A u-net-based multi-mode wind energy resource monthly scale prediction correction method and system

Abstract

The application discloses a kind of multi-mode wind energy resource month scale prediction revision method and system based on U-Net, belong to wind energy resource evaluation and climate numerical prediction cross technical field, for the month scale wind speed of climate mode output is revised.The method constructs climatic wind speed field using ERA5 reanalysis, calculates month scale 10m wind speed anomaly as revision benchmark;Obtain the month scale historical return data of multiple dynamic climate models, extract 10m wind speed and multiple layers meteorological elements, uniform interpolation and standardization, form sample set.Based on sample set, the U-Net revision model with encoding and decoding structure is constructed, and the feature combination and hyperparameter are optimized through cross-validation, to learn the nonlinear mapping from multi-mode prediction field to ERA5 anomaly field.The optimal model is used to revise the future month scale prediction, to generate wind speed products with more similar spatial structure and amplitude distribution to observations, to provide high credible wind energy climate information for wind power planning, power generation planning and power grid dispatching.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross technical field of wind energy resource assessment and numerical weather prediction, and relates to multi-mode climate model monthly scale wind speed prediction bias correction and fine comprehensive assessment of wind energy resources and operational wind power planning application, in particular to monthly scale prediction of wind energy resources, a wind energy resource monthly prediction correction method and system based on a U-Net artificial intelligence model are proposed, so as to effectively improve the prediction accuracy. BACKGROUND

[0002] Compared with fossil energy, wind power generation has significant intermittency and volatility. With wind turbines in larger areas and higher proportions of grid connection, the safe and stable operation of the power grid puts higher requirements on the spatiotemporal predictability of wind energy resources. The safe operation, supply-demand matching and dispatching optimization of the energy system are closely related to the climate conditions. Developing operational wind energy climate prediction, especially accurate monthly scale wind speed prediction, has important reference value in wind farm site selection demonstration, supporting medium-term planning of power systems, monthly power generation plan compilation, maintenance arrangement and power market transaction strategy formulation.

[0003] The dynamic climate model widely used in the current business can provide monthly climate prediction, but there are still systematic biases and spatial details in the near-surface and wind turbine hub height wind speed: first, the model has limited ability to depict wind fields under complex terrain and underlying surface conditions, resulting in significant regional bias; second, the amplitude distribution and spatial pattern of monthly scale wind speed often appear systematic drift, which directly used in wind power business will introduce large errors; third, the resolution of the climate model is relatively coarse, which cannot meet the demand of high resolution and high accuracy for wind power site selection evaluation, regional consumption and monthly power generation capacity.

[0004] The existing post-processing methods mainly based on statistical regression or simple bias correction are difficult to simultaneously consider the spatial structure correction, amplitude distribution correction and comprehensive constraints of static factors such as terrain roughness, resulting in that the monthly scale products after correction still have deficiencies in business stability and usability. With the development of multi-model ensemble prediction and multi-source data assimilation technology, the joint use of multi-climate model and multi-physical scheme output has gradually become an important direction to improve the prediction ability. Although this kind of method can alleviate the systematic bias and accidental error of single model to some extent, it generally simplifies the multi-model information into a small amount of statistical characteristics or weighting coefficients, ignoring the complementarity and difference of different models in multi-level dynamic field, thermal field and underlying surface response. It is still difficult to finely correct the monthly scale anomaly from the overall structure of the spatial field.

[0005] In summary, the existing climate model monthly wind speed prediction still has obvious deficiencies in system bias, spatial structure restoration and business application, and therefore how to build a multi-climate model oriented wind energy resource monthly prediction correction technology with high precision spatial correction on the monthly scale, stable performance in different regions, seasons and prediction time, has become a technical problem to be solved for improving wind power consumption efficiency and ensuring power grid operation safety. SUMMARY

[0006] (I) Invention purposes

[0007] To solve the above problems of the prior art, the present application aims to provide a multi-model wind energy resource monthly scale prediction correction method and system based on U-Net, which is used to correct the spatial structure and amplitude distribution of the monthly wind speed output by the dynamic climate model, generate monthly wind speed products that are closer to the observed characteristics and have higher regional applicability, and thus improve the efficiency and operation safety of wind power consumption in the power grid. Specifically, the present application takes global monthly near-surface wind speed as the research object, uses reanalysis data and historical return data of mainstream dynamic climate models to systematically evaluate the skill performance of the model in monthly prediction; on this basis, through comparison, screening and cross-validation, a spatial mapping correction framework based on U-Net is constructed and optimized, which integrates ground temperature, sea level pressure, mid-high level circulation and multi-year climatic characteristics to realize the distribution correction of near-surface wind speed anomaly; accordingly, an objective monthly wind energy resource correction technology is formed for business, so that the spatial correlation coefficient of the corrected monthly near-surface wind speed in the global main wind area is significantly greater than the direct output of the dynamic model, and has the ability of robust generalization and engineering rolling operation across models and regions.

