A method and system for predicting user load at edge nodes based on federated learning

Abstract

This invention discloses a method and system for predicting user load at edge nodes based on federated learning. First, a federated learning architecture for predicting user load at edge nodes is constructed. Then, based on this architecture, a local load prediction model is trained on a mobile edge computing node (MEC) using collected local user load data, generating local model parameters. These local model parameters are then encrypted to obtain encrypted model parameters. This invention achieves accurate prediction of user load at edge nodes by using local data that does not leave the domain and only transmitting encrypted model parameters. Furthermore, by introducing a dirty model suppression strategy based on a similarity matrix, the robustness of the federated learning framework in heterogeneous edge environments is significantly enhanced. Simultaneously, it avoids the damage to global performance caused by low-quality or malicious model updates, ensuring the privacy and security of user electricity data and meeting the stringent data compliance requirements of the power system.

Description

Technical Field

[0001] This invention relates to the field of power system user-side load forecasting technology, specifically to an edge node user load forecasting method and system based on federated learning. Background Technology

[0002] With the development of smart grids, user-side load data is showing a massive growth trend, and accurate load forecasting is the key to achieving optimized power dispatch and efficient energy utilization.

[0003] Currently, existing centralized load forecasting methods generally require the collection of raw load data from each user, which increases the server processing burden and poses a serious risk of user privacy leakage. At the same time, the load data of user-side edge nodes is highly volatile and of inconsistent quality, resulting in significant differences in the training effect of local models. This not only reduces the accuracy of the global model but also slows down the model convergence speed, seriously affecting the load forecasting effect. Therefore, it is necessary to design an edge node user load forecasting method and system based on federated learning. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and to better and more effectively address the problems of existing centralized load forecasting methods, which generally require the collection of raw load data from each user, thus increasing the server processing burden and posing a serious risk of user privacy leakage. Furthermore, the load data of user-side edge nodes exhibits large fluctuations and inconsistent quality, leading to significant differences in local model training effects. This not only reduces the accuracy of the global model but also slows down model convergence, severely impacting load forecasting performance. This invention provides a method and system for edge node user load forecasting based on federated learning. It achieves accurate prediction of edge node user load using a mechanism that allows local data to remain within the domain and only transmits encrypted model parameters. Moreover, by introducing a dirty model suppression strategy based on a similarity matrix, it significantly enhances the robustness of the federated learning framework in heterogeneous edge environments, while avoiding the damage to global performance caused by low-quality or malicious model updates. This not only protects the privacy and security of user electricity data but also meets the stringent data compliance requirements of the power system.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting user load at edge nodes based on federated learning includes the following steps:

[0007] Step A: Construct a federated learning architecture for edge node user load prediction;

[0008] Step B: Based on the federated learning architecture, the local load prediction model is trained on the mobile edge computing node (MEC) according to the collected local user load data, and local model parameters are generated.

[0009] Step C: Encrypt the local model parameters and obtain the encrypted model parameters, then upload the encrypted model parameters to the cloud server Cloud;

[0010] Step D: Receive encrypted model parameters from each mobile edge computing node (MEC) using the cloud server Cloud, then filter the uploaded encrypted model parameters based on the similarity matrix to obtain the filtered model parameters.

[0011] Step E: Use a dynamic weighting mechanism to perform weighted aggregation on the filtered model parameters and obtain the aggregated model parameters;

[0012] Step F: The aggregated model parameters are sent to each mobile edge computing node (MEC), and then the mobile edge computing node (MEC) updates its local load forecasting model based on the aggregated model parameters and obtains the updated local load forecasting model.

[0013] Step G involves using the updated local load forecasting model to predict the user load in the area where the edge node is located and outputting the load forecast values ​​for future periods.

