A method for campus power load forecasting based on federated learning

Abstract

The present application is applicable to the technical field of power load forecasting, and provides a method for predicting power load in a park based on federated learning, including: the client performs EMD decomposition on the power load data sequence of the park to obtain multiple IMF components; calculates the DTW distance between the meteorological temperature sequence of the park and the IMF component, and uses the DTW distance as the similarity between the meteorological temperature sequence and the IMF component; sorts multiple IMF components to obtain a power load decomposition component sequence; receives global model parameters sent by a central server, performs local training using the power load decomposition component sequence, uploads the trained local model parameters to the central server, and causes the central server to aggregate global parameters, and sends the aggregated global model parameters to the client, and the client uses the new global model parameters to perform the next round of local training until the model converges. The present application can improve the accuracy of power load forecasting under the federated learning framework.

Description

Technical Field

[0001] The present application belongs to the technical field of power load forecasting, and in particular, relates to a method for predicting power load in a park based on federated learning. Background Art

[0002] Under the distributed architecture of virtual power plants, using federated learning, multiple parks can collaboratively train an efficient power load forecasting model without sharing sensitive user data, thereby improving the accuracy and performance of predictions while protecting privacy. By increasing the number of effective training samples, virtual power plants can more accurately predict power loads and optimize power production and distribution. However, aggregating data from different parks will cause data heterogeneity problems, that is, the local data of different participants come from different distributions. Data heterogeneity leads to the collapse of the representation dimension, which will slow down the convergence of the model, thereby greatly limiting the expressive power of the model and resulting in low accuracy in power load prediction. Summary of the invention

[0003] The embodiment of the present application provides a method for predicting power load in a park based on federated learning, which can solve the problem of low accuracy of power load prediction under the federated learning framework.

[0004] The present application embodiment provides a method for predicting campus power load based on federated learning, including:

[0005] The client participating in federated learning obtains the power load data sequence of the park to which it belongs, and performs EMD decomposition on the power load data sequence to obtain multiple IMF components; the power load data sequence includes the power load data of the park in the historical time period;

[0006] The client obtains the meteorological temperature sequence of the park, which includes the meteorological temperature data of the park in the historical time period;

[0007] The client uses the DTW algorithm to calculate the DTW distance between the meteorological temperature series and the IMF component for each IMF component, and uses the DTW distance as the similarity between the meteorological temperature series and the IMF component;

[0008] The client sorts multiple IMF components based on similarity to obtain a sequence of power load decomposition components;

[0009] The client receives the global model parameters sent by the central server;

[0010] The client updates the local model according to the global model parameters. If the updated local model converges, the updated local model is used to predict the power load of the park. If the updated local model does not converge, the updated local model is locally trained using the power load decomposition component sequence, and the local model parameters of the trained local model are uploaded to the central server, so that the central server aggregates the global parameters based on the dynamic matching sorting cumulative distance uploaded by each client participating in the federated learning, and sends the aggregated global model parameters to the clients participating in the federated learning, and returns to the step of the client receiving the global model parameters sent by the central server. The dynamic matching sorting cumulative distance is related to the DTW distance calculated by the client.

[0011] Optionally, the client sorts multiple IMF components based on similarity to obtain a power load decomposition component sequence, including:

[0012] The multiple IMF components are sorted in order of similarity from small to large, and the sorted multiple IMF components are used as the power load decomposition component sequence.

[0013] Optionally, after calculating the DTW distance between the meteorological temperature series and the IMF component using the DTW algorithm, the park power load forecasting method further includes:

[0014] Calculate the sum of the DTW distances corresponding to all IMF components, and use the sum as the dynamic matching sorting cumulative distance of the client.

[0015] Optionally, the central server aggregates global parameters based on the dynamic matching ranking cumulative distances uploaded by each client participating in federated learning, including:

[0016] The central server aggregates global parameters using the following formula:

[0017] ;

[0018] in, represents the global model parameters after aggregation, Indicates the first The dynamic matching sorting cumulative distance of the clients, Indicates the number of clients participating in federated learning, , Indicates the first The local model parameters uploaded by the client.

[0019] Optionally, the local model includes a CNN network, an LSTM network, and a fully connected layer connected in sequence.