[0008] (II) Technical solutions

[0009] To achieve the purposes of the present application and solve its technical problems, the present application adopts the following technical solutions:

[0010] The first purpose of the present application is to provide a multi-model wind energy resource monthly scale prediction correction method based on U-Net, which is used to correct the spatial structure and amplitude distribution of the monthly wind speed output by the dynamic climate model, generate monthly wind speed products that are closer to the observed characteristics and have higher regional applicability, and the method at least includes the following steps when implemented:

[0011] S100. Baseline climatic state and observed anomaly preparation: extract the global monthly 10m wind speed data in the preset continuous multi-year period from the reanalysis dataset ERA5, calculate the multi-year average climatic state wind speed field of each grid point according to the same month of each year, and obtain the ERA5 wind speed anomaly field as the baseline and correction target by subtracting the climatic state wind speed of the corresponding month from the monthly wind speed of each year;

[0012] S200. Acquisition of multi-model multivariate return data: Acquire monthly historical return data of multiple climate models within the preset continuous multi-year time period, including at least 10-meter wind speed and geopotential height of multiple isobaric layers, zonal wind, meridional wind and surface air pressure, to form a multi-model forecast field;

[0013] S300. Data format standardization: The ERA5 wind speed anomaly field and the multi-model forecast field data are uniformly interpolated to a latitude and longitude grid of the same resolution, and the variables are standardized to form a sample set for model building training;

[0014] S400. Construction of U-Net deep learning correction model: Taking the multi-model forecast field as input and the ERA5 wind speed anomaly field as the correction target, a U-Net deep learning correction model based on an encoder-decoder structure convolutional neural network is constructed. By iteratively optimizing the network parameters, the model approximates the nonlinear spatial mapping relationship from the multi-model forecast field to the observed anomaly field.

[0015] S500. Time Series Cross-Validation and Model Optimization: The sample set is divided into multiple training and testing periods in chronological order. A continuous time period is selected as the test set in each period, and the remaining data is divided into training and validation sets. The correction model is trained using the training set. Based on the model's correction effect on the 10m wind speed anomaly on the validation set, the input feature combination and network hyperparameter configuration that maximize the improvement of the spatial correlation coefficient of the anomaly are selected as the optimal model parameters.

[0016] S600. Test set evaluation of the correction model effect: Based on the optimal model parameters, the multi-model forecast fields for multiple future months in the independent test set are corrected month by month, and the spatial correlation coefficient of anomalies is used to evaluate the improvement effect of the model output in different regions and months compared with the corresponding ERA5 wind speed anomaly field.

[0017] The second objective of this invention is to provide a monthly wind energy resource prediction correction system, based on the aforementioned U-Net-based multi-model monthly wind energy resource prediction correction method, comprising:

[0018] The baseline climatological and observational anomaly construction module is used to read global monthly 10m wind speed data for a preset continuous multi-year period from the reanalysis dataset ERA5, calculate the multi-year average climatological wind speed field for each grid point according to the same month of each year, and generate the ERA5 wind speed anomaly field by subtracting the corresponding monthly climatological wind speed from the wind speed at each year and month scale.

[0019] The multi-model multivariate return data acquisition module is used to acquire monthly-scale historical return data of multiple dynamic climate models within the stated time period to form a multi-model forecast field.

[0020] The data format standardization module is used to uniformly interpolate the ERA5 wind speed anomaly field and the multi-model forecast field to a latitude and longitude grid of the same resolution, and to perform uniform standardization processing on each physical quantity to form a sample set for model building training.

[0021] The U-Net deep learning correction model training module is used to construct a U-Net deep learning correction model based on an encoder-decoder structure convolutional neural network, with multi-model forecast fields as input and ERA5 wind speed anomaly fields as correction targets. The model approximates the nonlinear spatial mapping relationship from multi-model forecast fields to observed anomaly fields by iteratively optimizing network parameters.

[0022] The time-series cross-validation and prediction correction module is used to divide the sample set into multiple training and testing periods in chronological order. Within each period, the correction model is trained using the training set. Based on the model's correction effect on the 10m wind speed anomaly on the validation set, the optimal model parameters are determined by the combination of input features and network hyperparameters that maximize the improvement of the spatial correlation coefficient of the anomaly. Furthermore, based on the optimal model parameters, the multi-model forecast field of the independent test set and the target forecast period is corrected month by month, and the corrected monthly 10m wind speed anomaly prediction product is output.

[0023] (III) Technical Effects

[0024] This invention addresses the systematic bias in 10-meter wind speed anomaly predictions by multiple climate models. It employs a "multi-model output + deep learning correction" approach, constructing an error correction model based on the U-Net network. This effectively reduces the bias in global 10-meter wind speed anomaly predictions by climate models. Compared to direct model output, its advantages and effects are mainly reflected in the following aspects:

[0025] (1) This invention utilizes the nonlinear feature extraction and spatial mapping capabilities of the U-Net deep learning model to accurately learn and correct systematic errors in climate models. During training and testing, the model exhibited a good trend of loss convergence, with a significant decrease in mean square error. After correction, within a lead timeframe of 0 to 5 months, the spatial correlation coefficients of predicted anomalies for global, land, and ocean regions were significantly improved, indicating that the corrected wind speed anomalies are spatially closer to actual observations.

[0026] (2) The correction effect of this invention varies across different regions and seasons, demonstrating significant advantages. In particular, for the ocean region, which covers most of the Earth's surface, the spatial correlation coefficient of anomalies is improved most significantly, reaching a maximum of over 0.2, effectively compensating for the model's predictive shortcomings caused by sparse observational data over the ocean. At the same time, the model exhibits the strongest correction capability in winter months (such as January and December), indicating that this invention can effectively capture and correct the more significant errors of the model in the winter half-year, providing more reliable forecasting services for seasons with high wind speeds and high wind energy potential.

