Medium-and-long-term wind power probabilistic prediction method based on spatio-temporal feature fusion
By integrating spatiotemporal features and using probabilistic prediction, the problems of insufficient spatiotemporal feature mining and uncertainty representation in medium- and long-term wind power forecasting have been solved, achieving high-precision wind power forecasting and supporting the safe and efficient operation of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-03-31
AI Technical Summary
Existing wind power prediction methods are insufficient in mining spatiotemporal features on medium- and long-term scales, have weak uncertainty characterization capabilities, and lack long-term generalization performance, making it difficult to achieve reliable medium- and long-term wind power prediction.
A probabilistic prediction method for medium- and long-term wind power is adopted by fusion of spatiotemporal features. By acquiring multi-source data of wind farms and encoding temporal and spatial features, and using learnable spatial location coding and adaptive weighted fusion strategy, combined with a probabilistic prediction module and a decoder, the probabilistic prediction results of wind power are output, and interpretability analysis is performed through a surrogate model.
It significantly improves the accuracy and robustness of medium- and long-term wind power forecasting, can quantify forecast uncertainty, support grid risk dispatch and reserve capacity optimization decisions, and enhance the reliability and practicality of forecasting.
Smart Images
Figure CN121769826A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of wind power prediction, and in particular to a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion. Background Technology
[0002] Currently, wind power prediction is mainly divided into three categories: physical model methods, statistical model methods, and deep learning methods. Physical model methods, based on numerical weather prediction and atmospheric dynamic equations, combined with wind turbine power curves, offer physical interpretability and are suitable for medium- and long-term forecasts. However, they are sensitive to initial conditions, have high computational costs, and exhibit significant simulation biases in complex terrain, limiting prediction accuracy. Statistical model methods rely on historical data for modeling. They are simple and computationally efficient, performing well in short-term forecasts, but struggle to capture nonlinear dynamic characteristics in medium- and long-term forecasts, resulting in severe error accumulation and insufficient generalization ability.
[0003] In recent years, deep learning models have been widely used in wind power forecasting, effectively extracting temporal features and outperforming traditional methods in point prediction. However, most models focus only on single-point time series, neglecting the spatial correlation between wind farms and the spatiotemporal coupling characteristics of wind propagation. Furthermore, most models provide deterministic outputs, lacking probabilistic representations of prediction uncertainties, making it difficult to support risk decision-making. In addition, problems such as insufficient spatiotemporal information fusion and error accumulation in medium- and long-term forecasts remain unresolved. It is worth noting that achieving reliable medium- and long-term wind power forecasting requires a systematic understanding of long-term meteorological evolution patterns and the response characteristics of renewable energy generation. On the one hand, it is necessary to fully utilize decades of historical meteorological observation data, such as wind speed and irradiance, to identify typical climate modes and their impact mechanisms on regional wind power output. On the other hand, it is necessary to construct a nonlinear mapping relationship between meteorological elements covering a regional scale and renewable energy generation, breaking through the limitations of single-site modeling and fully exploring the spatial synergy and propagation characteristics of wind farm clusters. Summary of the Invention
[0004] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, addressing the issues of insufficient spatiotemporal feature mining, weak uncertainty characterization ability, and inadequate long-term generalization performance in existing wind power prediction methods at medium- and long-term scales.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for probabilistic prediction of medium- and long-term wind power by spatiotemporal feature fusion, comprising: acquiring multi-source data of a wind farm, and inputting the multi-source data into a spatiotemporal feature fusion module for... Power temporal coding and terrain spatial coding are used to extract temporal and spatial features. The time and spatial features are weighted and fused to generate a fused feature, and the fused feature is input into the probabilistic prediction module; The probabilistic prediction module extracts the fused features through an encoder to form a connection feature map, and inputs the connection feature map and a target sequence containing a start marker into the decoder for decoding to obtain the probabilistic prediction result of wind power. The prediction results are analyzed for interpretability using the results interpretability analysis module.
[0006] As a preferred embodiment of the probabilistic prediction method for medium- and long-term wind power by spatiotemporal feature fusion described in this invention, the step of performing interpretability analysis on the prediction results through the result interpretability analysis module includes: constructing a dataset for training a surrogate model using the prediction results and input features loaded from the original multi-source data. The proxy model is trained by taking the original input features as input and the prediction results as output. Based on the trained proxy model, the contribution value of each input feature is calculated and the feature contribution explanation result is output.
