Multi-point wind speed prediction model and method that simultaneously processes spatiotemporal wind speed information

Abstract

This invention discloses a multi-point wind speed prediction model that simultaneously processes spatiotemporal wind speed information. It utilizes sequentially connected modules A, B, C, and D, and leverages the feature channels F of a 1×1 convolutional 3D CNN layer I in module A. I The dimensional expansion only increases the dimension of the feature channels without extracting information from the spatial and temporal dimensions, thus preserving as much of the original input data's temporal and spatial information as possible. Therefore, the two ResNet-ConvLSTM modules in module B can more completely capture the nonlinear spatiotemporal information inherent in the data simultaneously. In module C, the first permute / reshape layer swaps the temporal and feature channel dimensions, and two 1DCNN layers are used to increase and decrease the temporal dimension, ensuring the time step matches the prediction step size K. The second permute / reshape layer swaps the data dimensions back to the original data. Finally, the 3DCNN layer II in module D restores the number of feature channels to the number F of output feature channels. O Thus, the prediction result is finally obtained.

Description

Technical Field

[0001] This invention belongs to the field of wind speed prediction technology, specifically a multi-point wind speed prediction model and method that simultaneously processes spatiotemporal wind speed information. Background Technology

[0002] Accurate short-term wind speed forecasting (minutes to hours) is crucial for improving wind power utilization and ensuring stable power supply. As onshore wind farms in China increasingly gravitate towards mountainous areas, the terrain and other factors create wake effects and complex flow phenomena such as flow separation and reattachment, and shear flow. Furthermore, wind speed itself exhibits significant randomness and volatility. This complex wind field variation poses a substantial challenge to the safe and stable operation of wind farms. With large-scale integration of wind power into the power grid, this unstable power supply can impact the grid, reduce the economic benefits of wind power, and even lead to curtailment. Therefore, developing accurate methods for forecasting short-term wind speeds is of paramount importance.

[0003] Because the governing equations describing fluid motion are continuous in both time and space, this indicates that fluids exhibit not only temporal correlation but also spatial correlation. Wind, as a fluid, shares these same physical properties. Traditional single-point short-term wind speed prediction methods often rely on wind speed signals from a single spatial location, neglecting the correlation between wind speeds at different measurement points, which to some extent limits the improvement of short-term wind speed prediction accuracy. With the continuous development and upgrading of wind speed measurement equipment and the expansion of wind farm scale, wind speed measurement is no longer limited to a single location; real-time detection of three-dimensional wind fields can now be achieved. Therefore, multi-point synchronous wind speed measurement data has been effectively collected and expanded. Considering that spatial wind speed distribution is influenced by multiple variables such as topography, wind direction, and temperature, wind speed information at different locations exhibits certain correlations in both time and space. Therefore, exploring the spatiotemporal relationships between wind speed information from multiple points is an effective way to improve wind speed prediction accuracy; such methods can be called multi-point spatiotemporal synchronous wind speed prediction methods.

[0004] Currently, common multi-point wind speed synchronous prediction methods often employ deep neural networks to extract the spatiotemporal correlation between wind speed time histories at different spatial locations. The basic idea is to treat each measurement point at a different spatial location as an image, arranging multiple consecutive images to form a dataset with spatiotemporal information. By first extracting spatial information from this spatiotemporal dataset and then performing temporal modeling, multi-point spatiotemporal synchronous prediction can be achieved. This modeling strategy that separates space and time is also known as "two-stage" modeling. While the "two-stage" modeling strategy utilizes the spatial correlation inherent between measurement points, significantly improving the accuracy of short-term wind speed prediction, some problems remain. For example, in the process of converting spatial information into temporal information, the spatial information needs to be flattened into a one-dimensional vector and input into the temporal model. This operation disrupts the relative spatial positions between measurement points, weakening the spatial correlation. Summary of the Invention

[0005] In view of this, in order to solve the problem that the existing "two-stage" modeling strategy destroys the spatial relative positional relationship between measuring points and weakens the spatial correlation, the purpose of this invention is to provide a multi-point wind speed prediction model and method that simultaneously processes wind speed spatiotemporal information.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention first proposes a multi-point wind speed prediction model that simultaneously processes spatiotemporal wind speed information, comprising the following sequentially connected components:

[0008] Module A: A 3D CNN layer I consisting of an input layer and a 1×1 convolution, wherein the input layer is used to input data in the form (J, X, Y, F). I The four-dimensional data; the convolution kernel of the 3DCNN layer I is (1, 1, 1) and is used to implement the feature channels F. I An extension of; where J represents the time dimension; X and Y represent the dimensions in two directions in the spatial dimension; F I Indicates the characteristic channel;

[0009] Module B: consists of two ResNet-ConvLSTM modules connected in sequence. Each ResNet-ConvLSTM module includes two ConvLSTM layers and a residual connection block, used to capture the nonlinear spatiotemporal information contained in the data center.