[0027] (3) This invention can be widely applied to climate prediction research, wind energy assessment, meteorological risk assessment and other fields. Through deep learning correction of multi-mode integrated output, it can provide more accurate 10-meter wind speed anomaly prediction products on monthly to seasonal scales, providing highly reliable climate information support for wind power production planning, energy storage configuration and monthly grid scheduling. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 Here is a flowchart of the U-Net-based multi-mode wind energy resource monthly-scale prediction correction method;

[0030] Figure 2 This is a schematic diagram of the U-Net deep learning network structure;

[0031] Figure 3 A schematic diagram showing the change of MSE for four climate models with the number of training cycles;

[0032] Figure 4 This diagram illustrates the changes in spatial correlation coefficients of four climate models a~d before and after correction in different regions (blue for global, yellow for land, and green for ocean). The horizontal axis represents different forecast lead times (months). global_ori represents the original model output for the global region (including all grid points for land and ocean), and global_pred represents the result after U-Net correction; land_ori represents the original model output for the land region, and land_pred represents the result after U-Net correction; ocean_ori represents the original model output for the ocean region, and ocean_pred represents the result after U-Net correction.

[0033] Figure 5This diagram illustrates the changes in global anomaly correlation coefficients for different months with a forecast lead time of one month. The horizontal axis represents the changes before and after corrections for different climate models, and the vertical axis represents the months. Detailed Implementation

[0034] This invention aims to provide a multi-model wind energy resource monthly-scale prediction correction method and system based on U-Net, used to perform integrated correction of the spatial structure and amplitude distribution of monthly-scale wind speeds output from dynamic climate models, generating monthly-scale wind speed products that more closely reflect observational characteristics and have higher regional applicability. To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. The described embodiments are some, but not all, embodiments of this invention, and are exemplary, intended to explain the invention, and should not be construed as limiting the invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0035] Example 1: A Multi-Model Wind Energy Resource Monthly Scale Forecast Correction Method Based on U-net

[0036] As a specific example, such as Figure 1 As shown in the embodiments of the present invention, the multi-model wind energy resource monthly-scale prediction correction method based on U-net mainly includes the following steps in its implementation: preparation of baseline climatological conditions and observational anomalies, acquisition of multi-model multivariate return data, data format standardization, construction of U-Net deep learning correction model, time series cross-validation and model optimization, and evaluation of the correction model effect using a test set. According to the actual technical operation process, this method can be divided into two stages: a preliminary data preparation and analysis stage and a modeling correction stage. The first stage includes the extraction and processing of monthly-scale return data and corresponding reanalysis data (ERA5) from 1993 to 2024 from four climate models (ECMWF, JMA, NCEP, and UKMO), and the analysis of model error characteristics. After integrating the required data, the second stage uses the U-Net deep learning method to correct the model output based on the model error distribution characteristics. The present invention provides an implementation example based on the U-Net deep learning method, which can effectively reduce the bias of climate models in predicting 10m wind speed anomalies globally. Specifically:

[0037] S100. Preparation of baseline climatological and observational anomaly data:

[0038] Global monthly 10m wind speed data for a predetermined continuous multi-year period were extracted from the ERA5 reanalysis dataset. The multi-year average climatological wind speed field for each grid point was calculated for the same month of each year. Anomalies were then processed by subtracting the corresponding month's climatological wind speed from the annual monthly wind speed to obtain the ERA5 wind speed anomaly field, serving as both a baseline and a correction target. The predetermined continuous multi-year period was selected to meet long-term stability requirements, with a length of no less than 20 consecutive calendar years. For ERA5 monthly 10m wind speed data with missing data or outliers, a quality control and imputation strategy combining spatial and temporal interpolation was employed. Outlier grid points that significantly deviated from the statistical distribution were removed or reconstructed to ensure that the resulting multi-year average climatological wind speed field and ERA5 wind speed anomaly field have statistical consistency and physical plausibility globally.

[0039] In this embodiment of the invention, global monthly 10-meter wind speed data for 30 years (1991-2020) is extracted from the ERA5 reanalysis dataset. The average value for each grid point and each month (e.g., all Januarys, all Februarys, etc.) over these 30 years is calculated to obtain the "1991-2020 climatological" data. Subsequently, for each month within the period of 1993-2024 that needs to be used, the actual ERA5 wind speed data for that month is subtracted from the corresponding climatological average value (e.g., the data for January 1995 minus the average value of 30 Januarys) to obtain the 10-meter wind speed anomaly field based on ERA5, which serves as the target value for model training and evaluation.

[0040] S200. Multimodal, multivariate return data collection:

[0041] Monthly historical return data from multiple climate models are acquired over a preset continuous multi-year period, including at least 10-meter wind speed and geopotential heights, zonal winds, meridional winds, and surface pressures from multiple isobars, forming a multi-model forecast field. The multiple climate models are selected from seasonal-subseasonal dynamic climate prediction systems from different forecast centers or different dynamic frameworks. The historical return data uses actual observations as the initial field, re-forecasting past dates to form a dataset containing several reporting times and different forecast lead times from 0 to 5 months. When acquiring monthly historical return data, the reporting times and forecast lead times of each model are uniformly organized and strictly aligned with the ERA5 wind speed anomaly field month by month in the time dimension. Furthermore, in addition to 10-meter wind speed, the historical return data also includes at least geopotential heights, zonal winds, meridional winds, and surface pressure variables from the 925hPa, 850hPa, 700hPa, and 500hPa isobars, constituting a multivariate input feature set characterizing the three-dimensional atmospheric circulation state.

[0042] In this embodiment of the invention, monthly historical forecast data released by four major climate prediction centers (ECMWF, JMA, NCEP, UKMO) between 1993 and 2024 were systematically collected. The forecast data are model "predictions" for past dates, with initial fields derived from historical observations. The collected variables include not only the target variable of 10-meter wind speed requiring correction, but also multiple levels of variables closely related to atmospheric circulation (such as geopotential height at 925 hPa, 850 hPa, 700 hPa, and 500 hPa, zonal wind, and meridional wind) and surface pressure. These variables collectively constitute a feature set describing the three-dimensional dynamic state of the atmosphere. A total of 379 forecast times were collected, each with a forecast lead of 6 periods, resulting in 2274 samples.

