Multi-station charging load distributed prediction method and system based on reservoir calculation

Abstract

The invention discloses a multi-station charging load distributed prediction method and system based on reservoir calculation, and the method comprises the steps: firstly obtaining the historical data of the charging loads of a plurality of electric vehicle charging stations, and carrying out the preprocessing; then, based on a similarity matrix of a Gaussian kernel function, space-time coupling characteristics between electric vehicle charging stations distributed at different places in space are evaluated, and a distributed station set is constructed according to the coupling characteristics; and finally, respectively establishing a distributed prediction network based on a reservoir calculation normal form for the associated distributed station set, taking station load data in the distributed station set as input and output, and performing prediction network distributed training by adopting a distributed learning algorithm based on an approximate synchronous parallel mechanism. And performing real-time prediction by using the trained prediction network. According to the method, efficient collaboration of multi-station data and charging load dynamic evolution modeling are achieved, and the method plays an important role in further promoting traffic electrification and power grid intellectualization.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy vehicle charging stations, and more specifically, to a multi-station charging load distributed prediction method and system based on reservoir calculation. Background Art

[0002] With the global energy transition, the number of new energy vehicles (NEVs) has exploded. Electric vehicles (EVs) have dominated the market, leading to a booming industry for EV charging equipment. Basic charging equipment includes AC slow chargers, DC slow chargers, DC fast chargers, and DC super chargers. DC fast chargers and DC super chargers are highly favored by consumers for their high charging power and speed, and are widely deployed in public charging stations.

[0003] With the continuous growth of electric vehicle sales, the surge in charging demand has led to increasing peak-to-valley variations in the distribution network and increasingly prominent problems such as overload in some areas. Accurate power forecasting of charging loads at multiple stations within a region has become a core requirement for smart grid scheduling. Traditional forecasting methods face significant challenges in addressing the spatiotemporal heterogeneity and dynamic coupling of loads across multiple stations: 1. Inadequate spatiotemporal feature extraction. Existing models often use single-time series analysis, which struggles to capture the load distribution characteristics of different stations due to differences in geographic location and user behavior. 2. Inefficient data collaboration. Traditional centralized forecasting frameworks are subject to data transmission delays and privacy risks between multiple stations, while single-station models cannot effectively leverage load correlations across multiple stations. 3. Dynamic adaptability is weak. Traditional statistical models (such as ARIMA) are sensitive to external disturbances such as sudden weather changes and holidays. While machine learning models (such as LSTM) can capture time series features, parameter training is complex and they struggle to adapt to heterogeneous data from multiple stations.

[0004] Among current technologies, reservoir computing-based approaches have demonstrated advantages in time series modeling. Leveraging the dynamic reservoir properties of reservoir computing, they can preserve the spatiotemporal characteristics of input signals through random mapping, offering new approaches for multi-station load forecasting. However, existing distributed forecasting systems often employ federated learning architectures, which protect data privacy but struggle to ensure real-time performance.

[0005] Therefore, there is an urgent need to build a distributed prediction system based on the reservoir computing paradigm to achieve efficient coordination of multi-station data and dynamic evolution modeling of charging loads, which plays a vital role in further developing my country's electric vehicle industry and promoting transportation electrification and power grid intelligence. Summary of the Invention

[0006] The purpose of the present invention is to address the deficiencies of the existing technology and propose a multi-station charging load distributed prediction method and system based on reservoir calculation.

[0007] The object of the present invention is achieved through the following technical solution: a multi-station charging load distributed prediction method based on reservoir calculation, the method comprising:

[0008] Step 1: Obtain historical data of charging loads of multiple electric vehicle charging stations and perform preprocessing;

[0009] Step 2: Evaluate the spatiotemporal coupling characteristics between electric vehicle charging stations distributed in different locations based on the similarity matrix of the Gaussian kernel function, and construct a distributed station set based on the coupling characteristics;

[0010] Step 3: Establish a distributed prediction network based on the reservoir computing paradigm, using the site load data in the distributed site set as input and the matrix weights as output, and use a distributed learning algorithm based on an approximately synchronous parallel mechanism to perform distributed training of the prediction network;

[0011] Step 4: For each associated distributed station set, a distributed prediction network is constructed and trained in turn according to step 3, and the trained prediction network is used for real-time prediction.

