Multi-time-effect short-temporary rainfall prediction method based on time-space joint feature extraction

Abstract

The invention discloses a multi-time-efficiency short-temporary rainfall prediction method based on time-space joint feature extraction, and the method comprises the steps: carrying out the down-sampling processing of preprocessed multi-dimensional meteorological data through employing an encoder based on a U-Net architecture, and carrying out the semantic embedding of input information through employing a time information embedding module, and the sub-module has the time sensing capability. And the features after down-sampling processing are input into a Transform-Mama hybrid module for efficiently extracting space and time features, and window division is used in front of each sub-module for parallel processing. And performing up-sampling reduction on the feature data by using a decoder, complementing information lost due to down-sampling by using jump connection, and finally inputting the information into a multi-layer perceptron for prediction to obtain a precipitation result of the target area. The method can effectively carry out multi-time local high-precision rainfall prediction in a short critical time, realizes the complexity of a lightweight model, and has excellent prediction capability.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and deep learning technology, and in particular to a multi-time-effect short-term precipitation prediction method based on temporal and spatial joint feature extraction. Background Art

[0002] Precipitation forecasting is crucial for disaster preparedness, urban planning, and public safety. Timely and accurate predictions can effectively reduce loss of life and property. Traditional numerical weather forecasting (NWF) 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, NWF suffers from insufficient real-time performance in short-term nowcasting (typically 0-3 hours). This is especially true in scenarios requiring high-resolution output, where the computational resource requirements increase significantly, making it difficult to meet the requirements of real-time applications.

[0003] In recent years, weather forecasting methods based on deep learning have gradually emerged. By leveraging neural networks and GPU-accelerated computing, they can generate forecast results within seconds and, to a certain extent, make up for the shortcomings of numerical weather forecasting. For example, PredRNNv2 captures spatiotemporal features through a recursive network, and DGMR uses a generative model to improve precipitation forecast accuracy. However, existing DLWP methods still face the following challenges: first, independent models are trained for each forecast time, resulting in high computational cost and lack of flexibility; second, the modeling capability of high-resolution spatiotemporal features is insufficient, making it difficult to simultaneously capture local details and global dynamics of precipitation; third, the balance between real-time performance and prediction accuracy still needs to be optimized, especially when processing large-scale meteorological datasets (such as radar reflectivity and ground station observation data). The computational load is large, limiting the widespread application of the model.

[0004] To address the above issues, there is an urgent need for an efficient and lightweight short-term precipitation prediction method that can reduce the demand for computing resources while improving the accuracy and real-time performance of multi-time predictions. Summary of the Invention

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

[0006] In order to achieve the above object, the technical solutions specifically adopted by the present invention are as follows:

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

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

[0009] Step 2: Use a lightweight encoder to perform feature extraction; the input end of the lightweight encoder is embedded with a time embedding module, and the input tensor is outputted as a time-aware output tensor through the time embedding module;

[0010] Step 3: Input the extracted features into a time-space feature extraction module to extract time-space features; the input end of the time-space 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 features with time perception;

[0011] Step 4: Using the spatiotemporal features as input, a lightweight decoder is used to obtain restored spatiotemporal features; the input end of the lightweight decoder is embedded with a time embedding module, the spatiotemporal features are input into the time embedding module to output spatiotemporal features with time perception, and the restored spatiotemporal feature multilayer perceptron outputs a prediction result;

[0012] The method for embedding time information by the time embedding module is:

[0013] The input information of each module is classified according to the time information category, and the classification results are extracted to obtain multi-dimensional features, which are then spliced ​​to obtain the overall features;

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

[0015] Preferably, the multi-dimensional information includes multi-element meteorological related 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; the prediction timeliness information refers to the future time steps that the model needs to predict, such as the prediction targets for the next 1 day, 1 week, or 1 month. These discrete time steps (such as "1 day" and "2 days") are converted into continuous feature vectors, which are called first features. The role of the embedding layer is to map the time steps to a vector space so that the model can 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 daily cycles (the progress of the day within 24 hours), weekly cycles (the day of the week within 7 days), annual cycles (the progress of the year within 365 days), or seasonal cycles (spring, summer, autumn, and winter). These cyclical features are converted into values ​​in the [0,1] interval by normalization to obtain the second feature; 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 is used to provide the context for the prediction. First, the window semantic information is obtained through the embedding layer. Then, the window semantic information is processed by a normalized exponential function and multiplied by a learnable weight parameter to obtain the third feature, thereby obtaining features in three dimensions. This processing method allows the model to dynamically focus on more important parts of historical data. For example, recent data may have more predictive value than long-term data.

