A thunderstorm prediction method based on spatiotemporal memory decoupled RNN

Abstract

The application discloses a thunderstorm prediction method based on a space-time memory decoupling RNN, and the method steps comprise the following steps: acquiring meteorological radar echo data, pre-processing the meteorological radar echo data to obtain a radar echo grayscale image, on the basis of an ST-LSTM unit, constructing a recurrent neural network model with a memory decoupling module, converting the meteorological radar echo grayscale image into a time sequence, and dividing the time sequence into a training set and a test set. The training set is substituted into the recurrent neural network with the memory decoupling module in step S2 for training, and the test set is used for verification in the training process, and the training is stopped when the training frequency reaches a preset frequency. The trained network is used for thunderstorm prediction on the test set, and a short-term near-time thunderstorm prediction image time sequence is obtained. The application is trained to focus on different aspects of space-time changes of thunderstorms in different memory states, does not learn redundant features, and improves the accuracy of thunderstorm prediction to a certain extent.

Description

Technical Field

[0001] This invention belongs to the basic field of meteorological radar signal processing, and in particular relates to a thunderstorm prediction method based on spatiotemporal memory decoupling RNN. Background Technology

[0002] Thunderstorms are localized severe convective weather events, typically accompanied by lightning, hail, heavy rainfall, tornadoes, and other violent weather phenomena. These phenomena can seriously threaten industrial and agricultural production and people's daily lives. Statistics show that approximately 44,000 thunderstorms occur globally every day, affecting about 1% of the world's surface. In 1987, the United Nations listed thunderstorms as one of the ten most serious natural disasters. Meanwhile, with advancements in technology, society has placed higher demands on the timeliness, effectiveness, and accuracy of weather forecasts; therefore, providing more accurate and efficient short-term forecasts is particularly important.

[0003] Due to the limited temporal and spatial resolution of conventional data, and the short lifecycle and small spatial scale of thunderstorms, detection and early warning are difficult. Doppler weather radar data, with its high temporal and spatial resolution, is a primary tool for detecting severe convective weather systems and providing short-term nowcasting. Traditional thunderstorm prediction methods mainly include the single-cell centroid method, optical flow method, and cross-correlation method. These methods only use radar echo images from a few consecutive moments to infer the location of the thunderstorm at the next moment, ignoring the fact that thunderstorm movement is usually nonlinear in reality. They also suffer from limitations in utilizing historical radar data and having short extrapolation timelines. In contrast, neural network models have powerful nonlinear mapping capabilities and significantly improve data utilization.

[0004] Continuous radar observation images are time-series data. To achieve good extrapolation results, it is necessary to fully consider the temporal correlation between radar reflectivity factors at adjacent time points. Recurrent neural networks (RNNs) are commonly used to model data related to time-series problems. RNNs not only have forward feedback but also backward feedback, utilizing the internal neural network multiple times through internal loops. Their input depends on the current input and the memory of historical inputs, which can persist key information in time-series data, giving them strong time-series processing capabilities. In recent years, the widespread application of deep learning in tasks such as natural language processing and image classification has attracted the attention of meteorological researchers, who are attempting to apply deep learning to meteorology. However, existing models such as RNNs and Long Short-Term Memory (LSTM) networks struggle to capture long-term memory features, and inefficient use of network parameters affects thunderstorm prediction results. Summary of the Invention

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A thunderstorm prediction method based on spatiotemporal memory decoupled RNN has the following steps:

[0007] S1. Acquire meteorological radar echo data, preprocess the meteorological radar echo data, and obtain radar echo grayscale image.

[0008] S2. Based on the ST-LSTM unit, construct a recurrent neural network model with a memory decoupling module;

[0009] S3. Convert the grayscale images of weather radar echoes into time series and divide them into training and test sets. Substitute the training set into the recurrent neural network with the memory decoupling module in S2 for training, and use the test set for classification and recognition verification during the training process. Stop training when the preset number of training iterations is reached. Use the trained network to predict thunderstorms on the test set to obtain the short-term near-term thunderstorm prediction image time series.

[0010] Furthermore, the meteorological radar echo data in S1 includes: filtering out isolated noise points, smoothing filtering, and constructing a high-pass filter to filter out reflectivity factor values ​​that are unrelated to precipitation.

[0011] Furthermore, the data preprocessing process in S1 involves normalizing the filtered weather radar echo data and converting it into a grayscale image. At the same time, grayscale images with an effective reflectivity factor value ratio less than a threshold are filtered out and cut into images with a resolution of 128×128.