[0020] Optionally, when the client uses the power load decomposition component sequence to perform local training on the updated local model, the loss function used is:

[0021] ;

[0022] in, represents the loss value calculated for the k-th client, represents the total number of time steps of the training data used by the kth client, represents the time step index, represents the power load value predicted by the kth client at time step t, Represents the actual power load value of the kth client at time step t.

[0023] Optionally, after the client updates the local model according to the global model parameters, the park power load prediction method further includes:

[0024] The client determines whether the updated local model meets the convergence condition. If so, it is determined that the updated local model is converged; otherwise, it is determined that the updated local model is not converged.

[0025] The convergence condition is that the local model is trained for a preset number of times or the prediction accuracy of the local model reaches a preset accuracy.

[0026] The above solution of the present application has the following beneficial effects:

[0027] In an embodiment of the present application, by performing EMD decomposition on the original power load data sequence, multiple IMF components representing different frequency components are obtained, thereby eliminating the impact of data non-stationarity. Subsequently, the meteorological temperature sequence is used as the reference center, and the IMF components are used as features extracted to compare different scales. The DTW algorithm is used for component matching to obtain the optimal order of power load decomposition component sequence, achieve data similarity alignment, weaken the impact of data heterogeneity in a distributed training environment, help improve the model convergence speed, enhance the model's expressiveness, and thereby improve the accuracy of power load prediction under the federated learning framework.

[0028] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 A flowchart of a method for predicting campus power load based on federated learning provided in one embodiment of the present application. DETAILED DESCRIPTION

[0031] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0032] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0033] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0034] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0035] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0036] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0037] In response to the current problem of low accuracy in power load prediction, an embodiment of the present application provides a park power load prediction method based on federated learning. By performing EMD decomposition on the original power load data sequence, multiple IMF components representing different frequency components are obtained, thereby eliminating the impact of data non-stationarity. Subsequently, the meteorological temperature sequence is used as the reference center, and the IMF components are used as features extracted to compare different scales. The DTW algorithm is used for component matching to obtain the optimal order of power load decomposition component sequence, achieve data similarity alignment, weaken the impact of data heterogeneity in a distributed training environment, help improve the model convergence speed, enhance the model's expressiveness, and thereby improve the accuracy of power load prediction under the federated learning framework.

[0038] The following is an illustrative description of the park power load forecasting method based on federated learning provided in the embodiment of the present application in conjunction with specific embodiments.

[0039] like Figure 1 As shown, the park power load forecasting method based on federated learning provided in the embodiment of the present application includes the following steps:

[0040] In step 11, the client participating in federated learning obtains the power load data sequence of the park to which it belongs, and performs EMD decomposition on the power load data sequence to obtain multiple IMF components.

[0041] The above-mentioned client may be a device with computing capabilities, such as a computer device, which is used to participate in federated learning to obtain a model for predicting the power load of the park to which it belongs.

[0042] The above-mentioned power load data sequence includes the power load data of the park in the historical time period. The park can be a high-tech industrial park, a characteristic industrial park, etc. The historical time period can be set according to the actual situation, such as the past week, the past month, and the past quarter. The power load data is the power load value. Exemplarily, the power load data sequence can include the power load value of the park in the past quarter. It is understandable that, in order to facilitate data processing, the power load value within a quarter can be sampled at a sampling frequency, and the multiple power load values ​​obtained by sampling are used as the power load data sequence.

[0043] In some embodiments of the present application, multiple intrinsic mode function (IMF) components and a residual term can be obtained by performing EMD decomposition on the power load data sequence. Each IMF component represents a different frequency component of the power load data sequence to eliminate the impact of data non-stationarity.

[0044] Step 12: The client obtains the meteorological temperature sequence of the park.

[0045] The above-mentioned meteorological temperature sequence includes meteorological temperature data of the park in a historical time period, which is the same period as the historical time period corresponding to the power load data sequence, and the meteorological temperature data can be the ambient temperature value of the park. It is understandable that the meteorological temperature sequence includes multiple ambient temperature values ​​of the park in the historical time period.

[0046] Step 13: The client calculates the DTW distance between the meteorological temperature sequence and the IMF component using the DTW algorithm for each IMF component, and uses the DTW distance as the similarity between the meteorological temperature sequence and the IMF component.

[0047] Step 14: The client sorts the multiple IMF components based on similarity to obtain a power load decomposition component sequence.

