Wind power prediction method and system based on multivariate combination prediction model, and medium

Abstract

The invention provides a wind power prediction method and system based on a multivariate combination prediction model and a medium, and relates to the technical field of machine learning, and the method comprises the steps: obtaining a historical wind power data time sequence and a historical numerical weather forecast data time sequence of a wind power plant; analyzing a nonlinear dependency relationship by using correlation, and optimally training a multi-modal parallel time sequence prediction model by using a variational mode decomposition algorithm in combination with a sliding window length; and based on the final sliding window length, constructing an input sample, inputting the input sample into a multi-mode parallel time sequence prediction model, outputting future prediction values of a plurality of intrinsic mode function components, and carrying out summation to obtain a wind power prediction result. According to the method and the device, the technical problem of insufficient wind power prediction precision caused by strong volatility and nonlinearity of a wind power sequence and limited learning ability of a single prediction model in the prior art can be solved, and the wind power prediction precision is improved by combining variational mode decomposition with a machine learning model for prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a wind power prediction method and system based on a multi-element combination prediction model and a medium. BACKGROUND

[0002] Wind energy, as a clean and renewable energy with low generation cost and no pollution in the generation process, has gradually become the main force for the sustainable development of energy and environment. Common wind power prediction methods can be divided into physical methods, statistical methods and artificial intelligence methods. However, the physical method has a huge amount of calculation, high modeling difficulty and cannot be used for real-time prediction; the statistical method finds the mapping relationship between the historical weather data collected by the numerical weather prediction system and the historical power data of the wind farm, but due to the nonlinear characteristics of the wind power sequence, the prediction result often has a large deviation when facing complex wind power data, which is difficult to meet the accuracy requirement of wind power prediction; a large number of deep learning models have been applied in the field of wind power prediction, but the strong volatility and nonlinearity of wind power sequence and the limited learning ability of single prediction model make it difficult to effectively extract features and mine information of wind power sequence, resulting in insufficient prediction accuracy.

[0003] In summary, the prior art has the technical problem of insufficient wind power prediction accuracy due to the strong volatility and nonlinearity of wind power sequence and the limited learning ability of single prediction model. SUMMARY

[0004] The purpose of the present application is to provide a wind power prediction method and system based on a multi-element combination prediction model and a medium, to solve the technical problem of insufficient wind power prediction accuracy due to the strong volatility and nonlinearity of wind power sequence and the limited learning ability of single prediction model in the prior art.

[0005] In order to achieve the above purpose, the present application provides a wind power prediction method and system based on a multi-element combination prediction model and a medium.

[0006] In a first aspect, the application provides a wind power prediction method based on a multi-element combination prediction model, which is implemented by a wind power prediction system based on a multi-element combination prediction model, wherein the wind power prediction method based on the multi-element combination prediction model comprises: obtaining historical wind power data time series and historical numerical weather forecast data time series of a wind farm in a first historical time zone; based on the historical wind power data time series and the historical numerical weather forecast data time series, using a correlation analysis to input a nonlinear dependence relationship between a variable and output power, using a variational mode decomposition algorithm combined with a sliding window length to optimize training of a multi-mode parallel time series prediction model; based on the final sliding window length obtained by training, constructing an input sample to be input to the multi-mode parallel time series prediction model, summing up future prediction values of a plurality of intrinsic mode function components to obtain a final wind power prediction result.

[0007] Optionally, based on the historical wind power data time series, a historical-future power correlation cycle iteration analysis is performed under a time series neighbor, a first historical window length and a first future window length under a stable correlation coefficient distribution are determined; a historical measured numerical weather forecast data time series corresponding to the historical numerical weather forecast data time series is collected, a prediction accuracy analysis of the weather forecast is performed, and a second window length with a prediction accuracy greater than a preset accuracy threshold is identified; the first future window length is optimized with the second window length, and a sliding window length is generated with the first historical window length and the optimized first future window length; based on the sliding window length, the historical wind power data time series is sequentially segmented and then subjected to variational mode decomposition, and the multi-mode parallel time series prediction model is trained in combination with the mode decomposition result and the historical numerical weather forecast data time series corresponding to the sliding window.

[0008] Optionally, the historical wind power data and the future wind power data in the sliding window segmented by the variational mode decomposition algorithm are subjected to mode decomposition respectively to obtain K historical intrinsic mode function components and K future intrinsic mode function components; the K historical intrinsic mode function components and the future numerical weather forecast data combination are used as input training samples, the K future intrinsic mode function components are used as output supervision values, the multi-mode parallel time series prediction model is generated based on the sliding window length by iterative training.

[0009] Optionally, during the iterative training process based on the sliding window length, the sliding window is adaptively adjusted in conjunction with the training loss. Specifically, this includes: if the training loss is less than a preset loss threshold for N consecutive times, increasing the future window length in the sliding window length by a preset step size; if the training loss is greater than the preset loss threshold for N consecutive times, decreasing the future window length in the sliding window length by a preset step size; wherein, the prediction loss of the finally generated multimodal parallel temporal prediction model must be less than or equal to the preset loss threshold.

[0010] Optionally, the multimodal parallel temporal prediction model includes a shared feature extraction layer, a temporal dependency processing layer, and a modality-specific attention prediction layer connected in sequence; the shared feature extraction layer consists of at least one layer of dilated causal convolution, used to extract common temporal features from the input data of all modalities; the temporal dependency processing layer consists of a gated recurrent unit network, used to capture long-term dependencies in the common temporal features; the modality-specific attention prediction layer is used to assign attention weights to each modality and output the future prediction value of each modality.

