Regional wind-solar combined power interval prediction method and system based on multi-task learning and hybrid neural network

Abstract

The invention discloses a regional wind-solar combined power interval prediction method and system based on multi-task learning and a hybrid neural network, and relates to the technical field of combined prediction. Meteorological data and wind-solar historical power data of a to-be-researched region are collected, and a historical data set is constructed; preprocessing the historical data set, and performing feature selection by adopting a predictive ability scoring algorithm to obtain screened meteorological data and historical power data; constructing a joint power interval prediction model; and performing combined power generation prediction based on the trained combined power interval prediction model. According to the method, the space-time correlation and the complementary relation of wind energy and photovoltaic energy are excavated, the joint power interval prediction model is constructed, and the power prediction precision is improved.

Description

Technical Field

[0001] This invention relates to the field of joint prediction technology, and more specifically to a method and system for regional wind and solar power joint prediction based on multi-task learning and hybrid neural networks. Background Technology

[0002] Both wind and solar power generation are characterized by volatility and intermittency, posing significant challenges to grid security and economic stability in the context of high renewable energy penetration. Therefore, improving the accuracy and efficiency of power generation forecasting is crucial for power systems. In recent years, researchers have explored various meteorological and environmental factors, historical data analysis, and real-time monitoring systems to improve forecast accuracy and reliability. However, when analyzing the key challenges of improving wind and solar power forecasting models, few studies have considered the correlations between different power generation resources. Traditional wind and solar power forecasting methods typically predict each energy source independently, neglecting their interrelationships and influences. However, within the same region, wind and solar power variables often exhibit complex coupling relationships. By establishing appropriate models and algorithms, accurate power generation forecasts for both energy sources can be achieved. Therefore, further research and innovation in this field are still needed.

[0003] Currently, wind and solar power forecasting methods are mainly divided into physical principle-based methods and data-driven methods. Physical methods are model-driven, utilizing mathematical models of the relationship between various meteorological factors and power generation, making them more suitable for medium- to long-term forecasts. Data-driven methods include traditional statistical methods and artificial intelligence methods, with artificial intelligence methods referring to forecasts using machine learning (ML) and deep learning (DL) models. Most traditional statistical methods used for solar / wind power forecasting are based on deterministic models and historical data to predict future time series and better reflect linear relationships, such as Autoregressive Moving Average (ARMA), Autoregressive Integrated Moving Average (ARIMA), and Kalman filter models. In recent years, artificial intelligence models have become increasingly popular due to their advantages in analyzing power generation uncertainties. Traditional machine learning models, including Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Extreme Learning Machines (ELM), have been applied to power forecasting. However, traditional machine learning models suffer from poor predictive performance because they cannot accurately capture the deep features of large-scale power generation data. Today, deep learning algorithms are receiving significant attention in the field of power forecasting. Various deep learning models have been introduced into power forecasting and have demonstrated good predictive performance. These algorithms are capable of extracting advanced information and identifying nonlinear correlations between massive amounts of data, demonstrating powerful mapping capabilities.

[0004] However, these methods primarily focus on forecasting single energy forms, while current trends in practical applications consider multiple power generation types within renewable energy clusters. Joint forecasting considering correlations in power systems mainly concentrates on the load side. In reality, weather systems within a specific region exhibit inertia, with strong coupling and correlations among various meteorological elements. However, many existing studies tend to model different forecasting factors independently, failing to consider the interrelationships between different renewable energy sources.

[0005] Therefore, how to provide a regional wind and solar power joint prediction method and system based on multi-task learning and hybrid neural networks, explore the spatiotemporal correlation and complementarity between wind energy and photovoltaic energy, and construct a joint power interval prediction model to improve the accuracy of power prediction are problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for regional wind and solar power joint prediction based on multi-task learning and hybrid neural networks, which explores the spatiotemporal correlation and complementarity between wind energy and photovoltaic energy, and constructs a joint power interval prediction model to improve the accuracy of power prediction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a regional wind and solar power joint range prediction method based on multi-task learning and hybrid neural networks, comprising:

[0008] Collect meteorological data and historical wind and solar power data of the area to be studied, and construct a historical dataset;

[0009] The historical dataset is preprocessed, and a predictive ability scoring algorithm is used for feature selection to obtain filtered meteorological data and historical power data.

[0010] A joint power range prediction model is constructed. The joint power range prediction model uses the PLE model as the prediction framework and combines the TCN-Attention expert network, the gating network and the QR-BiGRU tower network for prediction.

[0011] The joint power interval prediction model is trained based on the screened meteorological data and historical power data to obtain the trained joint power interval prediction model.

[0012] Joint power generation prediction is performed based on a trained joint power range prediction model.

[0013] Preferably, the joint power range prediction model includes: an input module, an extraction module, and an output module; the input module is used to receive historical power data and meteorological data, and to process the data using a sliding window; the extraction module is used to extract specific features and shared features of wind and solar power generation tasks; and the output module is used to output the predicted value for each task.

[0014] Preferably, the extraction module includes an expert module and a gating network, wherein the expert module includes a specific task expert module for wind power generation, a specific expert module for photovoltaic power generation, and a shared expert module;

[0015] The shared expert module extracts common feature information of wind and solar power generation tasks, while the specific expert module extracts unique feature information of wind and solar power generation tasks.

[0016] The gating network is used to dynamically select and weight the outputs of different experts based on the specific input of the task.

[0017] Preferably, the expert module uses a TCN-Attention expert network to process the input time-series data;

[0018] The TCN-Attention expert network uses TCN for local feature capture and a self-attention mechanism for global feature capture.

[0019] Preferably, the output module employs a QR-BiGRU network;

[0020] Time series prediction is performed using a BiGRU neural network, and confidence intervals are obtained using the QR algorithm.

[0021] Preferably, the BiGRU network consists of two GRUs in two directions; wherein, the forward GRU makes predictions based on time series data from front to back, and the backward GRU makes predictions based on time series data from back to front.

[0022] Simultaneously process the input sequences of the forward GRU and the backward GRU to obtain the forward and backward information of the sequences, and concatenate the outputs of the forward GRU and the backward GRU to obtain the concatenated result of the forward and backward hidden states.

[0023] Preferably, the QR algorithm can be used to regress the function parameters between the explanatory variable X and the explained variable Y at a certain quantile. The expression of the QR algorithm is as follows:

[0024] Q Y (τ∣X)=β(τ)X;

[0025] In the formula: Q Y (τ|X) represents Y in X=(x1,x2,…,x) k ) T The conditional quantile under the given condition, β(τ)=(β0(τ),β1(τ),β2(τ),…,β k (τ)) is the parameter vector, and τ is the quantile;

[0026] The parameter vector β(τ) is estimated using the following formula:

[0027]

[0028] In the formula, i is the sample index and N is the number of samples.

[0029] Preferably, a regional wind and solar joint power interval prediction system based on multi-task learning and hybrid neural networks includes:

[0030] The data acquisition module is used to collect meteorological data and historical wind and solar power data of the area under study to build a historical dataset;

[0031] The feature selection module is used to preprocess the historical dataset and use a predictive ability scoring algorithm to select features, thereby obtaining filtered meteorological data and historical power data.

[0032] The model building module is used to build a joint power range prediction model. The joint power range prediction model uses the PLE model as the prediction framework and combines the TCN-Attention expert network, the gating network and the QR-BiGRU tower network for prediction.

[0033] The model training module is used to train the joint power interval prediction model based on the screened meteorological data and historical power data, so as to obtain the trained joint power interval prediction model.

[0034] The prediction module is used to perform joint power generation prediction based on the trained joint power range prediction model.

