Method for predicting precipitation of multi-source meteorological elements based on spatio-temporal information conversion equation

Abstract

The present invention provides a method for predicting precipitation using multi-source meteorological elements based on a spatiotemporal information conversion equation, which relates to the field of precipitation prediction and includes: establishing a forward-inverse dual model based on the spatiotemporal information conversion equation; obtaining multiple groups of training samples, wherein the training samples include multi-source meteorological element data and the labels of the training samples are rainfall data; training the forward-inverse dual model based on the spatiotemporal information conversion equation based on an adaptive weighted gradient loss function and multiple groups of training samples; generating a multi-source meteorological element precipitation prediction model based on the trained forward-inverse dual model of the spatiotemporal information conversion equation; obtaining multi-source meteorological element data to be predicted; and generating a precipitation prediction image based on the multi-source meteorological element precipitation prediction model based on the multi-source meteorological element data to be predicted, which has the advantage of improving the accuracy of precipitation prediction.

Description

Technical Field

[0001] The present invention relates to the field of precipitation prediction, and in particular to a method for predicting precipitation using multi-source meteorological elements based on a time-space information conversion equation. Background Art

[0002] Precipitation forecasting refers to the quantitative or qualitative prediction of precipitation type (such as rain, snow, hail), intensity (such as light rain or heavy rain), and temporal and spatial distribution over a specific area within a specific time period, based on meteorological observation data, numerical models, statistical methods, and artificial intelligence technologies. Its core goal is to provide a scientific basis for disaster prevention and mitigation, agricultural planning, water resources management, and public health.

[0003] In existing technologies, numerical weather forecasting uses sub-grid-scale parameterization to approximate unresolved physical processes. Although reducing the grid size can improve spatial resolution, this requires an exponential increase in computing resources to resolve fine-scale atmospheric interactions. In addition, the sensitivity of numerical weather forecasting to initial condition perturbations causes the forecast trajectory to deviate rapidly. These limitations hinder its applicability to short-term precipitation forecasts. Precipitation forecasting methods based on radar echo extrapolation, under highly nonlinear and complex atmospheric conditions, will amplify position and intensity errors over time due to iterative extrapolation. This error propagation can lead to a significant decrease in forecast accuracy, especially for localized heavy precipitation events. Moreover, due to the rapid pace of climate change, long-term time series data are generally insufficient to accurately resolve precipitation forecasts for heavy precipitation weather events. Most current forecasting methods that rely on short-term data perform poorly in the prediction of heavy precipitation events due to insufficient short-term data information.

[0004] Therefore, it is necessary to provide a method for predicting precipitation using multi-source meteorological elements based on spatiotemporal information conversion equations to solve the problem of inaccurate forecasts in heavy rainfall areas caused by insufficient short-term data information. Summary of the Invention

[0005] The present invention provides a method for predicting precipitation using multi-source meteorological elements based on a spatiotemporal information conversion equation, comprising: establishing a forward-inverse dual model based on the spatiotemporal information conversion equation; obtaining multiple groups of training samples, wherein the training samples include multi-source meteorological element data, and the labels of the training samples are rainfall data; training the forward-inverse dual model based on the spatiotemporal information conversion equation based on an adaptive weighted gradient loss function and multiple groups of training samples; generating a multi-source meteorological element precipitation prediction model based on the trained forward-inverse dual model of the spatiotemporal information conversion equation; obtaining multi-source meteorological element data to be predicted; and generating a precipitation prediction image based on the multi-source meteorological element precipitation prediction model based on the multi-source meteorological element data to be predicted.

[0006] Furthermore, the forward-inverse dual model based on the space-time information conversion equation includes a data preprocessing module, a space-time converter and a space-time converter; the data preprocessing module is used to preprocess multi-source meteorological element data and rainfall data to generate a delay embedding matrix; the space-time converter is used to generate a precipitation prediction image based on the delay embedding matrix; the space-time converter is used to extract the time characteristics of the precipitation prediction image, and reconstruct the spatial distribution of multidimensional meteorological data based on the time characteristics of the precipitation prediction image.

[0007] Furthermore, the data preprocessing module preprocesses the multi-source meteorological element data and rainfall data to generate a delay embedding matrix, including: performing spatiotemporal alignment on the multi-source meteorological element data and rainfall data to generate spatiotemporal aligned multi-source meteorological element data and rainfall data; normalizing the spatiotemporal aligned multi-source meteorological element data and rainfall data to generate normalized multi-source meteorological element data and rainfall data; performing delay embedding transformation on the normalized multi-source meteorological element data to generate a delay embedding matrix.

[0008] Furthermore, the space-time converter includes a first encoder and a second encoder, wherein the first encoder is used to perform convolution feature extraction on the delay embedding matrix to generate a spatial feature map, and the second encoder is used to generate a precipitation prediction image based on the spatial feature map.

