - Wind farm weather forecast algorithm based on federal learning - Google Patents
Wind farm weather forecast algorithm based on federal learning Download PDFInfo
- Publication number
- CN117761802A CN117761802A CN202311532774.4A CN202311532774A CN117761802A CN 117761802 A CN117761802 A CN 117761802A CN 202311532774 A CN202311532774 A CN 202311532774A CN 117761802 A CN117761802 A CN 117761802A
- Authority
- CN
- China
- Prior art keywords
- model
- data
- federal learning
- weather
- weather forecast
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012549 training Methods 0.000 claims abstract description 40
- 238000012545 processing Methods 0.000 claims abstract description 16
- 238000012216 screening Methods 0.000 claims abstract description 4
- 238000000034 method Methods 0.000 claims description 15
- 238000000605 extraction Methods 0.000 claims description 12
- 238000012795 verification Methods 0.000 claims description 12
- 238000011156 evaluation Methods 0.000 claims description 9
- 238000010200 validation analysis Methods 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 238000003066 decision tree Methods 0.000 claims description 4
- 238000012360 testing method Methods 0.000 claims description 4
- 238000001514 detection method Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000001502 supplementing effect Effects 0.000 claims description 3
- 230000000694 effects Effects 0.000 description 16
- 230000002776 aggregation Effects 0.000 description 5
- 238000004220 aggregation Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 210000004027 cell Anatomy 0.000 description 4
- 238000012423 maintenance Methods 0.000 description 4
- 230000005540 biological transmission Effects 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000010248 power generation Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 208000025274 Lightning injury Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 230000008878 coupling Effects 0.000 description 1
- 238000010168 coupling process Methods 0.000 description 1
- 238000005859 coupling reaction Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000013468 resource allocation Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a wind farm weather forecast algorithm based on federal learning, which relates to the field of wind farm weather forecast, and comprises a weather forecast module, wherein the weather forecast module comprises a data processing module and a federal learning module; the weather prediction module comprises the following specific steps: step one, acquiring historical meteorological data of a multi-area meteorological site; step two, the historical weather is locally preprocessed by a data processing module; step three, constructing an iterative model by the federal learning module for iterative updating; and fourthly, screening the federal learning model with satisfactory performance as a final model to predict. Model training is directly completed by utilizing dimensionless data in local by utilizing federal learning, and then local model parameters are respectively weighted and aggregated, so that a global model capable of comprehensively learning global weather changes is obtained, and the model is enabled to better understand the mode and the association between weather data by combining data of cross sites, so that the accuracy and the reliability of a prediction model are improved.
Description
Technical Field
The invention relates to the field of wind farm weather prediction, in particular to a wind farm weather prediction algorithm based on federal learning.
Background
The application of the meteorological monitoring and forecasting technology in the field of wind power generation almost covers all stages of the full life cycle of a wind farm. From the beginning of site selection, a pre-selected area needs to be thoroughly researched and evaluated, future wind energy development potential of the area is predicted, and meanwhile, the corrosion degradation of the turbine materials, the influence of rain and snow factors on the blade rotation performance, the potential influence of damage caused by lightning stroke and scheduling of fault operation and maintenance work, which are caused by local conditions such as illumination radiance, ultraviolet intensity and rainfall condition, are considered, and the scheduling of fault operation and maintenance work is required to be dealt with by making strategies in advance based on weather prediction results. In addition, the most critical factor influencing the generation of the wind turbine generator is wind speed, under normal generation, the output power of the wind turbine generator is almost proportional to the cube of the wind speed, and the intermittent fluctuation of the wind speed and wind direction factors also influence the stable operation of the power grid. Therefore, weather conditions influence the instant output of the wind driven generator and the maintenance requirement and the overall service life of the wind driven generator, so that the research and development of a high-precision weather prediction technology with regional pertinence for a wind power plant is particularly important;
the current wind power plant meteorological data prediction mostly depends on an NWP meteorological prediction method requiring a high-performance computer, and wind energy conditions of a wind power plant are simulated again by collecting meteorological data of a meteorological station, or meteorological factors are predicted based on collected data of a wind power plant and a meteorological station which are built by the wind power plant.
