A complex terrain wind field prediction method based on space-time guidance

Abstract

The application discloses a complex terrain wind field prediction method based on space-time guidance, and steps include: S1, wind field data is acquired, preliminary feature extraction is carried out after preprocessing, and initial feature embedding in a high-dimensional feature space is obtained; S2, an input core feature extraction network is used to capture nonlinear space-time dynamic characteristics of a wind speed field near a mountain, and a deep space-time feature map of the complex terrain wind field is obtained; S3, transformation and integration are carried out, so that the output dimension is aligned with the space-time dimension corresponding to the target prediction step; S4, a freshness function is determined; S5, a space-time loss function is constructed based on the freshness function; and S6, the model is iteratively optimized, and a prediction result of the wind speed near the mountain is output. Through the design of the loss function, meteorological rules and engineering requirements are integrated into model training, and more reliable and more practical space-time prediction of the wind field is realized.

Description

Technical Field

[0001] This invention belongs to the field of wind field prediction technology, and in particular relates to a method for predicting wind fields in complex terrain based on spatiotemporal guidance. Background Technology

[0002] Within the global strategic framework of addressing climate change and achieving net-zero emissions, wind energy has become an indispensable pillar. However, the inherent intermittency and randomness of wind energy are the core bottlenecks for its large-scale grid integration. Unpredictable power fluctuations pose a severe challenge to the real-time supply and demand balance of the power grid, easily leading to system instability. Therefore, high-precision wind energy forecasting is a key enabling technology for ensuring grid security, increasing the proportion of renewable energy consumption, and achieving economic dispatch.

[0003] Traditional wind energy forecasting typically focuses on single-point time-series predictions. However, for critical applications such as wind resource assessment, wind turbine micro-situation, and wind farm operation control, more refined predictions of the spatiotemporal dynamics of the entire wind farm are needed. This challenge is particularly severe in areas with complex terrain. Unlike the relatively uniform flow fields on flat terrain or open seas, complex underlying surfaces such as mountains and hills induce intense nonlinear atmospheric physics phenomena. These phenomena dominate the flow characteristics within the near-surface boundary layer where wind turbines operate, including the airflow acceleration effect at the top of ridges, flow separation on the leeward side of obstacles, the formation of recirculation zones and wakes, and strong wind shear. Accurately capturing these topographically dominated microscale flow structures is crucial; any deviation will lead to significant errors in power generation assessment and potential load risks. Therefore, high-resolution, high-fidelity wind field prediction for complex terrain is a core challenge that urgently needs to be overcome in the field of wind energy science and engineering.

[0004] To address the challenges of wind field forecasting, existing methods are mainly divided into physics-driven and data-driven approaches. Physics-driven models, represented by Numerical Weather Prediction (NWP) and Computational Fluid Dynamics (CFD), are the cornerstone of modern weather forecasting. NWP models predict large-scale weather systems by solving atmospheric control equations, but their spatial resolution is typically on the order of kilometers, making it unable to resolve the terrain-induced micro-scale flows crucial for wind energy applications. Furthermore, they are computationally expensive, and their output exhibits systematic deviations from field measurements, usually requiring downscaling or bias correction. Computational Fluid Dynamics, particularly Large Eddy Simulation (LES), can accurately simulate turbulent fields over complex terrain at scales of meters or even finer detail. However, although LES is considered the "gold standard" for simulation, its enormous computational cost prevents its use for real-time operational forecasting.

[0005] In recent years, data-driven and deep learning methods have developed rapidly, demonstrating excellent performance in spatiotemporal wind field prediction due to their powerful nonlinear mapping and automatic feature extraction capabilities. However, most existing research focuses on network structure design, neglecting the fundamental impact of the loss function on model learning behavior. Currently, the vast majority of deep learning models use mean squared error (MSE) and its variants as the loss function. The fundamental assumption of MSE is that the prediction error weights are the same for all spatiotemporal points, leading to the following key drawbacks:

[0006] 1. Ignoring spatial heterogeneity: In wind energy applications, the prediction error in high wind speed areas (such as ridges) has a much greater impact on power generation assessment than in low wind speed areas, but MSE fails to reflect this difference in importance.

[0007] 2. Ignoring time dynamics: In multi-step forecasting, the closer the data is to the current time, the stronger its indicative power (i.e., wind speed has time-sensitive characteristics), but MSE does not distinguish between errors at different time steps.

