Minute-level rainfall nowcasting method based on deep learning

Abstract

The application discloses a minute-level rainfall nowcasting method based on deep learning, comprising the following steps: obtaining radar data, measured rainfall and rainfall forecast product of a preset area, obtaining original data, normalizing the original data, obtaining grid data of a rainfall analysis field, obtaining a basic forecast model through deep learning according to the grid data, comparing the forecast rainfall of the basic forecast model with the measured rainfall according to the forecast period of the rainfall forecast product, generating minute-level samples of corresponding regional equal distribution through a preset sample generator according to rainfall deviation, taking the rainfall forecast product as an additional input of a decoder of the basic forecast model based on the minute-level samples, training the basic forecast model based on the additional input samples, and obtaining an optimized model and a minute-level rainfall nowcasting product. The method uses the construction of the basic forecast model and the adjustment of the sample distribution to improve the CSI score index of the optimized model, and provides a forecast product with higher precision and stronger stability for short-time nowcasting of the preset area.

Description

Technical Field

[0001] This invention relates to the field of precipitation forecasting, and more particularly to a method for short-term rainfall forecasting at the minute level based on deep learning. Background Technology

[0002] For severe convective weather such as short-duration heavy rainfall and thunderstorms, the main reliance is on accurate rainfall forecasts at the hourly, minute, and kilometer levels. Short-term forecast products are directly related to the effectiveness of disaster prevention work such as flood control scheduling, traffic management, and emergency response. Among these, radar echo data, due to its high spatiotemporal resolution, has become a core influencing factor for short-term nowcasting of precipitation. However, single radar observations have regional blind spots, and the corresponding forecast accuracy still has room for further improvement. Currently, the field of meteorological forecasting has gradually shifted from traditional numerical model-driven approaches to a technical path that integrates multi-source observation data with deep learning. This involves first using simple variables and combining them with deep learning model structures to train a basic forecast model, and then improving the model's forecasting capabilities by adding variables and refining the model structure. Summary of the Invention

[0003] The purpose of this invention is to provide a method for short-term rainfall forecasting at the minute level based on deep learning.

[0004] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0005] This invention provides a method for minute-level short-term rainfall forecasting based on deep learning, including:

[0006] Acquire radar data, measured rainfall, and rainfall forecast products for a preset area; perform quality control on the radar data, measured rainfall, and rainfall forecast products to obtain raw data;

[0007] Based on the original data, the radar combined reflectivity factor, measured rainfall and rainfall forecast products are normalized to obtain gridded data of the rainfall analysis field.

[0008] Based on the grid data, non-sliding window sampling samples are extracted, and deep learning is performed on the sampling samples to obtain a basic prediction model with an encoding and decoding structure.

[0009] Based on the prediction period of the rainfall forecast product, the predicted rainfall and the measured rainfall of the basic forecast model are compared, and minute-level samples are generated equally in the corresponding area according to the rainfall deviation through a preset sample generator.

[0010] Based on the minute-level samples, the rainfall forecast product is used as an additional input to the decoder of the basic forecast model. The basic forecast model is trained based on the additional input samples, and the CSI score is calculated to obtain the optimized model and the minute-level short-term rainfall forecast product.

[0011] Furthermore, the method for obtaining the original data includes:

[0012] The system acquires raw radar echo observation data, minute-level measured rainfall from ground rain gauges, and rainfall forecast products within a preset area. It removes non-precipitation echo interference data and outliers from the raw radar echo observation data to obtain radar data. It removes invalid zero values ​​based on measured rainfall and completes short-term missing data through interpolation with neighboring stations. It matches the spatiotemporal range of the rainfall forecast products with the preset area and removes forecast data with misaligned spatiotemporal ranges.

[0013] The raw data is obtained by using the radar data, measured rainfall, and rainfall forecast products in a unified geographic coordinate system and timestamp format.

