A personalized federated learning prediction method for distributed photovoltaic power generation sites

Abstract

The present invention discloses a personalized federated learning prediction method for distributed photovoltaic power generation sites, which relates to the field of artificial intelligence technology, including a multi-layer time series decomposition module and a multi-layer perceptron module; specifically including the following steps: S1: data preprocessing, normalizing the data by the installed capacity of the site to eliminate the differences caused by the scales of different photovoltaic sites; S2: local training, the client trains the model locally; S3: personalized federated aggregation, the client uploads the multi-layer time series decomposition module to the server, the server performs average aggregation, and sends the aggregated global model to the server; S4: perform the next round of iteration until the final iteration round. The present invention adopts the above-mentioned personalized federated learning prediction method for distributed photovoltaic power generation sites, takes the last multi-layer perceptron layer as the personalized layer, and considers the differences between sites through the personalized method of hierarchical aggregation, which has higher accuracy than direct federated aggregation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a personalized federated learning prediction method for distributed photovoltaic power generation sites. Background Art

[0002] As the scale of photovoltaic power generation continues to expand, distributed photovoltaic power stations emerge in an endless stream. Most current methods are aimed at large photovoltaic sites and distributed small photovoltaic units that share data. However, considering the data security and privacy laws that have been enacted, centralized training methods will destroy the privacy of data, and it is impossible to train models using shared data. In addition, there is another very important issue. In the absence of sufficient training data, deep learning methods are prone to overfitting. The main problems are as follows:

[0003] 1) Newly built PV sites often lack data, and collecting data from multiple sites for centralized training will result in privacy leakage. When training only on local data, it will lead to lower model accuracy. Current research rarely focuses on newly built PV sites.

[0004] 2) Standardized federated learning does not take into account the differences in photovoltaic conversion efficiency between sites. The same model is used for all sites, and the prediction accuracy is not high enough.

[0005] Therefore, it is necessary to provide a personalized federated learning prediction method for distributed photovoltaic power generation sites to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to provide a personalized federated learning prediction method for distributed photovoltaic power generation sites. The multi-layer time series decomposition module decomposes the original input data layer by layer, and aggregates them after prediction, and finally forms deep features, which can efficiently process complex photovoltaic data; the personalized federal strategy, under the premise of protecting privacy, solves the problem of poor prediction effect of single training of new sites through the method of federated learning, and uses the last multi-layer perceptron layer as the personalized layer. Through the personalized method of hierarchical aggregation, the differences between sites are considered, which has higher accuracy than direct federal aggregation.

[0007] To achieve the above object, the present invention provides a personalized federated learning prediction method for distributed photovoltaic power generation sites, including a multi-layer time series decomposition module and a multi-layer perceptron module; specifically, the following steps are included:

[0008] S1: Data preprocessing, normalizing the data by the installed capacity of the site to eliminate the differences caused by different PV site sizes;

[0009] S2: Local training, the client trains the model locally;

[0010] S3: Personalized federated aggregation. The client uploads the multi-layer time series decomposition module to the server. The server performs average aggregation and sends the aggregated global model to the server.

[0011] S4: Perform the next round of iteration until the final iteration round.

[0012] Preferably, the multi-layer time series decomposition module uses a multi-layer perceptron as a network structure and combines it with a Gaussian filter to perform multi-layer time series decomposition. The multi-layer time series decomposition module is a global model, and the global model is a shared model used to extract underlying common features. The encoder and decoder in the multi-layer time series decomposition module are both composed of multi-layer perceptrons, which are superimposed in series through n multi-layer time series decomposition modules.

[0013] Preferably, the multi-layer perceptron module is a local model, which is used to extract the characteristics of the photovoltaic site itself and process the differences between photovoltaic sites. The local model is a private model, which maps the extracted underlying common features to photovoltaic power generation.

[0014] Preferably, in step S2, the following steps are specifically included:

[0015] S21: Input the data into the Gaussian filter to decompose the data to obtain the trend component and the volatility component, and then use the multi-layer perceptron to predict the trend component and the volatility component respectively;

[0016] S22: Calculate the residual between the prediction result and the input as the input of the next multi-layer time series decomposition module;

[0017] S23: Multiple multi-layer time series decomposition modules are superimposed to decompose complex photovoltaic power generation data;

[0018] S24: The prediction result of the multi-layer time series decomposition module is input into the final multi-layer perceptron module to obtain the final prediction result.

[0019] Preferably, all clients have models, and client i has models It consists of two parts: the global model and local models Written as:

[0020]

[0021] Global Model It is shared by all models and is used to extract the underlying common features and local models. It is private to each model and maps the extracted common features to the required PV power generation;

[0022] N models correspond to N photovoltaic sites. During the local training process, the site uses the local data set to train the corresponding model for E rounds, and then sends the global model to the server for aggregation. The server sends the aggregated global model to the corresponding client and enters the next round of iteration. During the aggregation process, the change values ​​of the parameters of the global model are averaged, which is expressed as:

[0023]

[0024] Here, k represents the k-th round of aggregation, and the aggregation results in the change value of the global model weight of the k+1th round that needs to be issued.

[0025] Therefore, the present invention adopts the above-mentioned personalized federated learning prediction method for distributed photovoltaic power generation sites, which has the following beneficial effects:

[0026] (1) The present invention has higher data security: through the method of federated learning, the original training data can be not transmitted during the model training process, which can well protect privacy.

[0027] (2) The present invention has strong complex data processing capabilities: by decomposing data through a multi-layer time series decomposition module, data features can be extracted well, and the problem of large data fluctuations can be alleviated. The model accuracy far exceeds that of the classic model, and while exceeding the accuracy of the latest model, the model training speed is greatly improved.

[0028] (3) The present invention can handle the differences in photovoltaic conversion efficiency of sites: Taking into account the different photovoltaic conversion efficiencies of different photovoltaic sites, each photovoltaic site is personalized by setting a personalized layer, which has better prediction accuracy.

[0029] (4) The present invention can effectively improve the accuracy of photovoltaic site prediction in terms of data security, complex data processing and site personalized processing.

[0030] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a personalized federated learning prediction method for distributed photovoltaic power generation sites of the present invention;

[0032] Figure 2 It is a flowchart of a photovoltaic power generation prediction model based on personalized federated learning of the present invention;

[0033] Figure 3 It is a comparison chart of RMSE experimental results of the present invention and the existing model;

[0034] Figure 4It is a comparison chart of MAE experimental results of the present invention and the existing model;

[0035] Figure 5 This is a comparison chart of the training time of the present invention and the existing model. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0037] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0038] The words "include" or "comprises" and the like used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms "inside", "outside", "upper", "lower", etc. is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly specified and limited, the terms "attachment" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral body; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0039] Example

[0040] like Figure 1 As shown, the present invention provides a personalized federated learning prediction method for distributed photovoltaic power generation sites, including a multi-layer time series decomposition module and a multi-layer perceptron module; specifically including the following steps, such as Figure 2 As shown:

[0041] S1: Data preprocessing, normalizing the data by the installed capacity of the site to eliminate the differences caused by different PV site sizes;

[0042] S2: Local training: The client trains the model locally. In step S2, the following steps are specifically included:

[0043] S21: Input the data into the Gaussian filter to decompose the data to obtain the trend component and the volatility component, and then use the multi-layer perceptron to predict the trend component and the volatility component respectively;

[0044] S22: Calculate the residual between the prediction result and the input as the input of the next multi-layer time series decomposition module;

[0045] S23: Multiple multi-layer time series decomposition modules are superimposed to decompose complex photovoltaic power generation data;

[0046] S24: The prediction result of the multi-layer time series decomposition module is input into the final multi-layer perceptron module to obtain the final prediction result.

[0047] S3: Personalized federated aggregation. The client uploads the multi-layer time series decomposition module to the server. The server performs average aggregation and sends the aggregated global model to the server.

[0048] S4: Perform the next round of iteration until the final iteration round.

[0049] The multi-layer time series decomposition module uses the multi-layer perceptron as the network structure and combines it with the Gaussian filter to perform multi-layer time series decomposition. The multi-layer time series decomposition module is a global model. The global model is a shared model used to extract the underlying common features. The encoder and decoder in the multi-layer time series decomposition module are composed of multi-layer perceptrons, which are superimposed in series through n multi-layer time series decomposition modules. Where n means that n multi-layer time series decomposition modules are superimposed in series, thereby increasing the complexity of the model and being able to process complex data.

[0050] The multi-layer perceptron module is a local model, which is used to extract the characteristics of the photovoltaic site itself and process the differences between photovoltaic sites. The local model is a private model that maps the extracted underlying common features to photovoltaic power generation. Figure 1 As shown in the figure, the multilayer perceptron module is a local model that does not require federated aggregation and can handle the differences between photovoltaic sites. The multilayer perceptron is a simple network with weak fitting ability. The prediction effect is poor when processing time series tasks by simple stacking, but the multilayer perceptron can predict individual signal quantities well. Since the complexity of the multilayer perceptron itself is relatively low, the complexity after multilayer stacking is still much lower than that of the Transformer model, and the prediction accuracy is greatly improved.