[0008] 3. Tendency to oversmooth: MSE tends to penalize large fluctuations, causing the model to generate an oversmoothed prediction field, thereby losing local details and extreme value information that are crucial for engineering applications.

[0009] Therefore, current research is gradually shifting from simply optimizing network architecture to embedding physical cognition and domain knowledge into the learning process. This involves designing loss functions with physical guidance to make the model focus more on the prediction accuracy of key spatiotemporal regions, thereby improving the model's physical consistency and application value. To this end, this invention proposes a spatiotemporally guided method for predicting wind fields in complex terrain, aiming to address the inherent deficiencies in the model training objective (MSE) of existing technologies through a customized loss function. Summary of the Invention

[0010] The purpose of this invention is to provide a method for predicting wind fields in complex terrain based on spatiotemporal guidance. By designing a loss function, meteorological laws and engineering requirements are integrated into the model training, thereby achieving more reliable and practical spatiotemporal prediction of wind fields.

[0011] The technical solution adopted in this invention is a method for predicting wind fields in complex terrain based on spatiotemporal guidance, the steps of which include:

[0012] S1. Obtain wind field data, perform preliminary feature extraction after preprocessing, and obtain the initial feature embedding in the high-dimensional feature space. ;

[0013] S2, will Inputting the core feature extraction network captures the nonlinear spatiotemporal dynamic features of the wind speed field near mountains, resulting in a deep spatiotemporal feature map of the wind field in complex terrain. ;

[0014] S3, will Transform and integrate the data to align its output dimensions with the target prediction step size. Corresponding spatiotemporal dimension alignment;

[0015] S4, Determine the freshness function;

[0016] S5, Constructing a spatiotemporal loss function based on the freshness function;

[0017] S6 iteratively optimizes the model and outputs the predicted wind speed near the mountains.

[0018] Furthermore, the specific steps of S1 are as follows:

[0019] S11, acquire wind field data. Use the large eddy simulation method to continuously collect transient data of wind speed field near the mountain. All transient data together constitute the spatiotemporal sequence data of wind speed near the mountain.

[0020] S12 preprocesses the transient data of wind speed fields near mountains, mapping irregular spatial sampling point data to a regular grid, and formatting the spatiotemporal sequence data of wind speed near mountain terrain into a dimensionless grid. space matrix tensor ,in, Indicates continuous time steps, These represent the number of grids in the spatial longitude and latitude directions, respectively. Indicates the number of feature channels. Represents the set of real numbers;

[0021] S13, the space matrix tensor The initial recurrent units or encoder modules of the input deep learning network are mapped to a high-dimensional feature space through convolution operations and activation processing to obtain the initial feature embedding of the wind field in complex terrain. .

[0022] Furthermore, in S2, the deep spatiotemporal feature map of the wind field in complex terrain... The methods for obtaining it are as follows:

[0023]

[0024] in, This represents the core feature extraction network. Represents network parameters, This represents the initial feature embedding.

[0025] Furthermore, in S3, a deep spatiotemporal feature map of the wind speed field near the mountains is obtained through a dimension transformation operator. The transformation and integration process yields the aligned spatiotemporal feature matrix.

[0026] Furthermore, in S4, the formula for the freshness function is as follows:

[0027]

[0028] in, This indicates the current time at the wind speed measurement point. This indicates the last time of the measured wind speed data. It is a hyperparameter greater than 0.

[0029] Furthermore, the freshness function hyperparameters The impact of different values ​​on the weight is as follows:

[0030] when hour, ;

[0031] when When the weight curve is close to The steepness of the curve can make the loss function extremely sensitive to recent data errors.

[0032] when At that time, the weight curve is relatively flat.

[0033] Furthermore, in S5, the specific steps for constructing the spatiotemporal loss function are as follows:

[0034] S51, Determine the spatial weighting function, as shown in the formula below:

[0035]

[0036] in, These represent the x and y coordinates of the spatial grid plane, respectively. express Spatial weighting function at the location, , These represent the x and y coordinates of the location of the maximum wind speed in the first channel of the wind speed distribution map at the current moment. This represents the spatial guidance sensitivity coefficient. Represents the natural constant;

[0037] S52, Constructing the Spatiotemporal Loss Function The formula is as follows:

[0038]

[0039]

[0040] in, Represents the integrated spatiotemporal weight function. Indicates the total forecast time step. Indicates a time step. This represents the total number of spatial grid points. Indicates the index of the spatial grid. This represents the predicted wind speed near the mountains obtained from the model's forward calculations. Represents the freshness function. express Spatial weighting function at the location, These represent the horizontal and vertical coordinates of the spatial grid plane, respectively.

