A low-altitude three-dimensional wind field prediction method based on a segmented prediction neural network architecture

Abstract

This invention discloses a low-altitude three-dimensional wind field prediction method based on a segmented prediction neural network architecture, belonging to the technical field of meteorological forecasting. The method includes: acquiring low-altitude meteorological data and constructing a low-altitude meteorological dataset; preprocessing the dataset and dividing it into a training set and a test set; inputting the training set into the constructed segmented prediction network model to generate segmented prediction results; training the segmented prediction network model using the segmented prediction results and the training set; testing the trained segmented prediction network model using the test set; and using the tested segmented prediction network model to complete low-altitude meteorological forecasts. This invention achieves accurate forecasting in complex meteorological tasks, increases the scale of forecasting without increasing resource consumption, and the designed time-segmented weighted loss function introduces a time-series importance differentiation mechanism, improving the model's ability to perceive neighboring segments and reducing the cumulative error of multiple segments, thereby improving the accuracy of the forecast.

Description

Technical Field

[0001] This invention belongs to the technical field of meteorological forecasting, specifically relating to a method for predicting low-altitude three-dimensional wind fields based on a segmented prediction neural network architecture. Background Technology

[0002] Spatiotemporal forecasting plays a vital role in modern society, with applications spanning urban planning, public safety, and environmental monitoring. Meteorological forecasting, a key branch of spatiotemporal forecasting, enables accurate weather prediction, thereby improving the efficiency of disaster prevention and mitigation, agricultural production, transportation, and energy dispatch. However, traditional numerical weather prediction (NWP) methods primarily rely on numerical calculations based on physical laws, typically requiring significant computational resources and lengthy simulation times, making it difficult to meet the timeliness requirements of real-time forecasting.

[0003] With the rapid development of artificial intelligence technology, machine learning-based methods have gradually gained widespread attention in the field of weather forecasting due to their ability to significantly reduce computational complexity. These data-driven methods, after training, can generate forecast results within seconds, far exceeding the efficiency of traditional numerical weather prediction systems. Because meteorological data and image data share similarities, various weather forecasting schemes based on improved visual architectures are currently available. Although general spatiotemporal forecasting methods have achieved continuous performance improvements, these methods are mainly designed for relatively simple forecasting tasks. When applied to multi-task long-sequence weather forecasting, they struggle to effectively capture the inherent complex spatiotemporal dependencies in meteorological data. For example, the SimVP scheme uses multi-scale convolutional kernels to enhance spatial feature extraction capabilities in video forecasting tasks; the LMC_memory scheme utilizes memory mechanisms to model long-term action dependencies to improve action prediction; and the WAST scheme uses wavelet transform to represent frequency domain data to improve traffic forecasting accuracy. However, such tasks typically involve fewer variables and have lower spatiotemporal complexity, making it difficult for these methods to achieve accurate predictions when dealing with complex weather forecasting tasks. Current spatiotemporal forecasting methods suffer from another significant drawback: as the scale of the predicted spatiotemporal data increases (manifested as longer prediction sequences or more predictor variables), the required training computational resources increase accordingly, but the prediction accuracy fails to improve and may even decrease. The conventional solution in meteorology is to invest even more resources. For example, while existing methods achieve accurate medium-range global weather forecasts, training a single model requires approximately 16 days on a cluster of 192 NVIDIA Tesla-V100 graphics cards; existing methods using graph neural networks to model meteorological dynamics require approximately four weeks of training on 32 Cloud TPU v4 devices. Such resource demands far exceed the capacity of ordinary research teams.

[0004] The shortcomings of existing technologies include:

[0005] 1. Traditional numerical weather prediction usually involves complex systems of equations and numerical calculations, requiring a large amount of computing resources. Especially when processing large-scale data, it is difficult to achieve efficient prediction.

[0006] Meteorological forecasting enables accurate weather prediction, thereby improving the efficiency of activities such as disaster prevention and mitigation, agricultural production, transportation, and energy dispatch. However, traditional numerical weather prediction (NWP) methods mainly rely on numerical calculations based on physical laws. These methods are computationally complex and time-consuming, resulting in low efficiency in real-time applications. Therefore, there is an urgent need to develop new forecasting models to improve the speed and accuracy of meteorological forecasting.

