New energy power short-term prediction method and system based on multi-source meteorological data fusion

Abstract

The application relates to the field of new energy power prediction, in particular to a new energy power short-term prediction method and system based on multi-source meteorological data fusion. The method comprises the following steps: collecting multi-source time-series meteorological data and generating standard data through preprocessing; constructing a third-order space-time meteorological tensor, generating a multi-dimensional feature matrix based on a non-negative tensor decomposition fusion model and an alternating least squares method iteration solution; inputting the multi-dimensional feature matrix into a joint prediction model based on an Informer architecture, performing space-time patch block segmentation, ProbSparse self-attention mechanism feature extraction, combining a decision-oriented loss function for training and optimization, and outputting a power point prediction value; calculating a confidence interval, determining an early warning level, identifying a climbing event, matching a dispatching suggestion, and integrating to generate a prediction result; and the method can overcome the defects of a shallow feature extraction layer of a traditional fusion method and insufficient long-time sequence processing capacity of a prediction model, and improve the accuracy and practicality of new energy power short-term prediction.

Description

Technical Field

[0001] This application relates to the field of new energy power prediction technology, and in particular to a method and system for short-term prediction of new energy power by fusing multi-source meteorological data. Background Technology

[0002] With the rapid development of new energy power generation technologies, power prediction technology based on multi-source meteorological data fusion has emerged. This technology integrates multi-source data such as ground observation, satellite remote sensing, and numerical weather prediction to compensate for the limited coverage and large errors of single data sources, improve the comprehensiveness and accuracy of meteorological input data, and thus enhance the reliability of new energy power prediction.

[0003] Traditional techniques typically employ weighted averaging or Kalman filtering algorithms to fuse acquired multi-source meteorological data, or use Principal Component Analysis (PCA) to extract features and reduce data dimensionality. For predictive models, Long Short-Term Memory (LSTM) networks or Support Vector Regression (SVR) models are often used to establish a mapping relationship between meteorological elements and power output, aiming to capture the dependency features in the time series.

[0004] However, current multi-source data fusion methods often remain at the level of simple linear weighted concatenation or shallow feature extraction, making it difficult to deeply explore the complex nonlinear complementary information of multi-source meteorological data in the temporal, spatial, and variable dimensions, resulting in limited feature quality after fusion. At the same time, existing prediction models often face the problem of high computational complexity or difficulty in effectively capturing long-distance dependencies when processing long-term meteorological data, limiting further improvement in prediction accuracy. Summary of the Invention

[0005] To address the aforementioned technical issues, this application proposes a method and system for short-term prediction of new energy power based on multi-source meteorological data fusion.

[0006] The technical solution adopted in this application is: a short-term prediction method for new energy power through multi-source meteorological data fusion, comprising the following steps:

[0007] S1. Collect multi-source time-series meteorological data covering the target new energy power station area, preprocess the multi-source time-series meteorological data, and generate standard multi-source meteorological data;

[0008] S2. Perform multimodal deep fusion of standard multi-source meteorological data based on tensor decomposition to obtain a multidimensional feature matrix;

[0009] S3. Input the multidimensional feature matrix into the preset joint prediction model and output the predicted value of new energy power points for future periods; wherein the joint prediction model includes an input layer, a feature extraction layer based on the Informer architecture, a time series modeling layer and a prediction output layer;

[0010] S4. Based on the predicted value of new energy power points in the future period, obtain the new energy power prediction result.

[0011] Furthermore, the multi-source time-series meteorological data includes numerical weather prediction grid data, satellite remote sensing observation data, and ground meteorological station observation data. Preprocessing methods include spatiotemporal alignment, anomaly removal, missing value supplementation, and normalization.

[0012] Furthermore, step S2 specifically includes:

[0013] S21. Map the numerical weather forecast grid data in the standard multi-source meteorological data to the geographic coordinates of the target new energy power station using the bilinear interpolation method to obtain the background meteorological vector sequence at the station scale.

[0014] S22. Based on satellite remote sensing observation data, remote sensing estimates of cloud optical thickness and surface incident shortwave radiation intensity of the target station area are obtained by inversion through a simplified radiative transfer model, and satellite-derived feature vector sequence is generated.

[0015] S23. Constructing a multidimensional spatiotemporal meteorological tensor: Arrange the background meteorological vector sequence, satellite-derived feature vector sequence, and ground meteorological station observation data sequence according to the time dimension, spatial dimension, and meteorological variable dimension to construct a third-order spatiotemporal meteorological tensor. , For the set of real numbers, This is the time step, corresponding to the length of the entire time series data, and should be consistent with the time resolution of the preprocessed data. This refers to the spatial location data, encompassing all ground meteorological station observation points deployed inside and around the target new energy power station, as well as the interpolated core coordinate points of the station. As a meteorological variable dimension, it integrates all core meteorological variables from the three types of data;

[0016] S24. Construct a fusion model based on nonnegative tensor decomposition; define the objective function of the fusion model based on nonnegative tensor decomposition as minimizing the Frobenius norm error between the reconstructed tensor and the original tensor, and add a regularization term constraint. The expression of the objective function is:

[0017] ;

[0018] In the formula, It is a third-order spacetime tensor. For the core tensor, For time pattern matrix, For spatial pattern matrix, For the variable pattern matrix, The regularization coefficient is . Represents the n-modulo product;

[0019] S25. The objective function is solved iteratively by using the alternating least squares method. Two of the factor matrices are fixed and the third factor matrix is ​​updated until convergence, and the converged core tensor is obtained.

[0020] S26. Reconstruct the converged core tensor and the corresponding factor matrix, and flatten the reconstructed tensor data to generate a multidimensional feature matrix.

[0021] Furthermore, step S25 specifically includes:

[0022] S251. Initialize the time pattern matrix, spatial pattern matrix, and variable pattern matrix;

[0023] S252. Fix the spatial pattern matrix and variable pattern matrix, and update the temporal pattern matrix: calculate the Khatri-Rao product of the third-order spatiotemporal meteorological tensor and the core tensor, solve the update temporal pattern matrix using the least squares method, and set the negative elements in the temporal pattern matrix to zero.

[0024] S253. Fixed-time pattern matrix and variable pattern matrix, update spatial pattern matrix: calculate the Khatri-Rao product of the third-order temporal-space meteorological tensor and the core tensor, use the least squares method to solve for the updated spatial pattern matrix, and set the negative elements in the spatial pattern matrix to zero.

[0025] S254. Fixed time pattern matrix and spatial pattern matrix, update variable pattern matrix: Calculate the Khatri-Rao product of the third-order temporal and spatial meteorological tensor and the core tensor, solve for the update variable pattern matrix using the least squares method, and set the negative elements in the variable pattern matrix to zero.

[0026] S255. Repeat steps S252 to S254 until the reconstruction error between two adjacent iterations is less than the preset threshold or the maximum number of iterations is reached, and obtain the converged core tensor.

[0027] Furthermore, the input layer in the joint prediction model is responsible for receiving the multidimensional feature matrix and adjusting it into the input format required by the model.

[0028] The feature extraction layer based on the Informer architecture consists of several stacked Informer encoder blocks. Each Informer encoder block is composed of a sparse self-attention layer, a feedforward neural network, layer normalization, and residual connections. The sparse self-attention layer adopts a probabilistic sampling strategy, which divides the time step into multiple intervals through bucketing. Within each interval, a portion of the time steps are randomly sampled to calculate attention. The number of attention heads and the dropout rate are set. The number of attention heads is used to extract attention features of different dimensions in parallel, and the dropout rate is used to suppress overfitting. The feedforward neural network consists of two fully connected layers with a GELU activation function inserted in between. Layer normalization is used to standardize the input features of each layer.

[0029] The temporal modeling layer is connected to the feature extraction layer, and several Informer decoder blocks are used. The Informer decoder blocks contain masked self-attention layers and cross-attention layers.

[0030] The prediction output layer consists of a fully connected layer, with the number of neurons corresponding to the length of the future prediction time step. A linear activation function is used, and the predicted value of the new energy power point for the future period is directly output.

[0031] Furthermore, step S3 specifically includes:

[0032] S31. Input the multidimensional feature matrix into the preset joint prediction model, and divide the multidimensional feature matrix into multiple non-overlapping spatiotemporal patch blocks through the input layer;

[0033] S32. Map each spatiotemporal patch block to a high-dimensional embedding vector, and add a learnable positional encoding to each high-dimensional embedding vector to generate a patch embedding sequence.

[0034] S33. Input the patch embedding sequence into the feature extraction layer based on the Informer architecture;

[0035] S34. Between the layers of the Informer encoder block, the output of the previous layer is used as the input of the next layer through the self-attention distillation layer operation, while retaining the context information of the long sequence, to obtain the deep feature tensor.

[0036] S35. Input the deep feature tensor into the time series modeling layer and the prediction output layer to generate a prediction of the power points of new energy sources in the future period.

[0037] Furthermore, the joint prediction model is trained and optimized using a decision-oriented loss function. The training process of the joint prediction model includes:

[0038] S361. During training, the joint prediction model uses a multi-dimensional feature matrix as the model input and the corresponding historical actual new energy power value as the supervision label.

[0039] S362. Define the basic prediction loss as the mean square error between the predicted power point sequence output by the model and the true power sequence.

[0040] S363. Define asymmetric economic penalty loss, which is used to apply different penalty weights to positive and negative prediction errors.

[0041] S364. The weighted sum of the basic forecast loss and the asymmetric economic penalty loss yields the decision-oriented loss function.

[0042] S365. Minimize the loss function using the backpropagation algorithm to train and optimize the parameters of the joint prediction model, thus obtaining a well-trained joint prediction model.

[0043] Furthermore, the expression for the decision-oriented loss function is as follows:

[0044] ;

[0045] in, The total loss function value is determined by the decision-making process. The weighting coefficients for the basic predicted loss; and These are the basic forecast loss and the asymmetric economic penalty loss, respectively.

[0046] The expression for the asymmetric economic penalty loss is as follows:

[0047] ;

[0048] in, The penalty weight for positive prediction error, The penalty weights for negative prediction errors are both non-negative values, and their values ​​are determined by calibration using the historical economic operation data of the target new energy power station. Used to extract positive prediction errors and mask negative errors. This is used to extract negative prediction errors and mask positive errors, ensuring that the penalty calculations for the two types of errors are independent of each other; The number of time steps in a single batch of training samples; For the first The historical actual new energy power value at each time step; The first output of the model Predicted power values ​​at each time step.