[0007] As a preferred embodiment of the spatiotemporal feature fusion method for medium- and long-term wind power probabilistic prediction described in this invention, the multi-source data includes terrain data, meteorological data, and historical power data. The multi-source data is input into the spatiotemporal feature fusion module for power time encoding and terrain spatial encoding. Before extracting time features and spatial features, the method further includes filling in the missing power values of the target wind turbine based on the power and wind direction data of the target wind turbine and its adjacent wind turbines in the time series. For each neighboring wind turbine, calculate the minimum circumferential angle between the wind direction of the neighboring wind turbine and the wind direction of the target wind turbine, and map the minimum circumferential angle to the interval [0,1] to obtain the wind direction weight factor; when the wind directions of the two wind turbines are completely the same or opposite, the wind direction weight factor is 1; when the wind directions are perpendicular, the wind direction weight factor is 0. Calculate the distance weighting factor based on the physical distance between the neighboring wind turbine and the target wind turbine; The initial comprehensive weight of each neighboring wind turbine is obtained by multiplying the wind direction weight and the distance weight. The initial weights of all neighboring wind turbines are then normalized to obtain the final dynamic weight.
[0008] As a preferred embodiment of the spatiotemporal feature fusion-based probabilistic prediction method for medium- and long-term wind power described in this invention, it further includes: The power values of all neighboring wind turbines at the missing time are weighted and summed using normalized dynamic weights to obtain the spatiotemporal state value at the missing time. Within the historical time window centered on the missing time, the historical spatiotemporal state value is calculated for each historical time in the window using the same dynamic weights. The absolute difference between each historical spatiotemporal state value and the spatiotemporal state value at the current missing time is also calculated. All historical moments are sorted in ascending order based on the absolute difference, and the top K historical moments with the smallest difference are selected. The power values of the corresponding wind turbines at those historical moments are averaged as the power missing value of the target wind turbine. The meteorological data and wind power data with missing values were normalized, and the dataset was divided into training and test sets.
[0009] As a preferred embodiment of the spatiotemporal feature fusion method for probabilistic prediction of medium- and long-term wind power according to the present invention, the multi-source data is input into the spatiotemporal feature fusion module for power temporal encoding and terrain spatial encoding, and the extraction of temporal and spatial features includes: Flatten the terrain data around the wind turbine into a vector. And it is mapped to a spatial location code aligned with the hidden layer dimension of the model through a multilayer perceptron, and the calculation formula is as follows: in, It is a spatial location encoding, and W1 and W2 are weight matrices. and It is a bias vector. It is terrain data; Historical time-series data is concatenated with spatially encoded data to construct an enhanced input sequence. Z And calculate the temporal attention and spatial attention corresponding to the sequence, using the following formula: in, This represents the concatenation of temporal features and spatial encoding along the feature dimension. and Indicates the parameters of the linear projection layer. Indicates the number after projection Enhanced feature representation at each time step, This represents an augmented input sequence consisting of multiple time steps. Indicates time-based attention output. These represent the query, key, and value matrices for temporal attention, respectively. The linear transformation weight matrix representing temporal attention. This indicates spatial attention output. , These represent the query, key, and value matrices for spatial attention, respectively. This represents the dimension of the key vector.
[0010] The advantages of this preferred technical solution are as follows: it introduces learnable spatial location coding to effectively characterize the geographical association of wind turbines; it adopts a spatiotemporal separation attention mechanism to model the long-term dependence in the time dimension and the correlation in the spatial dimension separately, thereby reducing computational complexity and improving feature extraction efficiency; through an adaptive weighted fusion strategy, it dynamically adjusts the contribution weights of spatiotemporal branches to achieve collaborative optimization and deep fusion of key information; this design enhances the model's ability to represent the medium- and long-term wind power evolution process, significantly improves the accuracy and robustness of predictions, and is suitable for wind power sequence modeling tasks on medium- and long-term time scales.
[0011] As a preferred embodiment of the probabilistic prediction method for medium- and long-term wind power by spatiotemporal feature fusion described in this invention, the weighted fusion of the temporal and spatial features to generate fused features includes: dynamically calculating the fusion weight based on the current time step state and global statistical features, and weighting and merging the temporal and spatial attention outputs according to the fusion weight, as shown in the following formula: in, It is an adaptive fusion weight. Represents the mean vector of the sequence. The standard deviation vector of the sequence. This indicates the fusion feature.