[0012] Furthermore, in step 1, the historical data of charging load of the electric vehicle charging station is the load data collected from each charging station; the preprocessing includes: cleaning outliers and missing values ​​in the data, normalizing the data; and performing time alignment and interpolation processing on the historical data of multiple charging stations.

[0013] Furthermore, in step 2, based on the historical load time series data of charging stations at multiple sites, the Gaussian kernel function values ​​and similarities between load data points are calculated to construct a similarity matrix.

[0014] Furthermore, in step 3, an echo state network is established based on the reservoir computing paradigm as a charging load prediction model for a single station, and a distributed prediction network is constructed for the load prediction problem of multiple charging station clusters.

[0015] Furthermore, in step 3, the charging station cluster needs to integrate the historical load data of different stations to establish the augmented Lagrangian function of the prediction network :

[0016]

[0017] in, Represents the first in the charging station cluster Historical data of charging station load, Represents the charging station cluster The hidden layer output of the neural network of the charging station, For the charging station cluster The output layer weight matrix of the neural network of the charging station, is the target value of the output layer weight matrix, , is the system hyperparameter, is the penalty term parameter of the Lagrangian function, Represents the number of charging stations in the distributed charging station cluster, is the Lagrange multiplier vector, T represents the transpose;

[0018] According to the target value of the output layer weight matrix, the prediction network is trained by the following method to obtain the parameters of the prediction network:

[0019]

[0020] in is the estimated value of the output layer weight matrix during the k-th iteration training process.

[0021] Furthermore, a distributed average consensus protocol based on an approximate synchronous parallel mechanism is established to realize distributed collaborative training of multiple charging station cluster systems. The specific process is as follows: randomly initialize the output layer weight matrix and target value and the Lagrange multiplier vector parameters , iteratively calculate the output layer weight matrix of the k+1th iteration at the local site , , to achieve local weight update:

[0022]

[0023] in is the Lagrange multiplier vector of the kth iteration, Represents the identity matrix.

[0024] Furthermore, in step 3, define and Respectively represent and The mean of the two is needed to be calculated during the distributed collaborative training process to obtain global information. The distributed average consensus protocol is used among the sub-sites to achieve global consensus:

[0025]

[0026] in, represents the communication network topology connection coefficient between sub-site i and sub-site j, n represents the number of communications, and according to the maximum degree of the charging station cluster communication network topology graph, , calculate the communication coefficient between different sites ,Conversely, when there is no communication between different sites, the value is 0; ,in Indicates the node communication topology diagram The degree of the charging station node; then when the number of communications , the data of each node and will converge to the mean of all data and ; Multiple communications are carried out between charging station clusters and according to the matrix The elements of the transmission and Update to get the mean.

[0027] Furthermore, in step 3, the distributed average consensus protocol introduces the Approximate Synchronous Parallel (ASP) mechanism. For each node, when each update is greater than the set threshold, communication occurs and distributed global consensus calculation is performed. Otherwise, it is necessary to confirm whether the node communication is normal; during the iteration process, the threshold is dynamically adjusted.

[0028] Furthermore, in step 3, after obtaining global consensus through multiple communications, the global consensus variable is updated at the local site. and local multipliers , until the convergence condition is met;

[0029]

[0030]

[0031] During distributed training, and All are updated locally, without the need to obtain global information, maintaining data privacy, only updating When it is necessary to obtain the global weight mean information, it only needs to transmit the output weight matrix with partial information and does not participate in the direct transmission of sensitive data.

[0032] On the other hand, the present invention also provides a multi-station charging load distributed prediction system based on reservoir calculation, the system comprising:

[0033] The acquisition module is used to obtain the historical load data of each electric vehicle charging station and perform preprocessing;

[0034] The module is selected and the spatiotemporal coupling characteristics between electric vehicle charging stations distributed in different locations are evaluated based on the similarity matrix of the Gaussian kernel function. A distributed station set is constructed based on the coupling characteristics.