[0043] S300. Data format standardization:

[0044] The ERA5 wind speed anomaly field and the multi-model forecast field data were uniformly interpolated to a latitude and longitude grid of the same resolution, and the variables were standardized to form a sample set for model building and training. Preferably, the bilinear interpolation isopreserving structure interpolation algorithm is used to uniformly interpolate the ERA5 wind speed anomaly field data obtained in step S100 and all model return data (i.e., multi-model forecast field data) collected in step S200 onto a latitude and longitude grid between 0.5°×0.5° and 2.5°×2.5°. Preferably, the data is uniformly resampled onto the same 1°×1° latitude and longitude grid. All training data are standardized. Based on the statistics of the training period (such as mean, standard deviation, or quantile scale), Z-score standardization is performed on each input variable and target variable, which removes the mean and scales according to the standard deviation, or normalization is performed with robust scaling estimation. This reduces the impact of differences in the dimensions and numerical magnitudes between different models on network training. The same standardized parameters are strictly reused in the prediction stage to ensure that the numerical distribution of the network input and output is consistent with that in the training stage and avoids performance degradation caused by distribution drift.

[0045] S400. Constructing the U-Net deep learning correction model:

[0046] Using multi-model forecast fields as input and ERA5 wind speed anomaly fields as correction targets, a U-Net deep learning correction model based on an encoder-decoder structured convolutional neural network is constructed. By iteratively optimizing the network parameters, the model approximates the nonlinear spatial mapping relationship from multi-model forecast fields to observed anomaly fields.

[0047] Preferably, the U-Net deep learning correction model includes a preprocessing layer, an encoder, a decoder, and an output layer. The preprocessing layer improves the spatial resolution of the input multi-mode forecast field through interpolation and performs preliminary feature fusion through convolution and batch normalization. The encoder is composed of stacked multi-stage residual blocks and extracts cross-scale features layer by layer through downsampling and channel expansion. The decoder upsamples through deconvolution and makes skip connections with the corresponding layer of the encoder to fuse high-level semantic information and shallow spatial details. The output layer compresses the multi-channel features into a single-variable global 10-meter wind speed anomaly prediction field through convolution. Furthermore, the U-Net deep learning correction model's encoder-decoder structure includes at least three layers of downsampling-upsampling cascaded units. Each encoder unit consists of a convolutional layer, a nonlinear activation function, and a downsampling operation, while each decoder unit consists of an upsampling operation, a convolutional layer, and a nonlinear activation function. In cross-layer feature connections, feature maps at the corresponding scale at the encoder end and feature maps at the same scale at the decoder end are concatenated along the channel dimension to simultaneously preserve large-scale circulation background and small-to-medium-scale spatial structure information. Moreover, the size of the convolutional kernel, the number of layers, and the number of channels are set according to the sample spatial resolution and variable dimension, enabling the network to have sufficient expressive power to learn the complex nonlinear mapping relationship from multi-model forecast fields to the ERA5 wind speed anomaly field.

[0048] In a further preferred embodiment, in step S400, the mean squared error (MSE) is used as the loss function, and the network parameters are iteratively updated based on a stochastic gradient descent optimization algorithm combined with a backpropagation mechanism. During training, the ERA5 wind speed anomaly field is used as the label, and the difference between the model output and the corrected wind speed anomaly field is calculated at each grid point, the squares are summed, and the average value is taken as the current batch loss. The weights and bias coefficients of each convolutional layer, residual block, and transposed convolutional layer are updated according to this loss until the loss values ​​on the training set and validation set converge and remain below the preset error level.

[0049] Furthermore, the loss function employed not only includes the grid-by-grid mean square error between the predicted ERA5 wind speed anomaly field and the ERA5 wind speed anomaly field, but can also further include constraints on large-scale spatial structure, such as at least a correlation measure or a spectral energy distribution difference measure within a predetermined spatial window. By constructing a composite loss function through a weighted combination of amplitude error and spatial structure error, both overall bias and structural bias are suppressed during backpropagation to update network parameters, thereby improving the ability of the correction results to recover the regional wind field anomaly distribution pattern. This makes the monthly wind speed product output by the model more closely resemble the ERA5 observation characteristics in terms of spatial correlation structure.

[0050] Specifically, in constructing and training the U-Net deep learning correction model, this embodiment of the invention first constructs a convolutional neural network with a U-Net structure. The input to the model is the forecast field of several climate models (e.g., three-dimensional data containing multiple variables such as 10-meter wind speed, geopotential height of various isobaric layers, and wind field). The output of the model is the corrected global 10-meter wind speed anomaly field. The training target is the ERA5 wind speed anomaly field prepared in step S100. Through training with a large number of model input-ERA5 target data pairs, the network parameters are optimized using the backpropagation algorithm to make the correction result output by the model as close as possible to the ERA5 observations. This network adopts a U-Net encoder-decoder structure (…). Figure 2 (), used to map multivariate input fields to univariate high-resolution prediction fields.

[0051] First, the input multi-source meteorological variable field (with len_vars channels and an original resolution of approximately 180×360) is preprocessed using a preprocessing layer. Interpolation is used to uniformly enlarge the spatial resolution to 192×368, followed by a 3×3 convolutional layer and batch normalization to initially extract local features and perform scale normalization and feature fusion for different variables. Subsequently, the network further reduces the resolution and increases the number of channels using a 7×7 convolutional layer with a stride of 2 (in conjunction with BN and ReLU) and a 3×3 max-pooling layer, providing a larger receptive field for subsequent deep feature extraction.