[0049] Furthermore, step S4 specifically includes:

[0050] S41. Based on the predicted value of new energy power points for future periods output by the joint prediction model, and combined with the prediction error distribution characteristics learned by the model during training, calculate the confidence interval of the predicted value at each prediction time.

[0051] S42. Compare the upper and lower boundaries of the predicted confidence interval with the pre-set power risk level threshold to determine the warning level;

[0052] S43. Based on the time series curve of the predicted power point of new energy in the future period, identify the ramping event segments with rapid power decline and rapid rise, calculate the start time, duration and ramping rate of the ramping event, and obtain ramping event data.

[0053] S44. Based on the warning level and ramp event data, and combined with the current operating status of the power grid, generate corresponding dispatching assistance suggestions from the preset dispatching rule base;

[0054] S45. Generate new energy power prediction results based on the predicted power points of new energy sources in the future period, the confidence interval of the predicted values, the early warning level, the ramp-up event data, and the scheduling assistance suggestions.

[0055] This application also proposes a short-term forecasting system for new energy power based on multi-source meteorological data fusion, used to implement the aforementioned method for short-term forecasting of new energy power based on multi-source meteorological data fusion. The system includes:

[0056] The multi-source data acquisition and standardization module is used to collect multi-source time-series meteorological data covering the target new energy power station area, and to preprocess the multi-source time-series meteorological data to generate standard multi-source meteorological data.

[0057] The data fusion module is used to perform multimodal deep fusion of standard multi-source meteorological data based on tensor decomposition to obtain a multidimensional feature matrix;

[0058] The new energy power prediction module is used to input a multi-dimensional feature matrix into a preset joint prediction model and output the predicted value of new energy power points for future periods. The joint prediction model includes an input layer, a feature extraction layer based on the Informer architecture, a time series modeling layer, and a prediction output layer.

[0059] The new energy power prediction result generation module is used to obtain the new energy power prediction result based on the predicted value of new energy power points in the future period.

[0060] The advantages of this application compared to existing technologies are as follows: This application effectively mines potential correlation features in time, space, and variable dimensions through refined preprocessing and non-negative tensor decomposition fusion of multi-source meteorological data, avoiding data redundancy and spatiotemporal misalignment problems, and improving the integrity and reliability of multi-source data fusion; relying on the Informer architecture and ProbSparse self-attention mechanism, it reduces the computational complexity of long-term data prediction, and combines a decision-oriented loss function to balance prediction accuracy with the economic rationality of actual scheduling; through confidence interval calculation, ramp-up event identification, and early warning scheduling suggestion generation, it enriches the dimensions of prediction results, achieving an upgrade from point prediction to multi-dimensional decision support; and it improves the stability and practicality of new energy power prediction, providing a scientific basis for grid safe operation, power plant power generation planning, and energy storage scheduling. Attached Figure Description

[0061] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0062] Figure 1 This is a flowchart illustrating a method for short-term forecasting of new energy power based on the fusion of multi-source meteorological data in one embodiment.

[0063] Figure 2 This is a schematic diagram of the structure of a short-term forecasting system for new energy power based on the fusion of multi-source meteorological data in one embodiment. Detailed Implementation

[0064] like Figure 1 and Figure 2 As shown, this application provides a method for short-term forecasting of new energy power based on multi-source meteorological data fusion. The method includes the following steps:

[0065] S1. Collect multi-source time-series meteorological data covering the target new energy power station area, and perform spatiotemporal alignment, anomaly removal, missing value supplementation and normalization on the multi-source time-series meteorological data to generate standard multi-source meteorological data.

[0066] Specifically, multi-source time-series meteorological data covering the target renewable energy power plant area will be collected. This multi-source data includes numerical weather prediction grid data, satellite remote sensing data, and ground meteorological station observation data. These three types of data provide meteorological support for renewable energy power prediction from different dimensions and must be acquired according to relevant industry standards and specifications to ensure data validity. Numerical weather prediction grid data can be obtained through global or regional numerical prediction models published by authoritative meteorological institutions. It covers core meteorological elements closely related to renewable energy power generation, including near-surface wind speed, wind direction, solar irradiance, air temperature, relative humidity, and air pressure. After acquisition, it is necessary to ensure that the spatiotemporal coverage of the data completely covers the target renewable energy power plant and its surrounding affected areas. Satellite remote sensing data will be selected from Level-2 products of high-resolution meteorological satellites, mainly collecting data such as solar irradiance and cloud image inversion parameters. This type of data can compensate for the shortcomings of ground observations in spatial coverage. After acquisition, it needs to undergo radiometric calibration, atmospheric correction, and geometric correction through professional satellite data processing procedures to eliminate sensor errors and atmospheric propagation effects, improving data accuracy to meet subsequent processing requirements. The ground meteorological station observation data comes from automatic meteorological stations deployed within and around the target new energy power station. By deploying an observation network, the meteorological parameters of the target new energy power station area are accurately captured, and parameters such as wind speed, wind direction, solar irradiance, and temperature are collected in real time. The data is uploaded to the data center through a stable wireless transmission module to ensure the real-time performance and integrity of the data transmission.

[0067] The collected multi-source time-series meteorological data were sequentially processed through spatiotemporal alignment, anomaly removal, missing value supplementation, and normalization to ultimately generate standard multi-source meteorological data. The spatiotemporal alignment process used the timestamps of ground meteorological stations as a benchmark, employing linear interpolation to unify the temporal resolution of numerical weather prediction grid data and satellite remote sensing data to match the observation data from ground meteorological stations. Spatially, bilinear interpolation was used to interpolate the numerical weather prediction grid data and satellite remote sensing data to the observation locations of various ground meteorological stations, achieving precise matching of multi-source data in the spatiotemporal dimensions and avoiding fusion errors caused by spatiotemporal misalignment from the outset.

[0068] Anomaly removal employs the 3σ criterion combined with domain knowledge for dual judgment. First, the mean and standard deviation of each meteorological parameter are calculated. Data exceeding the mean plus or minus three times the standard deviation are marked as suspected anomalies. Then, a second verification is performed based on the rationality of the meteorological scenario to determine whether the suspected anomalies are caused by factors such as sensor failure or extreme weather changes. Confirmed anomalies are directly removed.

[0069] Missing value imputation employs a differentiated strategy based on the single-parameter missing value rate to ensure that the imputed data does not introduce excessive errors. When the missing value rate is low, linear interpolation is used to imputate the missing value by fitting effective data from several time steps before and after it. When the missing value rate is in the moderate range, a random forest-based interpolation method is used, employing data with the same parameters from adjacent observation points and other relevant meteorological parameters as input features to construct a random forest model to predict the missing value, leveraging the model's nonlinear fitting capability to improve imputation accuracy. When the missing value rate is too high, data for the corresponding time period is discarded and supplemented with explanations to avoid ineffective imputation affecting the overall data quality.

[0070] The normalization process employs the Min-Max normalization method, mapping all meteorological parameters to a fixed interval to eliminate the interference of differences in the dimensions of different parameters on subsequent model training. The Min-Max normalization expression is as follows:

[0071] ;

[0072] In the formula, Raw data representing various meteorological parameters, This represents the normalized parameter data. This represents the minimum value of the meteorological parameter in the full preprocessed data. This represents the maximum value of the meteorological parameter in the full preprocessed data. By traversing the full preprocessed data, the minimum and maximum values ​​of each parameter are determined, and then substituted into the formula to complete the normalization calculation, generating standard multi-source meteorological data with uniform format and qualified quality.

[0073] S2. Perform multimodal deep fusion of standard multi-source meteorological data based on tensor decomposition to obtain a multidimensional feature matrix.

[0074] Specifically, a three-dimensional tensor is first constructed based on standard multi-source meteorological data. This tensor encompasses three dimensions: time, space, and variables. Meteorological data from different sources are then filled into different dimensions of the tensor according to their modal correspondences. Numerical weather forecast grid data, satellite remote sensing data, and ground meteorological station observation data are treated as different subcategories under the variable dimension, generating a clearly structured multimodal tensor structure. The three-dimensional tensor can be represented as follows: ,in The time dimension length corresponds to the number of time steps in the preprocessed time series data. This refers to the spatial dimension length, i.e., the number of ground-based meteorological station observation points; The variable dimension length covers all core meteorological parameters; It is the set of real numbers.

[0075] Tensor decomposition employs the Tucker decomposition algorithm, which can decompose high-dimensional tensors into a core tensor and three factor matrices. This effectively preserves the nonlinear correlation characteristics of data across time, space, and variable dimensions, overcoming the limitations of traditional linear fusion in uncovering deeper information. The expression for Tucker decomposition is:

[0076] ;

[0077] in, The core tensor has a dimension of , , , These are the decomposition ranks for time, space, and variable dimensions, respectively, and the core tensor carries the core cross-dimensional correlation information of multi-source data; The time factor matrix has dimensions of . , used to extract feature information in the time dimension; The spatial factor matrix has dimensions of . , used to extract feature information in the spatial dimension; The variable factor matrix has dimensions of . , used to extract feature information in the variable dimension; express Pattern product, i.e., by the first... Perform tensor-matrix multiplication operations in each dimension.

[0078] The decomposition rank is determined using cross-validation. The constructed tensor is divided into training and validation sets in a reasonable proportion, and different rank combination ranges are set. Tucker decomposition is performed on each set, with the goal of minimizing the reconstruction error of the fused features to determine the optimal rank combination. The expression for the reconstruction error is:

[0079] ;

[0080] Where err represents the reconstruction error. The F-norm is used to measure the square root of the sum of squares of the elements of a tensor or matrix. The reconstruction error is the square of the F-norm of the original tensor and the reconstructed tensor. The current reconstruction error is compared with the previous reconstruction error; the smaller the difference, the better the decomposition and the more complete the original information is preserved. During the decomposition process, the core tensor and factor matrices are iteratively solved using alternating least squares. The matrices are initialized as random matrices, with a fixed number of iterations and a convergence threshold. Iteration stops when the difference between the reconstruction errors of two adjacent iterations is less than the convergence threshold, resulting in a stable decomposition.

[0081] Subsequently, the factor matrices obtained from the decomposition are deeply fused. First, the time factor matrix and the spatial factor matrix are multiplied to obtain the spatiotemporal correlation feature matrix. Then, the Kronecker product is performed with the variable factor matrix to strengthen the cross-modal correlation between the variable dimension and the spatiotemporal dimension. Finally, a fully connected neural network is used to perform nonlinear transformation on the fused features. GELU is selected as the activation function to output a multidimensional feature matrix whose dimension matches the input dimension of the subsequent joint prediction model. This multidimensional feature matrix not only retains the core information of the multi-source meteorological data, but also mines the cross-dimensional nonlinear complementary features.

