A high-precision nowcasting method for precipitation in complex geographical environment

Abstract

The application discloses a high-precision near precipitation prediction method in a complex geographical environment and relates to the technical field of meteorological data processing. Precipitation data and various target variable data corresponding to a target period of a target region are obtained, a standard multi-source data sequence is formed through weighted fusion, constraint interpolation, data quality weighting and standardization processing, the standard multi-source data sequence is divided to form a standard multi-source sample set, a to-be-trained model containing a space-time feature extraction module, a reconstruction prediction module and a post-processing and refinement module is constructed, the standard multi-source sample set is used for training, a space-time attention precipitation prediction network is obtained, the standard multi-source data sequence to be predicted is input into the trained space-time attention precipitation prediction network, and a high-resolution precipitation prediction sequence corresponding to a target near period is obtained. The application can fully fuse local and global information, improve the space-time modeling capability and the space reconstruction capability, and realize accurate prediction of near multi-time-step precipitation.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, specifically to a high-precision nowcasting method for precipitation in complex geographical environments. Background Technology

[0002] In the field of meteorological forecasting, nowcasting (usually referring to short-term precipitation forecasts within 0–3 hours) is of great significance for urban disaster prevention and mitigation, rainstorm warnings, and flood management. However, precipitation is characterized by its suddenness, locality, and highly nonlinear evolution. Traditional physical or statistical models often exhibit limited prediction accuracy, insufficient spatial resolution, and insensitivity to abrupt weather changes in nowcasting. Furthermore, relying on a single data source is insufficient to fully capture the potential mechanisms of precipitation occurrence and development.

[0003] Meanwhile, breakthroughs in deep learning technology in image processing and time series modeling have brought new solutions to weather forecasting. However, how to effectively apply these technologies to weather forecasting, especially while maintaining physical consistency and interpretability while fusing multi-source heterogeneous data, remains a challenge for current research.

[0004] Therefore, there is an urgent need to propose a precipitation forecasting method that can integrate multi-source meteorological information and incorporate advanced deep learning technologies such as wavelet analysis, multi-scale convolution, and attention mechanisms. This method would improve the spatial detail and temporal coherence of precipitation forecasts, break through the bottlenecks of traditional methods in responding to extreme weather, and thus achieve significant improvements in forecast accuracy, resolution, and practicality, meeting the urgent needs of modern refined meteorological services. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a high-precision nowcasting method for precipitation in complex geographical environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A high-precision nowcasting method for precipitation in complex geographical environments comprises the following steps S1 to S3: constructing a spatiotemporal attention precipitation prediction network for predicting nowcast precipitation, and generating a high-resolution precipitation prediction sequence corresponding to the target nowcast period in step A.

[0008] Step S1: Obtain precipitation data and target variable data for the target area and the target time period. The precipitation data and target variable data are all single-source data types. Based on the target time resolution and target spatial resolution, align the time resolution and spatial resolution of each single-source data. Further perform data quality weighting and standardization on each single-source data, and then form a standard multi-source data sequence from the standardized single-source data.

[0009] Step S2: Divide the standard multi-source data sequence using a sliding time window to form various standard multi-source data sub-sequences; take the standard multi-source data sub-sequences as samples, each sample includes a basic multi-source data sequence and a target multi-source data sequence, and the last time step of the basic multi-source data sequence is earlier than the first time step of the target multi-source data sequence, thereby forming a standard multi-source sample set;

[0010] Step S3: Construct a training model that includes a spatiotemporal feature extraction module, a reconstruction prediction module, and a post-processing and refinement module; Based on a standard multi-source sample set and combined with a preset batch size, use the basic multi-source data sequence in the sample as input and the target precipitation data in the target multi-source data sequence in the sample as output to train the training model and obtain a spatiotemporal attention precipitation prediction network.

[0011] Step A: Use step S1 to generate a standard multi-source data sequence to be predicted, and input the standard multi-source data sequence to be predicted into the trained spatiotemporal attention precipitation prediction network to obtain a high-resolution precipitation prediction sequence corresponding to the target's immediate time period.

[0012] Furthermore, for each single-source data, step S1 performs temporal and spatial resolution alignment processing as follows:

[0013] Step S1.1: Divide the target region into a regular spatial grid according to the target spatial resolution. Calculate the horizontal spatial gradient of the single-source data at each spatial grid point using the following formula. Then, by integrating the horizontal spatial gradients of each spatial grid point, form the spatial gradient field of the single-source data with respect to the target region:

[0014] ;

[0015] in, For horizontal spatial gradient, For the longitude of the spatial grid points, The latitude of the spatial grid points For the partial derivative of the single-source data in the longitude direction, This represents the partial derivative of the single-source data in the latitudinal direction. For transpose;