[0012] As a preferred embodiment of the medium- and long-term wind power probabilistic prediction method for spatiotemporal feature fusion described in this invention, the probabilistic prediction module performs deep extraction of the fused features through an encoder to form a connection feature map, including: the encoder contains multiple stacked distillation layers and multi-head sparse self-attention layers, and the distillation layers and multi-head sparse self-attention layers are alternately connected. Information is modeled by multiple sparse self-attention layers and feedforward network layers to obtain hidden representation vectors. These hidden representation vectors are then input into the decoder, and the hidden state for future time steps is obtained by combining the temporal features of the decoder input with an attention mechanism.
[0013] As a preferred embodiment of the medium- and long-term wind power probabilistic prediction method based on spatiotemporal feature fusion described in this invention, wherein: the connection feature map and the target sequence containing the start marker are input into the decoder for decoding processing to obtain the wind power probabilistic prediction result includes: the decoder includes a multi-head attention layer and a masked multi-head sparse self-attention layer. The hidden state of the future time step is input into a Gaussian distribution head to predict the mean and logarithmic standard deviation of the future time step, respectively. An exponential transformation is used to ensure that the standard deviation is positive. The calculation formula is as follows: in, It is the decoder hidden state at time step t. , It is a learnable matrix. , It is a bias term. It is the mean. It is the standard deviation.
[0014] As a preferred embodiment of the spatiotemporal feature fusion-based probabilistic prediction method for medium- and long-term wind power described in this invention, the method further includes: during the training process, constructing a Gaussian-distributed negative log-likelihood loss function with the true normalized power as the target, used to optimize the model parameters, and the calculation formula is as follows: in, This is the true normalized data at step t. It is the total loss function.
[0015] The beneficial effects of this preferred technical solution are as follows: it outputs the predicted mean and logarithmic standard deviation of future time steps through a probabilistic decoder; based on the Gaussian distribution assumption and combined with the quantile table of the standard normal distribution, it dynamically calculates the predicted quantile values at different confidence levels, thereby realizing the output of the probability interval of wind power; this mechanism not only provides point estimation, but also quantifies the uncertainty of prediction, supports grid risk dispatch and reserve capacity optimization decisions, and significantly improves the reliability and practicality of medium- and long-term wind power prediction in practical engineering applications.
[0016] As a preferred embodiment of the spatiotemporal feature fusion-based probabilistic prediction method for medium- and long-term wind power described in this invention, the surrogate model is trained by taking the original input features as input and the prediction result as output, including: introducing a surrogate objective function with uncertainty-aware weights, the calculation formula of which is as follows: in, N This represents the total number of training samples. Indicates sample weights, This represents the i-th predicted value of the main model. This indicates the output of the proxy mode. Represents the regularization coefficient. This represents the complexity penalty term. Let represent the prediction variance of the i-th sample.
[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: In the spatiotemporal feature adaptive fusion module, this invention introduces learnable spatial location coding to effectively characterize the geographical associations of wind turbines; it employs a spatiotemporal separation attention mechanism to separately model long-term dependencies in the time dimension and correlations in the spatial dimension, thereby reducing computational complexity while improving feature extraction efficiency; through an adaptive weighted fusion strategy, it dynamically adjusts the contribution weights of spatiotemporal branches to achieve collaborative optimization and deep fusion of key information; this design enhances the model's ability to represent medium- and long-term wind power evolution processes, significantly improving prediction accuracy and robustness, and is suitable for wind power sequence modeling tasks on medium- and long-term time scales.
[0018] This invention, in the Prob-Informer module, outputs the predicted mean and logarithmic standard deviation for future time steps through a probabilistic decoder; based on the Gaussian distribution assumption and combined with the quantile table of the standard normal distribution, it dynamically calculates the predicted quantile values at different confidence levels, realizing the probabilistic interval output of wind power; this mechanism not only provides point estimation but also quantifies prediction uncertainty, supports grid risk dispatch and reserve capacity optimization decisions, and significantly improves the reliability and practicality of medium- and long-term wind power prediction in practical engineering applications.
