Comprehensive energy wind power prediction method based on modal decomposition denoising and GAT-BiLSTM

Abstract

This invention, based on mode decomposition denoising and GAT-BiLSTM, belongs to the field of renewable energy prediction technology. It aims to address the low accuracy issues in wind power prediction caused by high-frequency noise interference and insufficient mining of the spatiotemporal correlation of multidimensional meteorological factors. The method first collects wind power and meteorological data such as wind speed and direction. After preprocessing using the isolated forest algorithm, the wind power sequence is decomposed using ICEEMDAN. High-frequency noise modes are located by combining sample entropy, and the signal is reconstructed after filtering with a Savitzky-Golay filter. Then, a GAT-BiLSTM model is constructed. GAT is used to mine the feature space correlation, and BiLSTM is used to extract bidirectional time dependence, thus achieving wind power prediction. This invention achieves accurate noise reduction of wind power signals, deeply integrates the spatiotemporal correlation of wind power and meteorology, significantly improves prediction accuracy and robustness, and provides reliable decision support for the economic dispatch of integrated energy systems.

Description

Technical Field

[0001] This invention relates to the field of renewable energy prediction technology, and in particular to a comprehensive energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM. Background Technology

[0002] With the transformation of the global energy structure, renewable energy is playing an increasingly important role in integrated energy systems, with wind power, as one of the most representative renewable energy sources, experiencing continuous expansion in its development and utilization. As a key to achieving efficient utilization and complementary coordination of multiple energy forms, the economic efficiency and security of integrated energy systems are becoming increasingly important. However, the inherent volatility, intermittency, and strong randomness of wind power output pose significant challenges to the safe, stable operation and economic dispatch of integrated energy systems. Therefore, high-precision forecasting of wind power output has become a crucial prerequisite for ensuring system reliability, developing optimized dispatch strategies, and improving the renewable energy absorption rate.

[0003] Currently, wind power forecasting methods are mainly divided into physical model methods, traditional statistical methods, and artificial intelligence methods represented by deep learning. Physical model methods rely on complex numerical weather prediction and fluid dynamics models. Although these models can theoretically provide relatively accurate predictions, they are difficult to model, consume high computational resources, and are often affected by various uncertainties in practical applications, resulting in limited prediction accuracy. Traditional statistical methods, such as ARIMA (Autoregressive Integrated Moving Average) models, predict future wind power through statistical analysis of historical data. These methods are simple and easy to implement, but they struggle to capture the strong nonlinear and non-stationary characteristics of wind power sequences. Especially when facing complex and variable weather conditions and wind farm operating environments, their prediction accuracy is often unsatisfactory.

[0004] In recent years, artificial intelligence methods, represented by deep learning, have been widely applied in the field of wind power forecasting. Deep learning models such as Long Short-Term Memory (LSTM) networks and gated recurrent units (GRUs) have demonstrated significant advantages in handling time series forecasting problems by capturing long-term dependencies in time series data, effectively improving the accuracy of wind power forecasting. However, existing research still suffers from two key limitations, making it difficult to meet the practical needs of high-precision wind power forecasting: At the signal processing level, the raw wind power signal inevitably contains a large amount of high-frequency random noise, which severely interferes with the feature extraction capability of the prediction model. Although existing signal decomposition methods such as Empirical Mode Decomposition (EMD) and its improved methods can separate noise to some extent, they generally suffer from problems such as mode aliasing and baseline drift. Moreover, the processing of noise modes often adopts a direct rejection approach, which easily leads to the loss of useful information and seriously affects the accuracy of signal denoising.

[0005] At the feature mining level, wind power output is not only strongly correlated with its historical data, but also influenced by a variety of meteorological factors such as wind speed, wind direction, temperature, air pressure, and humidity. However, most existing prediction models treat multidimensional meteorological data as parallel input channels, neglecting the dynamic spatiotemporal correlations and coupling relationships between various input features. For example, the wind speed in one location may be lagging influencing the air pressure in another location. This spatial correlation is difficult for traditional time series models to capture, causing the models to fail to fully utilize the information implicit in the multidimensional data and limiting further improvements in prediction accuracy.

[0006] Existing related technical solutions have also failed to effectively solve the above problems. For example, CN116662790A discloses a wind turbine vibration feature prediction method that improves the NMS-RLM microbial community algorithm and optimizes CNN-BiLSTM. This method uses 1DCNN (1-Dimensional Convolutional Neural Network) to extract vibration feature parameters in the frequency domain, and then inputs them into BiLSTM (Bidirectional Long Short-Term Memory) for modeling. Although a deep learning model is used and has been applied in vibration feature prediction, this method still has the problem that the potential relationship of local features cannot be well represented when dealing with the wind power prediction problem. Although the traditional CNN (Convolutional Neural Network) model can extract local features, it has significant limitations in mining the cross-dimensional spatial correlation and dynamic potential relationship between wind power and multi-dimensional meteorological factors such as wind speed and air pressure.

[0007] Furthermore, CN111832825A discloses a wind power prediction method integrating Long Short-Term Memory (LSTM) networks and Extreme Learning Machine (ELM). This method recombines wind power sequences and meteorological feature data "according to frequency magnitude" into low-frequency and high-frequency combined input vectors, and uses LSTM and ELM for prediction respectively. Although this method considers the influence of different frequency components on wind power, its "recombination according to frequency magnitude" approach is a relatively coarse division, making it difficult to adaptively separate high-frequency noise from the true signal details. Moreover, it does not accurately identify and process high-frequency noise, resulting in limited noise reduction performance and an inability to effectively address the strong interference of high-frequency noise in the original signal, thus affecting the prediction accuracy of subsequent models.

[0008] In summary, existing technologies still have many shortcomings and challenges in areas such as accurate noise reduction of wind power signals and in-depth mining of the spatiotemporal correlation of multidimensional meteorological factors. Therefore, there is an urgent need to develop a more advanced, comprehensive, and refined wind power output forecasting method to overcome existing technological bottlenecks, significantly improve the accuracy and robustness of wind power forecasting, and provide scientific and reliable decision support for the efficient, low-carbon, and economical operation of integrated energy systems. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a comprehensive energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM. This method addresses the technical problems of low prediction accuracy and poor robustness in existing comprehensive energy systems, which are caused by high-frequency noise interference and difficulty in mining the spatiotemporal correlation of multidimensional meteorological factors. This invention achieves high-precision wind power prediction, providing reliable decision support for the economic dispatch and safe operation of comprehensive energy systems.