[0041] Furthermore, the specific steps of S6 are as follows:

[0042] S61, using the spatiotemporal weighted loss function constructed in S5 Instead of the standard mean squared error loss function, the error backpropagation algorithm is used to calculate and iteratively update the weights of the core features extracted from the network using training data in batches.

[0043] S62 inputs the data of the time to be predicted into the trained model, performs forward calculation, and restores it to the physical quantity space through 1×1 convolution to obtain the final high-fidelity, complex terrain multi-step wind speed prediction sequence.

[0044] The beneficial effects of this invention are:

[0045] 1. This invention constructs a customized weighted loss function by integrating a freshness function (time-guided) and a spatial weighting function (spatial-guided) into the loss function, which significantly improves the prediction accuracy of wind fields in complex terrains and solves the technical problem in the prior art where the traditional mean square error (MSE) loss function ignores the spatiotemporal heterogeneity of wind fields and leads to overly smooth prediction results.

[0046] 2. The spatiotemporal guided STG strategy proposed in this invention explicitly embeds the spatiotemporal correlation of wind fields and physical prior knowledge into the model training process, enabling the model to focus on optimizing the prediction accuracy of high-wind-speed areas and recent time steps that are crucial to wind energy engineering. This results in a performance improvement of up to 78% in multi-step prediction of complex terrain.

[0047] 3. This invention can serve as a general optimization framework, which can systematically enhance the performance of various deep learning architectures (ConvLSTM, PredRNN, U-Net) and effectively suppress error accumulation in long-term prediction, demonstrating strong engineering practicality and universality. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are 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.

[0049] Figure 1 This is a flowchart of the present invention;

[0050] Figure 2 These are numerical diagrams of the three functions in this invention, where (a) represents the time-guided strategy. Numerical diagram, (b) is under the spatial guidance strategy Numerical diagram, (c) is under the spatiotemporal guidance strategy Numerical diagram.

[0051] Figure 3 These are comparison diagrams of wind speeds along the centerline in complex terrain, where (a) is a single-step prediction and (b) is a multi-step prediction.

[0052] Figure 4 These are time series predictions of wind speed at the center of the terrain using different models, where (a) is a single-step prediction and (b) is a multi-step prediction.

[0053] Figure 5 These are box plots of single-step prediction errors for different models, where (a) is MAE, (b) is RMSE, (c) is MAPE, and (d) is R². 2 .

[0054] Figure 6 These are box plots of multi-step prediction errors for different models, where (a) is MAE, (b) is RMSE, (c) is MAPE, and (d) is R². 2 .

[0055] Figure 7 These are prediction error cloud maps of different models on complex terrain, where (a) is the prediction error cloud map of the ConvLSTM model, (b) is the prediction error cloud map of the STG-ConvLSTM model, (c) is the prediction error cloud map of the PredRNN model, (d) is the prediction error cloud map of the STG-PredRNN model, (e) is the prediction error cloud map of the Unet model, and (f) is the prediction error cloud map of the STG-Unet model. Detailed Implementation

[0056] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Embodiments of the present invention provide a method for predicting wind fields in complex terrain based on spatiotemporal guidance, the flowchart of which is shown below. Figure 1 As shown, the steps include:

[0058] Step S1: Acquire wind field data and preprocess it. Then, perform preliminary feature extraction on the preprocessed wind field data to obtain the initial feature embedding in the high-dimensional feature space. The specific steps are as follows:

[0059] S11. Acquire wind field data. Continuously collect transient data of the wind speed field near the mountains using the Large Eddy Simulation (LES) method. This transient data covers the spatial distribution of wind speed near the mountains at different times, completely recording the changes in wind speed magnitude and direction over time and space. All transient data together constitute the spatiotemporal sequence data of wind speed near the mountains. LES is a numerical simulation technique based on the filtered Navier-Stokes equations, capable of accurately resolving large-scale eddy structures in the flow field, and is suitable for wind field simulation in complex terrain.

