A wind power joint prediction method
By constructing a wind power prediction model based on a long short-term memory network and using a federated learning algorithm for collaborative training, the problems of spatial correlation and competition among wind farms were solved, achieving high-precision wind power prediction and data security protection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID SHANGHAI MUNICIPAL ELECTRIC POWER CO
- Filing Date
- 2023-02-17
- Publication Date
- 2026-05-01
AI Technical Summary
Existing wind power forecasting methods fail to effectively utilize the spatial correlation between wind farms and ignore the competitive relationship between geographically adjacent wind farms, resulting in insufficient forecasting accuracy.
A wind power prediction model based on long short-term memory networks is adopted, and the models of each wind farm are collaboratively trained through federated learning algorithms. Distributed multi-party inner product encryption technology is used to aggregate data samples without moving local data, taking into account the spatial correlation and competition between wind farms.
It improves the accuracy of wind power forecasting, protects the privacy and security of wind farm data, makes full use of the spatial correlation of data, and enhances the collaborative forecasting effect among multiple wind farms.
Smart Images
Figure SMS_1 
Figure SMS_4 
Figure SMS_6
Abstract
Description
Technical Field
[0001] This invention relates to the field of virtual power plant operation technology, and in particular to a wind power prediction model based on long short-term memory networks and a joint training algorithm based on horizontal federation. Background Technology
[0002] Wind power forecasting is a crucial prerequisite for the planning and operation of new power systems. Wind power forecasting primarily employs physical, statistical, probabilistic, and artificial intelligence methods. Due to the strong fitting capabilities of neural networks, this artificial intelligence approach is used to improve the accuracy of wind power forecasting. However, current methods based on neural networks for wind power forecasting often only consider the temporal correlation of historical data from individual wind farms, neglecting the spatial correlation of wind power output from geographically adjacent wind farms. Furthermore, methods that consider spatial correlation assume that training data can be directly aggregated, failing to account for potential competition between wind farms. Summary of the Invention
[0003] The purpose of this invention is to provide a joint wind power forecasting method that can not only improve the wind power forecasting accuracy of each wind farm, but also ensure the security of sensitive data of each wind farm. Under the premise of data immobilization, it can make full use of the spatial correlation between data to explore the value of data, thereby achieving the effect of directly aggregating data samples from multiple wind farms in the result.
[0004] To achieve the above objectives, the present invention provides a wind power joint forecasting method, comprising the following steps:
[0005] Step S1: Construct a wind power prediction model for each wind farm based on a long short-term memory network;
[0006] Step S2: Use the federated learning algorithm to collaboratively train the wind power prediction model of each wind farm. Through continuous iteration of local training updates, parameter encryption and secure aggregation, a global federated prediction model of the wind farm is obtained.
[0007] Step S1 includes:
[0008] Step S1.1: Construct a Long Short-Term Memory (LSTM) network layer, consisting of T memory units, each containing:
[0009] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )
[0010] f t =σ(W if x t +bif +W hf h t-1 +b hf )
[0011] g t =tanh(W ig x t +b ig +W hg h t-1 +b hg )
[0012] o t =σ(W io x t +b io +W ho h t-1 +b ho )
[0013] c t =f t ⊙c t-1 +i t ⊙g t
[0014] h t =o t ⊙tanh(c t )
[0015] In the formula, x t It is the input vector at time t, including wind power output and meteorological data, h t It is the hidden state vector at time t-1, i t and W ii b ii W hi b hi These are the information gating signals and their corresponding network parameters, f t and W if b if W hf b hf These are the forgetting gating signal and its corresponding network parameters, g t and W ig b ig W hg b hg For the candidate state vectors and their corresponding network parameters, o t and W io b io W ho b ho To output the gating signal and its corresponding network parameters, c t This is the state vector of the memory cell;
[0016] Step S1.2: Add a fully connected layer on top of the Long Short-Term Memory (LSTM) network layer to map the hidden layer output to the wind power output value.
[0017] y t =σ(W hy h t +b hy )
[0018] In the formula, y t W is the predicted wind power output at time t+τ, where τ is the prediction time interval. hy b hy These are the network parameters for the fully connected layer.
[0019] Step S2 includes:
[0020] Step S2.1: Each wind farm updates its local wind power processing prediction model using local data samples;
[0021] Step S2.2: Each wind farm uploads the calculated gradient values to the cloud.
[0022] Step S2.3: Aggregate the local gradient values uploaded by all wind farms in the cloud to obtain the global gradient:
[0023]
[0024] In the formula, It is the global gradient obtained by aggregation, a n,k This represents the weight of each wind farm, where N is the total number of wind farms.
[0025] Step S2.4: The cloud sends the aggregated global gradient back to each wind farm, and each wind farm updates the prediction model parameters as follows:
[0026]
[0027] In the formula, ω k+1 λ represents the global prediction model parameters at the (k+1)th iteration, where λ is the step size.
[0028] Repeat steps S2.1 to S2.4 iteratively until the global model converges and the inter-wind farm collaborative prediction is completed.
[0029] Step S2.1 includes:
[0030] Step S2.1.1: Calculate the loss function value of the current wind power prediction model:
[0031]
[0032] In the formula, y iIt is the wind power prediction value of the Long Short-Term Memory network. N represents the true value given by the data sample, i.e., the labeled value. b It refers to batch size, L. k (ω n，k ) is the local prediction network parameter ω for the nth wind farm in the kth iteration. n,k The loss function value calculated at that time;
[0033] Step S2.1.2: Calculate the gradient with respect to the network parameters using the loss function value:
[0034]
[0035] In the formula, The nth wind farm has a local prediction network parameter of ω in the kth iteration. n,k The local gradient is calculated in real time.
[0036] The beneficial effects of this invention include:
[0037] 1. This invention establishes a wind power prediction model based on neural networks. The original input time series is preprocessed, and then the predicted wind power output value is obtained through a neural network consisting of an input layer, a hidden layer, an output layer, and an activation function. It considers the correlation of various input features and their time series, and can fit the high-dimensional nonlinear relationship between meteorological data, historical power output data, and predicted power output data.
[0038] 2. This invention employs a horizontal federated learning algorithm based on distributed multi-party inner product encryption. Without moving the data samples stored locally at the wind farms, it effectively aggregates the data sample set, thereby improving the accuracy of wind power output prediction through collaboration. Specifically, each wind farm first updates its local model based on its local data samples. Then, it encrypts the updated parameters using a pre-generated key and uploads them to the cloud. Finally, the cloud aggregates the decryption keys from each wind farm, securely aggregates the updated local parameters, and decrypts them to obtain a new federated model. This process iterates continuously until convergence.
[0039] 3. This invention can achieve equivalent aggregation of training samples while ensuring the protection of privacy data of each wind farm, and fully considers the spatial correlation between wind farms, which can significantly improve the effect of collaborative prediction between wind farms. Detailed Implementation
[0040] This invention provides a wind power joint forecasting method, comprising the following steps:
[0041] Step S1: Construct a wind power prediction model for each wind farm based on a long short-term memory network. The wind power prediction model consists of a long short-term memory network layer and a fully connected layer.
[0042] Step S1.1: Construct a Long Short-Term Memory (LSTM) network layer, consisting of T memory units. Each memory unit consists of the following components:
[0043] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )
[0044] f t =σ(W if x t +b if +W hf h t-1 +b hf )
[0045] g t =tanh(W ig x t +b ig +W hg h t-1 +b hg )
[0046] o t =σ(W io x t +b io +W ho h t-1 +b ho )
[0047] c t =f t ⊙c t-1 +i t ⊙g t
[0048] h t =o t ⊙tanh(c t )
[0049] In the formula, x t It is the input vector at time t, including wind power output and meteorological data, h t It is the hidden state vector at time t-1, i t and W ii b ii W hi b hi These are the information gating signals and their corresponding network parameters, f t and W if b if W hfb hf These are the forgetting gating signal and its corresponding network parameters, g t and W ig b ig W hg b hg For the candidate state vectors and their corresponding network parameters, o t and W io b io W ho b ho To output the gating signal and its corresponding network parameters, c t This is the state vector of the memory cell;
[0050] Step S1.2: Add a fully connected layer on top of the Long Short-Term Memory (LSTM) network layer to map the hidden layer output to the wind power output value.
[0051] y t =σ(W hy h t +b hy )
[0052] In the formula, y t W is the predicted wind power output at time t+τ, where τ is the prediction time interval. hy b hy These are the network parameters for the fully connected layer;
[0053] Step S2: Collaboratively train the wind power prediction model for each wind farm using a federated learning algorithm. This federated learning algorithm consists of four parts: local training, parameter uploading, cloud aggregation, and parameter distribution.
[0054] Step S2.1: Each wind farm updates its local wind power processing prediction model using local data samples.
[0055] Step S2.1.1: Calculate the loss function value of the current wind power prediction model:
[0056]
[0057] In the formula, y i It is the wind power prediction value of the Long Short-Term Memory network. N represents the true value given by the data sample, i.e., the labeled value. b It refers to batch size, L. k (ω n，k ) is the local prediction network parameter ω for the nth wind farm in the kth iteration. n，k The loss function value calculated at that time;
[0058] Step S2.1.2: Calculate the gradient with respect to the network parameters using the loss function value:
[0059]
[0060] In the formula, The nth wind farm has a local prediction network parameter of ω in the kth iteration. n，k The local gradient is calculated in real time;
[0061] Step S2.2: Each wind farm uploads the calculated gradient values to the cloud.