[0010] To achieve the above technical objectives, the present invention adopts the following technical solution: A wind power prediction method for integrated energy systems based on mode decomposition denoising and GAT-BiLSTM comprises six steps: data acquisition, data preprocessing, mode decomposition, noise mode identification, signal denoising reconstruction, and spatiotemporal fusion prediction, as detailed below: Step 1, Data Acquisition: Through a sensor network deployed on the wind turbine itself, inside and around the wind farm, the power data of the wind turbine and external meteorological data in the integrated energy system are collected simultaneously. The external meteorological data includes at least wind speed, wind direction, temperature, humidity and air pressure. All data are collected and stored at a uniform time interval to ensure the consistency of the data sequence.

[0011] Step 2, Data Preprocessing: The Isolation Forest algorithm is used to identify outliers in the collected raw data. By utilizing the sparsity of outliers in the data space, a random hyperplane is constructed to cut the data space to achieve accurate identification of outliers. Interpolation is used to fill in missing values ​​in the data. Piecewise smooth polynomials are used to fit known data points to ensure the continuity and smoothness of the time series, resulting in a preprocessed regular dataset.

[0012] Step 3, Mode Decomposition: The preprocessed wind power sequence is decomposed using an improved adaptive noise complete ensemble empirical mode decomposition (ICEEMDAN) algorithm. By adding standard Gaussian white noise with zero mean to the original sequence multiple times and performing empirical mode decomposition, the corresponding mode components obtained from multiple decompositions are then ensemble averaged. This effectively overcomes the mode aliasing problem of traditional mode decomposition algorithms and decomposes the original wind power sequence into several inherent mode components and residual components with clear physical meaning.

[0013] Step 4, Noise Mode Identification: Combining the sample entropy algorithm, calculate the sample entropy value of each intrinsic mode component. Utilize the characteristic of sample entropy in representing sequence complexity, and automatically locate and identify the mode components containing high-frequency noise based on the abrupt change characteristics of entropy value, thereby achieving accurate differentiation between noise modes and effective signal modes.

[0014] Step 5, Signal Denoising and Reconstruction: Establish a Savitzky-Golay (SG) filter model, and perform denoising and filtering only on the identified high-frequency noise modes. This filter achieves smooth denoising through the local polynomial least squares fitting method within a sliding window, effectively filtering out noise while preserving the original form and useful information of the signal to the greatest extent. The denoised noise mode components are reconstructed with the remaining unprocessed effective mode components to obtain the denoised wind power sequence.

[0015] Step 6, Spatiotemporal Fusion Prediction: Construct a GAT-BiLSTM combined prediction model, using denoised wind power data and preprocessed multidimensional meteorological data as joint inputs. First, a Graph Attention Network (GAT) is used to treat wind power data and various meteorological factor data as graph nodes. A masked self-attention mechanism is used to dynamically assign weights to the node edges, deeply mining and quantifying the spatial correlation between various features, and outputting a feature sequence that integrates correlation weights. Then, this feature sequence is input into a Bidirectional Long Short-Term Memory Network (BiLSTM). The forward LSTM layer and the backward LSTM layer capture the past and future temporal information of the sequence, respectively, fully extracting the bidirectional temporal dependence of the feature sequence. Finally, a fully connected layer maps the final hidden state of the BiLSTM to the predicted wind power value, obtaining a high-precision wind power prediction result.

[0016] Furthermore, the interpolation method described in this invention employs cubic spline interpolation, constructing a cubic polynomial interpolation function between two adjacent known data points to achieve smooth filling of missing values ​​and avoid abrupt distortion of time-series data; the ICEEMDAN algorithm adaptively adjusts the noise amplitude coefficient to match the signal-to-noise ratio of the added noise with the current residual component, improving the accuracy and adaptability of the decomposition; the sample entropy algorithm quantifies the complexity of modal components by calculating the matching probability of similar patterns in different dimensional spaces of the sequence. The higher the entropy value, the stronger the randomness of the modal components, providing a quantitative criterion for the identification of high-frequency noise modes.

[0017] Furthermore, the GAT-BiLSTM combined prediction model described in this invention achieves dual feature mining in both spatial and temporal dimensions. The GAT layer enhances the features of the multidimensional input features at the same time and mines spatial correlations, realizing a deep representation of the coupling relationship between wind power and multidimensional meteorological factors. The BiLSTM layer captures the temporal dependence of the feature sequence output by GAT, comprehensively grasping the changing pattern of wind power in the time dimension. The model directly outputs the predicted wind power value for one or more future time steps through an end-to-end mapping method, improving the efficiency and practicality of prediction.

[0018] The core innovation of this invention lies in the organic combination of advanced signal decomposition and denoising technology with a spatiotemporal fusion deep learning model. It proposes targeted solutions to two major technical challenges in wind power prediction: Firstly, through a combined denoising strategy of ICEEMDAN + sample entropy + SG filter, accurate denoising of wind power signals is achieved, overcoming the mode aliasing problem of traditional decomposition algorithms and avoiding the loss of useful information caused by directly removing noise modes. Secondly, by constructing a GAT-BiLSTM combined model, it for the first time uses wind power and multidimensional meteorological factors as graph nodes for spatial correlation mining. Combined with the bidirectional temporal feature extraction capability of BiLSTM, it achieves deep fusion of spatiotemporal correlation between wind power and meteorology, solving the problems of traditional models neglecting dynamic coupling relationships between features and insufficient feature mining.

[0019] The integrated energy wind power prediction method based on modal decomposition denoising and GAT-BiLSTM provided by this invention has the following beneficial effects: 1. This invention accurately solves the core technical problems of wind power prediction in integrated energy systems being affected by high-frequency noise interference and insufficient mining of the spatiotemporal correlation of multidimensional meteorological factors, and significantly improves the accuracy and robustness of wind power prediction results, providing reliable data support for system operation decisions.