[0014] The aforementioned edge node user load forecasting method based on federated learning includes step A: constructing a federated learning architecture for edge node user load forecasting. This architecture comprises a smart meter (SM), a mobile edge computing node (MEC), a cloud server (Cloud), a power supplier (PS), and a certificate authority (CA). The smart meter (SM) collects electricity consumption data from the user side. The MEC performs local load forecasting model training and serves as a client training node for federated learning. The cloud server (Cloud) filters and aggregates relevant model parameters using federated learning. The power supplier (PS) publishes and manages short-term load forecasting services. The certificate authority (CA) authenticates the mobile edge computing node (MEC) and the smart meter (SM) and distributes secure communication certificates.

[0015] In the aforementioned edge node user load prediction method based on federated learning, step B involves training a local load prediction model on the mobile edge computing node (MEC) based on the collected local user load data, and generating local model parameters. Specifically, the local load prediction model adopts a lightweight neural network model, which includes two LSTM layers and one fully connected layer. The two LSTM layers are used to capture the short-term fluctuation characteristics and long-term periodic patterns of the load, respectively. The fully connected layer uses the ReLU activation function and is used for nonlinear mapping.

[0016] In the aforementioned edge node user load prediction method based on federated learning, step C involves encrypting the local model parameters to obtain encrypted model parameters, and then uploading the encrypted model parameters to the cloud server Cloud. Specifically, the encryption of the local model parameters employs homomorphic encryption or differential privacy mechanisms.

[0017] In the aforementioned edge node user load prediction method based on federated learning, step D involves using a cloud server (Cloud) to receive encrypted model parameters from each mobile edge computing node (MEC). The uploaded encrypted model parameters are then filtered based on a similarity matrix to obtain the filtered model parameters. Specifically, the filtering of uploaded encrypted model parameters involves calculating the cosine similarity between the encrypted model parameters of each mobile edge computing node MEC and constructing a similarity matrix. A dynamic threshold is then set, and encrypted model parameters that significantly differ from the update direction of the mainstream local load prediction model or whose similarity is below the dynamic threshold are removed. The specific steps are as follows.

[0018] Step D1: Calculate the cosine similarity between the parameters of the MEC encryption model of each mobile edge computing node and construct a similarity matrix, as shown in formula (1).

[0019] (1)

[0020] in, For the cosine similarity between the parameters of the encryption model, and The first The model parameter vectors of the j-th and j-th edge nodes;

[0021] Step D2: Set a dynamic threshold and remove parameters of the encrypted model that differ significantly from the update direction of the mainstream local load forecasting model and whose similarity is lower than the dynamic threshold, as shown in formula (2).

[0022] (2)

[0023] in, For dynamic thresholds, and These represent the average similarity and standard deviation of all edge computing nodes (MECs) in the current round. This is the adjustment coefficient.

[0024] In the aforementioned edge node user load prediction method based on federated learning, step E involves using a dynamic weighting mechanism to weight and aggregate the filtered model parameters to obtain the aggregated model parameters. Specifically, the dynamic weighting mechanism calculates the aggregation weights of each Mobile Edge Computing (MEC) node. Specifically, as shown in formula (3),

[0025] (3)

[0026] in, Let be the local effective data volume of the k-th mobile edge computing node (MEC). This is the inverse coefficient of variation of the gradient change during local training. and All of these are adjustable hyperparameters.

[0027] In the aforementioned edge node user load prediction method based on federated learning, step F involves distributing the aggregated model parameters to each mobile edge computing node (MEC). The MEC then updates its local load prediction model based on the aggregated model parameters, obtaining the updated local load prediction model. The updating of the local load prediction model based on the aggregated model parameters is shown in formula (4).

[0028] (4)

[0029] in, To update the model parameters of the local load forecasting model, To update the model parameters of the local load forecasting model before the update, These are the aggregated model parameters for the k-th Mobile Edge Computing (MEC) node.