However, although the accumulation speed of meteorological data is increased to thousands of TB level from the original MB level monthly nowadays, abnormal data is difficult to process due to the loss of the data per se in time period, a large amount of data is not reasonably utilized after being acquired, and especially, the phenomenon of data island that the data of an area is not connected with the area exists universally; the same is true from region to region of the wind farm, and the problems of data acquisition methods, differences in unit standards, differences in weather distribution in space and the like all make the integration and comparison of data very difficult. Therefore, we propose a wind farm weather forecast algorithm based on federal learning to solve the technical problems presented above.
Disclosure of Invention
In view of the above, the main objective of the present invention is to provide a wind farm weather forecast algorithm based on federal learning, which aims to solve the above problems.
In order to achieve the above purpose, the present invention adopts the following technical scheme: the system comprises a weather prediction module, wherein the weather prediction module comprises a data processing module and a federal learning module;
the weather prediction module comprises the following specific steps:
step one, acquiring historical meteorological data of a multi-area meteorological site;
step two, the historical weather is locally preprocessed by a data processing module;
step three, constructing an iterative model by the federal learning module for iterative updating;
step four, screening a federal learning model with satisfactory performance as a final model to predict;
the data processing module comprises the following steps:
s101, analyzing the missing condition of data, and deleting variables with excessively high missing degree based on a missing degree threshold value; filling variables with fewer missing values, and supplementing the missing values which are single-value missing and continuous missing by using nearest neighbor interpolation and decision tree interpolation respectively to ensure the data integrity;
s102, recoding at least one mode of single-heat coding, tag coding or custom numerical coding can be adopted for discrete variable data;
s103, converting the data into uniform scales;
s104, constructing a neural network layer for continuous variables to extract features;
the federal learning module includes the steps of:
s201, training a local model;
s202, after the local model training is completed, the local model of each participant is sent to a central server, and the central server aggregates the parameters of all the local models into a global model;
s203, evaluating the performance of the global model by using the test data set to ensure that the model can accurately predict.
In a preferred embodiment, in the step S101, the missing detection is performed on the collected data, the variable whose missing rate exceeds 45% is deleted, and the missing value filling is performed on the remaining variables.
As a preferred embodiment, in the step S102, the missing discrete variable is mainly filled in the forward direction, and the continuous variable is filled in the mean value.
In a preferred embodiment, in S103, the continuous variable is separated from the wind direction discrete variable, wind direction data is recoded, angle data of 0 ° to 360 ° is converted, and the continuous variable is normalized, and then the next feature extraction is performed.
In a preferred embodiment, in the step S104, the network such as the fully connected network, LSTM, biLSTM, etc. is selected to extract the high-dimensional features of the data, and the local model is trained together with the original discrete data after being spliced again.
As a preferable scheme, for continuous variables, a two-layer LSTM is adopted as a characteristic extraction network to acquire high-dimensional time sequence characteristics of the continuous variables, the original continuous variable sequence is predicted in advance by training a network model with two layers of LSTM and one layer of full-connection layer, and the iteration times are set to be 100;
the fitted model is reserved with parameters of two layers of LSTM networks, original meteorological data are input into the model, the expanded high-dimensional characteristics are extracted, the final extraction dimension is set to be 32, the characteristics and discrete variables are combined, data are segmented into a training set and a verification set through a leave-one-method, and the training set is input into a federal learning model for training.
In a preferred embodiment, in the step S201, the local model updates the model parameters based on the gradient descent method, and R is selected 2 As an evaluation criterion of the model, R predicted on a verification set after local model training is saved 2 And calculating a value.
As a preferred solution, in the step S202, an update formula for the global model is as follows:
wherein θ global Representing parameters, θ, of the global model i Local model parameters representing the ith party, w i Representing the weight of the i-th participant,representing +.o. of the ith participant local model on the verification set>A calculated value, N, representing the number of participants;
which is based on R on the validation set for each party's local model 2 Calculating value to weight average parameter of aggregate global model, weight w i By R of each participant 2 And (5) value determination.
In a preferred embodiment, in the step S203, the iteration returns to the step S201, and the global model is issued back to the local to continue training until the global model converges or the maximum number of iterations is reached.
In the first to fourth steps, the federal learning model with satisfactory performance is put into weather prediction, new data is input into the model, and a multi-step prediction result can be obtained.
Compared with the prior art, the invention has obvious advantages and beneficial effects, and particularly, the technical scheme mainly comprises the following steps;
according to the method, model training is directly completed by utilizing dimensionless data through federal learning, and then local model parameters are respectively weighted and aggregated, so that a global model capable of comprehensively learning global weather changes is obtained, and the model is enabled to better understand the mode and the association between weather data by combining data of cross sites, so that the accuracy and the reliability of a prediction model are improved.