[0020] 2. This invention uses the ICEEMDAN algorithm to perform mode decomposition on the original wind power sequence, which effectively overcomes the problems of mode aliasing and baseline drift in traditional EMD and EEMD algorithms. The inherent mode components obtained by decomposition have clear frequency layering and clear physical meaning, laying a high-quality data foundation for subsequent noise reduction processing.

[0021] 3. This invention combines the sample entropy algorithm to achieve automatic and accurate localization of high-frequency noise modes. It quantifies and distinguishes noise modes from effective signal modes by using the entropy value mutation characteristics, avoiding the subjectivity and error of manual identification and improving the accuracy and efficiency of noise identification.

[0022] 4. This invention uses a Savitzky-Golay filter to specifically filter and reduce noise in high-frequency noise modes, rather than directly eliminating noise modes. While effectively filtering out high-frequency random noise, it retains the useful information in the signal to the greatest extent, avoiding the decrease in prediction accuracy caused by information loss.

[0023] 5. The GAT model constructed in this invention uses wind power and multi-dimensional meteorological factors such as wind speed, wind direction, temperature, humidity, and air pressure as graph nodes to deeply explore the dynamic spatial correlation between various features, thus solving the problem of traditional models using meteorological data as parallel input and ignoring feature coupling relationships.

[0024] 6. This invention dynamically assigns weights to node edges using the masked self-attention mechanism of GAT, which can adaptively quantify the nonlinear influence of different meteorological factors on wind power at different times, and realize a refined characterization of the spatial coupling relationship between wind power and meteorology.

[0025] 7. The GAT-BiLSTM combined prediction model constructed in this invention achieves dual in-depth mining of spatial and temporal features. First, spatial correlation is fused through GAT, and then bidirectional temporal dependence is extracted through BiLSTM, forming a spatiotemporal fusion feature mining system that makes full use of the implicit information in multidimensional data.

[0026] 8. The BiLSTM network in this invention captures the past and future time-series information of the sequence through two LSTM layers, forward and backward, respectively, to fully grasp the changing pattern of wind power in the time dimension. Compared with the traditional unidirectional LSTM, it can more accurately capture the fluctuation characteristics of wind power.

[0027] 9. This invention uses the isolated forest algorithm to identify outliers in the data. It achieves accurate identification based on the sparsity of outliers and combines it with cubic spline interpolation to fill missing values, thus ensuring the continuity, smoothness and integrity of the time series and improving the quality of the input data.

[0028] 10. The prediction method of this invention achieves standardization of the entire process from data acquisition, preprocessing, noise reduction and reconstruction to model prediction. The algorithms of each step are closely connected and highly operable, requiring no complex manual intervention, and are suitable for wind power prediction scenarios of integrated energy systems with wind farms of different sizes.

[0029] 11. This invention organically combines signal decomposition and noise reduction technology with deep learning combined models to form a brand-new wind power prediction system, breaking through the bottlenecks of existing technologies in noise reduction accuracy and feature mining, and providing new technical ideas and implementation solutions for wind power prediction in integrated energy systems.

[0030] 12. The high-precision wind power prediction results output by the prediction method of the present invention can provide a reliable basis for the economic dispatch, unit combination and output optimization of the integrated energy system, which helps to improve the renewable energy consumption rate, reduce system operating costs and ensure the safe and stable operation of the system.

[0031] 13. The GAT-BiLSTM combined model of the present invention adopts an end-to-end mapping method, which directly maps the vector that integrates spatiotemporal features to the predicted wind power value, simplifies the model inference process, improves the prediction efficiency, and can meet the time requirements of real-time scheduling of integrated energy systems.

[0032] 14. The present invention has targeted processing of each modal component in the noise reduction process. It only filters the noise mode and keeps the effective signal mode in its original state, thus restoring the true characteristics of the wind power sequence to the greatest extent and enabling the prediction model to learn a more realistic wind power output pattern.

[0033] 15. The technical solution of the present invention has good scalability. It can add meteorological, environmental and other related features as nodes of the GAT model according to actual application needs without making major adjustments to the overall algorithm framework. It can adapt to the feature expansion needs under different application scenarios.

[0034] 16. The prediction method of the present invention significantly improves the anti-interference ability of wind power prediction. Even under complex meteorological conditions and slight data anomalies, it can still output stable prediction results. The model has stronger generalization ability and practical engineering applicability. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the GAT multi-head graph attention of the present invention; Figure 3 This is a schematic diagram of the BiLSTM structure of the present invention; Figure 4 This is an exploded view of the ICEEMDAN of the present invention; Figure 5 This is a graph showing the wind power prediction results of the present invention. Detailed Implementation

[0036] The technical solutions of the present invention will be further described below with reference to the embodiments and accompanying drawings: Example 1 like Figures 1 to 5 As shown, this embodiment provides a comprehensive energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM, which specifically includes the following steps: Step 1: Collect external meteorological data and wind turbine power data for the integrated energy system; Step 2: Perform data preprocessing on the collected data, handling outliers and missing values; Step 3: Apply the improved adaptive noise complete set empirical mode decomposition algorithm to the processed data to decompose the original wind power sequence containing high-frequency noise into several modal components. Step 4: Combine the sample entropy algorithm to calculate the entropy value of each modal component, and automatically locate and identify the high-frequency noise mode based on the entropy value mutation characteristics; Step 5: Establish a Savitzky-Golay filter model, perform filtering and noise reduction only on the high-frequency noise mode, and reconstruct the noise-reduced mode components with the other mode components to obtain the noise-reduced wind power sequence. Step 6: Construct a GAT-BiLSTM (Graph Attention Network-Bidirectional Long Short-Term Memory) combined prediction model, input the denoised wind power data and meteorological data into the model for prediction, and obtain the final wind power prediction result.

[0037] Furthermore, step 1 specifically involves using a sensor network deployed within and around the wind turbine generator itself, the wind farm, and its surroundings to acquire the wind turbine generator's output power sequence and related external meteorological data. The meteorological data specifically includes wind speed, wind direction, temperature, humidity, and air pressure, ensuring that all data sequences are collected and recorded synchronously at fixed time steps.