[0035] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for predicting regional wind and solar combined power ranges based on multi-task learning and hybrid neural networks. It introduces a multi-task learning range prediction model, which employs quantile regression and a bidirectional gated recurrent unit (BIGRU) network to construct a regional wind and solar combined power generation prediction model. This model can effectively capture and share the interdependence between the two power generation types. This invention has the following beneficial effects:

[0036] (1) Traditional correlation coefficients can only measure the linear relationship between variables and are difficult to detect nonlinear coupling relationships between variables. This invention uses an improved Predictive Power Score (PPS) for feature selection and uses the Extreme Gradient Boosting (XGBoost) algorithm to replace the original Decision Tree (DT) algorithm for prediction. This can explore the linear and nonlinear correlation between meteorological characteristics and power generation variables and select features with strong correlation.

[0037] (2) To extract more in-depth features of wind and solar power meteorological factors, a multi-task learning framework is introduced to better learn the nonlinear coupling information between wind and solar power features. The Temporal Convolutional Fusion Attention (TCN-Attention) mechanism is used as the expert network for extracting modules within the multi-task learning framework to improve the performance of time series modeling. TCN-Attention combines the local feature capture of TCN with the global feature capture of the attention mechanism, allowing the network to focus on both local patterns and effectively capture long-term dependencies and global patterns in time series data during learning. (The correlation between multivariate features is extracted using the framework.)

[0038] (3) Wind and solar power generation is highly dependent on weather conditions, and is inherently uncertain and intermittent. To address these issues, a quantile regression bidirectional gated recurrent unit (QR-BiGRU) neural network is used to provide more comprehensive prediction information. By using QR-BiGRU as the output module of the tower network in a multi-task learning framework, more accurate and reliable interval estimation of wind and solar power generation prediction is achieved. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 A schematic diagram of the overall framework for wind and solar combined power range prediction provided in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the TCN-Attention structure provided in an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of a gating network structure provided in an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram of the GRU structure provided in an embodiment of the present invention.

[0044] Figure 5 This is a schematic diagram of the TCN structure provided in an embodiment of the present invention.

[0045] Figure 6 This is a schematic diagram of a two-layer dilated causal convolution structure provided in an embodiment of the present invention.

[0046] Figure 7A schematic diagram of the BiGRU structure provided in an embodiment of the present invention.

[0047] Figure 8 The PPS heatmap between photovoltaic features and photovoltaic power generation is provided for embodiments of the present invention.

[0048] Figure 9 The PPS heat map between wind characteristics and wind power generation is provided for embodiments of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention discloses a method for predicting regional wind and solar power intervals based on multi-task learning and hybrid neural networks, including:

[0051] Collect meteorological data and historical wind and solar power data of the area to be studied, and construct a historical dataset;

[0052] The historical dataset is preprocessed, and a predictive ability scoring algorithm is used for feature selection to obtain filtered meteorological data and historical power data.

[0053] A joint power range prediction model is constructed. The joint power range prediction model uses the PLE model as the prediction framework and combines the TCN-Attention expert network, the gating network and the QR-BiGRU tower network for prediction.

[0054] The joint power interval prediction model is trained based on the screened meteorological data and historical power data to obtain the trained joint power interval prediction model.

[0055] Joint power generation prediction is performed based on a trained joint power range prediction model.

[0056] In one specific embodiment of the present invention, wind power and photovoltaic power generation, both renewable energy sources, are intermittent energy sources, exhibiting characteristics such as randomness, fluctuation, and intermittency in output, which significantly impacts the accuracy of output prediction. Existing prediction methods for wind and solar power output primarily focus on single-energy prediction, rarely considering the spatiotemporal correlation between wind and solar power output-related variables within the same region. Furthermore, the lack of historical data for newly built wind and solar power stations in the same region also affects prediction accuracy. To address these issues, this embodiment of the present invention proposes a method for joint regional wind and solar power prediction based on a multi-task learning framework and QR-BiGRU. First, the Predictive Power Score (PPS) method is used to select highly correlated wind and solar meteorological features from historical wind and solar data. Second, a Progressive Layered Extraction (PLE) multi-task learning framework is introduced to better learn the nonlinear coupling information between different wind and solar power generation features. Simultaneously, a Temporal Convolutional Fusion Attention (TCN-Attention) mechanism is used as the expert network for extracting modules within the multi-task learning framework to improve the performance of time series modeling. Finally, considering the uncertainty and intermittency of wind and solar photovoltaic power generation, a quantile regression bidirectional gated recurrent unit (QR-BiGRU) neural network is used as the tower network of the PLE multi-task framework for predicting combined wind and solar power generation, resulting in more accurate and reliable interval estimates. Experimental results show that the proposed method improves feature information extraction and reliability, effectively enhancing prediction accuracy and interval estimation performance.

[0057] Specifically, such as Figures 1-4 As shown in the figure, the overall framework for wind-solar combined power range prediction proposed in this embodiment of the invention is mainly divided into two parts, as follows: Figure 1 As shown, the first part is data preprocessing and feature selection. First, outliers and missing values ​​in historical meteorological and wind / solar power data are detected. Then, these outliers and missing values ​​are filled with data from the previous moment to ensure data continuity. Next, the data is normalized to maintain a consistent data scale, which helps eliminate the impact of dimensional differences between different data sets. Finally, the predictive power score (PPS) method is used to calculate the correlation between meteorological characteristics and regional power generation, and features with low correlation scores are then eliminated from the analysis.

[0058] The second part is the model prediction section. This invention proposes a wind-solar joint power range prediction model based on the PLE framework and the TCN-Attention-BiGRU network. This model uses a progressive hierarchical extraction (PLE) multi-task learning model as the prediction framework, combining components such as the TCN-Attention expert network, gating network, and BiGRU tower network, and introducing quantile regression (QR) prediction to achieve wind-solar joint power prediction and generate prediction ranges. The entire model can be divided into three main modules: an input module, an extraction module, and an output module.

[0059] First, the input module receives meteorological data and historical power data filtered by PPS. The input data is used to construct samples through a sliding time window, and then converted into a format suitable for training before being passed to the extraction module.

[0060] Secondly, the extraction module is responsible for extracting specific and shared features of wind and solar power generation tasks, and selecting specific and shared information beneficial to the task based on the final prediction task. It mainly consists of an expert module (ExpertModels) and a gate mechanism. The extraction module includes two task-specific expert modules and one shared expert module. The shared expert module extracts common features across all tasks, while the task-specific expert module extracts features relevant to the specific task. The expert modules use a TCN-Attention network to process the input time-series data. The TCN-Attention network embeds a self-attention layer into the TCN network, combining the local feature capture of TCN with the global feature capture of the self-attention mechanism, improving the performance of time-series modeling and enabling flexible learning and representation of complex patterns in meteorological and historical power data. The gate mechanism is responsible for dynamically selecting and weighting the outputs of different experts based on the specific input of the task, achieving adaptive feature fusion. The output of the gate mechanism is a probability distribution representing the weighting coefficients of different experts.

[0061] Finally, the main task of the output module is to generate the final interval prediction results for wind power and solar power based on the weighted expert output generated by the extraction module. This model uses a QR-BiGRU network as the tower network of the output module. The two-layer GRU network architecture enables bidirectional extraction of features from the input of the extraction module, improving the completeness and globality of the features and achieving more accurate power prediction. Simultaneously, quantile regression (QR) prediction technology is employed to obtain more comprehensive interval prediction information, replacing the model's deterministic prediction and reducing the impact of the intermittent uncertainty in wind and solar power generation.

[0062] In one specific embodiment of the present invention, meteorological characteristics have a significant impact on power generation. Due to the complexity of atmospheric dynamics, the quantitative relationship between power generation and meteorological factors is difficult to model accurately, and the mathematical expression of their coupling relationship faces challenges. To analyze the predictive power generation capability of different meteorological characteristics, this embodiment of the present invention uses an improved Predictive Power Score (PPS) for measurement.

