A multi-time short-impending precipitation prediction method based on time-space joint feature extraction

Abstract

The application discloses a multi-time short-impending precipitation prediction method based on time and space joint feature extraction, which comprises the following steps: through pretreated multi-dimensional meteorological data, an encoder based on a U-Net architecture is used for downsampling processing, and a time information embedding module is used for semantic embedding of input information, so that the sub-module has time perception capability; features after the downsampling processing are input into a Transform-Mamba hybrid module for efficiently extracting spatial and temporal features, and window division is used before each sub-module for parallel processing; a decoder is used for upsampling and restoring the feature data, and a jump connection is used for supplementing the lost information due to the downsampling, and finally, the features are input into a multi-layer perception machine for prediction to obtain the precipitation result of a target area. The method can effectively perform local high-precision precipitation prediction of multi-time in a short-impending time, realizes the complexity of a lightweight model, and has excellent prediction capability.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, specifically a multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction. Background Technology

[0002] Precipitation forecasting is crucial for disaster preparedness, urban planning, and public safety; timely and accurate forecasts can effectively reduce loss of life and property. Traditional numerical weather prediction relies on complex physical equations and numerical simulations, demonstrating high accuracy in medium- and long-term weather forecasts. However, due to its high computational complexity, numerical weather prediction faces real-time limitations in short-term nowcasting (typically 0-3 hours), especially in scenarios requiring high-resolution output, where computational resource demands increase significantly, making it difficult to meet the requirements of real-time applications.

[0003] In recent years, deep learning-based weather forecasting methods have emerged, utilizing neural networks and GPUs to accelerate computation and generate forecasts within seconds, thus mitigating some of the shortcomings of numerical weather prediction. For example, PredRNNv2 captures spatiotemporal features through recursive networks, while DGMR utilizes generative models to improve precipitation forecast accuracy. However, existing DLWP methods still face the following challenges: First, training independent models for each forecast lead time results in high computational costs and a lack of flexibility; second, insufficient modeling capability for high-resolution spatiotemporal features makes it difficult to simultaneously capture both local details and global dynamics of precipitation; and third, the balance between real-time performance and prediction accuracy still needs optimization, especially when processing large-scale meteorological datasets (such as radar reflectivity and ground station observation data), where the computational load is significant, limiting the widespread application of these models.

[0004] To address the aforementioned issues, there is an urgent need for an efficient and lightweight short-term precipitation forecasting method that can improve the accuracy and real-time performance of multi-time-dependent forecasts while reducing computational resource requirements. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction. Through the innovative design of a lightweight U-Net structure, a temporal embedding module, and a Transformer-Mamba hybrid feature extraction module, this invention achieves the goals of reducing computational complexity, supporting single-model multi-time-dependent prediction, improving real-time performance to meet minute-level response requirements, enhancing the ability to model high-resolution spatiotemporal features, and improving prediction accuracy.

[0006] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0007] A multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction includes the following steps:

[0008] Step 1: Collect historical meteorological data and preprocess it to generate an input tensor containing multi-dimensional information; the historical meteorological data includes radar data and station data.

[0009] Step 2: Feature extraction is performed using a lightweight encoder; the input end of the lightweight encoder is embedded with a time embedding module, which outputs a time-aware output tensor from the input tensor.

[0010] Step 3: Input the extracted features into the temporal-spatial feature extraction module to extract spatiotemporal features; the input end of the temporal-spatial extraction module is embedded with a time embedding module, and the features extracted in step 2 are input into the time embedding module to output time-aware features;

[0011] Step 4: The spatiotemporal features are used as input to obtain the restored spatiotemporal features through a lightweight decoder; the input end of the lightweight decoder is embedded with a time embedding module, the spatiotemporal features are input to the time embedding module and the output is time-aware spatiotemporal features, and the restored spatiotemporal features are output by a multilayer perceptron to predict the results;

[0012] The method by which the time embedding module embeds time information is as follows:

[0013] The input information of each module is classified according to the time information category. The classification results are then used to extract multi-dimensional features. The multi-dimensional features are then concatenated to obtain the total features.