[0011] Optionally, during mode decomposition, variational mode decomposition is performed for each candidate K value according to the search range of the preset number of modes K, and the sum of the envelope entropy of all decomposed modal components is calculated. The optimal number of modes K is determined based on the first inflection point of the sum of envelope entropy changing with the K value for final decomposition.

[0012] Optionally, based on the historical window length in the final sliding window length, starting from the current time, target historical power generation data is collected; target future numerical weather forecast data is collected with the future window length in the final sliding window length as a constraint; and the input sample is constructed using the target historical power generation data and the target future numerical weather forecast data.

[0013] Optionally, the weather forecast accuracy is analyzed using historical data, with the future window length in the final sliding window length as a constraint; the impact of weather forecast error on wind power is analyzed under the weather forecast accuracy; and the final wind power forecast result is labeled with error based on the error impact.

[0014] Secondly, this application also provides a wind power prediction system based on a multivariate combined prediction model, used to execute the wind power prediction method based on a multivariate combined prediction model as described in the first aspect. The wind power prediction system based on the multivariate combined prediction model includes: a time-series data acquisition module, used to acquire the time series of historical wind power data and historical numerical weather forecast data of a wind farm in a first historical time zone; a model construction module, used to, based on the time series of the historical wind power data and the time series of the historical numerical weather forecast data, utilize correlation analysis to determine the nonlinear dependency between input variables and output power, and utilize variational mode decomposition algorithm combined with sliding window length optimization to train a multimodal parallel time-series prediction model; and a wind power prediction module, used to, based on the final sliding window length obtained from the training, construct input samples and input them into the multimodal parallel time-series prediction model, summing the future predicted values ​​of several intrinsic mode function components to obtain the final wind power prediction result.

[0015] Thirdly, a computer-readable storage medium storing a computer program that, when executed, implements the steps of the wind power prediction method based on a multivariate combined prediction model as described in any of the first aspects above.

[0016] One or more technical solutions provided in this application have at least the following technical effects or advantages: By acquiring historical wind power data and historical numerical weather prediction data from wind farms in the first historical time zone, and based on these data, the nonlinear dependence between input variables and output power is analyzed using correlation analysis. A multimodal parallel time series prediction model is then trained using variational mode decomposition (MODED) combined with a sliding window length optimization. Based on the final sliding window length obtained during training, input samples are constructed and fed into the multimodal parallel time series prediction model. The future predicted values ​​of several intrinsic mode function components are summed to obtain the final wind power prediction result. In other words, by using variational mode decomposition to decompose and stabilize the wind power data sequence, and then constructing a multimodal parallel time series prediction model based on machine learning for wind power prediction, the accuracy of wind power prediction is improved.

[0017] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the wind power prediction method based on a multivariate combined prediction model proposed in this application.

[0020] Figure 2 This is a schematic diagram of the wind power prediction system based on a multivariate combined prediction model according to this application.

[0021] Figure labeling: Time series data acquisition module 11, model building module 12, wind power prediction module 13. Detailed Implementation

[0022] This application addresses the technical problem in existing technologies where the strong volatility and nonlinearity of wind power sequences, coupled with the limited learning ability of single prediction models, lead to insufficient wind power prediction accuracy. By utilizing variational mode decomposition to decompose and stabilize wind power data sequences, and constructing a multimodal parallel time-series prediction model based on machine learning, the accuracy of wind power prediction is improved.

[0023] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0024] Example 1, please refer to the appendix. Figure 1 This application provides a wind power prediction method based on a multivariate combined prediction model, wherein the wind power prediction method based on the multivariate combined prediction model is applied to a wind power prediction system based on the multivariate combined prediction model, and the wind power prediction method based on the multivariate combined prediction model specifically includes the following steps: Obtain the time series of historical wind power data and historical numerical weather forecast data of the wind farm in the first historical time zone.

[0025] Specifically, the historical wind power data time series of the wind farm is obtained in the first historical time zone. This time series consists of the actual active power output values ​​recorded and stored at fixed time intervals within the first historical time zone. The historical numerical weather prediction data time series is the historical weather forecast data corresponding to the wind farm location and hub height within the first historical time zone. The first historical time zone is a specific, past continuous period. To accurately define the data range used for model training and validation, and to ensure the time zone correspondence between the historical wind power data time series and the historical numerical weather prediction data time series, the historical power data and numerical weather prediction data of the wind farm are aligned according to timestamps, ensuring a one-to-one correspondence between meteorological data and power data at each time point. By precisely aligning the historical wind power data and numerical weather prediction data, ensuring a one-to-one correspondence between power output and meteorological data at each time point, the influence of meteorological factors such as wind speed and direction on wind power is captured, thereby improving the accuracy of the model in predicting wind power.

[0026] Data preprocessing, including data cleaning and normalization, was performed on historical wind power data and historical numerical weather prediction data. A lateral processing method was employed to remove and correct bad data in the historical power dataset, using formula x. t =(x t-1 +x t+1 ) / 2, where x t This represents the data that needs to be corrected, x. t-1 and x t+1 These represent the data from the previous and next time points, respectively. Since historical wind power data and historical numerical weather prediction data have different dimensions and orders of magnitude, these input data need to be normalized to eliminate data dimensions and accelerate the convergence of the training network. The min-max normalization method is used to map the data to the range [0,1], where x1 = (xx... min ) / (x max -x min ), where x1 represents the mapped value, x min and x max These are the minimum and maximum values ​​of the sample data, respectively.

[0027] Based on the historical wind power data time series and the historical numerical weather forecast data time series, the nonlinear dependency between input variables and output power is analyzed using correlation analysis, and a multimodal parallel time series prediction model is trained by using variational mode decomposition algorithm combined with sliding window length optimization.