[0007] 2. Although deep learning spatiotemporal prediction models can model change patterns, they are still insufficient to capture complex spatiotemporal dynamics in complex tasks such as weather forecasting.

[0008] Existing forecasting methods have shown continuous improvement in performance for spatiotemporal forecasting tasks involving fewer variables and lower spatiotemporal complexity, but they still have shortcomings when dealing with complex weather forecasting tasks. Due to the similarity between meteorological data and image data, several weather forecasting schemes based on improved visual architectures have been developed, but these methods are mainly aimed at relatively simple forecasting tasks. When applied to multi-task long-sequence weather forecasting, they struggle to effectively capture the inherent complex spatiotemporal features of meteorological data.

[0009] 3. Other significant drawbacks of current prediction methods: When the scale of the spatiotemporal data to be predicted increases (manifested as an extension of the prediction sequence or an increase in the number of prediction variables), the required training computational resources increase accordingly, but the prediction accuracy fails to improve or may even decrease.

[0010] The conventional solution in the meteorological field is to invest more resources, but such resource requirements far exceed the capacity of ordinary research teams. Summary of the Invention

[0011] The purpose of this invention is to address the aforementioned shortcomings of existing technologies by providing a low-altitude three-dimensional wind field prediction method based on a segmented predictive neural network architecture. This addresses the problem that existing methods primarily focus on pattern learning for relatively simple tasks, and struggle to effectively capture the inherent complex spatiotemporal characteristics of meteorological changes when faced with complex weather forecasting tasks. Furthermore, existing prediction strategies lead to continuously increasing resource consumption and performance degradation when training models to predict large-scale spatiotemporal data.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0013] A method for predicting low-altitude three-dimensional wind fields based on a piecewise predictive neural network architecture includes the following steps:

[0014] S1. Obtain low-altitude meteorological data and construct a low-altitude meteorological dataset;

[0015] S2. Preprocess the dataset and divide it into a training set and a test set;

[0016] S3. Input the training set into the constructed segmented prediction network model to generate segmented prediction results;

[0017] S4. Train the segmented prediction network model using the segmented prediction results and the training set;

[0018] S5. Use the test set to test the trained segmented prediction network model, and use the tested segmented prediction network model to complete the low-altitude weather prediction.

[0019] This invention reduces prediction complexity and cumulative error by designing separate prediction time segments and reallocating losses for these time segments. Addressing the limitations of existing spatiotemporal prediction models in capturing changing patterns in complex long-sequence weather forecasts, this invention reduces prediction complexity through a segmented prediction strategy and incorporates a spatiotemporal attention module to effectively capture spatiotemporal features. To address the complexity of long-sequence prediction, the segmented prediction strategy predicts only a small portion of the time series at a time, reducing the fitting difficulty of a single prediction. Furthermore, the spatiotemporal attention module focuses on local spatiotemporal regions, performing fine-grained evolution on local features to further enhance the network's ability to capture spatiotemporal features. Experiments on different meteorological datasets demonstrate that this invention achieves an average performance improvement of 27% compared to existing prediction methods.

[0020] Furthermore, in S1, a high-resolution near-surface meteorological dataset is constructed using the WRF-PALM large eddy simulation framework. This dataset includes wind speed, wind direction, temperature, humidity, and turbulent kinetic energy, and covers a variety of meteorological conditions and seasonal characteristics.

[0021] Furthermore, S2 specifically includes: taking low-altitude meteorological data as a sample with 5 frames of input and 30 frames of output, and then dividing the low-altitude meteorological dataset into a training set and a test set.

[0022] Furthermore, step S3 specifically includes the following sub-steps:

[0023] S31. Preprocess the training set to obtain the input data;

[0024] S32. Input the input data into the segmented prediction network model and output the segmented prediction results;

[0025] S33. Calculate the loss function of the segmented prediction network model.

[0026] Furthermore, in step S31, the training set is preprocessed to obtain input data, including:

[0027] The low-altitude meteorological data in the training set are randomly sampled sequentially. Ten time segments are segmented from the 30-frame sequence of the label of each sample. One time segment is randomly selected from the ten time segments as the true value of the label for segment prediction. The segment codes corresponding to the segment index are then merged as input data.