[0014] The total features are used as input to obtain the time rate of change and bias value using two pre-trained multilayer perceptrons. The time rate of change and bias value are then added to the input information to obtain time-aware data, which is then used as the input to the next module.

[0015] Preferably, the multi-dimensional information includes multi-element meteorological information, time series, and spatial distribution information.

[0016] Preferably, the preprocessing method includes normalizing the data, filtering out noise, and filling in missing values.

[0017] Preferably, the time information categories include prediction timeliness information, time information with cyclical patterns, and time window information with historical weights. Prediction timeliness information refers to the future time steps the model needs to predict, such as prediction targets for the next day, week, or month. These discrete time steps (e.g., "1 day," "2 days") are converted into continuous feature vectors, called the first feature. The embedding layer maps the time steps to a vector space, enabling the model to understand and utilize the predicted time span. Time information with cyclical patterns refers to the periodic patterns in the time data, such as daily cycles (the progress of the current day within 24 hours), weekly cycles (the days of the week within 7 days), annual cycles (the progress of the current year within 365 days), or seasonal cycles (spring, summer, autumn, winter). These cyclical features are converted into values ​​in the [0,1] interval through normalization, resulting in the second feature. Time window information with historical weights refers to the past time steps contained in the historical data input to the model, such as data from the past 7 or 30 days, used to provide context for prediction. First, the window semantic information is obtained through the embedding layer. Then, the window semantic information is processed by the normalized exponential function and multiplied by a learnable weight parameter to obtain the third feature, thus obtaining three-dimensional features. This processing method allows the model to dynamically focus on more important parts of historical data. For example, recent data may be more predictive than distant data.

[0018] Preferably, the time-aware data = time change rate · input information + bias value.

[0019] Preferably, the spatiotemporal feature extraction module includes four spatial extraction sub-modules and four temporal extraction sub-modules connected in sequence; each of the spatial and temporal sub-modules has a temporal embedding module embedded at its input end, and the data input to each of the spatial and temporal sub-modules needs to be windowed.

[0020] Preferably, the input of the first-layer submodule is radar data, the input of the second-layer submodule is the output of the first-layer submodule and the data after splicing radar data and site data, and the input of the third-layer submodule is the output of the second-layer submodule.

[0021] Preferably, the spatiotemporal feature extraction module includes four spatial extraction sub-modules and four temporal extraction sub-modules connected in sequence; each of the spatial and temporal sub-modules is embedded with a temporal embedding module, and the data input to each of the spatial and temporal sub-modules needs to be windowed.

[0022] Preferably, the spatial extraction submodule processes the spatial distribution features from the input feature meteorological data; the temporal extraction submodule captures the long-term temporal dependencies from the input feature meteorological data and generates a causal temporal feature representation.

[0023] Preferably, the lightweight decoder is symmetrical to the lightweight encoder. The three-layer sub-modules in the lightweight decoder include residual network blocks and upsampling, and each has a time embedding module embedded at the input. Preferably, the upsampling uses transposed convolutional upsampling, and the downsampling in the three-layer sub-modules of the lightweight encoder is skip-connected to the corresponding three-layer residual network blocks in the lightweight decoder.

[0024] This invention has the following characteristics and beneficial effects:

[0025] This invention boasts significant innovations and advantages. By optimizing the U-Net structure and hybrid modules, the total number of parameters is controlled within 23.9MB, significantly reducing computational requirements and improving model scalability. The temporal embedding module enables a single model to adapt to multi-time-series forecasts, avoiding the need for separate model training for each time-series forecast. The Transformer-Mamba hybrid module combines the advantages of spatial global modeling and temporal series modeling, enhancing high-resolution feature capture capabilities. Minute-to-kilometer level predictions achieve accuracy comparable to state-of-the-art deep learning models while possessing rapid response capabilities. Through the above technical solutions, this invention effectively addresses the shortcomings of traditional methods in short-term precipitation forecasting, providing a novel approach for real-time meteorological applications. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of a multi-time-effect short-term precipitation prediction method based on joint temporal and spatial feature extraction according to the present invention.