[0038] Furthermore, step 2 specifically involves applying the Isolation Forest algorithm to identify outliers in the data. This algorithm is based on the principle that outliers are more easily isolated in random hyperplane cutting due to their sparsity. Simultaneously, interpolation is used to impute missing data items. This method ensures the continuity and smoothness of the time series by fitting a piecewise smooth polynomial to the known data points. The Isolation Forest algorithm is specifically expressed as follows: (5); In the formula For data points The average path length across all isolated trees. The shorter the path, the better. The smaller; The size of the dataset used for training; This is an estimate of the average path length used for standardization; It is a harmonic number; For data points Abnormal scores; when When it is close to 1, it means The value is very small, so this point is considered an outlier.

[0039] The piecewise smooth polynomial used in this invention is a cubic spline value, which is applied to every two known data points. and Between each interval Above, construct a cubic polynomial, specifically represented as: (6); In the formula, For the first A cubic spline interpolation function over an interval, for any given interval... Values ​​such as the time points where data is missing, as long as the Located in the interval Within this range, the corresponding interpolation can be calculated using this formula. value; No. The coefficients to be determined in each interval.

[0040] Furthermore, step 3 specifically involves processing the wind power sequence using an improved adaptive complete ensemble empirical mode decomposition (EMD) algorithm. This algorithm injects specific zero-mean Gaussian white noise into the signal multiple times, performs EMD on the signal after each injection, and finally performs ensemble averaging on the obtained mode components. This aims to effectively suppress mode aliasing in traditional EMD, thereby improving the decomposition accuracy. The ICEEMDAN algorithm is specifically expressed as follows: No. Modal components Represented as: (7); No. Residual components Represented as: (8); The final decomposition is expressed as: (1); In the formula, For the first The intrinsic mode component is the _n ... The signal components extracted in the step; and The first Step and the first The residual components obtained after step calculation; For the first The noise amplitude coefficient of the stage is used to adjust the intensity of the added noise in each iteration to match the signal-to-noise ratio of the current residual component, thus achieving "adaptive noise". For the first The standard Gaussian white noise with a mean of zero and a variance of 1 was added in this experiment. This is the original wind power sequence; This represents the total number of modes derived from the decomposition.

[0041] Furthermore, step 4 specifically involves applying the sample entropy algorithm to calculate the entropy value of each modal component obtained from the previous step. This entropy value is used as a criterion to automatically locate and identify the modes representing high-frequency noise from all components.

[0042] Sample entropy can measure the complexity of a time series; it is calculated by transcribing the sequence in terms of... When similar in 3D space, in This is achieved by using the negative logarithm of the conditional probability that the elements in the dimensional space are still similar. Specifically: (2); In the formula, The sample entropy value is a non-negative number. The larger the value, the more complex and unpredictable the sequence. The embedding dimension represents the length of the template vector used for comparison; the smaller the value, the simpler and more regular the sequence. For length is The total number of similar vector pairs; For length is The total number of similar vector pairs.

[0043] Sample entropy, based on the reconstruction and pattern matching probability of a time series, aims to measure the "unpredictability" or complexity of a signal. If a sequence is highly regular, the probability of matching similar patterns in a high-dimensional space is high, resulting in a low entropy value; conversely, if the sequence is random or disturbed by noise, the matching probability is low, and the entropy value is high.

[0044] Furthermore, step 5 specifically involves using a Savitzky-Golay filter to smooth the high-frequency noise modes. This filter is based on local polynomial least squares fitting technology within a sliding window, which can efficiently filter out noise interference while preserving the original form and useful information of the signal to the greatest extent, overcoming the information loss problem caused by directly eliminating noise modes.

[0045] The core idea of ​​the SG (Savitzky-Golay) filter is indeed local polynomial least squares fitting. In practical applications, this fitting process is ultimately simplified to a convolution operation, specifically represented as: (3); In the formula, The output value after smoothing and noise reduction is located at the center point of the window. ; The coefficients of the SG filter are pre-calculated and depend on the lengths of the two parameter windows and the order of the fitted polynomial. The input data points are the original, noisy data points. The mode input with the largest sample entropy is filtered by an SG filter, and the final output is a smoothed and denoised mode sequence.

[0046] Furthermore, step 6 specifically involves constructing a GAT-BiLSTM combined prediction architecture, using the denoised wind power data and multidimensional meteorological data as its joint input. In this architecture, the Graph Attention Network (GAT) first processes the input data, treating wind power and various meteorological factors as graph nodes to deeply mine the spatial correlations between features and generate a feature sequence incorporating attention weights. Subsequently, this feature sequence is passed to a Bidirectional Long Short-Term Memory (BiLSTM) network, which captures the bidirectional dependencies of the sequence in the time dimension. The model ultimately outputs a high-precision wind power prediction value. Specifically: First, GAT is used to handle the spatial correlation between features: The function of GAT is to receive... The GAT layer takes all input features at any given time: denoised power, wind speed, temperature, and other weather characteristics, and treats them as nodes on a graph. It updates the feature representation of each node by calculating attention weights, thus incorporating the correlation weights of other factors into the wind power output. The output of the GAT layer is the updated node feature. By analyzing the characteristics of neighboring nodes The weighted sum is obtained as follows: (4); In the formula, for Time Node The output feature vector, after aggregating information from other nodes, is one element in the feature sequence that incorporates relevance weights in the GAT output; node For an input feature; For activation functions; For nodes In this invention, all meteorological factors are connected to the wind power to be predicted, meaning that all meteorological features are wind power nodes. Neighbors; A trainable weight matrix is ​​used to weight the original features of all nodes. Perform linear transformations to enhance its expressive power; for Time Node The original input feature vector; The attention weights calculated for GAT represent the attention weights in Time Node To what extent should we pay attention to nodes? The information specifically indicates: (9); In the formula, It is calculated through a shared attention mechanism, specifically: (10); In the formula, This indicates a splicing operation.

[0047] Next, the output of GAT is transformed into the input of BiLSTM to construct the combined model: By combining the outputs of GAT at each time step to form a complete feature sequence, this sequence is then used as the input to BiLSTM. Specifically, this is represented as follows: (11); In the formula, For BiLSTM The input vector at time step; The total number of input features; For the splicing operation, GAT is set to all The updated feature vector calculated from each feature node Concatenate them into a longer vector ,this Vectors contain The original information of the wind power nodes at any given time is combined with the spatial correlation between all features calculated by GAT.