[0052] The encoder (downsampling path) consists of four stacked residual blocks, similar in structure to ResNet-34: the first layer maintains 64 channels and stacks multiple residual blocks at the current scale; the second, third, and fourth layers expand the number of channels to 128, 256, and 512 respectively based on downsampling. Each layer contains multiple residual units, and residual connections mitigate the gradient vanishing problem while ensuring the expressive power of deeper networks, effectively capturing large-scale circulation background information and small-to-medium-scale spatial structure features. The decoder (upsampling path) is roughly symmetrical to the encoder, performing upsampling through layer-by-layer deconvolution (transposed convolution), and concatenating it with the feature maps of the corresponding encoding layers at each scale using skip connections to achieve the fusion of high-level semantic information and shallow spatial details. The upsampled features are then nonlinearly mapped and reconstructed through residual blocks, gradually restoring spatial resolution and refining the spatial structure of the target variable.

[0053] After upsampling to near the original resolution at the top layer, the network uses a 3×3 convolutional layer (with BN and ReLU) and a 13×9 convolutional layer to compress the number of channels to 1, and outputs a univariate 10m wind speed anomaly prediction field with a size of 180×360, achieving an end-to-end mapping from multivariate input to the spatial distribution of the target physical quantity. The entire network combines the deep residual feature extraction capability of ResNet with the multi-scale skip fusion advantage of U-Net, which can characterize the large-scale background field while preserving local detail information.

[0054] pass Figure 3 It can be observed that the training loss of each mode shows a rapid decreasing trend, converging quickly within the first 20 epochs and stabilizing in subsequent training. The losses of ECMWF, JMA, and UKMO decrease steadily, eventually converging to an extremely low error level (<0.01 m / s). Although the NCEP mode has good overall convergence performance, a significant increase in loss occurs around the 70th epoch, which may be related to instability or abnormal samples during training, but it is able to recover to a stable state afterward.

[0055] S500. Optimize the model through cross-validation:

[0056] To improve the model's generalization ability over time, time-series cross-validation was used to train and test the model. The sample set was divided into multiple training and testing periods in chronological order. Each period selected a continuous time segment as the test set, and the remaining data was divided into training and validation sets according to a preset ratio. The correction model was trained using the training set. Based on the model's correction effect on 10m wind speed anomalies on the validation set, the optimal model parameters were selected by combining the input feature combination and network hyperparameter configuration that maximized the improvement of the anomaly spatial correlation coefficient (ACC).

[0057] The time-series cross-validation and model optimization process employs a rolling window strategy to construct multiple training and testing cycles. In each cycle, the test set consists of several consecutive years, while the training and validation sets are divided chronologically by the remaining years. During model optimization, for different combinations of candidate input features and network hyperparameter configurations, multidimensional evaluation indicators, including at least spatial correlation coefficient, mean square error, and regional average deviation, are calculated on the validation set of each cycle. The scheme that performs stably and has the best comprehensive indicators in most cycles is selected through weighted comprehensive scoring or multi-objective ranking, ensuring that the obtained optimal model parameters have robust generalization ability in the time dimension.

[0058] In this embodiment of the invention, data from 1993 to 2024 is divided into training and validation sets. The model is trained on the training set and its performance is evaluated on the validation set. For each experiment, 60 months are selected chronologically as the test set (if less than 60 months, all available data are used), and the remaining data is used for training and validation (80% / 20%). For example, in the first experiment, January 1993 to December 1997 is the test set, January 1998 to July 2024 is the training + validation set, and so on, with 7 experiments conducted for each model. Different combinations of input features are systematically tested (e.g., using only 10-meter wind speed, adding geopotential height field, adding wind field, etc.), and network hyperparameters (such as network depth, learning rate, number of filters, etc.) are adjusted. Based on the model's correction effect on the validation set (the degree of improvement in the spatial correlation coefficient (ACC)), the optimal combination of input features and model parameters are selected.

[0059] S600. Evaluate the effectiveness of the corrected model using an independent test set:

[0060] Using the optimal model parameters obtained in step S500, the multi-model forecast fields for multiple future months in the independent test dataset (data not involved in training and validation) are corrected month by month. The spatial correlation coefficient of anomalies is used to evaluate the improvement effect of the model output in different regions and months compared with the corresponding ERA5 wind speed anomaly fields. The system evaluates the improvement of the corrected results compared with the original model forecast.

[0061] As a preferred approach, independent evaluation sets were constructed for each forecast lead time from 0 to 5 months, and statistical analysis was performed according to global, land, and ocean regions. For each forecast lead time and spatial region, the spatial correlation coefficient of the 10m wind speed anomaly before and after correction relative to the ERA5 wind speed anomaly was calculated in the test set. The variation characteristics of the correction effect with forecast lead time and spatial region were analyzed to verify the stability of the U-Net deep learning correction model under multiple lead time and multiple regional conditions. In addition, the preferred method for calculating the spatial correlation coefficient of anomaly is as follows: within a given evaluation region and evaluation month, the corrected 10m wind speed anomaly and the ERA5 wind speed anomaly are expanded into two sets of time series at all grid points. Pearson correlation coefficients are constructed based on covariance and variance, and area weights are used to perform weighted averaging of each grid point to obtain the spatial correlation coefficient of anomaly at the regional scale. During the model evaluation phase, the spatial correlation coefficients of anomaly for different months, different climate models, and different forecast lead times are statistically compared to quantify the degree of improvement in spatial pattern preservation and anomalous signal enhancement relative to the original model output.

