A method for constructing a neural network-based local short-term precipitation forecast model

Abstract

This invention discloses a method for constructing a local short-term precipitation forecast model based on a neural network, relating to the field of precipitation prediction technology. By utilizing radar echo sequences, multi-source environmental field data, and surface precipitation observation data of the target area, a spatiotemporally consistent training sample pair is constructed. A two-branch neural network model is established, including an encoder-evolution branch, a modulation branch, and a gated fusion decoder. With the goal of minimizing the composite loss function, the neural network model is trained using the training samples to obtain a deterministic forecast model. Using the output of the deterministic forecast model and the current environmental field as conditions, a conditional generation network is established to generate several probabilistic precipitation forecast fields, forming a probabilistic forecast set for quantifying forecast uncertainty. The trained deterministic forecast model and the conditional generation network are integrated and deployed, inputting real-time meteorological data to output deterministic and probabilistic local short-term precipitation forecasts.

Description

Technical Field

[0001] This invention relates to the field of precipitation prediction technology, and in particular to a method for constructing a local short-term precipitation forecast model based on neural networks. Background Technology

[0002] Short-term precipitation forecasting (SMF) typically refers to refined forecasts of the time, location, and intensity of precipitation occurring within the next 0-6 hours, particularly 0-2 hours. It is a key supporting technology in meteorological disaster prevention and mitigation, urban operation management, and major event support. With socio-economic development, unprecedented demands have been placed on the spatiotemporal accuracy and reliability of forecasts. Traditional SMF methods are mainly divided into three categories: extrapolation-based methods, numerical weather prediction-based methods, and methods based on the statistical relationship between radar echoes and precipitation. All of these face significant technical bottlenecks in practical applications. In recent years, artificial intelligence technologies, represented by deep learning, have brought new opportunities to SMF. Models based on convolutional neural networks (CNNs) and recurrent neural networks (RNNs) and their variants (such as ConvLSTM and TrajGRU) can learn complex spatiotemporal evolution patterns from historical radar sequences end-to-end, surpassing traditional extrapolation methods in forecast accuracy.

[0003] However, existing deep learning-based short-term forecast models still suffer from the following critical problems that urgently need to be addressed: First, most existing models are purely "data-driven" black boxes, with their learning objective typically limited to minimizing pixel-level prediction errors (such as mean squared error). This leads to models tending to output smooth forecast fields, weakening or omitting extreme heavy precipitation events. Furthermore, the meteorological fields they generate (such as precipitation fields) may violate basic meteorological physics laws, reducing the physical reliability and operational usability of the forecast results. Second, most existing models rely solely on radar echo sequences as a single input, neglecting the crucial modulating effects of key environmental factors such as wind, humidity, temperature, and topography on precipitation evolution. This results in insufficient forecasting capabilities for scenarios such as precipitation under complex terrain influences and new convection triggered by fronts. Finally, most existing deep learning models output single deterministic forecasts, failing to provide probabilistic information or ensemble forecast products that characterize forecast reliability, making it difficult to meet the demand for uncertainty information in risk warning and decision-making. Therefore, to address the aforementioned technical problems of existing models, a method for constructing a local short-term precipitation forecast model based on neural networks is proposed. Summary of the Invention