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

[0019] Preferably, the time-space feature extraction module includes 4 spatial extraction sub-modules and 4 time extraction sub-modules connected in sequence; the input end of each of the spatial sub-module and time sub-module is embedded with a time embedding module, and the data before inputting each of the spatial sub-module and time sub-module 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 spliced ​​from the radar data and the site data, and the input of the third-layer submodule is the output of the second-layer submodule.

[0021] Preferably, the time-space feature extraction module includes 4 spatial extraction sub-modules and 4 time extraction sub-modules connected in sequence; each of the spatial sub-module and time sub-module is embedded with a time embedding module, and the data before inputting each of the spatial sub-module and time sub-module needs to be windowed.

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

[0023] Preferably, the lightweight decoder is symmetrical with the lightweight encoder. The three-layer submodules in the lightweight decoder include a residual network block and an upsampling layer, and each layer has a time embedding module embedded at the input. Preferably, the upsampling layer uses transposed convolution upsampling, and the downsampling layers in the three-layer submodules in the lightweight encoder are skip-connected to the corresponding three-layer residual network blocks in the lightweight decoder.

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

[0025] The present invention has significant innovations and advantages. By optimizing the U-Net structure and hybrid module, the total number of parameters is controlled within 23.9MB, which significantly reduces computing requirements and improves model scalability. The time embedding module enables a single model to adapt to multi-time prediction, avoiding the need to train a separate model for each time. The Transformer-Mamba hybrid module combines the advantages of spatial global modeling and time series modeling to enhance high-resolution feature capture capabilities. The minute-to-kilometer level prediction is comparable to the most advanced deep learning model in accuracy, while also having rapid response capabilities. Through the above-mentioned technical solutions, the present invention effectively solves the shortcomings of traditional methods in short-term precipitation prediction, and provides a new route for real-time meteorological applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 The figure is a schematic diagram of the overall structure of a multi-time-effect short-term precipitation prediction method based on time-space joint feature extraction of the present invention.

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

[0028] Figure 3 The figure is a line chart comparing the MAE indicators of the present invention with other methods under 6 prediction time periods.

[0029] Figure 4 The figure is a line chart comparing the CSI (1 mm / h) index of the present invention with other methods under 6 prediction time periods.

[0030] Figure 5 The figure is a line chart comparing the CSI (4 mm / h) of the present invention with other methods under 6 prediction time periods.

[0031] Figure 6 The figure is a line chart comparing the CSI (8 mm / h) of the present invention with other methods under 6 prediction time periods. DETAILED DESCRIPTION

[0032] The present invention is described in detail below in conjunction with specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

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

[0034] Step 1: Collect historical meteorological data and preprocess the historical meteorological data to generate an input tensor containing multi-dimensional information; 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 estimation, covering the North American area of ​​20°-55°N and 60°-130°W, with a spatial resolution of 1 km and a temporal resolution of 2 minutes. The automatic ground observation system data comes from 1,679 ground stations, providing wind direction, wind speed, air temperature, dew point temperature, visibility, sea level pressure and cloud height observations with a temporal resolution of 5 minutes. In 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, which contains radar data and station data for 9 historical time steps.

[0036] Step 2: Use a lightweight encoder to perform feature extraction; the input end of the lightweight encoder is embedded with a time embedding module, and the input tensor is outputted as a time-aware output tensor through the time embedding module.