[0060] S12 preprocesses the transient data of wind speed fields near mountains. The irregular spatial sampling point data obtained from large eddy simulation is mapped to a regular grid using bilinear interpolation or nearest-neighbor interpolation. The spatiotemporal sequence data of wind speed near mountainous terrain is then formatted into a grid with dimension [missing information]. space matrix tensor .in, Indicates continuous time steps, These represent the number of grids in the longitude and latitude directions (set to 15×15 in this embodiment), corresponding to meter-level resolution in actual physical space; This indicates the number of characteristic channels, representing wind speed vector components (such as east-west wind speed, north-south wind speed, and vertical wind speed). This represents the set of real numbers. The purpose of formatting is to transform the physical information of the flow field in complex terrain into a tensor format that meets the input requirements of existing deep learning network architectures (ConvLSTM / PredRNN / U-Net).

[0061] S13, the spatial matrix tensor generated in S12 is processed through the initial recurrent unit or encoder module of a deep learning network (ConvLSTM, PredRNN, or U-Net). Mapping to a high-dimensional feature space yields the initial feature embedding of the wind field in complex terrain. The formula is as follows:

[0062]

[0063] in, This represents the weight matrix of the convolutional embedding kernel. This represents the convolution operation. Indicates the bias term. This represents the activation function (such as ReLU). This represents the initial feature embedding. These represent the number of grids in the spatial longitude and latitude directions, respectively. Indicates the embedding dimension. Indicates the number of feature channels. It represents the set of real numbers.

[0064] Step S2, embed the initial features The input core feature extraction network (composed of multiple recurrent units or encoder-decoder structures, such as recurrent layers of ConvLSTM or PredRNN, encoder-decoder structures of U-Net, and other existing network structures) captures the nonlinear spatiotemporal dynamic features of wind speed fields near mountains, resulting in a deep spatiotemporal feature map of wind fields in complex terrain. The formula is as follows:

[0065]

[0066] in, This represents the core feature extraction network. Represents network parameters, This represents the initial feature embedding.

[0067] Step S3: Create a deep spatiotemporal feature map of the wind speed field near the mountains. Transform and integrate the data to align its output dimensions with the target prediction step size. Corresponding spatiotemporal alignment is ensured to guarantee that the model output can cover the future. The formula for the overall wind speed distribution at each time step is as follows:

[0068]

[0069] in, Represents the dimensional transformation operator. This represents the aligned spatiotemporal feature matrix. Represents the set of real numbers. These represent the number of grids in the spatial longitude and latitude directions, respectively.

[0070] Step S4: Determine the freshness function. Wind speed, as a time-varying process, has informational value that evolves over time. The closer the observed data is to the current prediction time, the stronger its indicative power for future state evolution; that is, wind speed has time sensitivity. To quantify the importance that should be given during data mining, this invention introduces a freshness function. Based on existing technology (Chinese Patent Publication No. CN120337715A), a learnable time sensitivity parameter is further proposed. The purpose is to enable the weight decay rate of the freshness function to adaptively match the different requirements of historical data timeliness for different types of models (long / short time memory) and different prediction tasks (single-step / multi-step). The formula for the freshness function is as follows:

[0071]

[0072] in, This indicates the current time at the wind speed measurement point. This indicates the last time of the measured wind speed data. It is a hyperparameter greater than 0;

[0073] when hour, This function is close to the form in the prior art (CN120337715A);

[0074] when When the weight curve is close to The steepness of the curve becomes extremely high, which makes the loss function less sensitive to recent events (…). It is extremely sensitive to data errors and is suitable for models with short time memory, such as U-Net;

[0075] when When the curve is relatively flat, it is suitable for models such as LSTM that require the use of longer historical information.

[0076] Step S5: Construct a spatiotemporal loss function based on the freshness function. The specific steps are as follows:

[0077] S51, determine the spatial weighting function, using the Gaussian distribution function as the weighting function for the spatial guidance strategy (SG), and based on the wind speed values ​​near the mountains. (Based on measured data collected in step S1), identify the coordinates of the maximum wind speed in the dominant characteristic channel of the wind speed field at the current moment. By assigning different weights using the Gaussian distribution function, a penalty is applied to the error in high-wind-speed areas. The formula for calculating the spatial weighting function is as follows:

[0078]

[0079] in, These represent the x and y coordinates of the spatial grid plane, respectively. express Spatial weighting function at the location, , These represent the x and y coordinates of the location of the maximum wind speed in the first channel (dominant characteristic channel) of the wind speed distribution map at the current moment. This represents the spatial guidance sensitivity coefficient, used to control the range of radiation from the high-weight region to the area around the peak point. Its value range is (0, 5], with a preferred value of 1. Represents the natural constant.

