Rainfall nowcasting method based on deep learning

Abstract

The invention discloses a precipitation nowcasting method based on deep learning, and the method comprises the following steps: collecting historical radar echo data, carrying out the preprocessing of a collected radar echo image sequence, and constructing a time-space sequence data set; constructing a dynamic flow space-time generative adversarial network model which comprises a generator and a discriminator; designing a discriminator based on a Transform architecture, performing authenticity evaluation on the generated rainfall forecast sequence by using a self-attention mechanism, and evaluating the time-space consistency and visual quality of the generated sequence; a two-stage training strategy is adopted to optimize the network, in the first stage, reconstruction loss and decoupling loss are used to pre-train a generator, in the second stage, adversarial loss and feature matching loss are introduced to carry out joint training, and high-quality rainfall nowcasting is achieved; according to the invention, the modeling capability of the model for complex meteorological phenomena is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of weather forecasting, and in particular to a deep learning-based precipitation nowcasting method. BACKGROUND

[0002] Traditional precipitation forecasting methods mainly rely on radar echo extrapolation and numerical weather prediction (NWP) models. These methods have difficulties in dealing with complex atmospheric dynamics and limited spatial resolution, resulting in poor performance during severe weather events.

[0003] In recent years, deep learning-based methods have shown significant potential in addressing these challenges. ConvLSTM combines convolutional operations with a recurrent framework, better capturing the spatiotemporal features in radar echo sequences. PredRNN introduces ST-LSTM units to model spatial appearance and temporal changes simultaneously through additional memory transformations. However, existing models face two important limitations: they cannot adaptively model complex motion patterns with different regional variation trajectories. The fixed geometric sampling patterns of traditional convolutional operations limit their ability to capture the common non-rigid transformations in weather systems. The increasingly blurred prediction results are produced during long-term prediction processes. Regression-based training objectives prioritize average performance indicators rather than the perceptual quality of generated predictions. These limitations severely limit the practicality of current models in precipitation nowcasting applications. SUMMARY

[0004] The purpose of the present application is to provide a deep learning-based precipitation nowcasting method, which realizes accurate modeling of complex precipitation systems and high-quality prediction through a dynamic flow feature extraction mechanism and an adversarial learning framework, to solve the problems in the background art.

[0005] The technical scheme of the present application is as follows: (1) Collect historical radar echo data, preprocess the collected radar echo image sequence, and construct a spatiotemporal sequence dataset; (2) Construct a dynamic flow spatiotemporal generative adversarial network model, including a generator and a discriminator, wherein the generator adopts an encoder-prediction module-decoder architecture, and the prediction module is composed of multiple ST-LSTM++ units; adopt a hierarchical design strategy, design the dynamic flow feature extraction mechanism as a general component and a time sequence special component: the general dynamic flow component is applied to the encoder-decoder part for spatial adaptive sampling of static feature maps, and the time sequence special dynamic flow component is integrated in the ST-LSTM++ unit to predict the motion trend based on the previous time hidden state, realizing dynamic modeling of spatiotemporal coupling; (3) Design a discriminator based on the Transformer architecture and use the self-attention mechanism to evaluate the authenticity of the generated precipitation forecast sequence and assess the spatiotemporal consistency and visual quality of the generated sequence; (4) A two-stage training strategy is adopted to optimize the network. In the first stage, the generator is pre-trained using reconstruction loss and decoupling loss. In the second stage, adversarial loss and feature matching loss are introduced for joint training to achieve high-quality precipitation nowcasting.

[0006] Furthermore, the preprocessing in step (1) includes data cleaning, normalization, and basic feature extraction. The basic feature extraction uses traditional image processing methods to extract basic features from the original radar echo data, including using statistical methods to calculate the mean, variance, and maximum value of the echo intensity, using Gabor filters to extract texture features of different directions and scales, calculating geometric features such as the area, perimeter, and compactness of the echo region through morphological analysis, and constructing temporal features such as the intensity change rate and displacement vector between adjacent time points, providing rich initial feature representations for subsequent deep learning networks.

[0007] Furthermore, in step (2), the generator is responsible for generating future predictions based on historical radar sequences, and the discriminator is responsible for evaluating the authenticity of the generated sequences.