[0063] Pearson Predictive Power (PPS) is used to assess the predictive power between two variables. Traditional Pearson coefficients only measure linear relationships and are symmetrical. However, in the real world, the relationships between features are often asymmetrical; that is, the predictive power of A on B is different from the predictive power of B on A. This leads to the neglect of some features that appear to have low correlation coefficients but actually have high predictive power. PPS can handle not only linear relationships but also nonlinear relationships, and it measures an asymmetrical relationship. It can analyze the predictive power between variables in different directions and is applicable to different types of data. The core idea of ​​PPS is to evaluate the ability of a feature (variable) to predict a target variable, rather than simply measuring the linear relationship between them. PPS scores range from 0 to 1, where 0 indicates no predictive power and 1 indicates effective prediction. The relationship measured by PPS is asymmetrical. The principle of PPS is as follows:

[0064] (1) Single feature prediction: Only one feature is used to predict the target column, thereby avoiding the influence of multiple feature interactions.

[0065] (2) Cross-validation: Perform 4-fold cross-validation to calculate the score.

[0066] (3) Sampling limit: When the data volume is greater than 5000 rows, the data will be randomly sampled and limited to 5000 rows.

[0067] (4) Learning Algorithm: Extreme Gradient Boosting (XGBoost) is used as the prediction algorithm. XGBoost is a gradient boosting-based decision tree implementation with excellent performance and computational speed. It can handle large-scale data, is more sensitive to weakly correlated features, and can better capture complex nonlinear relationships. Moreover, the explicit regularization terms (L1 / L2) of XGBoost can prevent overfitting, making the PPS score more stable and reliable, and avoiding artificially high false correlations.

[0068] PPS uses Mean Absolute Error (MAE) as the evaluation metric, calculating the ratio of the MAE of the XGBoost model to the MAE of the baseline (the model predicting the median). The specific calculation process is as follows:

[0069]

[0070] In the formula, y represents the predicted value of one variable for another. i This represents the true value of the variable being predicted, and n represents the total number of samples.

[0071] Taking the screening of photovoltaic meteorological characteristics as an example, the detailed implementation steps are as follows:

[0072] (1) Data preparation

[0073] First, prepare raw data containing photovoltaic meteorological characteristics and photovoltaic power generation. The data includes meteorological characteristics such as the direct downward shortwave irradiance of the entire sky surface (Wh / m²). 2 The data includes parameters such as the target variable (photovoltaic power generation in kW), and outlier and missing value handling. Outliers are identified using box plots and treated as missing data, which are then filled in using data from the previous time step to ensure the continuity of the time series data. Finally, normalization is applied to eliminate the influence of different dimensions and magnitudes between different features.

[0074] (2) Calculate the score

[0075] The PPS was used to calculate the predictive power generation of each meteorological feature.

[0076] 1) Constructing the XGBoost model

[0077] The XGBoost model was used as the regression model, with the number of hyperparameter trees set to 30, the maximum tree depth to 3, and the learning rate to 0.001. The initial model's predictions were the mean of the target variable. The mean absolute error (MAE) was chosen as the loss function for the regression task, and the residual (the difference between the true and predicted values) for each sample was calculated as the target value for the new tree. Using the residuals as the target value, a new regression tree was constructed. The optimal split point was selected by calculating the gain, optimizing the tree structure. The model's predictions were updated based on the new tree's predictions, and the learning rate controlled the impact of the new tree on the model. This process of calculating residuals, constructing new trees, and updating the model was repeated until the preset number of trees was reached or the loss function converged. L1 / L2 regularization was used to prevent overfitting during this process. The final model consisted of multiple weighted decision trees, and the predictions from all trees were summed to obtain the final result.

[0078] 2) Cross-validation

[0079] The constructed model is used to perform 4-fold cross-validation on the data. Each time, three subsets are used to train the model, and the remaining subset is used for validation. The mean absolute error (MAE) of each validation is calculated, and the average value is taken as the final score of the model.

[0080] 3) Construct a benchmark model

[0081] A baseline model is constructed, using the median of the target variable as the predicted value for all samples. The mean absolute error (MAE) of the baseline model is calculated as the baseline score. In PPS, the baseline model serves as a reference point for measuring the performance of prediction models. The performance of the baseline model represents "how well the model can predict without any feature information." If a feature provides effective information, then an XGBoost model trained based on that feature will perform better than the baseline model. For example, an XGBoost model trained based on the all-sky surface shortwave downward direct irradiance (Wh / m^2) outperforms the baseline model, indicating that this feature contributes to photovoltaic power prediction.

[0082] 4) Calculate the predictive ability score

[0083] The MAE of the XGBoost model is compared with that of the baseline model. The predictive ability score is calculated using the PPS formula mentioned earlier. The higher the score, the stronger the predictive ability of the feature on the target variable. A PPS score of 0 indicates that the feature has no predictive ability on the target variable; a score of 1 indicates that the feature can effectively predict the target variable.

[0084] Through the above steps, PPS can capture the linear and nonlinear relationships between each photovoltaic meteorological feature and photovoltaic power generation.

[0085] Specifically, the joint power range prediction model includes an input module, an extraction module, and an output module; the input module is used to receive historical power data and meteorological data, and to process the data using a sliding window; the extraction module is used to extract specific and shared features of wind and solar power generation tasks; and the output module is used to output the predicted value for each task.

[0086] Specifically, multi-task learning enhances the learning process of each sub-task by utilizing shared information between tasks, thereby improving prediction accuracy and efficiency. In this embodiment, a multi-task learning (MTL) based model—Progressive Layered Extraction (PLE) network architecture—is used as the wind-solar combined power generation prediction algorithm. PLE is a novel MTL algorithm developed from the hard parameter sharing method. This framework includes three main parts: an input module, an extraction module, and an output module. The structure of the PLE model is as follows: Figure 1 As shown, it specifically includes:

[0087] (1) Input module: It is responsible for receiving historical power data of wind and solar power and key meteorological data, and processing the data with a sliding window to convert it into a format suitable for training as the basic input of the model.

[0088] (2) Extraction Module: Information is separated from the expert module through a gating network. The expert module includes shared experts and task-specific experts. Each expert is a feature extraction network, and the computation process of the expert sub-network is related to the selected network. The gating network is responsible for dynamically fusing the expert outputs. The expert outputs are input into the gating network to obtain the final output. The gating network calculation formula for task k is as follows:

[0089] g k (x)=w k (x)S k (x);

[0090]

[0091] Where x is the input; S represents the parameter matrix; k (x) is the selection matrix composed of the outputs of the shared network and the expert network, used for calculating task k:

[0092]

[0093] (3) Output Module: The output module outputs the predicted value for each task. Each task has a dedicated tower network. The output of the gating network is passed to the upper-level tower network, and the final prediction result is obtained through the data processing of the tower network. The calculation process of the tower network is as follows:

[0094] y k (x)=t k (g k (x));

[0095] Among them, t k This represents the tower network computation for task k; then, y k This represents the predicted output for task k.

[0096] In one specific embodiment of the present invention, the input module is responsible for receiving historical wind and solar power data and filtered key meteorological data, and processing the data using a sliding window to convert it into a format suitable for training, which serves as the basic input to the model and is then passed to the three expert modules of the extraction module. The shape of the input data is a 3-dimensional tensor (batch size = 128, time step = 12, feature dimension = 9), where the batch size is set to 128, i.e., 128 samples are input at a time, the time step is 12, and each time step contains 9 features (meteorological features + historical power).

[0097] The extraction module consists of two top-level gating networks and three bottom-level expert modules, used to extract specific and shared information about wind and solar power generation tasks.