[0027] Figure 2 This is a diagram illustrating the overall architecture of the time embedding module method of the present invention.

[0028] Figure 3 This is a line graph comparing the MAE index of this invention with other methods at six prediction timeframes.

[0029] Figure 4 This is a line graph comparing the CSI (1mm / h) index of this invention with other methods at six prediction lead times.

[0030] Figure 5 This is a line graph comparing the CSI (4mm / h) index of this invention with other methods at six prediction lead times.

[0031] Figure 6 This is a line graph comparing the CSI (8mm / h) index of this invention with other methods at six prediction lead times. Detailed Implementation

[0032] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0033] A multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction, such as... Figure 1 As shown, it includes the following steps:

[0034] Step 1: Collect historical meteorological data. By preprocessing the historical meteorological data, an input tensor containing multi-dimensional information is generated. The historical meteorological data includes radar data and station data.

[0035] Specifically, in this embodiment, the data used comes from a multi-radar / multi-sensor system and an automatic ground observation system. The multi-radar / multi-sensor system data includes radar precipitation rate and one-hour quantitative precipitation estimates, covering the North American region from 20°N to 55°N and 60°W to 130°W, with a spatial resolution of 1 km and a temporal resolution of 2 minutes. The automatic ground observation system data comes from 1679 ground stations, providing observations of wind direction, wind speed, air temperature, dew point temperature, visibility, sea level pressure, and cloud height, with a temporal resolution of 5 minutes. During the data preprocessing stage, the multi-radar / multi-sensor system and automatic ground observation system data are normalized to unify the dimensions of different variables. Finally, an input tensor is constructed, containing radar and station data from nine historical time steps.

[0036] Step 2: Use a lightweight encoder for feature extraction; the input end of the lightweight encoder is embedded with a time embedding module, which outputs a time-aware output tensor from the input tensor.

[0037] Specifically, the lightweight encoder includes three sequentially connected sub-modules. Each sub-module includes a residual network block and a downsampling module. The input of each residual network block has a time embedding module embedded in it. The output after downsampling is used as the input of the next sub-module, and the output of the last sub-module is used as the input of the temporal-space feature extraction module.

[0038] The implementation of this invention involves several steps. In the feature extraction stage, the lightweight encoder is the initial part, responsible for extracting high-level feature representations from the input data. The encoder consists of three sub-modules, each containing a residual network block and a downsampling operation. The residual network block adopts a depth of 3, mitigating the gradient vanishing problem through residual connections. The calculation formula is...

[0039]

[0040] The residual network block performs feature extraction using a transformation function consisting of convolutional layers, batch normalization, and ReLU activation. The input data is gradually downsampled from the original 512×512 spatial resolution to 64×64 to extract multi-scale feature information.

[0041] Furthermore, the input to the first-layer submodule is radar data, the input to the second-layer submodule is the output of the first-layer submodule plus the data obtained by concatenating radar data and station data, and the input to the third-layer submodule is the output of the second-layer submodule. In this embodiment, after the first submodule of the lightweight encoder, data from a multi-radar / multi-sensor system and an automatic ground observation system with a resolution of 4 kilometers are fused. These data are upsampled to the same resolution as the current feature map (i.e., 256×256) and concatenated with the encoder feature map in the channel dimension to enhance the model's ability to perceive spatial context.