[0008] Furthermore, in step (2), the general dynamic flow feature extraction mechanism is as follows: based on the input feature map, it is processed through a two-layer convolutional network. Generate displacement field : ; in, For displacement field, Network for generating displacement fields, For the input feature map, The parameters for the displacement field generation network; for each location in the feature map Calculate the dynamic sampling position: ; in, For the first Each sampling layer at location Dynamic sampling coordinates, , This is the spatial coordinate index of the feature map. For sampling layer index, For the displacement field in the th Layer position Offset at; Extracting features at dynamic sampling locations using bilinear interpolation: ; in, For the first Each sampling layer at location Feature values ​​extracted at the location, It is a bilinear interpolation function; Through feature aggregation network Fusion of multi-scale features: ; in, For position The final output features, For feature aggregation networks, arrive for Sampling features at different scales, This represents the total number of sampling locations. These are the parameters of the feature aggregation network.

[0009] Furthermore, in step (2), the time-specific dynamic flow component is integrated into the ST-LSTM++ unit. Based on the general component, it uses time dimension information to guide spatial sampling, including dynamic flow mechanism and memory decoupling mechanism. The specific algorithm is as follows: Time-specific dynamic flow mechanism: Generates a displacement field based on the hidden state of the previous time step. ; in, Let be the displacement field at time t. Network for generating displacement fields, For the first The hidden state at any given moment. Parameters for the displacement field generation network; Dynamically sample the hidden state: ; in, For the first after dynamic sampling Always in a hidden state. For dynamic sampling functions; Memory decoupling mechanism: maintaining two independent memory streams, cell state And memory state Mt: Candidate memory values: ; in, For the first Candidate memory values ​​at time point The hyperbolic tangent activation function is used. The weight matrix for candidate memory values. For the first dynamically sampled Always in a hidden state. For the first Input at any time Bias term for candidate memory values; Cell status update: ; in, For the first Cellular state at any given moment For the first The output of the forget gate at any moment, This represents element-level multiplication. For the first Cellular state at any given moment For the first Input gate output at any time, This is the weight matrix for cell states. This is a bias term for the cell state; Memory status update: ; in, For the first The state of memory at any given moment For the first The state of memory at any given moment; Gating unit calculation: ; ; in, It is the sigmoid activation function. Here is the weight matrix of the input gate. For the bias term of the input gate, Here is the weight matrix for the forget gate. For the bias term of the forget gate; Decoupling loss calculation: ; in, For decoupling loss, The total number of time steps. For summation, The cosine similarity function is used. For gradient operators, For the gradient of cell states, This represents the gradient of the memory state.

[0010] Furthermore, in step (4), the specific implementation of the two-stage training strategy is as follows: Phase 1: Only the generator is pre-trained, while the discriminator parameters remain frozen. The pre-training loss function is: ; in, This represents the total loss during the pre-training phase. The weighting coefficients for reconstructing the loss. To reconstruct the loss, These are the weighting coefficients for the decoupling loss. This is the decoupling loss; Reconstruction loss calculation: ; in, The total number of samples, For the first The true value of each sample For the first The predicted value for each sample, Represents the L2 norm; Phase 2: Adversarial training is performed on the generator and discriminator, alternately updating the parameters of the two networks. The adversarial training loss function is: Discriminator loss: ; in, For discriminator loss, For mathematical expectation, For real data, For the true data distribution, For the discriminator network, For generator networks, Input data; Generator loss: ; in, For generator loss; Feature matching loss: ; in, For feature matching loss, The number of layers in the discriminator network. For the first The number of features in a layer Indicates the discriminator's first... Layer feature representation, Represents the L1 norm; Generator total loss function: ; in, Let be the total loss function of the generator. Weighting coefficients to counteract losses, These are the weight coefficients for the feature matching loss.

[0011] The precipitation nowcasting system based on deep learning comprises: A data acquisition module is configured to acquire radar echo data in real time and perform preprocessing; A model training module is configured to construct and train a dynamic flow spatiotemporal generative adversarial network model; A prediction output module is configured to generate future precipitation prediction results and perform visual display; An evaluation analysis module is configured to calculate prediction performance indicators and perform error analysis.

[0012] The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of any of the methods when executing the program.

[0013] The computer readable storage medium stores a computer program, and the program is executed by the processor to implement the steps of any of the methods.