[0010] Module C: Includes two permute / reshape layers, with two 1DCNN layers between them; the permute / reshape layer connected to Module B is used to swap the time dimension and feature channel dimension in the output data of Module B, and transform the data into (X×Y×F) f,J), where,F f This indicates the number of convolutional kernels used in the previous neural network layer; two 1DCNN layers are used to increase and decrease the dimensionality of the time dimension, respectively, so that the time step matches the number of prediction steps K; K is any positive integer; another permute / reshape layer is used to restore the data to its original order;

[0011] Module D: Includes a 3DCNN layer II with 1×1 convolutions and an output layer. The number of convolutional kernels in the 3DCNN layer II is equal to the number F of the required output feature channels. O The final spatiotemporal wind speed prediction result is then output through the output layer.

[0012] Furthermore, the principle of the ConvLSTM layer is as follows:

[0013] f t =σ(W xf *x t +W hf *h t-1 +b f )

[0014] g t =tanh(W xg *x t +W hg *h t-1 +b g )

[0015] i t =σ(W xi *x t +W hi *h t-1 +b i )

[0016] c t =f t ⊙c t-1 +i t ⊙g t

[0017] o t =σ(W xo *x t +W ho *h t-1 +b o )

[0018] y t =h t =o t ⊙tanh(c t )

[0019] Where t and t-1 are the current and previous times, respectively; f tg t i t and o t These represent the controllers for the forget gate, the intermediate output of the LSTM network, the input gate, and the output gate, respectively; where x t Let c be the input vector. t For the long-term state, h t For short-term states, y t W is the output vector. xf W xg W xi W xo To be respectively with x t The connection weight matrix, W hf W hg W hi W ho To be respectively with h t -1 connects the weight matrix, b f ,b g ,b i ,bo is the bias term; σ(·) represents the activation function, * represents the convolution operation of the convolutional layer, and ⊙ represents the sequential multiplication of array elements.

[0020] Furthermore, the principle of the residual connection block is as follows:

[0021] h(x) = f(x) - x

[0022] Where f(x) represents the output of the residual connection block; x represents the input of the residual connection block; and h(x) represents the residual.

[0023] This invention also proposes a multi-point wind speed prediction method that simultaneously processes spatiotemporal wind speed information, comprising the following steps:

[0024] Step 1: Obtain the original spatiotemporal wind speed sequence and preprocess it to obtain four-dimensional wind speed data (Time steps, Row, Column, Features). Here, Time steps represents the number of image information at various time points in the four-dimensional wind speed data, representing the time dimension; Row and Column represent the height and width of the image, respectively, representing the spatial dimension; and Features represent the number of wind speed components, representing the feature channels.

[0025] Step 2: Construct and train the multi-point wind speed prediction model as described in claim 1;

[0026] Step 3: Input the four-dimensional wind speed data into the multi-point wind speed prediction model to obtain the spatiotemporal wind speed prediction results.

[0027] Furthermore, in step one, the original spatiotemporal wind speed sequence is normalized using the following method:

[0028] The wind speed vector at any time and location in the original spatiotemporal wind speed sequence is decomposed into three wind speed components: u, v, and w. Then, the three wind speed components u, v, and w are normalized respectively.

[0029] u / U ref ,v / U ref w / U ref

[0030] Among them, U ref This indicates the wind speed of the incoming airflow in the far field.

[0031] Furthermore, in step two, the method for training the multi-point wind speed prediction model is as follows:

[0032] 21) The preprocessed four-dimensional wind speed data U CH The training set, validation set, and test set are divided in a 6:2:2 ratio.

[0033] 22) Input the training set data into the multi-point wind speed prediction model and obtain the spatiotemporal wind speed prediction value, and calculate the loss function of the spatiotemporal wind speed prediction value compared with the actual spatiotemporal wind speed value.