[0080] S52, Constructing the Spatiotemporal Loss Function By combining the freshness function under the time-guided (TG) strategy with the spatial weighting function under the spatial-guided (SG) strategy, a comprehensive spatiotemporal weighting function is formed. The formula is as follows:

[0081]

[0082]

[0083] in, This represents the integrated spatiotemporal weighting function, used to characterize the contribution of different spatiotemporal points to the model training objective; Indicates the total forecast time step. Indicates a time step. This represents the total number of spatial grid points. Indicates the index of the spatial grid; This represents the predicted wind speed near the mountain obtained from the forward calculation of the model, and the physical quantity result generated by mapping the aligned spatiotemporal feature matrix obtained from S3 through the 1×1 convolutional layer of S6. This represents the true wind speed near the mountains, used to compare with the true value during the training phase to calculate the loss value. .

[0084] , represents the weighting parameters guided by time and space, respectively, used to balance the contributions of the two dimensions. Both parameters range from [0,1] and satisfy the following condition: , , The preferred value is 0.5; Represents the freshness function. express Spatial weighting function at the location, These represent the horizontal and vertical coordinates of the spatial grid plane, respectively.

[0085] Step S6: Iteratively optimize the model and output the predicted wind speed near the mountains. The specific steps are as follows:

[0086] S61, using the spatiotemporal weighted loss function constructed in S5 Instead of the standard mean squared error (MSE) loss function, this algorithm uses backpropagation to calculate and iteratively update the weights within the core feature extraction network (such as ConvLSTM, PredRNN, or U-Net) using batches of training data. Under its guidance, the model training process will adaptively increase the learning intensity for regions with high timeliness (recent) and high energy density (high wind speed).

[0087] S62. After completing the model training, the data at the time to be predicted (the feature map processed in S3) is input into the model for forward calculation. After 1×1 convolution, it is restored to the physical quantity space to obtain the final high-fidelity, complex terrain multi-step wind speed prediction sequence. This sequence can more accurately reflect the micro-scale flow characteristics induced by the terrain and can directly serve the real-time power prediction of wind farms and grid dispatch decision-making.

[0088] Experimental verification

[0089] Table 1 shows the evaluation metrics of different guidance strategies and the different optimization functions used. The results indicate that different guidance strategies all achieve good optimization effects for the original three neural network methods. In the time-guided strategy, compared with existing technologies, this invention introduces a learnable time sensitivity parameter, resulting in better fitting performance, a significantly reduced root mean square error (RMSE), and a higher coefficient of determination (R²). 2 The value is closer to 1. In the spatial guidance strategy, comparison with other distribution functions shows that the Gaussian distribution function used in this invention has the best fitting effect. Furthermore, compared with the individual temporal and spatial guidance strategies, the spatiotemporal guidance strategy significantly improves the fitting effect, achieving the lowest RMSE and the highest R². 2 .

[0090] Table 1. Comparison of evaluation metrics for different guiding strategies and the different optimization functions used.

[0091]

[0092] Comparison of wind speeds along the midline in complex terrain using different models is shown in the figure below. Figure 3 As shown, (a) is a single-step prediction, and (b) is a multi-step prediction. The time-series wind speed predictions at the terrain center point for different models are shown in the figure. Figure 4As shown, (a) represents single-step prediction, and (b) represents multi-step prediction. The results show that while the three baseline models (ConvLSTM, PredRNN, Unet) can roughly capture the trend, they exhibit significant numerical biases, especially in multi-step prediction. Conversely, the prediction accuracy of all models is significantly improved after applying the STG (Spatiotemporal Guided) strategy.

[0093] Numerical diagrams of the three functions in this invention are shown below. Figure 2 As shown, (a) is under the time-guided strategy. Numerical diagram, (b) is under the spatial guidance strategy Numerical diagram, (c) is under the spatiotemporal guidance strategy The numerical diagram shows that the prediction effect is best under the guidance of the spatiotemporal strategy. This is because the spatiotemporal guidance function constructed in this invention adopts an additive coupling mechanism, where the spatial weighting function... The purpose of establishing a system based on fixed, complex terrain features is to provide a stable spatial heterogeneity baseline for the loss function, while the freshness function... The evolution tends to level off in the later stages of the prediction, and the numerical increment is insufficient to change the weight distribution pattern dominated by topography. Figure 2 The values ​​in (c) show no significant change.