[0082] S3. Input the multidimensional feature matrix into the preset joint prediction model and output the predicted value of the new energy power point for the future period.

[0083] Specifically, the multidimensional feature matrix is ​​input into a pre-set joint prediction model, which outputs the predicted value of new energy power points for future periods. The joint prediction model is an end-to-end architecture that has been trained and optimized in advance, including an input layer, a feature extraction layer based on the Informer architecture, a time series modeling layer, and a prediction output layer. The structure and parameter settings of each layer are clear and reproducible.

[0084] The input layer is responsible for receiving the multidimensional feature matrix and adjusting it into the input format required by the model. The input format covers three dimensions: batch size, time series length, and feature dimension. The batch size is determined by cross-validation to balance memory usage and training stability. The time series length corresponds to the number of time steps in the historical data. The feature dimension is determined by the output dimension of the multidimensional feature matrix.

[0085] The feature extraction layer based on the Informer architecture is the core of the model. This architecture solves the problem of high computational complexity of traditional self-attention through a sparse self-attention mechanism, making it suitable for long-term meteorological data processing. This layer contains several stacked Informer encoder blocks, each consisting of a sparse self-attention layer, a feedforward neural network, layer normalization, and residual connections. The sparse self-attention layer adopts a probabilistic sampling strategy, focusing attention weight calculation on key time steps. By dividing the time step into multiple intervals through bucketing, attention is calculated by randomly sampling a portion of the time steps within each interval, significantly reducing computational complexity. Simultaneously, the number of attention heads and the dropout rate are set. The number of attention heads is used to extract attention features of different dimensions in parallel, and the dropout rate is used to suppress overfitting and enhance the model's generalization ability.

[0086] The feedforward neural network consists of two fully connected layers with a GELU activation function inserted in between. The number of neurons in the first layer is several times the feature dimension, while the second layer restores the feature dimension. Nonlinear transformations are used to enhance feature representation, and a dropout layer is added to further suppress overfitting. Layer normalization employs an appropriate normalization method to standardize the input features of each layer, accelerating model convergence. Residual connections are used to alleviate the vanishing gradient problem in deep network training, ensuring stable model training.

[0087] The temporal modeling layer is connected to the feature extraction layer, and several Informer decoder blocks are used. Each decoder block contains a masked self-attention layer and a cross-attention layer. The masked self-attention layer uses masking operations to shield information from future time steps, avoiding information leakage that could affect prediction accuracy. The cross-attention layer associates the features output by the encoder with the target temporal encoding input by the decoder, accurately capturing the temporal dependency between the input features and the predicted target.

[0088] The decoder takes time codes for future time periods as input, using a sinusoidal position coding method to convert time step information into vector features, which are then fused with the encoder output features and input into the time series modeling layer.

[0089] The prediction output layer consists of a single fully connected layer, with the number of neurons corresponding to the future prediction time step length. A linear activation function is used, directly outputting the predicted value of the renewable energy power point for the future period. The model training process uses historical standard multi-source meteorological data and actual renewable energy power data for the corresponding time period as training samples. The data time span needs to cover scenarios under different seasons and meteorological conditions to ensure the model's generalization ability.

[0090] During training, the loss function used is mean squared error, and the expression for mean squared error is:

[0091] ;

[0092] in, This represents the mean squared error loss value. Indicates the number of samples. Indicates the actual power value of new energy sources. The mean squared error (MSE) represents the power value predicted by the model. It effectively measures the deviation between the predicted and actual values, providing a basis for optimizing model parameters. The optimizer uses the Adam optimizer, setting an initial learning rate and employing a cosine annealing learning rate scheduling strategy to periodically adjust the learning rate and improve training performance. It also sets the number of training iterations and uses an early stopping strategy and regularization. The early stopping strategy is implemented by monitoring the validation set loss; when the validation set loss no longer decreases for several consecutive iterations, training stops and the optimal model parameters are saved. The regularization method uses L2 regularization, which suppresses overfitting by setting a weight decay coefficient, further improving the model's generalization ability and prediction stability.

[0093] S4. Based on the predicted value of new energy power points in the future period, obtain the new energy power prediction result.

[0094] Specifically, based on the predicted power point values ​​for future periods, the final predicted power output is obtained through post-processing, accuracy verification, and result output, ensuring the reliability and engineering applicability of the prediction results. First, the predicted power point values ​​output by the model are inversely normalized to restore the actual power values. The expression for the inverse normalization formula is as follows:

[0095] ;

[0096] in, This represents the actual power value after denormalization. This represents the normalized predicted value output by the model. This represents the maximum power of the renewable energy source in the training set. This represents the minimum power of new energy sources in the training set. The maximum and minimum values ​​are determined in advance by traversing the training set data, and then substituted into the formula to complete the inverse normalization calculation, ensuring that the result conforms to the actual power range.

[0097] Subsequently, anomaly corrections are performed on the prediction results. Based on the equipment operating parameters of the new energy power plants, predicted values ​​exceeding reasonable ranges are corrected. For example, if the predicted power exceeds the equipment's rated power, it is corrected to the rated power; if the predicted power is negative, it is corrected to 0. This avoids unreasonable results caused by model prediction bias and ensures that the prediction results closely match the actual equipment operating conditions. The accuracy verification stage uses multiple evaluation indicators to quantify the prediction accuracy. The core indicators include mean absolute error, root mean square error, and mean absolute percentage error. The expressions for each indicator are as follows:

[0098] ;

[0099] ;

[0100] ;

[0101] In the formula, MAE represents the mean absolute error, which measures the average absolute deviation between the predicted value and the actual value. The smaller the value, the higher the accuracy. RMSE represents the root mean square error, which is more sensitive to larger deviations and can reflect the dispersion of the predicted value. MAPE represents the mean absolute percentage error, which expresses the prediction deviation as a percentage and makes it easier to intuitively evaluate the accuracy. This represents the actual power value for a future time period. To predict power values, To predict the number of time steps.

[0102] The accuracy of the prediction results is determined based on the mean absolute percentage error. When the error meets the preset standard, the result can be output directly. When the error exceeds the preset standard, the rank combination of the tensor decomposition or the hyperparameters of the model are adjusted, and the fusion and prediction are repeated until the accuracy requirements are met. Due to the volatility of renewable energy power, the qualified prediction results are smoothed using a moving average method. This method eliminates short-term fluctuations by weighted averaging of the predicted values ​​at adjacent time steps, making the prediction results more consistent with actual operating scenarios. The final output of the renewable energy power prediction results includes power time-series data divided into fixed time intervals within a preset future period. The format uses a general data file format and a visualized time-series curve. The data file includes timestamps, predicted power values, and meteorological parameters for the corresponding time period, facilitating reading by the power station dispatch system. The visualized curve uses time as the horizontal axis and power as the vertical axis, marking power peaks, valleys, and corresponding time points, while also overlaying actual meteorological data trends, providing an intuitive reference for power station operation and dispatch. The prediction results also need to generate a reliability analysis report, explaining the accuracy indicators, applicable meteorological conditions, and possible sources of error in this prediction, providing a scientific basis for the formulation of power generation plans for renewable energy power stations and grid dispatch coordination.

[0103] The aforementioned method for short-term forecasting of renewable energy power through multi-source meteorological data fusion fully leverages the nonlinear correlation characteristics of time, space, and variables by refining the preprocessing of multi-source meteorological data and deeply fusing it with non-negative tensor decomposition. This avoids spatiotemporal misalignment and data redundancy, improving the quality of the fused data. Utilizing the Informer architecture and ProbSparse self-attention mechanism, the computational pressure of long-term data processing is reduced. Combined with a decision-oriented loss function, the method ensures forecast accuracy while considering the economic rationality of actual scheduling. Through confidence interval calculation, ramp-up event identification, and early warning scheduling suggestion generation, the method overcomes the limitations of single-power-point forecasting, achieving multi-dimensional decision support, improving the stability and engineering practicality of renewable energy power forecasting, providing a scientific basis for grid safety operation and power plant generation planning, and effectively alleviating the scheduling challenges caused by the volatility of renewable energy generation.

[0104] In an optional embodiment, standard multi-source meteorological data are subjected to multimodal deep fusion based on tensor decomposition to obtain a multidimensional feature matrix, including:

[0105] S21. The numerical weather forecast grid data in the standard multi-source meteorological data is mapped to the geographic coordinates of the target new energy power station through bilinear interpolation to obtain the background meteorological vector sequence at the station scale.

[0106] Optionally, the bilinear interpolation method, due to its high spatial interpolation accuracy and ability to effectively fit the mapping relationship between grid data and discrete geographic coordinates, is suitable for converting large-scale grid meteorological data into refined data at the station scale. During implementation, the core geographic coordinates of the target new energy power station and the coordinates of surrounding key observation points are first identified. These coordinates are used as interpolation targets, and the corresponding grid node data from the numerical weather prediction grid data is called. Each grid node contains complete meteorological parameter information. Bilinear interpolation calculates a weighted average of the four nearest grid node data around the target coordinate point. The weights are determined by the spatial distance between the target coordinate point and each grid node; the closer the distance, the greater the weight. The meteorological parameter value for each target coordinate point is obtained through weighted summation, ultimately generating a background meteorological vector sequence arranged in chronological order. Each vector in this sequence corresponds to a set of meteorological parameters for all target coordinate points at a given time step, covering core elements such as wind speed, wind direction, temperature, and air pressure.

[0107] S22. Based on satellite remote sensing observation data, remote sensing estimates of cloud optical thickness and surface incident shortwave radiation intensity in the target station area are obtained by inversion through a simplified radiative transfer model, and a satellite-derived feature vector sequence is generated.

[0108] Optionally, while satellite remote sensing data can provide meteorological information covering a wide spatial area, the raw data cannot be directly used for power prediction at power plants. Key parameters closely related to new energy power generation need to be obtained through inversion. A simplified version of the radiative transfer model suitable for short-term timescale prediction is selected to reduce computational complexity while ensuring inversion accuracy and avoiding the impact of complex models on overall prediction efficiency. During the inversion process, standard-processed satellite remote sensing data is used as input, including parameters such as radiance values ​​of the satellite observation channel, observation angle, and satellite altitude. The simplified radiative transfer model establishes a mapping relationship between observed values ​​and cloud optical thickness and surface incident shortwave radiation intensity, eliminating interference from atmospheric scattering and absorption on the remote sensing signal. This inversion yields the cloud optical thickness and surface incident shortwave radiation intensity at each spatial location within the target power plant area at each time step. The two key parameters obtained from the inversion are organized chronologically and associated with information from the corresponding time step and spatial location to generate a satellite-derived feature vector sequence. This sequence supplements cloud conditions and radiation intensity information that is difficult to cover by ground observations and numerical forecasts, completing the complementarity of multimodal data.