In order to more clearly illustrate the structural features and efficacy of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and examples.
Drawings
FIG. 1 is a block diagram of a weather forecast technique according to an embodiment of the present invention;
FIG. 2 is a general flow chart of weather forecast according to an embodiment of the present invention;
FIG. 3 is a graph of the predicted effect of an Ff validation set in accordance with an embodiment of the invention;
FIG. 4 is a diagram of DD verification set prediction effects according to an embodiment of the present invention;
FIG. 5 is a graph of the effect of variable T prediction in an embodiment of the present invention;
FIG. 6 is a graph of the predicted effects of variable Po according to an embodiment of the present invention;
FIG. 7 is a graph of the effect of variable P prediction in an embodiment of the invention;
FIG. 8 is a graph of the predicted effect of variable Pa in an embodiment of the invention;
FIG. 9 is a diagram of the effect of variable DD prediction in accordance with an embodiment of the present invention;
FIG. 10 is a graph of the predicted effect of variable Ff in an embodiment of the invention;
FIG. 11 is a graph of the predicted effect of variable ff3 in an embodiment of the invention;
FIG. 12 is a graph of variable Tx prediction effect according to an embodiment of the present invention;
fig. 13 is a view showing the effect of variable VV prediction in the embodiment of the present invention.
Detailed Description
For the purpose of making the technical solution and advantages of the present invention more apparent, the present invention will be further described in detail below with reference to the accompanying drawings and examples of implementation. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
It will be understood that when an element is referred to as being "fixed to" another element, it can be directly on the other element or intervening elements may also be present. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or intervening elements may also be present. The terms "vertical," "horizontal," "left," "right," and the like are used herein for illustrative purposes only.
Referring to fig. 1 to 13, an embodiment of the present invention provides a wind farm weather forecast algorithm based on federal learning, including a weather forecast module, where the weather forecast module includes a data processing module and a federal learning module;
the weather prediction module comprises the following specific steps:
step one, acquiring historical meteorological data of a multi-area meteorological site;
step two, the historical weather is locally preprocessed by a data processing module;
step three, constructing an iterative model by the federal learning module for iterative updating;
and fourthly, screening the federal learning model with satisfactory performance as a final model to predict.
The data processing module comprises the following steps:
s101, analyzing the missing condition of data, and deleting variables with excessively high missing degree based on a missing degree threshold value; filling variables with fewer missing values, and supplementing the missing values which are single-value missing and continuous missing by using nearest neighbor interpolation and decision tree interpolation respectively to ensure the data integrity;
s102, recoding at least one mode of single-hot coding, label coding or custom numerical coding can be adopted for discrete variable data, and corresponding coding modes are adopted for different types of data such as weather change degree, order and disorder according to the meaning of the variable so as to ensure that a predictable data format is obtained;
s103, converting the data into uniform scales for subsequent processing;
s104, constructing a neural network layer for feature extraction for continuous variables, specifically selecting a fully connected network, an LSTM (least squares), a BiLSTM (BiLSTM) and other networks to extract high-dimensional features of data, and performing local model training after re-splicing with original discrete data;
the module is mainly used for preprocessing meteorological data of various areas, a framework is a horizontal (transverse) federal learning model framework, and the same characteristics are required to be shared during model training, so that the data are firstly required to be cleaned and preprocessed, the alignment of the characteristics is realized so as to facilitate model training, the basic meteorological data are considered to comprise continuous data such as temperature, humidity, air pressure, air speed, air direction, cloud layer data and the like, and further discrete category information related to cloud layers, rainfall and snowfall, and short text description such as weather conditions, climate change and the like are provided with complex and various formats.
The federal learning module includes the steps of:
s201, training a local model, and training the local model: after the model is selected, as the data sets are not shared, all the participants receive the locally processed data and begin to train after initializing the model parameters, the local model can update the model parameters based on a gradient descent method, and R is selected 2 As an evaluation criterion of the model, R predicted on a verification set after local model training is saved 2 Calculating a value;
s202, updating a global model: after the local training is completed, the local model of each participant is sent to a central server, and the central server aggregates the parameters of all the local models into a global model, wherein the scheme is based on R of the local models on a verification set 2 Carrying out weighted average aggregation on the parameters of the model, and updating the parameters of the global model;
the update formula for the global model is as follows:
wherein θ global Representing parameters, θ, of the global model i Local model parameters representing the ith party, w i Representing the weight of the i-th participant,representing +.o. of the ith participant local model on the verification set>A calculated value, N, representing the number of participants;
which is based on R on the validation set for each party's local model 2 Calculating value to weight average parameter of aggregate global model, weight w i By R of each participant 2 The value determines, and therefore, the better the local model of the participant performs on the validation set, the greater its contribution to the global model parameters.