[0098] 1) The expert module includes one wind power generation task expert, one photovoltaic power generation task expert, and one shared expert. These three experts receive three inputs of the same shape. Each expert uses a TCN-Attention network to extract features. The TCN-Attention network consists of an input layer, two TCN residual blocks, a self-attention layer, and a fully connected layer. The input layer receives the 3D tensor transformed by the input module. A residual block consists of two layers of dilated causal convolutions and a nonlinear mapping. The two TCN residual blocks compute dilated causal convolutions one after another, progressively extracting higher-level temporal features while preserving the temporal dimension through residual connections. This conveys local temporal information while retaining the features of the original input through residual connections. The output of the TCN block is a 3D tensor that preserves the time step and batch size, and updates the feature dimension. Its shape is (batch size = 128, feature dimension = 64, time step size = 12). The feature dimension is increased to 64 because the number of TCN channels is set to 64 to capture richer feature representations. The output of the second TCN block is transposed into a 3D tensor (batch size = 128, time step size = 12, feature dimension = 64) and fed into the self-attention layer to further model global dependencies. The self-attention layer receives the output of the last TCN as input and outputs a tensor of the same shape, but the features at each time step have been integrated with global context information, enabling further integration of global temporal features. The fully connected layer maps the features extracted by the TCN and attention mechanism to the target output space to complete feature integration and output, and then passes it to the gating network. The output shape is (batch size = 128, time step size = 12, feature dimension = 32). Reducing the feature dimension from 64 to 32 dimensions reduces subsequent computation while retaining the most useful information. The computation process of the expert network is related to the selected network; the specific computation process is described later in the introduction to the TCN network and self-attention mechanism. Different expert networks are responsible for learning different information.

[0099] 2) The gating network is responsible for dynamically fusing expert outputs. Each task's gating network connects a task-specific expert and a shared expert. The experts' outputs are input into the gating network. Before being passed to the gating network, the outputs of all experts for each category are stacked to form a new tensor with the shape (number of experts = 1, batch size = 128, time step = 12, feature dimension = 32). The number of experts is set to 1, meaning there is one expert for wind power generation, one for photovoltaic power generation, and one for shared experts. The gating network generates a weight vector using Softmax and performs a weighted summation of the outputs of the task-specific expert and the shared expert, thus generating the final feature representation for each task. Taking wind power generation as an example, the specific operation consists of three steps. First, the expert outputs are spliced ​​together. The outputs of the wind power generation task experts and shared experts are spliced ​​along the dimension of the number of experts to form a complete expert output tensor. The purpose is to independently merge the outputs of specific experts and shared experts for the gating network to select. The shape of the spliced ​​output is (batch = 128, time step = 12, experts = 2, feature dimension = 32). The number of experts (2) represents that the outputs of two experts are merged together, but the internal structure of each expert output (batch = 128, time step = 12) remains unchanged. The gating network operates by continuously integrating expert outputs (batch size = 128, time step = 12, feature dimension = 9) into a 3-dimensional tensor (batch size = 128, time step = 12, feature dimension = 2). The first step involves calculating the weights. The gating network uses fully connected layers to calculate the weight score for each expert, then uses the Softmax activation function to transform the score into a probability distribution, generating normalized weights. The input is the original input features of the task (batch size = 128, time step = 12, feature dimension = 2), and the output is the expert weight tensor (batch size = 128, time step = 12, expert weight dimension = 2). Finally, a weighted summation is performed. The weights generated by the gating network are multiplied by the concatenated expert output tensor and summed to obtain the final weighted output. Through these steps, the gating network dynamically integrates expert outputs, thereby generating a task-specific final output, which is a 3-dimensional tensor (batch size = 128, time step = 12, feature dimension = 32) that integrates expert outputs.

[0100] The formula for calculating the gating network for task k is as follows:

[0101] g k (x)=w k (x)S k (x);

[0102]

[0103] Where x is the input; Represents the parameter matrix; w k (x) is the function for calculating the weight vector; S k (x) is the selection matrix composed of the outputs of the shared network and the expert network, used for calculating task k:

[0104]

[0105] in, This represents the output of the nth specific expert network for task k. This represents the output of the nth shared expert network.

[0106] The output module generates predicted values ​​for wind and solar power generation tasks. Each task has a dedicated tower network to avoid prediction interference between tasks. The tower network is a two-layer BiGRU network, consisting of two BiGRU layers and a fully connected layer. The input to the first-layer BiGRU network is the final feature representation generated by the gating network (batch size = 128, time step size = 12, feature dimension = 32). The first-layer BiGRU extracts the forward and backward outputs of each time step and concatenates them along the feature dimension (batch size = 128, time step size = 12, feature dimension = 32), then passes them to the next layer BiGRU network. At this point, each time step already contains the forward and backward context information. Although the hidden layer size of the BiGRU network is 16, meaning the feature output dimension of a unidirectional GRU is 16, because it is a bidirectional GRU, the final feature output dimension is still 32 (16*2 = 32). This allows for a smaller hidden layer to reduce computation, while retaining more information through bidirectionality. The second layer of the BiGRU network repeats the operations of the first layer while maintaining the output shape. It then averages the features from 12 time steps along the time-step dimension, compressing the time dimension and aggregating information from multiple time steps to generate a fixed-size feature representation. This reduces computational complexity. The tensor after time-step average pooling becomes 2-dimensional (batch = 128, feature dimension = 32). The fully connected layer maps the features processed by the BiGRU network to the final output space, outputting two quantile predictions. The output of the fully connected layer is a 3-dimensional tensor (batch = 128, number of quantiles = 2, output dimension = 1). The number of quantiles determines how many quantiles the model will output as predictions; here, it is set to 2, representing the lower and upper bounds of the interval. An output dimension of 1 indicates that the prediction for each quantile is 1-dimensional, meaning each tower network performs univariate prediction, predicting wind or solar power. Finally, for each quantile, a quantile loss function is introduced to calculate the quantile loss between the predicted value and the true value. The losses of all quantiles are summed to calculate the total quantile loss for wind power prediction and solar power prediction, respectively. The total loss for the two tasks is calculated by a dynamic weighted summation method. The prediction results of different quantiles are optimized by backpropagation, and the upper and lower boundaries of wind and solar power, i.e., the prediction interval, are finally determined.

[0107] Unlike traditional multi-task prediction models, this model explicitly separates shared experts from task-specific experts to avoid harmful parameter interference. Secondly, by incorporating gating units that include both task-specific shared and expert network processing results, the correlation between the task-specific pyramid network and expert networks for other tasks is eliminated, enabling different types of expert networks to learn different data features. Therefore, this model effectively overcomes the challenges of negative transfer and the "seesaw effect" faced by previous multi-task learning frameworks when dealing with weakly or complexly correlated tasks.

[0108] Specifically, the extraction module includes an expert module and a gating network. The expert module includes a specific task expert module for wind power generation, a specific expert module for photovoltaic power generation, and a shared expert module.

[0109] The shared expert module extracts common feature information of wind and solar power generation tasks, while the specific expert module extracts unique feature information of wind and solar power generation tasks.

[0110] The gating network is used to dynamically select and weight the outputs of different experts based on the specific input of the task.

[0111] Specifically, during training, the total quantile loss for wind power prediction and photovoltaic (PV) prediction, along with the weighted sum of the two task losses, is first calculated via forward propagation. During backpropagation, the wind power task expert receives the backward gradient of the wind power loss and updates its parameters through the optimizer, focusing more on the unique characteristics of the wind power task. Specifically, a stronger gradient signal for a feature (such as wind speed) indicates a greater impact on the loss, and the expert will pay more attention to it when updating parameters. For example, the wind power expert's gradient might strengthen the extraction of wind speed features while ignoring irradiance. The PV task expert receives the backward gradient of the PV loss and focuses on the unique characteristics of the PV task when updating parameters. The shared expert receives the backward gradient of the total loss from both tasks and learns features useful for both tasks.