[0109] S23. Constructing a multidimensional spatiotemporal meteorological tensor: Arrange the background meteorological vector sequence, satellite-derived feature vector sequence, and ground meteorological station observation data sequence according to the time dimension, spatial dimension, and meteorological variable dimension to construct a third-order spatiotemporal meteorological tensor. .

[0110] in, The time step corresponds to the length of the entire time series data, and is consistent with the time resolution of the preprocessed data to ensure the continuity of the time dimension. The spatial location data includes all ground meteorological station observation points deployed inside and around the target new energy power station, and the core coordinate points of the station obtained by interpolation, thus obtaining a complete spatial monitoring network; For the meteorological variable dimension, all core meteorological variables from three types of data are integrated, including wind speed, wind direction, temperature, and air pressure from the background meteorological vector sequence; cloud optical thickness and surface incident shortwave radiation intensity from the satellite-derived feature vector sequence; and solar irradiance and relative humidity from the ground meteorological station observation data sequence. Each variable is treated as an independent dimension component to ensure the integrity of variable information. During the arrangement process, the data sequences are strictly aligned according to the timestamps. Different meteorological variables at the same time step and the same spatial location correspond to different variable dimensions of the tensor, resulting in a well-structured and information-complete third-order spatiotemporal meteorological tensor.

[0111] S24. Construct a fusion model based on nonnegative tensor decomposition; define the objective function of the fusion model based on nonnegative tensor decomposition as minimizing the Frobenius norm error between the reconstructed tensor and the original tensor, and add a regularization term constraint. The expression of the objective function is:

[0112] ;

[0113] In the formula, It is a third-order spacetime tensor. For the core tensor, For time pattern matrix, For spatial pattern matrix, For the variable pattern matrix, The regularization coefficient is . It represents the n-modulo product.

[0114] Specifically, due to the non-negative nature of meteorological data, using a non-negative tensor decomposition algorithm to construct the fusion model can effectively preserve the physical meaning of the data and avoid information distortion caused by negative values ​​during the decomposition process. The core of the model is defining an objective function, which focuses on minimizing the Frobenius norm error between the reconstructed tensor and the original third-order spatiotemporal meteorological tensor. A regularization term is also added to constrain the model, suppressing overfitting and improving the generalization ability of the fused features. In the expression of the above objective function, The constructed third-order spatiotemporal meteorological tensor carries the original multi-source meteorological data information; As the core tensor, it is used to represent the potential correlation features in the three dimensions of time, space, and variables, and is the core information carrier of multimodal data fusion; This is the time pattern matrix, corresponding to the feature extraction matrix of the time dimension, used to mine the changing patterns of meteorological data over time series; This is the spatial pattern matrix, corresponding to the feature extraction matrix of the spatial dimension, used to capture the correlation characteristics of meteorological parameters at different spatial locations; The variable pattern matrix corresponds to the feature extraction matrix of the variable dimension, which is used to explore the intrinsic relationship between different meteorological variables; is the regularization coefficient, used to adjust the weight of the regularization term, balancing the model fitting accuracy and the risk of overfitting. Its value is determined through cross-validation. Represents the n-modulo product, i.e., according to the nth... Perform tensor-matrix multiplication operations in each dimension to achieve fusion calculation of tensors and various pattern matrices; The Frobenius norm is used to measure the square root of the sum of squares of the elements of a tensor or matrix, precisely quantifying the error between the original tensor and the reconstructed tensor.

[0115] S25. The objective function is solved iteratively using the alternating least squares method. Two factor matrices are fixed and the third factor matrix is ​​updated until convergence, and the converged core tensor is obtained.

[0116] Optionally, the alternating least squares method has the advantages of convergence stability and moderate computational complexity, making it suitable for solving the objective function of nonnegative tensor decomposition. The principle is to fix two factor matrices and update the third factor matrix by minimizing the objective function, iterating repeatedly until the convergence condition is met. Specifically, the time pattern matrix is ​​first initialized. Spatial pattern matrix and variable pattern matrix It is a non-negative random matrix, ensuring that the matrix elements conform to the non-negativity characteristic of meteorological data. The first round of iteration is fixed. and Transform the objective function into a function about For a single-variable optimization problem, the derivative of the objective function is derived by taking the derivative and setting it to zero. The update formula is used to calculate the updated result by substituting the data. The second round of iterations fixed the update. and Similarly, the derivation The updated formula was used to calculate the new formula. The third round of iterations after fixed updates and Derivation The updated formula was used to calculate the new formula. Three rounds of updates constitute one complete iteration. This process is repeated to continuously update the three factor matrices. Simultaneously, the objective function value is calculated in each iteration. A convergence threshold and a maximum number of iterations are set as stopping conditions. Iteration stops when the difference between the objective function values ​​of two adjacent iterations is less than the convergence threshold, or when the maximum number of iterations is reached. The converged core tensor is then obtained. And the corresponding three factor matrices.

[0117] S26. Reconstruct the converged core tensor and the corresponding factor matrix, and flatten the reconstructed tensor data to generate a multidimensional feature matrix.

[0118] Optionally, the converged core tensor With the corresponding time pattern matrix Spatial pattern matrix Variable pattern matrix Tensor reconstruction is performed according to the n-modulus product rule. The reconstruction process strictly follows the operational logic in the objective function, resulting in a tensor similar to the original third-order spacetime tensor. A dimensionally consistent reconstructed tensor is used. The reconstructed tensor retains the core information of the original multi-source meteorological data, while non-negative tensor decomposition uncovers potential correlations across time, space, and variable dimensions, eliminating redundant information and noise interference from the multi-source data. The reconstructed tensor data is then flattened, transforming the third-order tensor into a two-dimensional matrix, i.e., a multi-dimensional feature matrix, with rows corresponding to each time step and columns corresponding to all meteorological variable features at each spatial location. The dimensions match the input dimensions of the subsequent joint prediction model, providing high-quality fusion feature input for the model.

[0119] In an optional embodiment, the objective function is iteratively solved using alternating least squares, fixing two factor matrices and updating the third factor matrix until convergence, resulting in the converged core tensor, which includes:

[0120] S251. Initialize the time pattern matrix, spatial pattern matrix, and variable pattern matrix.

[0121] Optionally, since meteorological data inherently possesses non-negativity, and non-negative tensor decomposition requires all matrix elements to be non-negative to preserve physical meaning, the three model matrices must be initialized as non-negative random matrices during the initialization phase. During initialization, the range of matrix element values ​​is set to a reasonable non-negative interval to avoid initial values ​​being too large or too small, which could lead to slow iteration convergence or getting trapped in local optima. Simultaneously, the core tensor is also initialized with non-negative random values, its dimension determined by a preset decomposition rank to ensure it matches the dimensions of the three model matrices, providing basic initial parameters for subsequent iterative updates. After initialization, the matrices and tensors need to be validated to ensure there are no outliers or dimension mismatches, guaranteeing a smooth start to the iteration process.

[0122] S252. Fix the spatial pattern matrix and variable pattern matrix, and update the temporal pattern matrix: calculate the Khatri-Rao product of the third-order spatiotemporal meteorological tensor and the core tensor, solve for the update of the temporal pattern matrix using the least squares method, and set the negative elements in the temporal pattern matrix to zero.

[0123] Optionally, before updating, the Khatri-Rao product of the third-order spatiotemporal meteorological tensor and the core tensor needs to be calculated. The Khatri-Rao product, a commonly used joint product operation in tensor decomposition, effectively integrates column vector information from multiple matrices, providing accurate input data for pattern matrix updates. Specifically, the object of the Khatri-Rao product calculation here is the spatial pattern matrix. and variable pattern matrix By performing a column-directed Khatri-Rao product operation on the two, a joint matrix is ​​obtained. ,in This represents the Khatri-Rao product operator, whose operation rule is to... The Columns and The Perform outer product operations on the columns, and then stack all the outer product results column-wise to obtain the joint matrix. After the calculation is complete, convert the third-order spatiotemporal tensor... Expanding the time-dimensional matrix into a matrix form and combining it with the resulting Khatri-Rao product joint matrix, the time pattern matrix is ​​solved using the least squares method. The updated value. The core of the least squares method is to construct an overdetermined system of equations and obtain the optimal solution by minimizing the sum of squared residuals. Specifically, the update formula for the time pattern matrix is ​​derived as follows: ,in Tensor Expanding the matrix according to the first dimension (time dimension), This represents the Moore-Penrose pseudoinverse of the Khatri-Rao product joint matrix, used to solve the problem of non-invertibility of the equation system. After updating the time model matrix, the negative elements in the matrix need to be set to zero. Since meteorological data and decomposition features have no negative physical meaning, the presence of negative elements will lead to information distortion. The zeroing operation can ensure that the matrix always satisfies the non-negativity constraint and maintain the rationality of the decomposition process.

[0124] S253. Fixed-time pattern matrix and variable pattern matrix, update spatial pattern matrix: calculate the Khatri-Rao product of the third-order temporal-space image tensor and the core tensor, solve for the update spatial pattern matrix using the least squares method, and set the negative elements in the spatial pattern matrix to zero.

[0125] Optionally, the Khatri-Rao product of the third-order spatiotemporal imagery tensor and the core tensor is first calculated, with the time pattern matrix being the object of this operation. and variable pattern matrix Similarly, the joint matrix is ​​obtained through column-directed Khatri-Rao product operations. Then the third-order spacetime tensor Expand into a matrix form according to the second dimension (spatial dimension). Combined with joint matrix The update value of the spatial pattern matrix H is solved using the least squares method. The update formula is as follows: ,in joint matrix The Moore-Penrose pseudoinverse is calculated. After the update, negative elements in the spatial pattern matrix are set to zero, strictly adhering to the non-negativity constraint to ensure that the matrix elements conform to the physical characteristics of the meteorological data and to avoid introducing invalid or erroneous feature information due to negative elements, which would affect the accuracy of subsequent fusion. After the zeroing operation, the matching between the matrix dimension and the core tensor needs to be verified to ensure the consistency of subsequent iterative updates.

[0126] S254. Fixed time pattern matrix and spatial pattern matrix, update variable pattern matrix: Calculate the Khatri-Rao product of the third-order temporal and spatial meteorological tensor and the core tensor, solve for the updated variable pattern matrix using the least squares method, and set the negative elements in the variable pattern matrix to zero.