S203, evaluating the performance of the global model by using the test data set to ensure that the model can accurately predict, and evaluating the performance of the global model by using the test data set to ensure that the model can accurately predict. Repeating the iteration and returning to the step S201, issuing the global model back to the local and continuing to train until the global model converges or the maximum iteration number is reached;
finally, putting the federal learning model with satisfactory performance into weather prediction, and inputting new data into the model to obtain a multi-step prediction result;
meanwhile, to improve the performance of the model, the global model may be updated periodically and the performance re-evaluated after the update, and if the performance of the model is not improved or reduced, the training strategy may be adjusted, such as a loss function replacement to better reflect the data distribution, or the infrastructure model architecture may be re-selected.
In some embodiments, considering that the weather utilized by a general wind farm mainly originates from peripheral weather stations, the experiment collects the measured data of 16 weather stations in a certain province, the data set includes weather data during eight months of each weather station, and after the weather variable with higher quality is screened out, the data set includes 9 weather elements such as atmospheric temperature, air pressure, wind speed, wind direction, horizontal visibility and the like.
The predictable variables are shown in Table one below
List one
Firstly, carrying out missing detection on acquired data, deleting variables with the missing rate exceeding 45%, and carrying out missing value filling on the rest variables, wherein the missing of discrete variables mainly adopts forward filling, and continuous variables adopt mean filling; and separating continuous values such as air pressure, air temperature, wind speed and the like from wind direction discrete variables, recoding wind direction data, converting angle data of 0-360 degrees, normalizing the continuous variables, and performing the next feature extraction.
For continuous variables, a two-layer LSTM is adopted as a feature extraction network to acquire high-dimensional time sequence features of the continuous variables, the original continuous variable sequence is predicted in advance by training a network model with two layers of LSTM and one layer of full-connection layer, and the iteration times are set to be 100;
as shown in fig. 3-4, it can be observed that the fitting effect of a single LSTM model on meteorological variables is still insufficient, the model needs to be further deepened, the learning of the model on the sequence is enhanced, the fitted model retains the parameters of two layers of LSTM networks, original meteorological data are input into the model, the high-dimensional characteristics after expansion are extracted and obtained, the final extraction dimension is set to be 32, the characteristics and discrete variables are combined, the data are segmented into a training set and a verification set by a leave-one-method, and the training set is input into a federal learning model for training;
and after receiving the combined characteristic variables, respectively initializing global model parameters and local model parameters. The model here has a uniform structure, the core comprising four layers of BiLSTM network, full connectivity layer, dropout layer and finally output layer. Model training parameters are shown in Table II below
Watch II
After initializing model parameters, the local model starts to train by combining local participation local data respectively, and in the training process, the data firstly passes through a first BiLSTM layer, and the layer extracts initial characteristic association;
these features are then passed to a second tier of BiLSTM network, which is composed of two tiers of BiLSTM networks in parallel, which receives the features extracted from the previous tier and further processes to obtain a more complex, deep representation;
the third layer BiLSTM performs deeper extraction and coupling on the features again to acquire richer feature information;
after the three layers of feature extraction and processing, the data then passes through a full-connection layer and a Dropout layer, and the random loss rate is set to be 0.2, and the Dropout layer can randomly and temporarily close a part of neurons in the training process, so that the model is effectively prevented from being excessively dependent on certain specific neurons, and the overfitting is avoided;
the federal learning utilizes distributed computation to disperse model training tasks on a plurality of nodes, so that computing resources can be fully utilized, training efficiency is improved, in weather prediction, the federal learning can utilize data of a plurality of regional weather stations to simultaneously perform model training, and a more robust and accurate global model is generated by combining training results of the nodes.