[0034] 23) Determine if the loss function is less than the set threshold: if yes, the multi-point wind speed prediction model training is complete, and proceed to step 26); if no, increment the iteration count by 1, and proceed to step 24.

[0035] 24) Determine if the number of iterations has reached the set maximum number of iterations: if yes, the multi-point wind speed prediction model training is complete, and proceed to step 26); if no, proceed to step 25.

[0036] 25) Update the hyperparameters in the multi-point wind speed prediction model, and execute step 22);

[0037] 26) The prediction performance of the multi-point wind speed prediction model was verified and tested using the verification machine and the test set, respectively.

[0038] Furthermore, in step 23), the loss function is MAPE, MAE, or RMSE.

[0039] Furthermore, in step 25), the hyperparameters in the multi-point wind speed prediction model are updated using Adam as the optimizer.

[0040] The beneficial effects of this invention are as follows:

[0041] This invention presents a multi-point wind speed prediction model that simultaneously processes spatiotemporal wind speed information. It utilizes sequentially connected modules A, B, C, and D, and leverages the feature channels F of a 1×1 convolutional 3D CNN layer I in module A. IThe dimensional expansion only increases the dimension of the feature channels without extracting information from the spatial and temporal dimensions, thus preserving as much of the original input data's temporal and spatial information as possible. Therefore, the two ResNet-ConvLSTM modules in module B can more completely capture the nonlinear spatiotemporal information inherent in the data simultaneously. In module C, the first permute / reshape layer swaps the temporal and feature channel dimensions, and two 1DCNN layers are used to increase and decrease the temporal dimension, ensuring the time step matches the prediction step number K. The second permute / reshape layer swaps the data dimensions back to the original data. Finally, the 3DCNN layer II in module D restores the number of feature channels to the number F of output feature channels. O This leads to the final prediction result, which has the following advantages:

[0042] 1) It overcomes the shortcomings of the existing multi-point wind speed synchronous prediction method, which uses a "two-stage" modeling strategy that destroys the spatial relative positional relationship between measuring points and weakens spatial correlation.

[0043] 2) It overcomes the shortcomings of traditional multi-point wind speed synchronous prediction methods, which have a large resolution of the spatial wind speed matrix or require the simultaneous prediction of multiple wind speed components (i.e., predicting three-dimensional wind speed vectors), resulting in too many parameters and difficulty in training.

[0044] 3) By combining the advantages of ConvLSTM, Residual Network (ResNet) and 1×1 3D convolution, a prediction model that can simultaneously process wind speed spatiotemporal information can be developed without weakening or destroying the spatial relativity between measurement points. It can also use four-dimensional spatiotemporal wind speed data to carry out multi-point wind speed and direction synchronous prediction. Attached Figure Description

[0045] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0046] Figure 1 This is a framework diagram of an embodiment of the multi-point wind speed prediction model that simultaneously processes spatiotemporal wind speed information according to the present invention.

[0047] Figure 2 This is a schematic diagram of a 3DCNN.

[0048] Figure 3 This is a schematic diagram of the residual connection block;

[0049] Figure 4 This is a schematic diagram illustrating the principle of 1×1 convolution;

[0050] Figure 5 This is a diagram of the overall network framework of the ST3DCNN model;

[0051] Figure 6This is the overall network framework diagram of the STCL model. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0053] This embodiment describes a multi-point wind speed prediction method that simultaneously processes spatiotemporal wind speed information, including the following steps:

[0054] Step 1: Obtain the original spatiotemporal wind speed sequence and preprocess it to obtain four-dimensional wind speed data (Time steps, Row, Column, Features). Here, Time steps represents the number of image information at various time points in the four-dimensional wind speed data, representing the time dimension; Row and Column represent the height and width of the image, respectively, representing the spatial dimension; and Features represent the number of wind speed components, representing the feature channels.

[0055] Specifically, in this embodiment, the original spatiotemporal wind speed sequence is normalized using the following method:

[0056] The wind speed vector at any time and location in the original spatiotemporal wind speed sequence is decomposed into three wind speed components: u, v, and w. Then, the three wind speed components u, v, and w are normalized respectively.

[0057] u / U ref ,v / U ref w / U ref

[0058] Among them, U ref This indicates the wind speed of the incoming airflow in the far field.

[0059] Step 2: Build and train a multi-point wind speed prediction model.