[0127] Optionally, the Khatri-Rao product of the third-order spatiotemporal image tensor and the core tensor is first calculated, with the time pattern matrix as the operand. and spatial pattern matrix The joint matrix is ​​obtained by column-directed Khatri-Rao product operation. The third-order spacetime tensor Expand into a matrix form according to the third dimension (variable dimension). Combined with joint matrix Solving the variable pattern matrix using the least squares method The updated value is determined by the following formula: ,in joint matrix The Moore-Penrose pseudoinverse is used. After updating the variable model matrix, the negative elements are set to zero to maintain the non-negativity constraint, ensuring the physical rationality of the variable dimensional features and ensuring that the feature extraction of different meteorological variables conforms to the actual data patterns. This completes a full cycle of updating the three model matrices, simultaneously updating the core tensor. To adapt to the new pattern matrix and prepare for the next iteration.

[0128] S255. Repeat S252 to S254 until the reconstruction error between two adjacent iterations is less than the preset threshold or the maximum number of iterations is reached, and the converged core tensor is obtained.

[0129] Optionally, steps S252 to S254 are repeated. After each iteration, the reconstruction error of the current iteration is calculated. The reconstruction error is calculated using the Frobenius norm error in the objective function, i.e., by calculating the original third-order spacetime tensor. With the current core tensor The Frobenius norm error between the tensors reconstructed from the three model matrices is used. The reconstruction error of the current iteration is compared with that of the previous iteration. If the difference is less than a preset convergence threshold, the model has reached a stable state, and the iteration can stop. If the difference is still greater than the preset threshold, the iteration process continues. A maximum number of iterations is set as a fallback condition to avoid infinite iteration due to an excessively small convergence threshold. When the maximum number of iterations is reached, iteration stops regardless of whether the reconstruction error meets the threshold requirement, and the current core tensor is used as the converged result. Both the convergence threshold and the maximum number of iterations are determined through cross-validation to balance iteration efficiency and fusion accuracy. The resulting converged core tensor accurately reflects the potential correlation characteristics of multi-source meteorological data in the temporal, spatial, and variable dimensions.

[0130] In an optional embodiment, the multidimensional feature matrix is ​​input into a preset joint prediction model to output a prediction of the power points of new energy sources in the future, including:

[0131] S31. Input the multidimensional feature matrix into the preset joint prediction model, and divide the multidimensional feature matrix into multiple non-overlapping spatiotemporal patch blocks through the input layer.

[0132] Optionally, the dimensions of the multidimensional feature matrix correspond to the time step, spatial location, and meteorological variable features. The segmentation operation of the input layer must simultaneously consider the local correlations of the time and spatial dimensions to avoid disrupting the intrinsic connections between adjacent spatiotemporal information. During segmentation, the matrix is ​​divided into blocks of a fixed size. Each spatiotemporal patch block covers continuous time step segments, a fixed number of spatial locations, and all meteorological variable features, with no overlapping areas between blocks to ensure that feature information is not repeated or omitted. The segmentation size is determined through cross-validation, balancing local feature capture capability with model computational efficiency. This ensures that each patch block can carry sufficient spatiotemporal correlation information while avoiding excessively large block sizes that would increase the complexity of subsequent processing. After segmentation, each patch block undergoes dimensionality normalization, adjusting it to the standard format required for subsequent processing in the input layer, laying the foundation for mapping the patch block to a high-dimensional embedding vector.

[0133] S32. Map each spatiotemporal patch block to a high-dimensional embedding vector, and add a learnable positional encoding to each high-dimensional embedding vector to generate a patch embedding sequence.

[0134] Optionally, the mapping process is implemented through a fully connected neural network. The input dimension of this fully connected layer is the feature dimension of a single spatiotemporal patch, and the output dimension is a preset high-dimensional embedding dimension. Nonlinear transformation is used to upscale the low-dimensional spatiotemporal patch features to a high-dimensional space, enhancing feature representation capabilities. Learnable positional encoding is chosen instead of fixed positional encoding because it can adaptively capture the positional correlation characteristics of different spatiotemporal patches. The positional encoding vector is dynamically adjusted through model training, more accurately reflecting the relative positional information of the patch in the overall temporal and spatial order, avoiding the limitations of fixed encoding in adapting to complex spatiotemporal features. When adding positional encoding, each high-dimensional embedding vector is element-wise added to its corresponding positional encoding vector, preserving both the patch's own feature information and incorporating positional information. Finally, all processed vectors are arranged in the original spatiotemporal order to generate a patch embedding sequence.

[0135] S33. Input the patch embedding sequence into the feature extraction layer based on the Informer architecture.

[0136] Optionally, the feature extraction layer based on the Informer architecture comprises multi-level stacked Informer encoder blocks. This multi-level stacking structure enables hierarchical feature extraction; shallow Informer blocks capture basic spatiotemporal correlation features, while deeper blocks mine more complex high-order features, progressively improving the depth and effectiveness of feature representation. Each Informer block employs the ProbSparse self-attention mechanism, a core technology for adapting to long sequence inputs. This mechanism replaces the full computation of traditional self-attention with a sparse sampling strategy, significantly reducing computational complexity while focusing on key information. Specifically, the ProbSparse self-attention mechanism first evaluates the importance of query points in the patch embedding sequence, selecting key query points. Attention scores are calculated only for key query points and all key-value pairs, rather than calculating for each query point individually. This reduces the computational complexity of attention from quadratic to linear-logarithmic levels, perfectly solving the problem of excessive computation in long-term data processing. The attention score calculation is based on the similarity between the query point and the key-value pair, combined with Softmax function normalization to obtain the attention weights corresponding to each key query point. The attention output is then obtained through weighted summation of these weights. Each Informer block is also equipped with a feedforward neural network, layer normalization, and residual connections. The feedforward neural network enhances feature representation through nonlinear transformation, layer normalization standardizes features to accelerate model convergence, and residual connections alleviate the gradient vanishing problem caused by deep stacking, ensuring stable training and feature transfer at each Informer block level.

[0137] S34. Between the layers of the Informer encoder block, the output of the previous layer is used as the input of the next layer through the self-attention distillation layer operation, while retaining the context information of the long sequence, to obtain the deep feature tensor.

[0138] Optionally, the core function of the self-attention distillation layer is to filter and compress information during feature propagation, reducing redundant information and model parameter size while ensuring that global contextual information of long sequences is not lost. Specifically, the self-attention distillation layer first performs self-attention calculation on the feature sequence output by the upper Informer block, focusing on global correlation information across blocks. Then, it adjusts the dimension of the upper output to match the dimension of the lower input through feature mapping. Simultaneously, it employs an information preservation strategy guided by attention weights, assigning higher weights to key contextual features and weakening the influence of irrelevant information. This distillation operation is performed between each layer, enabling the feature sequence to remove redundancy and retain core information during progressive propagation. Finally, the last distillation layer outputs a deep feature tensor that integrates multi-level encoded features and global contextual information. The dimension of this deep feature tensor precisely matches the input requirements of subsequent temporal modeling layers.

[0139] S35. Input the deep feature tensor into the time series modeling layer and the prediction output layer to generate a prediction of the power points of new energy sources in the future period.

[0140] Optionally, the temporal modeling layer adopts an Informer decoder structure, working in conjunction with the Informer encoder block. The decoder includes a masked self-attention layer and a cross-attention layer. The masked self-attention layer masks information from future time steps through masking operations, completely avoiding information leakage and ensuring that the model predicts only based on historical and current features. The cross-attention layer correlates the deep feature tensor with the future time encoding input to the decoder, accurately capturing the temporal dependency between the encoded features and the prediction target, further enhancing feature mining in the time dimension. The future time encoding uses a sinusoidal positional encoding method, transforming the positional information of each future time step into vector features, which are complementary and fused with the temporal features of the deep feature tensor. The prediction output layer consists of a fully connected layer, with the input being the feature vector output from the temporal modeling layer. The output dimension corresponds to the length of the future prediction time step, and a linear activation function is used to directly map the feature vector to the predicted values ​​of the new energy power points at each future time step.

[0141] In an optional embodiment, the implementation process of the ProbSparse self-attention mechanism includes:

[0142] S331. Calculate the sparsity measure of the query vector. The expression for the sparsity measure is:

[0143] ;

[0144] In the formula, For the sparsity measure of the query vector, For the first A query vector, For a set of key vectors, The length of the key vector. For vector dimensions, This is the average of the dot products of all query keys.

[0145] In the above expression, For the first The sparsity metric of each query vector. The larger the value, the more effective the association between the corresponding query vector and the set of key vectors, and the more it needs to be retained for subsequent attention calculations. For the first Each query vector is derived from the output of the patch embedding sequence after linear transformation, and each query vector corresponds to the feature representation of a patch block. It is a set of key vectors, generated synchronously with the query vector, and obtained by linear transformation of the features of all patch blocks. The two have the same dimension to meet the requirements of dot product operation. This is the length of the key vectors, which is the total number of key vectors contained in the key vector set, and should be consistent with the number of query vectors. The vector dimension, i.e., the feature dimension of the query vector and the key vector, is used to scale the dot product result to avoid numerical inflation caused by excessive dimensionality. The average value of the dot product of all query and key vectors is calculated by iterating through the dot product results of all query vectors and key vectors, and taking the arithmetic mean as the benchmark threshold to filter valid related items. This is a simplified form of the ReLU activation function, used to filter out dot product terms below a benchmark threshold, retaining only positive and valid associations, thus enhancing the discriminative power of the sparsity metric. During the calculation, each query vector is first processed... With key vector set Each key vector in Perform dot product operation, divide by After scaling, subtract the baseline threshold. ,pass After filtering out invalid items, the average of all results is taken to obtain the sparsity metric of the query vector. .

[0146] S332. Sort all query vectors in descending order based on their sparsity metric, and select the top-ranked... The query vectors are used as a sparse query set.

[0147] Optionally, the core logic of descending order sorting is to prioritize retaining query vectors closely associated with the key vector set and discard query vectors with sparse associations, thereby sparsifying attention computation. The size of the sparse query set is determined through cross-validation, balancing computational complexity with feature capture accuracy. It is typically chosen as a certain proportion of the total number of query vectors to ensure that core relevance features are not lost while significantly reducing computational cost. During the sorting process, the sparsity metric of each query vector is used. Sort by size from largest to smallest, and extract the first few bytes. The query vectors form a sparse query set, which contains only the key query vectors that contribute significantly to the attention calculation. The attention score is then calculated based solely on this set, transforming the full calculation of traditional self-attention into sparse calculation.