[0009] Furthermore, the first encoder includes multiple convolutional layers and an encoding ConvLSTM layer, wherein the multiple convolutional layers are used to extract spatial information of the delay embedding matrix, and the encoding ConvLSTM layer is used to generate a spatial feature map based on the spatial information of the delay embedding matrix.

[0010] Furthermore, the first decoder includes multiple deconvolution layers and a decoding ConvLSTM layer, wherein the multiple deconvolution layers are used to generate precipitation distribution at future moments based on the spatial feature map, and the decoding ConvLSTM layer is used to generate a precipitation prediction image based on the precipitation distribution at future moments.

[0011] Furthermore, the spatiotemporal converter includes a second encoder and a second decoder, wherein the second encoder is used to extract the temporal characteristics of the precipitation prediction image, and the second decoder is used to reconstruct the spatial distribution of multidimensional meteorological data based on the temporal characteristics of the precipitation prediction image.

[0012] Furthermore, the adaptive weighted gradient loss function includes a mean absolute error function and a gradient difference loss function based on adaptive weights.

[0013] Furthermore, the adaptive weighted gradient loss function is:

[0014] ;

[0015] ;

[0016] ;

[0017] ;

[0018] in, is the adaptive weighted gradient loss function, is the mean absolute error function, is the gradient difference loss function based on adaptive weights, is the pixel value of the pixel in the i-th row and j-th column in the precipitation prediction image output by the forward-inverse dual model based on the spatiotemporal information conversion equation, is the total number of rows, is the total number of columns, is the gradient change rate of the precipitation prediction image in the horizontal direction, is the gradient change rate of the precipitation prediction image in the vertical direction, is the gradient change rate of the real precipitation image in the horizontal direction, is the gradient change rate of the real precipitation image in the vertical direction, is the pixel value of the pixel in the i-1th row and jth column in the precipitation prediction image, is the pixel value of the pixel in the i-th row and j-1-th column in the precipitation prediction image, is the adaptive weight.

[0019] Furthermore, the calculation formula of the adaptive weight is:

[0020] ;

[0021] in, is the precipitation prediction image of the i-th training sample, is the real precipitation image of the i-th training sample, is the number of training samples.

[0022] Compared with the existing technology, the method for predicting precipitation using multi-source meteorological elements based on the spatiotemporal information conversion equation provided by the present invention has at least the following beneficial effects:

[0023] Traditional precipitation forecasting methods typically rely on short-term historical data, making it difficult to fully capture the complex spatiotemporal evolution of heavy precipitation. This new method, using a spatiotemporal information conversion equation and a delayed embedding conversion mechanism, maps multidimensional spatial information into temporal information about future precipitation. This effectively integrates long-term spatiotemporal dependencies and addresses the issue of insufficient short-term data. This enables the model to more comprehensively understand the mechanisms that drive heavy precipitation, resulting in more accurate forecasts.

[0024] In long-term precipitation forecasts, errors tend to accumulate over time, causing the forecast results to gradually deviate from the actual situation. The present invention forms a closed-loop feedback mechanism through the design of a forward-inverse dual model: the space-time converter is responsible for predicting future precipitation information based on historical data. The space-time converter attempts to reconstruct the input features from the predicted precipitation information. This two-way constraint effectively suppresses the propagation of errors, allowing the model to maintain a low cumulative error in long-term time series forecasts and improve the stability of long-term precipitation forecasts. The design of the forward-inverse dual model forms a closed-loop constraint, avoiding the risk of overfitting of a single prediction model and improving the model's ability to generalize to unseen data. This enables the present invention to maintain stable performance under complex meteorological conditions (such as areas with complex terrain and extreme weather events), and has greater practical value.

[0025] Heavy precipitation regions often exhibit high spatial and temporal heterogeneity, and traditional methods tend to underestimate precipitation intensity due to the loss of local features. This method comprehensively captures the characteristics of heavy precipitation regions by fusing multi-source meteorological data. By constraining the gradient domain differences between the predicted precipitation image and the target image, the model enhances its capture of gradients at the edges and within heavy precipitation regions, thus avoiding underestimation of extreme precipitation.

[0026] Through the delayed embedding conversion mechanism, historical meteorological feature sequences are converted into delayed embedding matrices, explicitly encoding spatiotemporal dependencies. The synergistic effect of the space-time converter and the space-time converter enables bidirectional conversion from spatial features to time series, establishing a stable mapping relationship. This enables the model to learn more robust spatiotemporal feature representations, improving its generalization capabilities for complex precipitation patterns.