[0028] Furthermore, this application also includes the following steps: based on the historical wind power data time series, performing iterative analysis of the historical-future power correlation under adjacent time series, determining the first historical window length and the first future window length under a stable distribution of correlation coefficients; collecting the historical measured numerical weather forecast data time series corresponding to the historical numerical weather forecast data time series, performing weather forecast prediction accuracy analysis, and identifying a second window length with a prediction accuracy greater than a preset accuracy threshold; optimizing the first future window length with the second window length, and generating a sliding window length with the first historical window length and the optimized first future window length; based on the sliding window length, performing variational mode decomposition on the historical wind power data time series after sequential sliding segmentation, and training the multimodal parallel time series prediction model by combining the mode decomposition results with the historical numerical weather forecast data time series of the corresponding sliding window.

[0029] Furthermore, this application also includes the following steps: performing mode decomposition on historical wind power data and future wind power data within the sliding window after sliding segmentation using the variational mode decomposition algorithm, obtaining K historical intrinsic mode function components and K future intrinsic mode function components; using the combination of the K historical intrinsic mode function components and future numerical weather prediction data as input training samples, and using the K future intrinsic mode function components as output supervision values, performing iterative training based on the sliding window length to generate the multimodal parallel time series prediction model.

[0030] Specifically, based on historical wind power data time series, a cyclical iterative analysis of the historical-future power correlation under adjacent time series is performed. This involves iteratively analyzing the historical wind power data time series to quantify the correlation strength between power data of different past time lengths and power data of different future time lengths, calculating their correlation coefficients, and finding the optimal solution by iteratively combining different time lengths. The correlation coefficient is a statistic used to measure the strength and direction of the linear relationship between two variables, with a value range of [-1, 1]. The larger the absolute value, the stronger the correlation. The historical power time series is iteratively analyzed. For example, the test uses data from the past 1 hour, 2 hours, ..., 24 hours to predict the power for the next 15 minutes, 30 minutes, ..., 4 hours, respectively, and calculates the correlation coefficient for each combination. A sliding window method is used to select different time windows and calculate the correlation coefficients between them. The correlation under each window is analyzed until a stable correlation coefficient greater than a preset coefficient (e.g., 0.85) and a correlation coefficient greater than a preset proportion (e.g., 90%) is found. The first historical window length represents the time length of the historical data used to train the model, and the first future window length represents the time length of the future data to be predicted.

[0031] The process involves collecting historical numerical weather forecast data along with corresponding historical observed numerical weather forecast data, i.e., the actual weather data observed. By comparing the errors between historical meteorological data and observed meteorological data, it's determined which time periods have prediction accuracy greater than a preset threshold, thus determining the length of the second window. The sliding window is essentially a time window; it uses a specific amount of historical data to predict a specific amount of future data, resulting in more accurate predictions.

[0032] Based on the second window length, the first future window length is optimized, i.e., a more suitable time window is used as the first future window length. The first future window length obtained in the first step and the optimized first future window length are used as the sliding window length to sequentially slide and segment the historical wind power data time series. For each data sample generated by the sliding segmentation, variational mode decomposition (VMD) is used to perform mode decomposition on the historical and future wind power data respectively, obtaining K historical intrinsic mode function (IMF) components and K future IMF components. The complete process of mode decomposition is as follows: First, initialization is performed, setting the initial mode components, center frequency, and Lagrange multipliers to start the iteration process. In the iteration phase, each mode component is updated in each iteration. When updating the mode component, the current remaining signal (i.e., the original signal minus the sum of other mode components) is added to the influence of the Lagrange operator, and adjusted through a frequency-related function to obtain the new mode component. Simultaneously, the center frequency of each modal component is updated. The center frequency is calculated based on the energy distribution of that modal component in the frequency domain, i.e., by integrating the frequency-weighted energy to determine the new center frequency. The Lagrange operator is updated to reflect the difference between the current modal component and the original signal. The update of the Lagrange operator takes noise tolerance into account, balancing the accuracy of signal reconstruction and the impact of noise by adjusting parameters. The iteration termination condition is checked. The termination condition is based on the amount of change of the modal components in consecutive iterations. If the sum of the changes of all modal components is small relative to the previous iteration, below a preset threshold, the iteration stops, and multiple subsequences after decomposition are output; otherwise, the iterative update steps continue to be repeated until the termination condition is met.

[0033] Historical intrinsic mode function (IMF) components are each frequency component obtained from variational mode decomposition (MODE Decomposition), representing different periodic or trend components in the wind power sequence. Through variational mode decomposition, historical wind power data can be decomposed into multiple historical IMF components, each representing a different characteristic of the original data. Future IMF components are components obtained from future wind power data after variational mode decomposition, used to predict the changing characteristics of future wind power.

[0034] K historical intrinsic mode function (IMF) components and future numerical weather prediction data are combined as input training features to form training samples. K future IMF components are used as output supervision values ​​to train a multimodal parallel time series prediction model. Historical data within a sliding window length is used for prediction, and the model's weights and parameters are gradually adjusted to minimize prediction error. The model undergoes multiple iterations of training based on training samples of different sliding window lengths, continuously optimizing the weights until the training results converge, yielding the final multimodal parallel time series prediction model. After training, the multimodal parallel time series prediction model can predict new historical wind power data and meteorological data, outputting future wind power prediction results. By simultaneously performing mode decomposition on the input and output, the complex original power prediction problem is transformed into multiple relatively simple, homogeneous subsequence prediction problems, learning and capturing components at different time scales, significantly improving the accuracy and stability of predictions.