[0060] like Figure 1 As shown, the multi-point wind speed prediction model that simultaneously processes spatiotemporal wind speed information in this embodiment includes modules A, B, C, and D connected in sequence.

[0061] Module A: A 3D CNN layer I consisting of an input layer and a 1×1 convolution, wherein the input layer is used to input data in the form (J, X, Y, F). I The four-dimensional data; the convolution kernel of the 3DCNN layer I is (1, 1, 1) and is used to implement the feature channels F. I The extension of; where J represents the time dimension, corresponding to the Time steps in four-dimensional wind speed data; X and Y represent the dimensions in two directions in the spatial dimension, corresponding to the Row and Column in four-dimensional wind speed data; FI This represents the feature channels, corresponding to the Features in four-dimensional wind speed data. For example... Figure 2 As shown, the kernel size reveals that the 3DCNN layer I using 1×1 convolutions does not extract information from the spatial and temporal dimensions; it only increases the dimension of the feature channels. This operation preserves as much of the temporal and spatial information of the original input data as possible for the RC module to capture and extract.

[0062] Module B: consists of two ResNet-ConvLSTM modules connected in sequence. Each ResNet-ConvLSTM module includes two ConvLSTM layers and a residual connection block, used to capture the nonlinear spatiotemporal information contained in the data center.

[0063] Specifically, the principle of the ConvLSTM layer is as follows:

[0064] f t =σ(W xf *x t +W hf *h t-1 +b f )

[0065] g t =tanh(W xg *x t +W hg *h t-1 +b g )

[0066] i t =σ(W xi *x t +W hi *h t-1 +b i )

[0067] c t =f t ⊙c t-1 +i t ⊙g t

[0068] o t =σ(W xo *x t +W ho *h t-1 +b o )

[0069] y t =h t =o t ⊙tanh(c t )

[0070] Where t and t-1 are the current and previous times, respectively; f t g t i t and o t These represent the controllers for the forget gate, the intermediate output of the LSTM network, the input gate, and the output gate, respectively; where x t Let c be the input vector. t For the long-term state, h t For short-term states, y t W is the output vector. xf W xg W xi W xo To be respectively with x t The connection weight matrix, W hf W hg W hi W ho To be respectively with h t-1 The connection weight matrix, b f ,b g ,b i ,b o These are the bias terms for the four layers respectively; σ(·) represents the activation function, * represents the convolution operation of the convolutional layer, and ⊙ represents the sequential multiplication of array elements.

[0071] like Figure 3 As shown, the principle of the residual connection block is as follows:

[0072] h(x) = f(x) - x

[0073] Where f(x) represents the output of the residual connection block; x represents the input of the residual connection block; and h(x) represents the residual.

[0074] Module C: Includes two permute / reshape layers, with two 1DCNN layers between them; the permute / reshape layer connected to Module B is used to swap the time dimension and feature channel dimension in the output data of Module B, and transform the data into (X×Y×F) f ,J), where F f This indicates the number of convolutional kernels used in the previous neural network layer; two 1DCNN layers are used to increase and decrease the time dimension in turn, so that the time step matches the prediction step number K; K is any positive integer; another permute / reshape layer is used to restore the data to its original order.

[0075] Module D: Includes a 3DCNN layer II with 1×1 convolutions and an output layer. The number of convolutional kernels in the 3DCNN layer II is equal to the number F of the required output feature channels. OThe final spatiotemporal wind speed prediction result is then output through the output layer.

[0076] In this embodiment, the method for training the multi-point wind speed prediction model is as follows:

[0077] 21) The preprocessed four-dimensional wind speed data U CH The training set, validation set, and test set are divided in a 6:2:2 ratio.

[0078] 22) Input the training set data into the multi-point wind speed prediction model and obtain the spatiotemporal wind speed prediction value, and calculate the loss function of the spatiotemporal wind speed prediction value compared with the actual spatiotemporal wind speed value.

[0079] 23) Determine if the loss function is less than the set threshold: if yes, the multi-point wind speed prediction model training is complete, and proceed to step 26); if no, increment the iteration count by 1, and proceed to step 24.

[0080] Specifically, the loss function used is MAPE, MAE, or RMSE:

[0081]

[0082]

[0083]

[0084] Where: x(i,j) t For the true value, Represents the predicted value; N here represents the length of the spatiotemporal sequence used for error calculation; U SH It has three components: u, v, and w; x can be any of u, v, or w, and error calculations need to be performed separately for each wind speed component.