[0027] By optimizing the model with an adaptive weighted gradient loss function (e.g., combining MAE and GDLoss), this paper balances global error with local detail. Furthermore, a dual training framework is employed to enhance the model's interpretability and robustness. Ultimately, the model generates high-resolution, high-precision precipitation forecast images that accurately depict the spatial and temporal distribution of precipitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] This specification will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, like numbers represent like structures, wherein:

[0029] Figure 1 This is a flow chart of a method for predicting precipitation using multiple meteorological elements based on a spatiotemporal information conversion equation according to some embodiments of this specification;

[0030] Figure 2 is a schematic structural diagram of a forward-inverse dual model based on a spatiotemporal information conversion equation according to some embodiments of this specification;

[0031] Figure 3 is a schematic structural diagram of a first encoder according to some embodiments of this specification;

[0032] Figure 4 (a) is a schematic diagram of CSI scores under different precipitation thresholds according to all comparison methods shown in some embodiments of this specification;

[0033] Figure 4 (b) is a schematic diagram of HSS scores under different precipitation thresholds according to all comparative methods shown in some embodiments of this specification;

[0034] Figure 5 (a) is a schematic diagram of CSI scores at different prediction times under a high precipitation threshold according to all comparison methods shown in some embodiments of this specification;

[0035] Figure 5 (b) is a schematic diagram of HSS scores at different prediction times under a high precipitation threshold according to all comparative methods shown in some embodiments of this specification;

[0036] Figure 5 (c) is a schematic diagram of the FAR scores at different prediction times for all the comparison methods shown in some embodiments of this specification under a high precipitation threshold;

[0037] Figure 6 This is a visual comparison diagram of the overall effects of all comparison methods shown in some embodiments of this specification under the precipitation case;

[0038] Figure 7 (a) is a schematic diagram of CSI scores after eliminating the space-time converter according to some embodiments of this specification and adding the space-time converter by other methods;

[0039] Figure 7 (b) is a schematic diagram of HSS scores after eliminating the space-time converter according to some embodiments of this specification and adding the space-time converter by other methods;

[0040] Figure 8 (a) is a schematic diagram of a CSI score for ablation of a loss function according to some embodiments of this specification;

[0041] Figure 8 (b) is a schematic diagram of the HSS score for ablation of the loss function according to some embodiments of the present specification;

[0042] Figure 8 (c) is a schematic diagram of the BSS score of the loss function ablated according to some embodiments of this specification. DETAILED DESCRIPTION

[0043] To more clearly illustrate the technical solutions of the embodiments of this specification, the following briefly describes the drawings required for describing the embodiments. Obviously, the drawings described below are merely examples or embodiments of this specification. Those skilled in the art can apply this specification to other similar scenarios based on these drawings without inventive effort. Unless otherwise apparent from the context or otherwise noted, the same reference numerals in the figures represent the same structure or operation.

[0044] Figure 1 is a flow chart of a method for predicting precipitation using multiple meteorological elements based on a spatiotemporal information conversion equation according to some embodiments of this specification, such as Figure 1 As shown, the method for predicting precipitation using multi-source meteorological elements based on the spatiotemporal information conversion equation may include the following steps.

[0045] Step 110: Establish a forward-inverse dual model based on the spatiotemporal information conversion equation.

[0046] Figure 2 is a schematic structural diagram of a forward-inverse dual model based on a spatiotemporal information conversion equation according to some embodiments of this specification, such as Figure 2 As shown, in some embodiments, the forward-inverse dual model based on the space-time information conversion equation includes a data preprocessing module, a space-time converter and a space-time converter.

[0047] The data preprocessing module is used to preprocess multi-source meteorological element data and rainfall data to generate a delay embedding matrix, specifically including:

[0048] Performing spatiotemporal alignment on multi-source meteorological element data and rainfall data to generate spatiotemporal aligned multi-source meteorological element data and rainfall data;

[0049] Normalizing the multi-source meteorological element data and rainfall data after time and space alignment to generate normalized multi-source meteorological element data and rainfall data;

[0050] The normalized multi-source meteorological element data is subjected to delay embedding transformation to generate a delay embedding matrix.

[0051] Specifically, a set of multi-source meteorological element data with high correlation with precipitation events and a set of high-resolution precipitation data were prepared. Each set of meteorological data and precipitation data was jump-sampled at a time interval of 10 minutes, and then the spatial resolution was adjusted. The block size of all data was adjusted to 256×256 using bicubic interpolation to avoid increased complexity and unnecessary errors in the training process of the forward-inverse dual model based on the spatiotemporal information conversion equation.

[0052] Due to the differences in the attributes of different meteorological elements, there are huge differences in the numerical range. Directly inputting them into the space-time converter will cause training instability or difficulty in convergence. Therefore, each meteorological element data and precipitation data are normalized to the range of 0-1. In order to capture the short-term and medium- and long-term dependencies in the time series, the multi-source meteorological data are transformed into a delayed embedding matrix that conforms to the space-time information conversion equation through delayed embedding.

[0053] Performing a delay embedding transformation on the normalized multi-source meteorological element data to generate a delay embedding matrix may include the following steps:

[0054] S11. The multidimensional meteorological data input into the model is expressed in rectangular form. The expression is as follows:

[0055] ;

[0056] in, Represents the Radar echo image frames of different meteorological factors recorded by the radar sensor in the atmospheric system at the moment. is the number of variables in the atmospheric space system, is the observation length of the historical observation image sequence.