[0016] Step S1.2: Generate the theoretical gradient field of the target region based on the target spatial resolution, and then weight and fuse the spatial gradient field of the single-source data with the theoretical gradient field according to the following formula to generate the fused gradient field corresponding to the single-source data. Further, constrain and interpolate the fused gradient field to the target spatial resolution to generate the spatially aligned data of the single-source data with respect to the target region:

[0017] ;

[0018] ;

[0019] in, To fuse gradient fields, For preset weighting coefficients, For the spatial gradient field of single-source data, For the theoretical gradient field, Spatially aligned data; For gradient-preserving interpolation functions, These are the height and width of the target spatial resolution, respectively;

[0020] Step S1.3: Based on the spatially aligned data, align the temporal resolution of the single-source data with the target temporal resolution using the following formula to generate spatiotemporally aligned data of the single-source data about the target region:

[0021] ;

[0022] in, For spatiotemporally aligned data, It is a cubic interpolation function. It is a linear interpolation function. For the target time resolution, The absolute value of the rate of change of precipitation intensity. A preset threshold is used to distinguish between drastic and steady changes in precipitation.

[0023] Furthermore, based on the spatiotemporally aligned data, step S1 performs data quality weighting and standardization on each single-source data according to the following formula:

[0024] ;

[0025] ;

[0026] ;

[0027] in, For standardized single-source data, For the first Spatiotemporal alignment data of spatial grid points This is the quality-weighted average of the spatiotemporally aligned data across the spatial grid. For the first Integrity weight of each spatial grid point For the first The number of time steps when a spatial grid point is missing. For the first The total number of time steps for each spatial grid point This represents the total number of spatial grid points.

[0028] Furthermore, the input end of the spatiotemporal feature extraction module constitutes the input end of the spatiotemporal attention precipitation prediction network, and the output end of the spatiotemporal feature extraction module is sequentially connected in series with the reconstruction prediction module and the post-processing and refinement module. The output end of the post-processing and refinement module constitutes the output end of the spatiotemporal attention precipitation prediction network.

[0029] The spatiotemporal feature extraction module is used to receive basic multi-source data sequences of a preset batch size and generate multi-source fusion features corresponding to each basic multi-source data sequence; the reconstruction prediction module is used to merge the multi-source fusion features along the time dimension and generate refined features, i.e., the initial precipitation prediction sequence for the target's near time period; the post-processing and refinement module is used to correct the initial precipitation prediction sequence and generate deviation-corrected refined features, i.e., the high-resolution precipitation prediction sequence corresponding to the target's near time period.

[0030] Furthermore, the spatiotemporal feature extraction module includes a feature extraction module, a splicing enhancement module, and an adaptive fusion module;

[0031] The input end of the feature extraction module constitutes the input end of the spatiotemporal feature extraction module, and the output end of the feature extraction module is sequentially connected in series with the splicing enhancement module and the adaptive fusion module. The output end of the adaptive fusion module constitutes the output end of the spatiotemporal feature extraction module.

[0032] For each basic multi-source data sequence, the feature extraction module uses wavelet decomposition and 3D convolution to extract features from each basic single-source data sequence, generating multi-scale single-source features corresponding to each basic single-source data sequence at different time frequencies. The adaptive filtering module concatenates each multi-scale single-source feature along the channel dimension and calculates global channel attention and local spatial attention. It further adjusts the global channel attention and local spatial attention with precipitation gradient adaptive weights and generates attention-enhanced features corresponding to the basic multi-source data sequence after element-wise multiplication weighting. The adaptive fusion module uses multi-granularity adaptive gating to collaboratively filter key features in the attention-enhanced features through channel and spatiotemporal dual-dimensional gating, generating multi-source fusion features corresponding to the basic multi-source data sequence.

[0033] Furthermore, the reconstruction prediction module includes a spatiotemporal evolution feature extraction module, a semantic feature extraction module, and a preliminary prediction module. The spatiotemporal evolution feature extraction module includes a preliminary encoding module, a bi-branch convolutional learning module, and a concatenation module. The bi-branch convolutional learning module includes a depthwise separable convolutional module and a grouped convolutional module.

[0034] The input of the preliminary encoding module constitutes the input of the reconstruction prediction module. The inputs of the depthwise separable convolution module and the group convolution module are connected to the output of the preliminary encoding module. The outputs of the depthwise separable convolution module and the group convolution module constitute the output of the splicing module. The output of the splicing module is connected in series with the semantic feature extraction module and the preliminary prediction module. The output of the preliminary prediction module constitutes the output of the reconstruction prediction module.

[0035] The preliminary encoding module receives multi-source fusion features of a preset batch size and merges them along the time dimension. It then generates preliminary encoded features through 2D convolution, group normalization, and ReLU activation. The depthwise separable convolution module and the group convolution module simultaneously receive the preliminary encoded features and extract the depthwise separable convolution features and the group convolution features. The stitching module stitches the depthwise separable convolution features and the group convolution features along the channel dimension to generate precipitation spatiotemporal evolution features.