[0085] 24) Determine if the number of iterations has reached the set maximum number of iterations: if yes, the multi-point wind speed prediction model training is complete, and proceed to step 26); if no, proceed to step 25.

[0086] 25) Using Adam as the optimizer, update the hyperparameters in the multi-point wind speed prediction model and execute step 22);

[0087] 26) The prediction performance of the multi-point wind speed prediction model was verified and tested using the verification machine and the test set, respectively.

[0088] Step 3: Input the four-dimensional wind speed data into the multi-point wind speed prediction model to obtain the spatiotemporal wind speed prediction results.

[0089] This embodiment describes a multi-point wind speed prediction method that simultaneously processes spatiotemporal wind speed information. It utilizes sequentially connected modules A, B, C, and D, and leverages the feature channels F of the 1×1 convolutional 3D CNN layer I in module A.I The dimensional expansion only increases the dimension of the feature channels without extracting information from the spatial and temporal dimensions, thus preserving as much of the original input data's temporal and spatial information as possible. Therefore, the two ResNet-ConvLSTM modules in module B can more completely capture the nonlinear spatiotemporal information inherent in the data simultaneously. In module C, the first permute / reshape layer swaps the temporal and feature channel dimensions, and two 1DCNN layers are used to increase and decrease the temporal dimension, ensuring the time step matches the prediction step number K. The second permute / reshape layer swaps the data dimensions back to the original data. Finally, the 3DCNN layer II in module D restores the number of feature channels to the number F of output feature channels. O This leads to the final prediction result, which has the following advantages:

[0090] 1) It overcomes the shortcomings of the existing multi-point wind speed synchronous prediction method, which uses a "two-stage" modeling strategy that destroys the spatial relative positional relationship between measuring points and weakens spatial correlation.

[0091] 2) It overcomes the shortcomings of traditional multi-point wind speed synchronous prediction methods, which have a large resolution of the spatial wind speed matrix or require the simultaneous prediction of multiple wind speed components (i.e., predicting three-dimensional wind speed vectors), resulting in too many parameters and difficulty in training.

[0092] 3) By combining the advantages of ConvLSTM, Residual Network (ResNet) and 1×1 3D convolution, a prediction model that can simultaneously process wind speed spatiotemporal information can be developed without weakening or destroying the spatial relativity between measurement points. It can also use four-dimensional spatiotemporal wind speed data to carry out multi-point wind speed and direction synchronous prediction.

[0093] Instance verification

[0094] This embodiment uses a multi-point wind speed prediction model that simultaneously processes spatiotemporal wind speed information. The model is a spatiotemporal convolutional long short-term memory neural network with residual connections, hereinafter referred to as RSTCL. This embodiment also constructs two spatiotemporal models based on 3DCNN and ConvLSTM respectively as comparison models, named ST3DCNN and STCL, as follows: Figure 5 and 6 The neural network frameworks for the non-repeating parts of ST3DCNN, STCL, and RSTCL are shown respectively.

[0095] The changes made by ST3DCNN compared to RSTCL are as follows:

[0096] 1) The convolutional kernel of the first layer of the 3DCNN is increased to (3, 3, 3);

[0097] 2) The RC module was replaced with a two-layer 3DCNN.

[0098] This is because ST3DCNN is based on 3DCNN, so it can directly process spatiotemporal data from the beginning.

[0099] The changes of STCL compared to RSTCL are as follows:

[0100] 1) Replace the first layer of 3DCNN with ConvLSTM;

[0101] 2) The RC module is replaced with a two-layer ConvLSTM.

[0102] Since there is no need to consider changing the dimension of the feature channels to meet the input requirements of the RC module, the spatiotemporal information can be captured and learned directly by the ConvLSTM layer.

[0103] Finally, this study also introduces the persistence model (PR) as a benchmark for measuring prediction performance. Its principle is to directly use the raw data at time t as the predicted value at time t+1. Based on the above, we conduct single-step wind speed predictions and synchronous wind speed predictions for the next 1-5 time steps.