[0028] Furthermore, in S32, the segmented prediction network model includes an encoder and a decoder; the encoder includes four encoding layers, and the decoder includes four decoding layers; each encoding layer includes a convolutional feedforward network and a downsampling layer; each decoding layer includes an upsampling layer and an STA Block module.

[0029] The specific process of generating segmented prediction results by inputting the training set into the segmented prediction network model is as follows:

[0030] The input data enters the first encoding layer, and after processing by the convolutional feedforward network and the downsampling layer, the first feature map is obtained;

[0031] The first feature map enters the second encoding layer, and after processing by a convolutional feedforward network and a downsampling layer, the second feature map is obtained.

[0032] The second feature map enters the third encoding layer, and after processing by a convolutional feedforward network and a downsampling layer, the third feature map is obtained.

[0033] The third feature map enters the fourth coding layer, and after being processed by a convolutional feedforward network and a downsampling layer, the fourth feature map is obtained.

[0034] The fourth feature map enters the first decoding layer, and after upsampling and processing by the STA Block module, the fifth feature map is obtained;

[0035] The fifth feature map is added to the third feature map and then enters the second decoding layer. After upsampling and processing by the STA Block module, the sixth feature map is obtained.

[0036] The sixth feature map is added to the second feature map and then enters the third decoding layer. After upsampling and processing by the STA Block module, the seventh feature map is obtained.

[0037] The seventh feature map is added to the first feature map and then enters the fourth decoding layer. After upsampling and processing by the STA Block module, the eighth feature map is obtained. The eighth feature map is then processed by the STA Block module to obtain the segmented prediction result.

[0038] Furthermore, the STA Block module includes a convolutional feedforward network, a spatiotemporal attention module, and a convolutional feedforward network arranged in sequence.

[0039] The specific process of the convolutional feedforward network is as follows: the input feature map enters the convolutional feedforward network, and after passing through two consecutive Norm normalization layers, activation functions and convolutions, it is residually connected with the input feature map to obtain the output result;

[0040] The specific process of the spatiotemporal attention module is as follows: the input feature map enters the spatiotemporal attention module and is processed by depthwise separable convolution, dilated convolution and ordinary convolution to obtain query component, key component and value component;

[0041] The attention score is obtained by multiplying the query component and the key component element by element, and the attention score is then normalized using an activation function to obtain the attention weight.

[0042] The spatiotemporal feature map is output by multiplying the attention weights by the value components.

[0043] Furthermore, in S33, the loss function is the sum of the piecewise weighted errors of all samples in the batch, expressed as:

[0044]

[0045] in:

[0046]

[0047]

[0048] In the formula, L B represents the loss of the segmented prediction network model; B is the batch size, representing the total number of samples in the batch. To predict the mean squared error between the segment and the label sequence; The index representing the time segment corresponding to each random sample. Normalized weighting coefficients; The index represents the time segment of the random sampling, where, , , Indicates uniform distribution. It is an integer. The number of time segments set for the low-altitude meteorological dataset; It is a hyperparameter; This means traversing the same batch of samples and normalizing the segment weights of individual samples; For prediction segmentation; A sequence of labels; It is the Euclidean norm; , This is an index for the samples in the batch.

[0049] The low-altitude three-dimensional wind field prediction method based on a segmented predictive neural network architecture provided by this invention has the following beneficial effects:

[0050] 1. This invention enables accurate forecasting in complex meteorological tasks, and can increase the scale of forecasting without increasing resource consumption. Furthermore, the designed time-segmented weighted loss function introduces a time-series importance differentiation mechanism, which improves the model's ability to perceive neighboring segments, reduces the cumulative error of multiple segments, and thus improves the accuracy of forecasting.

[0051] 2. This invention introduces a spatiotemporal attention module (STA): In order to capture the dynamic spatiotemporal features in complex data, a spatiotemporal attention mechanism is introduced to enhance spatiotemporal modeling capabilities and achieve accurate capture of multidimensional spatiotemporal dependencies;

[0052] The segmented prediction strategy of this invention: For large-scale prediction tasks where resource consumption increases during training, a segmented prediction strategy is designed to significantly reduce computational cost and prediction complexity by processing shorter sequence segments;

[0053] The present invention provides a time-segmented weighted loss function: This function dynamically adjusts the network's attention distribution to different time segments, effectively suppressing the cumulative effect of prediction errors.