After the local model is trained, the verification set reserved by combining the local data is used for verifying the model, and R is reserved 2 Uploading the model parameters to a central server to participate in parameter aggregation of the global model, namely, the model parameters are evaluated according to an evaluation index R 2 Carrying out weighted summation and updating the global model;
based on the evaluation index R 2 The local model parameters are aggregated, so that the participation of each local model to the global model can be more effectively distributedDegree and reduce the impact of low quality models on the results by using R 2 The evaluation index adjusts the weights of different local models, and the federal learning system can automatically screen out models and parameters with better performance, so that the quality and accuracy of the global model are improved, and for weather prediction, federal learning can better utilize the data of a distributed weather site, eliminate the interference of a low-quality model and improve the overall prediction performance.
The global model is evaluated using the goodness of fit (R 2 ) The model prediction effect is evaluated by four indexes of Mean Square Error (MSE), root Mean Square Error (RMSE) and Mean Absolute Error (MAE), and the formula is as follows:
wherein y is i Andr is respectively the actual variable value and the federal learning model predicted value 2 The closer the value is to 1, the better the model performance, the smaller the MSE, the RMSE, and the smaller the MAE, the model prediction error.
Predicting by using a final global model, wherein the finally obtained evaluation indexes are shown in the following table III:
watch III
As shown in FIGS. 5-13, the weather data of the S54736 site is taken out to conduct weather rolling prediction 3 hours in advance, and compared with the actual value, the model can be observed to realize better prediction basically according to the result, and only larger errors exist on the wind direction, the horizontal visibility influenced by the change of cloud layer and other variables with more severe changes, but R above 0.78 can be realized 2 Is a predictive effect of (a).
Simultaneously comparing the global model to the global goodness of fit (R) of the meteorological data of 16 meteorological sites 2 ) Table IV below
Table four
As a result, it was found that the prediction effect of the global model on each region was also excellent, wherein the prediction effect on the region of the S54776 site was excellent, R 2 The minimum evaluation score can reach more than 0.93 and also reach 0.86 on the data set, which indicates the effectiveness of the federal learning model for weather prediction and the feasibility of the scheme for predicting future changes of weather.
In addition, in federal learning, the selection of the base model can be adjusted according to the type of the prediction variable so as to achieve better prediction performance, and when the prediction processing of a large number of weather texts is carried out, the replacement base model can be considered to be a classifier with good performance such as a decision tree, a support vector machine and the like, and the models have excellent performance in the aspect of processing text data; for continuous variable prediction, models such as LSTM or BiLSTM can be selected.
Based on the federal learning technology, the method can participate in the update of the global model only by uploading trained local model parameters, so that the problems of possible deletion, error and the like in the regional data transmission process are effectively avoided, the risks of data transmission and privacy disclosure are reduced while the original data are kept locally, and the data security and privacy protection level are improved;
because federal learning can fuse and summarize the models of a plurality of participants, a more comprehensive and accurate model can be obtained, so that the generalization capability of the model is improved, and in weather prediction, weather data and model parameters of different areas can be integrated by using federal learning, so that a more accurate and reliable weather prediction result is provided;
meanwhile, the scheme has obvious economic value: the federal learning can reduce the operation and maintenance cost of enterprises by reducing the cost of data transmission and storage, and meteorological sites only need to upload local model parameters without transmitting a large amount of original data, so that the expenditure of bandwidth and storage resources is saved, and the federal learning framework is utilized, so that the meteorological sites in all areas can participate in model training more efficiently, and the quality and accuracy of a prediction model are improved;
in addition, the accurate weather forecast result can also provide important reference basis for decision making of wind power generation, and can help enterprises and departments to optimize operation and resource allocation, thereby improving efficiency and profits.
To further illustrate, federal learning is a machine learning technique whose core idea is to allow multiple participants to share model updates instead of raw data. Through the local training of the model, the leakage and loss risks in the data circulation process are reduced, and only the model gradient or parameters of privacy security are transmitted to update the cooperative federal model, so that the data privacy security is ensured;
the federal learning framework used herein is lateral (horizontal) federal learning, i.e., each participant (locally) has a different row of data under the same column of data features, where the input features of the local model and the global model are the same. The model aggregation mode adopts a weighted aggregation method and is based on an evaluation index R 2 And carrying out weight aggregation of the weights, and repeating iteration to obtain the global model with optimal performance.
To further illustrate, biLSTM is a variant of a Recurrent Neural Network (RNN) that includes two parallel LSTM networks that communicate information in forward and reverse directions, respectively, and that is effective in capturing information before and after a sequence and in processing long sequences.
