Short-time rainfall prediction method based on radar image and reanalysis data fusion

Abstract

The invention discloses a short-time rainfall prediction method based on radar image and reanalysis data fusion, and the method comprises the following steps: collecting radar images and reanalysis data at continuous times, and generating input data in a unified grid format through spatial interpolation, time alignment and standardization processing; respectively extracting spatial and temporal features of the radar image and the reanalysis data by using a dual-channel encoder, and carrying out weighted fusion through a channel attention mechanism to generate a fusion feature tensor; inputting the fusion features into a ConvLSTM (Convolutional Long Short-Term Memory Neural Network), modeling a spatio-temporal evolution process of a rainfall system, and outputting a preliminary rainfall prediction image in 0-3 hours in the future; constructing a residual learning network, and performing deviation correction on the preliminary prediction result based on historical residual and observation information; when the radar image input is missing, the completeness of the input structure is maintained through the replacement feature generation module; generating a rainfall intensity image or a probability graph in 0-3 hours in the future; the method supports visual output.

Description

Technical Field

[0001] The present invention relates to the technical field of precipitation forecasting, and particularly relates to a short-term precipitation prediction method based on the fusion of radar images and reanalysis data. Background Art

[0002] In recent years, with the development of artificial intelligence technology, especially deep learning, more and more studies have explored introducing deep neural networks into the task of nowcasting precipitation. Structures such as convolutional neural networks (CNNs), convolutional long short-term memory networks (ConvLSTMs), generative adversarial networks (GANs), and Transformers have been widely applied to the modeling of radar echo sequences, and some results have surpassed the performance of traditional algorithms such as TREC in terms of prediction accuracy and structure preservation.

[0003] For example, models such as PredRNN and NowcastNet have achieved multi-step extrapolation of echo images by introducing spatio-temporal memory modules, showing good continuity and robustness; structures such as CIUnet have combined channel attention mechanisms to perform probabilistic identification of convective triggering signals in satellite cloud images; studies such as NNDA-VAE and Neural-Koopman have introduced the idea of data assimilation, modeled the state field through an encoder network, optimized the model residuals in the latent space, and improved prediction stability.

[0004] Although deep learning technology provides new possibilities, the existing research still faces the following main challenges: (1) Difficulty in fusing multi-source data: Radar images have high spatial resolution but strong timeliness, while reanalysis data such as ERA5 and FNL have low temporal resolution and coarse spatial granularity. Direct splicing and processing will introduce physical inconsistencies or scale aliasing problems; (2) Difficulty in controlling model errors: Neural networks have the problem of error accumulation during step-by-step prediction, especially in predicting echo intensity or boundary structures, which are prone to deviate from actual observations and are difficult to work stably for a long time; (3) Common problem of incomplete input: Affected by problems such as terrain occlusion, equipment maintenance, and signal loss at radar stations, image sequences are prone to frame loss or holes, affecting the operation of the model; (4) Insufficient interpretability and credibility: Most AI models are "black box structures", and it is difficult for business personnel to trace their prediction basis, restricting their use and promotion in actual weather warning systems.

[0005] Currently, there is still a lack of a unified modeling method with the following characteristics. Summary of the Invention

[0006] Objective of the Invention: The objective of the present invention is to provide a short-term precipitation prediction method based on the fusion of radar images and reanalysis data. Based on a dual-channel encoding structure and a convolutional long short-term memory neural network, it can fuse data at different scales within a unified structure, effectively control prediction residuals through a dynamic error correction network, has a function of filling in missing modalities, ensures continuous operation ability when radar data is unavailable, and is applicable to the refined intelligent forecasting requirements of short-term heavy precipitation processes.

[0007] Technical Solution: A short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to the present invention includes the following steps: (1) Collect radar images and reanalysis data at consecutive time intervals, and generate input data in a unified grid format through spatial interpolation, time alignment, and normalization processing; (2) Use a dual-channel encoder to extract the spatio-temporal features of radar images and reanalysis data respectively, and perform weighted fusion through a channel attention mechanism to generate a fused feature tensor; (3) Input the fused features into a Convolutional Long Short-Term Memory (ConvLSTM) network to model the spatio-temporal evolution process of the precipitation system, and output a preliminary precipitation prediction image for the next 0-3 hours; (4) Construct a residual learning network to correct the deviation of the preliminary prediction results based on historical residuals and observation information; (5) When the input of radar images is missing, maintain the integrity of the input structure through a substitute feature generation module; (6) Generate a precipitation intensity image or probability map for the next 0-3 hours, and adapt to the business system interface for output.