[0036] The semantic feature extraction module is used to receive the spatiotemporal evolution features of precipitation, extract the output features corresponding to different levels through an adaptive structure with multi-size convolutional kernels, and generate spliced ​​and compressed features after pooling dimensionality reduction, pointwise convolution to adjust channels, and splicing and compression. Then, the spliced ​​and compressed features are residually fused with the spatiotemporal evolution features of precipitation to generate high-level semantic features.

[0037] The preliminary prediction module is used to receive high-level semantic features, extract spatial reconstruction features through 2D convolution, group normalization, and LeakyReLU activation, and then perform upsampling, convolutional decoding, group normalization, and LeakyReLU processing on the spatial reconstruction features to generate refined features.

[0038] Furthermore, the post-processing and refinement module includes a filtering module, a sliding window averaging module, and a deviation correction module;

[0039] The input terminal of the filtering module constitutes the input terminal of the post-processing and refinement module, and the output terminal of the filtering module is connected in series with the sliding window averaging module and the deviation correction module. The output terminal of the deviation correction module constitutes the output terminal of the post-processing and refinement module.

[0040] The filtering module is used to receive refined features and generate filtered refined features through adaptive weighting processing based on gradient graphs; the sliding window averaging module is used to perform sliding window averaging on the filtered refined features, correct abrupt frame changes, and generate time-series consistency-corrected refined features; the deviation correction module is used to correct the time-series consistency-corrected refined features based on radar meteorological data and generate deviation-corrected refined features.

[0041] The beneficial effects of adopting the above technical solution are as follows:

[0042] (1) This invention constructs a theoretical gradient field through spatial gradient analysis and combines gradient-preserving interpolation technology to achieve unified spatial resolution while ensuring physical continuity. It also improves the accuracy and adaptability of time alignment through cubic and linear interpolation. At the same time, this invention adopts a quality weighting strategy to minimize the interference of low-quality data on the overall feature distribution, improve the stability and accuracy of modeling, and enhance the generalization ability and robustness of the model to complex meteorological scenarios.

[0043] (2) This invention effectively extracts the multi-scale features of precipitation at different time frequencies through wavelet decomposition and 3D convolution, thereby improving the modeling ability for sudden and intermittent precipitation.

[0044] (3) This invention introduces global channel attention and local spatial attention mechanisms to achieve dynamic perception of the importance of multi-source features, enhance the expressive ability of key regions, and further improve the model's response to regions with drastic precipitation changes by combining adaptive weighting guided by precipitation gradient.

[0045] (4) This invention fully integrates local and global information, improves the spatiotemporal modeling capability and spatial reconstruction capability, and can accurately output near-term multi-time step precipitation prediction sequences, thereby improving the resolution and reliability of short-term precipitation forecasts. Attached Figure Description

[0046] Figure 1 This is a flowchart of the present invention;

[0047] Figure 2 This is a structural diagram of the semantic feature extraction module of the present invention;

[0048] Figure 3 This is a spatial distribution diagram of the root mean square error of the present invention and the comparative model regarding the prediction results of near-term precipitation.

[0049] Figure 4 This is a comparison chart of the precipitation prediction results of the model of this invention and various comparative models within 3 hours from the target prediction time;

[0050] Figure 5 This is a comparison chart of the threat score changes between the model of this invention and various comparative models;

[0051] Figure 6 This is a comparison chart of the deviation changes between the model of this invention and various comparative models. Detailed Implementation

[0052] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] refer to Figure 1A high-precision nowcasting method for precipitation in complex geographical environments is proposed. The method involves constructing a spatiotemporal attention-based precipitation prediction network for predicting nowcast precipitation according to steps S1 to S3, and generating a high-resolution precipitation prediction sequence corresponding to the target nowcast period according to step A.

[0054] Step S1: The target area is defined as the region with a geographical range of 97°-106.6°E and 20.4°-30°N. Precipitation data, topographic data, and reanalysis wind field data for the target time period are obtained for the target area. All the obtained precipitation data, topographic data, and reanalysis wind field data are single-source data types. The target time resolution is set to 0.5 hours and the target spatial resolution is set to 0.1°×0.1°. The time and spatial resolution of each single-source data are aligned. The data quality of each single-source data is further weighted and standardized. Then, a standard multi-source data sequence is formed from the standardized single-source data.

[0055] Step S2: Based on the preset time series length, the standard multi-source data sequence is divided using a sliding time window to form various standard multi-source data sub-sequences; each standard multi-source data sub-sequence is used as a sample, and each sample includes a basic multi-source data sequence and a target multi-source data sequence, and the last time step of the basic multi-source data sequence is earlier than the first time step of the target multi-source data sequence, thereby forming a standard multi-source sample set;

[0056] Step S3: Construct a training model that includes a spatiotemporal feature extraction module, a reconstruction prediction module, and a post-processing and refinement module; Based on a standard multi-source sample set and combined with a preset batch size, use the basic multi-source data sequence in the sample as input and the target precipitation data in the target multi-source data sequence in the sample as output to train the training model and obtain a spatiotemporal attention precipitation prediction network.

