Wind power intelligent prediction method and system based on actually measured power and wind speed collaborative assimilation

Abstract

The invention discloses a wind power intelligent prediction method and system based on measured power and wind speed collaborative assimilation, and the method comprises the steps: constructing a two-stage collaborative'macroscopic trend constraint-microscopic observation correction 'bimodal learning architecture, and carrying out the three-dimensional residual convolution and multi-head self-attention modeling of NWP data through a Resnet3D-Attention model, a short-term prediction background field with meteorological dynamic constraints is generated, and a macroscopic trend reference is provided for ultra-short-term correction; based on a Resnet-BiLSTM double-branch assimilation network, cross-modal feature fusion and dynamic error correction of real-time field station actually-measured power and wind speed data and a prediction background field are realized in a power space, and the fitting capability of a prediction result to real power fluctuation can be enhanced especially in scenes such as severe convective weather and complex terrains.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a method and system for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed. Background Art

[0002] As the global energy mix shifts toward a low-carbon future, the large-scale integration of wind power, a core clean and renewable energy source, places higher demands on the stable operation of power grids. High-precision ultra-short-term wind power forecasting plays a key role in frequency regulation, optimal allocation of spinning reserve capacity, and economic dispatch.

[0003] Currently, wind power forecasting generally uses methods such as physical models, statistical models, artificial intelligence models, and hybrid models. Physical models often rely on numerical weather forecast data and are suitable for short-term predictions. Statistical models exploit time series characteristics based on historical data. In recent years, artificial intelligence models, represented by deep learning, have demonstrated great potential in wind power forecasting due to their ability to capture complex nonlinear characteristics. For example, models such as CNN, LSTM, and TCN have been shown to perform well in wind power regression forecasting. However, despite the great potential of artificial intelligence models in capturing complex nonlinear characteristics, their development often focuses on optimizing deep learning architectures or learning patterns from static historical datasets, thereby ignoring the dynamic, real-time characteristics of atmospheric conditions and power generation. This overemphasis on model architecture, rather than a thorough consideration of data dynamics and utilization efficiency, has limited their performance in practical applications.

[0004] On the other hand, the inherent reliance of existing methods on a single data source also leads to inaccurate forecasts: for example, models built solely on historical measured data struggle to fully capture the dynamic changes in the instantaneous atmospheric state. This is because historical data cannot fully reflect real-time, ever-changing meteorological conditions, resulting in a disconnect between forecast results and actual operating conditions. While large-scale NWP data provides broad atmospheric background information, they often lack the local accuracy required for specific wind farm areas. This inherent mismatch in spatial and temporal resolution leads to inaccurate forecasts, as large-scale meteorological forecasts may be biased in local areas.

[0005] Existing methods fail to effectively integrate and utilize real-time, high-frequency station measurement data to dynamically correct forecast results. This limits their ability to adapt to sudden changes or correct accumulated errors, making it impossible for forecast results to respond promptly to errors and changes in actual operation. However, solving this problem is not easy, because traditional data assimilation techniques are rooted in atmospheric dynamics and rely on complex physical equations and computationally intensive optimization schemes. Directly applying this approach to power systems would impose a huge computational burden, making it unable to meet the high real-time requirements of ultra-short-term forecasting. This computational complexity makes traditional methods difficult to promote in practical applications. Another core challenge is that meteorological variables and power system parameters belong to different physical domains. Constructing a unified, physics-based assimilation framework across these different spaces is inherently difficult, hindering the direct conversion of meteorological observation information into power forecast corrections. Summary of the Invention

[0006] The present invention aims to solve one of the technical problems existing in the related art at least to a certain extent.

[0007] One purpose of the present invention is to provide an intelligent wind power prediction method based on the coordinated assimilation of measured power and wind speed, aiming to overcome the dual bottlenecks of the existing technology, namely the spatial resolution limitation of meteorological model data and the inefficient dynamic utilization of real-time measured data at the site. By constructing a full-process technical chain of "cross-modal data fusion-dynamic error compensation", the spatiotemporal coordination of the macro trend constraints of numerical weather forecast data and the micro dynamic correction of measured power and wind speed data at the site is achieved, which is suitable for short-term and ultra-short-term prediction scenarios of wind power in power systems.

[0008] Another object of the present invention is to provide an intelligent wind power prediction system based on the coordinated assimilation of measured power and wind speed.

[0009] In order to achieve the above-mentioned object, the present invention provides, on one hand, a method for intelligent wind power prediction based on the collaborative assimilation of measured power and wind speed, comprising a first stage and a second stage of collaborative work;

[0010] The first stage generates the short-term prediction background field, including:

[0011] Obtain numerical weather forecast data and perform spatiotemporal alignment preprocessing;

[0012] The preprocessed data is processed by the Resnet3D-Attention model containing a three-dimensional residual convolution layer and a multi-head self-attention layer to generate a continuous time series power prediction background field;

[0013] The second phase is to conduct dynamic assimilation correction of ultra-short-term wind power, including:

[0014] Obtain historical measured power data, measured wind speed data, and short-term forecast background field data to construct a data set; and perform data cleaning and sliding window processing on the data set;

[0015] A Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture was constructed. The first branch of the Resnet-BiLSTM AI assimilation model extracted the mutation characteristics of the measured power and wind speed, while the second branch extracted the background field prediction characteristics. The dual-branch features were then fused using a bidirectional time series modeling unit.

[0016] Use the preprocessed data set to train the Resnet-BiLSTM AI assimilation model; input the real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model to output a rolling updated ultra-short-term power forecast sequence.