[0004] The main objective of this invention is to provide a method for constructing a local short-term precipitation forecasting model based on neural networks. By leveraging the powerful nonlinear fitting capabilities of deep learning, this method effectively incorporates prior meteorological and physical knowledge and comprehensively utilizes multi-source observation and model information to construct a next-generation intelligent local short-term precipitation forecasting model with higher accuracy, more reasonable physics, and richer information, which can effectively solve the problems in the background technology.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for constructing a local short-term precipitation forecasting model based on neural networks includes the following steps: S1: Acquire radar echo sequences, multi-source environmental field data and ground precipitation observation data of the target area, perform quality control, spatiotemporal alignment and gridding processing, and construct spatiotemporally consistent training sample pairs. The sample pairs include historical observation sequences and corresponding future precipitation field true values. S2: Construct a dual-branch neural network model, including an encoder-evolution branch that takes the radar echo sequence as input, a modulation branch that takes the multi-source environmental field at the current moment as input, and a gated fusion decoder; among which, The encoder-evolution branch is used to extract the spatiotemporal evolution features of the precipitation system; The modulation branch is used to extract multi-scale environmental features; The gated fusion decoder dynamically injects the environmental features extracted by the modulation branch into the spatiotemporal features of the evolution branch through a gated fusion mechanism, thereby generating a deterministic precipitation forecast field for future time periods. S3: Using the training sample pairs constructed in step S1, train the neural network model constructed in step S2 with the goal of minimizing the composite loss function to obtain a deterministic prediction model; the composite loss function includes a pixel-level error loss term, a physical constraint loss term, and a spectral domain matching loss term; S4: Construct and train a conditional generation network based on the output of the deterministic forecast model and the current environmental field, and use the conditional generation network to generate several probabilistic precipitation forecast fields to form a probabilistic forecast set for quantifying forecast uncertainty. S5: Integrate and deploy the trained deterministic forecast model with the conditional generation network, input real-time meteorological data, and output local short-term precipitation deterministic and probabilistic forecast products.

[0006] Furthermore, in step S2, the encoder-evolution branch adopts a Transformer architecture that includes a spatiotemporal self-attention mechanism, specifically including: Spatiotemporal position encoding is performed on the input radar echo sequence; Features are extracted by cascading spatiotemporal Transformer encoder blocks, each of which contains a spatiotemporal multi-head self-attention sublayer and a feedforward neural network sublayer. Output spatiotemporal evolution feature maps at different levels to the gated fusion decoder.

[0007] Furthermore, in step S2, the modulation branch uses a dilated convolutional pyramid structure to extract multi-scale environmental features, specifically including: Multiple dilated convolutional layers with different dilation rates are used in parallel to convolve the spliced ​​multi-source environmental field data to simultaneously capture local, mesoscale, and synoptic-scale environmental features. After upsampling the features obtained by global average pooling, the channels are concatenated with the outputs of each dilated convolutional layer; The concatenated features are fused and dimensionality reduced by a 1×1 convolutional layer to generate unified environmental modulation features.

[0008] Furthermore, in step S2, the gating fusion mechanism of the gated fusion decoder is implemented according to the following formula: ; ; in, The evolution branch features received by the i-th layer decoder are... For modulation branching characteristics, Indicates an upsampling operation. Indicates channel splicing. It is the Sigmoid activation function. Represents a 1×1 convolution. For the generated spatial adaptive gating weight graph, The projected environmental features This indicates element-wise multiplication. These are the features after fusion.

[0009] Furthermore, in step S3, the composite loss function The definition of is: ;in, A pixel-weighted Huber loss is used to strengthen the constraint on areas of heavy precipitation while balancing ordinary errors; Loss due to physical constraints; To achieve spectral matching loss, the spatial scale structure of the precipitation field is constrained by comparing the difference between the predicted field and the actual field in the Fourier space power spectrum. and To balance the hyperparameters.

[0010] Furthermore, the physical constraint loss Includes soft constraint terms related to mass conservation. and dynamic smoothing constraint terms ,in: The mass conservation soft constraint term Defined by the following formula: In the formula, This indicates summation of regions. This is the deterministic forecast output of the model, where M is the approximation term for the water vapor flux divergence in the lower atmosphere. It is a learnable scaling factor; The dynamic smoothing constraint term Defined by the following formula: In the formula, These represent the gradient operators in the spatial x-direction, y-direction, and time t-direction, respectively.

[0011] Furthermore, in step S4, the conditional generation network is a conditional generative adversarial network, which includes a generator G and a discriminator D. The generator G outputs the deterministic prediction model. The current environmental field E and the random noise vector z are taken as inputs, and the output is a probabilistic precipitation field. ; The input to the discriminator D is a real sample. Or generate samples The output is the probability that the input is real data; The conditional generative adversarial network minimizes the adversarial loss function between the generator and the discriminator. And simultaneously minimize the difference between the generated results and the actual precipitation field. Reconstruction loss Conduct joint training.