[0042] Further settings in this embodiment, such as Figure 2 As shown, the time embedding module includes a time information decomposition module, a time information classification and processing module, and a semantic tensor merging module. The time information decomposition module classifies the input time information of each submodule into three categories: predicted timeliness information, time information with cyclical patterns, and time window information with historical weights. After classification, the information is uniformly summarized into different vectors for subsequent processing.

[0043] The categorized information is input into the time information classification processing module. The predicted lead time information is mapped to a feature vector (e.g., 10 minutes to 180 minutes) through an embedding layer. Feature 1 = Embedding layer (predicted lead time information)

[0044] Periodic time information is encoded using sine and cosine functions to capture the periodic characteristics of weather phenomena (such as diurnal variations). For time information exhibiting cyclical patterns, it is normalized.

[0045] Feature 2 = Normalization (Time information of cyclical patterns)

[0046] For historical weighted windows, different weights are assigned to data from historical time steps, and a learnable weight vector is used to emphasize the importance of recent observations:

[0047] Window semantic information = Embedded layer (window information)

[0048] Feature 3 = Normalized exponential function (window semantic information) · Window information

[0049] It should be noted that the prediction timeliness information refers to the future time step that the model needs to predict, such as the prediction target of 1 day, 1 week, or 1 month in the future. These discrete time steps (such as "1 day" or "2 days") are converted into continuous feature vectors, which are called the first features. The role of the embedding layer is to map the time steps to a vector space, enabling the model to understand and utilize the predicted time span. The time information with cyclical patterns refers to the part of the time data with periodic patterns, such as the daily cycle (the progress of the current day within 24 hours), the weekly cycle (the days of the week within 7 days), the annual cycle (the progress of the current year within 365 days), or the seasonal cycle (spring, summer, autumn, and winter). These cyclical features are converted into values ​​in the range [0,1] through normalization to obtain the second features. The time window information with historical weights refers to the past time steps contained in the historical data input to the model, such as data from the past 7 days or 30 days, which are used to provide the context for prediction. First, the window semantic information is obtained through the embedding layer. Then, the window semantic information is processed by the normalized exponential function and multiplied by a learnable weight parameter to obtain the third feature, thus obtaining three-dimensional features. This processing method allows the model to dynamically focus on more important parts of historical data. For example, recent data may be more predictive than distant data.

[0050] Finally, these feature vectors are concatenated:

[0051] Total features = concatenation(feature1, feature2, feature3)

[0052] It should be noted that, through the semantic tensor module, the total features are used as input to obtain the temporal rate of change and bias values ​​using two pre-trained multilayer perceptrons. One multilayer perceptron is pre-trained on a time series trend prediction task, aiming to learn to capture the rate of change or trend of the time series; the other is pre-trained on a time series baseline estimation or offset correction task, aiming to learn to capture the baseline or static offset of the time series. These differences result in different weights and bias parameters for the two MLPs after training, thus generating different outputs (temporal rate of change and bias values) in the semantic tensor module.

[0053] The time rate of change and bias value are then added to the input tensor to obtain a time-aware output tensor, which is then used as the input to the next module.

[0054] Time-aware data = rate of change over time · input information + bias value.

[0055] Step 3: Input the extracted features into the temporal-spatial feature extraction module to extract spatiotemporal features; the input end of the temporal-spatial extraction module is embedded with a time embedding module, and the features extracted in step 2 are input into the time embedding module to output time-aware features.

[0056] Specifically, the spatiotemporal feature extraction module includes four spatial extraction sub-modules and four temporal extraction sub-modules connected sequentially. Each spatial and temporal sub-module has a temporal embedding module embedded at its input. The data input to each spatial (Transformer) and temporal (Mamba) sub-module needs to be windowed. Windowing involves taking the time-aware output tensor as input and using the windowing module to split the large-dimensional tensor data into multiple sub-tensors according to a predefined window size, adjusting their dimensions to ensure they meet the input requirements of the sub-modules.