[0057] Step A: Use step S1 to generate a standard multi-source data sequence to be predicted, and input the standard multi-source data sequence to be predicted into the trained spatiotemporal attention precipitation prediction network to obtain a high-resolution precipitation prediction sequence corresponding to the target's immediate time period.

[0058] Furthermore, for each single-source data, step S1 performs temporal and spatial resolution alignment processing as follows:

[0059] Step S1.1: Divide the target region into a 96×96 spatial grid according to the target spatial resolution. Calculate the horizontal spatial gradient of the single-source data at each spatial grid point using the following formula. Then, by integrating the horizontal spatial gradients of each spatial grid point, form the spatial gradient field of the single-source data with respect to the target region:

[0060] ;

[0061] in, For horizontal spatial gradient, For the longitude of the spatial grid points, The latitude of the spatial grid points For the partial derivative of the single-source data in the longitude direction, This represents the partial derivative of the single-source data in the latitudinal direction. For transpose;

[0062] Step S1.2: Generate the theoretical gradient field of the target region based on the target spatial resolution, and then weight and fuse the spatial gradient field of the single-source data with the theoretical gradient field according to the following formula to generate the fused gradient field corresponding to the single-source data. Further, constrain and interpolate the fused gradient field to the target spatial resolution to generate the spatially aligned data of the single-source data with respect to the target region:

[0063] ;

[0064] ;

[0065] in, To fuse gradient fields, For preset weighting coefficients, For the spatial gradient field of single-source data, For the theoretical gradient field, Spatially aligned data; For gradient-preserving interpolation functions, These are the height and width of the target spatial resolution, respectively;

[0066] Step S1.3: Based on the spatially aligned data, align the temporal resolution of the single-source data with the target temporal resolution using the following formula to generate spatiotemporally aligned data of the single-source data about the target region:

[0067] ;

[0068] in, For spatiotemporally aligned data, It is a cubic interpolation function. It is a linear interpolation function. For the target time resolution, It is the absolute value of the rate of change of precipitation intensity, calculated by dividing the difference in precipitation intensity between adjacent moments by the time interval; A preset threshold is used to distinguish between drastic and steady changes in precipitation.

[0069] Furthermore, based on spatiotemporally aligned data, integrity weights are generated according to the data quality of each spatial grid point. Through weighted calculation, a quality-weighted average reflecting the dominant trend of high-quality data and a weighted standard deviation reflecting the dispersion of high-quality data are obtained. This transforms the data of each spatial grid point into a fluctuation value at a uniform scale that retains the quality-weighted characteristics, thus completing data quality weighting and standardization. The formula is shown below:

[0070] ;

[0071] ;

[0072] ;

[0073] in, For standardized single-source data, For the first Spatiotemporal alignment data of spatial grid points This is the quality-weighted average of the spatiotemporally aligned data across the spatial grid. For the first Integrity weight of each spatial grid point For the first The number of time steps when a spatial grid point is missing. For the first The total number of time steps for each spatial grid point The total number of spatial grid points. This is the weighted standard deviation of all spatial grid point data in the spatiotemporally aligned data.

[0074] Furthermore, the input end of the spatiotemporal feature extraction module constitutes the input end of the spatiotemporal attention precipitation prediction network, and the output end of the spatiotemporal feature extraction module is sequentially connected in series with the reconstruction prediction module and the post-processing and refinement module. The output end of the post-processing and refinement module constitutes the output end of the spatiotemporal attention precipitation prediction network.

[0075] The spatiotemporal feature extraction module is used to receive basic multi-source data sequences of a preset batch size and generate multi-source fusion features corresponding to each basic multi-source data sequence; the reconstruction prediction module is used to merge the multi-source fusion features along the time dimension and generate refined features, i.e., the initial precipitation prediction sequence for the target's near time period; the post-processing and refinement module is used to correct the initial precipitation prediction sequence and generate deviation-corrected refined features, i.e., the high-resolution precipitation prediction sequence corresponding to the target's near time period.

[0076] Furthermore, the spatiotemporal feature extraction module includes a feature extraction module, a splicing enhancement module, and an adaptive fusion module;

[0077] The input end of the feature extraction module constitutes the input end of the spatiotemporal feature extraction module, and the output end of the feature extraction module is sequentially connected in series with the splicing enhancement module and the adaptive fusion module. The output end of the adaptive fusion module constitutes the output end of the spatiotemporal feature extraction module.