[0014] Furthermore, the method for obtaining the grid data includes:

[0015] Based on the raw data, the radar combined reflectivity factor is linearly normalized according to its physical value range. The measured rainfall and rainfall forecast products are normalized using the extreme precipitation thresholds for a preset area to obtain gridded data for the rainfall analysis field. The time decomposition formula for the gridded data is as follows:

[0016] ;

[0017] ;

[0018] in For the first Spatial coordinates within a 10-minute time period The spatial average value of the radar combined reflectivity factor in the corresponding region. The number of grid rows for the preset sub-region. The number of grid columns for the preset sub-region. This refers to the row offset of the grid points within the preset sub-region. This refers to the column offset of the grid points within a preset sub-region. Y is the baseline x-axis, and Y is the baseline y-axis. This is the normalized radar composite reflectivity factor value. For spatial grid points No. Rainfall forecast values ​​for 10-minute intervals, Spatial coordinates The 1-hour cumulative rainfall forecast value at the location, For the first A timestamp for a 10-minute time interval.

[0019] Furthermore, the method for obtaining the basic forecast model includes:

[0020] Based on the grid data, the radar combined reflectivity factor grid data and rainfall grid data of a fixed length period are used as historical observation input through a non-sliding window. The rainfall grid data of continuous periods outside the historical observation period are used as supervised learning labels to obtain sampling samples. Based on the sampling samples, they are randomly divided into training set, validation set and test set according to the time dimension.

[0021] A deep learning model with an encoding and decoding structure is constructed based on the training set to obtain a basic forecast model. The encoder and decoder are memory units of the same type, including ConvLSTM and ConvGRU. The memory units are stacked in two layers. Deep learning is performed on the basic forecast model based on the training set, and the model parameters are updated according to the mean squared error loss function until the validation set loss does not decrease for three consecutive rounds, thus obtaining the trained basic forecast model.

[0022] Furthermore, the method for obtaining the rainfall deviation includes:

[0023] Based on the rainfall forecast product, the forecast period is obtained. According to the forecast period, the grid point data of the rainfall analysis field in different forecast period lengths are randomly selected and input into the basic forecast model to obtain the predicted rainfall at each grid point. The predicted rainfall and the measured rainfall are compared to calculate the rainfall deviation, which includes absolute deviation and relative deviation.

[0024] Based on the rainfall deviation, the deviation area is identified, and the sampling frequency of the high deviation area is increased to twice to obtain the adjusted sample generation parameters. Based on the sample generation parameters, the corresponding area is evenly distributed into minute-level samples by a preset sample generator. The number of samples in the minute-level samples is evenly distributed between the high deviation area and the low deviation area according to the grid point ratio.

[0025] Furthermore, the method for optimizing the model includes:

[0026] Rainfall forecast products are embedded into feature vectors based on minute-level samples. These feature vectors are then adaptively fused with the spatiotemporal features output by the encoder. The fusion result is used as an additional input to the decoder. The gating calculation formula for the adaptive fusion is as follows:

[0027] ;

[0028] ;

[0029] in For the first Adaptive gating weights at each time step It is the Sigmoid activation function. This is the weight matrix of the gate layer. This refers to the historical hidden state output by the encoding end. Forecast characteristics of rainfall forecast products This is the bias vector of the gated layer. No. Hidden state of the decoding end at each time step For the first Hidden state of the decoding end at each time step For Hadamah accumulation;

[0030] Based on the updated basic forecast model trained with additional input samples, the difference between the updated predicted rainfall and the supervised learning labels is calculated according to the loss function, and the learnable parameters of the model are updated by backpropagating gradients through the optimizer. After each round of training, the CSI score of the prediction results is calculated on the sample test set until the CSI score improves by less than 1% for three consecutive rounds, thus obtaining the optimized model. Based on the optimized model, minute-level short-term rainfall forecast products are output.

[0031] Furthermore, the method for obtaining the CSI score includes:

[0032] Based on the prediction results and the actual measured rainfall, the number of grid time pairs where precipitation was predicted and actually occurred is taken as the hit count, the number of grid time pairs where precipitation was predicted but not actually occurred is taken as the false alarm count, and the number of grid time pairs where precipitation was predicted but not actually occurred is taken as the missed alarm count. The CSI score is calculated based on the hit count, false alarm count, and missed alarm count. The CSI score includes time-averaged CSI, spatial-averaged CSI, and global CSI. The time-averaged CSI is the average of the CSI scores for a single time step, and the spatial-averaged CSI is the average of the CSI scores for a single grid point.