[0035] Furthermore, this application also includes the following steps: during iterative training based on the sliding window length, adaptive adjustment of the sliding window is performed in conjunction with the training loss, specifically including: if the training loss is less than a preset loss threshold for N consecutive times, the future window length in the sliding window length is increased by a preset step size; if the training loss is greater than the preset loss threshold for N consecutive times, the future window length in the sliding window length is decreased by a preset step size; wherein, the prediction loss of the finally generated multimodal parallel temporal prediction model must be less than or equal to the preset loss threshold.

[0036] Specifically, during the iterative training of the multimodal parallel time-series prediction model, a moving average training loss value is continuously monitored to avoid misjudgments caused by random fluctuations in a single loss. When the training loss is detected to be consistently below the preset loss threshold for N consecutive times, it indicates that the current multimodal parallel time-series prediction model has sufficient capacity to learn more challenging tasks. Automatically increasing the future window length by a preset step size means that the multimodal parallel time-series prediction model needs to learn to predict future power modes over a longer period, thus expanding its prediction scope.

[0037] When the training loss exceeds the preset loss threshold N times consecutively, it indicates that the current learning task is too difficult, possibly due to the accumulation of time-series errors in numerical weather forecast data or a weakening of the long-range dependence of the power sequence itself. N is a positive integer greater than or equal to 1. To prevent instability in the training of the multimodal parallel time-series prediction model, the future window length is automatically reduced according to a preset step size, reducing the difficulty of the learning task and allowing the multimodal parallel time-series prediction model to refocus on accurate short-term predictions, thus solidifying its learning foundation.

[0038] The ultimate goal of the entire training process is to generate a final model whose predictive loss must be less than or equal to a preset loss threshold. Adaptive tuning continues until the performance of the multimodal parallel temporal prediction model stabilizes and meets the target. The final generated multimodal parallel temporal prediction model has a future window length that, given the data, can achieve the maximum predictable duration while still meeting accuracy requirements.

[0039] For example, suppose the initial sliding window is 6 hours of historical data and 2 hours of future data, with a preset loss threshold of 0.025, a successive iteration count (N) of 5, and a preset step size of 30 minutes. The multimodal parallel time series prediction model is trained with windows of 6 hours and 2 hours. At a certain stage, its evaluation loss for 5 consecutive iterations is: 0.022, 0.021, 0.023, 0.020, and 0.022, all < 0.025, triggering adjustment and increasing the future window length from 2 hours to 2.5 hours. Subsequent training samples will be built based on the historical data of 6 hours and the future data of 2.5 hours, and the multimodal parallel time series prediction model begins to learn longer-term predictions. Alternatively, the multimodal parallel time series prediction model is trained with windows of 6 hours and 2 hours. As the forecast duration increased, the time-series error and sequence uncertainty of the numerical weather prediction data increased, leading to a rise in training loss. The values ​​for five consecutive periods were 0.028, 0.031, 0.029, 0.030, and 0.027, all exceeding 0.025, triggering adjustment by reducing the future window length from 3 hours to 2.5 hours. This reduced the task difficulty, and the loss of the multimodal parallel time-series prediction model began to decline and gradually stabilize. After several such dynamic adjustments, the model stably maintained the training loss between 0.024 and 0.025 with a historical 6-hour and a future 2-hour and 45-minute configuration, meeting the requirement of a final loss ≤ 0.025. Therefore, the final multimodal parallel time-series prediction model was fixed at a future window length of 2 hours and 45 minutes.

[0040] While ensuring prediction accuracy, this study explores the boundaries of the predictive capability of multimodal parallel time-series prediction models, identifying the longest time range within which reliable predictions can be made given data and quality. This provides more ample decision-making time for power grid dispatching. Adaptively adjusting the window length avoids using inappropriate time windows during each training iteration, thereby improving training efficiency and enabling the model to reach the expected loss level more quickly.

[0041] Furthermore, this application also includes the following steps: the multimodal parallel temporal prediction model includes a shared feature extraction layer, a temporal dependency processing layer, and a modality-specific attention prediction layer connected in sequence; the shared feature extraction layer is composed of at least one layer of dilated causal convolution, used to extract common temporal features from the input data of all modalities; the temporal dependency processing layer is composed of a gated recurrent unit network, used to capture long-term dependencies in the common temporal features; the modality-specific attention prediction layer is used to assign attention weights to each modality and output the future prediction value of each modality.

[0042] Specifically, the multimodal parallel temporal prediction model processes time-series data from different data sources and predicts data from multiple modalities in parallel. Each modality represents different input features or data types, enabling the simultaneous processing of multiple inputs and combining their respective characteristics for prediction to improve accuracy. The multimodal parallel temporal prediction model consists of a shared feature extraction layer, a temporal dependency processing layer, and a modality-specific attention prediction layer connected in sequence.

[0043] The shared feature extraction layer is responsible for initially mining common, fundamental time-series patterns from all input modalities. This is used to extract global temporal information from the input data and share the feature extraction capability across multiple modalities, ensuring that the multimodal parallel temporal prediction model can identify common information between different modalities. The input data from all modalities are concatenated along the feature dimension to form a multi-channel temporal tensor. This multi-channel temporal tensor is fed into one or more stacked dilated causal convolutional layers. The first convolutional layer scans the input with a basic receptive field, and subsequent layers expand the receptive field exponentially with increasing dilation factors. Dilated causal convolution is a convolutional neural network technique designed to capture long-term dependencies when performing causal modeling on sequential data. It increases the receptive field by expanding the interval of the convolutional kernels while ensuring causality, meaning each output depends only on the previous input.