[0078] For each basic multi-source data sequence, the feature extraction module uses wavelet decomposition and 3D convolution to extract features from each basic single-source data sequence, generating multi-scale single-source features corresponding to each basic single-source data sequence at different time frequencies. The adaptive filtering module concatenates the multi-scale single-source features corresponding to each basic single-source data sequence along the channel dimension, calculates global channel attention and local spatial attention, and further adjusts the global channel attention and local spatial attention with precipitation gradient adaptive weights. After element-wise multiplication and weighting, attention-enhanced features corresponding to the basic multi-source data sequence are generated. The adaptive fusion module uses multi-granularity adaptive gating to collaboratively filter key features in the attention-enhanced features through channel and spatiotemporal dual-dimensional gating, generating multi-source fusion features corresponding to the basic multi-source data sequence.

[0079] Specifically, the adaptive filtering module calculates the global channel attention and local spatial attention according to the following expressions:

[0080] ;

[0081] ;

[0082] in, For activation function, For multilayer perceptrons, For three-dimensional global average pooling, It is a 1×1×1 three-dimensional convolution. This is a 3×3×3 local max pooling. For splicing operations, As a precipitation characteristic, As a topographic feature, This refers to wind field characteristics.

[0083] Furthermore, the adaptive fusion module filters key features from the attention-enhancing features according to the following expression:

[0084] ;

[0085] in, As a feature of multi-source fusion, For element-wise multiplication, For attention enhancement features, To perform a 1×1×1 3D convolution and normalize it in the channel dimension, To perform a 3×3×3 three-dimensional convolution and normalize it in the spatiotemporal dimension, For Hadama accumulation.

[0086] Furthermore, the reconstruction prediction module includes a spatiotemporal evolution feature extraction module, a semantic feature extraction module, and a preliminary prediction module. The spatiotemporal evolution feature extraction module includes a preliminary encoding module, a bi-branch convolutional learning module, and a concatenation module. The bi-branch convolutional learning module includes a depthwise separable convolutional module and a grouped convolutional module.

[0087] The input of the preliminary encoding module constitutes the input of the reconstruction prediction module. The inputs of the depthwise separable convolution module and the group convolution module are connected to the output of the preliminary encoding module. The outputs of the depthwise separable convolution module and the group convolution module constitute the output of the splicing module. The output of the splicing module is connected in series with the semantic feature extraction module and the preliminary prediction module. The output of the preliminary prediction module constitutes the output of the reconstruction prediction module.

[0088] The preliminary encoding module receives multi-source fusion features of a preset batch size and merges them along the time dimension. It then generates preliminary encoded features through 2D convolution, group normalization, and ReLU activation. The depthwise separable convolution module and the group convolution module simultaneously receive the preliminary encoded features and extract the depthwise separable convolution features and the group convolution features. The stitching module stitches the depthwise separable convolution features and the group convolution features along the channel dimension to generate precipitation spatiotemporal evolution features.

[0089] The depthwise separable convolutional module is specifically:

[0090] ;

[0091] in, The output of the depthwise separable convolution module is the depthwise separable convolution feature. For pointwise convolution, For preliminary encoding features, To apply a 3×3 convolution kernel with a stride of 1 independently to each channel of the initial encoded features.

[0092] The grouped convolution module is specifically:

[0093] ;

[0094] in, This is the output of the grouped convolution module, i.e., the grouped convolution features; To divide the input channels of the initial encoded features into Each group is independently subjected to a 3×3 convolution with a stride of 1.

[0095] refer to Figure 2 The semantic feature extraction module is used to receive the spatiotemporal evolution features of precipitation, extract the output features corresponding to different levels through an adaptive structure with multi-size convolutional kernels, and generate spliced ​​compressed features after pooling dimensionality reduction, pointwise convolution to adjust channels, and splicing and compression. Then, the spliced ​​compressed features are residually fused with the spatiotemporal evolution features of precipitation to generate high-level semantic features.

[0096] Referring to the following expression, the adaptive structure containing multi-size convolutional kernels includes three nested processing modules connected in series. First, each layer extracts local features using 3×3 convolution, introduces non-linearity through ReLU activation, and the first two layers compress the dimensions through max pooling downsampling. Subsequently, each layer adjusts the channels using 1×1 pointwise convolution. Finally, the output features after channel adjustment are concatenated and fused along the channel dimension to generate concatenated compressed features.

[0097] ;

[0098] ;

[0099] ;

[0100]

[0101] in, The spatiotemporal evolution characteristics of precipitation, The output features of level 1 The output features of the second level, For the output features of the 3rd level, To perform a 3×3 convolution on the spatiotemporal evolution characteristics of precipitation, To perform non-linear activation on the convolution result, For max pooling operation, This is a 1×1 pointwise convolution operation. This is a splicing and compression feature.

[0102] The preliminary prediction module is used to receive high-level semantic features, extract spatial reconstruction features through 2D convolution, group normalization, and LeakyReLU activation, restore the spatial reconstruction features to a high-resolution form through upsampling operation, and then generate refined features by using convolutional decoding, group normalization, and LeakyReLU processing.

