A method for high-frequency short-term wind speed prediction in complex terrain

Abstract

This invention discloses a high-frequency short-term wind speed prediction method for complex terrain. This method is based on collecting numerical weather prediction data from different isobaric surfaces and constructing a high-frequency short-term wind speed prediction model containing CNN and LSTM layers. The CNN layer captures the common spatial correlations of wind speed, wind direction, and related data, generating feature maps. Two LSTM layers extract temporal dynamic features from the wind speed and wind direction feature maps, respectively. After integrating the features, the fully connected layer outputs the wind speed prediction result. The model is trained using a training set, and the parameters are evaluated and adjusted using a validation set, with optimization achieved through a joint loss function. The high-frequency short-term wind speed prediction model of this invention exhibits superior performance in high-frequency short-term wind speed prediction, overcomes the spatiotemporal limitations of physical models, reduces reliance on historical data, and achieves more reliable and accurate short-term wind speed prediction in complex terrain. This has significant implications for wind energy management, grid stability, and energy production optimization.

Description

Technical Field

[0001] This invention relates to the field of wind speed prediction technology, and more specifically, to a method for short-term wind speed prediction in complex terrain based on spatiotemporal multi-task joint optimization. Background Technology

[0002] Wind energy is a crucial component of the global energy transition; however, its inherent volatility and intermittency pose significant challenges to its integration into the power grid. Since wind energy is proportional to the cube of wind speed, even small fluctuations in wind speed can lead to significant changes in wind energy. Therefore, accurate wind speed forecasting is essential for effective wind energy management, ensuring grid stability, optimizing energy production, and reducing operating costs. Traditional wind speed forecasting methods are mainly divided into two categories: physical models and data-driven methods. Physical models, particularly those based on numerical weather prediction (NWP), use real-time meteorological and geographical data to mathematically simulate the atmosphere and predict wind speed. NWP models perform well in medium- to long-term wind speed forecasting; however, this method involves a large amount of complex calculations, resulting in low spatiotemporal resolution and making it unsuitable for short-term and high-frequency forecasts. Data-driven models (including statistical and machine learning methods) rely on historical wind speed information for correlation prediction. Traditional statistical models, such as the Autoregressive Integral Moving Average (ARIMA) model and its extended models, have been widely used, but due to their inherent linear assumptions, they often struggle to capture the nonlinear trends of dynamic wind speed changes. Summary of the Invention

[0003] The present invention provides a method for predicting short-term wind speed in complex terrain based on spatiotemporal multi-task joint optimization, which can solve the above-mentioned problems.

[0004] To solve the above problems, the technical solution adopted by the present invention is as follows: A method for predicting high-frequency short-term wind speeds in complex terrain includes the following steps: (1) Collect numerical weather forecast data of different isobaric surfaces, then perform the first preprocessing to form model forecast input data, merge the model forecast input data and the data of the observation automatic station equipment, then perform the second preprocessing, and then divide the training set and validation set from them. (2) Construct a high-frequency short-term wind speed prediction model, which includes an input layer, a shared layer, independent layers, a fully connected layer, and an output layer. The shared layer is a CNN layer, used to capture the common spatial correlation between wind speed, wind direction, and numerical weather forecast data of different isobaric surfaces, and to obtain wind speed-related spatial feature maps and wind direction-related spatial feature maps. There are two independent layers, both of which are LSTM layers, used to capture the temporal dynamic features of wind speed from the wind speed-related spatial feature maps and the temporal dynamic features of wind direction from the wind direction-related spatial feature maps, respectively. The fully connected layer contains two hidden layers, used to obtain the integrated wind speed features based on the temporal dynamic features of wind speed and the integrated wind direction features based on the temporal dynamic features of wind direction, respectively. The output layer is used to obtain the wind speed prediction result based on the integrated wind speed features and wind direction features. (3) Based on the multi-prediction task joint optimization method, the high-frequency short-term wind speed prediction model is trained using the training set. During the training process, the high-frequency short-term wind speed prediction model is evaluated using the validation set and the model parameters are adjusted. The loss function used during the training process is the weighted loss function of the joint wind speed loss function and the wind direction loss function. (4) Use the trained high-frequency short-term wind speed prediction model to predict the future wind speed of the target site.