[0112] The gating network is used to dynamically select and weight the outputs of different experts based on the specific input of the task. Specifically, the gating network for wind power generation tasks receives the gradient from the wind power task loss, updates the parameters through the optimizer, and adjusts the weight allocation of experts and shared experts in the wind power generation task to make the contribution of each expert more reasonable. The gating network for photovoltaic power generation tasks works similarly.

[0113] Through forward propagation, the extraction module utilizes an expert network to extract wind, solar, and shared features, and then fuses these features through a gating network. Backpropagation calculates gradients using a loss function, and the optimizer updates parameters, thereby progressively improving the feature extraction and fusion capabilities. In this way, the extraction module can effectively separate and share features between wind and solar power generation, while preserving shared information to improve the overall performance of the model.

[0114] Specifically, the expert module uses a TCN-Attention expert network to process the input time-series data;

[0115] The TCN-Attention expert network uses TCN for local feature capture and a self-attention mechanism for global feature capture.

[0116] The embodiments of this invention can better improve the performance of time series modeling. Temporal Convolutional Networks (TCNs) are neural networks that integrate dilated convolutions, dilated causal convolutions (DCCs), and residual connections (RCs), and can be used for time series prediction. The network architecture is as follows: Figure 5 As shown.

[0117] Causal convolution cannot see future data; it can only output the current value of the current layer based on the current and previous values ​​of the lower layer. It is a unidirectional structure with causal relationships and is a strictly time-constrained model. Using causal convolution will not cause data leakage.

[0118] However, causal convolution still suffers from the problems of traditional neural network models, namely, it is limited by the size of the convolution kernel and has difficulty capturing features over longer time scales. Therefore, TCN introduces dilated convolution to sample the input intervals during convolution to obtain a larger receptive field. This allows for a broader capture of contextual information while reducing the number of parameters and computational complexity. Figure 6 This is a diagram of the two-layer dilated causal convolutional structure of the TCN network of this invention. The kernel size is 3, and the number of channels in each hidden layer is 64. The dilated convolution operation formula is as follows:

[0119]

[0120] Where F(t) represents the value of X. t The convolution result at point d; d is the dilation coefficient; k is the kernel size; X t-dv This indicates that convolution is performed on the input time-series data; f represents the TCN filter coefficients.

[0121] TCN uses residual connections to transmit information across layers, effectively solving the gradient vanishing and gradient exploding problems in deep network training, thus improving the training effect and performance of the network. TCN can process large-scale data in parallel. By increasing the number of layers and changing the dilation coefficient and filter size, its receptive field can be expanded, allowing for more flexible extraction of historical information over time. Furthermore, it avoids the gradient vanishing and gradient exploding problems in traditional RNN networks to some extent. Therefore, this embodiment of the invention uses TCN to extract local time-series features of wind and solar power generation characteristics. Through multiple convolutional and dilated convolutional operations, complex patterns in the time series are captured.

[0122] The computation process of the TCN residual block is as follows: First, the input feature vector is subjected to dilated causal convolution. Then, weight normalization is used to accelerate network training. Next, the ReLU activation function is used for non-linear calculation, and Dropout is added to prevent overfitting. Finally, the result is summed with the input to obtain the output vector. The calculation process is as follows:

[0123] S i =Conv(W i ×F j +b i );

[0124] {S0,S1,…,S t-1 ,S t} = WeightNorm({S0,S1,…,S) t-1 ,S t});

[0125] {S0,S1,…,S t-1 ,S t} = ReLU({S0,S1,…,S…) t-1 ,S t});

[0126] Among them, S i W represents the feature vector obtained by convolution at time i. i F represents the weight matrix calculated during convolution at time i. j Let b represent the convolution kernel of the j-th layer. i Represents the bias vector; ||w x || represents the magnitude of the weight w in ReLU(x) = max(0,x). Represents a unit vector in the same direction as w; {S0,S1,…,S t-1 ,S t} represents the feature map obtained after the j-th complete convolution.

[0127] Specifically, in traditional TCNs, the network primarily captures long-term dependencies in a sequence by stacking convolutional layers. However, due to the locality of convolution operations, the network may not effectively capture the relationships between different positions in the sequence. To better capture important temporal information from time series data, a self-attention layer is added after the TCN residual block to capture global dependencies in the time series data. The self-attention mechanism weights the output based on the input information, enabling the network to better capture the correlations between various positions in the sequence and aggregate information at different time steps. It can ignore the distance between vectors and perform global feature extraction on the time series, thereby improving the network's performance and generalization ability.

[0128] The formula for calculating attention mechanism is as follows:

[0129]

[0130] First, calculate the inner product of each row vector of the query matrix Q and the key matrix K. To prevent the inner product from becoming too large, divide by the dimension d. k First, the square root of the product is obtained; second, the result of the inner product is normalized using the Softmax function; finally, the Softmax matrix is ​​multiplied by the value V to obtain the final output.

[0131] Specifically, the output module adopts a QR-BiGRU network; in this embodiment of the invention, a bidirectional gated recurrent unit (QR-BiGRU) neural network is selected as the tower network, wherein the BiGRU neural network achieves more accurate time series prediction, while quantile regression (QR) achieves the acquisition of the confidence interval of the prediction result.

[0132] Specifically, this embodiment will perform power time series prediction based on historical meteorological and power time series data. LSTM neural networks are the most classic time series prediction method; however, due to their complex "gate" structure, they suffer from problems such as a large number of training parameters, low training efficiency, and susceptibility to overfitting. To address this issue, researchers have proposed a simpler gating mechanism, GRU. Compared to LSTM, GRU reduces the number of parameters by one-third, thereby accelerating the iteration process, reducing training time, lowering the probability of overfitting, and improving prediction efficiency.

[0133] Time series prediction is performed using a BiGRU neural network, and confidence intervals are obtained using the QR algorithm.

[0134] Specifically, such as Figure 7 As shown, the BiGRU network consists of two GRUs in two directions; the forward GRU makes predictions based on time series data from front to back, while the backward GRU makes predictions based on time series data from back to front.

[0135] Simultaneously process the input sequences of the forward GRU and the backward GRU to obtain the forward and backward information of the sequences, and concatenate the outputs of the forward GRU and the backward GRU to obtain the concatenated result of the forward and backward hidden states.

[0136] Specifically, BiGRU consists of two GRUs in two directions. The forward GRU predicts from front to back based on time series data, while the backward GRU predicts from back to front based on time series data. By processing both forward and backward input sequences simultaneously, it can effectively acquire forward and backward information of the sequence. By concatenating the outputs of the forward and backward GRU units, BiGRU can integrate contextual information at each time step, thereby improving the model's ability to model complex temporal features. Its working mechanism can be represented as follows:

[0137]

[0138]

[0139] In the formula, This represents the hidden state of the forward GRU at time step t. This represents the hidden state of the reverse GRU at time step t, x. t h is the input sequence at time step t. t This is the result of splicing the forward and reverse hidden states.

[0140] Specifically, the QR algorithm can be used to regress the function parameters between the explanatory variable X and the explained variable Y at a certain quantile. The expression for the QR algorithm is as follows:

[0141] Q Y (τ∣X)=β(τ)X;

[0142] In the formula: Q Y (τ|X) represents Y in X=(x1,x2,…,x) k ) T The conditional quantile under the given condition, β(τ)=(β0(τ),β1(τ),β2(τ),…,β k (τ)) is the parameter vector, and τ is the quantile;

[0143] The parameter vector β(τ) is estimated using the following formula:

[0144]

[0145] In the formula, i is the sample index and N is the number of samples.