[0035] Communication module, used for network communication when the distributed prediction system of electric vehicle charging station cluster performs collaborative prediction;

[0036] The training module is used to establish a distributed prediction network based on the reservoir computing paradigm, using site load data in a distributed site set as input and matrix weights as output, and adopting a distributed learning algorithm based on an approximately synchronous parallel mechanism to perform distributed training of the prediction network;

[0037] The prediction module is used to perform real-time prediction using the trained prediction network.

[0038] The beneficial effects of the present invention are:

[0039] 1. The present invention adopts an echo state network prediction method based on reservoir calculation. The system is lightweight and easy to deploy. It can be easily deployed and expanded in large-scale regional multi-station electric vehicle charging clusters.

[0040] 2. The distributed network prediction model designed by this invention takes into account the security and data transmission efficiency of the system information network transmission. During the process of obtaining global consensus, it does not directly transmit sensitive historical data, but only transmits the desensitized output weight matrix data. This meets the privacy protection requirements and the ability to deploy on resource-constrained edge devices.

[0041] 3. The multi-station electric vehicle charging cluster prediction system model of the present invention takes the system as a whole and is based on distributed collaborative training of local variable information interaction among multiple charging sites, which can effectively improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of the multi-station charging load distributed prediction method based on reservoir calculation provided by the present invention.

[0043] Figure 2 This is a structural diagram of the echo state network distributed prediction model based on reservoir calculation provided by the present invention.

[0044] Figure 3 This is an implementation diagram of the multi-station distributed charging load prediction model based on reservoir calculation provided by the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described below with reference to the following examples. The following examples are provided only to facilitate understanding of the present invention. It should be noted that, without departing from the principles of the present invention, it is possible for a person skilled in the art to make various modifications to the present invention, and such improvements and modifications fall within the scope of the claims of the present invention.

[0046] Example 1:

[0047] like Figure 1 As shown, the embodiment of the present application provides a multi-station charging load distributed prediction method based on reservoir calculation, which is applied to Figure 3 As shown in the figure, the charging load forecast of five stations distributed in different geographical locations is established as follows: Figure 2 The distributed prediction network based on the approximately synchronous parallel mechanism shown in FIG. 1 specifically includes:

[0048] Step 1: Obtain historical load data of five geographically dispersed electric vehicle charging stations and preprocess the collected data.

[0049] In step 1, the preprocessing includes: missing value processing, time alignment processing and data normalization processing.

[0050] The interpolation processing and time alignment processing for the missing data include: observing historical data, selecting a unified timestamp, performing secondary interpolation processing on the missing data; and discarding data with unaligned timestamps.

[0051] The normalization process, the normalization formula is as follows:

[0052]

[0053] It represents the true value of each data point of the sequence data. 、 Represent the maximum and minimum values ​​of the sequence data respectively. Represents the normalized data of each data point in the series data. The historical load data of the five stations are normalized in turn.

[0054] Step 2: Construct a similarity matrix based on the Gaussian kernel function to evaluate the spatiotemporal coupling characteristics between EV charging stations distributed across different locations. Based on the coupling characteristics, a distributed station cluster is constructed, and the load data of the stations in this cluster is used as input to the prediction model.

[0055] The historical load data of each electric vehicle charging station is expressed as: ,in, Indicates the serial number of the charging station. , Indicates the sequence number of the data point, Indicates charging station A collection of historical load data points.

[0056] Specifically, the coupling relationship analysis includes: Gaussian kernel function calculation, similarity calculation and similarity matrix calculation.

[0057] Iterative calculation of the inter-site The data similarity of the sampling points for:

[0058]

[0059] in, The data set The first part of the historical load time series data of the charging station The sample values, is the control range parameter, For two sites The Gaussian kernel function value between load data points, is the weight coefficient for iterative update.

[0060] In step 2, the initial value of the similarity matrix is ​​calculated based on the data similarity :

[0061]

[0062] in, is the identity matrix, Represents the number of sampling points, is a matrix of all 1s, is the calibration parameter. The iterative formula of the similarity matrix is:

[0063]

[0064] in, The first row represents the historical load time series data of site m and site n. The similarity of the sample values ​​is the element in the similarity matrix, is the threshold, is the iteration parameter. After multiple iterative calculations, when the coupling between sites is strong, the similarity increases with the number of iterations. It will gradually converge to 1; when the coupling between sites is weak, the similarity It will gradually converge to 0. According to the similarity matrix based on the Gaussian kernel function, the spatiotemporal coupling relationship between each site can be better analyzed.