[0012] Furthermore, the pixel value range of the grayscale image is [0, 255].

[0013] Furthermore, the grayscale images whose effective reflectance factor value is less than a threshold are filtered out. The threshold is 0.1, and the effective reflectance factor value is 20dBz.

[0014] Furthermore, the recurrent neural network model composed of ST-LSTM units in S2 has two types of memory units: time memory units. It is a layer within each ST-LSTM layer that transitions from the previous node t-1 to the current node at the same time; spatial memory unit Vertical transition from the lower l-1 ST-LSTM layer to the current layer at the same time point; especially for the bottom ST-LSTM with l=1, Assign to Because memory conversion functions for different propagation directions should be controlled by different signals, It adopts the original gate structure of standard LSTM and provides... Construct another set of input gates, forget gates, and input modulation gates.

[0015] Furthermore, the memory decoupling module in S2 is represented as follows:

[0016]

[0017]

[0018]

[0019] Among them, W d This represents a 1×1 convolution shared by all network layers. <·,·> represent dot products, and ‖·‖ represents the planarization of the feature map. Norm.

[0020] Furthermore, in S3, the weather radar echo grayscale image is converted into a time series. Each time series is 20 frames long, with a 6-minute interval between each frame. 10 frames are input and 10 frames are output.

[0021] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0022] (1) This invention introduces a dual-stream conversion mechanism ST-LSTM unit that combines the original memory stream and the new spatial memory stream, so that the network can simultaneously pursue short-term recursion depth and long-term correlation, and has a more detailed memory of thunderstorm changes.

[0023] (2) Based on the ST-LSTM model, this invention adds a memory decoupling structure, adds a convolutional layer to the increment of each memory state, and uses a new decoupling loss to explicitly expand the distance between the temporal memory state and the spatial memory state in the latent space. Different memory states are trained to focus on different aspects of the spatiotemporal changes of thunderstorms without learning redundant features. As a result, the network can more effectively capture long-term and short-term spatiotemporal features and improve the accuracy of thunderstorm prediction. Attached Figure Description

[0024] Figure 1 This is a flowchart of the thunderstorm prediction method of the present invention;

[0025] Figure 2 This is a flowchart of the meteorological radar data preprocessing process of the present invention;

[0026] Figure 3 This is a schematic diagram of the network model structure of the present invention;

[0027] Figure 4 This is a schematic diagram of the ST-LSTM cell structure of the present invention;

[0028] Figure 5 This is an example extrapolated image sequence of the thunderstorm prediction method based on spatiotemporal memory decoupled RNN of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 As shown, the steps of a thunderstorm prediction method based on a spatiotemporal memory decoupled RNN recurrent neural network are as follows:

[0031] S1. Acquire meteorological radar echo data, preprocess the meteorological radar echo data, and obtain radar echo grayscale image.

[0032] S2. Based on the ST-LSTM unit, construct a recurrent neural network model with a memory decoupling module;

[0033] S3. Convert the grayscale images of weather radar echoes into time series and divide them into training and test sets. Substitute the training set into the recurrent neural network with the memory decoupling module in S2 for training, and use the test set for classification and recognition verification during the training process. Stop training when the preset number of training iterations is reached. Use the trained network to predict thunderstorms on the test set to obtain the short-term near-term thunderstorm prediction image time series.

[0034] The meteorological radar echo data in S1 includes: isolated noise point filtering, smoothing filtering, and constructing a high-pass filter to filter out reflectivity factor values ​​that are unrelated to precipitation.

[0035] In S1, the data preprocessing process involves normalizing the filtered weather radar echo data and converting it into a grayscale image. At the same time, grayscale images with an effective reflectivity factor value ratio less than the threshold are filtered out and cut into images with a resolution of 128×128.

[0036] The pixel value range of a grayscale image is [0, 255].

[0037] Grayscale images with an effective reflectance factor value less than a threshold of 0.1 are filtered out. The effective reflectance factor value is 20 dBz.

[0038] The recurrent neural network model composed of ST-LSTM units in S2 has two types of memory units: time memory units. It is a layer within each ST-LSTM layer that transitions from the previous node t-1 to the current node at the same time; spatial memory unit Vertical transition from the lower l-1 ST-LSTM layer to the current layer at the same time point; especially for the bottom ST-LSTM with l=1, Assign to Because memory conversion functions for different propagation directions should be controlled by different signals, It adopts the original gate structure of standard LSTM and provides... Construct another set of input gates, forget gates, and input modulation gates.