[0012] Furthermore, in step S5, the probability forecast product includes at least: the ensemble average forecast calculated based on the members of the probability forecast ensemble, the probability of precipitation occurring, the probability of precipitation exceeding a preset threshold, and the ensemble discreteness product characterizing the forecast uncertainty.

[0013] A system for constructing a local short-term precipitation forecasting model based on a neural network, the system being used to implement a method for constructing a local short-term precipitation forecasting model based on a neural network, comprising: The data preprocessing and sample construction module is used to acquire radar echo sequences, multi-source environmental field data and ground precipitation observation data of the target area, perform quality control, spatiotemporal alignment and gridding on the data, and construct spatiotemporally consistent training sample pairs. The sample pairs include historical observation sequences and corresponding future precipitation field ground values. A neural network model building module is used to construct a two-branch neural network model that integrates multi-scale features and physical constraints; the model includes: The encoder-evolution branch, with radar echo sequences as the main input, is used to extract the spatiotemporal evolution features of the precipitation system through an encoder containing a spatiotemporal self-attention mechanism. The modulation branch takes the multi-source environmental field at the current moment as input and is used to extract environmental field features through a multi-scale feature extractor. A gated fusion decoder, connected to the encoder-evolution branch and the modulation branch, is used to dynamically inject the environmental features extracted by the modulation branch into the spatiotemporal features of the evolution branch through a gated fusion mechanism to generate a deterministic precipitation forecast field for future time periods. The model training and optimization module includes: A deterministic model training unit is used to train the dual-branch neural network model using the training sample pairs, with the goal of minimizing the composite loss function, to obtain a deterministic prediction model; the composite loss function includes at least a pixel-level error loss term, a physical constraint loss term, and a spectral domain matching loss term; The probabilistic generation model training unit is used to construct and train a conditional generation network, which generates multiple probabilistic precipitation forecast fields based on the output of the deterministic forecast model and the current environmental field, forming a probabilistic forecast set for quantifying forecast uncertainty. The model deployment and inference module is used to integrate the deterministic forecast model and the conditional generation network, receive real-time meteorological data input, and output local short-term precipitation deterministic forecast and probabilistic forecast products.

[0014] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements a method for constructing a local short-term precipitation forecast model based on a neural network.

[0015] The present invention has the following beneficial effects: Compared with existing technologies, this scheme constructs an encoder-evolution branch, which enables the model to effectively capture the long-range spatiotemporal dependence and nonlinear evolution characteristics of radar echo cells during generation, movement, merging and dissipation. This overcomes the limitation of traditional optical flow methods, which can only perform linear extrapolation, and significantly improves the forecast accuracy within the critical forecast lead time of 0-2 hours. Compared with existing technologies, this solution establishes a composite loss function that includes pixel-level error loss, physical constraint loss, and spectral domain matching loss, assigns higher weight to areas of heavy precipitation, guides the model to optimize resource allocation, and significantly improves the forecasting ability for extreme heavy precipitation events. This effectively alleviates the "smoothing" problem of forecast fields commonly found in deep learning models and improves the hit rate of heavy precipitation. Compared with existing technologies, this solution adopts a gating fusion mechanism, which enables the model to dynamically and physically correct the trend of pure radar echo extrapolation based on the current environmental field (such as wind shear and water vapor conditions) and geographical information (such as terrain). This fundamentally improves the accuracy and rationality of precipitation forecasts under the influence of complex terrain (such as mountains and coastlines) and specific weather systems (such as fronts and squall lines). Compared with existing technologies, this solution incorporates simplified physical equations such as mass conservation and dynamic smoothing as soft constraints into the training process, making the model's learning process conform to basic meteorological dynamics principles. This makes the generated forecast field more physically consistent and reasonable, reduces forecast results that violate common sense physics, and increases the trust of operational forecasters. Compared with existing technologies, this solution, through the construction of a conditional generative adversarial network, enables the model to generate a set of probabilistic forecasts. Based on the members of the probabilistic forecast set, it further calculates products including the ensemble average forecast, the probability of precipitation, the probability of precipitation exceeding a preset threshold, and the ensemble discreteness product representing forecast uncertainty. This not only realizes the operational quantitative expression of forecast uncertainty, but also provides more comprehensive and scientific information support for downstream decision-making such as meteorological risk warning and emergency management. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the construction method of a local short-term precipitation forecasting model based on neural networks according to the present invention. Figure 2 This is a schematic diagram of the structure of a system for constructing a local short-term precipitation forecasting model based on a neural network according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] Example 1: join Figure 1 The flowchart shown is a method for constructing a local short-term precipitation forecasting model based on a neural network according to the present invention, including the following steps: S1: Acquire radar echo sequences, multi-source environmental field data and ground precipitation observation data of the target area, perform quality control, spatiotemporal alignment and gridding processing, and construct spatiotemporally consistent training sample pairs. The sample pairs include historical observation sequences and corresponding future precipitation field ground values. S2: Construct a two-branch neural network model, including: The encoder-evolution branch, which takes radar echo sequences as input, is used to extract the spatiotemporal evolution characteristics of precipitation systems. The modulation branch, which takes the multi-source environmental field at the current moment as input, is used to extract multi-scale environmental features; The gated fusion decoder dynamically injects the environmental features extracted from the modulation branch into the spatiotemporal features of the evolution branch through a gated fusion mechanism, generating a deterministic precipitation forecast field for future time periods. The encoder-evolution branch employs a Transformer architecture that includes a spatiotemporal self-attention mechanism, specifically comprising: Spatiotemporal position encoding is performed on the input radar echo sequence; Features are extracted by cascading spatiotemporal Transformer encoder blocks, each of which contains a spatiotemporal multi-head self-attention sublayer and a feedforward neural network sublayer. Output spatiotemporal evolution feature maps at different levels to the gated fusion decoder.