[0017] A further preferred technical solution of the present invention is to obtain numerical weather forecast data and perform spatiotemporal alignment preprocessing, specifically including:

[0018] The numerical weather forecast data were spatially resampled to a resolution of 0.1°x0.1° and temporally interpolated to the same temporal resolution as the measured power data.

[0019] When creating samples, a sliding time window is constructed in the time dimension. Taking the target timestamp as the benchmark, the numerical weather forecast data of several time steps before and after are integrated to form a time series data of continuous time. At the same time, a 20x20 spatial grid area is selected around the target station to cover the terrain and meteorological field information around the station.

[0020] Extract the near-surface layer parameters and wind vector parameters from the numerical weather forecast data and form a five-dimensional tensor as the input of the Resnet3D-Attention model, which is the batch size, feature dimension, time step, spatial height and spatial width;

[0021] The five-dimensional tensor containing spatiotemporal information is Z-score normalized to eliminate dimensional differences. The Z-score normalization formula is:

[0022] ;

[0023] in, is the original data point, is the data mean, is the standard deviation.

[0024] Preferably, the Resnet3D-Attention model includes:

[0025] A 3D residual convolutional layer for extracting spatiotemporal correlation features from numerical weather forecast data, and a multi-head self-attention layer for quantifying the contribution differences of key meteorological variables at different time steps and modeling temporal dependencies;

[0026] The three-dimensional residual convolution layer adopts a residual block structure, and its output Expressed as:

[0027] ;

[0028] When the input and output dimensions are inconsistent, linear projection is used Perform dimension matching, expressed as:

[0029] ;

[0030] in, is the input of the residual block, represents the residual mapping that needs to be learned, is a learnable weight parameter;

[0031] The multi-head self-attention layer projects the input into multiple low-dimensional subspaces and calculates multiple attentions in parallel, and then concatenates the results, so that the Resnet3D-Attention model can focus on different aspects of the input time series at the same time. The formula is:

[0032] ;

[0033] Each attention head The calculation formula is:

[0034] ;

[0035] The calculation formula of the attention mechanism is:

[0036] ;

[0037] in, 、 、 are query, key, and value matrices respectively; 、 、 It is The linear projection matrix of the attention heads, is the final linear projection matrix, is the dimension of the key vector.

[0038] As a preference, data cleaning and sliding window processing are performed on the data set, specifically:

[0039] Data cleaning uses a multi-stage data cleaning method that combines physical constraints with data distribution characteristics. This includes identifying and eliminating abnormal power-wind speed data points that do not conform to the wind turbine operating mechanism based on physical rules, and identifying and eliminating outliers in the data distribution using a density clustering algorithm based on the joint distribution of wind speed and power.

[0040] Among them, the physical rule-based screening and identification includes logical verification based on the cut-in wind speed threshold and the minimum power threshold. When the wind speed is lower than the cut-in wind speed and the power is higher than the preset minimum power threshold, it is defined as an outlier. The density clustering algorithm uses the DBSCAN algorithm;

[0041] The sliding window processing refers to the use of a rolling time window method to generate samples. For each time point in the data set, , which is defined as the end time of the measured data acquisition; As the starting point, several hours of measured data from the site are selected as the input of the first branch of the Resnet-BiLSTM AI assimilation model; the power prediction background field of several hours in the future is selected as the input of the second branch of the Resnet-BiLSTM AI assimilation model; at the same time, the measured power data in the same future time period is selected as the target label that the Resnet-BiLSTM AI assimilation model needs to predict.

[0042] Preferably, the Resnet-BiLSTM AI assimilation model includes:

[0043] A dual-channel measured and background field feature extractor consists of two independent ResNet-based encoders as the first and second branches, respectively. The first branch extracts the mutation characteristics of the measured power and wind speed, while the second branch extracts the background field prediction characteristics. The ResNet structure performs one-dimensional convolution along the time dimension to capture local patterns and mutation characteristics in time series data, and connects them through residuals.

[0044] The temporal dependency modeling module uses a bidirectional long short-term memory network to capture the long-term dependency and bidirectional contextual information of the measured feature sequence and the background field feature sequence;

[0045] The information collaborative assimilation module concatenates the measured feature sequence and the background field feature sequence processed by their respective bidirectional long short-term memory network layers along the feature dimension to form a joint feature representation that integrates the two information streams. The joint feature is then input into a shared bidirectional long short-term memory network layer.

[0046] The assimilation output module maps the output features of the shared bidirectional long short-term memory network layer to the final corrected ultra-short-term power prediction value through a fully connected layer.

[0047] As a preference, the preprocessed dataset is used to train the Resnet-BiLSTM AI assimilation model, specifically:

[0048] The historical measured power data, measured wind speed data and short-term forecast background field data after data cleaning and sliding window processing are divided into training set, validation set and test set;

[0049] The stochastic gradient descent optimization algorithm is used to train the constructed Resnet-BiLSTM AI assimilation model with the training set. The hyperparameters used in training are set as follows: batch size is 32, number of training rounds is 200, and for model parameters , the update rule for gradient descent is:

[0050] ;

[0051] in, are model parameters, is the learning rate, Represents the objective function Parameters In a single training sample The gradient on

[0052] The initial learning rate for model training is , using the StepLR learning rate scheduler to implement dynamic adjustment, in the After decay steps, the learning rate for:

[0053] ;

[0054] in, represents the attenuation coefficient;

[0055] The loss function is the mean absolute error. , which measures the average absolute difference between the predicted value and the true value, is expressed as:

[0056] ;

[0057] in, is the sample size, It is True value, It is predicted values.