[0014] Advantages: Compared with the prior art, the present application has the following significant advantages: adaptive motion modeling: the dynamic flow mechanism adaptively adjusts the receptive field according to the underlying motion pattern, and can accurately track the change trajectory of the complex precipitation system. High-quality prediction: the adversarial learning framework generates sharp and real prediction results, and can maintain detailed information even in long-time prediction. Time consistency: the global receptive field of the Transformer discriminator can evaluate the continuity of the entire sequence and detect subtle inconsistencies in the motion pattern. Strong robustness: the memory decoupling mechanism effectively separates the dynamic characteristics of different time scales, improving the modeling ability of the model for complex meteorological phenomena. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is the overall architecture diagram of the DFST-GAN of the present application; Figure 2 is a schematic diagram of the dynamic sampling process of the present application; Figure 3 is the ST-LSTM++ unit architecture diagram of the present application; Figure 4 is the training process flowchart of the present application. DETAILED DESCRIPTION

[0016] The technical solutions of the present application will be further described below in combination with the drawings.

[0017] As Figure 1As shown, the embodiment of the present application provides a precipitation nowcasting method based on deep learning. Firstly, in order to solve the technical problem that the existing precipitation prediction method cannot adaptively model the complex motion mode and the prediction result is fuzzy, a dynamic flow feature extraction mechanism is constructed. By generating a displacement field and calculating a dynamic sampling position, the convolution sampling position is adaptively adjusted according to the motion mode of the precipitation system, the limitation that the traditional fixed convolution kernel cannot capture the non-rigid transformation is overcome, and the accurate tracking of the complex precipitation system trajectory is realized. Then, on the basis of the dynamic flow mechanism, an ST-LSTM++ unit is designed, which integrates the dynamic flow mechanism and the memory decoupling technology. By maintaining independent cell states and memory states , and introducing candidate memory values, the short-term and long-term temporal dependencies are effectively separated, and the orthogonality of the two memory flows is promoted by using the decoupling loss, which significantly enhances the modeling ability and prediction accuracy of the model for complex spatio-temporal dynamics. Finally, in order to solve the problem of fuzzy prediction generated by the traditional regression method and improve the visual quality of precipitation prediction, the present application introduces a generative adversarial learning framework based on the Transformer architecture. The global receptive field of the discriminator is used to evaluate the spatio-temporal consistency of the generated sequence, and a two-stage training strategy is adopted. In the first stage, only the generator is pre-trained to establish a stable foundation, and in the second stage, the generated high-quality and detailed precipitation prediction results are realized by jointly optimizing the adversarial loss and feature matching loss , which realizes the accurate analysis of the historical radar echo sequence and the high-fidelity prediction of the future precipitation, and provides reliable technical support for meteorological business; including the following steps: Step 1: implementation of dynamic flow feature extraction mechanism, as Figure 2 shown. The dynamic flow feature extractor is the core innovation of the present application. The specific implementation steps are as follows: Displacement field generation: using a two-layer convolutional network to process the input feature map , output the displacement field . The network contains a 3x3 convolution kernel, and the activation function is ReLU.

[0018] As shown in Figure 2 , dynamic sampling: for each position in the feature map, the new sampling position is calculated according to the displacement field: ;

[0019] Feature sampling is performed at non-integer positions using bilinear interpolation: ;

[0020] wherein denotes the sampling layer index, is the The feature values extracted at the positions of the sampling layers. .

[0021] Feature aggregation: aggregate the features from multiple sampling positions through a network . .

[0022] where is the total number of sampling positions, is the parameter of the feature aggregation network.

[0023] Step 2: Construction of ST-LSTM++ unit, as shown in Figure 3 , the ST-LSTM++ unit integrates the dynamic flow mechanism and the memory decoupling mechanism on the basis of the standard ST-LSTM: Dynamic flow integration: First, generate the displacement field from the hidden state at the previous time: .

[0024] Then dynamically sample the hidden state: .

[0025] Gate mechanism: Calculate the input gate and the forget gate: .

[0026] .

[0027] where is the sigmoid activation function, , is the weight matrix, , is the bias term.

[0028] Memory decoupling: Maintain two independent memory flows: cell state and memory state .

[0029] First, calculate the candidate memory value: .

[0030] Then update the two memory states: .

[0031] .

[0032] Promote the orthogonality of the two memory flows through the decoupling loss: ;

[0033] Output computation: Output gate: ;

[0034] Hidden state output:

[0035] Step 3: Implementation of the Transformer discriminator The Transformer discriminator adopts the Vision Transformer architecture: Image patch processing: The 480x480 radar image is divided into 16x16 image patches, each of which is linearly projected into a 768-dimensional embedding.