[0005] As a further description of the above technical solution, in step (1), the numerical weather prediction data includes the average wind speed, zonal wind component, meridional wind component, and average wind speed at the isobaric surfaces of 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, and 100hPa, respectively; the numerical weather prediction data are integrated into a unified 15-minute interval dataset according to different time resolutions.

[0006] As a further description of the above technical solution, in step (1), the method for collecting numerical weather forecast data of different isobaric surfaces includes: inputting the coordinates of the wind speed observation device, finding the coordinates near 3KM of the numerical model weather forecast as the grid center, finding the grid within the coordinates and the adjacent grids nearby, and then extracting the numerical weather forecast data of different isobaric surfaces.

[0007] As a further description of the above technical solution, in step (1), the first preprocessing includes handling missing and outlier values, as well as normalizing the variables; the second preprocessing is to remove model weather forecast data with a correlation coefficient less than the threshold of 0.2 by using the Pearson correlation coefficient.

[0008] As a further description of the above technical solution, step (4) also includes applying inverse normalization processing to the wind speed prediction results.

[0009] As a further description of the above technical solution, in step (2), the high-frequency short-term wind speed prediction model framework is determined based on grid search.

[0010] As a further description of the above technical solution, the CNN layer consists of two hidden layers. Each hidden layer of the CNN layer uses 16 channels and the convolution kernel size is 3×3; each hidden layer of the fully connected layer has 100 neurons.

[0011] As a further description of the above technical solution, during the training of the high-frequency short-term wind speed prediction model, the Dropout rate is set to 0.25; the weights of the wind speed loss function and the wind direction loss function are 0.7 and 0.3, respectively. During the training of the high-frequency short-term wind speed prediction model, the learning rate is adjusted using a multinomial decay strategy based on the Adam optimizer, with an initial learning rate of 0.001 and a weight decay parameter of 0.1. The high-frequency short-term wind speed prediction model is used for 6-hour, 12-hour, and 24-hour prediction tasks. The batch size for the 6-hour prediction task is 12, for the 12-hour prediction task it is 32, and for the 24-hour prediction task it is 64. The entire training of the high-frequency short-term wind speed prediction model is set to 500 cycles, and an early stopping mechanism is introduced to avoid model overfitting.

[0012] As a further description of the above technical solution, in step (3), the server configuration used for training the high-frequency short-term wind speed prediction model is: Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz CPU, Quadro RTX 6000 GPU, and Ubuntu 16.04.6 LTS operating system.

[0013] As a further description of the above technical solution, the model evaluation indicators used in step (3) include mean absolute error and root mean square error.

[0014] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention proposes an innovative model framework for high-frequency short-term wind speed prediction in complex terrain, which effectively solves the limitations of traditional wind speed prediction methods in short-term and high-frequency prediction. By combining CNN layers and LSTM layers, the model can simultaneously capture the spatial correlation and temporal dynamic characteristics between wind speed, wind direction, and numerical weather forecast data of different isobaric surfaces, thereby achieving more accurate wind speed prediction.

[0015] (2) The high-frequency short-term wind speed prediction model WindNet in this invention is trained by a weighted loss function that combines the wind speed loss function and the wind direction loss function, which further optimizes the model parameters. This enables WindNet to achieve more stable and accurate short-term wind speed prediction in complex terrain, providing strong support for grid stability, energy production and operation cost reduction, and has significant practical value and application prospects.