[0058] Another aspect of the present invention provides a wind power intelligent prediction system based on the coordinated assimilation of measured power and wind speed, comprising:

[0059] The short-term prediction module receives numerical weather forecast data and builds a Resnet3D-Attention model consisting of a three-dimensional residual convolution layer and a multi-head self-attention layer. Based on the numerical weather forecast data, it outputs a future short-term wind power forecast sequence as the background field for ultra-short-term prediction.

[0060] Data acquisition module, used to obtain historical measured power data, measured wind speed data and short-term forecast background field data;

[0061] The data preprocessing module is used to preprocess the historical measured power data, measured wind speed data, and short-term forecast background field data, including data set construction, sliding windowing, and multi-stage data cleaning;

[0062] The AI ​​assimilation model training module is used to build a Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture. The first branch of this Resnet-BiLSTM AI assimilation model extracts the mutation characteristics of measured power and wind speed, and the second branch extracts the background field prediction characteristics. The dual-branch features are then integrated based on the bidirectional time series modeling unit. The Resnet-BiLSTM AI assimilation model is trained using the preprocessed dataset.

[0063] The ultra-short-term assimilation prediction module is used to input real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model, and output a rolling updated ultra-short-term power prediction sequence.

[0064] Yet another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the above-mentioned wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed.

[0065] Another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed.

[0066] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed.

[0067] Beneficial effects: The present invention constructs a two-stage collaborative "macro-trend constraint-micro-observation correction" bimodal learning architecture, performs three-dimensional residual convolution and multi-head self-attention modeling on NWP data through the Resnet3D-Attention model, generates a 24-hour short-term forecast background field with meteorological dynamic constraints, and provides a macro-trend benchmark for ultra-short-term correction; relying on the Resnet-BiLSTM dual-branch assimilation network, it realizes cross-modal feature fusion and dynamic error correction of real-time station measured power and wind speed data with the forecast background field in the power space, especially in scenarios such as severe convective weather and complex terrain, which can enhance the prediction results' ability to fit real power fluctuations.

[0068] This paper constructs a multi-stage data cleaning system that includes physical constraint screening and density clustering denoising. Combining dual-channel feature decoupling with bidirectional time series modeling, it accurately captures the spatiotemporal propagation patterns of forecast errors. Compared to traditional single models, this paper improves the ability to suppress extreme errors, and its dynamic error compensation efficiency meets the minute-level response requirements of ultra-short-term forecasting.

[0069] By improving the accuracy and reliability of ultra-short-term wind power forecasts, this invention can provide a more precise basis for decision-making in power grid dispatch. High-precision ultra-short-term wind power forecasts are crucial for power grid frequency regulation, optimal allocation of spinning reserve capacity, and economic dispatch. By reducing forecast errors, particularly extreme errors, this invention helps power grids more effectively manage the volatility associated with wind power integration, thereby improving the stability and economic efficiency of grid operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is an overall flow chart of the wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed of the present invention;

[0071] Figure 2 This is a schematic diagram of the architecture of the Resnet3D-Attention model constructed in the first phase of the present invention;

[0072] Figure 3 This is a schematic diagram of the architecture of the Resnet-BiLSTM AI assimilation model constructed in the second phase of the present invention;

[0073] Figure 4 This is a comparison of the wind speed-power scatter plots of the station before and after data cleaning in Example 1;

[0074] Figure 5 This is a comparison chart of the measured power and the ultra-short-term predicted power after AI assimilation and correction on the training set in Example 1;

[0075] Figure 6This is the error analysis of the short-term prediction background field and the ultra-short-term prediction power after AI assimilation correction in Example 1 on the training set. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0077] The following combination Figures 1-6 The present invention describes a method and system for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed.

[0078] Example 1: This example provides a two-stage wind power intelligent prediction method based on AI assimilation, such as Figure 1 As shown, this method achieves high-precision wind power forecasting by integrating macro-trend constraints from numerical weather forecast (NWP) data with micro-dynamic corrections to measured power and wind speed data at the site. The following examples describe the implementation steps of the technical solution in detail to ensure that those skilled in the art can replicate it, and also include explanations of the results and principle analysis.

[0079] Phase 1: Generation of short-term prediction background fields

[0080] In this phase, the Resnet3D-Attention model is used to process NWP data, generating a power forecast background field for the next 24 hours with a time granularity of 15 minutes, providing a macro-trend benchmark for ultra-short-term corrections. This model utilizes three-dimensional residual convolution to capture local spatiotemporal features and multi-head self-attention to model long-range temporal dependencies, improving adaptability to complex terrain and meteorological changes. The specific method is as follows:

[0081] Step 1: NWP data preprocessing

[0082] Data preprocessing is a key step in ensuring the quality and efficiency of model training. The numerical weather forecast data used in this example were selected from the European Centre for Medium-Range Weather Forecasts, spatially resampled to 0.1° x 0.1°, and temporally interpolated to the same 15-minute temporal resolution as the field data.

[0083] When creating samples, this embodiment constructs a sliding time window in the time dimension. Taking the target timestamp as the benchmark, it integrates forecast data from four time steps before and after, forming data containing nine consecutive time periods. At the same time, a 20x20 spatial grid area is selected around the target station, covering the terrain and meteorological field information surrounding the station. In terms of feature engineering, near-surface layer parameters and wind vector parameters are extracted, ultimately forming a five-dimensional tensor for model input, representing batch size, feature dimension, time step, spatial height, and spatial width.

[0084] The five-dimensional tensor containing spatiotemporal information is Z-score normalized. Z-score normalization is a widely used data normalization technique that converts data values ​​into a distribution with a mean of 0 and a standard deviation of 1. The Z-score normalization formula is:

[0085] ;

[0086] in, is the original data point, is the data mean, is the standard deviation.