[0037] Specifically, the lightweight encoder includes three layers of sub-modules connected in sequence, each of which includes a residual network block and downsampling. The input end of the residual network block in each layer is embedded with a time embedding module; the downsampled output serves as the input of the next layer of sub-modules, and the output of the last layer of sub-modules serves as the input of the time-space feature extraction module.

[0038] The implementation of the present invention is divided into several steps. In the feature extraction stage, the lightweight encoder is the starting part, which is responsible for extracting high-level feature representations from the input data. The encoder consists of three submodules, each of which contains a residual network block and a downsampling operation. The residual network block adopts a depth of 3 and alleviates the gradient vanishing problem through residual connections. The calculation formula is

[0039]

[0040] The residual network block performs feature extraction and is a transformation function consisting of a convolutional layer, batch normalization, and ReLU activation function. 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 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 concatenation of radar data and station data, and the input of the third-layer submodule is the output of the second-layer submodule. In this embodiment, after the first submodule of the lightweight encoder, multi-radar / multi-sensor system and automatic ground observation system data with a resolution of 4 kilometers are fused. This data is adjusted to the same resolution as the current feature map (i.e., 256×256) through an upsampling operation and is concatenated with the encoder feature map in the channel dimension to enhance the model's perception of spatial context.

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

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

[0044] Periodic time uses sine and cosine functions to encode time information to capture the periodic characteristics of weather phenomena (such as day and night changes). For time information with cyclical patterns, it is normalized:

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

[0046] For the historical weighted window, different weights are assigned to the data of the historical time step, and a learnable weight vector is used to emphasize the importance of recent observation data:

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

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

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

[0050] Finally, these feature vectors are concatenated:

[0051] Total features = concatenation (feature 1, feature 2, feature 3)

[0052] It's important to note that the semantic tensor module uses the total features as input to obtain the temporal rate of change and bias values ​​using two pre-trained multilayer perceptrons. One multilayer perceptron was pre-trained on a time series trend prediction task, aiming to learn to capture the rate of change or trend of a time series; the other was pre-trained on a time series baseline estimation or offset correction task, aiming to learn to capture the baseline or static offset of a time series. These differences result in different weights and bias parameters in the two MLPs after training, resulting in 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 appended to the input tensor to obtain a time-aware output tensor, which serves as the input of the next module.

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

[0055] Step 3: Input the extracted features into the time-space feature extraction module to extract the time-space features; the input end of the time-space 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 features with time perception.

[0056] Specifically, the temporal and spatial feature extraction module includes four spatial extraction submodules and four temporal extraction submodules connected in sequence; the input end of each spatial submodule and temporal submodule is embedded with a time embedding module, and the data before inputting each spatial submodule (Transformer submodule) and temporal submodule (Mamba submodule) needs to be windowed. Windowing is to use the time-aware output tensor as input and split the large-dimensional tensor data into multiple sub-tensors according to the predefined window size through the windowing module, and adjust the size to ensure that it meets the input requirements of the submodule:

[0057]

[0058] The adjusted sub-tensors are input in parallel to the corresponding sub-modules so that data operations can be performed in parallel and at high speed. The window division module is used in conjunction with the window restoration module after the sub-module to restore the processed results to the input shape after the sub-module processing, which is the inverse process of window division:

[0059]

[0060] It should be noted that the operation of the time embedding module in which each spatial submodule and temporal submodule is embedded 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 in each window is processed independently.

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

[0063] Specifically, each Transformer submodule adopts a standard multi-head self-attention mechanism, and the calculation formula is:

[0064]

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

[0066] Through the time extraction submodule, the long-term temporal dependencies in the input feature meteorological data are captured using a state-space model within the sub-window after window division. The state-space model can effectively capture long-term dependencies in time series by defining state transitions and observation equations. For example, in meteorological data, precipitation may be affected by long-term factors such as seasonal changes (such as the rainy season) and climate patterns (such as the El Niño phenomenon). 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, it only relies on past and current data without "peeking" at future information, ultimately generating a causal temporal feature representation.