[0048] In some embodiments of the present application, the IMF components obtained by decomposing the meteorological temperature information and the historical energy consumption information using a dynamic time planning (DTW) algorithm are measured for similarity, and multiple IMF components are sorted based on the similarity to achieve data similarity alignment, thereby improving data utilization efficiency and weakening the impact of data heterogeneity in the distributed training environment of federated learning, which helps to improve the model convergence speed, enhance the model's expressiveness, and thus help improve the prediction accuracy of power load.

[0049] As a preferred example, multiple IMF components can be sorted in order of similarity from small to large, and the sorted multiple IMF components are used as the power load decomposition component sequence. By sorting and reorganizing the power load decomposition component sequence, the optimal sequence component sequence is obtained, so that sequences with similar periods can become the same feature, ensuring the alignment of each node feature in the feature extraction stage. Based on this, when sorting multiple IMF components, they can also be arranged in order from large to small to obtain the power load decomposition component sequence.

[0050] Because meteorological factors have a significant impact on power load, and the meteorological data of participating nodes in different parks in the same area are often consistent, this application uses meteorological temperature information as an indirect comparison feature to improve the prediction accuracy.

[0051] Step 15: The client receives the global model parameters sent by the central server.

[0052] In the related art, in a federated learning framework, a central server (which can be specifically a cloud server device) and a client work together to achieve the goal of distributed machine learning.

[0053] ‌The central server is mainly used for:

[0054] ‌Initialize the global model‌: The central server is responsible for saving the initialized sharable global data, which is the starting point of the global model;

[0055] ‌Coordination and aggregation‌: The server samples from a set of eligible clients, collects a summary of device updates, and updates the shared model locally based on these updates;

[0056] ‌Communication Coordination‌: Communication between the server and the clients is a key part of the federated learning process, through which the server sends the global model to a selected subset of clients and uploads aggregated model updates.

[0057] ‌The client is mainly used for:

[0058] ‌Local Training‌: Each selected device computes updates to the model locally by executing the training procedure;

[0059] ‌Model parameter upload‌: The client trains the local machine learning model based on local data, and uploads the trained model parameters and other data to the central server;

[0060] ‌Receive global model‌: The client downloads the current model weights from the central server and trains it on local data.

[0061] In some embodiments of the present application, at the beginning of federated learning, the central server performs initialization: initializing global model parameters , and send it to each client, and the clients participating in federated learning get the initialization parameters .

[0062] It is understandable that the above global model parameters are the model parameters of the model used for power load forecasting.

[0063] In step 16, the client updates the local model according to the global model parameters. If the updated local model converges, the updated local model is used to predict the power load of the park. If the updated local model does not converge, the updated local model is locally trained using the power load decomposition component sequence, and the local model parameters of the trained local model are uploaded to the central server, so that the central server aggregates the global parameters based on the dynamic matching sorting cumulative distance uploaded by each client participating in the federated learning, and sends the aggregated global model parameters to the clients participating in the federated learning, and returns to the step of the client receiving the global model parameters sent by the central server. The dynamic matching sorting cumulative distance is related to the DTW distance calculated by the client.

[0064] It is understandable that in federated learning, in order for the local model to converge, the client needs to perform multiple rounds of local training. Therefore, after the client updates the local model according to the global model parameters, the client needs to perform the following steps to determine whether the local model has converged:

[0065] The client determines whether the updated local model meets the convergence condition. If so, it is determined that the updated local model is converged. Otherwise, it is determined that the updated local model is not converged.

[0066] The above convergence condition may be that the number of training times of the local model reaches a preset number, or the prediction accuracy of the local model reaches a preset accuracy.

[0067] The prediction accuracy of the local model can be verified by verification data, which includes input data used to input the model for power load prediction (which can be the power load data of the park in a historical period of time) and the real power load data corresponding to the input data. Exemplarily, the input data can be input into the local model for power load prediction to obtain a prediction result, and then the prediction accuracy of the local model can be evaluated based on the prediction result and the real power load data.

[0068] It should be noted that the local model of the client and the global model of the central server are the same model. In the embodiments of the present application, they can both be models for power load forecasting.

[0069] In some embodiments of the present application, the above-mentioned local model includes a CNN network, an LSTM network and a fully connected layer connected in sequence.