[0033] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects:

[0034] This invention constructs a basic forecast model with an encoding-decoding structure, introduces numerical model precipitation forecast products as additional input to the decoder, and adaptively fuses radar observation details and large-scale model information through a gating mechanism to improve rainfall prediction bias and extend effective forecast lead time. By using a preset sample generator to flexibly adjust the sample distribution according to rainfall bias, the sampling frequency is increased in bias areas to enhance the model's learning ability for extreme precipitation. Spatiotemporal matching of model forecasts and radar observations is performed to improve the coordination and consistency of multi-source data, improve the CSI score index of the optimized model, and provide more accurate and stable forecast products for 0-2 hour short-term nowcasts in the preset area. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the steps of the minute-level short-term rainfall forecasting method based on deep learning in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Reference Figure 1 As shown, this invention provides a method for minute-level short-term rainfall forecasting based on deep learning, including:

[0038] Acquire radar data, measured rainfall, and rainfall forecast products for a preset area; perform quality control on the radar data, measured rainfall, and rainfall forecast products to obtain raw data;

[0039] In the actual assessment, the pre-defined area was the region with a high incidence of severe convective precipitation in summer. Radar echo data, measured rainfall data from ground rain gauges, and rainfall forecast products were acquired from June to July. The radar data included raw observation data of radar network combined reflectivity factors, with a resolution of 6 minutes and 1 km, a grid size of 800×800, and a value range of 0-70 dBZ. The measured rainfall was minute-level measured rainfall data from ground rain gauges. The rainfall forecast product was a 1-hour cumulative rainfall forecast product output from the RMAPS-RISE numerical model. Non-precipitation echo interference data, including ground object echoes and superrefractive echoes, was removed from the radar data using a thresholding method, with a value <5 dB. Z, outliers were removed using the 3σ criterion to obtain quality-controlled radar data. Abnormal zero values ​​during non-precipitation periods of the measured rainfall were removed. Short-term missing data of ≤3 10-minute intervals at a single station were supplemented by interpolation with neighboring stations to obtain quality-controlled measured rainfall. The spatiotemporal range of the rainfall forecast product was matched with the preset area of ​​114°-119° E longitude and 36°-42° N latitude. Forecast data with misaligned spatiotemporal ranges were removed, and forecast grid points that overlapped with the preset area were retained. The quality-controlled radar data, measured rainfall, and rainfall forecast product were unified into the WGS84 geographic coordinate system, and the timestamps were unified into the Beijing time minute-level format to obtain the raw data.

[0040] Based on the original data, the radar combined reflectivity factor, measured rainfall and rainfall forecast products are normalized to obtain gridded data of the rainfall analysis field.

[0041] In the actual assessment, the radar combined reflectivity factor was linearly normalized according to the physical value range of 0-70dBZ, and the data was mapped to the 0-1 interval. The measured rainfall and rainfall forecast products were normalized according to the preset extreme precipitation threshold of 20mm / 1h, and the data was mapped to the 0-1 interval. Rainfall exceeding the threshold was uniformly set to 1. The measured rainfall was mapped to spatial grid points consistent with the radar data through spatial interpolation. The rainfall forecast products were resampled to a unified grid. The 1-hour cumulative rainfall forecast products were converted into 10-minute rainfall forecast data through time decomposition formula to obtain 10-minute grid point data of the rainfall analysis field.

[0042] Based on the grid data, non-sliding window sampling samples are extracted, and deep learning is performed on the sampling samples to obtain a basic prediction model with an encoding and decoding structure.