[0054] 3. To capture the spatiotemporal features in complex tasks, this invention designs a secondary spatiotemporal attention mechanism for local feature points. By weighting the local spatiotemporal information, the model's ability to perceive spatiotemporal changes is enhanced, thereby improving prediction accuracy. Attached Figure Description

[0055] Figure 1 This is a network structure diagram of the low-altitude three-dimensional wind field prediction method based on a segmented prediction neural network architecture in the embodiment.

[0056] Figure 2 This is a structural diagram of the spatiotemporal attention module (STA) and the convolutional feedforward network (FFN) in the embodiment.

[0057] Figure 3 This is a qualitative result diagram of the low-altitude meteorological dataset in the example.

[0058] Figure 4 This is a qualitative result diagram of wave height forecast in the example.

[0059] Figure 5 This is a flowchart of the low-altitude three-dimensional wind field prediction method based on a segmented prediction neural network architecture, as shown in the embodiment. Detailed Implementation

[0060] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0061] In the embodiments of this invention or the accompanying drawings, STA refers to: spatiotemporal attention module.

[0062] FFN stands for Convolutional Feedforward Network.

[0063] Up refers to the upsample layer.

[0064] Down refers to the downsample layer.

[0065] Conv stands for Convolution layer.

[0066] T stands for Segment Encoding.

[0067] DSC stands for Depthwise Separable Convolution.

[0068] ASC stands for Atrous Convolution.

[0069] Norm refers to normalization.

[0070] This embodiment provides a low-altitude three-dimensional wind field prediction method based on a segmented prediction neural network architecture. It reduces prediction complexity through a segmented prediction strategy and incorporates a spatiotemporal attention module to effectively capture spatiotemporal features. Furthermore, the spatiotemporal attention module focuses on local spatiotemporal regions, performing fine-grained evolution on local features to further enhance the network's ability to capture spatiotemporal features. (Refer to...) Figure 1 and Figure 5 Specifically, it includes the following steps:

[0071] S1. Obtain low-altitude meteorological data and construct a low-altitude meteorological dataset;

[0072] In some embodiments, a high-resolution near-surface low-air meteorological dataset was constructed using the WRF-PALM large eddy simulation framework. This dataset includes key variables such as wind speed, wind direction, temperature, humidity, and turbulent kinetic energy (TKE), covering various meteorological conditions (e.g., calm winds, precipitation) and seasonal characteristics (e.g., summer heat waves, winter temperature inversions). Ultimately, 12 meteorological variables were selected, distributed across 8 pressure layers corresponding to altitudes below 300 meters; and the data was finally resampled to... Spatial resolution of grid points.

[0073] S2. Preprocess the dataset and divide it into training and test sets;

[0074] In some embodiments, the low-altitude meteorological dataset is preprocessed, and from start to finish, 5 frames of input and 30 frames of output are taken as a sample. The preprocessed low-altitude meteorological prediction dataset is then divided into a training set and a test set.

[0075] S3. Input the training set into the constructed segmented prediction network model to generate segmented prediction results;

[0076] S31. Preprocess the training set to obtain the input data;

[0077] In one specific embodiment, the preprocessed training set is randomly sampled sequentially, and the 30-frame sequence of each sample's label is divided into 10 time segments. One of the 10 time segments is randomly selected as the true value of the label for segment prediction; and the segment codes corresponding to the segment index are merged as input data.

[0078] S32. Input the input data into the segmented prediction network model and output the segmented prediction results;

[0079] In one specific embodiment, the segmented prediction network model includes an encoder and a decoder. The encoder includes four encoding layers; the decoder includes four decoding layers; each encoding layer includes a convolutional feedforward network and a downsampling layer; each decoding layer includes an upsampling layer and an STA Block module.

[0080] In one specific embodiment, the training set is input into the segmented prediction network model to generate segmented prediction results. The specific process is as follows:

[0081] The input data enters the first encoding layer, and after processing by the convolutional feedforward network and the downsampling layer, the first feature map is obtained;

[0082] The first feature map enters the second encoding layer, and after processing by a convolutional feedforward network and a downsampling layer, the second feature map is obtained.