[0104] Before training, a fixed random number seed was used to ensure that the training results were not affected by the initial random number, allowing for better comparison of performance between different models and parameters. The "Adam" optimizer was used during training, with the loss function being MSE and a maximum of 300 epochs. During training, the Checkpoint function was used to automatically save the neural network parameters with the smallest error on the validation set. A training strategy of dynamically adjusting the learning rate (lr) was adopted. Each neural network was adjusted based on the fluctuation of its error function during training, reducing the learning rate before significant fluctuations. The initial learning rate was 0.001, with a dynamic adjustment range of 0.0002-0.001. The chosen deep learning platform was GPU-based TensorFlow 2.3 with Python 3.7. The parameters of each layer in the three models were manually selected to ensure that the number of trainable parameters was roughly the same for all three models, thus avoiding potential accuracy differences due to large variations in parameter count.

[0105] In addition, the relative improvement ratio P of the three indicators RMSE, MAPE, and MAE is defined. metric To further quantitatively evaluate the performance of different models. metric The calculation formula is:

[0106]

[0107] Wherein, metric represents three indicators: MAPE, MAE, and RMSE, E a E bThis represents the prediction error of model a and model b.

[0108] Tables 1 to 3 show the single-step prediction error results for each model under complex mountainous terrain, while Tables 4 to 6 show the calculated prediction errors for each model under complex mountainous terrain in 1-5 steps. When calculating the error, the error calculation rule for the resultant velocity U is as follows: first, use u, v, w, and... Calculate U and respectively Then calculate the corresponding error index. Before analyzing the results, it is important to clarify that among u, v, w, and U, the most critical physical quantities are u and U. u represents the main component of wind speed, while v and w only account for a small part of the wind speed component. The results of the optimal error level in the table are marked in bold.

[0109] Table 1. One-step prediction results (RMSE) of the spatiotemporal wind speed prediction model based on CH data.

[0110]

[0111] Table 2. One-step prediction results (MAE) of the spatiotemporal wind speed prediction model based on CH data.

[0112]

[0113] Table 3. One-step prediction results of the spatiotemporal wind speed prediction model based on CH data (MAPE)

[0114]

[0115] Table 4. Prediction results (RMSE) of the spatiotemporal wind speed prediction model based on CH data (steps 1-5).

[0116]

[0117] Table 5. Prediction results (MAE) of the spatiotemporal wind speed prediction model based on CH data (steps 1-5).

[0118]

[0119] Table 6. Prediction results of the spatiotemporal wind speed prediction model based on CH data (steps 1-5) MAPE

[0120]

[0121]

[0122] The above results demonstrate that the three spatiotemporal wind speed prediction models proposed in this embodiment achieve high prediction accuracy in both single-step and 1-5-step prediction scenarios, and significantly outperform the PR model across all error metrics. In single-step prediction, the error improvement of the three models is significantly higher than that of the PR model. metricThe accuracy ranges from 78.83% to 87.90%; under 1-5 step prediction, the error improvement ratio is P. metric The accuracy rate is 61.00% to 75.08%, which fully demonstrates the effectiveness of the spatiotemporal wind speed prediction model proposed in this embodiment.

[0123] Specifically, in single-step prediction, RSTCL and STCL based on ConvLSTM outperform ST3DCNN across all error metrics, demonstrating that ConvLSTM offers better performance than 3DCNN in handling single-step spatiotemporal wind speed prediction. Furthermore, RSTCL outperforms STCL in the vast majority of error metrics, only slightly exceeding STCL in the MAPE metric for v and w. This sufficiently proves that the strategy of encapsulating ConvLSTM with ResNet is indeed feasible.

[0124] When the prediction step size increases to 1-5 steps into the future, the errors of all prediction models, including PR, increase significantly. This also indicates that the prediction performance of neural networks decreases as the prediction step size increases, but the overall error levels of RSTCL, ST3DCNN, and STCL models are still far superior to the PR model. The error level gap between RSTCL, ST3DCNN, and STCL models is similar to that in single-step prediction, with RSTCL still achieving the best overall prediction accuracy in 1-5 step predictions. Overall, the RSTCL model exhibits superior performance in spatiotemporal wind speed prediction tasks.