[0070] Convolutional Neural Network (CNN) - Long Short-Term Memory Network (LSTM) has a strong ability to extract data features. In particular, CNN can directly learn the internal representation of time series data, without the need for domain expertise to manually extract features, while LSTM has a strong ability to extract time features and can more accurately capture the implicit information in time series trends. Combining CNN and LSTM can fully utilize the advantages of these two neural networks, fully tap the potential characteristics of the data, and provide support for prediction.

[0071] The CNN-LSTM model itself has a high degree of complexity, especially when dealing with high-dimensional data and long time series, the depth and parameter scale of the model may increase significantly. In the federated learning framework, this complexity will be further amplified due to the need to distribute training among multiple devices or institutions. However, this application achieves data similarity alignment based on the EMD-DTW dynamic matching algorithm, which weakens the impact of data heterogeneity in a distributed training environment, thereby effectively solving the problem of the model synchronization efficiency being affected by excessive complexity in a distributed training synchronization environment.

[0072] The CNN-LSTM serial structure is the basic model of federated learning. This model first changes the source data set through EMD subsequences, merges the subsequences into a 1-dimensional sequence, and inputs it into the CNN-LSTM serial structure for prediction. The serial structure means that the data passes through the CNN network structure and the LSTM network structure in sequence. The CNN network structure consists of one or more CNN networks. In order to extract different features, the CNN network uses different convolution kernels. After the CNN network extracts features, multiple CNN networks are fused through the pooling layer or the fully connected layer. The fusion result is then input into the LSTM for training and learning, and finally the final result is obtained through the fully connected layer. The specific operations are as follows:

[0073] (1) Input data

[0074] Assume that the input is the component data of the power load data sequence after EMD decomposition ,in: is the time step, is the number of IMF components generated after EMD decomposition, is the characteristic dimension in each IMF component (e.g., the time series value of load data). This means that each time step Depend on components, representing different frequency bands.

[0075] (2) CNN part - feature extraction

[0076] For each IMF component ,in , CNN can be used for feature extraction. Each IMF component represents the time series features in different frequency bands, and CNN will extract spatial features from each IMF component.

[0077] The CNN for each IMF includes convolution and pooling operations:

[0078] Convolution operation: ,in, is the convolution kernel, It is Components at time step data, Represents the bias term, which is used to adjust the convolution result to make the network have better expressive ability. Represents the output after convolution (also called feature map), which represents the extracted features. It is an activation function. Commonly used ones include ReLU (Rectified Linear Unit), etc. It is used to introduce nonlinearity so that the model can learn more complex features.

[0079] Pooling operation: ,in, is the output after the pooling operation, which is The feature representation of . Represents a pooling operation.

[0080] After CNN extracts features, for each time step , and the characteristic representation of each IMF component is obtained. The final output characteristic matrix is ,in is the feature dimension of the convolutional layer output.

[0081] (3) LSTM part - sequence modeling

[0082] Output feature matrix of CNN part is each time step The LSTM receives the feature sequence and models the dependencies between time steps.

[0083] For each time step , combine all IMF features extracted by CNN as the input of LSTM:

[0084] Input LSTM: LSTM processes the IMF features at each time step:

[0085]

[0086] in, is the hidden state, is the memory cell state, represents the LSTM network, represents the hidden state of the previous time step, Represents the state of the memory cell at the previous time step.

[0087] LSTM models the temporal relationship of load data through multiple time steps and generates the final hidden state sequence , Represents the dimension of the hidden layer, that is, the dimension of the hidden state output by LSTM.

[0088] (4) Fully connected layer and output

[0089] After LSTM processing, the hidden state Represents the global characteristics of the entire time series. It is passed to the fully connected layer to generate the final prediction value (i.e., the prediction of power load): .

[0090] in, represents the power load forecast value output by the local model, represents the weight matrix of the fully connected layer, represents the bias term of the fully connected layer, Represents the final hidden state of LSTM. It should be noted that for multi-time step prediction, multiple fully connected layers can be used to output the predicted values ​​of multiple time steps.