[0057] S12. Transform the meteorological variable matrix Z into a delayed embedding matrix that conforms to the spatiotemporal information conversion equation. The transformed matrix Y is formulated as follows:

[0058] ;

[0059] in, The dimension is The delayed embedding matrix, is the number of future image frames of the precipitation variable that needs to be predicted. In the lower right corner, the target variable is at time point The subsequent sequence All image frames in are future image frames that need to be predicted;

[0060] The space-time information conversion equation is:

[0061] ;

[0062] in, For the forward conversion, the historical observation sequence Mapping to a delayed embedding matrix , For the reverse transformation, it is used to embed the matrix from the delay Recover the estimated value of the historical observation series in .

[0063] The space-time converter is used to generate precipitation forecast images based on the delay embedding matrix.

[0064] In some embodiments, the space-time converter includes a first encoder and a second encoder, wherein the first encoder is used to perform convolution feature extraction on the delay embedding matrix to generate a spatial feature map, and the second encoder is used to generate a precipitation prediction image based on the spatial feature map.

[0065] Figure 3 is a schematic diagram of the structure of the first encoder according to some embodiments of this specification, such as Figure 3 As shown, in some embodiments, the first encoder includes multiple convolutional layers and an encoding ConvLSTM layer, wherein the multiple convolutional layers are used to extract the spatial information of the delay embedding matrix, and the encoding ConvLSTM layer is used to generate a spatial feature map based on the spatial information of the delay embedding matrix. The encoding ConvLSTM layer can adapt to small batches or unevenly distributed meteorological characteristics by capturing the time information of the time series, thereby improving the model's ability to capture the connection between different feature factors.

[0066] like Figure 3 As shown, in some embodiments, the first decoder includes multiple deconvolution layers and a decoding ConvLSTM layer, wherein the multiple deconvolution layers are used to generate precipitation distribution at future moments based on the spatial feature map, and the decoding ConvLSTM layer is used to generate a precipitation prediction image according to the precipitation distribution at future moments, and the decoding ConvLSTM layer can ensure the temporal continuity of future prediction results.

[0067] The spatiotemporal converter is used to extract the temporal features of precipitation forecast images and reconstruct the spatial distribution of multidimensional meteorological data based on the temporal features of precipitation forecast images.

[0068] In some embodiments, the spatiotemporal converter includes a second encoder and a second decoder, wherein the second encoder is used to extract the temporal characteristics of the precipitation forecast image, and the second decoder is used to reconstruct the spatial distribution of the multidimensional meteorological data based on the temporal characteristics of the precipitation forecast image. The structure of the second encoder is similar to that of the first encoder, and the structure of the second decoder is similar to that of the first decoder, and will not be repeated here. The inverse reconstruction process of the spatiotemporal converter increases the prediction robustness of the forward prediction process.

[0069] Step 120: Obtain multiple groups of training samples.

[0070] The training samples include multi-source meteorological element data, and the labels of the training samples are rainfall data.

[0071] Step 130 : training a forward-inverse dual model based on the spatiotemporal information conversion equation based on the adaptive weighted gradient loss function and multiple sets of training samples.

[0072] Specifically, the adaptive weighted gradient loss function includes a mean absolute error function and a gradient difference loss function based on adaptive weights.

[0073] The adaptive weighted gradient loss function is:

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] in, is the adaptive weighted gradient loss function, is the mean absolute error function, is the gradient difference loss function based on adaptive weights, is the pixel value of the pixel in the i-th row and j-th column in the precipitation prediction image output by the forward-inverse dual model based on the spatiotemporal information conversion equation, is the total number of rows, is the total number of columns, is the gradient change rate of the precipitation prediction image in the horizontal direction, is the gradient change rate of the precipitation prediction image in the vertical direction, is the gradient change rate of the real precipitation image in the horizontal direction, is the gradient change rate of the real precipitation image in the vertical direction, is the pixel value of the pixel in the i-1th row and jth column in the precipitation prediction image, is the pixel value of the pixel in the i-th row and j-1-th column in the precipitation prediction image, is the adaptive weight, || is the operation of the absolute value function. 、 The calculation method of 、 The calculation method of is similar and will not be repeated here.

[0079] The calculation formula of adaptive weight is:

[0080] ;

[0081] in, is the precipitation prediction image of the i-th training sample, is the real precipitation image of the i-th training sample, is the number of training samples.

[0082] For the prediction loss of the precipitation prediction task, the mean absolute error (MAE) loss is used to calculate the average difference between the model's predicted value and the true value. This allows the average difference between the predicted precipitation value and the true value to gradually decrease during the model's training process, thereby improving the prediction accuracy. The specific formula is as follows:

[0083] ;

[0084] in, Represents the model prediction results, represents the true value, and n represents the number of samples.