[0103] Furthermore, the post-processing and refinement module includes a filtering module, a sliding window averaging module, and a deviation correction module;

[0104] The input terminal of the filtering module constitutes the input terminal of the post-processing and refinement module, and the output terminal of the filtering module is connected in series with the sliding window averaging module and the deviation correction module. The output terminal of the deviation correction module constitutes the output terminal of the post-processing and refinement module.

[0105] The filtering module is used to receive refined features and generate filtered refined features through adaptive weighted processing based on gradient maps. Specifically, it first identifies edge regions with drastic precipitation changes and regions with gradual changes based on gradient intensity maps. Then, it performs targeted filtering processing, that is, it applies strong filtering to regions with gradual changes to smooth noise, and uses weak filtering to preserve the true boundaries of edge regions. Finally, it outputs filtered refined features with smoother space, clearer edges, and suppressed pseudo-surge values.

[0106] The sliding window averaging module is used to perform a sliding window averaging on the filtered and refined features according to the following expression, correct abrupt frames, and generate time-consistency corrected refined features.

[0107] Referring to the following expression, the deviation correction module is used to correct the temporal consistency correction refinement features based on the precipitation cloud structure information provided by radar meteorological data, thereby more accurately predicting precipitation intensity and location, and outputting the deviation correction refinement features:

[0108] ;

[0109] ;

[0110] in, This is the precipitation deviation. For deviation correction and refining features, For radar coupling correction network, Radar reflectivity characteristics, Refined features for timing consistency correction. These are features for filtering and refining.

[0111] The technical effects of the present invention will be further explained in detail below with reference to comparative experiments.

[0112] Using a simple video prediction model as a comparison model, Figure 3The figure shows the spatial distribution of the root mean square error (RMSE) of the present invention and the comparative model for near-term precipitation prediction. (a) shows the near-term precipitation prediction result obtained using the simple video prediction model; (b) shows the near-term precipitation prediction result obtained after replacing the encoder of the simple video prediction model with the spatiotemporal feature extraction module of the present invention; (c) shows the near-term precipitation prediction result obtained after introducing the reconstruction prediction module of the present invention into the simple video prediction model; and (d) shows the near-term precipitation prediction result obtained after replacing the encoder with the spatiotemporal feature extraction module and simultaneously introducing and reconstructing the prediction module into the simple video prediction model. As can be observed from the figure, after replacing the encoder with the spatiotemporal feature extraction module, the area with an RMSE exceeding 0.4 is significantly reduced, with the maximum value decreasing from approximately 0.65 to 0.45. After introducing the reconstruction prediction module into the simple video prediction model, the overall error continues to decrease, which is better than the scheme of simply replacing the encoder with the spatiotemporal feature extraction module. When both the spatiotemporal feature extraction module and the reconstruction prediction module are introduced simultaneously, the maximum RMSE is approximately 0.35, with the main area below 0.2. The results show that the spatiotemporal feature extraction module and reconstruction prediction module designed in this invention work together to improve the accuracy of regional precipitation forecasts.

[0113] Let the target prediction time be T, and let 0.5 hours be the time step. Figure 4 The figures show a comparison of precipitation prediction results between the proposed model and various comparative models over multiple time steps within three hours from the target prediction time. (a) is the reference baseline figure; (b) shows the prediction results of the weather research and forecasting model at different time steps; (c) shows the prediction results of the simple video prediction model at different time steps; (d) shows the prediction results of the convolutional long short-term memory network at different time steps; (e) shows the prediction results of the neural basis extended analysis time series model at different time steps; and (f) shows the prediction results of the spatiotemporal attention precipitation prediction network at different time steps. As can be observed from the figures, the precipitation prediction results of the proposed spatiotemporal attention precipitation prediction network best match the reference baseline figure, and compared to the classic weather research and forecasting model, the accuracy of precipitation prediction by the spatiotemporal attention precipitation prediction network is significantly improved at different time steps.

[0114] In the first two time steps, the spatiotemporal attention precipitation prediction network and the neural-based extended analysis time series model proposed in this invention showed relatively good forecasting performance. In the second time step, both predicted localized heavy rainfall, but only the spatiotemporal attention precipitation prediction network proposed in this invention predicted heavy rainstorms and some extremely heavy rainstorms. As the forecast lead time increased, the forecasting performance of each model decreased, but the spatiotemporal attention precipitation prediction network proposed in this invention maintained high forecasting accuracy, and its decline in forecasting performance was relatively gradual. Compared to other methods, the spatiotemporal attention precipitation prediction network proposed in this invention has a significant advantage in forecasting large-scale precipitation, and it can still accurately predict the occurrence of heavy rainstorms in the last two time steps.