[0008] Further, in step (1), the radar images include reflectivity Z, differential reflectivity , specific differential phase and radial velocity V, with a time resolution of 6-10 minutes and a spatial resolution of 0.5 km×0.5 km or 1 km×1 km; the reanalysis data includes temperature T, relative humidity RH, horizontal wind speeds U, V, and geopotential height , with a time resolution of 1 hour and a spatial resolution of 0.25°×0.25°, and are aligned to the radar image grid through bilinear interpolation.

[0009] Further, in step (2), the dual-channel encoder includes: a radar image encoder and a reanalysis variable encoder; the radar image encoder is a three-layer convolutional neural network with the number of channels being 32→64→64 in sequence, and the activation function is ReLU; the reanalysis variable encoder adopts a 1×1 convolution and a fully-connected structure to maintain the independence of variable channels, and the number of output channels is 64; the channel attention weight is a learnable parameter, and the fusion formula is: ; where is the learnable channel attention weight, which adaptively allocates the fusion ratio according to the importance of the input features.

[0010] Further, the reanalysis variable encoder independently encodes the variables at each pressure level of 850 hPa, 700 hPa, and 500 hPa and then splices them, and upsamples them to the radar image resolution.

[0011] Further, in step (3), the ConvLSTM has a two-layer stacked structure, the number of input channels is 64, the number of output channels is 64, the convolution kernel size is 3×3, the stride is 1, and the activation functions are Tanh and Sigmoid; the temporal modeling module adopts a parallel multi-step decoding mechanism, and generates 18 future precipitation images through two convolutional networks, namely a 3×3 convolution kernel and a 1×1 convolution kernel, with a time interval of 10 minutes, and the output size is H×W.

[0012] Further, the Dropout mechanism is enabled in the training stage of the ConvLSTM unit and disabled in the inference stage.

[0013] Further, in step (3), the model is trained in an end-to-end manner, and simulated missing samples are constructed in the training set; the loss function is: ; where is the mean square error loss; M is the number of time steps of the predicted image; is the spatial dimension; is the predicted precipitation intensity value of the model at position (i, j) at time t.

[0014] Further, in step (4), the inputs of the dynamic error correction network include the preliminary predicted image, the historical residual and the reanalysis background variable; The correction formula is: ; where represents the preliminary precipitation image output by the ConvLSTM network; is the residual image predicted by the error correction network; is the final precipitation prediction result after correction; The loss function includes the mean squared error term and the L2 regularization term: ; where is the predicted value and the observed value the mean squared error between them; is the correction term amplitude control term; where λ is the adjustment weight.

[0015] Furthermore, in step (5), the modal missing completion includes: radar image validity detection, marking ; when , use the historical mean template or conditional generation network to generate alternative radar features ; the fusion weight is adjusted to 0.5 or adaptively adjusted by the attention module.

[0016] Furthermore, in step (6), the prediction output includes the real precipitation intensity map, the probability map after Sigmoid or Softmax conversion, and the accumulated precipitation map; the output format is NetCDF or GeoTIFF, with attached geographic coordinate metadata, supporting sliding window inference, that is, updating the future 3-hour prediction results every 10 minutes.

[0017] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: The present invention can fuse meteorological data from different sources and different scales (radar images + reanalysis background); while maintaining the detail accuracy of the images, introduce a dynamic error correction mechanism to control the output stability; can automatically switch to the "background-driven mode" to continue to complete the precipitation prediction task when radar data is unavailable; support interpretive visualization output to assist forecasters in understanding the model prediction logic. Brief Description of the Drawings

[0018] Figure 1 is the flow chart of the present invention; Figure 2 is the model flow chart of the present invention. Detailed Embodiments

[0019] The technical solution of the present invention will be further described below in conjunction with the drawings.