[0083] The second feature map enters the third encoding layer, and after processing by a convolutional feedforward network and a downsampling layer, the third feature map is obtained.

[0084] The third feature map enters the fourth coding layer, and after being processed by a convolutional feedforward network and a downsampling layer, the fourth feature map is obtained.

[0085] The fourth feature map enters the first decoding layer, and after upsampling and processing by the STA Block module, the fifth feature map is obtained;

[0086] The fifth feature map is added to the third feature map and then enters the second decoding layer. After upsampling and processing by the STA Block module, the sixth feature map is obtained.

[0087] The sixth feature map is added to the second feature map and then enters the third decoding layer. After upsampling and processing by the STA Block module, the seventh feature map is obtained.

[0088] The seventh feature map is added to the first feature map and then enters the fourth decoding layer. After upsampling and processing by the STA Block module, the eighth feature map is obtained. The eighth feature map is then processed by the STA Block module to obtain the segmented prediction result.

[0089] In one specific embodiment, the STA Block module includes a convolutional feedforward network (FFN), a spatiotemporal attention module (STA), and a convolutional feedforward network (FFN) arranged in sequence.

[0090] Among them, reference Figure 2 A convolutional feedforward network is a standard convolutional network. Its specific process is as follows: the input feature map enters the convolutional feedforward network, passes through two consecutive Norm normalization layers, activation functions and convolutions, and then is residually connected with the input feature map to obtain the output result.

[0091] refer to Figure 2 The specific process of the spatiotemporal attention module is as follows: the input feature map enters the spatiotemporal attention module, and after depthwise separable convolution, dilated convolution and ordinary convolution, the query component, key component and value component are obtained;

[0092] The attention score is obtained by multiplying the query component and the key component element by element, and the attention score is then normalized using an activation function to obtain the attention weight.

[0093] The spatiotemporal feature map is output by multiplying the attention weights by the value components.

[0094] S33. Calculate the loss function (time-segmented weighted loss function) of the segmented prediction network model.

[0095] In one specific embodiment, a batch of samples contains B samples, where B represents the batch size (i.e., the total number of samples in the batch); the sample index is denoted as b, and all samples are traversed, i.e., b ranges from 0 to B-1 (inclusive).

[0096] For each sample corresponding to index b, its label comes from a time segment uniformly randomly sampled from the complete 30-frame prediction sequence, and the segment index is... The range is the set of integers {0,1,2,…,n-1}, where n is the number of segments (10) set in the low-altitude meteorological data set.

[0097] Time segment index for each sample Calculate a normalized weighting coefficient; the normalized weighting coefficient is:

[0098]

[0099]

[0100] For each sample corresponding to index b, calculate an error term. This error term represents the mean square error (MSE) between the predicted segment and the label sequence;

[0101] The loss is obtained by reweighting the mean squared error (MSE) between the predicted segments and the label sequences of different time segments randomly sampled from the same batch of input models. This loss function (time segment weighted loss function) is then used to obtain the loss. L The sum of the piecewise weighted errors of all samples in the batch is expressed as:

[0102]

[0103] In the formula, L B represents the loss of the segmented prediction network model; B is the batch size, representing the total number of samples in the batch. To predict the mean squared error between the segment and the label sequence; The index representing the time segment corresponding to each random sample. Normalized weighting coefficients; The index represents the time segment of the random sampling, where, , , Indicates uniform distribution. It is an integer. The number of time segments set for the low-altitude meteorological dataset; It is a hyperparameter; This means traversing the same batch of samples and normalizing the segment weights of individual samples; For prediction segmentation; A sequence of labels; It is the L2 norm (Euclidean norm); , This is an index for the samples in the batch.

[0104] S4. Train the segmented prediction network model using the segmented prediction results and the training set;

[0105] In some embodiments, the mean squared error (MSE) between the predicted segments and the label sequence is calculated; the calculated MSE is then processed according to the batch dimension. The loss function L is calculated using a piecewise weighted approach, followed by gradient descent optimization to complete the training of the SegSTA (SegPredictive Network) model.

[0106] S5. Use the test set to test the trained segmented prediction network model, and use the tested segmented prediction network model to complete the low-altitude weather prediction.