[0043] In the actual evaluation, gridded data of radar combined reflectivity factors and measured rainfall at six consecutive time steps of fixed length were selected as the historical observation input of the model. Measured rainfall at 12 consecutive time steps after the historical observation period were selected as the supervised learning labels. Non-sliding window sampling was performed on the rainfall analysis field data in June, resulting in 310 valid samples. These samples were randomly divided into a training set of 248 samples, a validation set of 31 samples, and a test set of 31 samples in a time-division stratification ratio of 8:1:1. A deep learning model with an encoder-decoder structure was constructed based on the training set. Both the encoder and decoder used two layers of stacked ConvLSTM units. The ConvLSTM convolution kernel size was 3×3, the stride was 1, and the padding method was "same". The loss function in the training parameters was mean squared error, the optimizer was Adam, the initial learning rate was 1e-3, the batch size was 4, and the number of training epochs was 50. The validation set loss was monitored in real time during training. When the validation set loss did not decrease for three consecutive epochs, an early stopping mechanism was triggered, training was stopped, and the basic forecast model was saved.

[0044] Based on the prediction period of the rainfall forecast product, the predicted rainfall and the measured rainfall of the basic forecast model are compared, and minute-level samples are generated equally in the corresponding area according to the rainfall deviation through a preset sample generator.

[0045] In the actual assessment, complete grid data of the rainfall analysis field from 14:00 to 14:54 on July 1st was selected and input into the basic forecast model to obtain the grid-by-grid and 10-minute forecast results for the next 0-2 hours. The forecast results were converted into physically meaningful predicted rainfall through inverse normalization. The predicted rainfall was compared with the measured rainfall at the same time and resolution grid point by grid point to calculate the absolute and relative rainfall deviations. The grid points were divided into high-deviation and low-deviation regions with an absolute deviation of 0.5 mm / 10 min as the boundary. The high-deviation regions are mainly urban areas and surrounding areas with high incidence of severe convective precipitation, accounting for about 35% of the grid points. The low-deviation regions account for about 65% of the grid points. The sampling frequency of the high-deviation regions was increased to twice that of the low-deviation regions, while the original sampling frequency was kept unchanged to obtain the adjusted sample generation parameters. Based on the sample generation parameters, the regionally evenly distributed minute-level samples were generated through a preset sample generator. The sample structure was consistent with the sampled samples, and the number of samples in the high-deviation and low-deviation regions was evenly distributed according to the grid point ratio, generating a total of 500 minute-level samples.

[0046] Based on the minute-level samples, the rainfall forecast product is used as an additional input to the decoder of the basic forecast model. The basic forecast model is trained based on the additional input samples, and the CSI score is calculated to obtain the optimized model and the minute-level short-term rainfall forecast product.

[0047] In practical evaluation, the 10-minute rainfall forecast products from the minute-level samples are mapped into feature vectors using a multilayer perceptron (MLP). The dimension of these vectors matches the 256-dimensional historical hidden state output by the encoder, and feature fusion is performed using a gating calculation formula. The activation function is sigmoid, the gating layer weight matrix is ​​256×256, and the bias vector is 256-dimensional. The fused result is used as additional input to the decoder. The base forecast model is trained based on the updated additional input samples, with the learning rate decaying to 5e-4 and the batch size remaining at 4. After each training round, the CSI score of the forecast results is calculated on the test set at a rate of 0.1 mm / 10 min. To determine the precipitation threshold, the number of hits (TP), false alarms (FP), and missed alarms (FN) were statistically analyzed. Based on the CSI scoring formula CSI = TP / (TP + FP + FN), the time-averaged CSI, spatial-averaged CSI, and global CSI were calculated simultaneously to obtain an optimized model. Based on the optimized model, minute-level short-term rainfall forecasts for the next 0-2 hours were output, gridded every 10 minutes. The forecast products include a rainfall numerical grid field and a precipitation distribution map. The average absolute deviation between the forecast and measured rainfall decreased from 0.8 mm / 10 min to 0.3 mm / 10 min, and the predicted precipitation distribution was highly consistent with the actual precipitation. The global CSI... The CSI score is 2.5% higher than existing rainfall forecast products, and the forecast quality does not decrease significantly over time. The average CSI score for the next 2 hours remains above 0.65. The optimized model improves its ability to forecast severe convective precipitation in urban areas, and the spatial average CSI score for high-bias areas is improved by 3.2%. The average relative deviation of the optimized model for light rain (<0.5mm / 10min) scenarios is reduced from 45% to 22%, and the average absolute deviation for heavy rain (≥5mm / 10min) scenarios is reduced from 1.2mm / 10min to 0.4mm / 10min, meeting the forecasting needs of actual operations such as flood control scheduling and emergency response.