[0148] S333. Calculate the self-attention output based on the sparse query set to obtain the attention score.

[0149] Optionally, firstly, a simplified query matrix is ​​constructed based on the sparse query set, containing only the first few queries. One key query vector, while retaining the complete set of key vectors. Sum value vector set (The value vector set and key vector set are generated simultaneously, representing another dimension of the patch block features.) Then, the attention score is calculated by performing a dot product operation between each query vector in the sparse query matrix and all key vectors in the key vector set, and dividing by... After scaling, the data is normalized using the Softmax function to obtain the attention weights for each key vector corresponding to each key query vector. The sum of these weights is 1, quantifying the contribution of different key vectors to the current query vector. Finally, the normalized attention weights are weighted and summed with the set of value vectors to obtain the self-attention output. This result retains the association information between the key query vector and all key vectors while significantly reducing computational complexity through sparse sampling, thus reducing the computational complexity of attention from... ( (To reduce the total number of query / key vectors) to It perfectly adapts to the processing requirements of long-time sequence patch embedding sequences.

[0150] In an optional embodiment, the joint prediction model is trained and optimized using a decision-oriented loss function. The training process of the joint prediction model includes:

[0151] S361. During training, the joint prediction model uses a multi-dimensional feature matrix as the model input and the corresponding historical actual renewable energy power value as the supervision label.

[0152] Optionally, during model training, a multidimensional feature matrix is ​​used as the model input to ensure that the input data is completely consistent with the preprocessing process in the prediction stage, guaranteeing consistency between training and inference. Historical actual renewable energy power values ​​for the corresponding time period are used as supervision labels. These historical actual power values ​​are derived from real-time data collected by the SCADA system of the target renewable energy plant. After data cleaning (removing outliers caused by equipment failures or communication interruptions), the data is precisely aligned with the multidimensional feature matrix by timestamp to form one-to-one training sample pairs. The training samples need to cover different seasons, different meteorological conditions (such as sunny, cloudy, rainy, and windy), and different power generation load scenarios. The data time span needs to meet the model's generalization ability requirements. Simultaneously, the data is divided into training, validation, and test sets in a reasonable proportion for parameter training, hyperparameter optimization, and performance verification, respectively.

[0153] S362. Define the basic prediction loss as the mean square error between the predicted power point sequence output by the model and the actual power sequence.

[0154] Optionally, the base prediction loss is the mean square error between the predicted power point sequence and the true power sequence output by the model, expressed as follows:

[0155] ;

[0156] in, Based on the predicted loss value, This represents the number of time steps in a single batch of training samples. For the first Historical actual renewable energy power values ​​at each time step (monitoring label). The first output of the model The predicted power points at each time step. The mean squared error is more sensitive to larger deviations, effectively constraining the overall prediction accuracy of the model and providing a basic deviation measurement for the loss function.

[0157] S363. Define asymmetric economic penalty loss, which is used to apply different penalty weights to positive and negative prediction errors.

[0158] Optionally, an asymmetric economic penalty loss is defined, applying different penalty weights to positive and negative prediction errors to adapt to the actual economic needs of renewable energy power plant scheduling. Positive prediction error is defined as the deviation between the predicted power value and the actual power value. Such errors may lead to over-planning of power generation resources at power plants, resulting in power curtailment losses or redundant investment costs in energy storage equipment; negative prediction error is defined as the deviation where the predicted power value is less than the actual power value. Such errors can lead to insufficient power supply, resulting in grid dispatch compensation costs or load losses due to power outages. Based on the different economic impacts of the two types of errors, differentiated penalty weights are set, and the expression for the asymmetric economic penalty loss is as follows:

[0159] ;

[0160] The parameters and their setting logic in the formula are explained below: This represents the asymmetric economic penalty loss value. The penalty weight for positive prediction error, The penalty weights for negative prediction errors are both non-negative values, determined through calibration using historical economic operation data of the target renewable energy power plants. For example, when the power curtailment loss exceeds the power shortage loss, a penalty weight is set. Conversely, set ; Used to extract positive prediction errors and mask negative errors. This is used to extract negative prediction errors and mask positive errors, ensuring that the penalty calculations for the two types of errors are independent of each other; , , The definition is consistent with the basic prediction loss formula, ensuring the uniformity of loss calculation.

[0161] S364. Weighted summation of basic forecast loss and asymmetric economic penalty loss yields the decision-oriented loss function.

[0162] Optionally, a decision-oriented loss function can be constructed, which weights and sums the basic prediction loss and the asymmetric economic penalty loss to achieve synergistic optimization of prediction accuracy and economic rationality. The expression for the loss function is as follows:

[0163] ;

[0164] in, The total loss function value is determined by the decision-making process. The weighting coefficient for the basic prediction loss ranges from [0,1] and is determined through cross-validation. It is used to balance the importance of basic prediction accuracy and economic penalty, and is usually set to 0.6-0.8, prioritizing prediction accuracy while taking into account economic needs. and These are the basic forecast loss and the asymmetric economic penalty loss, as defined above.

[0165] S365. Minimize the loss function using the backpropagation algorithm to train and optimize the parameters of the joint prediction model, thus obtaining a well-trained joint prediction model.

[0166] Optionally, the model parameters are optimized by minimizing the total loss function using the backpropagation algorithm. During training, the Adam optimizer is used to update the parameters of each layer of the model. The initial learning rate is determined through grid search, and a cosine annealing learning rate scheduling strategy is adopted to gradually reduce the learning rate with each iteration, avoiding the model from getting trapped in local optima. Training set samples are input into the model in batches, and the total loss value of each batch is calculated using forward propagation. The gradient of the loss function with respect to each model parameter, including the weights of fully connected layers, attention mechanism parameters, and position encoding vectors, is solved using the backpropagation algorithm. The parameters are updated based on the gradient descent principle, so that the total loss value gradually decreases. An early stopping strategy is introduced during training, using the total loss value of the validation set as a monitoring indicator. When the validation set loss no longer decreases for several consecutive iterations, training is stopped and the optimal model parameters are saved to avoid overfitting. At the same time, the L2 regularization method is used to suppress excessively large model parameters by setting a weight decay coefficient, thereby improving the model's generalization ability. Finally, a joint prediction model that has been trained and is adapted to actual scheduling requirements is obtained.

[0167] In an optional embodiment, the renewable energy power prediction result is obtained based on the predicted renewable energy power point value for a future period, including:

[0168] S41. Based on the predicted value of new energy power points for future periods output by the joint prediction model, and combined with the prediction error distribution characteristics learned by the model during training, calculate the confidence interval of the predicted value at each prediction time.

[0169] Optionally, the predicted power points output by the model are first pre-processed to ensure data validity. First, inverse normalization is performed to restore the actual power values. The expression for the inverse normalization formula is:

[0170] ;

[0171] in, This represents the actual power value after denormalization. This represents the normalized predicted value output by the model. This represents the maximum power of the renewable energy source in the training set. This represents the minimum power output of new energy sources in the training set, pre-determined by traversing the training set data to ensure the result matches the actual power range. Subsequently, anomaly correction is performed, combining the equipment operating parameters of the new energy power plants to correct predicted values ​​that exceed reasonable ranges. For example, when the predicted power exceeds the equipment's rated power, it is corrected to the rated power; when the predicted power is negative, it is corrected to 0, avoiding unreasonable results caused by model prediction bias.

[0172] Specifically, based on the corrected predicted values ​​of renewable energy power points for future periods, and combined with the prediction error distribution characteristics learned by the model during training, the confidence interval is quantitatively calculated. During model training, the prediction residuals of each batch of training samples are recorded synchronously, i.e., the difference between the actual power value and the predicted power value. The probability distribution characteristics of the fitted residuals are statistically analyzed. Commonly used distribution types include normal distribution, t-distribution, or Gaussian mixture distribution. The specific distribution type is determined through the Kolmogorov-Smirnov test to ensure fitting accuracy. For each prediction time, centered on the predicted value at that time, based on the fitted error distribution characteristics, the upper and lower boundaries of the confidence interval are calculated according to a preset confidence level (set according to engineering requirements, usually 95%). If the error follows a normal distribution, the expression for the confidence interval is: ,in, The predicted power point value at a certain prediction time. This represents the standard normal distribution quantile at the corresponding confidence level. This represents the standard deviation of the error distribution obtained during the training process. The width of the confidence interval reflects the magnitude of prediction uncertainty; a narrower interval indicates higher prediction reliability, providing a quantitative basis for subsequent risk warnings.

[0173] S42. Compare the upper and lower boundaries of the confidence interval of the predicted value with the pre-set power risk level threshold to determine the warning level.

[0174] Optionally, the upper and lower boundaries of the confidence interval of the predicted value at each prediction time are compared with a pre-set power risk level threshold, and the warning level at the corresponding time is determined based on the interval coverage. The power risk level threshold is set based on the rated parameters of the target renewable energy power plant's equipment, grid safety operation constraints, and historical fault data, and is divided into multiple levels, typically four: no warning, blue warning, yellow warning, orange warning, and red warning, each level corresponding to a specific threshold range. For example, when the upper boundary of the confidence interval exceeds the grid acceptance upper limit threshold, or the lower boundary is lower than the power plant's guaranteed power supply threshold, the corresponding level of warning is triggered; if the confidence interval falls entirely within the safety threshold range, it is determined to be no warning. The determination of the warning level must take into account the degree of deviation between the interval boundary and the threshold; the greater the deviation, the higher the warning level. At the same time, the prediction time and specific reason for triggering the warning are recorded to provide risk indications for subsequent dispatch decisions.

[0175] S43. Based on the time series curve of the predicted power point of new energy in the future period, identify the ramping event segments with rapid power decline and rapid rise, calculate the start time, duration and ramping rate of the ramping event, and obtain ramping event data.

[0176] Optionally, based on the time-series curves of predicted renewable energy power points for future periods, and using a preset ramp rate threshold, ramp event segments characterized by rapid power decreases and increases are identified. The ramp rate threshold is determined by the performance of power plant equipment (such as the maximum power change rate of wind turbines and photovoltaic modules) and grid dispatch requirements, distinguishing between positive ramp (rapid power increase) and negative ramp (rapid power decrease) events. Specifically, the ratio of power change to time interval between adjacent prediction moments (i.e., instantaneous ramp rate) is calculated. When the instantaneous ramp rate exceeds the preset threshold for multiple consecutive time steps, and the duration reaches the minimum event duration requirement, it is determined as a ramp event. Subsequently, the core data of the ramp event is extracted: the start time is the prediction moment when the threshold is first exceeded; the duration is the time from the start time to the instantaneous ramp rate returning to the threshold; and the ramp rate is the average or maximum instantaneous ramp rate within the event segment. Simultaneously, the power peak, valley, and change amplitude within the event segment are recorded to obtain complete ramp event data, providing support for the grid to respond to power surges.