[0048] Finally, a BiLSTM model is used as input, receiving the vector extracted from the GAT spatial features, to capture the bidirectional time dependence of the sequence and perform the final wind power prediction. Specifically, it is represented as follows: (12); (13); (14); (15); In the formula, For forward LSTM in The hidden state at any given moment; For backward LSTM in The hidden state at any given moment; For BiLSTM in Combined output of moments; It is a fully connected layer; For BiLSTM in the last time step The combined output is then fed into a fully connected layer to map to the final predicted value. ; This is the final wind power prediction result.

[0049] Furthermore, the GAT model uses the denoised wind power sequence along with meteorological data sequences such as wind speed, wind direction, temperature, humidity, and air pressure as graph nodes. It dynamically assigns weights to the edges of the nodes in the graph using a masked self-attention mechanism. The aim is to adaptively learn and quantify the nonlinear influence of different meteorological factors on wind power at different times, thereby achieving in-depth mining of spatial correlations. Its core is the calculation node as described in formula (10). For neighboring nodes Attention coefficient .

[0050] Furthermore, the bidirectional long short-term memory network in the GAT-BiLSTM combined prediction model receives the feature sequence output from the GAT layer, which has been fused with spatial correlation weights, as input. Its core function is to deeply extract the bidirectional dependence of this feature sequence in the time dimension. Specifically, BiLSTM uses a forward LSTM layer and a backward LSTM layer. The forward layer captures positive time-series information from the past to the present, while the backward layer captures negative time-series information from the future to the present. By fusing the hidden states in both directions, the model can more comprehensively understand the dynamic contextual relationship between wind power and meteorological factors at the current moment, thereby more accurately grasping the complete variation law of wind power fluctuations and improving the accuracy of time-series prediction.

[0051] Furthermore, the input to the GAT-BiLSTM combined prediction model is the denoised wind power data within a historical time window and the corresponding multidimensional meteorological data features, and the output is the predicted wind power value for one or more future time steps. The data flow of this combined model is as follows: the model first performs feature enhancement and spatial correlation mining on the multidimensional input features at the same time through the GAT layer, and outputs a sequence that integrates feature weights; then, the BiLSTM layer captures the temporal dependencies of the feature sequence output by GAT in the time dimension; finally, a fully connected output layer (Dense Layer) maps the final hidden state of BiLSTM to the required predicted value, realizing an end-to-end mapping from spatiotemporal features to prediction results.

[0052] Example 2 In another preferred embodiment, based on Embodiment 1, this embodiment provides a comprehensive energy wind power prediction method based on modal decomposition denoising and GAT-BiLSTM. The implementation process, technical details, and results of each step are described in conjunction with the accompanying drawings, verifying the feasibility and effectiveness of this method in wind power prediction. This embodiment is implemented based on the actual operation data acquisition and processing system of wind farms in a comprehensive energy system. Python is used as the core development language, combined with a deep learning framework to construct and train the GAT-BiLSTM combined model. All algorithms follow the mathematical model and parameter setting requirements described in this invention, such as... Figure 1 As shown, the specific implementation steps are as follows: Step 1: Collect external meteorological data and wind turbine power data for the integrated energy system. This embodiment uses a sensor network and meteorological monitoring stations deployed in the wind turbine itself, inside and around the wind farm to simultaneously collect power output data of the wind turbine within the integrated energy system, as well as five core external meteorological data: wind speed, wind direction, temperature, humidity, and air pressure.

[0053] During implementation, all data are collected and stored synchronously at a uniform time interval, ensuring that the timestamps of power data and meteorological data are strictly matched. This avoids feature correlation deviations caused by time sequence misalignment and lays a time-series consistent original data foundation for subsequent data processing and model training.

[0054] Step 2: Perform data preprocessing on the collected data, handling outliers and missing values. This embodiment uses the isolated forest algorithm to identify outliers in the original collected data. Based on the core principle of the isolated forest algorithm, a random hyperplane is constructed to cut the data space. By utilizing the sparsity of outliers in the data space, the outliers in the power and meteorological data can be accurately identified and labeled. For missing values ​​in the dataset caused by sensor failure or communication interruption, interpolation is used to fill them in. Known valid data points are fitted by piecewise smooth polynomials to ensure the continuity and smoothness of the wind power time series and the meteorological time series.

[0055] The piecewise smooth polynomial used in this embodiment is a cubic spline interpolation polynomial. A cubic polynomial function is constructed between two adjacent known data points. The corresponding values ​​of missing time points are calculated through this function, which effectively avoids abrupt distortion of time series data after missing value filling. After preprocessing, a regular dataset without anomalies or missing values ​​is obtained.

[0056] Step 3: Wind power sequence mode decomposition based on ICEEMDAN algorithm like Figure 4The figure shown is the ICEEMDAN decomposition diagram in this embodiment. In this embodiment, the preprocessed wind power sequence is decomposed using the improved adaptive noise complete set empirical mode decomposition (ICEEMDAN) algorithm.

[0057] During implementation, standard Gaussian white noise with a mean of zero and a variance of 1 is added multiple times to the original wind power sequence. Empirical mode decomposition is performed on the sequence after each addition of noise, and the corresponding mode components obtained from multiple decompositions are then ensemble-averaged. Finally, the original wind power sequence is decomposed into several intrinsic mode functions (IMFs) and one residual component. The decomposition result satisfies formula (1): (1).

[0058] from Figure 4 As can be clearly seen in the ICEEMDAN decomposition diagram, the original wind power sequence is decomposed into modal components of different frequencies. The frequency layers of each component are clearly defined, which effectively overcomes the mode aliasing problem of the traditional EMD algorithm and provides a decomposition result with clear physical meaning for subsequent noise mode identification.

[0059] Step 4: Locate and identify high-frequency noise modes using the sample entropy algorithm. This embodiment uses the sample entropy algorithm to calculate the sample entropy value for each intrinsic mode component obtained by ICEEMDAN decomposition in step 3. The sample entropy calculation follows formula (2): (2); The sequence complexity and randomness of each modal component are quantified by the sample entropy value.

[0060] During implementation, based on the abrupt change characteristics of sample entropy values, modal components with significantly high entropy values ​​and strong sequence randomness are identified as high-frequency noise modes. This enables accurate differentiation between high-frequency noise modes and effective signal modes, providing a clear target for subsequent targeted noise reduction and avoiding misprocessing of effective signal modes.