[0094] Tables 2 and 3 show the error metrics and their relative improvement rates for single-step and multi-step prediction, respectively. In the tables, U represents the specific value of each error evaluation metric, P represents the performance improvement rate, ConvLSTM is a convolutional long short-term memory network, Predrnn is a predictive recurrent neural network, and Unet is a U-shaped encoder-decoder network. The results show that, in both single-step and multi-step prediction tasks, the optimized model significantly reduces all error metrics (MAE, RMSE, MAPE), and the coefficient of determination (R²) is closer to 1.

[0095] Table 2 Error evaluation index and performance improvement rate P of wind speed U in single-step wind field prediction in complex terrain

[0096]

[0097] Table 3 Error evaluation index and performance improvement rate P of wind speed U in multi-step wind field prediction in complex terrain

[0098]

[0099] Table 2 shows that the relative performance improvement brought by the STG strategy ranges from 5.7% to 64.15% in single-step prediction, while Table 3 shows that this advantage becomes even more pronounced in the more challenging multi-step prediction, with improvement rates ranging from 28.6% to 78.87%. Among them, STG-ConvLSTM has the highest relative improvement rate at 78.87%, R 2 The mean absolute error (MAE) can reach 0.994, indicating that the STG strategy has the best optimization effect on ConvLSTM. This strongly proves that explicitly fusing terrain spatial information and temporal evolution patterns is an efficient optimization path. Specifically, Table 2 shows that in the single-step prediction task, the MAE (0.0562) and RMSE (0.0809) of STG-PredRNN are the lowest among all models. The ConvLSTM series models show the most significant performance leap under the STG strategy, with their mean absolute error (MAE) increasing by 39.59% and 64.98% for single-step and multi-step predictions, respectively. In contrast, the Unet series models benefit relatively little from STG. Furthermore, by comparing the single-step and multi-step prediction results, it can be found that the prediction error of all models accumulates with the prediction step length, but the STG strategy is crucial for suppressing the accumulation of errors in long-term predictions. Taking STG-ConvLSTM as an example, its MAE improvement rate in multi-step prediction (64.98%) is much higher than that in single-step prediction (39.59%), indicating that explicit modeling of spatiotemporal dependence is the key to improving the long-term prediction stability of the model.

[0100] Box plots of single-step prediction errors for different models are shown below. Figure 5 As shown, (a) is MAE, (b) is RMSE, (c) is MAPE, and (d) is R. 2 Box plots of multi-step prediction errors for different models are shown below. Figure 6 As shown, (a) is MAE, (b) is RMSE, (c) is MAPE, and (d) is R. 2 .from Figure 5 and Figure 6 The error box plots show that the optimized model significantly improves across all metrics, reducing prediction error and narrowing the error fluctuation range. Compared to ConvLSTM, PredRNN, and Unet, STG-ConvLSTM, STG-PredRNN, and STG-Unet have higher R-values. 2 The values ​​are closer to 1, and MAE, RMSE, and MAPE (%) are closer to 0, with a reduced range of fluctuation for these parameters, indicating that this method has varying degrees of optimization effect on all three machine learning methods. Specifically, STG-PredRNN's R... 2 The largest value, and the smallest MAE, RMSE, and mean absolute percentage error (MAPE), indicate that STG-PredRNN has the best fitting effect. Figure 5 and Figure 6The comparison shows that the error of single-step prediction is much smaller than that of multi-step prediction. The improved model has significant optimization in both single-step and multi-step prediction, especially in multi-step prediction.

[0101] Prediction error contour maps of different models on complex terrain wind fields are shown below. Figure 7 As shown, (a) is the prediction error cloud map of the ConvLSTM model, (b) is the prediction error cloud map of the STG-ConvLSTM model, (c) is the prediction error cloud map of the PredRNN model, (d) is the prediction error cloud map of the STG-PredRNN model, (e) is the prediction error cloud map of the Unet model, and (f) is the prediction error cloud map of the STG-Unet model. Figure 7 The study revealed the non-uniformity of the error distribution, with high error areas mainly concentrated on the windward, top, and leeward sides of hills. Comparative analysis showed that the improved model effectively reduced flow field prediction errors near mountains for all machine learning methods (ConvLSTM, PredRNN, Unet), demonstrating significant results. This distribution pattern highly aligns with fluid physics principles (such as flow separation, wake recirculation, and reattachment): in these regions, large wind speed gradients and high turbulence intensity increase prediction difficulty. Therefore, the spatial distribution characteristics of the errors not only validate the physical rationality of spatiotemporal correlation prediction but also indicate that the complexity of local terrain is a key factor affecting model prediction performance.