A single LSTM cell contains three gates (input gate, forget gate and output gate) and one cell state. Together, these gates and cell states determine the way information flows in the LSTM cell. For forward LSTM, the hidden state h is calculated using the following formula t ：
First updating the current forget door state f t ：
f t ＝σ(W f ·x t +W f ·h (t-1) +b f ),
Updating the current input gate result i t ：
i t ＝σ(W i ·x t +W i ·h (t-1) +b t ),
Updating candidate value c by combining forget gate and input gate results t ：
Wherein isRepresenting the temporary candidate value, multiplying the temporary candidate value with the input gate result point and adding a state update value:
then calculate the current output door state o t ：
o t ＝σ(W o ·x t +W o ·h (t-1) +b o ),
Finally combine with the output door o t And hidden state c t Calculating the external state, updating the final output h t ：
h t ＝o t ·tanh(c t )
For reverse LSTM, the hidden state h is calculated using the same formula t ' but hidden state h of BiLSTM t Is made up ofSpliced to the hidden states of the LSTM and the reverse LSTM, and finally spliced to obtain the hidden state h t ＝[h t ；h t ′]。
At the same time, output y of BiLSTM t Also formed by the output concatenation of forward LSTM and reverse LSTM, denoted y t ＝[y t ；y t ′]。
In summary, the model training is directly completed by utilizing dimensionless data through federal learning, and then local model parameters are respectively weighted and aggregated, so that a global model capable of comprehensively learning global weather changes is obtained, and the model is learned by combining data of cross sites, so that the model can better understand the mode and the association between weather data, and the accuracy and the reliability of the prediction model are improved.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the invention, but any modifications, equivalents, improvements, etc. within the principles of the present invention should be included in the scope of the present invention.
Claims (10)
1. A wind farm weather forecast algorithm based on federal learning is characterized in that: the system comprises a weather prediction module, wherein the weather prediction module comprises a data processing module and a federal learning module;
the weather prediction module comprises the following specific steps:
step one, acquiring historical meteorological data of a multi-area meteorological site;
step two, the historical weather is locally preprocessed by a data processing module;
step three, constructing an iterative model by the federal learning module for iterative updating;
step four, screening a federal learning model with satisfactory performance as a final model to predict;
the data processing module comprises the following steps:
s101, analyzing the missing condition of data, and deleting variables with excessively high missing degree based on a missing degree threshold value; filling variables with fewer missing values, and supplementing the missing values which are single-value missing and continuous missing by using nearest neighbor interpolation and decision tree interpolation respectively to ensure the data integrity;
s102, recoding at least one mode of single-heat coding, tag coding or custom numerical coding can be adopted for discrete variable data;
s103, converting the data into uniform scales;
s104, constructing a neural network layer for continuous variables to extract features;
the federal learning module includes the steps of:
s201, training a local model;
s202, after the local model training is completed, the local model of each participant is sent to a central server, and the central server aggregates the parameters of all the local models into a global model;
s203, evaluating the performance of the global model by using the test data set to ensure that the model can accurately predict.
2. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in the step S101, the acquired data is subjected to missing detection, the variables with the missing rate exceeding 45% are deleted, and the remaining variables are subjected to missing value filling.
3. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in the step S102, the missing discrete variable is mainly filled in forward direction, and the continuous variable is filled in by mean value.
4. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in S103, separating the continuous numerical value from the wind direction discrete variable, recoding wind direction data, converting angle data of 0-360 degrees, normalizing the continuous variable, and performing the next feature extraction.
5. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in the step S104, the high-dimensional characteristics of the data extracted by the networks such as the fully connected network, the LSTM, the BiLSTM and the like are selected, and the local model is trained together after the high-dimensional characteristics are spliced with the original discrete data again.
6. A federally learned wind farm weather forecast algorithm according to claim 5, wherein: for continuous variables, a two-layer LSTM is adopted as a feature extraction network to acquire high-dimensional time sequence features of the continuous variables, the original continuous variable sequence is predicted in advance by training a network model with two layers of LSTM and one layer of full-connection layer, and the iteration times are set to be 100;
the fitted model is reserved with parameters of two layers of LSTM networks, original meteorological data are input into the model, the expanded high-dimensional characteristics are extracted, the final extraction dimension is set to be 32, the characteristics and discrete variables are combined, data are segmented into a training set and a verification set through a leave-one-method, and the training set is input into a federal learning model for training.
7. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in the step S201, the local model updates model parameters based on the gradient descent method, and selects R 2 As an evaluation criterion of the model, R predicted on a verification set after local model training is saved 2 And calculating a value.
8. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in the step S202, the update formula for the global model is as follows:
wherein θ global Representing parameters, θ, of the global model i Local model parameters representing the ith party, w i Representing the weight of the i-th participant,representing +.o. of the ith participant local model on the verification set>A calculated value, N, representing the number of participants;
which is based on R on the validation set for each party's local model 2 Calculating value to weight average parameter of aggregate global model, weight w i By R of each participant 2 And (5) value determination.
9. A federally learned wind farm weather forecast algorithm according to claim 1 or 7, wherein: in the step S203, the iteration returns to the step S201, and the global model is issued back to the local to continue training until the global model converges or the maximum iteration number is reached.
10. A federally learned wind farm weather forecast algorithm according to claim 1, wherein: in the first to fourth steps, the federal learning model with satisfactory performance is put into weather prediction, new data is input into the model, and then a multi-step prediction result can be obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311532774.4A CN117761802A (en) | 2023-11-16 | 2023-11-16 | Wind farm weather forecast algorithm based on federal learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311532774.4A CN117761802A (en) | 2023-11-16 | 2023-11-16 | Wind farm weather forecast algorithm based on federal learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117761802A true CN117761802A (en) | 2024-03-26 |
Family
ID=90320892
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311532774.4A Pending CN117761802A (en) | 2023-11-16 | 2023-11-16 | Wind farm weather forecast algorithm based on federal learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117761802A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118465875A (en) * | 2024-07-10 | 2024-08-09 | 国网江西省电力有限公司信息通信分公司 | A method and system for predicting meteorological disasters in power areas based on convolution algorithm |
-
2023
- 2023-11-16 CN CN202311532774.4A patent/CN117761802A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118465875A (en) * | 2024-07-10 | 2024-08-09 | 国网江西省电力有限公司信息通信分公司 | A method and system for predicting meteorological disasters in power areas based on convolution algorithm |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2024051524A1 (en) | Joint prediction method and apparatus for hydraulic, wind and photovoltaic generation power | |
| CN102184337B (en) | Dynamic combination analysis method of new energy generating capacity influenced by meteorological information | |
| Liao et al. | Ultra-short-term interval prediction of wind power based on graph neural network and improved bootstrap technique | |
| CN113657662B (en) | Downscaling wind power prediction method based on data fusion | |
| CN106650784A (en) | Feature clustering comparison-based power prediction method and device for photovoltaic power station | |
| CN108596449A (en) | It is a kind of to consider distribution network reliability prediction technique of the weather to distribution network failure impact probability | |
| CN102999786A (en) | Method for short-term predicting of photovoltaic generation power on the basis of similar day feature classification and extreme learning machine | |
| CN112149879A (en) | A new energy medium and long-term electricity forecast method considering macro-volatility classification | |
| CN115267945A (en) | Thunder and lightning early warning method and system based on graph neural network | |
| CN115374995A (en) | Distributed photovoltaic and small wind power station power prediction method | |
| CN111612244B (en) | QRA-LSTM-based method for predicting nonparametric probability of photovoltaic power before day | |
| CN109636054A (en) | Solar energy power generating amount prediction technique based on classification and error combination prediction | |
| CN115965132A (en) | Power prediction method for distributed photovoltaic digital twin system based on GA-BP neural network | |
| CN110909994A (en) | Power generation forecast method of small hydropower group based on big data | |
| CN114021830A (en) | A multi-time-scale wind speed prediction method based on CNN-LSTM | |
| CN114357670A (en) | Power distribution network power consumption data abnormity early warning method based on BLS and self-encoder | |
| Chen et al. | Research on wind power prediction method based on convolutional neural network and genetic algorithm | |
| CN115034485A (en) | A data space-based wind power interval prediction method and device | |
| CN114399081A (en) | A weather classification-based photovoltaic power generation power prediction method | |
| CN116611702A (en) | Integrated learning photovoltaic power generation prediction method for building integrated energy management | |
| CN117761802A (en) | Wind farm weather forecast algorithm based on federal learning | |
| CN118070010A (en) | Renewable energy scene generation method based on generation countermeasure network | |
| CN116316615A (en) | Data enhancement-based distributed light Fu Qun short-term power prediction method and system | |
| CN114004405B (en) | Photovoltaic power prediction method and system based on Elman neural network and satellite cloud image | |
| CN118469285A (en) | Weather warning method and system based on substation and transmission line |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |