Wind power prediction method based on cloud computing and data privacy protection

Abstract

With the support of our government’s commitment to the energy sector, the installed capacity of wind power will continue to grow. However, due to the instability of wind power, accurate prediction of wind power output is essential for effective grid dispatch. In addition, data privacy and protection have become paramount in today’s society. Traditional wind forecasting methods rely on centralized data, which raises concerns about data privacy and data silos. To address these challenges, we propose a hybrid approach that combines federated learning and deep learning for wind power forecasting. In our proposed method, we use a bidirectional long short-term memory (BILSTM) neural network as the basic prediction model to improve the prediction accuracy. Then, the model is integrated into the federated learning framework to form the Fed-BILSTM prediction method. In addition, we have introduced cloud computing technology into the Fed-BILSTM method, using cloud resources for model training and parameter update. Participants share model parameters instead of sharing raw data, which solves data privacy concerns. We compared Fed-BILSTM with traditional forecasting methods. Experimental results show that the proposed Fed-BILSTM is better than the traditional prediction method in terms of prediction accuracy. What’s more, Fed-BILSTM can effectively protect data privacy compared to traditional centralized forecasting methods while ensuring prediction performance.

Introduction

Literature review

Wind power forecasting mainly predicts the amount of electricity generated by wind power plants at different time scales. Compared to traditional fossil fuels, wind power is a green, safe, and renewable energy source. However, wind energy is a resource with high variability and poor persistence, which leads to the instability and high frequency fluctuations of wind power generation. To mitigate this instability, wind power plants need to be supplemented by other forms of energy generation. However, other energy sources are not easily deployable in the short term and require sufficient time to start up. Therefore, in this context, wind power forecasting becomes a crucial issue for the integration and operation of wind power, aiming to achieve grid stability and power supply security by improving the accuracy of wind power predictions [1]. Data security and privacy protection are among the most prominent issues in today’s society. With the increasing awareness of copyright and data privacy protection, it is an important challenge in the academic community to balance the privacy of various parties’ data while ensuring the accuracy of prediction models and the security of user data [2]. According to different modeling approaches, wind power forecasting methods can be summarized as: physical methods, statistical methods, artificial intelligence methods, and hybrid methods [3]. Among them, physical methods construct a physical model corresponding to power based on the information of wind farms and their surrounding environments, but they are greatly influenced by environmental factors. Statistical methods predict based on the mapping relationship between historical data of wind farms and wind power, enabling analysis of the temporal characteristics of data but not capturing nonlinear relationships well. Common methods include Kalman filtering, autoregressive moving average (ARMA), and grey prediction [4]. For wind speed prediction, autoregressive and moving average (ARIMA) models have been proposed to enhance prediction accuracy [5]. A novel probabilistic wind power forecasting model called Neural GOA DeepAR (NGOA- DeepAR) has been developed based on autoregressive recursive neural networks [6]. The Kalman filtering (KF) technique has been utilized to predict wind speed and power [7]. Grey prediction has also been employed for wind power forecasting [8]. Artificial intelligence methods mainly involve machine learning approaches. These methods establish the relationship between historical wind power data and wind power through autonomous learning by computers. Neural network models are particularly effective in capturing nonlinear relationships and deep features in historical data, making them widely used in wind power forecasting. Beta-LSTM, a neural network model based on the Beta distribution, has been proposed, demonstrating that deep neural networks achieve higher prediction accuracy than shallow machine learning models. However, it also highlights the longer training time and increased complexity of deep neural networks [9]. A new framework called heteroscedastic support vector regression (SVR) has been designed for wind power prediction [10]. An improved empirical mode decomposition (EMD) combined with a generalized algorithm backpropagation (GA-BP) neural network has been introduced for wind speed prediction [11]. With the rise and wider application of deep learning, more complex neural networks have been applied to forecast wind speed and power, and hybrid models have emerged as well. Subseries prediction models combining long short-term memory (LSTM) and deep belief networks based on particle swarm optimization (PSO-DBN) have been constructed. These subseries models are then integrated using a nonlinear weighted combination method based on PSO-DBN, forming a hybrid model for short-term wind power forecasting [12]. A high-accuracy hybrid method has been proposed, utilizing historical wind farm data and numerical weather prediction (NWP) data. The method includes steps such as anomaly detection, wavelet transform-based time series decomposition, effective feature selection, and prediction using multilayer perceptron (MLP) neural networks [13].

Massive data is the core asset of enterprises. Different stakeholders of wind farms are unwilling to share their wind power data with each other [14]. Therefore, when training wind power forecasting models using multi-party generation power data, the central node database must consider data security and privacy protection [15]. In 2006, Dwork from Microsoft Research proposed a new privacy protection model that addresses two major drawbacks of traditional privacy protection models: (1) it defines a rigorous attack model that does not care about how much background knowledge attackers possess. Even if an attacker has access to all records except one (i.e., the maximum background knowledge assumption), the privacy of that particular record cannot be disclosed; (2) it provides a rigorous definition and quantitative evaluation method for privacy protection levels. In [16], a privacy protection method for wind power probability prediction was proposed, which employs the alternating direction method of multipliers (ADMM) structure. However, in this method, the central node can recover private data, which may lead to confidentiality breaches and data privacy leakage. With the widespread use of green clean energy, the number of wind farms worldwide will gradually increase, and high-accuracy wind power forecasting is crucial for the secure dispatch of wind farms [17]. High-precision wind power forecasting models require massive wind power data support. However, massive data is a critical asset for enterprises. Therefore, in this context, it is essential to address the issue of improving the accuracy of wind power forecasting models while protecting enterprise data privacy [18]. Existing research has made significant contributions to wind power forecasting, but there are still several issues that need to be addressed.

• 1. Enhancing the accuracy of wind power forecasting models is crucial. Existing research mostly adopts a centralized prediction approach, where a unified dataset is used for model training.

• 2. Each wind farm has its own wind power dataset, which includes private data of each wind farm. They are unwilling to share their data, resulting in data silos.

• 3. How to train models and improve prediction accuracy without sharing data and protecting the privacy of each wind farm’s data.

Contribution of this paper

To address the aforementioned issues, this paper proposes a Fed-BILSTM model that combines bidirectional long short-term memory (BILSTM) neural networks and federated learning. Federated learning is a distributed machine learning technique that allows multiple parties to train models collaboratively without sharing raw data. This method not only effectively protects data privacy, but also improves the generalization ability of the model by using multi-source data. However, federated learning still faces limitations in computing resources and communication efficiency in practical applications. As a powerful computing resource management and distributed processing platform, cloud computing can provide necessary infrastructure support for federated learning. Cloud computing can significantly improve the performance of federated learning through its elastic computing resources and efficient data management capabilities. First, cloud computing platforms provide powerful storage and computing resources to support large-scale data processing and complex model training. Secondly, the distributed architecture and dynamic resource scheduling function of cloud computing can efficiently manage the collaborative computing tasks of multiple participants and ensure the smooth progress of the federated learning process. In addition, cloud computing’s secure communication and privacy-preserving technologies, such as encrypted communication, homomorphic encryption, and differential privacy, provide a solid security guarantee for federated learning. The main contributions of this paper are as follows:

• 1. In order to solve the challenge of low accuracy in wind power prediction, we propose a novel method that combines bidirectional long short-term memory (BILSTM) neural network with federated learning to form a Fed-BILSTM prediction method. As a widely used model in time series forecasting, BILSTM has the ability to effectively capture the forward and backward dependencies in series data, thereby improving the accuracy of prediction.

• 2. In order to solve the problem of data privacy protection, we propose the Fed-BILSTM prediction method, in which cloud computing technology is introduced into the Fed-BILSTM method, we combine federated learning with deep learning model BILSTM to improve the prediction accuracy of the model, and use cloud computing technology to train the model and update parameters to improve the efficiency of federated training. Experimental results show that compared with the traditional prediction methods, the proposed method not only protects user privacy, solves the problem of data islanding, but also achieves high prediction accuracy.

• 3. To validate the effectiveness of the proposed Fed-BILSTM prediction approach, we conduct simulation experiments using real historical wind power data from wind farms. We design three validation cases that demonstrate the higher prediction performance of our approach compared to traditional prediction approaches. To demonstrate the wide applicability of the proposed approach, we design a case study to compare it with traditional approaches, showing its excellent generalizability.

Organization of this paper

The rest of this paper is summarized as following: “Introduction” section introduction the literature review. “Related works” section describes the related work. “Experiment and analysis” section lists the evaluation metrics and the results of model experiments based on real-world data are discussed and analyzed. Finally, “Conclusion” section summarizes the entire work presented in this paper.

Related works

Federated learning

As the public and policymakers become increasingly aware of the importance of privacy, there is a growing demand for privacy-preserving machine learning in data practices. Access to data is subject to increasing scrutiny, and striking the right balance between maintaining model performance and efficiency while achieving the desired level of privacy and security poses significant technical challenges [19]. Federated learning, also known as collaborative learning, allows for large-scale training on devices where the data is generated, with sensitive data remaining with the data owners for local collection and training [20]. After local training, a central training coordinator obtains the updates of the distributed models to aggregate the training contributions from each node but does not access the actual sensitive data. Government regulations such as the European Union’s General Data Protection Regulation (GDPR) or the California Consumer Privacy Act (CCPA) have made privacy protection strategies like federated learning valuable tools for enterprises wishing to maintain lawful operations [21]. Ideally, federated learning enables collaboration among data stakeholders while preserving the privacy of individuals and organizations, as previously trade secrets, private health information, or risks of data breaches often made such collaboration difficult or even impossible.

We refer to each participating enterprise in collaborative modeling as a party, and based on the different data distributions among multiple parties, federated learning can be categorized into three types: horizontal federated learning, vertical federated learning, and federated transfer learning [22]. the essence of horizontal federated learning is the joint learning of samples, suitable for scenarios where participants have similar business domains but different customer reach, i.e., high feature overlap and low user overlap, such as different banks in different regions. They have similar business operations (similar features) but different users (different samples). the essence of vertical federated learning is the joint learning of features, suitable for scenarios with high user overlap and low feature overlap, such as supermarkets and banks in the same region. They target users who are residents of the same region (same samples) but have different business operations (different features). Federated transfer learning refers to the process of applying models learned from the source domain to the target domain based on data, tasks, or model similarities.

The goal of federated learning is to achieve common modeling and improve the effect of machine learning models on the basis of ensuring data privacy security and legal compliance. In the process of using the federated learning framework, ensuring the performance of the model is a basic requirement. Compared with traditional centralized model training methods, federated learning methods need to meet the following requirements.

$$\begin{aligned} \mid A_F-A_C\mid <\delta \end{aligned}$$

(1)

Where in $A_F$ is the evaluation standard of the federated learning model, $A_C$ is the evaluation standard of the model built by the data aggregation method, and $\delta$ is a bounded positive number.

Federated averaging algorithm

In federated learning, the communication rate is unstable, the capacity of the central server is limited, and the number of communications with the server and the number of terminals is limited. The federated averaging algorithm uses increased terminal computation, limits communication frequency, and performs multiple local gradient descent iterations before uploading updated gradients. The federated averaging algorithm randomly selects m terminals for sampling, averages the gradient updates of these m terminals to form a global update, and replaces unsampled terminals with the current global model, which greatly reduces the communication cost. Federated average algorithm terminal sampling, as shown in the following formula:

$$\begin{aligned} \theta ^{t+1}=\sum \limits _{i\in S_t}\frac{n_i}{M}\theta ^{t+1}_i+\sum \limits _{i\in S_t}\frac{n_i}{M}\theta ^t \end{aligned}$$

(2)

• 1. In each iteration, $S_t$ uniform sampling is performed on a subset of clients randomly selected to participate, and the current global model parameters are distributed, and the gradient is trained $\theta ^{t}$ locally on the terminal, and uploaded to the server for average formation of updated parameters $\theta ^{t+1}$.

• 2. The update of terminals that do not belong to the sampling subset is replaced by the current global model parameters $\theta ^{t}$.

• 3. Then the server averages to generate new global parameters $\theta ^{t+1}$.

Bidirectional long short-term memory network (BILSTM)

Long short-term memory network (LSTM) is a memory network based on deep recurrent neural network (RNN). It is an advanced version of traditional RNN and is designed to deal with the problem that traditional CNN cannot handle long-term memory.As shown in Fig. 1, LSTM controls information flow through gate mechanism, including forget gate, input gate and output gate. The forget gate decides whether the information from the previous timestamp is to be remembered or irrelevant and can be discarded from the cell state. The input gate decides what new information should be added to the cell state. The output gates decide which parts of the updated cell state should be passed from the current timestamp to the next. The calculation formulas and functions involved in these gate mechanisms under the timestamp t are as follows:

$$\begin{aligned} f_t=\sigma (W_{lf}l_t+W_{m_f}h_{t-1}+b_f) \end{aligned}$$

(3)

$$\begin{aligned} i_t=\sigma (W_{li}l_t+W_{mi}h_{t-1}+b_i) \end{aligned}$$

(4)

$$\begin{aligned} o_t=\sigma (W_{lo}l_t+W_{mo}h_{t-1}+b_o) \end{aligned}$$

(5)

$$\begin{aligned} a_t=\tanh {(W_{la}l_t+W_{ma}h_{t-1}+b_a)} \end{aligned}$$

(6)

$$\begin{aligned} c_t=c_{t-1}*f_t+i_t*a_t \end{aligned}$$

(7)

$$\begin{aligned} m_t=o_t*\tanh c_t \end{aligned}$$

(8)

where $\sigma$ is the logistic sigmoid function, $i_t$, $o_t$, $c_t$, and $a_t$ denote the forget gate, input gate, output gate, memory cell, and hidden vector, respectively. $W_{l*} = (W_{lf}, W_{li}, W_{la}, W_{lo})$ and $W_{m*} = (W_{mf}, W_{mi}, W_{ma}, W_{mo})$ represent the trainable weights of the respective gates, while $b_f$, $b_i$, $b_o$, and $b_a$ are the output biases. Lastly, the operator $*$ defines the Hadamard product.

Fig. 1

The LSTM model structure diagram

BILSTM is a variant of LSTM, As shown in Fig. 2, which is characterized by LSTM in two directions. One processes the input sequence from front to back, and the other processes the input sequence from back to front. This structure enables BLSTM to capture the contextual information around the current moment to better predict the next output. At each time step, BLSTM takes the input vector and the output vector of the previous time step as the input of the forward and backward LSTM, and stitches their outputs together to preserve the past and future contextual information, thereby improving the model performance. accuracy. The formulas and calculation functions involved are as follows:

$$\begin{aligned} y_i=\sigma (LF_i(x_i),LB_i(x_i))\quad \forall i\in [0,1,2,\ldots ,n] \end{aligned}$$

(9)

where $LF_i$ and $LB_i$ denotes the outputs of the forward and backward hidden layers, $x_i$ and $y_i$ denotes the input and output put of the BLSTM model, respectively.

Fig. 2

The BILSTM model structure diagram

Fed-BILSTM

In this model structure, we introduce cloud computing and the PySyft federated learning training framework to further enhance data privacy and security. Here’s the updated workflow:

First, the central server assigns the initial global model to each wind farm terminal and creates a PySyft virtual worker to represent each terminal. These virtual workers are hosted on cloud servers, ensuring the security and privacy of data during model training.

Second, Each endpoint then uses local data to train the initial model, BILSTM, but the training process at this point is wrapped in the PySyft framework to ensure that the data is safely processed and trained locally. This means that the raw data does not leave the local environment where the endpoint is located, thus protecting user privacy and data security. After the local training is completed, the endpoint sends the new model parameter weights back to the central server through the PySyft secure communication mechanism. The central server averages the parameters of the new model using the secure aggregation protocol provided by PySyft, and updates the parameters of the global model with the average. This ensures the security and privacy of the data during parameter updates.

Finally, the central server transfers the parameters of the updated global model to the PySyft virtual worker for each endpoint so that the endpoint can continue to train and update the model locally. This process is iterative until the global model meets the convergence conditions or reaches the maximum number of iterations.

The design of this model structure not only protects user privacy and data security, but also enables the data of each terminal to participate in the training of the global model. By combining cloud computing and the PySyft federated learning training framework, we have implemented an efficient, secure, and privacy-preserving horizontal federated learning system, providing a reliable solution for wind power prediction. As shown in Fig. 3.

Fig. 3

The Fed-BILSTM model structure diagram

Cloud computing

Cloud computing, also translated as network computing, is an Internet-based computing method, in this way, the shared software and hardware resources and information can be provided to various computer terminals and other equipment on demand, and the computer infrastructure provided by the service provider is used for computing and resources. The characteristics of cloud computing services on the Internet are similar to the cloud and water cycles in nature, so the cloud is a fairly appropriate metaphor. According to the National Institute of Standards and Technology, cloud computing services should have the following characteristics:

• On-demand self-service.

• Access anytime, anywhere, on any network device.

• Multiple people share a pool of resources.

• Rapid redeployment flexibility.

Evaluation criteria

Currently, there is no unified international standard for comparing prediction models. Common evaluation metrics include Mean Error (ME), Mean Absolute Error (MAE), and Root Mean Square Error (MRSE). In order to better and more comprehensively evaluate the accuracy of the proposed models and reduce the errors of wind power prediction models, this paper adopts MAE and RMSE as the evaluation metrics for prediction accuracy. The formula for calculating MRSE is as follows:

$$\begin{aligned} RMSE=\sqrt{\frac{1}{N}\sum \limits _{i=0}^N\mid y_i-\hat{y}_i\mid ^2} \end{aligned}$$

(10)

where N represents the number of data series, $y_i$ and $\hat{y}$ denote the prediction and real values, The range of MRSE values is $[0,+\infty ]$. When the MRSE value is 0, it indicates that the predicted values are identical to the true values. However, in practical wind power prediction, due to the high instability of wind power, achieving this ideal scenario is often not possible. In this paper, the accuracy of the predictions is judged by comparing the magnitude of the MRSE values. MRSE is negatively correlated with prediction accuracy, with smaller MRSE values indicating smaller prediction errors.

The formula for calculating MAE is as follows:

$$\begin{aligned} MAE=\frac{1}{N}\sum \limits _{i=1}^N\mid y_i-\hat{y}_i\mid \end{aligned}$$

(11)

where N represents the number of data series, $y_i$ and $\hat{y}$ denote the prediction and real values, its value directly represents the average difference between the predicted values and the true values. MAE is more interpretable for model comparison and more suitable for statistical application scenarios in practical problems.

Experiment and analysis

Data preprocessing

In this experiment, we used a real-world dataset consisting of wind farm data from a specific region in the United States (with a sampling interval of 1 hour) as the local datasets for six clients. The dataset was divided into six different client datasets. When each wind farm used its own private dataset for local training, the ratio of training set to test set was 7:3, with each client having 973 data points in the training set and 417 data points in the test set. In the data preprocessing stage, we used mean interpolation to replace missing values in the dataset. We also removed outliers and duplicate values from the dataset. Finally, we normalized the data using the MinMaxScaler() function. Figure 4 illustrates the characteristics of each dataset, and the statistical information of the datasetssix clients is presented in the Table 1 below.

Fig. 4

Local datasets for six clients

Table 1 The statistical information of the six clients

Table 2 The statistical information of the six clients

Prediction results of different wind clients

The Figs. 5 and 6 presented below depicts the prediction results of the Fed-BILSTM prediction scheme for the datasets of six clients. It can be observed from the figure that the prediction results for each wind farm dataset exhibit a high level of accuracy compared to the actual wind power dataset. It is evident that by employing the Fed-BILSTM prediction scheme without sharing private datasets among the wind farm clients, it is possible to obtain highly accurate prediction models. This approach not only resolves the challenges of data isolation and data privacy preservation but also produces a reliable wind power prediction model.

Fig. 5

The prediction results of client1, client2 and client3

Fig. 6

The prediction results of client4, client5 and client6

We calculated the relative error between the predicted values and the true values. Due to the large amount of data, we averaged the relative errors for each client over every 40 hours, resulting in the average relative error values. The graph shows the average relative errors for the six clients. Overall, as is shown in the Fig. 7, the predicted relative errors for all clients fall within the range of -10% to 10%, highlighting the high accuracy of the Fed-BILSTM prediction approach.

Fig. 7

The predicted relative errors for all clients

Comparison of different cases

In this study, three scenarios were designed for conducting simulation experiments using real historical datasets. The prediction results were compared with the actual data to validate the proposed model’s data privacy and generalizability, as shown in the Table 2 below. Additionally, different parameters were designed to test the robustness of the proposed model scheme. The experiments were conducted on a computer in a laboratory environment, and all code implementations were based on Python 3.6, PyTorch 2.0, and Pandas 1.5.3.

Firstly, traditional deep learning model training is mostly centralized and non-distributed. To compare with the proposed model, a centralized and non-distributed model training scheme was proposed, where all datasets were consolidated and trained on a single client. However, this approach compromises the privacy of the datasets due to the need for data consolidation. Thus, this scheme was used for comparing the prediction accuracy with the proposed model scheme.

Secondly, a fully privatized training scheme was designed, where each client performed model training locally in isolation from other clients. In this scheme, the models learned the features specific to each client’s local dataset, preventing cross-learning of features among models. This ensured that the models were unique to each client.

Finally, the initial global model was distributed to each wind farm client by the central server. Each client then trained the initial BILSTM model using its local data. After local training was completed, the client sent the new model parameter weights back to the central server. The central server employed the federated averaging algorithm to calculate the average of the new model parameters and updated the global model’s parameters accordingly. Lastly, the central server distributed the updated parameters of the global model to each client.

The centralized non-distributed model training approach involves consolidating the data from various clients into a centralized location for model training, allowing the model to learn more data features. This results in higher accuracy and better precision compared to localized and federated learning methods, as shown in Tables 3, 4 and 5. The MAE value of the centralized non-distributed model is 25.4% higher than the localized model and 7.52% higher than the federated model. The RMSE value is 18.70% higher than the localized model and 4.40% higher than the federated model. It is evident that the centralized non-distributed approach achieves better prediction accuracy. In Figs. 8 and 9, the data can be compared more clearly. However, the centralized non-distributed model training approach does not protect user data privacy and does not solve the data silo problem.

The localized model training method, on the other hand, involves each client training their model based on their own private dataset. The resulting models are more suitable for the local dataset but lack generalizability. When these models are applied to other clients’ private datasets, their accuracy decreases.

Both the proposed Fed-BILSTM and the centralized non-distributed models demonstrate excellent prediction performance and accurately describe the wind power output of wind farms. The accuracy of wind power prediction using the proposed Fed-BILSTM is comparable to that of the centralized non-distributed model. It is worth noting that using the centralized non-distributed approach requires collecting data from each wind farm, which may raise concerns about sensitive user privacy. If privacy data is leaked, it could cause severe losses for each wind farm. The method proposed in this paper allows each wind farm’s data to be kept locally for training and achieves similar prediction accuracy without transmitting the raw data. This indicates that the proposed Fed-BILSTM can achieve prediction accuracy standards without sacrificing user privacy.

Table 3 Forecasting performance evaluations of the centralized non-distributed model

Table 4 Forecasting performance evaluations of the localized model training method

Table 5 Forecasting performance evaluations of the Fed-BILSTM

Fig. 8

The MAE of the cases

Fig. 9

The $RMSE$ of the cases

Model generalizability

We validated the generalizability of the three proposed scenarios on two additional data sets that were not used during model training. The table below shows the performance metrics of the models in different scenarios.

As shown in Tables 6, 7 and 8. when the models were applied to previously unseen data, their overall performance decreased. In this case, the centralized model training method still had a significant advantage over the local model method. However, it should be noted that the centralized model training method compromises data privacy and exposes user-sensitive information. In this scenario, the federated model training method outperformed both the centralized and local training methods. Compared to the centralized model training method, the federated method achieved a 25.90% reduction in MAE and a 24.86% reduction in $RMSE$. Compared to the local model training method, the federated method achieved a 56.31% reduction in MAE and a 50.37% reduction in RMSE. In Figs. 10 and 11, the data can be compared more clearly. These results indicate that the federated model training method proposed in this paper exhibits better generalizability compared to the centralized and local training methods.

Table 6 Forecasting performance evaluations of the centralized non-distributed model

Table 7 Forecasting performance evaluations of the localized model training method

Table 8 Forecasting performance evaluations of the Fed-BILSTM

Fig. 10

The MAE of the cases

Fig. 11

The RMSE of the cases

Comparison between Fed-BILSTM and centralized forecasting methods

In this paper, the proposed Fed-BILSTM prediction method is compared with four other models: Mixed Logistic Regression (MLR), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM). All models are used to predict wind power in the same wind farm dataset in order to evaluate the prediction accuracy of the proposed model.

As shown in the Table 9, different algorithms exhibit different characteristics. From the perspective of MAE and RMSE values, MLR and LSTM models perform well and demonstrate good prediction accuracy for wind power. However, the table highlights the superior performance of BILSTM in terms of average results. Compared to MLR and LSTM models, BILSTM achieved a reduction in MAE and RMSE values by 13.88%, 15.61%, and 0.04%, 12.68%, respectively. These results indicate that BILSTM exhibits higher prediction accuracy compared to MLR and LSTM models.

Table 9 Forecasting performance evaluations of different models for the testing set in Midatl

Conclusion

With the increasing focus on renewable energy, such as hydropower, wind power, and solar energy, it has become a hot research topic for many scholars. Supported by relevant energy policies in our country, wind power development has grown rapidly, and the installed wind power capacity in China is expected to further increase. To protect the stability of the power grid, improve the prediction accuracy of wind power models, and address data privacy and data silo issues, this paper combines federated learning with deep learning to propose a data privacy-preserving federated deep learning method called Fed-BILSTM for wind power prediction. BILSTM is used as the basic prediction model, which participates in training and parameter updates to enhance prediction accuracy. Unlike traditional wind power prediction methods, in the Fed-BILSTM approach, wind farm clients only need to train the model on their local private datasets, and then upload the model parameters to a central server. The central server performs parameter updates using federated averaging algorithms and redistributes the updated parameters to each client. This approach ensures the accuracy of wind power prediction, resolves data silo issues, and effectively protects user privacy. The results consistently demonstrate that Fed-BILSTM outperforms MLR and LSTM models, achieving reductions in MAE and RMSE of 13.88%, 15.61%, 0.04%, and 12.68%, respectively. Therefore, Fed-BILSTM exhibits superior prediction accuracy. Most significantly, our proposed model effectively addresses the challenges associated with data privacy protection and data silos that are prevalent in traditional prediction models. By employing a federated learning approach and incorporating privacy-preserving mechanisms, we ensure that sensitive data remains secure while achieving accurate wind power predictions. In future work, we will focus on improving the training efficiency of the model. In this study, we did not consider communication delays and network packet loss issues, which will be the focus of our future work.