[0020] As Figure 1 shown, an embodiment of the present invention provides a short-term precipitation prediction method based on the fusion of radar images and reanalysis data, including the following steps: (1) Acquisition and preprocessing of source meteorological data: Acquire multi-source meteorological data required for model input, and perform standardization processing on the original data in the spatial, temporal, and physical dimensions to construct a data sample set with a unified structure, unified dimension, and suitable for input into a deep learning model. The input data includes the following two categories: Radar image data: It is an image sequence of variables such as reflectivity Z, differential reflectivity , specific phase , and radial velocity V obtained by a ground-based Doppler weather radar in the target area, with high temporal resolution (such as once every 6 minutes or 10 minutes) and high spatial resolution (such as 0.5km × 0.5km or 1km × 1km); Numerical reanalysis data: It is reanalysis field data from global or regional numerical weather models (such as ERA5, NCEPFNL, CMADS, etc.). The temporal resolution is generally 1 hour, and the spatial resolution is 0.25°× 0.25°. The main variables include temperature T, relative humidity RH, horizontal wind speeds U, V, and geopotential height etc., covering multiple standard pressure levels (such as 850hPa, 700hPa, 500hPa).

[0021] The above two types of data have different temporal resolutions, spatial distribution ranges, and physical meanings. To ensure the consistency between the data, co-training can be carried out in the neural network, and the following preprocessing processes are performed on the data: (11) Spatial range cropping and resampling: Determine the unified grid range according to the target prediction area (such as 64km × 64km), and crop the corresponding spatial areas from the radar image and reanalysis data with this area as the center: For radar image data, directly select the original resolution image block within the target area; For reanalysis data, use bilinear interpolation to resample it from the original 0.25° × 0.25° resolution to a grid structure consistent with the radar image (such as 128 × 128 or 64 × 64 pixels), and align it with the radar grid.

[0022] (12) Time series alignment and interpolation filling: Since the radar image data is usually at intervals of 6–10 minutes, while the reanalysis data is at intervals of 1 hour, to synchronously form a time-consistent input sequence, the following time alignment method is adopted in the present invention: Use linear interpolation to interpolate the reanalysis variables on the time axis to match the radar image time series; If there is missing reanalysis data, use forward filling or time mean filling methods for processing; Construct a unified sliding time window. For example, with the current time t as the reference, select the past N time steps (such as 6 frames, a total of 60 minutes) as the input sequence.

[0023] (13) Physical variable standardization and channel fusion: Normalize the input variables to enhance the training stability and cross-variable compatibility of the neural network. The standardization methods adopted include: performing min-max normalization or logarithmic transformation on radar variables (such as Z, Kdp, etc.); performing z-score standardization on reanalysis variables, that is: ; where μ is the historical mean and σ is the standard deviation, both pre-statistically obtained in the training set; The final single-timeframe input tensor formed is a multi-channel image structure, containing multiple radar and reanalysis variable channels; (14) Constructing input-output sample pairs: Construct the training sample set in a time-series sliding window manner. Each sample consists of an input sequence X and an output sequence Y: The input sequence is the images of the N frames before the current moment and the reanalysis variables ; The output sequence is the precipitation images at the M target time steps in the future ; The label data can be sourced from the interpolated map of ground rain gauge observations, merged precipitation analysis products (such as IMERG, CaPA), or numerical model forecast precipitation data; All samples are constructed in batches to form a four-dimensional tensor structure and input into the model for training. The construction form of the data samples can be expressed as: ; where represents the radar image variables, represents the reanalysis variables, represents the interpolation operation that has been uniformly processed in space and time.

[0024] The time resolution of the radar image data is 6 minutes, and the spatial resolution is 0.5 km × 0.5 km. The time resolution of the reanalysis data is 1 hour, and the spatial resolution is 0.25° × 0.25°. After spatial bilinear interpolation, it is aligned with the radar image grid, and the output size is 64 × 64; The radar variables include reflectivity Z, specific phase Kdp, radial velocity V, etc.; The reanalysis variables include temperature T, relative humidity RH, wind speed U / V, and geopotential height Zg at the 850 hPa, 700 hPa, and 500 hPa height levels.

[0025] The multi-source data preprocessing mechanism of the present invention ensures the consistency of different data sources in physical structure and spatial alignment, providing a stable and reliable input basis for the subsequent feature extraction module and time-series modeling module, which is a key pre-step for the present invention to achieve end-to-end intelligent precipitation prediction.