[0087] This embodiment performs Z-score normalization on NWP data, which can eliminate the dimensional differences between different meteorological variables and avoid the adverse effects of dimensional differences on model training, thereby accelerating model convergence and improving training stability.

[0088] Step 2: Build Resnet3D-Attention model and train

[0089] The constructed Resnet3D-Attention is as follows Figure 2 As shown in the figure, this embodiment uses a hybrid model based on Resnet-Attention for short-term wind power AI forecasting. This model can extract deep spatiotemporal features from multi-dimensional numerical weather forecast data and use a multi-head self-attention mechanism to quantify the cross-time contribution differences of key meteorological variables, thereby generating reliable baseline power forecasts. The model architecture integrates the ResNet and Multi-Head Attention mechanisms of 3DCNN. Its design is to extract effective information from complex NWP data and generate the background field forecast required for subsequent ultra-short-term assimilation processes. Specifically, it includes:

[0090] The three-dimensional residual convolution layer (ResNet3D) is used to extract spatiotemporal correlation features from numerical weather forecast data. This embodiment uses ResNet based on 3D CNN as the backbone encoder to process the input five-dimensional NWP data tensor. Unlike 2DCNN which only processes spatial information, 3D CNN can perform convolution operations directly on the spatiotemporal three-dimensional cube, effectively capturing local spatiotemporal patterns, that is, simultaneously correlating the feature changes of adjacent grid points in space at consecutive time steps. Through the ResNet structure, the introduction of residual connections greatly alleviates the gradient vanishing problem in deep network training, enabling the model to learn deeper abstract features. The core idea of ​​the residual block is to let the stacked nonlinear layers fit the residual function , instead of fitting the unreferenced map directly The original mapping is refactored to For a residual block, its output It can be expressed as: ;

[0091] When the input and output dimensions are inconsistent, linear projection is used Perform dimension matching, expressed as:

[0092] ;

[0093] in, is the input of the residual block, represents the residual mapping that needs to be learned, are learnable weight parameters.

[0094] The innovation of this embodiment's three-dimensional residual convolutional layer lies in its direct processing of the five-dimensional NWP tensor, effectively capturing the complex temporal and spatial correlations of wind field meteorological elements. Key residual blocks also integrate spatial and channel attention mechanisms, enabling the model to adaptively focus on feature channels and spatiotemporal regions that are most informative for downstream forecasting tasks.

[0095] The multi-head self-attention layer is used to quantify the differences in the contributions of key meteorological variables at different time steps and model temporal dependencies. After ResNet3D extracts deep spatiotemporal features, the multi-head self-attention layer processes the transformed feature sequence, focusing on capturing the complex temporal dependencies and contributions between time steps and different meteorological variables within the NWP forecast time series. The multi-head self-attention layer utilizes the multi-head self-attention mechanism derived from the Transformer model. Compared to traditional recurrent neural networks and their variants, the self-attention mechanism can directly model dependencies between time steps of arbitrary distance, unconstrained by the distance of information transmission. This is particularly important for pattern data, as forecast information from earlier times may still affect wind power changes far into the future. Through learned attention weights, the model dynamically assigns different importance to different time steps or features in the sequence, thereby more effectively integrating information to generate forecasts. The core of the self-attention mechanism is to calculate the similarity between the query (Q), key (K), and value (V), and then perform a weighted sum of the values ​​based on the similarity.

[0096] The calculation formula of the attention mechanism is:

[0097] ;

[0098] in, is the dimension of the key vector.

[0099] The multi-head attention mechanism projects the input into multiple low-dimensional subspaces, calculates multiple attentions in parallel, and then concatenates the results, so that the model can focus on different aspects of the input sequence at the same time. Its formula is:

[0100] ;

[0101] Each attention head Calculated as:

[0102] ;

[0103] in, , , It is The linear projection matrix of the attention heads, is the final linear projection matrix. The application of the multi-head self-attention mechanism in this embodiment enables quantifying the contributions of key meteorological variables at different time steps and modeling their complex temporal dependencies. This composite model, combining the Attention mechanism with 3D ResNet, is designed based on an in-depth analysis of the data characteristics and prediction challenges specific to wind power forecasting, resulting in an innovative design of the model structure.

[0104] The following describes the training process of the Resnet3D-Attention model of this embodiment:

[0105] In practical applications, the ResNet3D-Attention model must be trained on large-scale historical numerical weather forecast data to ensure that it can learn and generalize to wind power forecast patterns across different seasons and complex weather conditions. In this example, training is performed using one year of numerical weather forecast data (January 1, 2023, to January 1, 2024). The short-term forecast sequence output from this stage serves as the background field input for the second-stage AI assimilation model.

[0106] The training hyperparameters used a batch size of 256 and a number of training epochs of 200. The optimizer used was AdamW, with an initial learning rate of 1e-04 and a weight decay coefficient of 2e-0. The StepLR learning rate scheduler was used, with the learning rate decayed by 0.1 every 80 epochs. The mean absolute error loss function was chosen. Training was terminated if the loss function did not improve after 10 consecutive epochs on the validation set, and the best model was saved as the final result.

[0107] Phase 2: Ultra-short-term wind power AI assimilation correction

[0108] This phase aims to build an innovative AI-based assimilation model, the Resnet-BiLSTM model. By deeply integrating the short-term power forecast generated in the first phase with real-time, high-frequency field data, this model dynamically corrects the background field and generates a more accurate ultra-short-term wind power forecast for the future. This approach implements a process similar to traditional data assimilation in the power space, leveraging field data to update and optimize forecast results. Its principle lies in dual-channel feature decoupling to capture local mutations and trend characteristics, BiLSTM modeling of bidirectional temporal dependencies, and the information collaboration module learning nonlinear coupling to improve the real-time performance and accuracy of forecasts.