[0057]

[0058] The adjusted subtensors are input in parallel into the corresponding submodules to enable high-speed parallel data processing. The window partitioning module works in conjunction with the window restoration module after the submodules to restore the processing results back to their original input shape after processing by the submodules, which is the reverse process of window partitioning.

[0059]

[0060] It should be noted that the operation of the time embedding module embedded in each spatial submodule and temporal submodule is the same as the operation of the time embedding module in step 2.

[0061] Furthermore, to improve computational efficiency, the spatial dimension of the feature tensor is divided into 8×8 windows, and the data within each window is processed independently.

[0062] After the spatial extraction submodule is divided into windows, an attention mechanism is used within the sub-windows to process the spatial distribution features from the input feature meteorological data. The attention mechanism automatically assigns weights to each location by calculating the correlation between different spatial locations within the sub-window, thereby highlighting areas that are more important for precipitation prediction.

[0063] Specifically, each Transformer submodule uses a standard multi-head self-attention mechanism, calculated as follows:

[0064]

[0065] Where Q, K, and V are the query, key, and value matrices, respectively, and d h The number of attention heads determines the number of long-range spatial dependencies in the feature map. Through the self-attention mechanism, the Transformer can capture these long-range spatial dependencies.

[0066] The time extraction submodule captures long-term temporal dependencies in the input meteorological data using a state-space model within the divided sub-windows. The state-space model, by defining state transition and observation equations, effectively captures long-term dependencies in time series. For example, precipitation in meteorological data may be influenced by long-term factors such as seasonal variations (e.g., the rainy season) and climate patterns (e.g., El Niño). SSM can identify and model these relationships. By using causal convolution, the generated features are ensured to be causal during modeling; that is, when predicting future precipitation, only past and current data are relied upon, without "peeking" into future information, ultimately generating a causal temporal feature representation.

[0067] Specifically, the Mamba submodule is based on a structured state-space sequence model, suitable for processing time series data. Its core calculation formula is:

[0068] h′(t)=Ah(t)+Bx(t)

[0069] y(t)=Ch(t)

[0070] Here, h(t) represents the hidden state, a state vector maintained internally by the model when processing sequential data. It is updated based on the current input and previous hidden states, playing a role in conveying information and extracting features during sequence processing. A, B, and C are learnable parameters, and x(t) is the input data. To enhance the model's ability to model temporal features, causal convolution is introduced in the Mamba submodule to ensure temporal consistency of information. Temporal information is embedded at the input of each submodule to ensure that the model fully considers temporal factors when extracting spatiotemporal features. The output of each submodule is processed by layer normalization and a multilayer perceptron before being passed to the next layer, ultimately generating a feature representation that integrates spatial and temporal information.

[0071] Step 4: The spatiotemporal features are used as input to obtain the restored spatiotemporal features through a lightweight decoder; the input end of the lightweight decoder is embedded with a time embedding module, the spatiotemporal features are input to the time embedding module and the output is time-aware spatiotemporal features, and the restored spatiotemporal features are output by a multilayer perceptron for prediction results.

[0072] Specifically, the lightweight decoder is responsible for progressively restoring the features output by the mixing module to the original resolution and generating the final precipitation prediction. The lightweight decoder also consists of three sub-modules. Each sub-module uses transposed convolutions for upsampling and fuses features from the encoder through skip connections to preserve high-frequency details. Specifically, the lightweight decoder upsamples the feature map from 64×64 to 128×128, then to 256×512, and finally restores it to 512×512. In each upsampling step, skip connections concatenate the features of the corresponding layer of the encoder with the upsampled features along the channel dimension to enhance feature richness. Finally, the feature tensor output by the lightweight decoder is mapped to the target space through a multilayer perceptron to generate the precipitation prediction.

[0073] The processing method of the time embedding module embedded at the input end of the lightweight decoder is the same as that of the time embedding module in step 2.