[0107] In some embodiments, when testing the capability of a validation model, in order to output a complete sequence rather than individual time segments, for a batch of validation samples, the validation samples are sequentially encoded with the segment codes. The input model is merged to obtain predictions for each segment from 1 to 15. The final prediction result is obtained by merging the predictions for all time segments.

[0108] In some embodiments, the effectiveness of the network structure (SegSTA) of the present invention is verified by the following:

[0109] To address the problem that existing prediction methods have difficulty fitting complex tasks, this invention proposes a segmented prediction network model, SegSTA. Figure 1 This is a schematic diagram of the network structure proposed in this invention. Its input data... It consists of multiple variables and multiple pressure layers from five frames of low-altitude weather, then combined with segmented coding. The merged data serves as the final input. This input data, after passing through four encoders consisting of FFN and Down encoders, is then fed into the decoder containing the spatiotemporal attention module. Finally, the predicted output for the i-th time segment is obtained. The truth value of the corresponding i-th time segment is Frames It involves segmenting from a complete sequence. By randomly sampling a time segment and splitting the prediction sequence, the difficulty of fitting is reduced.

[0110] Figure 2This is the network structure diagram of the STA spatiotemporal attention module and the FFN convolutional feedforward network, where the spatiotemporal attention module captures spatiotemporal features. For the input features, they are processed by DSC depthwise separable convolution, ASC dilated convolution, and ordinary convolution to obtain the key component K and query component Q. The input features are then processed by ordinary convolution to obtain the value component V. Q and K are then multiplied element-wise. The attention score is obtained, and then normalized using the sigmoid activation function to obtain the attention weight. Finally, the attention weight is obtained by element-wise multiplication with V to get the final spatiotemporal attention output. The FFN convolutional feedforward network is a standard residual network. The input features pass through two Norm normalization layers, activation functions, and ordinary convolutions, and are then residually connected to the input to obtain the final result. It is responsible for the basic computational mapping.

[0111] To further illustrate the effectiveness of the method of this invention, it is compared with other existing methods. For a fair comparison, the officially released code of the other methods is used, and their experimental settings are followed, where all methods are implemented in the same computational environment, and both quantitative and qualitative analyses are performed simultaneously. The six methods are as follows:

[0112] Method 1 is the FouCastNet method, which uses an adaptive Fourier neural operator as its main architecture and can generate a large number of sub-models at a very low cost for ensemble prediction.

[0113] Method 2, SimVP, achieves excellent performance with a simple network structure, proving that complex networks such as LSTM may not be the best choice for spatiotemporal prediction.

[0114] Method 3 is SimVPv2, which achieves a performance breakthrough by completely abandoning the complex decoding structure through pure convolutional layer stacking and an innovative gated spatiotemporal attention mechanism.

[0115] Method 4, TAU, uses an attention mechanism to process temporal evolution in parallel. A novel differential dispersion regularization method is also proposed to optimize the loss function for spatiotemporal prediction learning.

[0116] Method 5 is PredFormer, which completely abandons the traditional recursive structure and CNN encoder-decoder framework. It is the first to propose a pure Transformer spatiotemporal prediction model. Through innovative gated Transformer blocks and 3D attention mechanism optimization, it achieves a breakthrough improvement in accuracy and efficiency.

[0117] Method 6, Earthfarseer, replaces complex customized models with a simplified design of "local convolution + global Fourier transform," providing a new paradigm for predicting spatiotemporal physical systems that combines powerful performance with engineering practicality.

[0118] Since the method of this invention is mainly aimed at complex low-altitude meteorological forecasting tasks, Table 1 presents the quantitative evaluation of all methods on the low-altitude meteorological dataset, using mean squared error (MSE), mean absolute error (MAE), and mean absolute deviation (MAD). The smaller the values ​​of MAE, MSE, and MAD, the better the performance.