[0062] In this embodiment of the invention, the evaluation includes: analyzing the changes in correction effectiveness from 0 to 5 months with different forecast lead times; and analyzing the model's correction effectiveness in different geographical regions, such as tropical, mid-latitude, and high-latitude areas, as well as on land and at sea. Figure 4 This reflects the changes in anomaly correlation coefficients before and after correction across different regions. It can be seen that for all models, across the global, land, and ocean regions, the corrected spatial correlation coefficients are generally higher than the original ones. The spatial correlation coefficient scores for the ocean region are significantly higher than those for the land region, especially in ECMWF and UKMO. For example, in the ECMWF model with a 0-month lead time, the corrected spatial correlation coefficient for the ocean is close to 0.2, while the corrected spatial correlation coefficient for the land region is less than 0.1. The decrease in the corrected spatial correlation coefficient with forecast lead time is relatively gradual, indicating that the model has a certain degree of cross-timeframe generalization ability. Figure 5 The changes in spatial correlation coefficients of anomalies were compared for different months with a forecast lead time of one month. Compared to the original model output, the corrected spatial correlation coefficients of anomalies in most months were significantly improved for all models, with the color changing from light to dark. This indicates that the deep learning model (U-Net) can effectively improve the correlation of global wind speed forecasts for different months. The winter half-year (November, December, and January) generally showed higher spatial correlation coefficient values, with the improvement after correction being particularly significant. In areas with low spatial correlation coefficients in the original model, such as June–August, the corrected values, although still relatively low, showed a significant darkening of the color, indicating that the U-Net correction mechanism still has a positive improving effect during periods of low forecast capability.

[0063] S700. Conversion of correction results to wind energy business applications (optional):

[0064] In a preferred embodiment of the present invention, a further step S700 for converting the correction results into wind energy business applications may be included. First, based on the monthly 10m wind speed anomaly field corrected by the U-Net deep learning correction model, the multi-year average climatological wind speed field of the corresponding month is superimposed to recover the corrected absolute wind speed field. Then, in the target wind farm or planning area, the absolute wind speed field is statistically processed in combination with local altitude and air density parameters, and converted into monthly wind power density or equivalent wind energy resource abundance index according to the functional relationship between wind power density and wind speed. Finally, the unit capacity factor and the corresponding monthly power generation are estimated by combining the power curve, installed capacity and operating constraints of the wind turbine, and the power generation sequence is provided as input to the monthly supply and demand balance analysis module of the power generation production planning, energy storage device capacity configuration and / or grid dispatching department, so as to realize the quantitative connection between the wind energy resource climate correction results and power system engineering applications.

[0065] Example 2: Monthly-scale Wind Energy Resource Prediction Correction System

[0066] In addition to the above-mentioned example of a multi-model wind energy resource monthly-scale prediction correction method based on U-Net, this embodiment of the invention further provides a wind energy resource monthly-scale prediction correction system corresponding to the above method. This system mainly includes several module units such as a baseline climatology and observation anomaly construction module, a multi-model multivariate return data acquisition module, a data format standardization module, a U-Net deep learning correction model training module, and a time series cross-validation and prediction correction module, wherein:

[0067] The baseline climatological and observational anomaly construction module is used to read global monthly 10m wind speed data for a preset continuous multi-year period from the reanalysis dataset ERA5, calculate the multi-year average climatological wind speed field for each grid point according to the same month of each year, and generate the ERA5 wind speed anomaly field by subtracting the corresponding monthly climatological wind speed from the wind speed at each year and month scale.

[0068] The multi-model, multivariate return data acquisition module is used to acquire monthly historical return data of multiple dynamic climate models (such as ECMWF, JMA, NCEP, UKMO) within the specified time period, including 10-meter wind speed and geopotential height, zonal wind, meridional wind and surface pressure of multiple isobaric layers (such as 925 / 850 / 700 / 500hPa), forming a multi-model, multivariate, and multi-time-dependent forecast field dataset.

[0069] The data format standardization module is used to uniformly interpolate the ERA5 wind speed anomaly field and the multi-model forecast field to a latitude and longitude grid of the same resolution, and to perform uniform standardization processing on each physical quantity to form a sample set for training the U-Net correction model.

[0070] The U-Net deep learning correction model training module is used to construct a U-Net deep learning correction model based on an encoder-decoder structure convolutional neural network, with multi-model forecast fields as input and ERA5 wind speed anomaly fields as correction targets. The model approximates the nonlinear spatial mapping relationship from multi-model forecast fields to observed anomaly fields by iteratively optimizing network parameters.

[0071] The time-series cross-validation and prediction correction module is used to divide the sample set into multiple training and testing periods in chronological order. Within each period, the correction model is trained using the training set. Based on the model's correction effect on the 10m wind speed anomaly on the validation set, the optimal model parameters are determined by the combination of input features and network hyperparameters that maximize the improvement of the spatial correlation coefficient of the anomaly. Furthermore, based on the optimal model parameters, the multi-model forecast field of the independent test set and the target forecast period is corrected month by month, and the corrected monthly 10m wind speed anomaly prediction product is output.