[0146] Specifically, quantile regression (QR) considers the probability distribution of a sample and obtains the regression results at each quantile under that distribution. In particular, QR can regress the function parameters between the explanatory variable X and the explained variable Y at a certain quantile; thus, when X takes the value x*, it can obtain the value of y* at that quantile. Assuming the error X follows a Gaussian distribution, if the values ​​of Y at the 0.05 and 0.95 quantiles are obtained based on QR, then a 90% confidence interval for Y is considered to have been obtained.

[0147] Assume the quantile regression model is as follows:

[0148] Q Y (τ∣X)=β(τ)X;

[0149] In the formula: Q Y (τ|X) represents Y in X=(x1,x2,…,x) k ) T The conditional quantile under the given condition, β(τ)=(β0(τ),β1(τ),β2(τ),…,β k (τ)) is a parameter vector, and τ is a quantile.

[0150] Estimating the parameter vector β(τ) can be transformed into the following least squares optimization problem:

[0151]

[0152] In the formula, i is the sample index and N is the number of samples.

[0153] In summary, QR is both a regression on the quantiles of Y and an estimate of the confidence interval of Y. In this embodiment of the invention, the above equation is embedded as a loss function into a BiGRU neural network to obtain a QR-BiGRU neural network, thereby achieving regression prediction at each quantile and obtaining the confidence interval for state prediction.

[0154] In one specific embodiment of the present invention, a regional wind-solar power interval prediction system based on multi-task learning and hybrid neural networks includes:

[0155] The data acquisition module is used to collect meteorological data and historical wind and solar power data of the area under study to build a historical dataset;

[0156] The feature selection module is used to preprocess the historical dataset and use a predictive ability scoring algorithm to select features, thereby obtaining filtered meteorological data and historical power data.

[0157] The model building module is used to build a joint power range prediction model. The joint power range prediction model uses the PLE model as the prediction framework and combines the TCN-Attention expert network, the gating network and the QR-BiGRU tower network for prediction.

[0158] The model training module is used to train the joint power interval prediction model based on the screened meteorological data and historical power data, so as to obtain the trained joint power interval prediction model.

[0159] The prediction module is used to perform joint power generation prediction based on the trained joint power range prediction model.

[0160] In one specific embodiment of the present invention, Golmud City, Qinghai Province, China, was selected as the research area for wind / photovoltaic power generation prediction. This region is rich in solar and wind energy resources. Relevant wind and photovoltaic power generation data and meteorological data are sourced from the Meteonorm and NASA systems. The meteorological data includes 10-meter wind speed (m / s) and total sky surface shortwave direct normal irradiance (Wh / m²). 2 The experiment collected 61,368 meteorological data observations from January 1, 2016 to December 31, 2022, with a time resolution of 1 hour (i.e., 24 observation points per day). The data was divided as follows: data from 2016 to 2021 constituted a training set of 43,848 observations and a validation set of 8,760 observations, while the entire year of 2022 was used as a test set to evaluate prediction accuracy. The data was also divided seasonally and then merged to improve the reliability and persuasiveness of the prediction results.

[0161] Data preprocessing

[0162] This embodiment uses box plots to identify outliers and treat them as missing data, filling in the gaps with data from the previous time step to ensure the continuity of the time series data. Compared to commonly used anomaly detection methods such as the 3-sigma criterion and z-score method, the advantage of box plots is that they do not require the data to follow a normal distribution.

[0163] The data in this embodiment includes two categories: meteorological features and power generation. Different features often have differences in dimensions and magnitudes; without normalization, some indicators may be weakened, thus affecting the accuracy of data analysis. Therefore, a max-min normalization method is used to scale the data to the [0,1] interval:

[0164]

[0165] In the formula, x and x′ represent the original sequence and normalized sequence of the variable, respectively; min(x) represents the minimum value of the sequence, and max(x) represents the maximum value of the sequence.

[0166] To process time series data, a sliding window method is used to preprocess the data after feature selection, converting the original data into a 3D vector form of (batch, time step, feature dimension) for better training. The sliding window size is 12. The experimental goal is to predict the power generation of the next hour based on the data sequence of the previous 12 hours.

[0167] Evaluation indicators

[0168] To determine the superiority and effectiveness of the model in this embodiment, this embodiment selects the normalized root mean square error (NRMSE), normalized mean absolute error (NMAE), and goodness of fit (R-squared, R0). 2 The model in this embodiment is evaluated using five evaluation metrics: three point prediction metrics and two interval prediction metrics: Prediction Interval Coverage Probability (PICP) and Prediction Interval Average Width (PINAW).

[0169] NRMSE is the square root of the normalized mean square error, and its calculation formula is as follows:

[0170]

[0171] NMAE is the mean of the normalized absolute error, and its calculation formula is as follows:

[0172]

[0173] Goodness of fit R 2 R is one of the most commonly used evaluation metrics in regression analysis. Its purpose is to examine the density of sample points near the regression curve, thereby evaluating the goodness of fit. 2 The maximum value of R is 1, and the minimum value is 0. 2 The closer the value is to 1, the better the fitting result. The calculation formula is as follows:

[0174]

[0175] In the formula: The predicted value representing power generation, y i N represents the actual value of power generation, and N represents the total number of samples.

[0176] PICP is mainly used to reflect the probability that the actual value falls within the prediction interval. The higher the coverage probability, the more actual values ​​will fall within the constructed prediction interval, that is, the better the prediction effect. The calculation formula is as follows:

[0177]

[0178] PINAW is defined as the width between the upper and lower bounds of the prediction interval. At the same confidence level, the smaller the value, the narrower the interval, and the more predictive information it contains. The calculation formula is as follows:

[0179]

[0180] In the formula: B n This represents a Boolean value, which is set to 1 if the actual value falls within the specified range, and 0 otherwise; E represents the difference between the maximum and minimum actual values; P represents the difference between the maximum and minimum actual values. U and P D This indicates the upper and lower bounds of the prediction interval.

[0181] loss function

[0182] The effectiveness of a multi-task learning model highly depends on the weight allocation strategy among the loss functions of different tasks. This embodiment uses an uncertainty weighting method to automatically balance (optimize) the multi-task loss, and the weights are dynamically adjusted according to the learning stage of the prediction task.

[0183]

[0184] Where M is the number of training tasks, w m and L m Let w represent the weight matrix and loss function for the m-th task, respectively. m It adaptively adjusts during the training phase to address the issues of imbalanced loss magnitude and differences in convergence speed in multi-task learning.

[0185] Single task loss L m The quantile loss function described above is suitable for regression tasks. After the total loss is calculated, the Adam optimizer can be used to update the model parameters.

[0186] Feature selection

[0187] The meteorological data used in this embodiment were extracted from the Meteonorm and NASA systems, and included 7 meteorological features for wind power generation, 9 meteorological features for photovoltaic power generation, and 2 common features, namely temperature (°C) and relative humidity (%). The specific meteorological features for wind and solar power are shown in Table 1.

[0188] Table 1. Symbols and meanings of specific meteorological features of wind and light

[0189]

[0190]

[0191] To improve efficiency and conserve resources, it is crucial to identify key meteorological features that affect regional power generation. Figure 8 The graph displays the predictive power score between normalized photovoltaic meteorological characteristics and photovoltaic power generation. Each element in the graph represents the predictive power (PPS score) of the horizontal axis relative to the vertical axis. As can be seen from the graph, the total sky surface shortwave direct irradiance (Wh / m²)... 2 The ability to predict photoelectric power is the strongest, followed by the total photosynthetically active irradiance (W / m²) of the entire sky surface. 2 ) and total sky surface ultraviolet irradiance (W / m 2 The predictive power of photovoltaic features is above 0.7, while the predictive power of other features is weak. Figure 9 The graph displays the predictive power score between normalized wind meteorological features and wind power generation. For wind direction features, the sine function is used to transform the wind direction angle value. As shown in the graph, the 50-meter wind speed (m / s) and the wind turbine hub wind speed (m / s) have significantly greater predictive power for wind power than other features, and are strongly correlated with wind power.