[0067] Specifically, the Mamba submodule is based on a structured state space sequence model and is 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) is the hidden state, a state vector maintained internally by the model when processing sequential data. It is updated based on the current input and the previous hidden state, playing a role in transmitting 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. Time information is embedded at the input of each submodule to ensure that the model fully considers time when extracting spatiotemporal features. The output of the submodule undergoes layer normalization and multi-layer perceptron processing before being passed to the next layer, ultimately generating a feature representation that integrates spatial and temporal information.

[0071] Step 4: Using the spatiotemporal features as input, a lightweight decoder is used to obtain restored spatiotemporal features; the input end of the lightweight decoder is embedded with a time embedding module, and the spatiotemporal features are input into the time embedding module to output spatiotemporal features with time perception, and the restored spatiotemporal feature multilayer perceptron outputs the prediction result.

[0072] Specifically, the lightweight decoder is responsible for gradually restoring the features output by the hybrid module to their original resolution and generating the final precipitation forecast. The lightweight decoder also consists of three submodules, each of which uses transposed convolution for upsampling and fuses features from the encoder via 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 to 512×512. In each upsampling step, a skip connection concatenates the features of the corresponding encoder layer with the upsampled features in the channel dimension to enhance feature richness. Finally, the feature tensor output by the lightweight decoder is mapped to the target space through a multi-layer perceptron to generate the precipitation forecast.

[0073] The processing method of the time embedding module embedded in the input end of the lightweight decoder is the same as the processing method 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 phase, to ensure the prediction accuracy and generalization ability of the model, the mean square error is used as the loss function. The error is calculated only within the valid data area to avoid interference of invalid or missing data in the training process. The loss function uses the root mean square error. In terms of optimization configuration, a stochastic gradient descent optimizer is used, the initial learning rate is set to 1×10^(-5), and the cosine annealing learning rate scheduler is used for dynamic adjustment. The model is trained on the GPU for 300 rounds, and an early stopping strategy is used to prevent overfitting.

[0075] In terms of implementation details and advantages, the present invention focuses on lightweight and high efficiency in design, and the total parameter volume is controlled within 23.90MB, which is comparable to the total parameter volume of traditional deep learning model methods. However, through the innovative design of the time embedding module, the model can adapt to multiple prediction time periods in a single training, avoiding the redundancy of training the model separately for each time period, so it has a lower total parameter volume under multiple prediction time period requirements. In addition, the combination of the lightweight U-Net structure and the Transformer-Mamba hybrid module enables the model to maintain a low parameter volume and complexity when processing high-resolution data, and can be quickly inferred and run to meet the needs of real-time weather forecasting. The present invention achieves high-precision, low-cost short-term precipitation forecasting through efficient spatiotemporal feature extraction and flexible time embedding mechanism, which is suitable for various weather forecast scenarios and provides strong technical support for real-time weather monitoring and disaster 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 data set is randomly divided into a training set (80%), a validation set (15%), and a test set (5%). The prediction time includes 10, 20, 30, 60, 120, and 180 minutes. The comparison methods include PredRNNv2, DGMR, and pySTEPS. The evaluation indicators are the mean absolute error (MAE) and the critical success index (CSI). For the CSI, the threshold settings used are 1 mm / h, 4 mm / h, and 8 mm / h, respectively.

[0077] Table 1 Comparison of method parameters

[0078] method Parameter quantity Number of models under multiple time periods pySTEPS - ×N PredRNNv2 21.59MB ×N DGMR 28.96MB ×N The present invention 23.90MB ×1

[0079] In this example, the parameters of all models are shown in Table 1. pySTEPS is a Python framework for short-term forecasting systems that does not require training, so its parameter count cannot be calculated. In contrast, the method of the present invention can handle all forecast timeframes simultaneously, reducing the total number of parameters from N models to just one, significantly reducing the total number of method parameters required for multiple timeframes.

[0080] Table 2 Prediction performance of the method under MAE and CSI indicators

[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 The present invention 0.371873 0.350295 0.205924 0.157116

[0082] Figures 3 to 6 The prediction performance of all models at different prediction time frames is shown, and Table 2 shows the average prediction performance of all methods at six prediction time frames. For MAE (mean absolute error), our two model versions perform well at 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 can maintain relatively stable MAE performance at longer lead times due to the stability of its numerical calculation method. This method shows excellent prediction performance in CSI (critical success index) at three different thresholds, and is only slightly inferior to other comparison models at the 4mm / h threshold.