[0044] The temporal dependency processing layer uses gated recurrent units (ROUs) to process the input temporal data and capture long-term temporal dependencies. It receives abstract features from the previous layer and further learns the temporal evolution of these features. The gated recurrent unit network is a variant of a recurrent neural network. Similar to LSTM, it controls the flow of information through a mechanism of updating and resetting gates, mitigating the gradient vanishing / exploding problem of traditional neural network models, thereby learning long-term dependencies in long sequences. Gated recurrent unit networks are generally simpler in structure, have fewer parameters, and train faster than LSTMs. The gated recurrent unit network processes these features sequentially step by step. For each time step, the gated recurrent unit network, based on the current input and its internal hidden state (containing condensed information from all past time steps), determines how much old information to retain and how much new information to add through a gating mechanism, updating its hidden state so that the final hidden state of the gated recurrent unit network contains the long-term contextual information of the entire input sequence.

[0045] The modality-specific attention prediction layer assigns different weights to each modality through an attention mechanism, thereby enabling differentiated predictions based on the characteristics of each modality. The attention mechanism automatically adjusts the importance of each modality in the final prediction, focusing more on important modalities and ignoring irrelevant or noisy data. K different attention weights are generated for each of the K future intrinsic mode function (IMF) components to be predicted. The attention mechanism calculates the correlation score between the hidden state of each historical time step in the output sequence of the gated recurrent unit network (GRN) and the current prediction task. These scores are normalized to a weight distribution, i.e., the attention weights, using a softmax function. A weighted summation of all hidden states of the GRN network is then performed to obtain a context vector. This vector is adaptive; for predicting high-frequency future IMF components, it focuses on recent sharp fluctuations; for predicting low-frequency future IMF components, it focuses on the gradual trend within the entire window. The predicted values ​​for each modality are weighted and summed to obtain the final wind power prediction result. By extracting global features through a shared feature extraction layer, capturing long-term dependencies through a temporal dependency processing layer, and adjusting the weights of a modality-specific attention prediction layer, the advantages of multimodal data are fully utilized, thereby improving the accuracy of wind power prediction.

[0046] Furthermore, this application also includes the following steps: when performing mode decomposition, according to the search range of the preset number of modes K, for each candidate K value, variational mode decomposition is performed, and the sum of the envelope entropy of all decomposed modal components is calculated. Based on the first inflection point of the sum of envelope entropy changing with the K value, the optimal number of modes K is determined for final decomposition.

[0047] Specifically, a search range for K values ​​is defined, which represents the range of selectable modes. Generally, the range of K values ​​is determined by the properties of the data and the required decomposition precision. For each candidate K value, a variational mode decomposition algorithm is executed to decompose the original time-series data into K intrinsic mode function (EMF) components. The envelope entropy of the K EMF components is calculated and summed to obtain the total envelope entropy corresponding to the current K value. Envelope entropy is an index used to quantify the randomness and sparsity of a signal. It is obtained by performing a Hilbert transform on the signal to obtain its envelope signal; normalizing the envelope signal to obtain a probability distribution sequence; and calculating the information entropy of this distribution. The higher the entropy value, the more chaotic the signal, the greater the uncertainty, and the more complex the information it contains, potentially including noise; the lower the entropy value, the more regular and pure the signal.

[0048] Plot the total envelope entropy as a function of the K value and observe the downward trend of the curve. Initially, as the K value increases, the decomposition becomes more refined, separating noise and fluctuations at different scales mixed in with the main modes, making each component more regular, i.e., reducing entropy, so the total entropy decreases rapidly. When the K value reaches a certain point, meaningful physical modes have been basically separated. Further increasing the K value may only separate some meaningless minor fluctuations or noise from the existing components. These new components themselves are still very chaotic, causing the rate of decrease in the total entropy to decrease sharply. This turning point is the first inflection point. The first inflection point is the point where the downward trend of the curve first shows a significant slowdown, meaning that the improvement in information purity brought by adding another mode is negligible. Continuing decomposition may only be decomposing noise rather than meaningful physical modes.

[0049] The K value corresponding to the inflection point is determined as the optimal number of modes in this decomposition, and this K value is used for the final variational mode decomposition. For example, suppose a power sequence of 24 data points over 6 hours is decomposed, with a preset search range K=[2,3,4,5,6]. When K=2, the variational mode decomposition yields intrinsic mode function component 1 and intrinsic mode function component 2, with envelope entropy 1 calculated to be 4.2, envelope entropy 2 to be 3.8, and total envelope entropy to be 8.0. When K=3, an additional intrinsic mode function component 3 is added, with envelope entropy 1 calculated to be 3.9, envelope entropy 2 to be 3.5, and envelope entropy 3 to be 3.0, total envelope entropy to be 7.4, and entropy decreases by 0.6. When K=4, an additional intrinsic mode function component 4 is added, with envelope entropy 1 calculated to be 3.8, and envelope entropy 2 to be 3. 4. Envelope entropy 3 is 2.9, envelope entropy 4 is 2.8, and the total envelope entropy is 6.9, with the entropy sum decreasing by 0.5. When K=5, an additional intrinsic mode function component 5 is added, and the calculated envelope entropy 1 is 3.8, envelope entropy 2 is 3.4, envelope entropy 3 is 2.9, envelope entropy 4 is 2.8, and envelope entropy 5 is 2.7, with the total envelope entropy being 6.7, and the entropy sum decreasing by 0.2, a significantly smaller decrease. When K=6, an additional intrinsic mode function component 6 is added, and the total envelope entropy is 6.6, with the entropy sum decreasing by 0.1. From K=3 to K=4, the entropy sum decreases by 0.5; from K=4 to K=5, the entropy sum decreases by only 0.2. The decrease slows significantly; therefore, K=4 is identified as the first inflection point. Using K=4 as the optimal number of modes, the final variational mode decomposition of the 6-hour power sequence is performed, yielding 4 intrinsic mode function components for subsequent training.