[0192] In this embodiment, the best and most efficient results were achieved by selecting wind and solar meteorological features with a prediction capability score of 0.7 or higher, along with two common features: temperature (°C) and relative humidity (%). These features were: 50-meter wind speed (m / s), wind turbine hub wind speed (m / s), and total sky surface shortwave direct irradiance (Wh / m²). 2 Total effective photosynthetic irradiance of the entire sky surface (W / m²) 2 ) and total sky surface ultraviolet irradiance (W / m 2 The forecast uses seven features, including temperature (°C), relative humidity (%), and so on. In addition, historical wind and solar power data are also used as time-series features in the forecast.

[0193] Analysis of Experimental Results

[0194] This experiment was conducted on an Ubuntu server using an NVIDIA GeForce RTX 3090 GPU. It was implemented in Python, and the power generation prediction algorithm was designed and trained using the Pythorch deep learning framework. The parameters of the experimental model were as follows: 100 epochs, 128 batch sizes, a window size of 12, a learning rate of 0.0001, an Adam optimization model, Dropout regularization of 0.2, and an input feature dimension of 9 (7 meteorological features + 2 historical power features). Each task had one dedicated expert and one shared expert. The network component parameters of the model in this embodiment and the comparison model are shown in Table 2.

[0195] Table 2 Network Component Parameters

[0196]

[0197] Verify the effectiveness of the BiGRU tower network

[0198] In this embodiment, the BiGRU network is selected as the tower network of the PLE multi-task learning framework. In order to verify the prediction performance of BiGRU, it is compared with three powerful and widely used prediction networks: CNN, LSTM and GRU.

[0199] In this experiment, to compare the performance of the four networks, a baseline expert network with two layers of Multilayer Perceptron (MLP) as the extraction module was used. The four networks were then used as tower networks for joint wind and solar power prediction. Each layer of the MLP expert network contained 64 neurons. The power prediction results for the four seasons are shown in Tables 3-7.

[0200] Table 3 Spring Wind and Solar Power Forecast Results

[0201]

[0202] Table 4 Summer Wind and Solar Power Forecast Results

[0203]

[0204] Table 5 Autumn Wind and Solar Power Forecast Results

[0205]

[0206] Table 6. Winter Wind and Solar Power Forecast Results

[0207]

[0208] Table 7 Annual Average

[0209]

[0210]

[0211] Experimental results show that BiGRU, as a tower network, exhibits significant advantages in joint wind and solar forecasting. Looking at the annual average indicators, BiGRU's wind power forecast NMAE (4.23 × 10⁻⁶) is significantly higher. -2 Compared to the suboptimal model GRU, the photoelectric prediction R is reduced by 33%. 2 (96.17%), an increase of 2.13 percentage points. Its two-way structure effectively captures the lag effects of wind and solar power generation (such as the cumulative effect of preceding wind speeds on winter wind power) by integrating forward and backward time-series characteristics. Notably, BiGRU exhibits less seasonal fluctuation in photovoltaic forecasting (summer NMAE is only 3.36 × 10⁻⁶). -2 This demonstrates its robustness to sudden changes in light intensity. In contrast, LSTM is prone to overfitting when the amount of data is insufficient due to parameter redundancy (Autumn Optoelectronics R). 2 Only 93.10% (while the local perception characteristics of CNNs make it difficult to model long-range meteorological correlations). Experimental results verify the advantages of BiGRU in balancing bidirectional feature extraction capability and parameter efficiency as a tower network in the PLE framework.

[0212] Validating the effectiveness of TCN-Attention expert network

[0213] This embodiment aims to verify the effectiveness of the improved extraction module in capturing shared information from wind and solar power generation and enhancing prediction performance. To this end, the BiGRU network, which demonstrated the best performance in the previous section, is used as the tower network. The TCN-Attention network used in this embodiment is used as the expert network, and compared with the baseline network, TCN network, and TCN-Channel network from the previous section. The TCN-Channel network combines the TCN network with a channel attention mechanism, a common attention mechanism used in CNNs and other networks with feature channels. Unlike the self-attention mechanism used in this embodiment, it focuses on information from different channels. Furthermore, without using the PLE multi-task learning framework, the BiGRU network is used alone to predict power for wind and solar power to verify the effectiveness of the PLE multi-task learning framework, as shown in Tables 8-12.

[0214] Table 8 Spring Wind and Solar Power Forecast Results

[0215]

[0216] Table 9 Summer Wind and Solar Power Forecast Results

[0217]

[0218]

[0219] Table 10 Autumn Wind and Solar Power Forecast Results

[0220]

[0221] Table 11 Winter Wind and Solar Power Forecast Results

[0222]

[0223]

[0224] Table 12 Annual Average Wind and Solar Power Forecast Results

[0225]

[0226] The introduction of the TCN-Attention expert network significantly improved feature extraction capabilities. Year-round data showed that the TCN-Attention-BiGRU reduced wind / solar NMAE by 21.8% and 29.1% respectively compared to the baseline MLP expert network, demonstrating the effectiveness of its spatiotemporal feature fusion. The dilated convolution of the TCN successfully captured the multi-scale periodic features of meteorological factors (such as 12-hour wind speed fluctuations), while the self-attention mechanism highlighted key time points through dynamic weighting (e.g., greater sensitivity to sudden changes in irradiance caused by cloud cover). Comparative experiments showed that channel attention performed poorly in wind power prediction (winter NMAE 2.79×10⁻⁶). -2 This is because it overemphasizes feature channels and ignores temporal dependencies. Furthermore, the PLE multi-task framework reduced the wind and solar prediction error (NMAE) by 45.2% and 47.0% respectively compared to the single-task model, validating the collaborative optimization mechanism of shared features and task-specific features, which can improve the prediction accuracy of both tasks.

[0227] Interval forecasting analysis

[0228] To address the uncertainty in quantification point prediction, a QR-TCN-Attention-BiGRU interval prediction model based on quantile regression was constructed on top of the TCN-Attention-BiGRU model. Using 90% and 80% confidence intervals, interval prediction methods combining QR with TCN-BiGRU, TCN-Channel-BiGRU, TCN-Attention-CNN, TCN-Attention-LSTM, and TCN-Attention-GRU were used as comparison models. The prediction results of each model in different seasons are shown in Tables 13-17.

[0229] Table 13 Spring Wind and Solar Power Range Forecast Results

[0230]

[0231]

[0232] Table 14 Prediction Results of Summer Wind and Solar Power Ranges

[0233]

[0234] Table 15 Autumn Wind and Solar Power Range Forecast Results

[0235]

[0236] Table 16 Prediction Results of Wind and Solar Power Ranges in Winter

[0237]

[0238]

[0239] Table 17 Annual Average Wind and Solar Power Range Forecast Results

[0240]

[0241] As shown in Table 17, the annual average forecast results indicate that, in both wind power and solar power forecasting tasks, the model in this embodiment exhibits the highest interval coverage at both the 80% and 90% confidence levels. Furthermore, the model in this embodiment has a lower PINAW value than the comparison model, indicating that its prediction interval is narrower, achieving higher interval coverage within a narrower prediction interval and demonstrating a more accurate fit.