[0072] In summary, the proposed solution uses monthly raw forecast fields from climate models as the main input, incorporating climatological features such as surface air temperature, sea level pressure, mid-to-upper-level circulation, and multi-year climatology to construct an end-to-end U-Net spatial mapping network. This enables multi-source information fusion and nonlinear bias correction for monthly-scale wind speed fields. The method can: learn the spatial correspondence between model-predicted wind speeds and reanalysis data wind speeds, correcting regional systematic biases and structural errors; restore small-to-medium-scale details while maintaining large-scale background consistency through encoding-decoding and cross-layer feature splicing mechanisms; improve the statistical consistency and physical rationality of corrected wind speeds by combining amplitude distribution constraints and climatological constraints; and scalably integrate multi-model inputs and training from different time periods to achieve robust generalization across models, regions, and seasons. Through these processes, the invention can output monthly near-surface / hub-height wind speed products that match operational needs and further provide uncertainty assessments, offering highly reliable climate information support for wind power production planning, energy storage configuration, and monthly grid scheduling.

[0073] The objectives of this invention have been fully and effectively achieved through the above embodiments. Those skilled in the art will understand that this invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments described above. Although the invention has been described with reference to what is currently considered the most practical and preferred embodiments, it should be understood that the invention is not limited to the disclosed embodiments, and any modifications that do not depart from the functional and structural principles of the invention will be included within the scope of the claims.

Claims

1. A multi-model wind energy resource monthly-scale prediction correction method based on U-Net, characterized in that, It should include at least the following steps: S100. Extract global monthly 10m wind speed data for a preset continuous multi-year period from the reanalysis dataset ERA5, calculate the multi-year average climatological wind speed field for each grid point, and obtain the ERA5 wind speed anomaly field as the baseline and correction target by subtracting the corresponding month's climatological wind speed from the wind speed at each year and month scale. S200. Obtain monthly historical return data of multiple climate models within the preset continuous multi-year time period, including at least 10-meter wind speed and geopotential height of multiple isobaric layers, zonal wind, meridional wind and surface air pressure, to form a multi-model forecast field; S300. Interpolate the ERA5 wind speed anomaly field and the multi-model forecast field data to a latitude and longitude grid of the same resolution, and standardize each variable to form a sample set for model building training; S400. Using the multi-model forecast field as input and the ERA5 wind speed anomaly field as the correction target, a U-Net deep learning correction model based on an encoder-decoder structure convolutional neural network is constructed. The network parameters are iteratively optimized to make the model approximate the nonlinear spatial mapping relationship from the multi-model forecast field to the observed anomaly field. S500. Divide the sample set into multiple training and testing periods. Take a continuous time period as the test set in each period. Divide the remaining data into training set and validation set. Use the training set to train the correction model. Based on the correction effect of the model on the 10m wind speed anomaly on the validation set, select the input feature combination and network hyperparameter configuration that maximizes the improvement of the spatial correlation coefficient of the anomaly as the optimal model parameters. S600. Based on the optimal model parameters, the multi-model forecast fields for multiple future months in the independent test set are corrected month by month, and the spatial correlation coefficient of anomalies is used to evaluate the improvement effect of the model outputs for different regions and different months compared with the corresponding ERA5 wind speed anomaly fields.

2. The method according to claim 1, characterized in that, In step S100, the selection of a continuous multi-year time period is preset to meet the long-term stability requirements, and its length is not less than 20 consecutive natural years. For ERA 5-month-scale 10m wind speed data with missing measurements or outliers, a quality control and imputation strategy combining spatial interpolation and temporal interpolation is adopted to remove or reconstruct outlier grid points that deviate significantly from the statistical distribution after verification.

3. The method according to claim 1, characterized in that, In step S200, multiple climate models are selected from seasonal-subseasonal dynamic climate prediction systems of different forecast centers or different dynamic frameworks; historical return data uses actual observations as the initial field to re-forecast past dates, forming a data set containing several reporting times and different forecast lead times of 0 to 5 months. When acquiring monthly-scale historical return data, the reporting time and forecast lead time of each model are uniformly organized and strictly aligned with the ERA5 wind speed anomaly field month by month in the time dimension. In addition to 10m wind speed, the historical return data also includes at least the geopotential height of the 925hPa, 850hPa, 700hPa and 500hPa isobaric layers, zonal wind, meridional wind and surface air pressure variables, constituting a multivariate input feature set characterizing the three-dimensional atmospheric circulation state.

4. The method according to claim 1, characterized in that, In step S300, a structure-preserving interpolation algorithm is used to achieve spatial resampling from the original grids of different modes to a unified regular latitude and longitude grid, and the spatial resolution of the unified grid is between 0.5°×0.5° and 2.5°×2.5°; in the variable standardization process, Z-score standardization with mean removal and scaling by standard deviation is performed on each input variable and target variable based on the statistics of the training time period, or normalization with robust scaling estimation is performed.

5. The method according to claim 1, characterized in that, In step S400, the U-Net deep learning correction model includes a preprocessing layer, an encoder, a decoder, and an output layer. The preprocessing layer improves the spatial resolution of the input multi-mode forecast field through interpolation and performs preliminary feature fusion through convolution and batch normalization. The encoder is composed of stacked multi-stage residual blocks and extracts cross-scale features layer by layer through downsampling and channel expansion. The decoder upsamples through deconvolution and makes skip connections with the corresponding layer of the encoder to fuse high-level semantic information and shallow spatial details. The output layer compresses the multi-channel features into a single-variable global 10-meter wind speed anomaly prediction field through convolution.

6. The method according to claim 5, characterized in that, In step S400, the encoder-decoder structure of the U-Net deep learning correction model includes at least three layers of downsampling-upsampling cascaded units. Each encoder unit consists of a convolutional layer, a non-linear activation function, and a downsampling operation, while each decoder unit consists of an upsampling operation, a convolutional layer, and a non-linear activation function. In cross-layer feature connections, the feature map at the corresponding scale of the encoder end and the feature map at the same scale of the decoder end are concatenated along the channel dimension to simultaneously preserve large-scale circulation background and small-to-medium-scale spatial structure information. Furthermore, the size, number of layers, and number of channels of each convolutional kernel are set according to the sample space resolution and variable dimension.