[0050] The complexity of the power sequence varies for different sliding window samples, and the optimal K value may also differ. By analyzing the curve of the sum of envelope entropy as a function of K, the optimal number of modes K can be automatically selected, avoiding subjective errors that may arise from manually selecting the K value. The optimal K value can effectively capture different frequency components in the time series data, avoiding the accuracy degradation caused by too many or too few mode decompositions. A reasonable K value ensures that each intrinsic mode function component can well represent different characteristics of the signal, improving the accuracy of subsequent prediction models.

[0051] Based on the final sliding window length obtained from training, the input samples are constructed and fed into the multimodal parallel time series prediction model. The future predicted values ​​of several intrinsic mode function components are summed to obtain the final wind power prediction result.

[0052] Furthermore, this application also includes the following steps: based on the historical window length in the final sliding window length, starting from the current time, collect target historical power generation data; based on the future window length in the final sliding window length as a constraint, collect target future numerical weather forecast data; and construct the input sample using the target historical power generation data and the target future numerical weather forecast data.

[0053] Furthermore, this application also includes the following steps: using the future window length in the final sliding window length as a constraint, analyzing the weather forecast accuracy through historical data; analyzing the impact of weather forecast error on wind power under the weather forecast accuracy; and labeling the final wind power forecast result with error based on the error impact.

[0054] Specifically, based on the historical window length within the final sliding window length, target historical power generation data is collected starting from the current moment. In other words, wind power data for a past period is collected, with the current moment as the starting point; the length of this period is determined by the historical window length. Simultaneously, constrained by the future window length within the final sliding window length, target future numerical weather forecast data for a period forward from the current moment is obtained. The target historical power generation data is the actual power data obtained by tracing back to the final determined historical window length, representing the most recent known power state; the target future numerical weather forecast data is the weather forecast data obtained by extending forward to the final determined future window length.

[0055] Historical power generation data and future numerical weather prediction data of the target are combined to form a time-series dataset, which is then combined, aligned, and normalized in the same way, serving as input to a multimodal parallel time-series forecasting model. Each input sample contains two parts of data: historical power generation data and future numerical weather prediction data.

[0056] The prepared input samples are fed into a pre-trained multimodal parallel time-series prediction model. Based on the complex mapping relationships it has learned, the model performs forward computation, and its output is not the final power value, but rather the predicted values ​​of K intrinsic mode function (IMF) components within a future window. The predicted values ​​of these K IMF components at each future time point are summed to obtain the final readable result, which is the wind power prediction value for each future time point, representing a comprehensive prediction of future wind power.

[0057] Using the future window length within the final sliding window length as a constraint, historical data is used to analyze the accuracy of weather forecasts. Historical weather forecast data is compared with actual weather observation data to assess forecast accuracy. Known weather forecast errors and actual weather data are used to analyze how these errors affect wind power forecasting. A quantitative model of the impact of weather forecast errors on power forecasting errors is established using historical power data and wind turbine power curves. The theoretical power curve of the wind turbine is used to analyze the magnitude of power change caused by wind speed errors near specific power points.

[0058] Based on the relationship between weather forecast errors and wind power errors, error labels are provided for wind power forecast results to further optimize or adjust wind power dispatch. If the forecast error is large for a certain period, dynamic adjustments are made based on the error label results, such as using weighted forecasting methods or error correction algorithms to correct the forecast values. For example, assuming an analysis of all 2-hour wind speed forecasts over the past year, compared with measured values, the RMSE of the 2-hour wind speed forecast is 0.8 m / s. The error distribution approximates a normal distribution with a mean of 0 and a standard deviation of 0.7 m / s. Around the rated power of 130 MW for wind turbines, according to the power curve, for every 1 m / s change in wind speed, the power changes by approximately 12 MW. Therefore, a standard deviation error of 0.7 m / s in wind speed results in a standard deviation of approximately 8.4 MW in power. For the current forecast, the multimodal parallel time-series forecast model outputs a forecast result of an average power of 125 MW for the next 30 minutes. Based on the predicted wind speed sequence, 1000 data points conforming to N(0, 0.7) are generated. 2 The perturbation wind speed sequence is used to calculate the error distribution. 1000 wind speed sequences are input into the power prediction model, resulting in 1000 predicted power values, forming a probability distribution. The statistic for this distribution is calculated, with a mean of 125 MW and a standard deviation of 7.2 MW. Taking the 5th and 95th quantiles, a 90% confidence interval of [112 MW, 138 MW] is obtained. The final output predicted value is 125 MW, with the error labeled as a 90% confidence interval of [112 MW, 138 MW] or an uncertainty of ±13 MW.

[0059] By analyzing the impact of weather forecast errors on wind power prediction, the uncertainties in the prediction results are highlighted, thus helping the dispatch system to take these uncertainties into account when scheduling wind power. Through analysis of prediction errors over different time periods and weather conditions, the prediction accuracy is dynamically adjusted, thereby improving the adaptability of the multimodal parallel time-series prediction model to various weather conditions. This is especially important when wind farms encounter significant weather changes, ensuring reasonable prediction and scheduling of wind power.

[0060] In summary, the wind power prediction method based on a multivariate combined prediction model provided in this application has the following technical advantages: By acquiring historical wind power data and historical numerical weather prediction data from wind farms in the first historical time zone, and based on these data, the nonlinear dependence between input variables and output power is analyzed using correlation analysis. A multimodal parallel time series prediction model is then trained using variational mode decomposition (MODED) combined with a sliding window length optimization. Based on the final sliding window length obtained during training, input samples are constructed and fed into the multimodal parallel time series prediction model. The future predicted values ​​of several intrinsic mode function components are summed to obtain the final wind power prediction result. In other words, by using variational mode decomposition to decompose and stabilize the wind power data sequence, and then constructing a multimodal parallel time series prediction model based on machine learning for wind power prediction, the accuracy of wind power prediction is improved.