[0030] In the aforementioned edge node user load prediction method based on federated learning, step G involves using an updated local load prediction model to predict the user load in the area where the edge node is located and outputting the load prediction value for the future period. The local load prediction model includes an input layer and an output layer. The input layer of the local load prediction model is used to receive historical load sequences, time characteristics, and meteorological data from the past 96 time steps. The output layer of the local load prediction model is used to output the load prediction value for the next 24 hours.

[0031] A federated learning-based edge node user load prediction system includes an architecture building module, a parameter generation module, a parameter encryption module, a parameter filtering module, a parameter aggregation module, a model update module, and a load prediction module. The architecture building module constructs a federated learning architecture for edge node user load prediction. The parameter generation module trains a local load prediction model on a mobile edge computing node (MEC) based on collected local user load data and generates local model parameters. The parameter encryption module encrypts the local model parameters to obtain encrypted model parameters, which are then uploaded to a cloud server (Cloud). The parameter filtering module utilizes the cloud server (Cloud) to... The system receives encrypted model parameters from each Mobile Edge Computing Node (MEC), filters the uploaded encrypted model parameters based on a similarity matrix, and obtains the filtered model parameters. The parameter aggregation module uses a dynamic weighting mechanism to perform weighted aggregation on the filtered model parameters and obtains the aggregated model parameters. The model update module distributes the aggregated model parameters to each Mobile Edge Computing Node (MEC), and the Mobile Edge Computing Node (MEC) updates its local load forecasting model based on the aggregated model parameters and obtains the updated local load forecasting model. The load forecasting module uses the updated local load forecasting model to predict the user load in the area where the edge node is located and outputs the load forecast value for future periods.

[0032] The beneficial effects of this invention are as follows: The edge node user load prediction method and system based on federated learning first constructs a federated learning architecture for edge node user load prediction. Then, based on the federated learning architecture, a local load prediction model is trained on a mobile edge computing node (MEC) using collected local user load data, generating local model parameters. Next, the local model parameters are encrypted to obtain encrypted model parameters, which are then uploaded to a cloud server (Cloud). Subsequently, the cloud server (Cloud) receives the encrypted model parameters from each mobile edge computing node (MEC), filters the uploaded encrypted model parameters based on a similarity matrix, and obtains filtered model parameters. Then, a dynamic weighting mechanism is used to weight and aggregate the filtered model parameters to obtain aggregated model parameters. Finally, the aggregated model parameters are distributed to each mobile edge computing node (MEC), and the mobile edge computing node (MEC) updates its local load prediction model based on the aggregated model parameters. The system obtains an updated local load forecasting model and then uses this model to predict the user load in the area where the edge node is located, outputting the load forecast value for future periods. This effectively realizes the edge node user load forecasting method and system's ability to accurately predict edge node user load using a mechanism that allows local data to remain within the domain and only transmits encrypted model parameters. Furthermore, by introducing a dirty model suppression strategy based on a similarity matrix, the robustness of the federated learning framework in heterogeneous edge environments is significantly enhanced, while avoiding the destruction of global performance by low-quality or malicious model updates. The dynamic weight aggregation mechanism fully considers the differences in data distribution and historical performance of each edge node, thereby improving the convergence speed and final prediction accuracy of the local load forecasting model. This not only ensures the privacy and security of user electricity data but also meets the stringent data compliance requirements of the power system, making it suitable for distributed load forecasting tasks in smart distribution networks, virtual power plants, and demand response scenarios. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of an edge node user load prediction method based on federated learning according to the present invention;

[0034] Figure 2 This is a schematic diagram of the federated learning architecture of the present invention;

[0035] Figure 3 This is a schematic diagram illustrating the principle of edge node user load prediction in this invention;

[0036] Figure 4 This is a flowchart of the operation of an edge node user load prediction system based on federated learning according to the present invention.

[0037] Figure 5 This is a schematic diagram illustrating the operating principle of the local load forecasting model of the present invention. Detailed Implementation

[0038] The present invention will now be further described with reference to the accompanying drawings.