[0109] Step 3: Ultra-short-term assimilation dataset construction and feature engineering

[0110] Before training the AI ​​assimilation model, historical data must be collected and integrated. This data includes historical short-term wind power forecasts, historical site-measured power data, and historical site-measured wind speed data. This data is collected at 15-minute intervals. The data coverage should be long enough to capture dynamic characteristics across different seasons and weather conditions.

[0111] In order to ensure the quality and efficiency of the model training, the acquired historical data needs to be preprocessed. In the two-stage prediction-assimilation AI framework proposed in this embodiment, the quality and reliability of the data, especially the power data measured at the site for the second-stage ultra-short-term wind power AI assimilation correction, are crucial to the accuracy of the final ultra-short-term prediction. However, there are common problems with the actual site power data collected: first, active maintenance, communication interruption or site power rationing will cause non-physical outliers to be mixed into the data; second, abnormal meteorological conditions and erroneous records will also cause outliers. If these outliers are used directly for model training, especially as "observation" information in the assimilation process, they will introduce erroneous information, causing the model to learn dynamic relationships that deviate from the actual situation, destroying the true coupling between meteorological elements and power, thereby seriously affecting the performance of the prediction model, especially in the ultra-short-term correction stage, which may lead to correction errors and reduce prediction accuracy.

[0112] Based on the above two issues, this embodiment adopts a multi-stage data cleaning solution that combines physical constraints and data distribution characteristics to systematically identify and remove outliers in the original wind speed-power data records. The cleaning process mainly includes the following two stages:

[0113] Physical rule screening based on operating mechanisms: Based on physical constraints such as the wind turbine's cut-in wind speed threshold and minimum power threshold, basic logical checks are applied to the measured power and wind speed data to identify and eliminate abnormal data points that do not conform to basic energy conversion laws. For example, if the wind speed is lower than the cut-in wind speed but the power is higher than the preset minimum power threshold, it is defined as a physical anomaly.

[0114] Density Clustering Anomaly Detection Based on the Joint Wind Speed-Power Distribution: Building on physical screening, this example uses density clustering algorithms such as DBSCAN to analyze the joint wind speed-power distribution, uncovering hidden complex anomaly patterns and outliers. DBSCAN constructs clusters by defining core points, reachable points, and noise points.

[0115] DBSCAN core parameters: The neighborhood radius is defined, Defines the minimum number of neighborhood points required to form a core point.

[0116] DBSCAN point types include:

[0117] Core point: If a point of There are at least point, then this point is the core point.

[0118] Direct density up to: point If at the core of In the neighborhood, Direct density can be reached .

[0119] Density reachable: If there exists a point chain , , , = , = , and each The density can be directly ,but The density can reach .

[0120] Density connection: If there is a core point , making and The density can reach ,but and It is density connected.

[0121] Noise point: a point that does not belong to any cluster, that is, a point that is not reachable by any core point density.

[0122] This method, employed in this embodiment, can eliminate distribution outliers caused by complex weather conditions or recording errors, preventing these irregular data from interfering with subsequent model learning of meteorological-power statistical relationships and error correction patterns. The cleaned data points are more closely clustered near the wind power curve, making the data more consistent with the physical characteristics of the wind power curve. This provides high-quality field observation data for subsequent model training, especially for ultra-short-term assimilation and correction.

[0123] In addition, in this embodiment, missing values ​​in the data can be filled by using methods such as linear interpolation.

[0124] After the above two stages of cleaning process, the scatter diagram of the actual wind speed and actual power generation before and after cleaning in this embodiment is as follows: Figure 4 shown.

[0125] When constructing a data set, this embodiment uses a rolling time window method to generate samples. , which is defined as the end time of obtaining measured information. As the starting point, the measured data of the site for 6 hours is selected as the historical information input of the model; the short-term predicted power for the next 6 hours is selected as the background field input sequence of the model; at the same time, the actual power data of the same future time period is selected as the target label that needs to be predicted by the Resnet-BiLSTM AI assimilation model.

[0126] Step 4: Build the Resnet-BiLSTM AI assimilation model and train it

[0127] This embodiment uses the Resnet-BiLSTM model as the AI ​​assimilation model. The model architecture is as follows: Figure 3 As shown in Figure 2, the Resnet-BiLSTM AI assimilation model can effectively integrate short-term power forecasts with real-time, high-frequency field data, and perform real-time corrections to the background field. This architecture integrates a deep residual network and a bidirectional long short-term memory network. The model includes:

[0128] Dual-channel measurement and background feature extractor: Two independent ResNet-based encoders are first used to process the input historical measurement sequence and background prediction sequence, respectively. The ResNet structure performs one-dimensional convolution along the time dimension, effectively capturing local patterns and mutation characteristics in time series data. Residual connections enhance the network's deep feature extraction capabilities and robustness.

[0129] Temporal dependency modeling module: uses a bidirectional long short-term memory network to capture the long-term dependency and bidirectional contextual information of the measured feature sequence and the background field feature sequence; unit calculation includes:

[0130] Forget Gate ): Determine the cell state from the previous moment What information is forgotten in the equation is expressed as:

[0131] ;

[0132] Input Gate ( ): Determine the current input What information is added to the cell state in , represented as:

[0133] ;

[0134] Candidate cell states ( ): Generate new candidate values ​​that may be added to the cell state, represented as:

[0135] ;

[0136] Cell state update ( ): Combine the forget gate and the input gate to update the cell state, expressed as:

[0137] ;

[0138] Output Gate( ): Determine the current cell state Which information is output to the hidden state is expressed as:

[0139] ;

[0140] Hidden state output ( ): The final hidden state output is expressed as:

[0141] ;

[0142] in, is the Sigmoid activation function, is the Tanh activation function, is element-wise multiplication, is the weight matrix, is the bias vector, is the hidden state at the previous moment, is the input at the current moment.