[0074] It should be noted that in this embodiment, the model constructed in steps 2-4 is trained and optimized. During the model training and optimization phases, to ensure the model's prediction accuracy and generalization ability, mean squared error is used as the loss function. Error is calculated only within the valid data region to avoid interference from invalid or missing data in the training process. The root mean squared error is used as the loss function. Regarding optimization configuration, a stochastic gradient descent optimizer is used, with an initial learning rate set to 1×10^(-5), dynamically adjusted using a cosine annealing learning rate scheduler. The model is trained on the GPU for 300 epochs, and an early stopping strategy is used to prevent overfitting.

[0075] In terms of implementation details and advantages, this invention emphasizes lightweight design and efficiency, keeping the total number of parameters below 23.90MB, comparable to the total number of parameters in traditional deep learning models. However, through the innovative design of the temporal embedding module, the model can adapt to multiple prediction timeframes in a single training iteration, avoiding the redundancy of training the model separately for each timeframe. Therefore, it has a lower total number of parameters for multiple prediction timeframes. Furthermore, the combination of the lightweight U-Net structure and the Transformer-Mamba hybrid module allows the model to maintain a low number of parameters and complexity when processing high-resolution data, enabling rapid inference and execution, meeting the needs of real-time weather forecasting. This invention achieves high-precision, low-cost short-term precipitation forecasting through efficient spatiotemporal feature extraction and a flexible temporal embedding mechanism, applicable to various weather forecasting scenarios, and providing strong technical support for real-time weather monitoring and disaster early warning.

[0076] Finally, to verify the effectiveness of the present invention, this embodiment uses data from the US Multi-Radar / Multi-Sensor System and Automatic Ground Observation System from 2019 to 2022 for experiments. The dataset is randomly divided into a training set (80%), a validation set (15%), and a test set (5%). The prediction lead time includes 10, 20, 30, 60, 120, and 180 minutes. The comparison methods include PredRNNv2, DGMR, and pySTEPS. The evaluation metrics selected are the mean absolute error (MAE) and the critical success index (CSI). For CSI, the threshold settings are 1 mm / h, 4 mm / h, and 8 mm / h, respectively.

[0077] Table 1 Comparison of Method Parameters

[0078] method Parameters Number of models under multiple time periods pySTEPS - ×N PredRNNv2 21.59MB ×N DGMR 28.96MB ×N This invention 23.90MB ×1

[0079] In this embodiment, the parameters of all models are shown in Table 1. pySTEPS is a Python framework for short-term prediction systems that does not require training; therefore, its parameter count cannot be calculated. In contrast, the method of this invention can handle all prediction timeframes simultaneously, thus reducing the total number of parameters from the parameters of N models to the parameters of one model, significantly reducing the total number of parameters required for methods across multiple timeframes.

[0080] Table 2 shows the predictive performance of the methods under the MAE and CSI indices.

[0081] method MAE↓ CSI (1mm / h)↑ CSI (4mm / h)↑ CSI (8mm / h)↑ pySTEPS 0.313100 0.257003 0.143201 0.096138 PredRNNv2 0.527527 0.273717 0.211647 0.124541 DGMR 0.615793 0.281918 0.215486 0.129790 This invention 0.371873 0.350295 0.205924 0.157116

[0082] Figures 3-6 The prediction performance of all models at different prediction lead times is shown in Table 2, which displays the average prediction performance of all methods across six prediction lead times. For MAE (Mean Absolute Error), our two model versions perform well with shorter lead times. However, as the lead time increases, the MAE of our model and other deep learning benchmark models increases significantly. In contrast, pySTEPS maintains relatively stable MAE performance even with longer lead times due to the stability of its numerical computation method. Our method demonstrates excellent prediction performance for CSI (Critical Success Index) at three different thresholds, only slightly lagging behind other comparative models at the 4 mm / h threshold.