[0065] Step 3: Build an echo state network based on the reservoir computing paradigm. This consists of a randomly generated input layer, a reservoir layer, and an output layer to be trained. The reservoir layer has the characteristics of a recurrent neural network. The specific network structure is as follows:

[0066]

[0067]

[0068]

[0069] in for The input vector at time t, for The state vector at time t, for The reservoir output vector at time , is a nonlinear function, for The output vector at time , They are the input layer weight matrix, reservoir weight matrix and output layer weight matrix respectively.

[0070] Based on the reservoir computing paradigm, an echo state network is established to establish a single station charging load prediction model. The output layer of the prediction network mathematical model is expressed as:

[0071]

[0072] Where, is the hidden layer output of the neural network, is the output layer weight matrix of the neural network, Represents the network output.

[0073] In order to improve the robustness of the prediction network model, a regularization term is introduced to improve the robustness of the model, as shown below:

[0074]

[0075] Where, Represents the historical load data of the charging station, Represents the hyperparameters of the model. In this example, there are five charging stations, so . Constructed a distributed prediction network model:

[0076]

[0077]

[0078] Where, Represents the first in the charging station cluster Historical data of charging station load, Represents the charging station cluster The hidden layer output of the neural network of the charging station, For the charging station cluster The output layer weight matrix of the neural network of the charging station, is the target value of the output layer weight matrix, is the system hyperparameter, Represents the number of charging stations in the distributed charging station cluster.

[0079] In step 4, the charging station cluster needs to integrate the historical load data of different stations. To balance the needs of big data training and data privacy protection, the augmented Lagrangian function of the prediction network is established as follows:

[0080]

[0081] in, , is the penalty term parameter of the Lagrangian function, is the Lagrange multiplier vector. It can be seen that if the target value of the output layer weight matrix is ​​estimated, then for the For each site, the prediction network can be trained and the parameters of the prediction network can be obtained by the following method:

[0082]

[0083] in is the estimated value of the output layer weight matrix during the k-th iteration training process.

[0084] The learning algorithm of the distributed average consensus protocol based on the approximately synchronous parallel mechanism is used to transform complex problems into solving problems about variables. , and The three sub-problems are solved by optimizing the three variables respectively to realize the distributed collaborative training of multiple charging station cluster systems. After enough iterations, the optimal weight matrix of the neural network is finally obtained. The training process of the neural network model is as follows:

[0085] 1. Randomly initialize the output layer weight matrix and target value and Lagrange multiplier vector parameters , iteratively calculate the output layer weight matrix of the k+1th iteration at the local site , , implement sub-site The prediction network local weight update:

[0086]

[0087] 2. During the distributed collaborative training process, it is necessary to calculate the mean of the two to obtain global information. In order to achieve good distributed prediction results, the distributed average consensus protocol is used between the sub-sites to implement global consensus variable updates:

[0088]

[0089] in, and Respectively represent and Therefore, each site needs the following communication iterations to obtain the result:

[0090]

[0091]

[0092] in, is the topology connection matrix of the communication network between sub-sites, Indicates the node communication topology diagram The degree of each charging station node, Indicates the maximum degree in the charging station cluster communication network topology. , the data of each node and will converge to the mean of all data and The charging station clusters communicate multiple times and follow the matrix The elements of the transmission and Update to get the mean.

[0093] To improve communication efficiency, the distributed average consensus protocol introduces an approximate synchronous parallel mechanism (ASP). When each update is less than a specific value, the update is postponed to reduce communication volume. When an event is triggered, a distributed global consensus calculation is performed; if the conditions are not met, a lightweight heartbeat packet (a 1-bit flag) is sent to confirm whether the node communication is normal. The trigger conditions are:

[0094]

[0095] in, is the threshold parameter, which is dynamically updated with the number of iterations, and can reduce the communication volume as the algorithm converges exponentially:

[0096]

[0097] 3. After obtaining global consensus through multiple communications, update the global consensus variables at the local site and local multipliers :

[0098]

[0099] The above training process is used to perform multiple iterations of the prediction model of the distributed electric vehicle charging station cluster to obtain the required weight matrix. and All updates are done locally, without the need to obtain global information, thus maintaining data privacy. When it is necessary to obtain the global weight mean information, it only needs to transmit the output weight matrix with partial information and does not participate in the direct transmission of sensitive data. When calculating the load, the distributed average consensus protocol is used to enable the site cluster to achieve automatic iterative calculation without a central node, and finally a network model for load prediction of a distributed electric vehicle charging station cluster is obtained.