7. The method according to claim 6, characterized in that, In step S400, the mean squared error (MSE) is used as the loss function, and the network parameters are iteratively updated based on the stochastic gradient descent optimization algorithm combined with the backpropagation mechanism. During the training process, the ERA5 wind speed anomaly field is used as the label, and the difference between the model output and the corrected wind speed anomaly field is calculated at each grid point, the squares are summed, and the average value is taken as the current batch loss. The weights and bias coefficients of each convolutional layer, residual block, and transposed convolutional layer are updated according to the loss value until the loss values ​​on the training set and validation set converge and remain below the preset error level.

8. The method according to claim 1, characterized in that, Step S500 uses a rolling window strategy to construct multiple training and testing cycles. In each cycle, the test set consists of several consecutive years, and the training set and validation set are divided by the remaining years in chronological order. During model optimization, for different combinations of candidate input features and network hyperparameter configurations, multidimensional evaluation indicators, including at least spatial correlation coefficient, mean square error, and regional average deviation, are calculated on the validation set of each period. The scheme with stable performance and optimal comprehensive indicators in most periods is selected by weighted comprehensive scoring or multi-objective ranking.

9. The method according to claim 1, characterized in that, In step S600, when evaluating the correction effect of the optimal model, independent evaluation sets are constructed for each forecast lead period from 0 to 5 months, and statistical analysis is performed according to global overall region, land region and ocean region. For each forecast lead period and spatial region, the spatial correlation coefficient of the 10m wind speed anomaly field before and after correction with respect to the ERA5 wind speed anomaly field is calculated in the test set. The characteristics of the change of correction effect with forecast lead time and spatial region are analyzed to verify the stability of the U-Net deep learning correction model under multiple lead periods and multiple regions.

10. The method according to claim 1 or 9, characterized in that, In step S600, the calculation of the spatial correlation coefficient of anomalies includes: within a given assessment area and assessment month, the corrected 10m wind speed anomaly field and the ERA5 wind speed anomaly field are expanded into two sets of time series at all grid points. The Pearson correlation coefficient is constructed based on the covariance and variance, and the area weight is used to perform a weighted average of each grid point to obtain the spatial correlation coefficient of anomalies at the regional scale. In the model evaluation stage, the spatial correlation coefficients of anomalies for different months, different climate models, and different forecast lead times are statistically compared to quantify the degree of improvement in spatial pattern preservation and anomalous signal enhancement relative to the original model output.

11. The method according to claim 1, characterized in that, It also includes step S700 for converting the correction results into wind energy business applications, including: The corrected absolute wind speed field is obtained by superimposing the multi-year average climatological wind speed field of the corresponding month on the monthly 10m wind speed anomaly field corrected by the U-Net deep learning correction model. Within the target wind farm or planning area, the absolute wind speed field is statistically processed in conjunction with local altitude and air density parameters, and converted into monthly wind power density or equivalent wind energy resource abundance index based on the functional relationship between wind power density and wind speed. The unit capacity factor and corresponding monthly power generation are estimated by combining the power curve, installed capacity and operating constraints of the wind turbine. The power generation sequence is then used as input to the monthly supply and demand balance analysis module of the power generation production planning, energy storage device capacity configuration and / or grid dispatching departments.

12. A wind energy resource monthly-scale prediction correction system, based on the U-Net-based multi-model wind energy resource monthly-scale prediction correction method according to any one of claims 1 to 11, characterized in that, include: The baseline climatological and observational anomaly construction module is used to extract global monthly 10m wind speed data for a preset continuous multi-year period in the reanalysis dataset ERA5, calculate the multi-year average climatological wind speed field for each grid point, and obtain the ERA5 wind speed anomaly field as the baseline and correction target by subtracting the corresponding monthly climatological wind speed from the wind speed at each year and month scale. The multi-model multivariate return data acquisition module is used to acquire monthly historical return data of multiple climate models within the preset continuous multi-year time period, including at least 10-meter wind speed and geopotential height of multiple isobaric layers, zonal wind, meridional wind and surface air pressure, to form a multi-model forecast field. The data format standardization module is used to interpolate the ERA5 wind speed anomaly field and multi-model forecast field data to a latitude and longitude grid of the same resolution, and to standardize each variable to form a sample set for model building training. The U-Net deep learning correction model training module is used to construct a U-Net deep learning correction model based on an encoder-decoder structure convolutional neural network, taking the multi-model forecast field as input and the ERA5 wind speed anomaly field as the correction target. By iteratively optimizing the network parameters, the model can approximate the nonlinear spatial mapping relationship from the multi-model forecast field to the observed anomaly field. The time-series cross-validation and prediction correction module is used to divide the sample set into multiple training and testing periods. Each period takes a continuous time period as the test set, and the remaining data is divided into training and validation sets. The correction model is trained using the training set. Based on the model's correction effect on the 10m wind speed anomaly on the validation set, the input feature combination and network hyperparameter configuration that maximizes the improvement of the anomaly spatial correlation coefficient are selected as the optimal model parameters. Based on the optimal model parameters, the multi-model forecast fields for multiple months in the independent test set are corrected month by month, and the anomaly spatial correlation coefficient is used to evaluate the improvement effect of the model output in different regions and months compared with the corresponding ERA5 wind speed anomaly field.

Images