[0085] To alleviate the MAE loss's tendency to underestimate and overestimate precipitation in areas of light and heavy rainfall, the Gradient Difference Loss (GDLoss) is introduced. This calculates the spatial variation in precipitation between the predicted image and the target image, as well as the gradient difference in precipitation amount. This allows the model to better learn edge contours and details in the precipitation radar echo map. Large gradient differences indicate inconsistent intensity variations between the predicted and target precipitation information, and the GDL loss penalizes this discrepancy, ensuring that the predicted precipitation information is consistent with the true value during training. The GDLoss loss is adjusted using an adaptive weight α, ultimately resulting in the loss function for both the forward and reverse processes.

[0086] Combining the MAE loss function with the GDL loss function can not only use MAE to ensure the balanced accuracy of the overall prediction results, but also enhance edge information through GDL loss and focus on the gradient differences in complex precipitation areas; α represents a dynamic parameter for adaptively weighting the gradient, which is used to enhance the model's sensitivity to predictions in heavy precipitation areas. When the model underestimates precipitation in complex precipitation areas, α is greater than 1, which overall amplifies the GDL loss's ability to capture high-frequency precipitation information, dynamically optimizes the model's prediction accuracy for heavy precipitation areas, and reduces the model's cumulative error in precipitation prediction tasks.

[0087] The loss of the preliminary precipitation forecast results and the actual precipitation value is calculated. The loss of the reconstruction results of the multidimensional meteorological data is calculated with the true value of the real meteorological data. The forward and reverse processes are optimized through the gradient backpropagation mechanism until the loss of the entire training process converges to a certain level, and then the training can be stopped.

[0088] Step 140: Generate a multi-source meteorological element precipitation prediction model based on the trained forward-inverse dual model of the spatiotemporal information conversion equation.

[0089] Specifically, the multi-source meteorological element precipitation prediction model includes a delayed embedding transformation unit of a data preprocessing module of a forward-inverse dual model of a trained spatiotemporal information conversion equation and a space-time converter.

[0090] Step 150: Obtain multi-source meteorological element data to be predicted.

[0091] Step 160 : Generate a precipitation prediction image based on the multi-source meteorological element data to be predicted using a multi-source meteorological element precipitation prediction model.

[0092] The following experiments are combined to illustrate the effect of the method of predicting precipitation using multi-source meteorological elements based on the spatiotemporal information conversion equation.

[0093] The following experiments all use the multimodal radar dataset SEVIR (Storm Event Imagery), which records approximately 10,000 weather events (including storms and random meteorological events) that occurred in a specific region between 2017 and 2019. These weather events are recorded using high-temporal and spatial resolution imagery from the GOES-16 geostationary satellite and the NEXRAD weather radar. Each weather event consists of a four-hour image sequence, with precipitation data having a temporal resolution of 5 minutes and a spatial resolution of approximately 0.01°, equivalent to a horizontal resolution of 1 km × 1 km, providing relatively detailed regional meteorological information. This dataset combines and aligns different meteorological sensing methods into a single dataset, avoiding the large size of traditional radar and meteorological satellite datasets, which makes batch processing and computation difficult. This rich dataset includes four satellite sensors and records five different data types, as shown in Table 1.

[0094] Table 1

[0095]

[0096] Four types of data other than lightning events were selected for the experiment because the spatiotemporal resolution of lightning data differs significantly from that of the rest of the data. Normalization with other data will produce large errors, and the uncertainty of lightning on precipitation is too large, which has minimal impact on the learning ability of the model. Furthermore, in the VIL precipitation data, the value range of each image is 0-254, with 255 representing a missing value. The pixel values of each image are converted to the real units of VIL using the following conversion formula. It can be seen from the formula that the pixel values of VIL are positively correlated with the real units of precipitation. The formula is as follows:

[0097] ;

[0098] Where x is the pixel value of each precipitation radar echo image in the VIL data.

[0099] Finally, the dataset was partitioned using a ratio of 8:1:1 for the 10,000 weather events. The first 8,000 events, after random ordering, were used as the training dataset. Events 8,000-9,000 (a total of 1,000 events) were used for validation, and the final 1,000 events were used to test the model's performance. Furthermore, due to the 5-minute resolution of precipitation data, the difference between the previous and next images is minimal. To achieve greater differentiation in precipitation maps within each frame, skip sampling (sampling every 10 minutes) was performed for the four sensor types. The 60-minute meteorological data was then used to predict precipitation for the next 60 to 180 minutes.