[0115] Figure 5 and Figure 6 The graphs show the changes in threat scores and biases between the proposed model and various comparative models as the number of iterations increases. As can be seen from the graphs, the threat scores and biases of each model continuously improve with increasing iteration count. Compared to other models, the spatiotemporal attention precipitation prediction network proposed in this invention exhibits more stable prediction performance throughout the training process, and the model tends to converge after 25 iterations.

[0116] After training, the threat scores, biases, and root mean square errors of each model are shown in Table 1. WRF is the weather research and forecasting model, SimVP is the simple video prediction model, ConvLSTM is the convolutional long short-term memory network, and N-BEATS is the neural basis extended analysis time series model. Table 1 shows that the spatiotemporal attention precipitation prediction network proposed in this invention achieves the highest threat scores and biases of approximately 0.65 and 0.83, respectively, among all models. This indicates that the model designed using the method of this invention has more stable performance in nowcasting precipitation.

[0117] Table 1 Threat score, bias, and root mean square error for each model

[0118]

[0119] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A high-precision nowcasting method for precipitation in complex geographical environments, characterized in that, Construct a spatiotemporal attention precipitation prediction network for predicting precipitation in the near future according to steps S1 to S3 below, and generate a high-resolution precipitation prediction sequence corresponding to the target near future period according to step A: Step S1: Obtain precipitation data and target variable data for the target area and the target time period. The precipitation data and target variable data are all single-source data types. Based on the target time resolution and target spatial resolution, align the time resolution and spatial resolution of each single-source data. Further perform data quality weighting and standardization on each single-source data, and then form a standard multi-source data sequence from the standardized single-source data. Step S2: Divide the standard multi-source data sequence using a sliding time window to form various standard multi-source data sub-sequences; take the standard multi-source data sub-sequences as samples, each sample includes a basic multi-source data sequence and a target multi-source data sequence, and the last time step of the basic multi-source data sequence is earlier than the first time step of the target multi-source data sequence, thereby forming a standard multi-source sample set; Step S3: Construct a training model that includes a spatiotemporal feature extraction module, a reconstruction prediction module, and a post-processing and refinement module; Based on a standard multi-source sample set and combined with a preset batch size, use the basic multi-source data sequence in the sample as input and the target precipitation data in the target multi-source data sequence in the sample as output to train the training model and obtain a spatiotemporal attention precipitation prediction network. Step A: Use step S1 to generate a standard multi-source data sequence to be predicted, and input the standard multi-source data sequence to be predicted into the trained spatiotemporal attention precipitation prediction network to obtain a high-resolution precipitation prediction sequence corresponding to the target's immediate time period.

2. The high-precision nowcasting method for complex geographical environments according to claim 1, characterized in that, For each single source of data, step S1 performs temporal and spatial resolution alignment as follows: Step S1.1: Divide the target region into a regular spatial grid according to the target spatial resolution. Calculate the horizontal spatial gradient of the single-source data at each spatial grid point using the following formula. Then, by integrating the horizontal spatial gradients of each spatial grid point, form the spatial gradient field of the single-source data with respect to the target region: ； in, For horizontal spatial gradient, For the longitude of the spatial grid points, The latitude of the spatial grid points For the partial derivative of the single-source data in the longitude direction, This represents the partial derivative of the single-source data in the latitudinal direction. This is a transpose operation; Step S1.2: Generate the theoretical gradient field of the target region based on the target spatial resolution, and then weight and fuse the spatial gradient field of the single-source data with the theoretical gradient field according to the following formula to generate the fused gradient field corresponding to the single-source data. Further, constrain and interpolate the fused gradient field to the target spatial resolution to generate the spatially aligned data of the single-source data with respect to the target region: ； ； in, To fuse gradient fields, For preset weighting coefficients, For the spatial gradient field of single-source data, For the theoretical gradient field, Spatially aligned data; For gradient-preserving interpolation functions, These are the height and width of the target spatial resolution, respectively; Step S1.3: Based on the spatially aligned data, align the temporal resolution of the single-source data with the target temporal resolution using the following formula to generate spatiotemporally aligned data of the single-source data about the target region: ； in, For spatiotemporally aligned data, It is a cubic interpolation function. It is a linear interpolation function. For the target time resolution, The absolute value of the rate of change of precipitation intensity. A preset threshold is used to distinguish between drastic and steady changes in precipitation.

3. The high-precision nowcasting method for complex geographical environments according to claim 2, characterized in that, Based on the spatiotemporally aligned data, step S1 performs data quality weighting and standardization on each single-source data according to the following formula: ； ； ； in, For standardized single-source data, For the first Spatiotemporal alignment data of spatial grid points This is the quality-weighted average of the spatiotemporally aligned data across the spatial grid. For the first Integrity weight of each spatial grid point For the first The number of time steps when a spatial grid point is missing. For the first The total number of time steps for each spatial grid point This represents the total number of spatial grid points.