[0119] Table 1

[0120]

[0121] By employing a segmented prediction strategy, the method of this invention effectively reduces the complexity of each prediction output and demonstrates strong performance in such challenging scenarios. As shown in Table 1, the quantitative results show that this model achieves state-of-the-art performance across all evaluation metrics: compared to the strongest comparison method 4, the method of this invention improves the MSE metric by 38%, MAE by 26%, and MAD by 29%. Figure 3 Qualitative results demonstrate that by assigning differentiated weights to time segments, the method of this invention can accurately capture fine-grained details of early frames while maintaining a stable trend in later frames. Most baseline methods can only reconstruct the local structure of neighboring frames (e.g., at t=6), while the method of this invention maintains stable and accurate predictions throughout the entire sequence; these results fully demonstrate the robustness and applicability of the method of this invention in complex prediction tasks.

[0122] Table 2 presents the quantitative evaluation of different methods on the wave forecast dataset, using mean squared error (MSE), mean absolute error (MAE), and R² score. Lower values ​​for MAE and MSE indicate better performance, while higher R² scores indicate better performance.

[0123] Table 2

[0124]

[0125] Wave height prediction involves high complexity because the dataset needs to learn complex wave patterns from multiple input variables. Current popular spatiotemporal prediction methods based on direct prediction strategies often struggle to cope with complex ocean environments, leading to high relative errors. In contrast, the method of this invention captures spatiotemporal features more effectively. As shown in Table 2, the method of this invention achieves the best performance across all metrics. Specifically, compared to the strongest comparison method 2, the method of this invention improves the MSE metric by 16%, MAE by 10%, and R2 score by 1%. These improvements are attributed to the use of the Spatiotemporal Attention (STA) module—which enhances spatiotemporal representation capabilities—and the application of a segmented prediction strategy—which reduces prediction complexity. Qualitative results are as follows: Figure 4As shown, compared with other methods, the model of this invention exhibits a more refined prediction pattern, highlighting its potential value in operational ocean forecasting.

[0126] Table 3 validates the effectiveness of each proposed module through ablation experiments. The ablation experiments evaluated the MSE metrics on two datasets, showing that the segmented prediction strategy significantly improves performance, demonstrating the substantial practical application value of the proposed method. Removing the spatiotemporal attention module reveals a performance degradation on different datasets, indicating the crucial role of spatiotemporal attention in capturing spatiotemporal features. Furthermore, segmented loss can be improved by adjusting the focus of the proposed method on different time segments, suggesting that the focus should be prioritized for different time segments.

[0127] Table 3

[0128]

[0129] In summary, accurate spatiotemporal modeling is essential for weather forecasting; however, existing methods primarily focus on learning patterns from relatively simple tasks. For example, when faced with complex weather forecasting tasks, they often struggle to capture the inherent complex spatiotemporal characteristics of weather changes. Furthermore, training these models to predict large-scale spatiotemporal data leads to increasing resource consumption and performance degradation. To address these challenges, this invention proposes a novel spatiotemporal segmented prediction framework, SegSTA. Specifically, to capture the dynamic spatiotemporal features in complex data, a spatiotemporal attention mechanism (STA) is introduced to enhance spatiotemporal modeling. For large-scale prediction tasks with high resource consumption during training, a segmented prediction strategy is designed, significantly reducing computational cost and prediction complexity by operating on shorter sequence segments. Additionally, a time-time weighted loss function is introduced to dynamically recalibrate the network's attention distribution across different time periods, effectively reducing accumulated prediction errors. The method was evaluated on multiple spatiotemporal prediction datasets, achieving a 27% accuracy improvement. Furthermore, integrating the segmented prediction strategy into other spatiotemporal models can continuously improve their performance. These results demonstrate the great potential of the framework of this invention in practical spatiotemporal prediction applications.

[0130] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for predicting low-altitude three-dimensional wind fields based on a piecewise predictive neural network architecture, characterized in that, Includes the following steps: S1. Obtain low-altitude meteorological data and construct a low-altitude meteorological dataset; S2. Preprocess the dataset and divide it into a training set and a test set; S3. Input the training set into the constructed segmented prediction network model to generate segmented prediction results; S3 specifically includes the following steps: S31. Preprocess the training set to obtain the input data; S32. Input the input data into the segmented prediction network model and output the segmented prediction results; S33. Calculate the loss function of the segmented prediction network model; in S33, the loss function is the sum of the segmented weighted errors of all samples in the batch, expressed as: in: In the formula, L B is the loss of the segmented prediction network model; B is the batch size, representing the total number of samples in the batch. To predict the mean squared error between the segment and the label sequence; The index representing the time segment corresponding to each random sample. Normalized weighting coefficients; The index represents the time segment of the random sampling, where, , , Indicates uniform distribution. It is an integer. The number of time segments set for the low-altitude meteorological dataset; It is a hyperparameter; This means traversing the same batch of samples and normalizing the segment weights of individual samples; For prediction segmentation; A sequence of labels; It is the Euclidean norm; , This is an index for samples within a batch; S4. Train the segmented prediction network model using the segmented prediction results and the training set; S5. Use the test set to test the trained segmented prediction network model, and use the tested segmented prediction network model to complete the low-altitude weather prediction.