[0242] The prediction results of each model in different seasons are shown in Tables 13-16. The model in this embodiment exhibits the highest PICP and the lowest PINAW at different confidence levels in wind and solar power predictions during spring, summer, and winter. In autumn, the model in this embodiment does not perform best across all indicators. In wind power prediction at the 80% confidence level, the model in this embodiment has a PINAW 0.022 higher than the QR-TCN-Attention-GRU model, but its PICP is 0.215 higher, resulting in a significant improvement in interval coverage. Therefore, when the prediction intervals are similar, the model in this embodiment is more accurate. In photovoltaic power prediction at the 90% confidence level, the model in this embodiment has a PINAW 0.014 higher than the QR-TCN-Attention-LSTM model, and its PICP is 0.051 higher, indicating that the prediction performance of the two models is not significantly different. Overall, the model in this embodiment achieves good prediction results in all four seasons.

[0243] This invention establishes a wind and solar energy cluster, effectively coupling wind and solar energy and fully utilizing local power characteristics and climate complementarity. Specifically, when wind farms and solar power plants are located in the same area, by exploring the complex correlation between these two energy sources and combining wind and solar power data, prediction accuracy can be greatly improved. In summary, exploring the spatiotemporal correlation and complementarity between wind and solar energy and constructing a joint power range prediction model to improve power prediction accuracy is of great significance. Furthermore, the joint prediction model only requires one model to obtain multiple types of prediction results, which is easier to test and maintain compared to building a separate model for each energy source.

[0244] To more intuitively demonstrate the performance of the embodiments of the present invention in addressing the challenges of the volatility and intermittency of wind and photovoltaic power generation, while ensuring interval coverage, the wind and solar interval prediction curves of four models—QR-TCN-BiGRU, QR-TCN-Attention-LSTM, QR-TCN-Channel-BiGRU, and QR-TCN-Attention-BiGRU—were plotted in four seasons under an 80% confidence interval to verify the predictive effect of the models when meteorological conditions fluctuate drastically.

[0245] While ensuring coverage, the embodiments of the present invention achieve the narrowest interval width in both wind and photovoltaic power generation interval prediction. Specifically, when the actual power is stable, the interval widths of the four models show little difference, indicating that all four models perform well under stable meteorological conditions. When the actual power fluctuates significantly, i.e., when meteorological conditions change abruptly, the embodiments of the present invention still achieve a relatively narrow interval width, while other models require wider interval widths to achieve high coverage, demonstrating that the embodiments of the present invention can achieve more accurate and reliable interval estimation. Furthermore, the QR-TCN-BiGRU and QR-TCN-Channel-BiGRU models exhibit a significant lag between their predicted interval trends and the actual power when power fluctuations are large, but the QR-TCN-Attention-LSTM and QR-TCN-Attention-BiGRU models largely do not exhibit this phenomenon, indicating that the combination of TCN networks and self-attention mechanisms can effectively capture time-series patterns.

[0246] Traditional wind and solar power forecasting methods primarily focus on single-energy predictions, rarely considering the interrelationships and influences between wind and solar power variables within the same region. Furthermore, both wind and solar power exhibit volatility and intermittency. This invention proposes a regional joint wind and solar power forecasting method based on a multi-task learning framework and QR-BiGRU, applicable to power generation systems in the same region where wind and solar energy are correlated. First, a forecasting capability score is used to screen highly correlated meteorological features. Then, a TCN-Attention network is used as the expert network within a PLE multi-task learning framework to extract coupling information between wind and solar power features. Validation of the extracted network shows that this embodiment achieves the best evaluation index, demonstrating its effectiveness in feature extraction. Finally, a BiGRU network with a quantile regression loss function is introduced as the tower network for wind and solar power joint generation interval estimation. Experimental results show that this embodiment maintains high interval coverage while keeping the interval width relatively narrow, indicating that the method can obtain more accurate and reliable interval estimates even under complex weather conditions. Through example analysis and comparison with conventional forecasting models, the accuracy of the proposed forecasting model is significantly improved, demonstrating high value in practical engineering applications.

[0247] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0248] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in these embodiments may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A regional wind-solar power joint prediction method based on multi-task learning and hybrid neural networks, characterized in that, include: Collect meteorological data and historical wind and solar power data of the area to be studied, and construct a historical dataset; The historical dataset is preprocessed, and a predictive ability scoring algorithm is used for feature selection to obtain filtered meteorological data and historical power data. A joint power range prediction model is constructed. The joint power range prediction model uses the PLE model as the prediction framework and combines the TCN-Attention expert network, the gating network and the QR-BiGRU tower network for prediction. The joint power interval prediction model is trained based on the screened meteorological data and historical power data to obtain the trained joint power interval prediction model. Joint power generation prediction is performed based on a trained joint power range prediction model.

2. The regional wind and solar power joint prediction method based on multi-task learning and hybrid neural networks according to claim 1, characterized in that, The joint power range prediction model includes an input module, an extraction module, and an output module. The input module receives historical power data and meteorological data and processes the data using a sliding window. The extraction module extracts specific and shared features of wind and solar power generation tasks. The output module outputs the predicted value for each task.

3. The regional wind and solar power joint prediction method based on multi-task learning and hybrid neural networks according to claim 2, characterized in that, The extraction module includes an expert module and a gating network. The expert module includes a specific task expert module for wind power generation, a specific expert module for photovoltaic power generation, and a shared expert module. The shared expert module extracts common feature information of wind and solar power generation tasks, while the specific expert module extracts unique feature information of wind and solar power generation tasks. The gating network is used to dynamically select and weight the outputs of different experts based on the specific input of the task.

4. The regional wind-solar power interval prediction method based on multi-task learning and hybrid neural networks according to claim 3, characterized in that, The expert module uses a TCN-Attention expert network to process the input time-series data. The TCN-Attention expert network uses TCN for local feature capture and a self-attention mechanism for global feature capture.

5. The regional wind and solar power joint prediction method based on multi-task learning and hybrid neural networks according to claim 2, characterized in that, The output module uses a QR-BiGRU network; Time series prediction is performed using a BiGRU neural network, and confidence intervals are obtained using the QR algorithm.

6. The regional wind-solar power interval prediction method based on multi-task learning and hybrid neural networks according to claim 5, characterized in that, The BiGRU network consists of two GRUs in two directions: the forward GRU predicts data from front to back based on time series data, and the backward GRU predicts data from back to front based on time series data. Simultaneously process the input sequences of the forward GRU and the backward GRU to obtain the forward and backward information of the sequences, and concatenate the outputs of the forward GRU and the backward GRU to obtain the concatenated result of the forward and backward hidden states.

7. The regional wind-solar power interval prediction method based on multi-task learning and hybrid neural networks according to claim 5, characterized in that, The QR algorithm can be used to regress the function parameters between the explanatory variable X and the explained variable Y at a certain quantile. The expression of the QR algorithm is as follows: Q Y (τ∣X)=β(τ)X; In the formula: Q Y (τ|X) represents Y in X=(x1,x2,…,x) k ) T The conditional quantile under the given condition, β(τ)=(β0(τ),β1(τ),β2(τ),...,β k (τ)) is the parameter vector, and τ is the quantile; The parameter vector β(τ) is estimated using the following formula: In the formula, i is the sample index and N is the number of samples.

8. A regional wind-solar power interval prediction system based on multi-task learning and hybrid neural networks, characterized in that, include: The data acquisition module is used to collect meteorological data and historical wind and solar power data of the area under study to build a historical dataset; The feature selection module is used to preprocess the historical dataset and use a predictive ability scoring algorithm to select features, thereby obtaining filtered meteorological data and historical power data. The model building module is used to build a joint power range prediction model. The joint power range prediction model uses the PLE model as the prediction framework and combines the TCN-Attention expert network, the gating network and the QR-BiGRU tower network for prediction. The model training module is used to train the joint power interval prediction model based on the screened meteorological data and historical power data, so as to obtain the trained joint power interval prediction model. The prediction module is used to perform joint power generation prediction based on the trained joint power range prediction model.

Images