[0061] Step 5: Signal Denoising and Reconstruction Based on Savitzky-Golay Filter In this embodiment, a Savitzky-Golay filter model is established, and only the high-frequency noise modes identified in step 4 are filtered and denoised without any modification to the effective signal modes. The Savitzky-Golay filter achieves smooth denoising by performing local polynomial least squares fitting on the data points within a sliding window. The filtering process is simplified to a convolution operation, as shown in formula (3): (3); While effectively filtering out high-frequency noise, it retains the useful signal information contained in the noise mode to the greatest extent, avoiding the information loss caused by the traditional direct removal of noise modes.

[0062] After filtering the high-frequency noise modes, the noise-reduced noise mode components are linearly superimposed with the remaining unprocessed effective signal mode components and residual components to reconstruct the wind power signal, resulting in a noise-reduced wind power sequence. This sequence removes high-frequency noise interference while fully preserving the core features of the original signal.

[0063] Step 6: Construct a GAT-BiLSTM combined prediction model to achieve wind power prediction. This embodiment uses the denoised wind power data from step 5 and the preprocessed multidimensional meteorological data from step 2 as joint inputs to construct a GAT-BiLSTM combined prediction model for wind power prediction. The core structure of the model and the implementation process are combined. Figure 2 , Figure 3 Expanding on the final prediction results, as follows: Figure 5 As shown.

[0064] Step 1: Mining feature space correlations in the AT layer like Figure 2 The diagram shown is a schematic of GAT multi-head graph attention in this embodiment. In this embodiment, six types of features, namely wind power data, wind speed, wind direction, temperature, humidity and air pressure, are used as graph nodes of the GAT model. All meteorological feature nodes are set as neighbor nodes of the wind power node, and attention weights are dynamically assigned to the edges of each node through a mask self-attention mechanism.

[0065] In implementation, the original features of all nodes are first linearly transformed using a trainable weight matrix W, and then the attention coefficients between nodes are calculated. With normalized attention weights Finally, through formula (4): (4); The node features are updated, spatial correlations between features are mined, and a feature sequence incorporating correlation weights is output. From Figure 2 The attention weight allocation and feature fusion process of wind power nodes to each meteorological feature node can be clearly seen, realizing a deep representation of the spatial coupling relationship between wind power and meteorological features.

[0066] Step 6.2: Extract bidirectional time dependencies from the BiLSTM layer. like Figure 3 The diagram shown is a schematic of the BiLSTM structure in this embodiment. In this embodiment, the feature sequence with fused spatial correlation weights output by the GAT layer is concatenated and used as the input of the BiLSTM model to extract the bidirectional time dependency of the feature sequence.

[0067] The BiLSTM model consists of a forward LSTM layer and a backward LSTM layer. The forward LSTM layer captures the forward temporal information of the sequence from the past to the present time and outputs the forward hidden state. The backward LSTM layer captures the reverse temporal information of the sequence from the past to the present time and outputs the backward hidden state. The hidden states in the two directions are concatenated to obtain the combined output of the BiLSTM. Finally, the combined output of the last time step is processed through a fully connected layer. This is mapped to predicted wind power values, enabling a comprehensive capture of the bidirectional variation patterns in the time series. From Figure 3 The schematic diagram of the BiLSTM structure clearly shows the bidirectional temporal feature extraction and output process of the input sequence.

[0068] Step 6.3: Output of Model Prediction Results like Figure 5 The figure shown is a graph illustrating the wind power prediction results of this embodiment. The graph compares the original wind power data with the predicted data output by the model. From... Figure 5 The comparison results clearly show that the predicted data curve output by the GAT-BiLSTM combined prediction model constructed in this invention closely matches the original wind power data curve, accurately capturing the fluctuation pattern of wind power and effectively restoring the temporal variation characteristics of wind power, demonstrating the high-precision prediction effect of this method.

[0069] This embodiment completes the entire process of the integrated energy system wind power prediction method based on modal decomposition denoising and GAT-BiLSTM through the above steps, combined with Figures 1 to 5 The results of each step show that the combined noise reduction strategy of ICEEMDAN + sample entropy + Savitzky-Golay filter achieves accurate noise reduction of wind power signals and provides high-quality input data for the prediction model. The GAT-BiLSTM combined model deeply mines the spatial correlation of wind power-meteorological features through the GAT layer and fully extracts the bidirectional temporal dependence of the sequence through the BiLSTM layer, thus realizing dual feature mining in the spatial and temporal dimensions.

[0070] The implementation results of this embodiment verify that the method of the present invention can effectively solve the core problems of existing wind power prediction being affected by high-frequency noise interference and insufficient mining of spatiotemporal correlation of multidimensional meteorological factors. It significantly improves the accuracy and robustness of wind power prediction, and the output prediction results can provide reliable decision support for the economic dispatch, safe operation and renewable energy consumption of integrated energy systems.

[0071] This embodiment is only a preferred implementation of the present invention. In actual engineering applications, relevant parameters can be adjusted within the algorithm framework described in the present invention according to the equipment configuration, data characteristics, and prediction time scale of the wind farm to adapt to different integrated energy system wind power prediction scenarios.

[0072] In the preferred embodiment, the data collection in step 1 is achieved through sensors deployed in and around the wind farm and on the wind turbine itself. The external meteorological data includes wind speed, wind direction, temperature, humidity, and air pressure. All data is collected and stored synchronously at uniform time intervals. This setup enables comprehensive and accurate acquisition of information about the wind farm's operation and surrounding environment. Synchronous collection and storage at uniform time intervals ensures data integrity and consistency, providing a reliable foundation for subsequent analysis. Based on this rich and accurate data, a deeper understanding of the relationship between wind power and various meteorological factors can be achieved. This allows the prediction model to learn patterns based on comprehensive information, effectively improving prediction accuracy and providing strong support for the stable operation and rational scheduling of the wind farm, thus reducing economic losses caused by inaccurate predictions.