[0019] In this invention, the Surrogate-SHAP module constructs a training dataset using historical feature matrices and Prob-Informer output. After partitioning, XGBoost is used as a surrogate model, and an objective function with uncertainty-aware weights is introduced to enhance the interpretability and robustness of high-error samples. TreeExplainer is used to analyze the model's decision path, efficiently calculate the SHAP value of each feature, realize attribution analysis of feature variables, and improve the interpretability and engineering credibility of prediction results. Attached Figure Description
[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic diagram of the method flow for a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, according to an embodiment of the present invention. Figure 2This is a diagram illustrating the overall structure of a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, as described in one embodiment of the present invention. Figure 3 This is a prediction result diagram of a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, as described in an embodiment of the present invention. Figure 4 This is a comparative effect diagram of a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, as described in one embodiment of the present invention. Figure 5 This is a comparison chart of the mean square error of a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, as described in one embodiment of the present invention. Detailed Implementation
[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0022] Example 1, referring to Figures 1-2 As one embodiment of the present invention, this embodiment provides a probabilistic prediction method for medium- and long-term wind power based on spatiotemporal feature fusion, comprising: S100: Acquire multi-source data from the wind farm and input the multi-source data into the spatiotemporal feature fusion module for power time coding and terrain spatial coding, and extract time and spatial features; S200: Weighted fusion of temporal and spatial features to generate fused features, and inputting the fused features into the probabilistic prediction module; S300: The probabilistic prediction module extracts the fused features through the encoder to form a connection feature map. The connection feature map and the target sequence containing the start marker are input into the decoder for decoding to obtain the probabilistic prediction result of wind power. S400: Perform interpretability analysis on the prediction results through the result interpretability analysis module.
[0023] It should be noted that existing wind power prediction models only focus on single-point time series, neglecting the spatial correlation between wind farms and the spatiotemporal coupling characteristics of wind propagation. Furthermore, these models are mostly deterministic, lacking a probabilistic expression of prediction uncertainties, making them unsuitable for supporting risk decision-making. In addition, problems such as insufficient spatiotemporal information fusion and error accumulation in medium- and long-term predictions remain unresolved. Existing wind power prediction methods suffer from insufficient spatiotemporal feature mining, weak uncertainty characterization capabilities, and inadequate long-term generalization performance at medium- and long-term scales. To achieve reliable medium- and long-term wind power prediction, a systematic understanding of long-term meteorological evolution patterns and the response characteristics of new energy power generation is necessary. On the one hand, decades of historical meteorological observation data, such as wind speed and irradiance, should be fully utilized to identify typical climate modes (e.g., high-wind years and low-wind years) and their impact mechanisms on regional wind power output. On the other hand, a nonlinear mapping relationship between meteorological elements and new energy power generation covering regional scales (e.g., provincial or grid dispatch zones) needs to be constructed to overcome the limitations of single-site modeling and fully explore the spatial synergy and propagation characteristics of wind farm clusters. Therefore, there is an urgent need to build a new forecasting framework that deeply integrates medium- and long-term meteorological forecast information with historical power generation data, achieves regional-scale collaborative forecasting through spatiotemporal joint modeling, outputs results in a probabilistic form, and supports continuous optimization of the model as data accumulates. This will significantly improve the accuracy, reliability, and practicality of medium- and long-term wind power forecasting, and provide solid support for the safe, economical, and flexible operation of high-proportion renewable energy power systems.
[0024] This invention utilizes a Spatio-Temporal Adaptive Fusion Transformer (STAT) module, introducing spatial location encoding and adaptive weighted fusion spatio-temporal separation attention to obtain the spatio-temporal features of the sequence data. The obtained spatio-temporal feature sequence is input into a Prob-Informer model, outputting the mean and logarithmic standard deviation of wind power prediction. Based on the quantiles of the normal distribution, the quantile prediction values for each future time step are calculated. The Surrogate-SHAP module is used to quantify the contribution of each input element to the wind power prediction, enhancing the model's interpretability. This invention fully utilizes a feature extraction module to extract sequence features and introduces quantile prediction and interpretability analysis, enabling high-precision medium- and long-term wind power prediction, supporting the safe and efficient operation of the power system, promoting intelligent management of the transmission network, and serving the energy security strategy.
[0025] In this embodiment of the invention, the multi-source data in step S100 includes terrain data, meteorological data and historical power data. The multi-source data is input into the spatiotemporal feature fusion module for power time encoding and terrain spatial encoding. Before extracting time features and spatial features, the method further includes: filling in the missing power values of the target wind turbine based on the power and wind direction data of the target wind turbine and its adjacent wind turbines in the time series. In an optional embodiment, in step S100 of the present invention, missing values of power data can be filled by linear interpolation.
[0026] In another optional embodiment, in step S100 of the present invention, missing values of power data can also be filled by the spatiotemporal weighted nearest neighbor filling method.