[0083] Experimental results show that the present invention is highly competitive in predictive performance across multiple prediction timeframes, and requires fewer parameters to meet the demand for multiple prediction timeframes. The model's dynamic adaptability and lightweight nature also reduce its environmental requirements.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction, characterized in that, Includes the following steps: Step 1: Collect historical meteorological data and preprocess it to generate an input tensor containing multi-dimensional information; the historical meteorological data includes radar data and station data. Step 2: Feature extraction is performed using a lightweight encoder; the input end of the lightweight encoder is embedded with a time embedding module, which outputs a time-aware output tensor from the input tensor. Step 3: Input the extracted features into the temporal-spatial feature extraction module to extract spatiotemporal features; the input end of the temporal-spatial feature extraction module is embedded with a time embedding module, and the features extracted in step 2 are input into the time embedding module to output time-aware spatiotemporal features; Step 4: Use the spatiotemporal features as input and pass them through a lightweight decoder to obtain the restored spatiotemporal features; The lightweight decoder has a time embedding module embedded in its input end. The spatiotemporal features are input to the time embedding module and output time-aware restored spatiotemporal features. The restored spatiotemporal features are then used by a multilayer perceptron to output prediction results. The method by which the time embedding module embeds time information is as follows: The input information of each time embedding module is classified according to the time information category. The classification results are then used to extract features to obtain multi-dimensional features. The multi-dimensional features are then concatenated to obtain the total features. The total features are used as input to obtain the time rate of change and bias value using two pre-trained multilayer perceptrons. The time rate of change and bias value are then added to the input information to obtain time-aware data, which is then used as the input to the next module.

2. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 1, characterized in that, The multi-dimensional information includes multi-element meteorological information, time series, and spatial distribution information.

3. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 1, characterized in that, Preprocessing methods include data normalization, noise filtering, and missing value imputation.

4. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 1, characterized in that, The time information categories include predicted timeliness information, time information with cyclical patterns, and time window information with historical weights. The predicted timeliness information obtains a first feature through an embedding layer. The time information with cyclical patterns obtains a second feature through normalization. The time window information with historical weights first obtains window semantic information through an embedding layer, and then processes the window semantic information through a normalized exponential function and multiplies it by a learnable weight parameter to obtain a third feature, thus obtaining three-dimensional features.

5. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 4, characterized in that, The time-aware data = time change rate · input information + bias value.

6. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 5, characterized in that, The lightweight encoder includes three sequentially connected sub-modules. Each sub-module includes a residual network block and a downsampling module. The input of each residual network block has a time embedding module embedded in it. The output after downsampling is used as the input of the next sub-module, and the output of the last sub-module is used as the input of the time-space feature extraction module.

7. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 6, characterized in that, The input to the first-layer submodule is radar data. The input to the second-layer submodule is the output of the first-layer submodule and the data obtained by stitching together radar data and site data. The input to the third-layer submodule is the output of the second-layer submodule.

8. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 5, characterized in that, The temporal and spatial feature extraction module includes four spatial extraction sub-modules and four temporal extraction sub-modules connected in sequence; each spatial extraction sub-module and temporal extraction sub-module has a temporal embedding module embedded at its input end, and the data input to each spatial extraction sub-module and temporal extraction sub-module needs to be windowed.

9. A multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 8, characterized in that, The spatial extraction submodule processes the spatial distribution features from the input meteorological data by dividing it into windows; the temporal extraction submodule captures the long-term temporal dependencies from the input meteorological data and generates a causal temporal feature representation.

10. A multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 5, characterized in that, The lightweight decoder is symmetrical to the lightweight encoder. The three-layer sub-modules in the lightweight decoder include residual network blocks and upsampling, and each has a time embedding module embedded at the input.

11. The multi-time-dependent short-term precipitation prediction method based on joint temporal and spatial feature extraction according to claim 10, characterized in that, The upsampling uses transposed convolution upsampling, and the downsampling in the three-layer sub-module of the lightweight encoder is skipped to the corresponding three-layer residual network block in the lightweight decoder.

Images