[0019] The modulation branch uses a dilated convolutional pyramid structure to extract multi-scale environmental features, specifically including: Multiple dilated convolutional layers with different dilation rates are used in parallel to convolve the spliced ​​multi-source environmental field data to simultaneously capture local, mesoscale, and synoptic-scale environmental features. After upsampling the features obtained by global average pooling, the channels are concatenated with the outputs of each dilated convolutional layer; The concatenated features are fused and dimensionality reduced by a 1×1 convolutional layer to generate unified environmental modulation features.

[0020] The gating fusion mechanism of the gated fusion decoder is implemented according to the following formula: ; ; in, The evolution branch features received by the i-th layer decoder are... For modulation branching characteristics, Indicates an upsampling operation. Indicates channel splicing. It is the Sigmoid activation function. Represents a 1×1 convolution. For the generated spatial adaptive gating weight graph, The projected environmental features This indicates element-wise multiplication. These are the features after fusion.

[0021] S3: Using the training sample pairs constructed in step S1, train the neural network model constructed in step S2 with the goal of minimizing the composite loss function to obtain a deterministic prediction model. The composite loss function includes a pixel-level error loss term, a physical constraint loss term, and a spectral domain matching loss term, and is defined as follows: ;in, A pixel-weighted Huber loss is used to strengthen the constraint on areas of heavy precipitation while balancing ordinary errors; Loss due to physical constraints; To achieve spectral matching loss, the spatial scale structure of the precipitation field is constrained by comparing the difference between the predicted field and the actual field in the Fourier space power spectrum. and To balance the hyperparameters.

[0022] Physical constraint loss Includes soft constraint terms related to mass conservation. and dynamic smoothing constraint terms ,in: Mass conservation soft constraint term Defined by the following formula: In the formula, This indicates summation of regions. This is the deterministic forecast output of the model, where M is the approximation term for the water vapor flux divergence in the lower atmosphere. It is a learnable scaling factor; Dynamic smoothing constraint Defined by the following formula: In the formula, These represent the gradient operators in the spatial x-direction, y-direction, and time t-direction, respectively.