[0016] (3) The WindNet model of this invention exhibits superior performance under various complex terrains and extreme wind speed variations. Compared with benchmark models such as the traditional NWP model, LSTM model, and CNN-LSTM model, WindNet demonstrates higher accuracy in 6-hour, 12-hour, and 24-hour predictions. This not only overcomes the shortcomings of physical models in terms of spatiotemporal resolution but also reduces the reliance on historical data, providing more reliable prediction results for wind energy management.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, embodiments of the present invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the high-frequency short-term wind speed prediction method for complex terrain described in the embodiments of the present invention; Figure 2 This is a framework diagram of the high-frequency short-term wind speed prediction model described in the embodiments of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0021] Please refer to Figure 1 and Figure 2 This invention provides a method for predicting high-frequency short-term wind speeds in complex terrain, comprising the following steps: (1) Input the coordinates of the wind speed observation equipment, find the coordinates of the numerical model weather forecast near 3KM as the grid center, find the grid within the coordinates and the adjacent grids nearby, extract and collect numerical weather forecast data of different isobaric surfaces, and then perform the first preprocessing to form the model forecast input data; merge the model forecast input data and the data of the observation automatic station equipment, and then perform the second preprocessing, and divide them into training set, validation set and test set.

[0022] In this embodiment, numerical weather prediction data were collected from three automatic weather stations located near the wind farm between January 1 and July 31, 2024. Stations S1 and S2 are located in the lower reaches of the Jinsha River in Sichuan Province, with station S1 situated in a mountainous area and station S2 in a canyon. These stations are located in the dry-hot valley region of the Jinsha River, an area prone to localized convective weather and high temperatures. Station S3 is located in a plain near Panzhihua City, Sichuan Province, in a subtropical monsoon climate zone. All automatic weather stations are managed by the National Atmospheric Monitoring Centre of China. Raw wind speed (m / s) and wind direction (degrees) data were recorded at 15-minute intervals. The numerical weather prediction (NWP) data used in this invention was generated by the Weather Research and Forecasting (WRF) model (version 3.9.1) with a horizontal resolution of 3 km × 3 km and a temporal resolution of 1 hour. NWP forecasts are provided every 24 hours (i.e., NWP predicts weather conditions for the next 24 hours). The forecast includes the average wind speed (m / s), zonal wind component U and meridional wind component V at different isobaric surfaces (850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 100hPa), as well as the average wind speed (m / s) at a height of 10 meters (i.e., the height of the monitoring station).

[0023] In this embodiment, the method for the first preprocessing of the collected numerical weather forecast data includes: 1) Handling missing and outlier values; 2) Furthermore, due to the differences in order of magnitude and units among different meteorological variables, this invention also uses the min-max normalization method to normalize the variables during preprocessing, thereby ensuring that all training features have the same weights and accelerating model training. The specific formula is as follows: , Where, x min and x max Let x represent the minimum and maximum values ​​of each variable in the sample data, respectively. Let x represent a sample from the original data. n This represents the normalized sample. After normalization, the values ​​of all meteorological variables are within the consistent range of [0, 1].

[0024] The normalized data is the input data for the model forecast. (It should be noted that because normalization is performed here, after obtaining the wind speed forecast results using the high-frequency short-term wind speed forecast model, inverse normalization needs to be applied to the wind speed forecast results to ensure that the final wind speed forecast results have physical meaning.)

[0025] In this embodiment, in order to improve the temporal resolution of the data, the present invention synthesizes the model forecast input data and the data from the automatic observation station equipment according to a 15-minute time resolution.

[0026] The synthesized data underwent a second preprocessing step, in which model weather forecast data with a correlation coefficient less than the threshold of 0.2 were removed. The final numerical weather forecast dataset contained 20,352 records.

[0027] After the second preprocessing, samples can be generated using a sliding window based on the obtained numerical weather forecast dataset. The window slides forward in one step (15 minutes) at a time, dividing the dataset into training, validation, and test sets. Specifically, data from the first five months (January 1 to May 31) is used for training, data from the sixth month (June 1 to June 30) is used for validation, and data from the last month (July 1 to July 31) is used for testing.

[0028] (2) Construct a high-frequency short-term wind speed prediction model, which includes an input layer, a shared layer, an independent layer, a fully connected layer and an output layer.