[0026] (2)Construction of multi-source feature extraction and fusion module: Extract features from the preprocessed multi-source meteorological input data and achieve structural fusion in the intermediate latent space. The core objective of this step is to extract the spatio-temporal features of radar images and reanalysis background fields, compress redundancy while maintaining the original physical information, enhance the semantic expression ability, and provide a fused feature input in a unified format for the subsequent time series modeling module. The feature extraction module adopts a parallel dual-channel neural network structure, corresponding to radar observation images and reanalysis meteorological variables respectively. The outputs of the two are upsampled, dimensionally aligned, and weighted fused to form a fused feature tensor with a unified spatial size and channel depth. The feature extraction module adopts a dual-channel neural network structure, including: The radar image encoder adopts a three-layer convolutional network with the number of channels being 32→64→64 in sequence, and the activation function is ReLU; the reanalysis variable encoder adopts a 1×1 convolution and fully connected structure to maintain the independence of variable channels, and the final output number of channels is 64; the two encoding results are weighted and fused through a channel attention mechanism, and the weight α is a learnable parameter; including the following steps: (21)Construction of radar image encoder: The present invention adopts a convolutional neural network (CNN) structure to construct a radar image encoder module for extracting spatial structure features and echo morphology information in radar sequence images. This module can effectively capture precipitation boundaries, strong echo centers, and organizational structures at different scales, and provide high-quality representations for subsequent feature fusion and time series modeling. The radar image encoder consists of multiple convolutional layers, batch normalization layers (BatchNorm), and non-linear activation functions (ReLU), forming a shallow encoding network. Each convolutional unit structure includes: 1. A two-dimensional convolutional layer (Conv2D) with a kernel size of 3×3 and a stride of 1 for extracting local spatial features; 2. A batch normalization layer (BatchNormalization) for normalizing the activation outputs in a small batch of samples, effectively alleviating internal covariate shift, improving training stability, and accelerating the convergence speed; 3. A rectified linear unit (ReLU) activation function with the expression , for introducing non-linear modeling capabilities, enhancing the response to echo gradient changes, and preventing the gradient vanishing problem; the above structure is repeated several times, and the number of channels increases layer by layer to gradually extract the semantic features from weak to strong and from shallow to deep in the image. The input of the radar image encoder is a four-dimensional tensor formed by stacking multiple frames of images, with the size of: where T represents the number of time frames, H and W represent the height and width of the image, and C represents the number of channels (such as variables like Z, Kdp, etc.). The output is the fused intermediate feature tensor .

[0027] (22)Re - analysis variable encoder construction: The re - analysis data is a structured grid tensor, containing three - dimensional tensor information of multiple meteorological variables and multiple pressure levels. A lightweight CNN or a combination of 1×1 convolution + fully - connected layers is required to construct the encoder. The re - analysis encoder designed in the present invention performs convolution processing on each variable channel without sharing between channels, maintaining the independence of variables, and finally outputs a feature tensor , whose size is consistent with the output of the radar encoder.

[0028] When the re - analysis input is of low resolution (e.g., 32 × 32), an up - sampling (bilinear or transposed convolution) method is used to interpolate it to the same resolution as the radar image (e.g., 64 × 64 or 128 × 128) to ensure spatial dimension matching.

[0029] (23)Feature fusion strategy: The present invention adopts a feature fusion method based on channel attention mechanism to perform weighted combination on the radar and re - analysis coding results. The fusion expression is: , where is the learnable channel attention weight, which adaptively allocates the fusion ratio according to the importance of the input features.

[0030] (24)Fusion result output and structure matching: The fused feature tensor is output in a unified - dimension structure, with the size of: , serving as the input interface for the subsequent temporal modeling module (ConvLSTM) and error correction network.

[0031] The feature extraction and fusion structure of the present invention supports flexible expansion, is applicable to different radar data structures (such as PPI, CAPPI, RHI) and various re - analysis data forms (such as ERA5, FNL, regional self - built models), and has good generality and engineering deployment capabilities.

[0032] This step realizes the unification of spatial resolution, the matching of feature channels, the retention of information structure, and the fusion of trainable weights, providing a high - quality input basis for the subsequent dynamic modeling and error regulation of the model in the time dimension, and is a crucial key step in the multi - source information fusion prediction system.