[0023] S4: Construct and train a conditional generation network based on the output of the deterministic forecast model and the current environmental field. Use the conditional generation network to generate several probabilistic precipitation forecast fields to form a probabilistic forecast set for quantifying forecast uncertainty. The conditional generative network is a conditional generative adversarial network, which includes a generator G and a discriminator D. The generator G outputs the deterministic prediction model. The current environmental field E and the random noise vector z are taken as inputs, and the output is a probabilistic precipitation field. ; The input to discriminator D is the real sample. Or generate samples The output is the probability that the input is real data; Conditional generative adversarial networks minimize the adversarial loss function between the generator and the discriminator. And simultaneously minimize the difference between the generated results and the actual precipitation field. Reconstruction loss Conduct joint training.

[0024] S5: Integrate and deploy the trained deterministic forecast model with the conditional generation network, input real-time meteorological data, and output local short-term precipitation deterministic and probabilistic forecast products.

[0025] Among them, the probability forecast products include at least: the ensemble average forecast calculated based on the members of the probability forecast ensemble, the probability of precipitation, the probability of precipitation exceeding a preset threshold, and the ensemble discreteness product characterizing the forecast uncertainty.

[0026] The technical solution of the present invention will be further explained below with specific examples. Taking the Yangtze River Delta urban agglomeration as a demonstration area, a 0-2 hour short-term precipitation forecast model covering the region (118°-123° east longitude, 29°-33° north latitude) will be constructed.

[0027] 1. Data used Radar data: A mosaic reflectivity product composed of an S-band Doppler weather radar network in the Yangtze River Delta region, with a spatiotemporal resolution of 6 minutes and 1 kilometer. Historical data for three years, from 2020 to 2022, were collected.

[0028] Environmental field data: ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF), with a spatiotemporal resolution of 1 hour and 0.25°×0.25°. The U / V wind field, specific humidity, temperature, and sea level pressure field at 850hPa, 700hPa, and 500hPa were extracted.

[0029] Ground truth data: Hourly precipitation observation data from automatic ground stations of the China Meteorological Administration, which are fused with radar quantitative precipitation estimates using the optimal interpolation method to generate an hourly precipitation grid analysis field with a resolution of 1 km as training labels.

[0030] Static data: SRTM 90-meter resolution digital elevation model.

[0031] 2. Data Preprocessing Spatiotemporal grid unification: All the above data are unified to a fixed latitude and longitude grid through bilinear interpolation (range: 118°E-123°E, 29°N-33°N; number of grids: 500×400; resolution: 0.01°≈1km).

[0032] Sample construction: Input: Radar echo sequence Z[t-9min,t] from the past hour (10 frames) + ERA5 environmental field E_t and terrain elevation H at the current time t.

[0033] Tags: Grid analysis field of precipitation for the next 1 hour (10 frames) R[t+6min,t+60min].

[0034] Approximately 150,000 valid sample pairs were generated from 3 years of data using a sliding window (6-minute step).

[0035] Dataset partitioning: Data from 2020-2021 is the training set (approximately 100,000 samples), data from January to June 2022 is the validation set (approximately 30,000 samples), and data from July to December 2022 is the independent test set (approximately 20,000 samples).

[0036] 3. Implementation methods of neural network models The model is implemented using the PyTorch 1.12 framework.

[0037] include: 3.1) Encoder - Evolution Branch: The input radar sequence (10, 400, 500, 1) is first processed by a Conv3d(1, 64, kernel) size=(3,3,3) ,padding=1).

[0038] This is then followed by four cascaded spatiotemporal Transformer coding blocks. Within each block: The 3D feature maps are reshaped into sequences, and learnable spatiotemporal location encodings are added.

[0039] First, perform time-dimensional self-attention (treating spatial location as a batch dimension), then perform spatial-dimensional self-attention (treating time dimension as a batch dimension).

[0040] After layer normalization (LayerNorm) and feedforward network (FFN, two fully connected layers).

[0041] Output four spatiotemporal feature maps at different scales {F e1 ,F e2 ,F e3 ,F e4}

[0042] 3.2) Modulation branch: The input is the concatenated current environmental field and terrain field (400, 500, 14).

[0043] Pyramid convolution pooling with a void space: include: Four parallel dilated convolutional layers: Conv2d(14,32,kernel size=3 ,dilation=1 / 6 / 12 / 18,padding=same).

[0044] A global average pooling layer upsamples the features back to their original size.