[0048] In this embodiment, the method for obtaining the original data includes:

[0049] The system acquires raw radar echo observation data, minute-level measured rainfall from ground rain gauges, and rainfall forecast products within a preset area. It removes non-precipitation echo interference data and outliers from the raw radar echo observation data to obtain radar data. It removes invalid zero values ​​based on measured rainfall and completes short-term missing data through interpolation with neighboring stations. It matches the spatiotemporal range of the rainfall forecast products with the preset area and removes forecast data with misaligned spatiotemporal ranges.

[0050] The raw data is obtained by using the radar data, measured rainfall, and rainfall forecast products in a unified geographic coordinate system and timestamp format.

[0051] In this embodiment, the method for obtaining the grid data includes:

[0052] Based on the raw data, the radar combined reflectivity factor is linearly normalized according to its physical value range. The measured rainfall and rainfall forecast products are normalized using the extreme precipitation thresholds for a preset area to obtain gridded data for the rainfall analysis field. The time decomposition formula for the gridded data is as follows:

[0053] ;

[0054] ;

[0055] in For the first Spatial coordinates within a 10-minute time period The spatial average value of the radar combined reflectivity factor in the corresponding region. The number of grid rows for the preset sub-region. The number of grid columns for the preset sub-region. This refers to the row offset of the grid points within the preset sub-region. This refers to the column offset of the grid points within a preset sub-region. Y is the baseline x-axis, and Y is the baseline y-axis. This is the normalized radar composite reflectivity factor value. For spatial grid points No. Rainfall forecast values ​​for 10-minute intervals, Spatial coordinates The 1-hour cumulative rainfall forecast value at the location, For the first A timestamp for a 10-minute time interval.

[0056] In this embodiment, the method for obtaining the basic forecast model includes:

[0057] Based on the grid data, the radar combined reflectivity factor grid data and rainfall grid data of a fixed length period are used as historical observation input through a non-sliding window. The rainfall grid data of continuous periods outside the historical observation period are used as supervised learning labels to obtain sampling samples. Based on the sampling samples, they are randomly divided into training set, validation set and test set according to the time dimension.

[0058] A deep learning model with an encoding and decoding structure is constructed based on the training set to obtain a basic forecast model. The encoder and decoder are memory units of the same type, including ConvLSTM and ConvGRU. The memory units are stacked in two layers. Deep learning is performed on the basic forecast model based on the training set, and the model parameters are updated according to the mean squared error loss function until the validation set loss does not decrease for three consecutive rounds, thus obtaining the trained basic forecast model.

[0059] In this embodiment, the method for obtaining the rainfall deviation includes:

[0060] Based on the rainfall forecast product, the forecast period is obtained. According to the forecast period, the grid point data of the rainfall analysis field in different forecast period lengths are randomly selected and input into the basic forecast model to obtain the predicted rainfall at each grid point. The predicted rainfall and the measured rainfall are compared to calculate the rainfall deviation, which includes absolute deviation and relative deviation.

[0061] Based on the rainfall deviation, the deviation area is identified, and the sampling frequency of the high deviation area is increased to twice to obtain the adjusted sample generation parameters. Based on the sample generation parameters, the corresponding area is evenly distributed into minute-level samples by a preset sample generator. The number of samples in the minute-level samples is evenly distributed between the high deviation area and the low deviation area according to the grid point ratio.