[0125] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A multi-point wind speed prediction system that simultaneously processes spatiotemporal wind speed information, characterized in that: Including those connected sequentially: Module A: A 3DCNN layer I consisting of an input layer and a 1×1 convolutional layer, wherein the input layer is used to input data in the form of (…). J , X , Y , F I The 3DCNN layer I has four-dimensional data; the convolution kernel of the 3DCNN layer I is (1, 1, 1) and is used to implement the feature channels. F I The extension; among which, J Indicates the time dimension; X and Y Represents the dimensions in two directions in a spatial dimension; F I Indicates the characteristic channel; Module B: consists of two ResNet-ConvLSTM modules connected in sequence. Each ResNet-ConvLSTM module includes two ConvLSTM layers and a residual connection block, used to capture the nonlinear spatiotemporal information contained in the data center. Module C: Includes two permute / reshape layers, with two 1DCNN layers between them; the permute / reshape layer connected to Module B is used to swap the time dimension and feature channel dimension in the output data of Module B, and transform the data into ( X × Y × Ff , J ),in, Ff This indicates the number of convolutional kernels used in the previous layer of the neural network; the two 1DCNN layers are used sequentially to increase and decrease the dimensionality of the time dimension, making the time step match the number of prediction steps. K ; K It can be any positive integer; another permute / reshape layer is used to restore the data to its original order; Module D: Includes a 3DCNN layer II with 1×1 convolutions and an output layer. The number of convolutional kernels in the 3DCNN layer II is equal to the number of output feature channels required. F O The final spatiotemporal wind speed prediction result is then output through the output layer. The principle of the ConvLSTM layer is as follows: in, t and t -1 represents the current time and the previous time, respectively; , , and These represent the controllers for the forget gate, the intermediate output of the LSTM network, the input gate, and the output gate, respectively; where... For the input vector, For the long term, This is a short-term state. For the output vector, To be respectively with The connection weight matrix, To be respectively with The connection weight matrix, For bias terms; This represents the activation function. The convolution operation represents the convolutional layer. This indicates that the array elements are multiplied sequentially.

2. The multi-point wind speed prediction system that simultaneously processes spatiotemporal wind speed information according to claim 1, characterized in that: The principle of the residual connection block is as follows: h ( x )= f ( x )- x in, f ( x ) indicates the output of the residual connection block; x Indicates the input of the residual connect block ；h ( x ) represents the residual.

3. A multi-point wind speed prediction method that simultaneously processes spatiotemporal wind speed information, characterized in that: Includes the following steps: Step 1: Obtain the original spatiotemporal wind speed sequence and preprocess it to obtain four-dimensional wind speed data (Time steps, Row, Column, Features). Here, Time steps represent the number of image information at various time points in the four-dimensional wind speed data, representing the time dimension; Row and Column represent the height and width of the image, respectively, representing the spatial dimension; and Features represent the number of wind speed components, representing the feature channels. Step 2: Construct and train the multi-point wind speed prediction model as described in claim 1 or 2; Step 3: Input the four-dimensional wind speed data into the multi-point wind speed prediction model to obtain the spatiotemporal wind speed prediction results; In step one, the original spatiotemporal wind speed sequence is normalized using the following method: The wind speed vector at any time and location in the original spatiotemporal wind speed sequence is decomposed into... Three wind speed components, and respectively for The three wind speed components were normalized: in, Indicates the far-field incoming wind speed; In step two, the method for training the multi-point wind speed prediction model is as follows: 21) The preprocessed four-dimensional wind speed data U CH The training set, validation set, and test set are divided in a 6:2:2 ratio. 22) Input the training set data into the multi-point wind speed prediction model and obtain the spatiotemporal wind speed prediction value, and calculate the loss function of the spatiotemporal wind speed prediction value compared with the actual spatiotemporal wind speed value; 23) Determine if the loss function is less than the set threshold: if yes, the multi-point wind speed prediction model training is complete, and proceed to step 26); if no, increment the iteration count by 1, and proceed to step 24). 24) Determine if the number of iterations has reached the set maximum number of iterations: If yes, the multi-point wind speed prediction model training is complete, and proceed to step 26); if no, proceed to step 25). 25) Update the hyperparameters in the multi-point wind speed prediction model, and execute step 22). 26) The prediction performance of the multi-point wind speed prediction model was verified and tested using the validation machine and the test set, respectively.

4. The multi-point wind speed prediction method that simultaneously processes spatiotemporal wind speed information according to claim 3, characterized in that: In step 23), the loss function is MAPE, MAE, or RMSE.

5. The multi-point wind speed prediction method that simultaneously processes spatiotemporal wind speed information according to claim 3, characterized in that: In step 25), the hyperparameters in the multi-point wind speed prediction model are updated using Adam as the optimizer.

Images