[0073] In the preferred embodiment, step 2, data preprocessing, employs the Isolation Forest algorithm to identify outliers. This is achieved by constructing a random hyperplane to divide the data space and recognizing outliers based on their sparsity. Interpolation is used to fill in missing values, and piecewise smooth polynomials are used to fit known data points, ensuring the continuity and smoothness of the time series. These settings enhance data usability, laying the foundation for building accurate prediction models and improving the reliability of wind power prediction, thus ensuring the efficient operation of wind farms. The Isolation Forest algorithm quickly and accurately identifies outliers in the data, avoiding interference from them in subsequent analysis and improving data quality. Interpolation, using piecewise smooth polynomial fitting, can reasonably infer missing values ​​based on known data, ensuring the continuity and smoothness of the time series and making the data more consistent with reality.

[0074] In the preferred embodiment, the piecewise smooth polynomial is a cubic spline interpolation polynomial. A cubic polynomial is constructed between adjacent known data points to calculate missing values. This configuration ensures the integrity and accuracy of the time series, providing high-quality data for subsequent model training, thereby improving the accuracy of wind power prediction and providing a reliable basis for the scientific scheduling of wind farms. The cubic spline interpolation polynomial has good smoothness and continuity. Constructing a cubic polynomial between adjacent known data points to calculate missing values ​​can better fit the local variation trend of the data. Compared with other interpolation methods, it can more accurately reflect the actual characteristics of the data, making the filled data closer to the true value.

[0075] In the preferred embodiment, the improved adaptive noise-complete set empirical mode decomposition algorithm in step 3 involves adding Gaussian white noise with zero mean multiple times and performing empirical mode decomposition, then averaging the corresponding mode components obtained from the multiple decompositions. This setting more accurately decomposes the wind power signal into components of different frequencies, facilitating subsequent targeted processing and analysis of each component. It helps to deeply explore the inherent patterns of the wind power signal, improves the predictive model's adaptability to complex signals, and thus enhances the accuracy of wind power prediction. Adding Gaussian white noise multiple times and averaging the decompositions effectively overcomes the mode aliasing problem of traditional mode decomposition, making the decomposed mode components more accurately reflect the true characteristics of the signal.

[0076] In the preferred embodiment, the Savitzky-Golay filter in step 5 achieves noise reduction through local polynomial least squares fitting within a sliding window, simplifying the filtering process to a convolution operation. This configuration results in higher quality data after noise reduction, providing cleaner data for subsequent model training and contributing to improved performance and prediction accuracy of the wind power prediction model. Utilizing local polynomial least squares fitting within a sliding window, the Savitzky-Golay filter effectively filters out noise while preserving important signal features. Simplifying the filtering process to a convolution operation improves computational efficiency and reduces computation time. This enables rapid noise reduction processing when handling large-scale wind power data, ensuring data timeliness and validity.

[0077] In the preferred embodiment, in step 6, the GAT-BiLSTM combined prediction model first mines the spatial correlation between features using GAT, and then inputs the feature sequence output by GAT into BiLSTM to extract bidirectional temporal dependencies. GAT uses wind power data and multidimensional meteorological data as graph nodes, and dynamically assigns weights to the node edges through a masked self-attention mechanism to quantify the influence of different meteorological factors on wind power. This setup fully leverages the respective advantages of GAT and BiLSTM, enabling the model to learn features from both spatial and temporal dimensions, improving its ability to capture complex changes in wind power, thereby increasing prediction accuracy and facilitating optimized wind farm scheduling. GAT can deeply mine the spatial correlation between wind power and multidimensional meteorological factors, and dynamically assigns weights through a masked self-attention mechanism to accurately quantify the influence of each factor, providing the model with richer feature information. BiLSTM can extract the bidirectional temporal dependencies of feature sequences, comprehensively grasping the changing patterns of wind power over time.

[0078] In the preferred embodiment, the BiLSTM captures past and future information of the sequence through forward LSTM and backward LSTM layers, respectively. The two hidden states are then fused and input into a fully connected layer to obtain the predicted wind power value, achieving dual feature mining in both spatial and temporal dimensions. This configuration allows the model to better adapt to the complex changes in wind power, improving the accuracy and stability of predictions and providing strong support for the safe operation and efficient management of wind farms. The forward and backward LSTM layers capture past and future information of the sequence, comprehensively considering the temporal dependencies of wind power without omitting any crucial information. The fusion of the two hidden states and input into the fully connected layer integrates features from both spatial and temporal dimensions, making the model's wind power prediction more closely aligned with reality.

[0079] In summary, the wind power output prediction method for integrated energy systems based on mode decomposition denoising and GAT-BiLSTM proposed in this invention has achieved significant results in solving the key problems of wind power prediction in integrated energy systems. Its specific advantages are reflected in the following aspects: In the field of wind power forecasting for integrated energy systems, traditional methods face two major challenges. First, wind power signals are highly susceptible to high-frequency noise interference, and existing technologies struggle to effectively filter out this noise, resulting in poor signal quality and severely impacting forecast accuracy. Second, complex and dynamic spatiotemporal correlations exist between multidimensional meteorological factors and wind power, and traditional models struggle to fully exploit these relationships, leading to significant discrepancies between predicted and actual values. This invention focuses on these two core issues, conducting in-depth research to improve the accuracy and reliability of wind power forecasting.

[0080] In the signal preprocessing stage, this invention employs a unique and rational combination of technologies. For the first time, it combines an improved adaptive noise-complete ensemble empirical mode decomposition (ICEEMDAN), sample entropy, and a Savitzky-Golay filter to preprocess wind power sequences. Traditional mode decomposition methods, such as empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and complete ensemble empirical mode decomposition (CEEMDAN), generally suffer from mode aliasing, which leads to unclear physical meaning of the decomposed modal components and affects subsequent processing. The ICEEMDAN algorithm, by introducing adaptive noise, fundamentally overcomes the mode aliasing problem, giving the decomposed modal components clearer physical meaning and laying a solid foundation for accurate subsequent processing. The use of sample entropy enables automatic localization of high-frequency noise modes, avoiding the loss of useful information compared to traditional noise reduction methods that directly remove high-frequency modes. The Savitzky-Golay filter further smooths and denoises the signal, achieving accurate noise reduction of the original signal, effectively improving signal quality, and providing high-quality data support for subsequent feature mining and prediction.