[0027] In this embodiment of the invention, in step S100, for each neighboring wind turbine, the minimum circumferential angle between the wind direction of the neighboring wind turbine and the wind direction of the target wind turbine is calculated, and the minimum circumferential angle is mapped to the interval [0,1] to obtain the wind direction weight factor; when the wind directions of the two wind turbines are completely consistent or opposite, the wind direction weight factor is 1; when the wind direction is perpendicular, the wind direction weight factor is 0. Calculate the distance weighting factor based on the physical distance between the neighboring wind turbine and the target wind turbine; The initial comprehensive weight of each neighboring wind turbine is obtained by multiplying the wind direction weight and the distance weight. The initial weights of all neighboring wind turbines are then normalized to obtain the final dynamic weight.
[0028] It should be noted that, in the embodiments of the present invention, the range of the minimum included angle is between [0°, 180°].
[0029] Specifically, the minimum included angle of the circle is mapped to a wind direction weighting factor between [0,1]. The minimum circular angle mapping is achieved using an improved cosine function, and the calculation formula is as follows: in, It is the smallest included angle of the circumference. It is the wind direction of the target wind turbine. It's the wind direction of the neighbor's wind turbine. It is a weighting factor.
[0030] Furthermore, based on the physical distance between neighboring wind turbines and the target wind turbine, a distance weighting factor is calculated. The wind direction weight and the distance weight are multiplied to obtain the initial comprehensive weight of each neighboring wind turbine. The initial weights of all neighboring wind turbines are then normalized to obtain the final dynamic weight. The calculation formula is as follows: in, It is a distance weighting factor. It refers to the physical distance between the wind turbines; This is the initial overall weight, and N is the total number of neighboring wind turbines.
[0031] In this embodiment of the invention, step S100 further includes: The power values of all neighboring wind turbines at the missing time are weighted and summed using normalized dynamic weights to obtain the spatiotemporal state value at that missing time. Within the historical time window centered on the missing time, the historical spatiotemporal state value is calculated for each historical time in the window using the same dynamic weights. The absolute difference between each historical spatiotemporal state value and the spatiotemporal state value at the current missing time is also calculated. Sort all historical moments in ascending order based on the absolute difference, select the K historical moments with the smallest difference, and take the average of the power values of the corresponding wind turbines at each historical moment as the missing power value of the target wind turbine. The meteorological data and wind power data with missing values were normalized, and the dataset was divided into training and test sets.
[0032] Specifically, the spatiotemporal state value at the missing moment is represented as follows: in, It is the spatiotemporal state value at the missing moment. It is a moment The corresponding power.
[0033] Specifically, for each historical moment in the window, its historical spatiotemporal state value is calculated using the same dynamic weight; and the absolute difference between each historical spatiotemporal state value and the spatiotemporal state value of the current missing moment is calculated using the following formula: in, It is the spatiotemporal state value of a historical moment within a time window.
[0034] In this embodiment of the invention, step S200 involves inputting multi-source data into a spatiotemporal feature fusion module for power temporal encoding and terrain spatial encoding, extracting temporal and spatial features, including: Flatten the terrain data around the wind turbine into a vector. And it is mapped to a spatial location code aligned with the hidden layer dimension of the model through a multilayer perceptron, and the calculation formula is as follows: in, It is a spatial location encoding, and W1 and W2 are weight matrices. and It is a bias vector. It is terrain data; Historical time-series data is concatenated with spatially encoded data to construct an enhanced input sequence. Z And calculate the temporal attention and spatial attention corresponding to the sequence, using the following formula: in, This represents the concatenation of temporal features and spatial encoding along the feature dimension. and Indicates the parameters of the linear projection layer. Indicates the number after projection Enhanced feature representation at each time step, This represents an augmented input sequence consisting of multiple time steps. Indicates time-based attention output. These represent the query, key, and value matrices for temporal attention, respectively. The linear transformation weight matrix representing temporal attention. This indicates spatial attention output. , These represent the query, key, and value matrices for spatial attention, respectively. This represents the dimension of the key vector.
[0035] It should be noted that this invention introduces learnable spatial location coding into the STAT module to effectively characterize the geographical association of wind turbines; it adopts a spatiotemporal separation attention mechanism to model the long-term dependence in the time dimension and the correlation in the spatial dimension separately, thereby reducing computational complexity while improving feature extraction efficiency; through an adaptive weighted fusion strategy, it dynamically adjusts the contribution weights of spatiotemporal branches to achieve collaborative optimization and deep fusion of key information; this design enhances the model's ability to represent the medium- and long-term wind power evolution process, significantly improves the accuracy and robustness of predictions, and is suitable for wind power sequence modeling tasks on medium- and long-term time scales.