[0039] The memory decoupling module in S2 is represented as follows:

[0040]

[0041]

[0042]

[0043] Among them, W d This represents a 1×1 convolution shared by all network layers. <·,·> represent dot products, and ‖·‖ represents the planarization of the feature map. Norm.

[0044] In S3, the grayscale image of the weather radar echo is converted into a time series. Each time series is 20 frames long, with a 6-minute interval between each frame. 10 frames are input and 10 frames are output.

[0045] In each ST-LSTM cell, we have:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053]

[0054]

[0055] Where C represents the temporal memory unit, M represents the spatial memory unit, H represents the hidden state, and x represents the input sequence. i represents the temporal input gate, f represents the temporal forget gate, and g represents the temporal input modulation gate; i′ represents the spatial input gate, f′ represents the spatial forget gate, and g′ represents the spatial input modulation gate. The subscript represents the step size, and the superscript represents the specific layer number in the network.

[0056] Finally, the hidden states between each node are analyzed. Depending on the combination of horizontal and zigzag memory states, and Connect them together and reduce the dimensionality using a 1×1 convolutional layer.

[0057]

[0058] The overall memory decoupling module can be represented as follows:

[0059]

[0060]

[0061]

[0062] Among them W d This represents a 1×1 convolution shared by all network layers. <·,·> represent dot products, and ‖·‖ represents the planarization of the feature map. Norms. By defining the decoupling loss using cosine similarity, the advantages of both C and M in long-term and short-term dynamic modeling are leveraged.

[0063] Figure 2 This is a flowchart of the meteorological radar data preprocessing process. To ensure the quality of the sample data, quality control of the radar data is required, including filtering out isolated noise points, smoothing filtering, and constructing a high-pass filter to filter out reflectivity factor values ​​unrelated to precipitation. Meteorological radar data is usually stored in polar coordinates, with data points denser closer to the radar and sparser farther away, resulting in uneven spatial distribution. To facilitate subsequent data processing, the radar data needs to be gridded. Data from each elevation layer is selected, and their maximum reflectivity is projected onto Cartesian grid points to obtain a combined reflectivity image. For accurate thunderstorm prediction, due to the complex atmospheric physics, echoes often have non-rigid shapes and may move, accumulate, or dissipate rapidly, making it crucial to learn dynamic features within a unified spatiotemporal feature space. Before generating the time series, the combined reflectivity factor values ​​need to be uniformly mapped to pixel values ​​and cropped into a grayscale image with a resolution of 128×128. The pixel value mapping formula is as follows:

[0064]

[0065] Where x is the reflectivity factor value of the grid point combination, here x min x max The values ​​were set to 0dBz and 80dBz, respectively. Echoes with excessively small areas are generally weaker and evolve faster, having a smaller positive impact on the prediction results. Therefore, during the dataset construction process, the cropped grayscale images need further filtering, retaining only images with a valid pixel value ratio exceeding 10%. Subsequently, a sliding window was used to slice the continuous sequence of radar echo images. Here, "continuous" is defined as continuous scanning radar volume scan data at 6-minute intervals, resulting in a time series.

[0066] like Figure 3As shown, the network model proposed in this invention consists of four layers of ST-LSTM units. Compared to single-layer units, the stacked structure increases the network depth, enabling it to better capture the temporal features of thunderstorms. The specific structure of the ST-LSTM unit is as follows: Figure 4 As shown, a single ST-LSTM unit consists of a traditional LSTM gating unit C and a spatial memory unit M. C is horizontally passed from the previous node to the current node within each layer, and M is vertically passed from lower layers to the current layer at the same time point. Finally, C and M are concatenated and dimensionality is reduced using a 1×1 convolutional layer to obtain the hidden state H of each node. The use of the 1×1 convolutional layer ensures that H and the memory state have the same dimension.

[0067] To verify the effectiveness of the thunderstorm prediction method based on spatiotemporal memory decoupled RNN, experiments were conducted on the collected thunderstorm dataset, and the method was compared with the ST-LSTM network without memory decoupling. Radar echo thresholds of 20dBz, 30dBz, and 40dBz were set, and three thunderstorm prediction scoring factors were used to evaluate the effect: Critical Success Index (CSI), Probability of Detection Success (POD), and False Alarm Ratio (FAR).