[0062] In this embodiment, the method for optimizing the model includes:

[0063] Rainfall forecast products are embedded into feature vectors based on minute-level samples. These feature vectors are then adaptively fused with the spatiotemporal features output by the encoder. The fusion result is used as an additional input to the decoder. The gating calculation formula for the adaptive fusion is as follows:

[0064] ;

[0065] ;

[0066] in For the first Adaptive gating weights at each time step It is the Sigmoid activation function. This is the weight matrix of the gate layer. This refers to the historical hidden state output by the encoding end. Forecast characteristics of rainfall forecast products This is the bias vector of the gated layer. No. Hidden state of the decoding end at each time step For the first Hidden state of the decoding end at each time step For Hadamah accumulation;

[0067] Based on the updated basic forecast model trained with additional input samples, the difference between the updated predicted rainfall and the supervised learning labels is calculated according to the loss function, and the learnable parameters of the model are updated by backpropagating gradients through the optimizer. After each round of training, the CSI score of the prediction results is calculated on the sample test set until the CSI score improves by less than 1% for three consecutive rounds, thus obtaining the optimized model. Based on the optimized model, minute-level short-term rainfall forecast products are output.

[0068] In this embodiment, the method for obtaining the CSI score includes:

[0069] Based on the prediction results and the actual measured rainfall, the number of grid time pairs where precipitation was predicted and actually occurred is taken as the hit count, the number of grid time pairs where precipitation was predicted but not actually occurred is taken as the false alarm count, and the number of grid time pairs where precipitation was predicted but not actually occurred is taken as the missed alarm count. The CSI score is calculated based on the hit count, false alarm count, and missed alarm count. The CSI score includes time-averaged CSI, spatial-averaged CSI, and global CSI. The time-averaged CSI is the average of the CSI scores for a single time step, and the spatial-averaged CSI is the average of the CSI scores for a single grid point.

[0070] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A method for minute-level short-term rainfall forecasting based on deep learning, characterized in that, Includes the following steps: Acquire radar data, measured rainfall, and rainfall forecast products for a preset area; perform quality control on the radar data, measured rainfall, and rainfall forecast products to obtain raw data; Based on the original data, the radar combined reflectivity factor, measured rainfall and rainfall forecast products are normalized to obtain gridded data of the rainfall analysis field. Based on the grid data, non-sliding window sampling samples are extracted, and deep learning is performed on the sampling samples to obtain a basic prediction model with an encoding and decoding structure. Based on the prediction period of the rainfall forecast product, the predicted rainfall and the measured rainfall of the basic forecast model are compared, and minute-level samples are generated equally in the corresponding area according to the rainfall deviation through a preset sample generator. Based on the minute-level samples, the rainfall forecast product is used as an additional input to the decoder of the basic forecast model. The basic forecast model is trained based on the additional input samples, and the CSI score is calculated to obtain the optimized model and the minute-level short-term rainfall forecast product.

2. The method for minute-level short-term rainfall forecasting based on deep learning according to claim 1, characterized in that, The method for obtaining the raw data includes: The system acquires raw radar echo observation data, minute-level measured rainfall from ground rain gauges, and rainfall forecast products within a preset area. It removes non-precipitation echo interference data and outliers from the raw radar echo observation data to obtain radar data. It removes invalid zero values ​​based on measured rainfall and completes short-term missing data through interpolation with neighboring stations. It matches the spatiotemporal range of the rainfall forecast products with the preset area and removes forecast data with misaligned spatiotemporal ranges. The raw data is obtained by using the radar data, measured rainfall, and rainfall forecast products in a unified geographic coordinate system and timestamp format.

3. The method for minute-level short-term rainfall forecasting based on deep learning according to claim 1, characterized in that, A method for obtaining the grid data includes: Based on the raw data, the radar combined reflectivity factor is linearly normalized according to its physical value range. The measured rainfall and rainfall forecast products are normalized using the extreme precipitation thresholds for a preset area to obtain gridded data for the rainfall analysis field. The time decomposition formula for the gridded data is as follows: ； ； in For the first Spatial coordinates within a 10-minute time period The spatial average value of the radar combined reflectivity factor in the corresponding region. The number of grid rows for the preset sub-region. The number of grid columns for the preset sub-region. This refers to the row offset of the grid points within the preset sub-region. This refers to the column offset of the grid points within a preset sub-region. Y is the baseline x-axis, and Y is the baseline y-axis. This is the normalized radar composite reflectivity factor value. For spatial grid points No. Rainfall forecast values ​​for 10-minute intervals, Spatial coordinates The 1-hour cumulative rainfall forecast value at the location, For the first A timestamp for a 10-minute time interval.