[0091] In some embodiments of the present application, when the client uses the power load decomposition component sequence to train the updated local model, the power load decomposition component sequence can be segmented according to the time step, and then the segmented data can be input into the local model as training data for training. In the specific training process, the following loss function can be used to calculate the loss value, and the loss value can be used to back-propagate and optimize the model parameters to complete the local training:

[0092]

[0093] in, Represents the loss value calculated for the kth client. The loss value is an indicator that measures the difference between the model prediction result and the true value, usually calculated by square error. Represents the total number of time steps of training data used by the kth client. It refers to the number of samples participating in training on this client. Represents the time step index, which is used to traverse the data of each time step of the client. For all time steps, represents the power load value predicted by the kth client at time step t, Represents the actual power load value of the kth client at time step t. That is, the actual observed data is used to compare with the predicted value.

[0094] It should be noted that after the client completes local training, the local model parameters obtained by training will be uploaded to the central server so that the central server can perform global aggregation to obtain global model parameters. However, since global parameter aggregation is required based on the dynamic matching and sorting cumulative distance of each client in this application, after the client uses the DTW algorithm to calculate the DTW distance between the meteorological temperature series and each IMF component, it is also necessary to calculate the sum of the DTW distances corresponding to all IMF components, and use the sum as the dynamic matching and sorting cumulative distance of the client, and upload the dynamic matching and sorting cumulative distance to the central server.

[0095] Based on this, the central server can perform global parameter aggregation through the following formula:

[0096]

[0097] in, represents the global model parameters after aggregation, Indicates the first The dynamic matching sorting cumulative distance of the clients, Indicates the number of clients participating in federated learning, , Indicates the first The local model parameters uploaded by the client.

[0098] It can be seen that after receiving the local model parameters of all clients, the central server uses weighted averaging to aggregate the parameters. The reason why the weight of each client is determined by the cumulative distance of its dynamic matching sorting is that the nodes with higher similarity (that is, the clients corresponding to the park) have a higher similarity period with the meteorological cycle, and the data quality can be more effective for the prediction task of the target node. Therefore, the cumulative distance is used as the weight.

[0099] In some other embodiments of the present application, the central server may also have a cumulative distance according to the dynamic matching sorting. Total cumulative distance The weighted value is assigned in inverse proportion to the distance The smaller the proportion, the greater the weighting parameter given to it.

[0100] It should be noted that after obtaining the global model parameters, the central server will send the global model parameters to each client so that the client can use the new global model parameters for the next round of local training.

[0101] Until the local model converges. Among them, the clients participating in federated learning belong to different parks and are used to predict power loads for different parks.

[0102] It should be further explained that when the client uses a converged local model to predict power load, the power load data of the park's historical time period (such as the past 7 days) can be input into the local model for prediction, and the power load forecast data for the future time (such as the day after the past 7 days) can be obtained.

[0103] In summary, the park power load forecasting method provided in the embodiment of the present application builds a new type of federated learning park load forecasting model based on the "data similarity dynamic matching-feature alignment extraction" structure, and uses this model to predict the future load of the park. In the data similarity dynamic matching stage, the power load data is first decomposed into multiple intrinsic mode functions using the empirical mode decomposition algorithm, each function represents a different frequency component of the original signal, eliminating the influence of data non-stationarity, and then the weather feature time series is used as the reference center, and the power load decomposition component is used as a comparison of feature extraction of different scales. The dynamic time planning algorithm is used for component matching to obtain the optimal order component sequence, realize data similarity alignment, and improve data utilization efficiency; in the feature alignment extraction stage, a convolutional neural network is used to extract different component combination features to generate more representative features; finally, an ordered federated learning framework is introduced to summarize all node processing results in a "data out of the domain, visible but unavailable" manner, increase the total amount of training data, and input the long short-term memory network to further mine the dynamic time series characteristics of the past and future time series to improve the exchange rate prediction accuracy.

[0104] The park power load forecasting method of the present application is exemplified below with reference to specific experimental data.

[0105] This experiment selected three industrial parks, all of which are food production and manufacturing industrial parks. They are identified as Industrial Park A, Industrial Park B, and Industrial Park C in the dataset. One of the industrial parks was selected as the target building, and the other two industrial parks were selected to form the building transmission network. The operating data of the three industrial parks were selected from January 1, 2016 to December 31, 2017 to ensure that all buildings were in the same time dimension.