[0036] In this embodiment of the invention, step S200, which involves weighted fusion of temporal and spatial features to generate fused features, includes: dynamically calculating fusion weights based on the current time step's state and global statistical features, and then weighted merging the temporal and spatial attention outputs based on the fusion weights. The calculation formula is as follows: in, It is an adaptive fusion weight. Represents the mean vector of the sequence. The standard deviation vector of the sequence. This indicates the fusion feature.
[0037] In this embodiment of the invention, in step S300, the probabilistic prediction module performs deep extraction of the fused features through the encoder to form a connection feature map, including: the encoder contains multiple stacked distillation layers and multi-head sparse self-attention layers, and the distillation layers and multi-head sparse self-attention layers are alternately connected. Information is modeled by multiple sparse self-attention layers and feedforward network layers to obtain hidden representation vectors. These hidden representation vectors are then input into the decoder, and the hidden state for future time steps is obtained by combining the temporal features of the decoder input with an attention mechanism.
[0038] In this embodiment of the invention, step S300 involves inputting the connection feature map and the target sequence containing the start marker into the decoder for decoding to obtain the wind power probabilistic prediction result. The decoder includes a multi-head attention layer and a masked multi-head sparse self-attention layer. The hidden states at future time steps are input into a Gaussian distribution head to predict the mean and log-standard deviation of future time steps, respectively. An exponential transformation is used to ensure that the standard deviation is positive. The calculation formula is as follows: in, It is the decoder hidden state at time step t. , It is a learnable matrix. , It is a bias term. It is the mean. It is the standard deviation.
[0039] In this embodiment of the invention, step S300 further includes: during the training process, using the true normalized power as the target, constructing a Gaussian-distributed negative log-likelihood loss function to optimize the model parameters, the calculation formula of which is as follows: in, This is the true normalized data at step t. It is the total loss function.
[0040] In this embodiment of the invention, step S400, which involves performing interpretability analysis on the prediction results through the result interpretability analysis module, includes: using the prediction results and input features loaded from the original multi-source data to construct a dataset for training the surrogate model; The surrogate model is trained by taking the original input features as input and the prediction results as output. Based on the trained surrogate model, the contribution value of each input feature is calculated and the feature contribution explanation result is output.
[0041] In an optional embodiment, a feature matrix is loaded from a historical dataset, which includes wind turbine IDs, spatial coordinates, historical wind speed sequences, NWP weather fields, etc., and combined with the power values output by the probabilistic prediction module to construct a surrogate model training dataset. An XGBoost regressor is used as the surrogate model for training. Taking advantage of the tree structure characteristics of the surrogate model, TreeExplainer is used to efficiently calculate the SHAP contribution value of each feature in each sample and output the feature contribution explanation results.
[0042] In another optional embodiment, a feature matrix is loaded from a historical dataset, which includes wind turbine IDs, spatial coordinates, historical wind speed sequences, NWP weather fields, etc., and combined with the power values output by the probabilistic prediction module to construct a surrogate model training dataset. CatBoost is used as the surrogate model for training. Taking advantage of the tree structure characteristics of the surrogate model, TreeExplainer is used to efficiently calculate the SHAP contribution value of each feature in each sample and output the feature contribution explanation results.
[0043] It should be noted that in the Prob-Informer module, this invention outputs the predicted mean and logarithmic standard deviation for future time steps through a probabilistic decoder; based on the Gaussian distribution assumption and combined with the quantile table of the standard normal distribution, it dynamically calculates the predicted quantile values at different confidence levels, thereby realizing the output of the probability interval of wind power; this mechanism not only provides point estimation but also quantifies the uncertainty of prediction, supports grid risk dispatch and reserve capacity optimization decisions, and significantly improves the reliability and practicality of medium- and long-term wind power prediction in practical engineering applications.
[0044] In this embodiment of the invention, step S400, where the proxy model is trained using the original input features as input and the prediction result as output, includes: introducing a proxy objective function with uncertainty-aware weights, calculated as follows: in, N This represents the total number of training samples. Indicates sample weights, This represents the i-th predicted value of the main model. This indicates the output of the proxy mode. Represents the regularization coefficient. This represents the complexity penalty term. Let represent the prediction variance of the i-th sample.