[0036] Position encoding: Add a learnable position embedding to maintain spatial information.

[0037] Transformer layer: Use a 6-layer Transformer encoder, each layer containing: multi-head self-attention mechanism (8 attention heads), layer normalization, MLP feedforward network (hidden layer dimension 3072) Classification output: Calculate the authenticity probability by the final representation of the classification token.

[0038] Step 4: As shown in Figure 4 , the specific implementation of the training process is as follows:

[0039] Data preprocessing: Input: 10 frames of historical radar images (480x480x1); Output: 20 frames of future radar images; Data normalized to the range [0,1].

[0040] First stage (generator pre-training): Training object: only the generator network; Discriminator state: parameter frozen, not involved in training; Learning rate: ; batch size: 8; optimizer: Adam; loss function: ; loss weight: =1.0, =0.1; training rounds: 20000; establish a stable generator.

[0041] Second stage (adversarial training): Training object: generator and discriminator; Training method: alternately update the parameters of the two networks; Learning rate: generator , discriminator ; discriminator loss: ; generator total loss: ; loss weight: =1.0, =0.1, =0.01, =10.0; training rounds: 80000 to improve the quality and authenticity of generation.

[0042] Scheduled sampling: gradually increase the probability of using model prediction as the input for the next time during training, linearly from 0.0 to 0.8.

[0043] Step 5: Model evaluation: evaluate model performance using the following indicators: regression indicators: mean square error (MSE); structural similarity index (SSIM); meteorological indicators: critical success index (CSI); Heidke skill score (HSS); detection probability (POD); false alarm rate (FAR); calculate the above indicators at different precipitation intensity thresholds (dBZ≥30, 40, 50).

[0044] Step 6: System deployment: the invention can be deployed as a real-time precipitation forecasting system: hardware requirements: GPU: NVIDIA A100 or equivalent performance; Memory: 32GB or more; Storage: SSD 1TB or more; Software environment: Operating system: Ubuntu 20.04; Deep learning framework: PyTorch 1.12; Python 3.8; Real-time forecasting process: receive real-time radar data; data preprocessing and quality control; model inference to generate forecasts; result post-processing and visualization; distribution of forecast products.

Claims

1. A deep learning-based precipitation nowcasting method, characterized in that, Includes the following steps: (1) Collect historical radar echo data, preprocess the collected radar echo image sequences, and construct a spatiotemporal sequence dataset; (2) Construct a dynamic flow spatiotemporal generative adversarial network model, including a generator and a discriminator. The generator adopts an encoder-prediction module-decoder architecture, and the prediction module is composed of multiple ST-LSTM++ units. A hierarchical design strategy is adopted to design the dynamic flow feature extraction mechanism as a general component and a time-specific component: the general dynamic flow component is applied to the encoder-decoder part to perform spatial adaptive sampling of the static feature map, and the time-specific dynamic flow component is integrated in the ST-LSTM++ unit to predict the motion trend based on the hidden state of the previous time step, so as to realize the dynamic modeling of spatiotemporal coupling. (3) Design a discriminator based on the Transformer architecture and use the self-attention mechanism to evaluate the authenticity of the generated precipitation forecast sequence and assess the spatiotemporal consistency and visual quality of the generated sequence; (4) A two-stage training strategy is adopted to optimize the network. In the first stage, the generator is pre-trained using reconstruction loss and decoupling loss. In the second stage, adversarial loss and feature matching loss are introduced for joint training to achieve precipitation nowcasting.

2. The precipitation nowcasting method based on deep learning according to claim 1, characterized in that, The preprocessing in step (1) includes data cleaning, normalization and basic feature extraction. The basic feature extraction uses traditional image processing methods to extract basic features from the original radar echo data, including using statistical methods to calculate the mean, variance and maximum value of the echo intensity, using Gabor filters to extract texture features of different directions and scales, calculating the area, perimeter and compactness of the echo region through morphological analysis, and constructing the intensity change rate and displacement vector between adjacent time points.

3. The precipitation nowcasting method based on deep learning according to claim 1, characterized in that, In step (2), the generator is responsible for generating future predictions based on historical radar sequences, and the discriminator is responsible for evaluating the authenticity of the generated sequences.