[0100] To quantitatively evaluate the forward-inverse dual model based on the spatiotemporal information conversion equation, evaluation metrics from the image field of computer vision and meteorological science were used, as shown in Table 2. Specifically, the mean square error (MSE) was used as the evaluation metric, measuring the overall pixel value deviation between the predicted and true values using the average error between the predicted and true values. The peak signal-to-noise ratio (PSNR) analyzes the signal-to-noise ratio between images and assesses the image quality of the predicted images. The continuous ranked probability score (CRPS) measures the cumulative error between the distribution of the predicted and true images and is used to evaluate the model's ability to model uncertainty in precipitation forecasts. Furthermore, in meteorological science, different levels of precipitation are required, and accurate forecasts for these levels are extremely valuable. Therefore, based on the pixel values of the VIL, the indicator discrimination thresholds were set to [16, 74, 133, 160, 181, 219]. The corresponding true precipitation values were [0.15, 0.78, 3.53, 7.07, 12.14, 32.23], expressed in kg / m². Furthermore, the Critical Success Index (CSI), False Alarm Rate (FAR), Heidke Skill Score (HSS), and Brier Skill Score (BSS) were used to measure the false alarm rate and accuracy of precipitation forecasts for all methods at different thresholds.

[0101] Table 2

[0102]

[0103] in, is the total number of samples, represents the true value of the radar echo image, Indicates the corresponding predicted value; Represents the maximum value of all pixel values; Represents the probability distribution of real-time images; is the unit step function; is the hit number, indicating that the model predicted precipitation and precipitation actually occurred. The specific mathematical representation is prediction = 1, true value = 1; is the number of falsely reported precipitation events, which means that the model predicted precipitation but no precipitation actually occurred. The specific mathematical representation is prediction = 1, true value = 0; is the number of missed precipitation events, mathematically expressed as prediction = 0, true value = 1; True negative, indicating that there is no precipitation in the forecast and no precipitation in reality, mathematically expressed as prediction = 0, true value = 0; is the predicted probability for day i, is the actual result on day i, is the benchmark Blair score, is the Blair score of the prediction model.

[0104] This study compared precipitation prediction performance with other methods, using multi-source meteorological data from the previous hour as input to predict precipitation for the next 60 minutes. Six frames of forecast images were output, each with a 10-minute interval. To fully validate the quality and performance of the forward-inverse dual model based on the spatiotemporal information transformation equation (STI-DRNN) for high-resolution short-term precipitation forecasting and to measure the STI-DRNN model's applicability for short-term forecasting of different precipitation levels, a quantitative comparison was conducted using leading-edge methods in the field of spatiotemporal forecasting. These methods include ConvLSTM, PredRNN, PhyNet, Rainformer, Earthformer, SimVP, TAU, and PastNet. These are primarily based on CNN, RNN, and Transformer-based spatiotemporal forecasting methods. To ensure the accuracy of the experimental results, all methods were evaluated and tested on the same test set, using the model with the lowest loss function after training. Table 3 shows the one-hour forecast results of the forward-inverse dual model based on the spatiotemporal information conversion equation and other methods on the SEVIR dataset.

[0105] Table 3

[0106]

[0107] The quantitative evaluation results in the table above show that the forward-inverse dual model based on the spatiotemporal information conversion equation demonstrates superior performance in short-term forecasting and produces higher prediction accuracy than other methods. In the basic 1-hour precipitation forecast, the RNN-based ConvLSTM model and the CNN-based SimVP model achieve relatively large MSE scores and relatively small PSNR scores, indicating that the CNN- and RNN-based models are less capable of extracting precipitation information. Furthermore, the Rainformer and Earthformer models achieve higher MSE values and lower PSNR values than the STI-DRNN model, indicating that the forward-inverse dual model based on the spatiotemporal information conversion equation outperforms the Transformer approach. Furthermore, to further verify the efficient uncertainty modeling capabilities of the forward-inverse dual model based on the spatiotemporal information conversion equation, the CRPS metric was used to measure the error between the probability distributions of the predicted images and the true images obtained by different methods. As shown in the table, the STI-DRNN model improves CRPS by 9.39% compared to the worst model and by 4.06% compared to the next-best PredRNN model. This indicates that the precipitation forecast results using the STI-DRNN model have a probability distribution that is most similar to the real image, indicating that the model has a stronger uncertainty modeling capability. Next, the average FAR, CSI, and HSS scores of all methods at a 60-minute lead time were calculated. The STI-DRNN model has the lowest FAR-M metric, indicating the lowest false alarm rate, while the STI-DRNN model has the highest CSI-M and HSS-M metrics, indicating that the STI-DRNN model has the best overall forecast performance and is more accurate. Specifically, for short-term forecasts, the STI-DRNN model improves CSI-M by 10.03%, HSS-M by 11.07%, and FAR-M by 9.71% compared to the best-performing PredRNN method.