[0045] It should be noted that in the Surrogate-SHAP module of this invention, a training dataset is constructed by using the historical feature matrix and the Prob-Informer output. After partitioning, XGBoost is used as a surrogate model, and an objective function with uncertainty-aware weights is introduced to enhance the interpretability and robustness of high-error samples. TreeExplainer is used to analyze the model decision path, efficiently calculate the SHAP value of each feature, realize the attribution analysis of feature variables, and improve the interpretability and engineering credibility of the prediction results.
[0046] Example 2, refer to Figures 3-5 This is one embodiment of the present invention. This embodiment verifies the beneficial effects of the present invention through testing experiments.
[0047] The experimental data in this embodiment uses two months of recorded data from a wind farm in China. The data includes wind turbine data and historical meteorological data, including wind speed, wind direction, and air pressure. Wind power data is sampled every hour. The data from the first 54 days is used as training data, and the data from the last 6 days is used as test data. Table 1 shows the quantitative analysis results of the method of this invention.
[0048] Table 1 Quantitative Analysis Results
[0049] Experiments were conducted on each model on the test set defined in this invention. R², RMSE, MSE, and MAE were all calculated as the average of 50 experiments. Table 1 shows that the method of this invention significantly outperforms the other comparative models. In terms of MSE, the method of this invention achieves 0.012, which is 0.079 lower than Informer, 0.075 lower than ConvLSTM, and 0.071 lower than CNN_Transformer, demonstrating extremely strong error control capabilities. In terms of R², the method of this invention also leads, reaching 0.95, higher than Informer (0.91), ConvLSTM (0.89), and CNN_Transformer (0.85), indicating that it has a stronger explanatory power for the target variable and a higher correlation between the predicted results and the true values. In terms of RMSE, the method of this invention achieves 0.109, outperforming Informer (0.132), ConvLSTM (0.15), and CNN_Transformer (0.105). While close to CNN_Transformer, it still slightly surpasses it. In terms of MAE, the method of this invention significantly outperforms other models with 0.063, 0.029 lower than Informer, 0.064 lower than ConvLSTM, and 0.026 lower than CNN_Transformer, indicating a clear advantage in mean absolute error control and predictions that more closely approximate the true values. Overall, the method of this invention achieves the best performance across all four core evaluation metrics, demonstrating the most stable and comprehensive superior performance.
[0050] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A spatiotemporal feature fusion medium and long term wind power probabilistic forecasting method, characterized in that, The method comprises the following steps: acquiring multi-source data of a wind farm and inputting the multi-source data into a space-time feature fusion module for power time encoding and terrain space encoding to extract time features and space features; weighting and fusing the time features and the space features to generate fused features, and inputting the fused features into a probabilistic prediction module; the probabilistic prediction module extracts deep features from the fused features through an encoder to form a connection feature map, inputs the connection feature map and a target sequence containing a starting mark into a decoder for decoding processing to obtain a wind power probabilistic prediction result; the result explainability analysis module analyzes the prediction result for explainability.
2. The spatio-temporal feature fusion-based medium and long term wind power probabilistic forecasting method of claim 1, wherein, The result explainability analysis module analyzes the prediction result for explainability, which comprises the following steps: using the prediction result and input features loaded from original multi-source data to construct a data set for training a proxy model; 3. The spatio-temporal feature fusion-based medium and long term wind power probabilistic forecasting method of claim 2, wherein, training the proxy model with original input features as input and the prediction result as output, calculating the contribution value of each input feature based on the trained proxy model, and outputting a feature contribution explanation result. The multi-source data comprises terrain data, meteorological data, and historical power data, and before the power time encoding and the terrain space encoding to extract the time features and the space features, the method further comprises the following steps: based on the power and wind direction data of a target wind turbine and its adjacent wind turbines in a time sequence, filling the power missing values of the target wind turbine; for each neighbor wind turbine, calculating the minimum circumferential angle between the wind direction of the neighbor wind turbine and the wind direction of the target wind turbine, and mapping the minimum circumferential angle to the interval [0, 1] to obtain a wind direction weight factor; when the wind directions of the two wind turbines are completely consistent or opposite, the wind direction weight factor is 1; when the wind directions are perpendicular, the wind direction weight factor is 0; 4.The spatio-temporal feature fusion-based medium and long term wind power probabilistic forecasting method of claim 3, wherein, based on the physical distance between the neighbor wind turbine and the target wind turbine, calculating a distance weight factor; multiplying the wind direction weight and the distance weight to obtain an initial comprehensive weight of each neighbor wind turbine, and normalizing