[0045] By concatenating the above 5 outputs along the channel dimension, we obtain the feature (400, 500, 160).

[0046] Through a Conv2d(160,128,kernel) size=1 The environmental modulation features F are obtained by fusion and dimensionality reduction. m .

[0047] 3.3) Gated fusion decoder: Specifically, it is a typical U-Net-type decoding structure, which contains 4 upsampling layers.

[0048] A gated fusion unit is set up before each upsampling layer: The last decoding layer is followed by a Conv2d(64,10,kernel) size=1 Using the Softplus activation function, it outputs a deterministic precipitation forecast for the next 10 frames. .

[0049] 4. Determining the loss function The composite loss function designed in the above steps is adopted; Training parameters: The AdamW optimizer was used (with an initial learning rate of 5e-4), a batch size of 8, and training was performed for 100 epochs on four NVIDIA A100 GPUs. Cosine annealing was used for learning rate scheduling and gradient clipping.

[0050] 5. Probability Generation and Post-processing Network structure: Conditional U-Net is used as the generator, and PatchGAN is used as the discriminator.

[0051] Training: Freeze the pre-trained deterministic model. When training a conditional GAN, the generator input is: , where z is random Gaussian noise. The discriminator input is or .

[0052] Adversarial loss: Use least squares GAN (LSGAN) loss and add L1 reconstruction loss (weight 100).

[0053] Inference: After training, for the same set of inputs (Z, E) t The generator samples 20 times (injecting different noise z each time) to obtain 20 set members. .

[0054] 6. Business Deployment and Product Development Model solidification: Use TorchScript to script and export deterministic models and probability generators.

[0055] Real-time process: Every 5 minutes, the system automatically retrieves the latest radar mosaic and rapidly updated numerical model analysis field.

[0056] The data goes through a preprocessing pipeline that is exactly the same as during training.

[0057] The solidified model is used for reasoning to generate a deterministic precipitation intensity field with a resolution of 1 kilometer for the next hour, and 20 ensemble member fields every 5 minutes.

[0058] Product generation includes: Deterministic products: Directly output gridded precipitation intensity fields for use in creating forecast maps.

[0059] Probability product: Based on 20 members, calculate and output: Ensemble average: the best estimate after smoothing.

[0060] Probability of precipitation: The proportion of members whose precipitation is >0.1 mm / 5 min.

[0061] Probability of heavy precipitation: the proportion of members with precipitation >5 mm / h.

[0062] Ensemble quantiles: such as the 90th quantile, used for extreme precipitation warnings.

[0063] Performance Verification: During the independent testing period from July to December 2022, the model in this embodiment was compared with the business optical flow method and the benchmark deep learning model. The key scores are shown in the table below:

[0064] Note: The CSI score for the probability ensemble is calculated using the ensemble average. A lower Continuous Ranking Probability Score (CRPS) indicates better probability forecast performance.

[0065] 7. Conclusion This embodiment fully demonstrates the implementation process of the technical solution of the present invention. By introducing a gated fusion multi-branch network structure, a composite loss function containing physical and spectral constraints, and probabilistic post-processing based on conditional GAN, the constructed forecasting model significantly outperforms traditional methods and benchmark deep learning models in terms of forecast accuracy, physical rationality, and uncertainty quantification capability. It successfully achieves a closed loop from data to business, and has significant practical value and creativity.