[0062] Step S2.3: Aggregate the local gradient values uploaded by all wind farms in the cloud to obtain the global gradient:
[0063]
[0064] In the formula, It is the global gradient obtained by aggregation, a n,k This represents the weight of each wind farm, where N is the total number of wind farms.
[0065] Step S2.4: The cloud sends the aggregated global gradient back to each wind farm, and each wind farm updates the prediction model parameters as follows:
[0066]
[0067] In the formula, ω k+1 λ represents the global prediction model parameters at the (k+1)th iteration, where λ is the step size.
[0068] Repeat steps S2.1 to S2.4 iteratively until the global model converges and the inter-wind farm collaborative prediction is completed.
[0069] The features and beneficial effects of this invention are as follows:
[0070] 1. This invention establishes a wind power prediction model based on neural networks. The original input time series is preprocessed, and then the predicted wind power output value is obtained through a neural network consisting of an input layer, a hidden layer, an output layer, and an activation function. It considers the correlation of various input features and their time series, and can fit the high-dimensional nonlinear relationship between meteorological data, historical power output data, and predicted power output data.
[0071] 2. This invention employs a horizontal federated learning algorithm based on distributed multi-party inner product encryption. Without moving the data samples stored locally at the wind farms, it effectively aggregates the data sample set, thereby improving the accuracy of wind power output prediction through collaboration. Specifically, each wind farm first updates its local model based on its local data samples. Then, it encrypts the updated parameters using a pre-generated key and uploads them to the cloud. Finally, the cloud aggregates the decryption keys from each wind farm, securely aggregates the updated local parameters, and decrypts them to obtain a new federated model. This process iterates continuously until convergence.
[0072] 3. This invention can achieve equivalent aggregation of training samples while ensuring the protection of privacy data of each wind farm, and fully considers the spatial correlation between wind farms, which can significantly improve the effect of collaborative prediction between wind farms.
[0073] It should be noted that, in the embodiments of the present invention, the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientation or positional relationships, are used only for the convenience of describing the embodiments and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0074] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.
Claims
1. A joint wind power forecasting method, characterized in that, Includes the following steps: Step S1: Construct a wind power prediction model for each wind farm based on a long short-term memory network; Step S2: Use the federated learning algorithm to collaboratively train the wind power prediction model of each wind farm. Through continuous iteration of local training update, parameter encryption and secure aggregation, a global federated prediction model of the wind farm is obtained. Specifically, each wind farm first updates its local wind power prediction model based on local data samples, then encrypts the updated parameters using a pre-generated key and uploads them to the cloud. Finally, the cloud aggregates the decryption keys from each wind farm, securely aggregates the updated local parameters, and decrypts them to obtain a new global federated prediction model. This process iterates continuously until convergence. Step S2 includes: Step S2.1: Each wind farm updates its local wind power processing prediction model using local data samples; Step S2.2: Each wind farm uploads the calculated gradient values to the cloud. Step S2.3: Aggregate the local gradient values uploaded by all wind farms in the cloud to obtain the global gradient: In the formula, It is the global gradient obtained by aggregation. It is the weight of each wind farm. This represents the total number of wind farms. Step S2.4: The cloud sends the aggregated global gradient back to each wind farm, and each wind farm updates the prediction model parameters as follows: In the formula, It is the first Global prediction of model parameters during round iteration. It is the step size; Repeat steps S2.1 to S2.4 iteratively until the global model converges and the inter-wind farm collaborative prediction is completed; Step S2.1 includes: Step S2.1.1: Calculate the loss function value of the current wind power prediction model: In the formula, It is the wind power prediction value of the Long Short-Term Memory network. These are the true values given by the data samples, i.e., the labeled values. It's the batch size. It is the first The wind farm in the first During each iteration, the network parameters are predicted locally. The loss function value calculated at that time; Step S2.1.2: Calculate the gradient with respect to the network parameters using the loss function value: In the formula, It is the first The wind farm in the first During each iteration, the network parameters are predicted locally. The local gradient is calculated in real time.
2. The wind power joint forecasting method as described in claim 1, characterized in that, Step S1 includes: Step S1.1: Construct a Long Short-Term Memory (LSTM) network layer, by... It consists of several memory units, each containing: In the formula, It is the first The input vector at any given time includes wind power output and meteorological data. It is the first The hidden state vector at time step 1. and These are the information gating signals and their corresponding network parameters. and These are the forgetting gating signal and its corresponding network parameters. and For the candidate state vectors and their corresponding network parameters, and To output the gating signal and its corresponding network parameters, This is the state vector of the memory cell; Step S1.2: Add a fully connected layer on top of the Long Short-Term Memory (LSTM) network layer to map the hidden layer output to the wind power output value. In the formula, It is the first Forecast wind power output at any given time For predicting time intervals, These are the network parameters for the fully connected layer.