the initial weights of all neighbor wind turbines to obtain a final dynamic weight. The method further comprises the following steps: using the normalized dynamic weight to weight and sum the power values of all neighbor wind turbines at the missing time to obtain a space-time state value at the missing time; in a historical time window centered on the missing time, for each historical time in the window, using the same dynamic weight to calculate its historical space-time state value; and calculating the absolute difference between each historical space-time state value and the space-time state value at the current missing time; 5. The spatio-temporal feature fusion-based medium and long term wind power probabilistic forecasting method of claim 4, wherein, sorting all historical times in ascending order according to the absolute difference, selecting the first K historical times with the smallest difference, and taking the average of the power values of the corresponding wind turbines at the historical times as the power missing value of the target wind turbine; normalizing the meteorological data and the wind power data with filled missing values, and dividing the data set into a training set and a test set. The method for extracting the time features and the space features comprises the following steps: Flattening fan surrounding terrain data into vectors and mapped to spatial position encodings aligned with the model hidden layer dimensions by a multilayer perceptron, computed as follows: wherein, is a spatial position encoding, W1 and W2 are weight matrices, and is a bias vector, is terrain data; The historical time series data is spliced with spatial coding to construct an enhanced input sequence Z And the time attention and spatial attention corresponding to the sequence are calculated, and the calculation formula is as follows: wherein, denotes the concatenation of the temporal feature and the spatial encoding on the feature dimension, and denotes the linear projection layer parameters, denotes the projected enhanced feature representation at the th time step, denotes the enhanced input sequence, consisting of multiple time steps, denotes the temporal attention output, denotes the query, key, value matrices of the temporal attention, respectively, denotes the linear transformation weight matrix of the temporal attention, denotes the spatial attention output, , denotes the query, key, value matrices of the spatial attention, respectively, denotes the dimension of the key vector.
6. The spatio-temporal feature fusion-based medium and long-term wind power probabilistic forecasting method of claim 5, wherein, The weighting and fusing of the time feature and the space feature to generate a fused feature comprises: dynamically calculating a fusion weight according to a state of a current time step and a global statistical feature, and weighting and merging time and space attention outputs according to the fusion weight, and a calculation formula is as follows: wherein, is an adaptive fusion weight, denotes the mean vector of the sequence, denotes the standard deviation vector of the sequence, denotes the fused feature.
7. The spatio-temporal feature fusion-based medium and long term wind power probabilistic forecasting method of claim 6, wherein, The probabilistic prediction module forms a connection feature map by deep extraction of the fused feature through an encoder, and the encoder comprises a plurality of stacked distillation layers and multi-head sparse self-attention layers, and the distillation layers and the multi-head sparse self-attention layers are alternately connected. The information is modeled through a plurality of layers of sparse self-attention layers and a feedforward network layer to obtain a hidden representation vector, the hidden representation vector is input into a decoder, and a hidden state of a future time step is obtained through an attention mechanism in combination with a time sequence feature input into the decoder.
8. The spatio-temporal feature fused mid-long term wind power probabilistic forecasting method of claim 7, wherein, The connection feature map and a target sequence containing a start mark are input into a decoder for decoding processing to obtain a wind power probabilistic prediction result, and the decoder comprises multi-head attention and a mask multi-head sparse self-attention layer. The hidden state of the future time step is input into a Gaussian distribution head to respectively predict a mean value and a logarithmic standard deviation of the future time step, and an exponential transformation is performed to ensure that the standard deviation is positive, and a calculation formula is as follows: wherein, is the decoder hidden state at the t-th time step, , is a learnable matrix, , is a bias term, is a mean, is a standard deviation.
9. The spatio-temporal feature fusion medium and long term wind power probabilistic forecasting method of claim 8, wherein, Further comprising: In the training process, a negative log-likelihood loss function of a Gaussian distribution is constructed with a real normalized power as a target, and is used to optimize model parameters, and a calculation formula is as follows: wherein, is the real normalized data at step t, is the total loss function.
10. The spatio-temporal feature fused medium and long term wind power probabilistic forecasting method of claim 2 or 9, wherein, The proxy model is trained with original input features as inputs and prediction results as outputs, and a calculation formula of a proxy objective function with uncertainty perception weight is introduced as follows: wherein, N denotes the total number of training samples, denotes the sample weight, denotes the i-th prediction value of the master model, denotes the output of the agent model, denotes the regularization coefficient, denotes the complexity penalty term, denotes the i-th sample prediction variance.