4. The method for minute-level short-term rainfall forecasting based on deep learning according to claim 1, characterized in that, The method for obtaining the basic forecast model includes: Based on the grid data, the radar combined reflectivity factor grid data and rainfall grid data of a fixed length period are used as historical observation input through a non-sliding window. The rainfall grid data of continuous periods outside the historical observation period are used as supervised learning labels to obtain sampling samples. Based on the sampling samples, they are randomly divided into training set, validation set and test set according to the time dimension. A deep learning model with an encoding and decoding structure is constructed based on the training set to obtain a basic forecast model. The encoder and decoder are memory units of the same type, including ConvLSTM and ConvGRU. The memory units are stacked in two layers. Deep learning is performed on the basic forecast model based on the training set, and the model parameters are updated according to the mean squared error loss function until the validation set loss does not decrease for three consecutive rounds, thus obtaining the trained basic forecast model.

5. The method for minute-level short-term rainfall forecasting based on deep learning according to claim 1, characterized in that, The method for obtaining the rainfall deviation includes: Based on the rainfall forecast product, the forecast period is obtained. According to the forecast period, the grid point data of the rainfall analysis field in different forecast period lengths are randomly selected and input into the basic forecast model to obtain the predicted rainfall at each grid point. The predicted rainfall and the measured rainfall are compared to calculate the rainfall deviation, which includes absolute deviation and relative deviation. Based on the rainfall deviation, the deviation area is identified, and the sampling frequency of the high deviation area is increased to twice to obtain the adjusted sample generation parameters. Based on the sample generation parameters, the corresponding area is evenly distributed into minute-level samples by a preset sample generator. The number of samples in the minute-level samples is evenly distributed between the high deviation area and the low deviation area according to the grid point ratio.

6. The method for minute-level short-term rainfall forecasting based on deep learning according to claim 1, characterized in that, The method for optimizing the model includes: Rainfall forecast products are embedded into feature vectors based on minute-level samples. These feature vectors are then adaptively fused with the spatiotemporal features output by the encoder. The fusion result is used as an additional input to the decoder. The gating calculation formula for the adaptive fusion is as follows: ； ； in For the first Adaptive gating weights at each time step It is the Sigmoid activation function. This is the weight matrix of the gate layer. This refers to the historical hidden state output by the encoding end. Forecast characteristics of rainfall forecast products, This is the bias vector of the gated layer. No. Hidden state of the decoding end at each time step For the first Hidden state of the decoding end at each time step For Hadamah accumulation; Based on the updated basic forecast model trained with additional input samples, the difference between the updated predicted rainfall and the supervised learning labels is calculated according to the loss function, and the learnable parameters of the model are updated by backpropagating gradients through the optimizer. After each round of training, the CSI score of the prediction results is calculated on the sample test set until the CSI score improves by less than 1% for three consecutive rounds, thus obtaining the optimized model. Based on the optimized model, minute-level short-term rainfall forecast products are output.

7. The method for minute-level short-term rainfall forecasting based on deep learning according to claim 1, characterized in that, The method for obtaining the CSI score includes: Based on the prediction results and the actual measured rainfall, the number of grid time pairs where precipitation was predicted and actually occurred is taken as the hit count, the number of grid time pairs where precipitation was predicted but not actually occurred is taken as the false alarm count, and the number of grid time pairs where precipitation was predicted but not actually occurred is taken as the missed alarm count. The CSI score is calculated based on the hit count, false alarm count, and missed alarm count. The CSI score includes time-averaged CSI, spatial-averaged CSI, and global CSI. The time-averaged CSI is the average of the CSI scores for a single time step, and the spatial-averaged CSI is the average of the CSI scores for a single grid point.

Images