[0029] In this embodiment, CNN-LSTM is used as the underlying architecture to build an ensemble model called "WindNet" for short-term wind speed prediction in complex terrain. The basic structure of the WindNet model is as follows: Figure 1 As shown. In WindNet, this invention uses NWP forecast data for surrounding areas on different isobaric surfaces prior to the target timestamp as input. For example, if the current timestamp is 13:00, and this invention needs to predict wind speed for the next 6 hours (i.e., from 13:00 to 19:00, with a target timestamp of 19:00), then the input data is NWP forecast data every 15 minutes from 13:00 to 19:00 (i.e., an input length of 6 × 4 = 24). These inputs allow this invention to fully utilize existing forecast results and provide future time-series wind speed data that cannot be obtained using only historical time-series wind speed data.

[0030] Unlike simple CNN-LSTM frameworks, this invention designs a multi-task learning (MTL) model to jointly optimize short-term wind speed and direction predictions, taking into account the shared local spatial correlations between wind speed and direction. Specifically, the convolutional layers are designed as shared layers to capture the common spatial correlations between wind speed, wind direction, and NWP prediction data on different isobaric surfaces, resulting in wind speed-related spatial feature maps and wind direction-related spatial feature maps. This shared CNN layer can efficiently extract spatial features related to both data types from the surrounding area and extract the complex nonlinear relationship between station height wind speed and different isobaric surfaces.

[0031] Subsequently, the LSTM was constructed as independent layers. These two independent LSTM layers can capture the temporal dynamics of wind speed and wind direction respectively, thus effectively modeling the temporal dependency of each feature.

[0032] In this embodiment, the optimal model structure of WindNet is determined through grid search. The CNN architecture used in this invention consists of two hidden layers without pooling layers, each with 16 channels and a 3×3 kernel size. The fully connected layer contains two hidden layers, used to obtain integrated wind speed features based on temporal dynamic features of wind speed, and to obtain integrated wind direction features based on temporal dynamic features of wind direction. Each hidden layer of the fully connected layer has 100 neurons.

[0033] The output layer is used to obtain wind speed prediction results based on the integrated wind speed and wind direction characteristics.

[0034] (3) Based on the joint optimization method of multiple prediction tasks, the high-frequency short-term wind speed prediction model is trained using the training set. During the training process, the high-frequency short-term wind speed prediction model is evaluated using the validation set and the model parameters are adjusted. The loss function used during the training process is a weighted loss function of the joint wind speed loss function and the wind direction loss function.

[0035] To further integrate the temporal information of wind speed and direction, this invention employs multi-task joint optimization, combining information from both the wind speed and wind direction loss functions to further improve wind speed prediction capabilities. The model's final output consists of multiple consecutive predicted values ​​up to the target timestamp.

[0036] Specifically, this invention takes wind speed prediction as the primary task and wind direction prediction as the secondary task, i.e., the weighted loss function F = 0.7L + 0.3L1, where L is the wind speed loss function and L1 is the wind direction loss function.

[0037] The wind speed loss function L is: , Where i represents the i-th training sample, n is the number of training samples, f(x) represents the predicted wind speed or wind direction, and y represents the actual wind speed or wind direction.

[0038] During the training of the high-frequency short-term wind speed prediction model, the dropout rate was set to 0.25. The learning rate was adjusted using a polynomial decay strategy based on the Adam optimizer, with an initial learning rate of 0.001 and a weight decay parameter of 0.1. The high-frequency short-term wind speed prediction model was used for 6-hour, 12-hour, and 24-hour prediction tasks. The batch size varied depending on the task: 12 for the 6-hour prediction task, 32 for the 12-hour prediction task, and 64 for the 24-hour prediction task. The entire training of the high-frequency short-term wind speed prediction model was set to 500 cycles, and an early stopping mechanism was introduced to avoid model overfitting.

[0039] In this embodiment, the evaluation metrics used for the prediction task (regression) include mean absolute error (MAE) and root mean square error (RMSE), wherein, yes The actual wind speed at any given moment yes Wind speed forecast for the current moment: Mean Absolute Error (MAE): , Root Mean Square Error (RMSE): , (4) Use the trained high-frequency short-term wind speed prediction model to predict the future wind speed of the target site, and apply inverse normalization to the wind speed prediction results.