2. The low-altitude three-dimensional wind field prediction method based on a piecewise predictive neural network architecture according to claim 1, characterized in that, In S1, a high-resolution near-surface meteorological dataset is constructed using the WRF-PALM large eddy simulation framework. This dataset includes wind speed, wind direction, temperature, humidity, and turbulent kinetic energy, and covers a variety of meteorological conditions and seasonal characteristics.

3. The low-altitude three-dimensional wind field prediction method based on a piecewise predictive neural network architecture according to claim 1, characterized in that, S2 specifically includes: taking low-altitude meteorological data as a sample with 5 frames of input and 30 frames of output, and then dividing the low-altitude meteorological dataset into a training set and a test set.

4. The low-altitude three-dimensional wind field prediction method based on a piecewise predictive neural network architecture according to claim 1, characterized in that, In step S31, the training set is preprocessed to obtain input data, including: The low-altitude meteorological data in the training set are randomly sampled sequentially. Ten time segments are segmented from the 30-frame sequence of the label of each sample. One time segment is randomly selected from the ten time segments as the true value of the label for segment prediction. The segment codes corresponding to the segment index are then merged as input data.

5. The low-altitude three-dimensional wind field prediction method based on a piecewise predictive neural network architecture according to claim 1, characterized in that, In S32, the segmented prediction network model includes an encoder and a decoder; the encoder includes four encoding layers, and the decoder includes four decoding layers; each encoding layer includes a convolutional feedforward network and a downsampling layer; each decoding layer includes an upsampling layer and an STA Block module. The specific process of generating segmented prediction results by inputting the training set into the segmented prediction network model is as follows: The input data enters the first encoding layer, and after processing by the convolutional feedforward network and the downsampling layer, the first feature map is obtained; The first feature map enters the second encoding layer, and after processing by a convolutional feedforward network and a downsampling layer, the second feature map is obtained. The second feature map enters the third encoding layer, and after processing by a convolutional feedforward network and a downsampling layer, the third feature map is obtained. The third feature map enters the fourth coding layer, and after being processed by a convolutional feedforward network and a downsampling layer, the fourth feature map is obtained. The fourth feature map enters the first decoding layer, and after upsampling and processing by the STA Block module, the fifth feature map is obtained; The fifth feature map is added to the third feature map and then enters the second decoding layer. After upsampling and processing by the STA Block module, the sixth feature map is obtained. The sixth feature map is added to the second feature map and then enters the third decoding layer. After upsampling and processing by the STA Block module, the seventh feature map is obtained. The seventh feature map is added to the first feature map and then enters the fourth decoding layer. After upsampling and processing by the STA Block module, the eighth feature map is obtained. The eighth feature map is then processed by the STA Block module to obtain the segmented prediction result.

6. The low-altitude three-dimensional wind field prediction method based on a piecewise predictive neural network architecture according to claim 5, characterized in that, The STA Block module includes a convolutional feedforward network, a spatiotemporal attention module, and a convolutional feedforward network arranged in sequence. The specific process of the convolutional feedforward network is as follows: the input feature map enters the convolutional feedforward network, and after passing through two consecutive Norm normalization layers, activation functions and convolutions, it is residually connected with the input feature map to obtain the output result; The specific process of the spatiotemporal attention module is as follows: the input feature map enters the spatiotemporal attention module and is processed by depthwise separable convolution, dilated convolution and ordinary convolution to obtain query component, key component and value component; The attention score is obtained by multiplying the query component and the key component element by element, and the attention score is then normalized using an activation function to obtain the attention weight. The spatiotemporal feature map is output by multiplying the attention weights by the value components.

Images