[0108] Since the meteorological field usually focuses on the model performance under different precipitation thresholds, it is necessary to measure the performance of all methods under different levels of precipitation. Therefore, the CSI scores and HSS scores of each method under different thresholds are measured separately, such as Figure 4 (a) and Figure 4 (b) As shown in Figure 2, the Rainformer method is used as the baseline, and the percentage of improvement in CSI and HSS scores of all other methods compared to the Rainformer method is measured. As the precipitation threshold gradually increases, the difficulty of reconstructing high-intensity radar echoes for all methods also increases. Figure 4As can be seen in the figure, STI-DRNN shows positive improvements over Rainformer at all thresholds, with the maximum improvement at each threshold. In particular, as the precipitation threshold increases, the difficulty of precipitation forecasting increases. However, the percentage improvement of the STI-DRNN model shows a positive correlation with the precipitation threshold, indicating that the STI-DRNN model is more suitable for forecasting moderate-intensity precipitation.

[0109] From the above experimental analysis, we can see that STI-DRNN has better overall prediction performance in medium-to-heavy precipitation levels. To this end, we draw time-effect analysis diagrams of indicators with precipitation thresholds of 181 and 219 to further verify the prediction details of STI-DRNN in medium-to-heavy precipitation levels, such as Figure 5 (a) to Figure 5 (c). Figure 5 (a) to Figure 5 (c) As can be seen, as the forecast time increases, the model's uncertainty about the future also increases, causing the forecast performance of all methods to gradually decline over time. However, the STI-DRNN model outperforms the other methods overall across all thresholds, with a smaller downward trend in the forecast indicators. Conversely, the performance gap between STI-DRNN and other methods widens as the forecast time increases. Specifically, as shown in Table 4, compared with PredRNN, a model with better overall forecasting capabilities, the CSI score improves by 27.77% at a threshold of 181 and by 82.64% at a threshold of 219. The HSS score also improves by 25.75% at a threshold of 181 and by 84.06% at a threshold of 219. This demonstrates that STI-DRNN has a stronger ability to capture high-intensity precipitation levels, further validating the model's superiority.

[0110] Table 4

[0111]

[0112] Among them, 181 and 219 correspond to the precipitation threshold of 12.14 kg / m 2 and 32.23kg / m 2 .

[0113] To validate STI-DRNN's ability to predict real-world precipitation, we randomly selected extreme precipitation events, which are rare in everyday life, for testing. Furthermore, to clearly demonstrate whether the model's predictions overestimate or underestimate the actual image, we added a heat map of the error between the sixth frame of the predicted image and the sixth frame of the actual image to the visualization. Red areas indicate underestimation, while blue areas indicate overestimation.

[0114] Depend on Figure 6 As can be seen, as the forecast time increases, all methods gradually lose detailed information about precipitation, especially the PhydNet method, which suffers the most severe loss of information about precipitation events. Furthermore, compared to other models, while other methods achieve basic visual quality, they all experience more underestimation and overestimation of the actual precipitation image. Furthermore, these methods exhibit greater discrepancies between the specific location and general shape of the precipitation area and the actual image, resulting in blurred predictions and inaccurate positions in darker areas (yellow and above). In contrast, while STI-DRNN inevitably still suffers from underestimation and overestimation, its predictions are closer to the true values, and both overestimation and underestimation are reduced. This demonstrates that the STI-DRNN method efficiently transforms detailed information in spatiotemporal data, resulting in superior spatiotemporal modeling capabilities for moderate to heavy precipitation scenarios compared to other methods.

[0115] As the forecast lead time gradually increases, STI-DRNN's accuracy in predicting weather events remains optimal, and this holds true across different precipitation thresholds. Furthermore, STI-DRNN's power spectrum curve is nearly consistent with the true value within the wavelength range of 64 km to 256 km, while other models deviate from the true value at a wavelength of 128 km. This indicates that the amount of information provided by STI-DRNN is nearly consistent with the true value at large spatial scales. At small scales, all methods produce low PSD scores, which appear to be independent of the forecast lead time. This suggests that deep learning methods introduce smoother and more blurred precipitation fields at small scales, but that STI-DRNN's blurring effect is lower than that of other methods. Overall, STI-DRNN is more applicable to large-scale extreme precipitation events.

[0116] The SEVIR dataset includes various additional meteorological variables, such as infrared brightness temperature and visible light. In the experiments above, a total of four different meteorological variables were used as model input. To verify the impact of using additional meteorological variables on the STI training framework, the type and number of input meteorological variables were systematically eliminated. The experimental results are shown in Table 5. According to the delay theorem, as the spatial dimensions of the input data increase, the model can learn more spatial interaction features, thereby expanding the model's nonlinear mapping of future precipitation temporal information. The gradual reduction in the number of meteorological variables is correlated with a downward trend in the performance of the STI-DRNN across all evaluation metrics. This decline is particularly pronounced in the FAR and CSI metrics, indicating that using more meteorological variables helps the model effectively map spatial information to the temporal information of future precipitation variables. It also enables the model to capture the complex interactions and dependencies between atmospheric processes involved in precipitation, improving the accuracy of precipitation forecasts.