[0106] In the experiment, the operating data of industrial park buildings were extracted as benchmark training data, with a time span from January 1, 2016 to December 31, 2016. One year was used as the training benchmark, and the operating power load of the target buildings in the industrial park in the next year was predicted as the test set, that is, from January 2017 to December 2017.

[0107] The comparison models are as follows:

[0108] Baseline federated learning model: The CNN-LSTM hybrid federated model is trained using the original decomposed component data without alignment and parameter aggregation is performed using parameter average.

[0109] Ordered federated learning model: The CNN-LSTM hybrid federated model is trained using the aligned original decomposed component data and the similarity-weighted average is used for parameter aggregation.

[0110] The evaluation indicators are root mean square error (RMSE) and mean absolute error (MAE). Specifically, the error results of the baseline federated learning model and the ordered federated learning model for park power load forecasting are shown in Table 1.

[0111] Table 1

[0112]

[0113] It can be seen from Table 1 that the accuracy of the park power load forecasting method based on the ordered federated learning model provided in this application is higher than the accuracy of the power load forecasting based on the benchmark federated learning model.

[0114] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for predicting park power load based on federated learning, characterized in that: include: The client participating in the federated learning obtains the power load data sequence of the park to which it belongs, and performs EMD decomposition on the power load data sequence to obtain multiple IMF components; The power load data sequence includes the power load data of the park in a historical time period; The client obtains a meteorological temperature sequence of the park, where the meteorological temperature sequence includes meteorological temperature data of the park within the historical time period; The client calculates the DTW distance between the meteorological temperature sequence and the IMF component using the DTW algorithm for each of the IMF components, and uses the DTW distance as the similarity between the meteorological temperature sequence and the IMF component; The client sorts the multiple IMF components based on the similarity to obtain a power load decomposition component sequence; The client receives the global model parameters sent by the central server; The client updates the local model according to the global model parameters. If the updated local model converges, the updated local model is used to predict the power load of the park; if the updated local model does not converge, the updated local model is locally trained using the power load decomposition component sequence, and the local model parameters of the trained local model are uploaded to the central server, so that the central server aggregates the global parameters based on the dynamic matching sorting cumulative distance uploaded by each client participating in the federated learning, and sends the aggregated global model parameters to the clients participating in the federated learning, and returns to the step of the client receiving the global model parameters sent by the central server, and the dynamic matching sorting cumulative distance is related to the DTW distance calculated by the client.

2. The method for predicting the power load of a park according to claim 1, characterized in that: The client sorts the multiple IMF components based on the similarity to obtain a power load decomposition component sequence, including: The multiple IMF components are sorted in order of similarity from small to large, and the sorted multiple IMF components are used as a power load decomposition component sequence.

3. The park power load forecasting method according to claim 1, characterized in that: After calculating the DTW distance between the meteorological temperature sequence and the IMF component by using the DTW algorithm, the park power load prediction method further includes: The sum of the DTW distances corresponding to all IMF components is calculated, and the sum is used as the dynamic matching sorting cumulative distance of the client.

4. The method for predicting the power load of a park according to claim 3, characterized in that: The central server aggregates global parameters based on the dynamic matching sorting cumulative distance uploaded by each client participating in the federated learning, including: The central server aggregates global parameters using the following formula: in, represents the global model parameters after aggregation, Indicates the first The dynamic matching sorting cumulative distance of the clients, Indicates the number of clients participating in federated learning, , Indicates the first The local model parameters uploaded by the client.

5. The method for predicting the power load of a park according to claim 1, characterized in that: The local model includes a CNN network, an LSTM network and a fully connected layer connected in sequence.

6. The method for predicting the power load of a park according to claim 5, characterized in that: When the client uses the power load decomposition component sequence to perform local training on the updated local model, the loss function used is: in, represents the loss value calculated for the k-th client, represents the total number of time steps of the training data used by the kth client, represents the time step index, represents the power load value predicted by the kth client at time step t, Represents the actual power load value of the kth client at time step t.

7. The method for predicting park power load according to claim 1, characterized in that: After the client updates the local model according to the global model parameters, the park power load prediction method further includes: The client determines whether the updated local model meets the convergence condition, and if so, determines that the updated local model converges; otherwise, determines that the updated local model does not converge; The convergence condition is that the number of training times of the local model reaches a preset number, or the prediction accuracy of the local model reaches a preset accuracy.

Images