[0100] A neural network was used to mine the nonlinear relationship between the loads of five selected electric vehicle charging stations. A continuous 30-day historical data set was selected as the training set, and load data for the next 1, 3, and 7 days was predicted. The results are shown in Tables 1-3. Since there was no previous step information during testing, the output data of the first 100 steps was discarded and the prediction data was padded backward. A grid search method was used to find the optimal hyperparameters on the training set. .

[0101] Table 1. One-day prediction results of distributed charging load at multiple stations

[0102]

[0103] Table 2. Three-day prediction results of distributed charging load at multiple stations

[0104]

[0105] Table 3. Seven-day forecast results of distributed charging load at multiple stations

[0106]

[0107] The accuracy indicators in model training include normalized root mean square error (RMSE), mean square error (MSE), and mean absolute error (MAE) for evaluation. The expressions are:

[0108]

[0109] in, represents the number of samples, Represents the electric vehicle charging station The load forecast value, Represents the electric vehicle charging station The smaller the error index is, the smaller the prediction error is.

[0110] Step 5. Use the trained model to perform real-time prediction. The prediction results are shown in Tables 1 to 3. The data with the best overall effect are shown in bold. The data source is the open source ST-EVCDP V2 data set of charging stations in a certain city, and the error indicator mainly uses RMSE. The control groups are L-RC and C-RC, which represent the centralized iterative solution training reservoir calculation model and the centralized analytical solution training reservoir calculation model, respectively. The reservoir calculation model for distributed multi-station electric vehicle charging load prediction of the present invention is D-RC. In the case where all data are not obtained locally for training, the effect of D-RC is better than L-RC and is basically close to C-RC.

[0111] Example 2:

[0112] Based on Example 1, this embodiment of the present application provides a multi-station charging load distributed prediction system based on reservoir calculation, including:

[0113] The acquisition module is used to obtain the historical load data of each electric vehicle charging station and perform preprocessing;

[0114] A selection module is used to construct a similarity matrix based on the Gaussian kernel function to test the correlation between the load data of each electric vehicle charging station, build a distributed station set based on the coupling characteristics, and use the load data of the stations in the set as the input of the prediction model;

[0115] Communication module, used for network communication when the distributed prediction system of electric vehicle charging station cluster performs collaborative prediction;

[0116] The training module is used to calculate the output matrix weights in the distributed neural network model training process based on the approximately synchronous parallel mechanism;

[0117] The prediction module is used to perform real-time predictions using the trained model.

[0118] Specifically, the system provided in this embodiment is a system corresponding to the method provided in Example 1. Therefore, the parts in this embodiment that are the same or similar to those in Example 1 can be referenced to each other and will not be repeated in this application.

[0119] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modifications and changes made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A distributed prediction method for multi-station charging load based on reservoir calculation, characterized in that: The method includes: Step 1: Obtain historical data of charging loads of multiple electric vehicle charging stations and perform preprocessing; Step 2: Evaluate the spatiotemporal coupling characteristics between electric vehicle charging stations distributed in different locations based on the similarity matrix of the Gaussian kernel function, and construct a distributed station set based on the coupling characteristics; Step 3: Establish a distributed prediction network based on the reservoir computing paradigm, using the site load data in the distributed site set as input and the matrix weights as output, and use a distributed learning algorithm based on an approximately synchronous parallel mechanism to perform distributed training of the prediction network; Step 4: For each associated distributed station set, a distributed prediction network is constructed and trained in turn according to step 3, and the trained prediction network is used for real-time prediction.

2. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 1, characterized in that: In step 1, the historical data of charging load of electric vehicle charging stations is the load data of each charging station collected; the preprocessing includes: cleaning outliers and missing values ​​in the data, normalizing the data; and performing time alignment and interpolation processing on the historical data of multiple charging stations.

3. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 1, characterized in that: In step 2, based on the historical load time series data of charging stations at multiple sites, the Gaussian kernel function values ​​and similarities between load data points are calculated to construct a similarity matrix.

4. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 1, characterized in that: In step 3, an echo state network is established as a charging load prediction model for a single station based on the reservoir computing paradigm, and a distributed prediction network is constructed for the load prediction problem of multiple charging station clusters.

5. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 4 is characterized in that: In step 3, the charging station cluster needs to integrate the historical load data of different stations to establish the augmented Lagrangian function of the prediction network : in, Represents the first in the charging station cluster Historical data of charging station load, Represents the charging station cluster The hidden layer output of the neural network of the charging station, For the charging station cluster The output layer weight matrix of the neural network of the charging station, is the target value of the output layer weight matrix, , is the system hyperparameter, is the penalty term parameter of the Lagrangian function, Represents the number of charging stations in the distributed charging station cluster, is the Lagrange multiplier vector, T represents the transpose; According to the target value of the output layer weight matrix, the prediction network is trained by the following method to obtain the parameters of the prediction network: in is the estimated value of the output layer weight matrix during the k-th iteration training process.

6. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 5 is characterized in that: Establish a distributed average consensus protocol based on an approximate synchronous parallel mechanism to achieve distributed collaborative training of multiple charging station cluster systems. The specific process is as follows: randomly initialize the output layer weight matrix and target value and the Lagrange multiplier vector parameters , iteratively calculate the output layer weight matrix of the k+1th iteration at the local site , , to achieve local weight update: in is the Lagrange multiplier vector of the kth iteration, Represents the identity matrix.

7. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 6 is characterized in that: In step 3, define and Respectively represent and The mean of the two is needed to be calculated during the distributed collaborative training process to obtain global information. The distributed average consensus protocol is used to achieve global consensus among the sub-sites: in, represents the communication network topology connection coefficient between sub-site i and sub-site j, n represents the number of communications, and according to the maximum degree of the charging station cluster communication network topology graph, , calculate the communication coefficient between different sites ,Conversely, when there is no communication between different sites, the value is 0; ,in Indicates the node communication topology diagram The degree of the charging station node; then when the number of communications , the data of each node and will converge to the mean of all data and ; Multiple communications are carried out between charging station clusters and according to the matrix The elements of the transmission and Update to get the mean.

8. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 6 is characterized in that: In step 3, the distributed average consensus protocol introduces the Approximate Synchronous Parallel (ASP) mechanism. For each node, when each update is greater than the set threshold, communication occurs and distributed global consensus calculation is performed. Otherwise, it is necessary to confirm whether the node communication is normal; during the iteration process, the threshold is dynamically adjusted.

9. The method for distributed prediction of multi-station charging load based on reservoir calculation according to claim 7, characterized in that: In step 3, after multiple communications to obtain global consensus, the global consensus variables are updated at the local site. and local multipliers , until the convergence condition is met; During distributed training, and All are updated locally, without the need to obtain global information, maintaining data privacy, only updating When it is necessary to obtain the global weight mean information, it only needs to transmit the output weight matrix with partial information and does not participate in the direct transmission of sensitive data.

10. A distributed prediction system for charging loads at multiple stations based on reservoir calculation that implements the distributed prediction method according to any one of claims 1 to 9, characterized in that: The system includes: The acquisition module is used to obtain the historical load data of each electric vehicle charging station and perform preprocessing; The module is selected and the spatiotemporal coupling characteristics between electric vehicle charging stations distributed in different locations are evaluated based on the similarity matrix of the Gaussian kernel function. A distributed station set is constructed based on the coupling characteristics. Communication module, used for network communication when the distributed prediction system of electric vehicle charging station cluster performs collaborative prediction; The training module is used to establish a distributed prediction network based on the reservoir computing paradigm, using site load data in a distributed site set as input and matrix weights as output, and adopting a distributed learning algorithm based on an approximately synchronous parallel mechanism to perform distributed training of the prediction network; The prediction module is used to perform real-time prediction using the trained prediction network.

Images