[0081] In terms of multidimensional feature mining, the graph attention network model constructed in this invention exhibits unique advantages. Traditional models often struggle to effectively mine the dynamic spatial correlations among multidimensional input features. This invention uses denoised wind power along with various meteorological factors such as wind speed, temperature, and air pressure as graph nodes. Through the self-attention mechanism of the graph attention network (GAT), the influence weight of each factor on wind power can be dynamically quantified. This mechanism can automatically focus on factors that have a significant impact on wind power and assign them higher weights, thereby solving the shortcomings of traditional models in mining multidimensional feature spatial correlations. It achieves a deep representation of spatial coupling relationships and provides richer and more representative feature information for subsequent predictions.

[0082] In terms of predictive model construction, the GAT-BiLSTM combined prediction model adopted in this invention has significant advantages. The feature sequence output by GAT, which incorporates spatial correlation weights, is input into a bidirectional long short-term memory network (BiLSTM). BiLSTM, with its bidirectional nature, can simultaneously consider past and future information of the sequence, deeply capturing the temporal dependencies of the sequence. This combined model achieves feature mining in both spatial and temporal dimensions. Compared to single models or simple combinations, it can more comprehensively and accurately grasp the changing patterns of wind power, significantly improving the model's predictive accuracy and better adapting to the complex dynamic changes in wind power.

[0083] This invention organically integrates multiple techniques, including ICEEMDAN precise noise reduction, automatic sample entropy localization, GAT spatial correlation mining, and BiLSTM time dependency capture, to form a complete method for predicting wind power output in integrated energy systems. This method not only overcomes the limitations of existing technologies such as incomplete noise reduction and insufficient feature mining, but also improves the overall performance and robustness of the prediction model. Implementing the method proposed in this invention can significantly improve the accuracy of wind power prediction. In practical applications, accurate wind power prediction has significant real-world implications. In terms of economic dispatch, it can help dispatchers rationally arrange power generation plans, reduce power generation costs, and improve energy utilization efficiency. In terms of safe operation, it can predict changes in wind power output in advance, enabling dispatchers to take timely measures to avoid system failures and ensure the stable operation of the power system. In terms of renewable energy consumption, it helps improve the utilization rate of wind power, reduce wind curtailment, and promote the effective use of renewable energy. Therefore, this invention has significant economic benefits and broad application prospects.

[0084] This invention excels in precise signal denoising, spatiotemporal coupling mining of multidimensional features, and construction of combined deep learning models. It provides an efficient and reliable prediction solution for the field of integrated energy system optimization and scheduling, and is expected to promote the technological development and practical application of this field.

Claims

1. A comprehensive energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM, characterized in that: Includes the following steps: Step 1: Collect external meteorological data and wind turbine power data for the integrated energy system; Step 2: Perform data preprocessing on the collected data, handling outliers and missing values; Step 3: Apply the improved adaptive noise complete set empirical mode decomposition algorithm to the processed data to decompose the original wind power sequence containing high-frequency noise into several modal components. Step 4: Combine the sample entropy algorithm to calculate the entropy value of each modal component, and automatically locate and identify the high-frequency noise mode based on the entropy value mutation characteristics; Step 5: Establish a Savitzky-Golay filter model, perform filtering and noise reduction only on the high-frequency noise modes, and reconstruct the noise-reduced mode components with the other mode components to obtain the noise-reduced wind power sequence. Step 6: Construct a GAT-BiLSTM combined prediction model, inputting the denoised wind power data and meteorological data into the model for prediction, and obtain the final wind power prediction result.

2. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 1, characterized in that: The data collection in step 1 is achieved by collecting data from sensors deployed in and around the wind farm and the wind turbine itself. The external meteorological data includes wind speed, wind direction, temperature, humidity and air pressure. All data are collected and stored synchronously at a uniform time interval.

3. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 1, characterized in that: Step 2, data preprocessing, involves using the isolated forest algorithm to identify outliers. This is achieved by constructing a random hyperplane to divide the data space and recognizing outliers based on their sparsity. Interpolation is used to fill in missing values, and piecewise smooth polynomials are used to fit known data points to ensure the continuity and smoothness of the time series.

4. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM as described in claim 3, characterized in that: The piecewise smooth polynomial is a cubic spline interpolation polynomial, which is used to construct a cubic polynomial between adjacent known data points to calculate missing values.

5. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 1, characterized in that: The improved adaptive noise-complete set empirical mode decomposition algorithm in step 3 adds Gaussian white noise with zero mean multiple times and performs empirical mode decomposition. The corresponding mode components obtained from multiple decompositions are averaged, and the final decomposition expression is: （1）； In the formula, For the first One intrinsic modal component; For the first Step residual components; This is the original wind power sequence; This represents the total number of modes derived from the decomposition.

6. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 1, characterized in that, The calculation expression for the sample entropy algorithm in step 4 is as follows: （2）； In the formula, For embedded dimensions, For similarity tolerance, For sequence length, For length is The total number of similar vector pairs, For length is The total number of similar vector pairs is used to determine the high-frequency noise mode based on the sample entropy value.

7. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 1, characterized in that, In step 5, the Savitzky-Golay filter achieves noise reduction through local polynomial least squares fitting within a sliding window. The filtering process is simplified to a convolution operation, expressed as: （3）； In the formula, The output value after smoothing and noise reduction; These are the coefficients of the SG filter; These are the original, noisy input data points.

8. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 1, characterized in that: In step 6, the GAT-BiLSTM combined prediction model first mines the spatial correlation between features through GAT, and then inputs the feature sequence output by GAT into BiLSTM to extract bidirectional time dependence. GAT uses wind power data and multidimensional meteorological data as graph nodes, and dynamically assigns weights to node edges through a masked self-attention mechanism to quantify the impact of different meteorological factors on wind power.

9. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 8, characterized in that, The update expression for node features in the GAT layer is as follows: （4）； In the formula, for Time Node The output feature vector after aggregating information from other nodes; For activation functions; For nodes The neighboring nodes; is a trainable weight matrix; for Time Node The original input feature vector; Attention weights calculated for GAT.

10. The integrated energy wind power prediction method based on mode decomposition denoising and GAT-BiLSTM according to claim 8, characterized in that: The BiLSTM captures past and future information of the sequence through forward LSTM and backward LSTM layers respectively. After fusing the two hidden states, it inputs them into a fully connected layer to obtain the wind power prediction value, thus realizing dual feature mining in the spatial and temporal dimensions.

Images