[0061] Example 2: Based on the same inventive concept as the wind power prediction method based on a multivariate combined prediction model in Example 1, this application also provides a wind power prediction system based on a multivariate combined prediction model. Please refer to the appendix. Figure 2 The wind power prediction system based on a multivariate combined prediction model includes: The time-series data acquisition module 11 is used to acquire the historical wind power data time series and historical numerical weather forecast data time series of the wind farm in the first historical time zone; the model construction module 12 is used to, based on the historical wind power data time series and the historical numerical weather forecast data time series, use correlation analysis to analyze the nonlinear dependence between input variables and output power, and use variational mode decomposition algorithm combined with sliding window length optimization to train a multimodal parallel time-series prediction model; the wind power prediction module 13 is used to, based on the final sliding window length obtained from training, construct input samples and input them into the multimodal parallel time-series prediction model, and sum the future predicted values ​​of several intrinsic mode function components to obtain the final wind power prediction result.

[0062] Furthermore, the model building module 12 in the wind power prediction system based on the multivariate combined prediction model is also used for: performing iterative analysis of the correlation between historical and future power under adjacent time series based on the historical wind power data time series, determining the first historical window length and the first future window length under a stable distribution of correlation coefficients; collecting the historical measured numerical weather forecast data time series corresponding to the historical numerical weather forecast data time series, performing weather forecast prediction accuracy analysis, and identifying a second window length with a prediction accuracy greater than a preset accuracy threshold; optimizing the first future window length with the second window length, and generating a sliding window length with the first historical window length and the optimized first future window length; performing variational mode decomposition on the historical wind power data time series after sequential sliding segmentation based on the sliding window length, and training the multimodal parallel time series prediction model by combining the mode decomposition results with the historical numerical weather forecast data time series of the corresponding sliding window.

[0063] Furthermore, the model building module 12 in the wind power prediction system based on the multivariate combined prediction model is also used to: perform mode decomposition on historical wind power data and future wind power data within the sliding window after sliding segmentation using the variational mode decomposition algorithm, to obtain K historical intrinsic mode function components and K future intrinsic mode function components; use the combination of the K historical intrinsic mode function components and future numerical weather prediction data as input training samples, and use the K future intrinsic mode function components as output supervision values, and perform iterative training based on the sliding window length to generate the multimodal parallel time series prediction model.

[0064] Furthermore, the model building module 12 in the wind power prediction system based on the multivariate combined prediction model is also used to: adaptively adjust the sliding window in combination with the training loss during the iterative training process based on the sliding window length, specifically including: if the training loss is less than a preset loss threshold for N consecutive times, increasing the future window length in the sliding window length according to a preset step size; if the training loss is greater than the preset loss threshold for N consecutive times, decreasing the future window length in the sliding window length according to a preset step size; wherein, the prediction loss of the finally generated multimodal parallel time series prediction model must be less than or equal to the preset loss threshold.

[0065] Furthermore, the model building module 12 in the wind power prediction system based on the multivariate combined prediction model is also used for: the multimodal parallel temporal prediction model includes a shared feature extraction layer, a temporal dependency processing layer, and a modality-specific attention prediction layer connected in sequence; the shared feature extraction layer is composed of at least one layer of dilated causal convolution, used to extract common temporal features from the input data of all modalities; the temporal dependency processing layer is composed of a gated recurrent unit network, used to capture long-term dependencies in the common temporal features; the modality-specific attention prediction layer is used to assign attention weights to each modality and output the future prediction value of each modality.

[0066] Furthermore, the model building module 12 in the wind power prediction system based on the multivariate combined prediction model is also used to: when performing mode decomposition, perform variational mode decomposition for each candidate K value according to the search range of the preset number of modes K, calculate the sum of the envelope entropy of all decomposed mode components, and determine the optimal number of modes K for final decomposition based on the first inflection point of the sum of envelope entropy changing with the K value.

[0067] Furthermore, the wind power prediction module 13 in the wind power prediction system based on the multivariate combined prediction model is also used to: collect target historical power generation data based on the historical window length in the final sliding window length, starting from the current time; collect target future numerical weather forecast data with the future window length in the final sliding window length as a constraint; and construct the input sample with the target historical power generation data and the target future numerical weather forecast data.

[0068] Furthermore, the wind power prediction module 13 in the wind power prediction system based on the multivariate combined prediction model is also used to: analyze the weather prediction accuracy using historical data with the future window length in the final sliding window length as a constraint; analyze the error impact of weather prediction error on wind power under the weather prediction accuracy; and mark the error of the final wind power prediction result according to the error impact.

[0069] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. The wind power prediction method and specific examples based on the multivariate combined prediction model in Embodiment 1 are also applicable to the wind power prediction system based on the multivariate combined prediction model in this embodiment. Through the foregoing detailed description of the wind power prediction method based on the multivariate combined prediction model, those skilled in the art can clearly understand the wind power prediction system based on the multivariate combined prediction model in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here. As for the systems / devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant details can be found in the method section.

[0070] In Embodiment 3, based on the same inventive concept as the wind power prediction method based on a multivariate combined prediction model in Embodiment 1, this application also provides a computer-readable storage medium storing a computer program that, when executed, implements the steps of the wind power prediction method based on a multivariate combined prediction model described in any one of Embodiment 1.