[0039] like Figure 1 As shown, the present invention provides a method for predicting the user load of edge nodes based on federated learning, which includes the following steps:

[0040] like Figure 2 As shown, step A involves constructing a federated learning architecture for edge node user load forecasting. This architecture includes a smart meter (SM), a mobile edge computing node (MEC), a cloud server (Cloud), a power supplier (PS), and a certificate authority (CA). The smart meter (SM) collects electricity consumption data from the user side. The mobile edge computing node (MEC) performs local load forecasting model training and serves as a client training node for federated learning. The cloud server (Cloud) filters and aggregates the corresponding model parameters using federated learning. The power supplier (PS) publishes and manages short-term load forecasting services. The certificate authority (CA) authenticates the mobile edge computing node (MEC) and the smart meter (SM) and distributes secure communication certificates.

[0041] Step B involves training a local load prediction model and generating local model parameters on the mobile edge computing node (MEC) based on the collected local user load data, using a federated learning architecture. Specifically, the local load prediction model employs a lightweight neural network model, which includes two LSTM layers and one fully connected layer. The two LSTM layers are used to capture the short-term fluctuation characteristics and long-term periodic patterns of the load, respectively. The fully connected layer uses the ReLU activation function and is used for nonlinear mapping.

[0042] Step C involves encrypting the local model parameters and obtaining the encrypted model parameters, then uploading the encrypted model parameters to the cloud server Cloud. Specifically, the encryption of the local model parameters uses homomorphic encryption or differential privacy mechanisms.

[0043] like Figure 3 As shown, in step D, the cloud server (Cloud) receives the encrypted model parameters from each mobile edge computing (MEC) node. Then, based on the similarity matrix, the uploaded encrypted model parameters are filtered to obtain the filtered model parameters. Specifically, the filtering of uploaded encrypted model parameters involves calculating the cosine similarity between the encrypted model parameters of each MEC node and constructing a similarity matrix. A dynamic threshold is then set, and encrypted model parameters that significantly differ from the update direction of the mainstream local load forecasting model and whose similarity is below the dynamic threshold are removed. The specific steps are as follows:

[0044] Step D1: Calculate the cosine similarity between the parameters of the MEC encryption model of each mobile edge computing node and construct a similarity matrix, as shown in formula (1).

[0045] (1)

[0046] in, For the cosine similarity between the parameters of the encryption model, and The first The model parameter vectors of the j-th and j-th edge nodes;

[0047] Step D2: Set a dynamic threshold and remove parameters of the encrypted model that differ significantly from the update direction of the mainstream local load forecasting model and whose similarity is lower than the dynamic threshold, as shown in formula (2).

[0048] (2)

[0049] in, For dynamic thresholds, and These represent the average similarity and standard deviation of all edge computing nodes (MECs) in the current round. This is the adjustment coefficient.

[0050] Step E involves using a dynamic weighting mechanism to weight and aggregate the filtered model parameters to obtain the aggregated model parameters. Specifically, the dynamic weighting mechanism calculates the aggregation weight of each Mobile Edge Computing (MEC) node. Specifically, as shown in formula (3),

[0051] (3)

[0052] in, Let be the local effective data volume of the k-th mobile edge computing node (MEC). This is the inverse coefficient of variation of the gradient change during local training. and All of these are adjustable hyperparameters.

[0053] Step F involves distributing the aggregated model parameters to each Mobile Edge Computing Node (MEC). The MEC then updates its local load forecasting model based on the aggregated model parameters, obtaining the updated local load forecasting model. The update of the local load forecasting model based on the aggregated model parameters is shown in formula (4).

[0054] (4)

[0055] in, To update the model parameters of the local load forecasting model, To update the model parameters of the local load forecasting model before the update, These are the aggregated model parameters for the k-th Mobile Edge Computing (MEC) node.