[0143] Final output of BiLSTM is the probability vector of the forward LSTM and the probability vector of the reverse LSTM Combination of:

[0144] ;

[0145] BiLSTM can effectively capture long-term dependencies and bidirectional contextual information in time series data and fully understand the correlation of data in the time dimension.

[0146] In the information collaborative assimilation module, the measured feature sequences and background field feature sequences, after their respective BiLSTM processing, are concatenated along the feature dimension to form a joint feature representation that fuses the two information streams. This joint feature is then input into a shared BiLSTM layer, referred to in this embodiment as the fused BiLSTM layer. This layer not only processes the temporal dependencies of the concatenated sequences but, more importantly, learns the complex interactions and nonlinear coupling relationships between the measured data and the background field prediction data. In particular, it models how the measured features at the site are related to the background field prediction error, thereby enabling error correction based on the current measured information. This layer performs an "assimilation" function within the power prediction space through deep learning, incorporating the influence of measured observation information into the background field prediction.

[0147] The assimilation output module integrates the features output by the BiLSTM layer and maps them through a fully connected layer to the final, corrected ultra-short-term power forecast. This output sequence represents a more accurate power forecast after collaborative assimilation and correction with real-time measured data.

[0148] The following describes the training process of the Resnet-BiLSTM AI assimilation model constructed in this embodiment:

[0149] The pre-processed historical data is divided into a training set, a validation set, and a test set in chronological order. This embodiment adopts a rolling monthly partitioning strategy, using data from the 1st to the 20th of each month for training, data from the 21st to the 25th for validation, and data from the 26th to the end of the month for testing.

[0150] The hyperparameters used in training are: batch size (Batch Size) is 32; number of training rounds (Epochs) is 200. The present invention uses the stochastic gradient descent (SGD) optimization algorithm. For the model parameters , the update rule of SGD is:

[0151] ;

[0152] in, are model parameters, is the learning rate, is the objective function Parameters In a single training sample The gradient on .

[0153] Use StepLR learning rate scheduler to implement dynamic adjustment: every The learning rate is decayed by the coefficient. Assume that the initial learning rate is , in decay steps (i.e. after After one round of training), the learning rate for:

[0154] ;

[0155] In this embodiment, , , .

[0156] The loss function chosen is the mean absolute error (MAE). MAE measures the average of the absolute differences between the predicted value and the true value, expressed as:

[0157] ;

[0158] in, is the sample size, It is True value, It is predicted values.

[0159] Set up an early stopping mechanism: monitor performance through the validation set. If the validation loss does not improve for 10 consecutive rounds, stop training and save the best performing model on the validation set.

[0160] Step 5: Obtain real-time forecast data and conduct ultra-short-term wind power forecasts

[0161] When performing actual ultra-short-term forecasts, real-time input data is required, including the real-time short-term wind power forecast background field, real-time station-measured power data, and real-time station-measured wind speed data. This real-time data requires the same preprocessing as historical data to match the input format of the Resnet-BiLSTM AI assimilation model.

[0162] The pre-processed real-time short-term wind power forecast background field, real-time station measured power data, and real-time station measured wind speed data are used as input and put into the trained Resnet-BiLSTM AI assimilation model. The Resnet-BiLSTM AI assimilation model will output a more accurate future ultra-short-term wind power forecast sequence after the real-time measured data is collaboratively assimilated and corrected, such as Figure 5 As shown in the figure, it is a line graph showing the change of the predicted power and actual power at the 4th hour of the ultra-short-term forecast with time points, as shown in the figure. Figure 6 As shown in the figure, it is a line graph showing the background field and the prediction error of the 4th hour after assimilation as a function of time. The error performance is significantly improved compared to the background field.

[0163] Embodiment 2: This embodiment provides a wind power intelligent prediction system based on the coordinated assimilation of measured power and wind speed, including:

[0164] The short-term prediction module receives numerical weather forecast data and builds a Resnet3D-Attention model consisting of a three-dimensional residual convolution layer and a multi-head self-attention layer. Based on the numerical weather forecast data, it outputs a future short-term wind power forecast sequence as the background field for ultra-short-term prediction.

[0165] Data acquisition module, used to obtain historical measured power data, measured wind speed data and short-term forecast background field data;

[0166] The data preprocessing module is used to preprocess the historical measured power data, measured wind speed data, and short-term forecast background field data, including data set construction, sliding windowing, and multi-stage data cleaning;

[0167] The AI ​​assimilation model training module is used to build a Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture. The first branch of this Resnet-BiLSTM AI assimilation model extracts the mutation characteristics of measured power and wind speed, and the second branch extracts the background field prediction characteristics. The dual-branch features are then integrated based on the bidirectional time series modeling unit. The Resnet-BiLSTM AI assimilation model is trained using the preprocessed dataset.

[0168] The ultra-short-term assimilation prediction module is used to input real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model, and output a rolling updated ultra-short-term power prediction sequence.