[0071] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0072] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A wind power prediction method based on a multivariate combined prediction model, characterized in that, include: Obtain the time series of historical wind power data and historical numerical weather forecast data of the wind farm in the first historical time zone; Based on the historical wind power data time series and the historical numerical weather forecast data time series, the nonlinear dependence between input variables and output power is analyzed by correlation analysis, and the multimodal parallel time series prediction model is trained by combining variational mode decomposition algorithm with sliding window length optimization. Based on the final sliding window length obtained from training, the input samples are constructed and fed into the multimodal parallel time series prediction model. The future predicted values ​​of several intrinsic mode function components are summed to obtain the final wind power prediction result.

2. The wind power prediction method based on a multivariate combined prediction model as described in claim 1, characterized in that, Based on the historical wind power data time series and the historical numerical weather forecast data time series, correlation analysis is used to analyze the nonlinear dependency between input variables and output power. A multimodal parallel time series prediction model is trained using variational mode decomposition algorithm combined with sliding window length optimization, including: Based on the historical wind power data time series, a cyclic iterative analysis of the historical-future power correlation under adjacent time series is performed to determine the first historical window length and the first future window length under a stable distribution of correlation coefficients. Collect the historical measured numerical weather forecast data time series corresponding to the historical numerical weather forecast data time series, perform weather forecast prediction accuracy analysis, and identify the second window length whose prediction accuracy is greater than a preset accuracy threshold; The length of the first future window is optimized using the second window length, and the length of the sliding window is generated using the length of the first historical window and the optimized length of the first future window. Based on the sliding window length, the historical wind power data time series is sequentially segmented by sliding and then subjected to variational mode decomposition. The multimodal parallel time series prediction model is trained by combining the mode decomposition results with the historical numerical weather forecast data time series corresponding to the sliding window.

3. The wind power prediction method based on a multivariate combined prediction model as described in claim 1, characterized in that, Based on the sliding window length, the historical wind power data time series is sequentially segmented using sliding mode decomposition. The multimodal parallel time series prediction model is then trained by combining the mode decomposition results with the corresponding historical numerical weather forecast data time series of the sliding window, including: Using the variational mode decomposition algorithm, the historical wind power data and future wind power data within the sliding window after sliding segmentation are decomposed into K historical intrinsic mode function components and K future intrinsic mode function components. The multimodal parallel time series prediction model is generated by using the K historical intrinsic mode function components and the combination of future numerical weather prediction data as input training samples, and the K future intrinsic mode function components as output supervision values, and iterative training based on the sliding window length.

4. The wind power prediction method based on a multivariate combined prediction model as described in claim 3, characterized in that, During iterative training based on the aforementioned sliding window length, the sliding window is adaptively adjusted in conjunction with the training loss, specifically including: If the training loss is less than the preset loss threshold for N consecutive training iterations, the future window length in the sliding window length is increased and adjusted according to the preset step size. If the training loss is greater than the preset loss threshold for N consecutive training iterations, the future window length in the sliding window length is reduced by a preset step size. The prediction loss of the final generated multimodal parallel temporal prediction model must be less than or equal to the preset loss threshold.

5. The wind power prediction method based on a multivariate combined prediction model as described in claim 4, characterized in that, The multimodal parallel temporal prediction model includes a shared feature extraction layer, a temporal dependency processing layer, and a modality-specific attention prediction layer connected in sequence. The shared feature extraction layer consists of at least one layer of dilated causal convolution, used to extract common temporal features from input data of all modalities; the temporal dependency processing layer consists of a gated recurrent unit network, used to capture long-term dependencies in the common temporal features. The modality-specific attention prediction layer is used to assign attention weights to each modality and output future prediction values ​​for each modality.

6. The wind power prediction method based on a multivariate combined prediction model as described in claim 3, characterized in that, During modal decomposition, variational modal decomposition is performed for each candidate K value according to the search range of the preset number of modes K, and the sum of the envelope entropy of all decomposed modal components is calculated. The optimal number of modes K is determined based on the first inflection point of the sum of envelope entropy as a function of K value, and the final decomposition is performed.

7. The wind power prediction method based on a multivariate combined prediction model as described in claim 1, characterized in that, Based on the sliding window length obtained during training, input samples are constructed, including: Based on the historical window length in the final sliding window length, the target historical power generation data is collected starting from the current moment; Collect target future numerical weather forecast data, using the future window length in the final sliding window length as a constraint. The input sample is constructed using the target's historical power generation data and the target's future numerical weather forecast data.

8. The wind power prediction method based on a multivariate combined prediction model as described in claim 7, characterized in that, After obtaining the final wind power prediction results, the following are included: Using the future window length within the final sliding window length as a constraint, the accuracy of weather forecasts is analyzed through historical data. The impact of weather forecast error on wind power output is analyzed under the aforementioned weather forecast accuracy. The final wind power prediction result is labeled with errors based on the aforementioned error impact.

9. A wind power prediction system based on a multivariate combined prediction model, characterized in that, The steps for implementing the wind power prediction method based on a multivariate combined prediction model according to any one of claims 1 to 8, wherein the wind power prediction system based on the multivariate combined prediction model comprises: The time-series data acquisition module is used to acquire the time-series data of historical wind power and historical numerical weather forecast data of wind farms in the first historical time zone. The model building module is used to train a multimodal parallel time series prediction model based on the historical wind power data time series and the historical numerical weather forecast data time series, using correlation analysis to analyze the nonlinear dependence between input variables and output power, and using variational mode decomposition algorithm combined with sliding window length optimization. The wind power prediction module is used to construct input samples based on the final sliding window length obtained from training, input them into the multimodal parallel time series prediction model, and sum the future predicted values ​​of several intrinsic mode function components to obtain the final wind power prediction result.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the steps of the wind power prediction method based on a multivariate combined prediction model as described in any one of claims 1 to 8.

Images