[0056] like Figure 5 As shown, in step G, the updated local load forecasting model is used to forecast the user load in the area where the edge node is located and output the load forecast value for the future period. The local load forecasting model includes an input layer and an output layer. The input layer of the local load forecasting model is used to receive historical load sequences, time characteristics and meteorological data of the past 96 time steps. The output layer of the local load forecasting model is used to output the load forecast value for the next 24 hours.

[0057] like Figure 4 As shown, an edge node user load prediction system based on federated learning includes an architecture building module, a parameter generation module, a parameter encryption module, a parameter filtering module, a parameter aggregation module, a model update module, and a load prediction module. The architecture building module constructs a federated learning architecture for edge node user load prediction. The parameter generation module trains a local load prediction model on the mobile edge computing node (MEC) based on collected local user load data and generates local model parameters. The parameter encryption module encrypts the local model parameters to obtain encrypted model parameters, which are then uploaded to a cloud server (Cloud). The parameter filtering module utilizes the cloud server (Cloud) to... The ud receives encrypted model parameters from each mobile edge computing node (MEC), then filters the uploaded encrypted model parameters based on a similarity matrix to obtain filtered model parameters. The parameter aggregation module uses a dynamic weighting mechanism to perform weighted aggregation on the filtered model parameters to obtain aggregated model parameters. The model update module distributes the aggregated model parameters to each mobile edge computing node (MEC), and then the mobile edge computing node (MEC) updates its local load forecasting model based on the aggregated model parameters to obtain the updated local load forecasting model. The load forecasting module uses the updated local load forecasting model to predict the user load in the area where the edge node is located and outputs the load forecast value for future periods.

[0058] To better illustrate the effectiveness of the present invention, a specific embodiment of using the method of the present invention to predict the user load of edge nodes is described below.

[0059] This embodiment designs three sets of comparative experiments to compare the method of the present invention with two typical prior art techniques:

[0060] Centralized learning methods, representing traditional approaches, involve uploading the raw load data of all edge nodes to a central server, where a global model is trained uniformly. Standard federated averaging algorithms, representing basic federated learning methods, involve each edge node training locally and then uploading its model parameters. The central server performs a simple weighted average aggregation without any filtering or dynamic weighting.

[0061] The dataset uses real smart meter data from 12 typical residential areas in a provincial capital city, with a sampling frequency of 15 minutes and a time span of 6 months. The data package contains information such as power consumption, temperature, and humidity.

[0062] The implementation scenario simulates an edge computing environment, with each residential area corresponding to an edge node (MEC) and the central server (Cloud) responsible for model aggregation;

[0063] Evaluation metrics include root mean square error (RMSE), which measures prediction accuracy (the smaller the value, the better); and mean absolute error (MAE), which measures prediction bias (the smaller the value, the better).

[0064] Communication Rounds: The number of iterations required for the model to converge. A smaller value indicates faster convergence.

[0065] Privacy protection level, qualitatively assessing the risk of privacy leakage during data transmission;

[0066] In the test conditions, during the 30th round of training, abnormal data (such as constant zero values ​​and periodic impulse noise caused by equipment failure) were artificially injected into two edge nodes (Node 7 and Node 11) to simulate a "dirty model" interference scenario.

[0067] The comparison results are shown in Table 1 below.

[0068] Table 1. User load forecast results for edge nodes;

[0069]

[0070] Results analysis;

[0071] (1) Compared with centralized learning, the method of this invention has comparable accuracy but better privacy: the RMSE (8.6 kW) and MAE (6.4 kW) of the method of this invention are very close to those of centralized learning (8.1 kW, 6.0 kW), indicating that its prediction accuracy is competitive. However, the most critical difference lies in privacy protection: centralized learning requires uploading the original electricity consumption data of all users, which poses a serious risk of privacy leakage; while the method of this invention only transmits encrypted model parameters throughout the process, and the original data does not leave the local domain, thus fundamentally solving the privacy problem.