[0033] (3)The temporal modeling module is constructed to simulate the evolution process of the precipitation system over time. This module takes the spatio - temporal feature tensor generated by fusion in the second step as input, and completes the prediction of precipitation intensity at future target time steps through the recursive propagation of information in the time dimension. For example Figure 2As shown in the figure, the present invention uses a Convolutional Long Short-Term Memory Network (ConvLSTM) as the main temporal modeling unit. Compared with the traditional LSTM structure, ConvLSTM uses convolutional operations instead of fully connected operations inside the unit, which can capture the dynamic correlations in both time and space simultaneously, and is especially suitable for the temporal modeling tasks of radar image sequences or spatial tensors.

[0034] The temporal modeling module includes two stacked Convolutional Long Short-Term Memory Networks (ConvLSTM) for modeling the spatio-temporal evolution characteristics of the precipitation system.

[0035] (31)ConvLSTM Structure Design and Parameter Setting: The ConvLSTM network structure includes multiple stacked ConvLSTM units, and each layer is responsible for modeling the spatio-temporal change trends at different semantic levels. The network structure is shown as follows: ConvLSTM Layer 1: Input channels D, output channels D′, convolutional kernel size 3×3, stride 1; ConvLSTM Layer 2: Further abstract the spatial semantics and increase the receptive field; LayerNorm is used for regularization and normalization control.

[0036] At each time step t, ConvLSTM receives the fused feature input and updates based on the hidden state and the memory state as follows: ; Finally, a sequence of hidden states of a time series is output for generating future precipitation image predictions.

[0037] (32)Multi-step Output Mechanism Design: The present invention supports continuous output predictions for multiple time steps, that is, the model can output the precipitation field distributions at several future moments at once, such as the target time steps of the next 30 minutes, 1 hour, 2 hours, 3 hours, etc.

[0038] To achieve multi-step prediction, the present invention designs the following two optional strategies: a Direct Multi-step Output Method: After ConvLSTM outputs the hidden sequence, it is connected to the decoder network to directly output the precipitation prediction images for all future target time steps (parallel prediction); b Auto-regressive Prediction Method: The current output is used as the input for the next moment to gradually advance the prediction process (recursive prediction).

[0039] The present invention preferentially adopts the direct multi-step output method because of its higher computational efficiency and less error accumulation. The corresponding formula is as follows: ; where is the predicted precipitation image at the th moment, is a set of shareable or independent convolutional decoding layers.

[0040] (33) Decoder decoding structure design: The decoding module of the present invention adopts a shallow convolutional neural network structure, which consists of 1 to 2 layers of convolutional inverse transforms and is used to map the hidden state H t to a precipitation image of the target size. The output image dimension is the same as that of the original radar Figure 1 For example: ; where M is the number of output time steps, and H and W are the height and width of the image.

[0041] Decoder output type: A real-valued precipitation intensity map with the unit of millimeters per hour (mm / h), which is used to characterize the precipitation intensity per pixel unit in the prediction area; A probabilistic precipitation map, which generates pixel-level precipitation probabilities by applying a Sigmoid or Softmax function transformation to the output result of the decoder, and is preferably used to judge whether the precipitation exceeds a set threshold (such as 5 mm / h, 10 mm / h), etc.; The output method can be flexibly configured according to different business requirements to support various application scenarios such as quantitative forecasting or graded early warning.

[0042] (34) Training method and loss function selection: The training objective of the model of the present invention is to minimize the error between the predicted image and the real precipitation image. The model is trained in an end-to-end manner, and simulated missing samples are constructed in the training set; The optional loss function is: Mean Squared Error (MSE): ; where is the mean squared error loss; M is the number of time steps of the predicted image; is the spatial dimension; is the predicted precipitation intensity value of the model at the position (i, j) at time t.

[0043] This step realizes the deep modeling of temporal features, captures the non-linear evolution trend of the precipitation system over time, and has the ability to model processes such as movement, generation, enhancement, and attenuation. It is the key step for the present invention to convert from static perception to dynamic prediction.

[0044] (4) Construction of the dynamic error correction module: Perform a secondary adjustment on the initially predicted precipitation result to reduce the prediction deviation caused by model non-linear deviation, input uncertainty, or multi-source error transmission, and improve the accuracy and credibility of the precipitation image.