[0177] S44. Based on the warning level and ramp event data, and combined with the current operating status of the power grid, generate corresponding dispatch assistance suggestions from the preset dispatch rule base.

[0178] Optionally, based on the determined warning level and extracted ramp-up event data, combined with the current grid operating status, including real-time data such as grid load level, reserve capacity, transmission line load rate, and energy storage device charging and discharging status, corresponding dispatching assistance suggestions are matched from a preset dispatching rule base. The dispatching rule base is constructed based on the operating experience of new energy power plants, grid dispatching procedures, and optimization algorithms. It includes standardized dispatching strategies corresponding to different warning levels, ramp-up event types, and grid operating status combinations, while reserving flexible adjustment interfaces to adapt to special scenarios. For example, for a red warning accompanied by a positive ramp-up event, if the grid reserve capacity is insufficient, the rule base matches the suggestion of "starting energy storage device charging + reducing part of the generation load"; for a negative ramp-up event with a high grid load, it matches the suggestion of "calling backup power + energy storage device discharging to replenish energy". The suggestions must clearly define the operation target, execution timing, adjustment range, and safety constraints, and possess engineering operability.

[0179] S45. Generate new energy power prediction results based on the predicted power points of new energy sources in the future period, the confidence interval of the predicted values, the early warning level, the ramp-up event data, and the scheduling assistance suggestions.

[0180] Optionally, the system integrates multi-dimensional information, including predicted power points for future periods (including inversely normalized and anomaly-corrected data), confidence intervals for predicted values ​​at each prediction time (including upper and lower boundaries and confidence levels), warning levels at each time (including triggering reasons), ramp-up event data (including event type, start and end times, ramp-up rates, etc.), and generated dispatching assistance suggestions. It also supplements prediction accuracy indicators (such as mean absolute percentage error) and reliability analysis, clarifying the applicable scenarios and potential error sources (such as prediction deviations caused by extreme weather conditions or equipment failures). The final output includes standardized data files (containing all core data for easy reading by the power plant dispatching system) and a visualization report. The visualization report displays predicted power values, confidence intervals, and ramp-up event segments using time-series curves, marking warning levels and key event nodes, and simultaneously presenting the core content of dispatching assistance suggestions. This provides comprehensive and accurate decision-making basis for the formulation of power generation plans for new energy power plants, grid dispatching coordination, and safe and stable operation.

[0181] The aforementioned method for short-term forecasting of renewable energy power by fusing multi-source meteorological data preprocesses the multi-source meteorological data and achieves deep fusion by combining non-negative tensor decomposition. This fully explores the deep correlation features in time, space, and variable dimensions, solving the problems of spatiotemporal misalignment and data redundancy, and improving the integrity and reliability of the fused data. Relying on the Informer architecture and ProbSparse self-attention mechanism, the processing pressure of long-term time-series data is reduced. Combined with a decision-oriented loss function, the method improves the stability of prediction while taking into account the economic rationality of actual scheduling. Through confidence interval calculation, ramp-up event identification, and early warning scheduling suggestion generation, the method breaks through the limitations of single power point prediction, extending from prediction to decision support, enhancing engineering practicality, providing a scientific basis for the safe operation of the power grid and the formulation of power generation plans for power plants, and alleviating the scheduling difficulties caused by the volatility of renewable energy power generation.

[0182] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0183] Based on the same inventive concept, this application also provides a system for short-term forecasting of new energy power by fusing multi-source meteorological data, used to implement the aforementioned method for short-term forecasting of new energy power by fusing multi-source meteorological data. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the system for short-term forecasting of new energy power by fusing multi-source meteorological data provided below can be found in the limitations of the method for short-term forecasting of new energy power by fusing multi-source meteorological data above, and will not be repeated here.

[0184] In one exemplary embodiment, such as Figure 2 As shown, a schematic diagram of the structure of a short-term forecasting system 10 for new energy power based on multi-source meteorological data fusion is provided, including:

[0185] The multi-source data acquisition and standardization module 11 is used to collect multi-source time-series meteorological data covering the target new energy power station area, and to perform spatiotemporal alignment, anomaly removal, missing value supplementation and normalization on the multi-source time-series meteorological data to generate standard multi-source meteorological data; the multi-source time-series meteorological data includes numerical weather prediction grid data, satellite remote sensing observation data and ground meteorological station observation data;

[0186] Data fusion module 12 is used to perform multimodal deep fusion of standard multi-source meteorological data based on tensor decomposition to obtain a multidimensional feature matrix;

[0187] The new energy power prediction module 13 is used to input the multi-dimensional feature matrix into the preset joint prediction model and output the predicted value of new energy power points in the future period. The joint prediction model includes an input layer, a feature extraction layer based on the Informer architecture, a time series modeling layer and a prediction output layer.

[0188] The new energy power prediction result generation module 14 is used to obtain the new energy power prediction result based on the predicted value of new energy power points in the future period.

[0189] Furthermore, the data fusion module 12 is also used for:

[0190] S21. Map the numerical weather forecast grid data in the standard multi-source meteorological data to the geographic coordinates of the target new energy power station using the bilinear interpolation method to obtain the background meteorological vector sequence at the station scale.

[0191] S22. Based on satellite remote sensing observation data, remote sensing estimates of cloud optical thickness and surface incident shortwave radiation intensity of the target station area are obtained by inversion through a simplified radiative transfer model, and satellite-derived feature vector sequence is generated.

[0192] S23. Construct a multidimensional spatiotemporal meteorological tensor; arrange the background meteorological vector sequence, satellite-derived feature vector sequence, and ground meteorological station observation data sequence according to the time dimension, spatial dimension, and meteorological variable dimension to construct a third-order spatiotemporal meteorological tensor. ,in For time step, For spatial location number, For meteorological variables;

[0193] S24. Construct a fusion model based on nonnegative tensor decomposition; define the objective function of the fusion model based on nonnegative tensor decomposition as minimizing the Frobenius norm error between the reconstructed tensor and the original tensor, and add a regularization term constraint. The expression of the objective function is:

[0194] ;

[0195] In the formula, It is a third-order spacetime tensor. For the core tensor, For time pattern matrix, For spatial pattern matrix, For the variable pattern matrix, The regularization coefficient is . Represents the n-modulo product;

[0196] S25. The objective function is solved iteratively by using the alternating least squares method. Two of the factor matrices are fixed and the third factor matrix is ​​updated until convergence, and the converged core tensor is obtained.

[0197] S26. Reconstruct the converged core tensor and the corresponding factor matrix, and flatten the reconstructed tensor data to generate a multidimensional feature matrix.

[0198] Furthermore, the data fusion module 12 is also used for:

[0199] S251. Initialize the time pattern matrix, spatial pattern matrix, and variable pattern matrix;

[0200] S252. Fix the spatial pattern matrix and variable pattern matrix, and update the temporal pattern matrix: calculate the Khatri-Rao product of the third-order spatiotemporal meteorological tensor and the core tensor, solve the update temporal pattern matrix using the least squares method, and set the negative elements in the temporal pattern matrix to zero.

[0201] S253. Fixed-time pattern matrix and variable pattern matrix, update spatial pattern matrix: calculate the Khatri-Rao product of the third-order temporal-space meteorological tensor and the core tensor, use the least squares method to solve for the updated spatial pattern matrix, and set the negative elements in the spatial pattern matrix to zero.

[0202] S254. Fixed time pattern matrix and spatial pattern matrix, update variable pattern matrix: Calculate the Khatri-Rao product of the third-order temporal and spatial meteorological tensor and the core tensor, solve for the update variable pattern matrix using the least squares method, and set the negative elements in the variable pattern matrix to zero.

[0203] S255. Repeat S252 to S254 until the reconstruction error between two adjacent iterations is less than the preset threshold or the maximum number of iterations is reached, and the converged core tensor is obtained.

[0204] Furthermore, the new energy power prediction module 13 is also used for:

[0205] S31. Input the multidimensional feature matrix into the preset joint prediction model, and divide the multidimensional feature matrix into multiple non-overlapping spatiotemporal patch blocks through the input layer;

[0206] S32. Map each spatiotemporal patch block to a high-dimensional embedding vector, and add a learnable positional encoding to each high-dimensional embedding vector to generate a patch embedding sequence.

[0207] S33. Input the patch embedding sequence into the feature extraction layer based on the Informer architecture. The feature extraction layer contains multi-level stacked Informer encoder blocks. Each Informer encoder block uses the ProbSparse self-attention mechanism to sparsely sample long sequence inputs and calculate attention scores for key query points.

[0208] S34. Between the layers of the Informer encoder block, the output of the previous layer is used as the input of the next layer through the self-attention distillation layer operation, while retaining the context information of the long sequence, to obtain the deep feature tensor.

[0209] S35. Input the deep feature tensor into the time series modeling layer and the prediction output layer to generate a prediction of the power points of new energy sources in the future period.

[0210] Furthermore, the new energy power prediction module 13 is also used for:

[0211] S331. Calculate the sparsity measure of the query vector. The expression for the sparsity measure is:

[0212] ;

[0213] In the formula, For the sparsity measure of the query vector, For the first A query vector, For a set of key vectors, The length of the key vector. For vector dimensions, The average of all query-key dot products;

[0214] S332. Sort all query vectors in descending order according to sparsity metric, and select the top-ranked... The query vectors are used as a sparse query set;

[0215] S333. Calculate the self-attention output based on the sparse query set to obtain the attention score.

[0216] Furthermore, the new energy power prediction module 13 is also used for:

[0217] S361. During training, the joint prediction model uses a multi-dimensional feature matrix as the model input and the corresponding historical actual new energy power value as the supervision label.

[0218] S362. Define the basic prediction loss as the mean square error between the predicted power point sequence output by the model and the true power sequence.

[0219] S363. Define asymmetric economic penalty loss, which is used to apply different penalty weights to positive and negative prediction errors.

[0220] S364. The weighted sum of the basic forecast loss and the asymmetric economic penalty loss yields the decision-oriented loss function.

[0221] S365. Minimize the loss function using the backpropagation algorithm to train and optimize the parameters of the joint prediction model, thus obtaining a well-trained joint prediction model.