[0068] Table 1 presents the results of the thunderstorm scoring factor index on the test set for two networks with and without memory decoupling structures. The test results are given for different time steps of 6 minutes, 30 minutes, and 60 minutes, with thresholds of 30dBz, 40dBz, and 50dBz.

[0069] The data comparison in Table 1 shows that, under the same testing conditions, the network with the memory decoupling structure outperforms the conventional ST-LSTM, with an average improvement of 0.05–0.1 in each prediction score factor. As the forecast time progresses, both networks exhibit a trend where the critical success index and prediction success probability decrease over time, while the false alarm rate increases. Furthermore, Table 1 shows the score factor values ​​for the same forecast time, indicating that as the reflectivity threshold increases, the critical success index and prediction success probability decrease, while the false alarm rate increases. This is because as the threshold increases, it focuses more on the evolution of strong echo regions, but strong echo regions change relatively faster and have greater uncertainty, leading to a decrease in prediction factors. Considering that even with relatively large echo intensities, areas above 50 dBz in thunderstorms are concentrated within the thunderstorm body and constitute a small proportion of the total area, the prediction results of this network still have practical value.

[0070] Table 1. Test results of different network test sets

[0071]

[0072] Figure 5 This is an example of an extrapolated image sequence for thunderstorm prediction based on a spatiotemporal memory decoupled RNN. Figure 5 It can be seen that the obtained thunderstorm prediction image is very similar to the real image in terms of echo location distribution, intensity, and trend of change. It correctly predicts the strong echo regions located in the echo contour and interior, as well as the trend of the entire echo region moving eastward. This shows that the thunderstorm prediction method based on spatiotemporal memory decoupled RNN has high accuracy.

[0073] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A thunderstorm prediction method based on spatiotemporal memory decoupled RNN, characterized in that, Includes the following steps: S1. Acquire meteorological radar echo data, preprocess the meteorological radar echo data, and obtain radar echo grayscale image. S2. Based on the spatiotemporal-long short-term memory (ST-LSTM) network unit, construct a recurrent neural network model with a memory decoupling module; the memory decoupling module is represented as: ； ； ； in, A 1×1 convolution shared by all network layers. For dot product, For planar feature maps Norm; S3. Convert the weather radar echo grayscale image into a time series and divide it into a training set and a test set. Substitute the training set into the recurrent neural network with the memory decoupling module in S2 for training, and use the test set for classification and recognition verification during the training process. Stop training when the training number reaches the preset number. Use the trained network to predict thunderstorms on the test set to obtain the short-term near-term thunderstorm prediction image time series.

2. The thunderstorm prediction method based on spatiotemporal memory decoupled RNN according to claim 1, characterized in that, The meteorological radar echo data in S1 includes: isolated noise point filtering, smoothing filtering, and constructing a high-pass filter to filter out reflectivity factor values ​​that are unrelated to precipitation.

3. The thunderstorm prediction method based on spatiotemporal memory decoupled RNN according to claim 1, characterized in that, The data preprocessing process in S1 involves normalizing the filtered weather radar echo data and converting it into a grayscale image. At the same time, grayscale images with an effective reflectivity factor value ratio less than a threshold are filtered out and cut into images with a resolution of 128×128.

4. The thunderstorm prediction method based on spatiotemporal memory decoupled RNN according to claim 3, characterized in that, The pixel value range of the grayscale image is [0, 255].

5. A thunderstorm prediction method based on spatiotemporal memory decoupled RNN according to claim 3, characterized in that, The grayscale images that are filtered out if the percentage of effective reflectance factor values ​​is less than a threshold of 0.1 are excluded. The effective reflectance factor value is 20 dBz.

6. The thunderstorm prediction method based on spatiotemporal memory decoupled RNN according to claim 1, characterized in that, The recurrent neural network model composed of ST-LSTM units in S2 has two types of memory units: temporal memory units and spatial memory units, wherein: temporal memory units It is a layer within each ST-LSTM layer that transitions from the previous node t-1 to the current node at the same time; spatial memory unit Vertical transition from the lower l-1 ST-LSTM layer to the current layer at the same time point; especially for the bottom ST-LSTM with l=1, Assign to , It adopts the original gate structure of standard LSTM and provides... Construct another set of input gates, forget gates, and input modulation gates.

7. The thunderstorm prediction method based on spatiotemporal memory decoupled RNN according to claim 1, characterized in that, In S3, the weather radar echo grayscale image is converted into a time series. Each time series is 20 frames long, with a 6-minute interval between each frame. 10 frames are input and 10 frames are output.

Images