[0045] Different from relying on the single output result of the main model in traditional methods, the present invention proposes a dynamic error correction network based on a residual learning structure, which learns the error term between the predicted value and the true value, and uses this error term to correct the initial output to achieve "fine-tuning enhancement" of the predicted image. The method includes the following steps: (41) Design of the residual modeling structure: The dynamic error correction network adopts a lightweight convolutional neural network structure, and its input includes the preliminary predicted image of the main model , the historical residual image (such as the residual at the previous moment ), and optional background variables, such as the humidity gradient and wind vector change in the reanalysis field. The network structure may include the following components: Convolutional layer Conv1: The input channel is N, and the output channel is D; BatchNorm + ReLU activation; Convolutional layer Conv2: The output channel is 1, which is mapped to the residual map; A residual connection or skip connection mechanism is set in the convolutional structure to improve the feature extraction ability and network stability. The residual prediction module outputs a correction term image , which is added to the original predicted image to form the final output: ; Among them, represents the preliminary precipitation image output by the ConvLSTM network; is the residual image predicted by the error correction network; is the final precipitation prediction result after correction.

[0046] (42) Input composition and feature encoding method: To enhance the generalization ability of the error correction module, the present invention combines and expands the input of the correction network, specifically including: the current output of the main model ; the difference between the historical prediction and observation (i.e., the residual map): Optional reanalysis background quantities, such as , wind field convergence, etc.; All inputs are stacked to form a multi-channel image, which is jointly encoded by a feature extraction module (shallow convolutional network), and then enters the residual prediction subnet to complete the correction term calculation.

[0047] (43) Correction strategy and constraint mechanism: To avoid structural damage caused by overcorrection, the present invention introduces the following mechanisms to control the correction intensity and direction: Set the maximum amplitude limit of the correction term (such as ±10 mm / h); Add a regularization term to the loss function to punish excessive correction; Use a sliding window to smooth the prediction trend and enhance temporal consistency; (44) Training method and loss combination: The present invention cascades or trains the error correction network and the main model in stages. A combined loss function is used during training: ; Among them, is the predicted value The mean square error with the observed value ; is the correction term amplitude control term; where λ is the adjustment weight.

[0048] The correction network can be deployed in ways such as separate training, end-to-end joint training, or online incremental learning, and can be flexibly selected according to the actual business system requirements. In this step, by constructing a lightweight and structurally stable dynamic error correction network, the model is enabled to automatically adjust the output according to historical deviations and current states, thereby reducing the error accumulation effect, improving the accuracy, coherence, and credibility of the predicted image, and enhancing the deployability and fault tolerance performance of the model in the actual meteorological business system. This is one of the important technical innovations of the present invention.

[0049] (5) Construction of the modality missing completion mechanism is used to solve the problem of the stable operation of the model in the case where some input data is unavailable (such as radar image missing, occluded, quality abnormal, etc.).

[0050] In the actual meteorological business system, radar observation images may be missing for a short time due to equipment maintenance, communication failures, regional blind spots, or strong precipitation shielding. The neural network model is highly sensitive to the input structure. If missing data is forced to be input, it is likely to cause model performance degradation or operation interruption. Therefore, to cope with the situation of key modality missing, the present invention introduces an input perception and dynamic completion mechanism to automatically switch to the completion mode to ensure the stability of the inference process. It includes the following steps: (51) Input modality detection mechanism: The present invention embeds modality validity detection logic in the data reading module to perform integrity and quality assessment operations on each frame of radar image: If the radar image data is complete and the structure is reasonable, it is marked as "valid modality"; If the image is empty, the structure is severely damaged, or the noise ratio exceeds the threshold, it is marked as "modality missing". The detection module dynamically evaluates the current input situation at each time step and generates a modality status flag ; where, indicates that the radar image is available; indicates that the radar image is missing or unavailable.

[0051] (52) Alternative feature generation strategy: When , the present invention activates the "modality completion module" to generate alternative radar features . This module is designed with two typical implementation methods: Static template filling: Using historical statistical means, radar image mean templates, or blank tensors as alternative inputs, suitable for lightweight and fast deployment scenarios; Dynamic generation filling: Introduce a small auxiliary network (such as a conditional generation network, mapping network), and predict the estimated radar features according to the current reanalysis variable features , input it into the main model as a completion value. The alternative features need to be consistent with the normal radar encoder output in terms of spatial dimensions and channel dimensions to ensure the normal operation of the subsequent feature fusion module.