[0051] All clients have models, and client i has models It consists of two parts: the global model and local models Written as:

[0052]

[0053] Global Model It is shared by all models and is used to extract the underlying common features and local models. It is private to each model and maps the extracted common features to the required PV power generation;

[0054] N models correspond to N photovoltaic sites. During the local training process, the site uses the local data set to train the corresponding model for E rounds, and then sends the global model to the server for aggregation. The server sends the aggregated global model to the corresponding client and enters the next round of iteration. During the aggregation process, the change values ​​of the parameters of the global model are averaged, which is expressed as:

[0055]

[0056] Among them, k represents the kth round of aggregation, and the aggregated value is the change value of the global model weight in the k+1th round that needs to be issued. The remaining models are local and do not need to be aggregated, which can make the model personalized and able to handle the problem of different photovoltaic conversion efficiencies at different sites.

[0057] Comparative Example

[0058] The comparison of the root mean square error (RMSE) experimental results of this scheme and the existing model is shown in the figure below. Figure 3 As shown in the figure, the comparison of the mean absolute error (MAE) experimental results of the present invention and the existing model is shown in Figure 4 As shown in the figure, MLP represents the multi-layer perceptron model, LSTM represents the long short-term memory network model, CNNLSTM represents the model combining the convolutional neural network and the long short-term memory network, and Koopa is a model based on the Koopaman operator. L, F, and P represent training only locally, training under ordinary federated aggregation conditions, and training under personalized federated aggregation conditions, respectively. The smaller the values ​​of RMSE and MAE, the better the model prediction effect.

[0059] This solution and the existing model are trained with the same 6-month training data. The comparison of the training time for one round is shown in the figure below. Figure 5 As shown in Figure 2, the training time of this scheme is much shorter than that of the Koopa algorithm. Although the training time of this scheme is longer than that of other algorithms, both RMSE and MAE have significant advantages.

[0060] From the experimental results, it can be found that the results of this scheme are significantly better than other models in the case of local training and ordinary federated training, and the results are significantly better than local training and ordinary federated training in the case of personalized federated training, which proves the effectiveness of this scheme.

[0061] Therefore, the present invention adopts the above-mentioned personalized federated learning prediction method for distributed photovoltaic power generation sites, which can effectively improve the accuracy of photovoltaic site prediction in terms of data security, complex data processing and site personalized processing.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A personalized federated learning prediction method for distributed photovoltaic power generation sites, characterized by: It includes a multi-layer time series decomposition module and a multi-layer perceptron module; specifically includes the following steps: S1: Data preprocessing, normalizing the data by the installed capacity of the site to eliminate the differences caused by different PV site sizes; S2: Local training, the client trains the model locally; All clients have models, and client i has models , It consists of two parts: the global model and local models , written as: Global Model It is shared by all models and is used to extract the underlying common features and local models. It is private to each model and maps the extracted common features to the required PV power generation; N models correspond to N photovoltaic sites. During the local training process, the site uses the local data set to train the corresponding model for E rounds, and then sends the global model to the server for aggregation. The server sends the aggregated global model to the corresponding client and enters the next round of iteration. During the aggregation process, the change values ​​of the parameters of the global model are averaged, which is expressed as: Where k represents the k-th round of aggregation, and the aggregated value is the change value of the global model weight in the k+1th round that needs to be issued; S3: Personalized federated aggregation. The client uploads the multi-layer time series decomposition module to the server. The server performs average aggregation and sends the aggregated global model to the server. S4: proceed to the next round of iteration until the final round of iteration; The multi-layer time series decomposition module uses the multi-layer perceptron as the network structure and combines it with the Gaussian filter to perform multi-layer time series decomposition. The multi-layer time series decomposition module is a global model. The global model is a shared model used to extract the underlying common features. The encoder and decoder in the multi-layer time series decomposition module are both composed of multi-layer perceptrons, which are superimposed in series through n multi-layer time series decomposition modules.

2. According to claim 1, a personalized federated learning prediction method for distributed photovoltaic power generation sites is characterized by: The multi-layer perceptron module is a local model, which is used to extract the characteristics of the photovoltaic site itself and process the differences between photovoltaic sites. The local model is a private model, which maps the extracted underlying common features to photovoltaic power generation.

3. According to claim 1, a personalized federated learning prediction method for distributed photovoltaic power generation sites is characterized by: In step S2, the following steps are specifically included: S21: Input the data into the Gaussian filter to decompose the data to obtain the trend component and the volatility component, and then use the multi-layer perceptron to predict the trend component and the volatility component respectively; S22: Calculate the residual between the prediction result and the input as the input of the next multi-layer time series decomposition module; S23: Multiple multi-layer time series decomposition modules are superimposed to decompose complex photovoltaic power generation data; S24: The prediction result of the multi-layer time series decomposition module is input into the final multi-layer perceptron module to obtain the final prediction result.

Images