4. The high-precision nowcasting method for complex geographical environments according to claim 1, characterized in that, The input end of the spatiotemporal feature extraction module constitutes the input end of the spatiotemporal attention precipitation prediction network. The output end of the spatiotemporal feature extraction module is sequentially connected in series with the reconstruction prediction module and the post-processing and refinement module. The output end of the post-processing and refinement module constitutes the output end of the spatiotemporal attention precipitation prediction network. The spatiotemporal feature extraction module is used to receive basic multi-source data sequences of a preset batch size and generate multi-source fusion features corresponding to each basic multi-source data sequence; the reconstruction prediction module is used to merge the multi-source fusion features along the time dimension and generate refined features, i.e., the initial precipitation prediction sequence for the target's near time period; the post-processing and refinement module is used to correct the initial precipitation prediction sequence and generate deviation-corrected refined features, i.e., the high-resolution precipitation prediction sequence corresponding to the target's near time period.

5. The high-precision nowcasting method for complex geographical environments according to claim 4, characterized in that, The spatiotemporal feature extraction module includes a feature extraction module, a splicing and enhancement module, and an adaptive fusion module; The input end of the feature extraction module constitutes the input end of the spatiotemporal feature extraction module, and the output end of the feature extraction module is sequentially connected in series with the splicing enhancement module and the adaptive fusion module. The output end of the adaptive fusion module constitutes the output end of the spatiotemporal feature extraction module. For each basic multi-source data sequence, the feature extraction module is used to extract features from each basic single-source data sequence using wavelet decomposition and 3D convolution, generating multi-scale single-source features corresponding to each basic single-source data sequence at different time frequencies; the adaptive filtering module is used to concatenate each multi-scale single-source feature along the channel dimension, and calculate global channel attention and local spatial attention, further adjusting the global channel attention and local spatial attention with precipitation gradient adaptive weights, and generating attention-enhanced features corresponding to the basic multi-source data sequence after element-wise multiplication weighting; The adaptive fusion module is used to generate multi-source fusion features corresponding to basic multi-source data sequences by using multi-granularity adaptive gating to collaboratively filter key features in attention enhancement features through channel and spatiotemporal dual-dimensional gating.

6. The high-precision nowcasting method for complex geographical environments according to claim 5, characterized in that, The reconstruction prediction module includes a spatiotemporal evolution feature extraction module, a semantic feature extraction module, and a preliminary prediction module. The spatiotemporal evolution feature extraction module includes a preliminary encoding module, a bi-branch convolutional learning module, and a concatenation module. The bi-branch convolutional learning module includes a depthwise separable convolutional module and a grouped convolutional module. The input of the preliminary encoding module constitutes the input of the reconstruction prediction module. The inputs of the depthwise separable convolution module and the group convolution module are connected to the output of the preliminary encoding module. The outputs of the depthwise separable convolution module and the group convolution module constitute the output of the splicing module. The output of the splicing module is connected in series with the semantic feature extraction module and the preliminary prediction module. The output of the preliminary prediction module constitutes the output of the reconstruction prediction module. The preliminary encoding module receives multi-source fusion features of a preset batch size and merges them along the time dimension. It then generates preliminary encoded features through 2D convolution, group normalization, and ReLU activation. The depthwise separable convolution module and the group convolution module simultaneously receive the preliminary encoded features and extract the depthwise separable convolution features and the group convolution features. The stitching module stitches the depthwise separable convolution features and the group convolution features along the channel dimension to generate precipitation spatiotemporal evolution features. The semantic feature extraction module is used to receive the spatiotemporal evolution features of precipitation, extract the output features corresponding to different levels through an adaptive structure with multi-size convolutional kernels, and generate spliced ​​and compressed features after pooling dimensionality reduction, pointwise convolution to adjust channels, and splicing and compression. Then, the spliced ​​and compressed features are residually fused with the spatiotemporal evolution features of precipitation to generate high-level semantic features. The preliminary prediction module is used to receive high-level semantic features, extract spatial reconstruction features through 2D convolution, group normalization, and LeakyReLU activation, and then perform upsampling, convolutional decoding, group normalization, and LeakyReLU processing on the spatial reconstruction features to generate refined features.

7. The high-precision nowcasting method for complex geographical environments according to claim 6, characterized in that, The post-processing and refinement module includes a filtering module, a sliding window averaging module, and a deviation correction module. The input terminal of the filtering module constitutes the input terminal of the post-processing and refinement module, and the output terminal of the filtering module is connected in series with the sliding window averaging module and the deviation correction module. The output terminal of the deviation correction module constitutes the output terminal of the post-processing and refinement module. The filtering module is used to receive refined features and generate filtered refined features through adaptive weighting processing based on gradient graphs; the sliding window averaging module is used to perform sliding window averaging on the filtered refined features, correct abrupt frame changes, and generate time-series consistency-corrected refined features; the deviation correction module is used to correct the time-series consistency-corrected refined features based on radar meteorological data and generate deviation-corrected refined features.

Images