[0052] (53) Adaptive processing of the fusion module: After receiving the alternative radar features at a certain moment, the feature fusion module still performs channel weighted fusion operations: ; where can be the default weight (0.5) for quick deployment scenarios, or be adaptively learned by the attention module. In the case of missing radar images, the fusion mechanism can further tend to increase the relative weight of the reanalysis variable channels, thereby enhancing the model's dependence on the background meteorological field and ensuring the stability and integrity of the prediction results.

[0053] (54) Modal completion compatible structure: The overall model structure of the present invention supports modal-aware operation, that is, if the radar modality is available, the system runs the complete fusion process; if the radar modality is missing, the input integrity is maintained through the alternative mechanism, and the output result is slightly worse but the structure is stable. The modal completion mechanism is embedded before the model entrance and the encoder, and has the characteristics of structural decoupling, clear logic, easy deployment, etc., and is suitable for a variety of actual business scenarios, including: areas with unstable radar signals in remote mountainous areas, sudden equipment offline scenarios, data channel frame loss, etc.

[0054] (55) Joint training strategy: The present invention supports the simultaneous training of data samples in the complete mode and the missing mode, that is, constructing some "simulated missing" samples in the training set so that the model can work normally in both modes, improving generalization and robustness.

[0055] By introducing the input modality detection, alternative feature generation and fusion adaptation mechanism, this step realizes the goal that the model can still run stably in the case of unavailable key observation modalities, and solves the problem that traditional deep learning models are highly dependent on data and are prone to failure. It is one of the important technical designs for the present invention to improve engineering practicability and system continuity.

[0056] (6) Construction of prediction output and visualization mechanism: Format the precipitation prediction results output by the model, and provide various types of business output interfaces and visualization functions to meet the actual application scenarios of different meteorological business requirements, forecasting models, and interpretive analysis. The core goal of this step is to convert the intermediate features (precipitation images at multiple time steps) output by the model into standard products that can be directly used in forecasting systems, visualization platforms or decision support systems. It includes the following steps: (61) Real-valued precipitation intensity map generation: After the model output tensor is processed by the decoder, a pixel-by-pixel precipitation intensity image at a certain time in the future is obtained, in millimeters per hour (mm / h), with the following structure: ; Where H, W represent the spatial dimensions of the image, and M is the number of output time steps (e.g. 6 steps correspond to the next 3 hours), which can be used to generate cumulative precipitation maps of any time step (10min / 30min / 60min).

[0057] The present invention can optionally configure a cumulative precipitation output mode, that is, superimpose continuous outputs to generate products such as "total precipitation in the next hour" and "total precipitation in the next 3 hours", and the format is consistent with the actual business system requirements.

[0058] (62) Generation of precipitation probability maps and threshold products: To enhance the business interpretability of the model, the present invention supports converting the real-value precipitation output into a probability image via a Sigmoid function or a Softmax classifier, which is used to generate the following types of products: pixel-level precipitation probability maps (such as “the probability of precipitation > 5 mm / h at this point in the next hour”); pixel-by-pixel category maps (such as “no precipitation / light rain / moderate rain / heavy rain” multi-classification maps); zoning statistical maps (such as the percentage of area in a certain area where the probability of precipitation exceeds a certain threshold); the above image structures are highly compatible with actual meteorological service platforms (such as early warning visualization systems and meteorological release platforms).

[0059] (63) Multi-time step output interface design: To support sliding window continuous forecasting, the model of the present invention supports rolling input of the latest data at fixed time intervals (e.g., 10 minutes), outputs precipitation forecast images for the next 3 hours, and automatically updates. Each round of forecast output can be packaged in standard NetCDF format with geographic coordinate metadata, which is convenient for direct embedding into GIS platforms, business systems, or disaster warning interfaces.

[0060] (64) The output results can be connected to existing meteorological business systems, such as forecast duty stations, urban rainstorm warning systems, airport operation and dispatch systems, water conservancy emergency management systems, etc., and can be configured as multi-resolution, multi-variable parallel output to meet the accuracy requirements of different business systems.

[0061] The present invention provides a complete prediction result output mechanism, supports the generation of real-valued, probabilistic and multi-category precipitation images, and has good interpretability and business compatibility, forming an effective bridge from model reasoning results to meteorological business services, ensuring the goals of the present invention's deployment, rapid integration and continuous optimization.