[0169] Embodiment 3: This embodiment provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions causing a computer to execute a wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed, the method comprising a first stage and a second stage of coordinated operation;

[0170] The first stage generates the short-term prediction background field, including:

[0171] Obtain numerical weather forecast data and perform spatiotemporal alignment preprocessing;

[0172] The preprocessed data is processed by the Resnet3D-Attention model containing a three-dimensional residual convolution layer and a multi-head self-attention layer to generate a continuous time series power prediction background field;

[0173] The second phase is to conduct dynamic assimilation correction of ultra-short-term wind power, including:

[0174] Obtain historical measured power data, measured wind speed data, and short-term forecast background field data to construct a data set; and perform data cleaning and sliding window processing on the data set;

[0175] A Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture was constructed. The first branch of the Resnet-BiLSTM AI assimilation model extracted the mutation characteristics of the measured power and wind speed, while the second branch extracted the background field prediction characteristics. The dual-branch features were then fused using a bidirectional time series modeling unit.

[0176] Use the preprocessed data set to train the Resnet-BiLSTM AI assimilation model; input the real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model to output a rolling updated ultra-short-term power forecast sequence.

[0177] Embodiment 4: This embodiment provides an electronic device, which may include: a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may call logic instructions in the memory to execute a wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed, the method comprising a first stage and a second stage of coordinated operation;

[0178] The first stage generates the short-term prediction background field, including:

[0179] Obtain numerical weather forecast data and perform spatiotemporal alignment preprocessing;

[0180] The preprocessed data is processed by the Resnet3D-Attention model containing a three-dimensional residual convolution layer and a multi-head self-attention layer to generate a continuous time series power prediction background field;

[0181] The second phase is to conduct dynamic assimilation correction of ultra-short-term wind power, including:

[0182] Obtain historical measured power data, measured wind speed data, and short-term forecast background field data to construct a data set; and perform data cleaning and sliding window processing on the data set;

[0183] A Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture was constructed. The first branch of the Resnet-BiLSTM AI assimilation model extracted the mutation characteristics of the measured power and wind speed, while the second branch extracted the background field prediction characteristics. The dual-branch features were then fused using a bidirectional time series modeling unit.

[0184] Use the preprocessed data set to train the Resnet-BiLSTM AI assimilation model; input the real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model to output a rolling updated ultra-short-term power forecast sequence.

[0185] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0186] Embodiment 5: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed. The method includes a first stage and a second stage of coordinated operation.

[0187] The first stage generates the short-term prediction background field, including:

[0188] Obtain numerical weather forecast data and perform spatiotemporal alignment preprocessing;

[0189] The preprocessed data is processed by the Resnet3D-Attention model containing a three-dimensional residual convolution layer and a multi-head self-attention layer to generate a continuous time series power prediction background field;

[0190] The second phase is to conduct dynamic assimilation correction of ultra-short-term wind power, including:

[0191] Obtain historical measured power data, measured wind speed data, and short-term forecast background field data to construct a data set; and perform data cleaning and sliding window processing on the data set;

[0192] A Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture was constructed. The first branch of the Resnet-BiLSTM AI assimilation model extracted the mutation characteristics of the measured power and wind speed, while the second branch extracted the background field prediction characteristics. The dual-branch features were then fused using a bidirectional time series modeling unit.

[0193] Use the preprocessed data set to train the Resnet-BiLSTM AI assimilation model; input the real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model to output a rolling updated ultra-short-term power forecast sequence.

[0194] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0195] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed, characterized in that: Includes Phase I and Phase II of collaborative work; The first stage generates the short-term prediction background field, including: Obtain numerical weather forecast data and perform spatiotemporal alignment preprocessing; The preprocessed data is processed by the Resnet3D-Attention model containing a three-dimensional residual convolution layer and a multi-head self-attention layer to generate a continuous time series power prediction background field; The second phase is to conduct dynamic assimilation correction of ultra-short-term wind power, including: Obtain historical measured power data, measured wind speed data, and short-term forecast background field data to construct a data set; and perform data cleaning and sliding window processing on the data set; A Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture was constructed. The first branch of this Resnet-BiLSTM AI assimilation model extracts the mutation characteristics of measured power and wind speed, while the second branch extracts the background field prediction characteristics. The two-branch features are then fused using a bidirectional time series modeling unit. Use the preprocessed data set to train the Resnet-BiLSTM AI assimilation model; input the real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model to output a rolling updated ultra-short-term power forecast sequence.

2. The method for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed according to claim 1 is characterized in that: Acquire numerical weather forecast data and perform spatiotemporal alignment preprocessing, including: The numerical weather forecast data were spatially resampled to a resolution of 0.1°x0.1° and temporally interpolated to the same temporal resolution as the measured power data. When creating samples, a sliding time window is constructed in the time dimension. Taking the target timestamp as the benchmark, the numerical weather forecast data of several time steps before and after are integrated to form a time series data of continuous time. At the same time, a 20x20 spatial grid area is selected around the target station to cover the terrain and meteorological field information around the station. Extract the near-surface layer parameters and wind vector parameters from the numerical weather forecast data and form a five-dimensional tensor as the input of the Resnet3D-Attention model, which is the batch size, feature dimension, time step, spatial height and spatial width; The five-dimensional tensor containing spatiotemporal information is Z-score normalized to eliminate dimensional differences. The Z-score normalization formula is: ； in, is the original data point, is the data mean, is the standard deviation.