[0066] Example 2: This invention also provides a system for constructing a neural network-based local short-term precipitation forecasting model to implement the method in Embodiment 1, see [link to documentation]. Figure 2 As shown, the system includes: The data preprocessing and sample construction module is used to acquire radar echo sequences, multi-source environmental field data and ground precipitation observation data of the target area, perform quality control, spatiotemporal alignment and gridding on the data, and construct spatiotemporally consistent training sample pairs. The sample pairs include historical observation sequences and corresponding future precipitation field ground values. The neural network model building module is used to construct a two-branch neural network model that integrates multi-scale features and physical constraints; the model includes: The encoder-evolution branch, with radar echo sequences as the main input, is used to extract the spatiotemporal evolution features of the precipitation system through an encoder containing a spatiotemporal self-attention mechanism. The modulation branch takes the multi-source environmental field at the current moment as input and is used to extract environmental field features through a multi-scale feature extractor. The gated fusion decoder, connected to the encoder-evolution branch and the modulation branch, is used to dynamically inject the environmental features extracted by the modulation branch into the spatiotemporal features of the evolution branch through the gated fusion mechanism, thereby generating a deterministic precipitation forecast field for future time periods. The model training and optimization module includes: The deterministic model training unit is used to train a two-branch neural network model using training sample pairs with the goal of minimizing the composite loss function, thereby obtaining a deterministic prediction model; the composite loss function includes at least a pixel-level error loss term, a physical constraint loss term, and a spectral domain matching loss term; The probabilistic generative model training unit is used to build and train a conditional generative network to generate multiple probabilistic precipitation forecast fields based on the output of the deterministic forecast model and the current environmental field, forming a probabilistic forecast set for quantifying forecast uncertainty. The model deployment and inference module is used to integrate deterministic forecast models and conditional generation networks, receive real-time meteorological data input, and output local short-term precipitation deterministic and probabilistic forecast products.

[0067] Example 3: The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for constructing a local short-term precipitation forecast model based on a neural network as described in Embodiment 1.

[0068] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a local short-term precipitation forecasting model based on neural networks, characterized in that, Includes the following steps: S1: Acquire radar echo sequences, multi-source environmental field data and ground precipitation observation data of the target area, perform quality control, spatiotemporal alignment and gridding processing, and construct spatiotemporally consistent training sample pairs. The sample pairs include historical observation sequences and corresponding future precipitation field true values. S2: Construct a dual-branch neural network model, including an encoder-evolution branch that takes the radar echo sequence as input, a modulation branch that takes the multi-source environmental field at the current moment as input, and a gated fusion decoder; among which, The encoder-evolution branch is used to extract the spatiotemporal evolution features of the precipitation system; The modulation branch is used to extract multi-scale environmental features; The gated fusion decoder dynamically injects the environmental features extracted by the modulation branch into the spatiotemporal features of the evolution branch through a gated fusion mechanism, thereby generating a deterministic precipitation forecast field for future time periods. S3: Using the training sample pairs constructed in step S1, train the neural network model constructed in step S2 with the goal of minimizing the composite loss function to obtain a deterministic prediction model. S4: Construct and train a conditional generation network based on the output of the deterministic forecast model and the current environmental field, and use the conditional generation network to generate several probabilistic precipitation forecast fields to form a probabilistic forecast set for quantifying forecast uncertainty. S5: Integrate and deploy the trained deterministic forecast model with the conditional generation network, input real-time meteorological data, and output local short-term precipitation deterministic and probabilistic forecast products.

2. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 1, characterized in that, In step S2, the encoder-evolution branch adopts a Transformer architecture that includes a spatiotemporal self-attention mechanism, specifically including: Spatiotemporal position encoding is performed on the input radar echo sequence; Features are extracted by cascading spatiotemporal Transformer encoder blocks, each of which contains a spatiotemporal multi-head self-attention sublayer and a feedforward neural network sublayer. Output spatiotemporal evolution feature maps at different levels to the gated fusion decoder.

3. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 1, characterized in that, In step S2, the modulation branch uses a dilated convolutional pyramid structure to extract multi-scale environmental features, specifically including: Multiple dilated convolutional layers with different dilation rates are used in parallel to convolve the spliced ​​multi-source environmental field data to simultaneously capture local, mesoscale, and synoptic-scale environmental features. After upsampling the features obtained by global average pooling, the channels are concatenated with the outputs of each dilated convolutional layer; The concatenated features are fused and dimensionality reduced by a 1×1 convolutional layer to generate unified environmental modulation features.

4. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 1, characterized in that, In step S2, the gating fusion mechanism of the gated fusion decoder is implemented according to the following formula: ; ; in, For the evolution branch features received by the i-th layer decoder, For modulation branching characteristics, Indicates an upsampling operation. Indicates channel splicing. It is the Sigmoid activation function. Represents a 1×1 convolution. For the generated spatial adaptive gating weight graph, The projected environmental features This indicates element-wise multiplication. These are the features after fusion.

5. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 1, characterized in that, In step S3, the composite loss function includes a pixel-level error loss term, a physical constraint loss term, and a spectral domain matching loss term, defined as: ;in, A pixel-weighted Huber loss is used to strengthen the constraint on areas of heavy precipitation while balancing ordinary errors. Loss due to physical constraints; To achieve spectral matching loss, the spatial scale structure of the precipitation field is constrained by comparing the difference between the predicted field and the actual field in the Fourier space power spectrum. and To balance the hyperparameters.

6. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 5, characterized in that, The physical constraint loss Includes soft constraint terms related to mass conservation. and dynamic smoothing constraint terms ,in: The mass conservation soft constraint term Defined by the following formula: In the formula, This indicates summation of regions. This is the deterministic forecast output of the model, where M is the approximation term for the water vapor flux divergence in the lower atmosphere. It is a learnable scaling factor; The dynamic smoothing constraint term Defined by the following formula: In the formula, These represent the gradient operators in the spatial x-direction, y-direction, and time t-direction, respectively.

7. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 1, characterized in that, In step S4, the conditional generation network is a conditional generative adversarial network, which includes a generator G and a discriminator D; The generator G outputs the deterministic prediction model. The current environmental field E and the random noise vector z are taken as inputs, and the output is a probabilistic precipitation field. ; The input to the discriminator D is a real sample. Or generate samples The output is the probability that the input is real data; The conditional generative adversarial network minimizes the adversarial loss function between the generator and the discriminator. And simultaneously minimize the difference between the generated results and the actual precipitation field. Reconstruction loss Conduct joint training.

8. The method for constructing a local short-term precipitation forecasting model based on a neural network according to claim 1, characterized in that, In step S5, the probability forecast product includes at least: the ensemble average forecast calculated based on the members of the probability forecast ensemble, the probability of precipitation occurring, the probability of precipitation exceeding a preset threshold, and the ensemble discreteness product characterizing the forecast uncertainty.

9. A system for constructing a local short-term precipitation forecasting model based on neural networks, characterized in that, The system is used to implement a method for constructing a local short-term precipitation forecast model based on a neural network as described in any one of claims 1-8, comprising: The data preprocessing and sample construction module is used to acquire radar echo sequences, multi-source environmental field data and ground precipitation observation data of the target area, perform quality control, spatiotemporal alignment and gridding on the data, and construct spatiotemporally consistent training sample pairs. The sample pairs include historical observation sequences and corresponding future precipitation field ground values. A neural network model building module is used to construct a two-branch neural network model that integrates multi-scale features and physical constraints; the model includes: The encoder-evolution branch, with radar echo sequences as the main input, is used to extract the spatiotemporal evolution features of the precipitation system through an encoder containing a spatiotemporal self-attention mechanism. The modulation branch takes the multi-source environmental field at the current moment as input and is used to extract environmental field features through a multi-scale feature extractor. A gated fusion decoder, connected to the encoder-evolution branch and the modulation branch, is used to dynamically inject the environmental features extracted by the modulation branch into the spatiotemporal features of the evolution branch through a gated fusion mechanism to generate a deterministic precipitation forecast field for future time periods. The model training and optimization module includes: A deterministic model training unit is used to train the dual-branch neural network model using the training sample pairs, with the goal of minimizing the composite loss function, to obtain a deterministic prediction model; the composite loss function includes at least a pixel-level error loss term, a physical constraint loss term, and a spectral domain matching loss term; The probabilistic generation model training unit is used to construct and train a conditional generation network, which generates multiple probabilistic precipitation forecast fields based on the output of the deterministic forecast model and the current environmental field, forming a probabilistic forecast set for quantifying forecast uncertainty. The model deployment and inference module is used to integrate the deterministic forecast model and the conditional generation network, receive real-time meteorological data input, and output local short-term precipitation deterministic forecast and probabilistic forecast products.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a method for constructing a local short-term precipitation forecast model based on a neural network as described in any one of claims 1 to 8.

Images