[0040] To demonstrate the effectiveness of the WindNet model of this invention, a multi-model cross-comparison was conducted. For comparative analysis, the WindNet model of this invention was compared with NWP, LSTM, and CNN-LSTM models using the same input data. The optimal structure of these models was also determined through grid search. All experiments were conducted using the Keras deep learning framework. The server configuration for model training was as follows: Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz CPU, Quadro RTX 6000 GPU, and Ubuntu 16.04.6 LTS operating system. This invention compared the performance of NWP, LSTM, CNN-LSTM, and WindNet in wind speed prediction tasks for the next 6, 12, and 24 hours (time resolution of 15 minutes). As shown in Table 1, model performance varies across different sites, with site S3 performing significantly better, likely due to its location in a plain area. All models exhibited a gradual decrease in prediction accuracy in tasks with longer prediction step lengths.

[0041] Table 1. Results of multi-model cross-comparison

[0042] Among all models, the LSTM model performed the worst, with a MAE of 1.39 and an RMSE of 1.69 for 6-hour predictions, 1.45 for 12-hour predictions, and 1.79 for RMSE, and 1.60 for 24-hour predictions, with an RMSE of 2.03. In contrast, the simple CNN-LSTM model performed better in short-term wind speed predictions, reducing the MAE by 13.67%, 8.28%, and 8.13% for 6-hour, 12-hour, and 24-hour predictions, respectively, and reducing the RMSE by 8.88%, 8.38%, and 12.32%, respectively. This improvement is mainly attributed to the model's ability to process sequential data and further extract high-dimensional features from the input data. Among all models, "WindNet" performed best, significantly outperforming other benchmark models, with MAE reductions of 20.83%, 15.79%, and 10.20%, and RMSE reductions of 20.13%, 14.02%, and 9.55%, respectively. In particular, the model's ability to capture sudden fluctuations and extreme values ​​in wind speed highlights its superior predictive power. This improvement is primarily attributed to the model's joint optimization strategy, which effectively utilizes the local spatial correlation between wind speed and direction, thereby enhancing prediction accuracy. These results underscore the robustness of the WindNet model in this invention, especially its ability to provide accurate, high spatiotemporal resolution predictions under diverse and complex terrain conditions, making it an important tool for accurate short-term wind speed forecasting.

[0043] In summary, the WindNet model in this embodiment of the invention can predict wind speed for target stations over 6 hours, 12 hours, and 24 hours. WindNet combines the advantages of NWP and deep learning by utilizing numerical weather prediction (NWP) data from different isobaric surfaces in the nearby area as input features. This integration allows WindNet to incorporate real-time meteorological information. The multi-task learning framework introduced in this invention captures the complex interdependence between wind speed and wind direction, jointly optimizing wind speed and wind direction predictions, enhancing the model's generalization ability and reducing its dependence on training samples. Furthermore, this invention performs meticulous variable selection for each station, retaining only the most relevant predictors from multiple isobaric surfaces, ensuring the model focuses on physically meaningful inputs while eliminating noise and redundancy.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting high-frequency short-term wind speed in complex terrain, characterized in that, Includes the following steps: (1) Collect numerical weather forecast data of different isobaric surfaces, then perform the first preprocessing to form model forecast input data, merge the model forecast input data and the data of the observation automatic station equipment, then perform the second preprocessing, and then divide the training set and validation set from them. (2) Construct a high-frequency short-term wind speed prediction model, which includes an input layer, a shared layer, an independent layer, a fully connected layer and an output layer; the shared layer is a CNN layer, which is used to capture the common spatial correlation between wind speed, wind direction and numerical weather forecast data of different isobaric surfaces, and obtain wind speed related spatial feature map and wind direction related spatial feature map. There are two independent layers, both of which are LSTM layers, used to capture the temporal dynamic features of wind speed from the wind speed-related spatial feature map and the temporal dynamic features of wind direction from the wind direction-related spatial feature map, respectively. The fully connected layer contains two hidden layers: one for obtaining the integrated wind speed feature based on the temporal dynamic features of wind speed, and the other for obtaining the integrated wind direction feature based on the temporal dynamic features of wind direction. The output layer is used to obtain wind speed prediction results based on the integrated wind speed and wind direction characteristics; (3) Based on the multi-prediction task joint optimization method, the high-frequency short-term wind speed prediction model is trained using the training set. During the training process, the high-frequency short-term wind speed prediction model is evaluated using the validation set, and the model parameters are adjusted. The loss function used during the training process is a weighted loss function of the joint wind speed loss function and the wind direction loss function. Among them, wind speed prediction is taken as the main task and wind direction prediction is taken as the auxiliary task. The weighted loss function is F=0.7L+0.3L1, where L is the wind speed loss function and L1 is the wind direction loss function. The wind speed loss function L is: Where i represents the training sample number, n is the number of training samples, f(x) represents the wind speed prediction result, and y represents the actual wind speed value; (4) Use the trained high-frequency short-term wind speed prediction model to predict the future wind speed of the target site.