[0072] (2) Compared with the standard federated average, the method of the present invention converges faster, requiring only 60 rounds to converge, which reduces communication overhead by about 30% compared with the 85 rounds of the standard federated average. This is due to the dynamic weight mechanism, which gives nodes with high data quality and large contributions greater say in the aggregation, thereby accelerating the optimization process of the global model.

[0073] The method of this invention has higher prediction accuracy. Under normal, interference-free conditions, the RMSE and MAE of this method are superior to those of the standard federated method. This indicates that the similarity matrix filtering eliminates potential low-quality updates, and the combination of dynamic weighting makes the aggregation process more scientific, avoiding the drag from invalid or negative updates.

[0074] The results of this comparative embodiment clearly demonstrate that, compared with the prior art, the edge node user load prediction method based on federated learning provided by the present invention has the following significant advantages: in terms of privacy protection, it is superior to centralized learning, achieving "data available but not visible"; in terms of convergence speed and prediction accuracy, it is superior to standard federated learning methods; in terms of system robustness, it can effectively resist "dirty model" interference and ensure reliable operation in complex and heterogeneous edge environments.

[0075] In summary, the edge node user load prediction method and system based on federated learning of the present invention first constructs a federated learning architecture for edge node user load prediction. Then, based on the federated learning architecture, a local load prediction model is trained on the mobile edge computing node (MEC) using collected local user load data, generating local model parameters. Next, the local model parameters are encrypted to obtain encrypted model parameters, which are then uploaded to a cloud server (Cloud). Subsequently, the cloud server receives the encrypted model parameters from each mobile edge computing node (MEC), filters the uploaded encrypted model parameters based on a similarity matrix, and obtains filtered model parameters. Then, a dynamic weighting mechanism is used to weight and aggregate the filtered model parameters to obtain aggregated model parameters. Finally, the aggregated model parameters are distributed to each mobile edge computing node (MEC), and the mobile edge computing node (MEC) updates its local load prediction model based on the aggregated model parameters. This invention effectively achieves the function of accurately predicting the user load of edge nodes by using a mechanism that allows local data to remain within the domain and only transmits encrypted model parameters. Furthermore, by introducing a dirty model suppression strategy based on a similarity matrix, the robustness of the federated learning framework in heterogeneous edge environments is significantly enhanced, while avoiding the destruction of global performance by low-quality or malicious model updates. The dynamic weight aggregation mechanism fully considers the differences in data distribution and historical performance of each edge node, thereby improving the convergence speed and final prediction accuracy of the local load prediction model. This not only ensures the privacy and security of user electricity data but also meets the strict requirements of power system data compliance, making it suitable for distributed load prediction tasks in smart distribution networks, virtual power plants, and demand response scenarios.