[0102] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for predicting wind fields in complex terrain based on spatiotemporal guidance, characterized by the following steps: include: S1. Obtain wind field data, perform preliminary feature extraction after preprocessing, and obtain the initial feature embedding in the high-dimensional feature space. ; S2, will Inputting the core feature extraction network captures the nonlinear spatiotemporal dynamic features of the wind speed field near mountains, resulting in a deep spatiotemporal feature map of the wind field in complex terrain. ; S3, will Transform and integrate the data to align its output dimensions with the target prediction step size. Corresponding spatiotemporal dimension alignment; S4, Determine the freshness function; S5, Constructing a spatiotemporal loss function based on the freshness function; S6 iteratively optimizes the model and outputs the predicted wind speed near the mountains.

2. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, acquire wind field data. Use the large eddy simulation method to continuously collect transient data of wind speed field near the mountain. All transient data together constitute the spatiotemporal sequence data of wind speed near the mountain. S12 preprocesses the transient data of wind speed fields near mountains, mapping irregular spatial sampling point data to a regular grid, and formatting the spatiotemporal sequence data of wind speed near mountain terrain into a dimensionless grid. space matrix tensor ,in, Indicates continuous time steps, , These represent the number of grids in the spatial longitude and latitude directions, respectively. Indicates the number of feature channels. Represents the set of real numbers; S13, the space matrix tensor The initial recurrent units or encoder modules of the input deep learning network are mapped to a high-dimensional feature space through convolution operations and activation processing to obtain the initial feature embedding of the wind field in complex terrain. .

3. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 1, characterized in that, In S2, the deep spatiotemporal characteristic map of the wind field in complex terrain. The methods for obtaining it are as follows: ； in, This represents the core feature extraction network. Represents network parameters, This represents the initial feature embedding.

4. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 1, characterized in that, S3, through a dimension transformation operator, generates a deep spatiotemporal feature map of the wind speed field near the mountains. The transformation and integration process yields the aligned spatiotemporal feature matrix.

5. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 1, characterized in that, In S4, the freshness function The formula is as follows: ； in, This indicates the current time at the wind speed measurement point. This indicates the last time of the measured wind speed data. It is a hyperparameter greater than 0. Represents the natural constant.

6. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 5, characterized in that, The freshness function hyperparameters The impact of different values ​​on the weight is as follows: when hour, ; when When the weight curve is close to The steepness of the curve can make the loss function extremely sensitive to recent data errors. when At that time, the weight curve is relatively flat.

7. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 1, characterized in that, In S5, the specific steps for constructing the spatiotemporal loss function are as follows: S51, Determine the spatial weighting function, as shown in the formula below: ； in, These represent the x and y coordinates of the spatial grid plane, respectively. express Spatial weighting function at the location, These represent the x and y coordinates of the location of the maximum wind speed in the first channel of the wind speed distribution map at the current moment. This represents the spatial guidance sensitivity coefficient. Represents the natural constant; S52, Constructing the Spatiotemporal Loss Function The formula is as follows: ； ； in, Represents the integrated spatiotemporal weight function. Indicates the total forecast time step. Indicates a time step. This represents the total number of spatial grid points. Indicates the index of the spatial grid. This represents the predicted wind speed near the mountains obtained from the model's forward calculations. Represents the freshness function. express Spatial weighting function at the location, These represent the x and y coordinates of the spatial grid plane, respectively. This represents the actual wind speed near the mountains.

8. The method for predicting wind fields in complex terrain based on spatiotemporal guidance according to claim 1, characterized in that, The specific steps of S6 are as follows: S61, using the spatiotemporal weighted loss function constructed in S5 Instead of the standard mean squared error loss function, the error backpropagation algorithm is used to calculate and iteratively update the weights of the core features extracted from the network using training data in batches. S62 inputs the data of the time to be predicted into the trained model, performs forward calculation, and restores it to the physical quantity space through 1×1 convolution to obtain the final high-fidelity, complex terrain multi-step wind speed prediction sequence.

Images