2. The high-frequency short-term wind speed prediction method for complex terrain according to claim 1, characterized in that, In step (1), the numerical weather prediction data includes the average wind speed, zonal wind component, meridional wind component, and average wind speed at the isobaric surfaces of 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, and 100hPa, as well as the average wind speed at a height of 10 meters monitoring station. The numerical weather prediction data are integrated into a unified 15-minute interval dataset according to different time resolutions.

3. The high-frequency short-term wind speed prediction method for complex terrain according to claim 1, characterized in that, In step (1), the method for collecting numerical weather forecast data of different isobaric surfaces includes: inputting the coordinates of the wind speed observation equipment, finding the coordinates near 3KM of the numerical model weather forecast as the grid center, finding the grid within the coordinates and the nearby adjacent grids, and then extracting the numerical weather forecast data of different isobaric surfaces.

4. The high-frequency short-term wind speed prediction method for complex terrain according to claim 1, characterized in that, In step (1), the first preprocessing includes handling missing and outlier values, as well as normalizing the variables; the second preprocessing is to remove model weather forecast data with a correlation coefficient less than the threshold of 0.2 by using the Pearson correlation coefficient.

5. The high-frequency short-term wind speed prediction method for complex terrain according to claim 4, characterized in that, Step (4) also includes applying inverse normalization to the wind speed prediction results.

6. The high-frequency short-term wind speed prediction method for complex terrain according to claim 1, characterized in that, In step (2), the high-frequency short-term wind speed prediction model framework is determined based on grid search.

7. The high-frequency short-term wind speed prediction method for complex terrain according to claim 6, characterized in that, A CNN layer consists of two hidden layers. Each hidden layer in a CNN layer uses 16 channels and has a 3×3 kernel size. Each hidden layer in a fully connected layer has 100 neurons.

8. The high-frequency short-term wind speed prediction method for complex terrain according to claim 7, characterized in that, During the training of the high-frequency short-term wind speed prediction model, the dropout rate was set to 0.

25. A multinomial decay strategy was used to adjust the learning rate based on the Adam optimizer, with an initial learning rate of 0.001 and a weight decay parameter of 0.

1. The high-frequency short-term wind speed prediction model was used for 6-hour, 12-hour, and 24-hour prediction tasks. The batch size for the 6-hour prediction task was 12, for the 12-hour prediction task it was 32, and for the 24-hour prediction task it was 64. The entire training of the high-frequency short-term wind speed prediction model was set to 500 cycles, and an early stopping mechanism was introduced to avoid model overfitting.

9. The high-frequency short-term wind speed prediction method for complex terrain according to claim 1, characterized in that, In step (3), the server configuration used for training the high-frequency short-term wind speed prediction model is: Intel(R) Xeon(R) Gold 5218 CPU @2.30GHz CPU, Quadro RTX 6000 GPU, and Ubuntu 16.04.6 LTS operating system.

10. The high-frequency short-term wind speed prediction method for complex terrain according to claim 1, characterized in that, In step (3), the model evaluation metrics used include mean absolute error and root mean square error.

Images