4. The precipitation nowcasting method based on deep learning according to claim 1, characterized in that, The specific implementation of the general dynamic flow component in step (2) is as follows: based on the input feature map, it is processed through a two-layer convolutional network. Generate displacement field : ； in, For displacement field, Network for generating displacement fields, For the input feature map, The parameters for the displacement field generation network; for each location in the feature map Calculate the dynamic sampling position: ； in, For the first Each sampling layer at location Dynamic sampling coordinates, , This is the spatial coordinate index of the feature map. For sampling layer index, For the displacement field in the th Layer position Offset at; Extracting features at dynamic sampling locations using bilinear interpolation: ； in, For the first Each sampling layer at location Feature values ​​extracted at the location, It is a bilinear interpolation function; Through feature aggregation network Fusion of multi-scale features: ； in, For position The final output features, For feature aggregation networks, arrive for Sampling features at different scales, This represents the total number of sampling locations. These are the parameters of the feature aggregation network.

5. The deep learning-based precipitation nowcasting method according to claim 4, characterized in that, In step (2), the time-specific dynamic flow component is integrated into the ST-LSTM++ unit. Based on the general component, it uses time dimension information to guide spatial sampling, including dynamic flow mechanism and memory decoupling mechanism. The specific algorithm is as follows: Time-specific dynamic flow mechanism: Generates a displacement field based on the hidden state of the previous time step. ； in, Let be the displacement field at time t. Network for generating displacement fields, For the first The hidden state at any given moment. Parameters for the displacement field generation network; Dynamically sample the hidden state: ； in, For the first after dynamic sampling Always in a hidden state. For dynamic sampling functions; Memory decoupling mechanism: maintaining two independent memory streams, cell state And memory state Mt: Candidate memory values: ； in, For the first Candidate memory values ​​at time point The hyperbolic tangent activation function is used. The weight matrix for candidate memory values. For the first dynamically sampled Always in a hidden state. For the first Input at any time Bias term for candidate memory values; Cell status update: ； in, For the first Cellular state at any given moment For the first The output of the forget gate at any moment, This represents element-level multiplication. For the first Cellular state at any given moment For the first Input gate output at any time, This is the weight matrix for cell states. This is a bias term for the cell state; Memory status update: ； in, For the first The state of memory at any given moment For the first The state of memory at any given moment; Gating unit calculation: ； ； in, It is the sigmoid activation function. Here is the weight matrix of the input gate. For the bias term of the input gate, Here is the weight matrix for the forget gate. For the bias term of the forget gate; Decoupling loss calculation: ； in, For decoupling loss, The total number of time steps. For summation, The cosine similarity function is used. For gradient operators, For the gradient of cell states, This represents the gradient of the memory state.

6. The deep learning-based precipitation nowcasting method according to claim 5, characterized in that, In step (4), the specific implementation of the two-stage training strategy is as follows: Phase 1: Only the generator is pre-trained, while the discriminator parameters remain frozen. The pre-training loss function is: ； in, This represents the total loss during the pre-training phase. The weighting coefficients for reconstructing the loss. To reconstruct the loss, These are the weighting coefficients for the decoupling loss. This is the decoupling loss; Reconstruction loss calculation: ； in, The total number of samples, For the first The true value of each sample For the first The predicted value for each sample, Represents the L2 norm; Phase 2: Adversarial training is performed on the generator and discriminator, alternately updating the parameters of the two networks. The adversarial training loss function is: Discriminator loss: ； in, For discriminator loss, For mathematical expectation, For real data, For the true data distribution, For the discriminator network, For generator networks, Input data; Generator loss: ； in, For generator loss; Feature matching loss: ； in, For feature matching loss, The number of layers in the discriminator network. For the first The number of features in a layer Indicates the discriminator's first... Layer feature representation, Represents the L1 norm; Generator total loss function: ； in, Let be the total loss function of the generator. Weighting coefficients to counteract losses, These are the weight coefficients for the feature matching loss.

7. A precipitation nowcasting system based on deep learning, characterized in that, include: The data acquisition module is used to acquire radar echo data in real time and perform preprocessing. The model training module is used to build and train dynamic flow spatiotemporal generative adversarial network models; The prediction output module is used to generate and visualize future precipitation forecasts. The evaluation and analysis module is used to calculate forecast performance indicators and perform error analysis.

8. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the program to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images