Claims

1. A short-term precipitation prediction method based on the fusion of radar images and reanalysis data, characterized in that, It includes the following steps: (1) Collect radar images and reanalysis data at consecutive times, and generate input data in a unified grid format through spatial interpolation, time alignment, and normalization processing; (2) Use a dual-channel encoder to extract the spatio-temporal features of radar images and reanalysis data respectively, and perform weighted fusion through a channel attention mechanism to generate a fused feature tensor; (3) Input the fused features into a convolutional long short-term memory neural network (ConvLSTM) to model the spatio-temporal evolution process of the precipitation system, and output a preliminary precipitation prediction image for the next 0 - 3 hours; (4) Construct a residual learning network to correct the bias of the preliminary prediction results based on historical residuals and observational information; (5) When the input of radar images is missing, maintain the integrity of the input structure through a substitute feature generation module; (6) Generate a precipitation intensity image or probability map for the next 0 - 3 hours, and adapt to the business system interface for output.

2. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, characterized in that, In step (1), the radar image includes reflectivity Z, differential reflectivity , specific phase and radial velocity V, with a time resolution of 6 - 10 minutes and a spatial resolution of 0.5 km × 0.5 km or 1 km × 1 km; the reanalysis data includes temperature T, relative humidity RH, horizontal wind speeds U, V and geopotential height , with a time resolution of 1 hour and a spatial resolution of 0.25° × 0.25°, and is aligned to the radar image grid through bilinear interpolation.

3. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, wherein In step (2), the dual-channel encoder includes: a radar image encoder and a reanalysis variable encoder; the radar image encoder is a three-layer convolutional neural network with the number of channels being 32 → 64 → 64 in sequence, and the activation function is ReLU; the reanalysis variable encoder adopts a 1×1 convolution and a fully connected structure to maintain the independence of variable channels, and the number of output channels is 64; the channel attention weight is a learnable parameter, and the fusion formula is: ； Among them, is the learnable channel attention weight, which adaptively assigns the fusion ratio according to the importance of the input features.

4. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 3, wherein, The reanalysis variable encoder independently encodes the variables at each pressure level of 850 hPa, 700 hPa, and 500 hPa and then splices them, and upsamples them to the resolution of the radar image.

5. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, characterized in that, In step (3), ConvLSTM has a two-layer stacked structure, the number of input channels is 64, the number of output channels is 64, the convolution kernel size is 3×3, the stride is 1, and the activation functions are Tanh and Sigmoid; the temporal modeling module adopts a parallel multi-step decoding mechanism to generate 18 frames of precipitation images for the future through two convolutional networks, namely a 3×3 convolution kernel and a 1×1 convolution kernel, with a time interval of 10 minutes, and the output size is H×W.

6. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 5, characterized in that, The Dropout mechanism is enabled in the training stage of the ConvLSTM unit and disabled in the inference stage.

7. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, wherein In step (3), the model is trained in an end-to-end manner, and simulated missing samples are constructed in the training set; the loss function is: ； Among them, is the mean squared error loss; M is the number of time steps of the predicted image; is the spatial dimension; is the predicted precipitation intensity value of the model at position (i, j) at time t.

8. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, wherein In step (4), the inputs of the dynamic error correction network include the preliminary prediction image, the historical residuals and the reanalysis background variables; The correction formula is: ； Among them, represents the preliminary precipitation image output by the ConvLSTM network; is the residual image predicted by the error correction network; is the final precipitation prediction result after correction; The loss function includes a mean squared error term and an L2 regularization term: ; where is the predicted value and the observed value the mean squared error between; is the correction term amplitude control term; where λ is the adjustment weight.

9. The short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, wherein In step (5), the modal missing completion includes: radar image validity detection, marking ; when occurs, use the historical mean template or conditional generation network to generate alternative radar features ; the fusion weight is adjusted to 0.5 or adaptively adjusted by the attention module.

10. A short-term precipitation prediction method based on the fusion of radar images and reanalysis data according to claim 1, characterized in that, In step (6), the prediction output includes a real-value precipitation intensity map, a probability map after Sigmoid or Softmax conversion, and an accumulated precipitation map; the output format is NetCDF or GeoTIFF, with attached geographical coordinate metadata, and supports sliding window inference, that is, the future 3-hour prediction results are updated every 10 minutes.

Images