3. The method for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed according to claim 1 is characterized in that: The Resnet3D-Attention model includes: A 3D residual convolutional layer for extracting spatiotemporal correlation features from numerical weather forecast data, and a multi-head self-attention layer for quantifying the contribution differences of key meteorological variables at different time steps and modeling temporal dependencies; The three-dimensional residual convolution layer adopts a residual block structure, and its output Expressed as: ； When the input and output dimensions are inconsistent, linear projection is used Perform dimension matching, expressed as: ； in, is the input of the residual block, represents the residual mapping that needs to be learned, is a learnable weight parameter; The multi-head self-attention layer projects the input into multiple low-dimensional subspaces and calculates multiple attentions in parallel, and then concatenates the results, so that the Resnet3D-Attention model can focus on different aspects of the input time series at the same time. The formula is: ； Each attention head The calculation formula is: ； The calculation formula of the attention mechanism is: ； in, 、 、 are query, key, and value matrices respectively; 、 、 It is The linear projection matrix of the attention heads, is the final linear projection matrix, is the dimension of the key vector.

4. The method for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed according to claim 1, characterized in that: Perform data cleaning and sliding window processing on the data set, specifically: Data cleaning uses a multi-stage data cleaning method that combines physical constraints with data distribution characteristics. This includes identifying and eliminating abnormal power-wind speed data points that do not conform to the wind turbine operating mechanism based on physical rules, and identifying and eliminating outliers in the data distribution using a density clustering algorithm based on the joint distribution of wind speed and power. Among them, the physical rule-based screening and identification includes logical verification based on the cut-in wind speed threshold and the minimum power threshold. When the wind speed is lower than the cut-in wind speed and the power is higher than the preset minimum power threshold, it is defined as an outlier. The density clustering algorithm uses the DBSCAN algorithm; The sliding window processing refers to the use of a rolling time window method to generate samples. For each time point in the data set, , which is defined as the end time of the measured data acquisition; As the starting point, several hours of measured data from the site are selected as the input of the first branch of the Resnet-BiLSTM AI assimilation model; the power prediction background field of several hours in the future is selected as the input of the second branch of the Resnet-BiLSTM AI assimilation model; at the same time, the measured power data in the same future time period is selected as the target label that the Resnet-BiLSTM AI assimilation model needs to predict.

5. The method for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed according to claim 4 is characterized in that: The Resnet-BiLSTM AI assimilation model includes: A dual-channel measured and background field feature extractor consists of two independent ResNet-based encoders as the first and second branches, respectively. The first branch extracts the mutation characteristics of the measured power and wind speed, while the second branch extracts the background field prediction characteristics. The ResNet structure performs one-dimensional convolution along the time dimension to capture local patterns and mutation characteristics in time series data, and connects them through residuals. The temporal dependency modeling module uses a bidirectional long short-term memory network to capture the long-term dependency and bidirectional contextual information of the measured feature sequence and the background field feature sequence; The information collaborative assimilation module concatenates the measured feature sequence and the background field feature sequence processed by their respective bidirectional long short-term memory network layers along the feature dimension to form a joint feature representation that integrates the two information streams. The joint feature is then input into a shared bidirectional long short-term memory network layer. The assimilation output module maps the output features of the shared bidirectional long short-term memory network layer to the final corrected ultra-short-term power prediction value through a fully connected layer.

6. The method for intelligent wind power prediction based on the coordinated assimilation of measured power and wind speed according to claim 5, characterized in that: Use the preprocessed dataset to train the Resnet-BiLSTM AI assimilation model, specifically: The historical measured power data, measured wind speed data and short-term forecast background field data after data cleaning and sliding window processing are divided into training set, validation set and test set; The stochastic gradient descent optimization algorithm is used to train the constructed Resnet-BiLSTM AI assimilation model with the training set. The hyperparameters used in training are set as follows: batch size is 32, number of training rounds is 200, and for model parameters , the update rule for gradient descent is: ； in, are model parameters, is the learning rate, Represents the objective function Parameters In a single training sample The gradient on The initial learning rate for model training is , using the StepLR learning rate scheduler to implement dynamic adjustment, in the After decay steps, the learning rate for: ； in, represents the attenuation coefficient; The loss function is the mean absolute error. , which measures the average absolute difference between the predicted value and the true value, is expressed as: ； in, is the sample size, It is True value, It is predicted values.

7. An intelligent wind power prediction system based on the coordinated assimilation of measured power and wind speed, characterized in that: include: The short-term prediction module receives numerical weather forecast data and builds a Resnet3D-Attention model consisting of a three-dimensional residual convolution layer and a multi-head self-attention layer. Based on the numerical weather forecast data, it outputs a future short-term wind power forecast sequence as the background field for ultra-short-term prediction. Data acquisition module, used to obtain historical measured power data, measured wind speed data and short-term forecast background field data; The data preprocessing module is used to preprocess the historical measured power data, measured wind speed data, and short-term forecast background field data, including data set construction, sliding windowing, and multi-stage data cleaning; An AI assimilation model training module is used to build a Resnet-BiLSTM AI assimilation model using a dual-branch deep learning architecture. The first branch of this Resnet-BiLSTM AI assimilation model extracts the mutation characteristics of measured power and wind speed, while the second branch extracts the background field prediction characteristics. The two-branch features are then integrated using a bidirectional time series modeling unit. And use the preprocessed dataset to train the Resnet-BiLSTM AI assimilation model; The ultra-short-term assimilation prediction module is used to input real-time background field data and real-time measured data into the trained Resnet-BiLSTM AI assimilation model, and output a rolling updated ultra-short-term power prediction sequence.

8. A non-transitory computer-readable storage medium, characterized in that Computer instructions are stored thereon, and the computer instructions enable the computer to execute the wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed as described in any one of claims 1-6.

9. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed as described in any one of claims 1-6.

10. A computer program product, characterized in that The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the wind power intelligent prediction method based on the coordinated assimilation of measured power and wind speed according to any one of claims 1 to 6.

Images