[0222] Furthermore, the new energy power prediction result generation module 14 is also used for:

[0223] S41. Based on the predicted value of new energy power points for future periods output by the joint prediction model, and combined with the prediction error distribution characteristics learned by the model during training, calculate the confidence interval of the predicted value at each prediction time.

[0224] S42. Compare the upper and lower boundaries of the predicted confidence interval with the pre-set power risk level threshold to determine the warning level;

[0225] S43. Based on the time series curve of the predicted power point of new energy in the future period, identify the ramping event segments with rapid power decline and rapid rise, calculate the start time, duration and ramping rate of the ramping event, and obtain ramping event data.

[0226] S44. Based on the warning level and ramp event data, and combined with the current operating status of the power grid, generate corresponding dispatching assistance suggestions from the preset dispatching rule base;

[0227] S45. Generate new energy power prediction results by combining the predicted power points of new energy sources in the future, the confidence interval of the predicted values, the early warning level, the ramp-up event data, and the scheduling assistance suggestions.

[0228] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the multi-source meteorological data fusion method for short-term forecasting of new energy power as described above.

[0229] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0230] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for short-term prediction of renewable energy power through multi-source meteorological data fusion, characterized in that: Includes the following steps: S1. Collect multi-source time-series meteorological data covering the target new energy power station area, preprocess the multi-source time-series meteorological data, and generate standard multi-source meteorological data; S2. Perform multimodal deep fusion of standard multi-source meteorological data based on tensor decomposition to obtain a multidimensional feature matrix; S3. Input the multidimensional feature matrix into the preset joint prediction model and output the predicted value of new energy power points for future periods; wherein the joint prediction model includes an input layer, a feature extraction layer based on the Informer architecture, a time series modeling layer and a prediction output layer; S4. Based on the predicted value of new energy power points in the future period, obtain the new energy power prediction result.

2. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 1, characterized in that: Multi-source time-series meteorological data include numerical weather prediction grid data, satellite remote sensing observation data, and ground meteorological station observation data. Preprocessing methods include spatiotemporal alignment, anomaly removal, missing value supplementation, and normalization.

3. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 2, characterized in that: Step S2 specifically includes: S21. Map the numerical weather forecast grid data in the standard multi-source meteorological data to the geographic coordinates of the target new energy power station using the bilinear interpolation method to obtain the background meteorological vector sequence at the station scale. S22. Based on satellite remote sensing observation data, remote sensing estimates of cloud optical thickness and surface incident shortwave radiation intensity of the target station area are obtained by inversion through a simplified radiative transfer model, and satellite-derived feature vector sequence is generated. S23. Constructing a multidimensional spatiotemporal meteorological tensor: Arrange the background meteorological vector sequence, satellite-derived feature vector sequence, and ground meteorological station observation data sequence according to the time dimension, spatial dimension, and meteorological variable dimension to construct a third-order spatiotemporal meteorological tensor. , For the set of real numbers, This is the time step, corresponding to the length of the entire time series data, and should be consistent with the time resolution of the preprocessed data. This refers to the spatial location data, encompassing all ground meteorological station observation points deployed inside and around the target new energy power station, as well as the interpolated core coordinate points of the station. As a meteorological variable dimension, it integrates all core meteorological variables from the three types of data; S24. Construct a fusion model based on nonnegative tensor decomposition; define the objective function of the fusion model based on nonnegative tensor decomposition as minimizing the Frobenius norm error between the reconstructed tensor and the original tensor, and add a regularization term constraint. The expression of the objective function is: ； In the formula, It is a third-order spacetime tensor. For the core tensor, For time pattern matrix, For spatial pattern matrix, For the variable pattern matrix, The regularization coefficient is . Represents the n-modulo product; S25. The objective function is solved iteratively by using the alternating least squares method. Two of the factor matrices are fixed and the third factor matrix is ​​updated until convergence, and the converged core tensor is obtained. S26. Reconstruct the converged core tensor and the corresponding factor matrix, and flatten the reconstructed tensor data to generate a multidimensional feature matrix.

4. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 3, characterized in that: Step S25 specifically includes: S251. Initialize the time pattern matrix, spatial pattern matrix, and variable pattern matrix; S252. Fix the spatial pattern matrix and variable pattern matrix, and update the temporal pattern matrix: calculate the Khatri-Rao product of the third-order spatiotemporal meteorological tensor and the core tensor, solve the update temporal pattern matrix using the least squares method, and set the negative elements in the temporal pattern matrix to zero. S253. Fixed-time pattern matrix and variable pattern matrix, update spatial pattern matrix: calculate the Khatri-Rao product of the third-order temporal-space meteorological tensor and the core tensor, use the least squares method to solve for the updated spatial pattern matrix, and set the negative elements in the spatial pattern matrix to zero. S254. Fixed time pattern matrix and spatial pattern matrix, update variable pattern matrix: Calculate the Khatri-Rao product of the third-order temporal and spatial meteorological tensor and the core tensor, solve for the update variable pattern matrix using the least squares method, and set the negative elements in the variable pattern matrix to zero. S255. Repeat steps S252 to S254 until the reconstruction error between two adjacent iterations is less than the preset threshold or the maximum number of iterations is reached, and obtain the converged core tensor.

5. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 1, characterized in that: The input layer in the joint prediction model is responsible for receiving the multidimensional feature matrix and adjusting it into the input format required by the model. The feature extraction layer based on the Informer architecture consists of several stacked Informer encoder blocks. Each Informer encoder block is composed of a sparse self-attention layer, a feedforward neural network, layer normalization, and residual connections. The sparse self-attention layer adopts a probabilistic sampling strategy, which divides the time step into multiple intervals through bucketing. Within each interval, a portion of the time steps are randomly sampled to calculate attention. The number of attention heads and the dropout rate are set. The number of attention heads is used to extract attention features of different dimensions in parallel, and the dropout rate is used to suppress overfitting. The feedforward neural network consists of two fully connected layers with a GELU activation function inserted in between. Layer normalization is used to standardize the input features of each layer. The temporal modeling layer is connected to the feature extraction layer, and several Informer decoder blocks are used. The Informer decoder blocks contain masked self-attention layers and cross-attention layers. The prediction output layer consists of a fully connected layer, with the number of neurons corresponding to the length of the future prediction time step. A linear activation function is used, and the predicted value of the new energy power point for the future period is directly output.

6. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 5, characterized in that: Step S3 specifically includes: S31. Input the multidimensional feature matrix into the preset joint prediction model, and divide the multidimensional feature matrix into multiple non-overlapping spatiotemporal patch blocks through the input layer; S32. Map each spatiotemporal patch block to a high-dimensional embedding vector, and add a learnable positional encoding to each high-dimensional embedding vector to generate a patch embedding sequence. S33. Input the patch embedding sequence into the feature extraction layer based on the Informer architecture; S34. Between the layers of the Informer encoder block, the output of the previous layer is used as the input of the next layer through the self-attention distillation layer operation, while retaining the context information of the long sequence, to obtain the deep feature tensor. S35. Input the deep feature tensor into the time series modeling layer and the prediction output layer to generate a prediction of the power points of new energy sources in the future period.

7. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 5, characterized in that: The joint prediction model is trained and optimized using a decision-oriented loss function. The training process of the joint prediction model includes: S361. During training, the joint prediction model uses a multi-dimensional feature matrix as the model input and the corresponding historical actual new energy power value as the supervision label. S362. Define the basic prediction loss as the mean square error between the predicted power point sequence output by the model and the true power sequence. S363. Define asymmetric economic penalty loss, which is used to apply different penalty weights to positive and negative prediction errors. S364. The weighted sum of the basic forecast loss and the asymmetric economic penalty loss yields the decision-oriented loss function. S365. Minimize the loss function using the backpropagation algorithm to train and optimize the parameters of the joint prediction model, thus obtaining a well-trained joint prediction model.

8. The method for short-term prediction of new energy power based on multi-source meteorological data fusion according to claim 7, characterized in that: The expression for the decision-oriented loss function is as follows: ； in, The total loss function value is determined by the decision-making process. The weighting coefficients for the basic predicted loss; and These are the basic forecast loss and the asymmetric economic penalty loss, respectively. The expression for the asymmetric economic penalty loss is as follows: ； in, The penalty weight for positive prediction error, The penalty weights for negative prediction errors are both non-negative values, and their values ​​are determined by calibration using the historical economic operation data of the target new energy power station. Used to extract positive prediction errors and mask negative errors. This is used to extract negative prediction errors and mask positive errors, ensuring that the penalty calculations for the two types of errors are independent of each other; The number of time steps in a single batch of training samples; For the first The historical actual new energy power value at each time step; The first output of the model Predicted power values ​​at each time step.

9. A method for short-term forecasting of new energy power based on multi-source meteorological data fusion according to claim 8, characterized in that: Step S4 specifically includes: S41. Based on the predicted value of new energy power points for future periods output by the joint prediction model, and combined with the prediction error distribution characteristics learned by the model during training, calculate the confidence interval of the predicted value at each prediction time. S42. Compare the upper and lower boundaries of the predicted confidence interval with the pre-set power risk level threshold to determine the warning level; S43. Based on the time series curve of the predicted power point of new energy in the future period, identify the ramping event segments with rapid power decline and rapid rise, calculate the start time, duration and ramping rate of the ramping event, and obtain ramping event data. S44. Based on the warning level and ramp event data, and combined with the current operating status of the power grid, generate corresponding dispatching assistance suggestions from the preset dispatching rule base; S45. Generate new energy power prediction results based on the predicted power points of new energy sources in the future period, the confidence interval of the predicted values, the early warning level, the ramp-up event data, and the scheduling assistance suggestions.

10. A short-term forecasting system for new energy power based on multi-source meteorological data fusion, characterized in that: For implementing the short-term forecasting method for new energy power based on multi-source meteorological data fusion as described in any one of claims 1-9, the system comprises: The multi-source data acquisition and standardization module is used to collect multi-source time-series meteorological data covering the target new energy power station area, and to preprocess the multi-source time-series meteorological data to generate standard multi-source meteorological data. The data fusion module is used to perform multimodal deep fusion of standard multi-source meteorological data based on tensor decomposition to obtain a multidimensional feature matrix; The new energy power prediction module is used to input a multi-dimensional feature matrix into a preset joint prediction model and output the predicted value of new energy power points for future periods. The joint prediction model includes an input layer, a feature extraction layer based on the Informer architecture, a time series modeling layer, and a prediction output layer. The new energy power prediction result generation module is used to obtain the new energy power prediction result based on the predicted value of new energy power points in the future period.

Images