[0076] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting user load at edge nodes based on federated learning, characterized in that: Includes the following steps, Step A: Construct a federated learning architecture for edge node user load forecasting. This architecture includes a smart meter (SM), a mobile edge computing node (MEC), a cloud server (Cloud), a power supplier (PS), and a certificate authority (CA). The smart meter (SM) collects electricity consumption data from the user side. The mobile edge computing node (MEC) performs local load forecasting model training and serves as a client training node for federated learning. The cloud server (Cloud) filters and aggregates the corresponding model parameters using federated learning. The power supplier (PS) publishes and manages short-term load forecasting services. The certificate authority (CA) authenticates the mobile edge computing node (MEC) and the smart meter (SM) and distributes secure communication certificates. Step B involves training a local load prediction model and generating local model parameters on the mobile edge computing node (MEC) based on the collected local user load data, using a federated learning architecture. Specifically, the local load prediction model employs a lightweight neural network model, which includes two LSTM layers and a fully connected layer. The two LSTM layers are used to capture the short-term fluctuation characteristics and long-term periodic patterns of the load, respectively. The fully connected layer uses the ReLU activation function and is used for nonlinear mapping. Step C involves encrypting the local model parameters and obtaining the encrypted model parameters, then uploading the encrypted model parameters to the cloud server Cloud. Specifically, the encryption of the local model parameters uses either homomorphic encryption or differential privacy mechanisms. Step D involves using the cloud server (Cloud) to receive the encrypted model parameters from each Mobile Edge Computing (MEC) node. Then, based on a similarity matrix, the uploaded encrypted model parameters are filtered to obtain the filtered model parameters. Specifically, the filtering process involves calculating the cosine similarity between the encrypted model parameters of each MEC node and constructing a similarity matrix. A dynamic threshold is then set, and encrypted model parameters that significantly differ from the update direction of the mainstream local load forecasting model or have a similarity below the dynamic threshold are removed. The specific steps are as follows: Step D1: Calculate the cosine similarity between the parameters of the MEC encryption model of each mobile edge computing node and construct a similarity matrix, as shown in formula (1). （1） in, To determine the cosine similarity between parameters of the encryption model, and The first The model parameter vectors of the j-th and j-th edge nodes; Step D2: Set a dynamic threshold and remove parameters of the encrypted model that differ significantly from the update direction of the mainstream local load forecasting model and whose similarity is lower than the dynamic threshold, as shown in formula (2). （2） in, For dynamic thresholds, and These represent the average similarity and standard deviation of all edge computing nodes (MECs) in the current round. This is the adjustment coefficient; Step E involves using a dynamic weighting mechanism to weight and aggregate the filtered model parameters to obtain the aggregated model parameters. Specifically, the dynamic weighting mechanism calculates the aggregation weight of each Mobile Edge Computing (MEC) node. Specifically, as shown in formula (3), （3） in, Let be the local effective data volume of the k-th mobile edge computing node (MEC). This is the inverse coefficient of variation of the gradient change during local training. and All are adjustable hyperparameters; Step F involves distributing the aggregated model parameters to each Mobile Edge Computing Node (MEC). The MEC then updates its local load forecasting model based on the aggregated model parameters, obtaining the updated local load forecasting model. The update of the local load forecasting model based on the aggregated model parameters is shown in formula (4). （4） in, To update the model parameters of the local load forecasting model, To update the model parameters of the local load forecasting model before the update, The aggregated model parameters for the k-th mobile edge computing node (MEC); Step G: The updated local load forecasting model is used to forecast the user load in the area where the edge node is located and output the load forecast value for the future period. The local load forecasting model includes an input layer and an output layer. The input layer of the local load forecasting model is used to receive historical load sequences, time characteristics and meteorological data of the past 96 time steps. The output layer of the local load forecasting model is used to output the load forecast value for the next 24 hours.

2. An edge node user load prediction system based on federated learning, wherein the specific prediction process of the edge node user load prediction system is based on the edge node user load prediction method of claim 1, characterized in that: It includes an architecture building module, a parameter generation module, a parameter encryption module, a parameter filtering module, a parameter aggregation module, a model update module, and a load prediction module. The architecture building module constructs a federated learning architecture for edge node user load prediction. The parameter generation module is used to train a local load prediction model and generate local model parameters at the mobile edge computing node (MEC) based on the collected local user load data. The parameter encryption module is used to encrypt the local model parameters and obtain the encrypted model parameters, and then upload the encrypted model parameters to the cloud server Cloud. The parameter filtering module is used to receive encrypted model parameters from each mobile edge computing node (MEC) using the cloud server Cloud, and then filter the uploaded encrypted model parameters based on the similarity matrix to obtain the filtered model parameters. The parameter aggregation module is used to perform weighted aggregation of the filtered model parameters using a dynamic weighting mechanism and obtain the aggregated model parameters. The model update module is used to send the aggregated model parameters to each mobile edge computing node (MEC), and then the mobile edge computing node (MEC) updates the local load prediction model according to the aggregated model parameters and obtains the updated local load prediction model. The load forecasting module is used to predict the user load in the area where the edge node is located using the updated local load forecasting model and output the load forecast value for future periods.

Images