[0083] Experimental results demonstrate that the proposed method is highly competitive in terms of predictive performance across multiple timeframes. Furthermore, the proposed method requires fewer parameters for these scenarios. The model's dynamic adaptability and lightweight design also reduce its environmental requirements.

[0084] The above shows and describes 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 above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-time effect short-term precipitation prediction method based on temporal and spatial joint feature extraction, characterized in that: The steps include: Step 1: Collect historical meteorological data and pre-process the historical meteorological data to generate an input tensor containing multi-dimensional information; the historical meteorological data includes radar data and station data; Step 2: Use a lightweight encoder to perform feature extraction; the input end of the lightweight encoder is embedded with a time embedding module, and the input tensor is outputted into a time-aware output tensor through the time embedding module; Step 3: Input the extracted features into a time-space feature extraction module to extract time-space features; the input end of the time-space 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 features with time perception; Step 4: Using the spatiotemporal features as input, a lightweight decoder is used to obtain restored spatiotemporal features; the input end of the lightweight decoder is embedded with a time embedding module, the spatiotemporal features are input into the time embedding module to output spatiotemporal features with time perception, and the restored spatiotemporal feature multilayer perceptron outputs a prediction result; The method for embedding time information by the time embedding module is: The input information of each module is classified according to the time information category, and the classification results are extracted to obtain multi-dimensional features, which are then spliced ​​to obtain the overall features; The total feature is used as input to obtain the time change rate and bias value using two pre-trained multi-layer perceptrons. The time change rate and bias value are then appended to the input information to obtain time-aware data, which is used as the input of the next module.

2. The method for predicting short-term precipitation based on time-space joint feature extraction according to claim 1 is characterized in that: The multi-dimensional information includes multi-element meteorological related information, time series and spatial distribution information.

3. The method for predicting short-term precipitation based on time-space joint feature extraction according to claim 1 is characterized in that: The preprocessing method includes normalizing the data, filtering out noise, and filling in missing values.

4. The method for predicting short-term precipitation based on time-space joint 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, thereby obtaining features in three dimensions.

5. The method for predicting short-term precipitation based on time-space joint feature extraction according to claim 4 is characterized in that: The time-perceived data = time change rate·input information+bias value.

6. The method for predicting short-term precipitation based on time-space joint feature extraction according to claim 5, characterized in that: The lightweight encoder includes three layers of sequentially connected sub-modules, each of which includes a residual network block and downsampling. The input end of each residual network block is embedded with a time embedding module; the downsampled output serves as the input of the next layer of sub-modules, and the output of the last layer of sub-modules serves as the input of the time-space feature extraction module.

7. The method for predicting short-term precipitation based on time-space joint feature extraction according to claim 1, characterized in that: 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 spliced ​​from radar data and site data, and the input of the third-layer submodule is the output of the second-layer submodule.

8. The method for predicting multi-effect short-term precipitation 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 submodules and four temporal extraction submodules connected in sequence; the input end of each of the spatial submodule and temporal submodule is embedded with a time embedding module, and the data before inputting each of the spatial submodule and temporal submodule needs to be windowed.

9. The method for predicting short-term precipitation 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 feature meteorological data after window division; the time extraction submodule captures the long-term time dependency from the input feature meteorological data and generates a causal time feature representation.

10. The method for predicting short-term precipitation with multiple time effects based on joint temporal and spatial feature extraction according to claim 5, characterized in that: The lightweight decoder is symmetrical with the lightweight encoder. The three-layer submodules in the lightweight decoder include a residual network block and upsampling, and both are embedded with a time embedding module at the input end.

11. The method for predicting short-term precipitation with multiple time effects 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 submodule in the lightweight encoder is jump-connected with the three-layer residual network block in the corresponding lightweight decoding.

Images