[0117] Table 5

[0118]

[0119] To clearly validate the optimization capabilities of the spatiotemporal transformer for auxiliary training of precipitation forecasts, the effectiveness of the spatiotemporal transformer was verified by ablating it and adding it to other methods with good forecasting performance. Although current deep learning methods can achieve good forecast results, due to the rapid pace of climate change, past meteorological information may not accurately reflect current atmospheric conditions. Existing precipitation forecast methods using short-term time data have poor prediction performance for heavy precipitation. Introducing the spatiotemporal transformation equation into precipitation forecasting methods can convert multidimensional spatial data into future time information, solving the problem of insufficient short-term data information and ultimately improving the performance of short-term precipitation forecasts. The experimental results are shown in Table 6.

[0120] Table 6

[0121]

[0122] Among them, ConvLSTM-STI is ConvLSTM with a spatiotemporal converter, PredRNN-STI is PredRNN with a spatiotemporal converter, PastNet-STI is PastNet with a spatiotemporal converter, and STI-DRNN-NoSTI is STI-DRNN with the spatiotemporal converter removed.

[0123] Experimental results show that after incorporating the space-time converter, all methods have achieved significant improvements in each quantitative indicator. In particular, the ConvLSTM method has achieved improvements of 13.05% and 9.20% in MSE and CSI scores, respectively. The CRPS score has also increased by 3.58% after combining with STI. These results demonstrate the effectiveness and versatility of the space-time converter in spatiotemporal dynamic modeling. The model incorporating the space-time converter can significantly improve prediction accuracy and robustness, thereby producing more accurate precipitation forecast results.

[0124] In addition, in order to verify the improvement effect of the spatiotemporal converter on precipitation events of different intensities, the performance improvement scatter plots of the HSS and CSI indicators of several methods under all threshold conditions are drawn, such as Figure 7 (a) and Figure 7 (b) Figure 7 (a) and Figure 7 As can be seen in (b), after incorporating the spatiotemporal converter, all methods produced positive improvements in CSI and HSS, especially in the cases of moderate and high precipitation, where the improvement effect was most significant. This proves the effectiveness of the STI dual training structure, which can make up for the problem of insufficient short-term data information and improve the overall prediction performance of the model.

[0125] Due to the rapid climate change in the real world, commonly used regression loss functions often ignore sudden rainstorms, resulting in more underestimation of moderate to heavy rainfall events. To this end, an adaptive weighted gradient loss function (ADGLoss) is proposed, which can effectively optimize the model's sensitivity to heavy rainfall events and improve prediction accuracy. In order to verify the effectiveness of ADGLoss, the model is trained using L1Loss, L2Loss, and ADGLoss without weights as loss functions to evaluate its effectiveness. The experimental results are shown in Tables 7 and 8. Figure 8 (a) to Figure 8 (c) shown.

[0126] Table 7

[0127]

[0128] The quantitative evaluation results of the above ablation loss functions show that the adaptive weighted gradient loss function does not achieve optimal results in terms of MSE loss. This is due to the long-tail distribution of precipitation data. The optimal solution of L1Loss may show more robust results in the squared error space of MSE. Although the adaptive weighted gradient loss function did not achieve optimal results in terms of MSE, it achieved greater advantages in meteorological precipitation indicators, especially the FAR score, which improved by 6.41% compared to the suboptimal loss function. In addition, the indicator curves for different time periods also show that ADGLoss leads the way in high-threshold precipitation conditions and has the least performance degradation with increasing prediction time. This indicates that the adaptive weighted gradient loss function can better alleviate the problem of increasing model error as prediction time increases, thereby improving the prediction accuracy of moderate to heavy precipitation.

[0129] Finally, it should be understood that the embodiments described in this specification are intended only to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, by way of example and not limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly described and illustrated in this specification.

Claims

1. A method for predicting precipitation using multi-source meteorological elements based on a spatiotemporal information conversion equation, characterized in that: include: Establish a forward-inverse dual model based on the space-time information conversion equation; Acquire multiple sets of training samples, wherein the training samples include multi-source meteorological element data, and the labels of the training samples are rainfall data; Based on the adaptive weighted gradient loss function and multiple sets of training samples, the forward-inverse dual model based on the spatiotemporal information conversion equation is trained; Generate a multi-source meteorological element precipitation prediction model based on the forward-inverse dual model of the trained spatiotemporal information conversion equation; Obtain multi-source meteorological element data to be predicted; Generate a precipitation forecast image based on the multi-source meteorological element data to be predicted using a multi-source meteorological element prediction model; The forward-inverse dual model based on the space-time information conversion equation includes a data preprocessing module, a space-time converter and a space-time converter; The data preprocessing module is used to preprocess multi-source meteorological element data and rainfall data to generate a delay embedding matrix; The space-time converter is used to generate a precipitation prediction image based on the delay embedding matrix; The spatiotemporal converter is used to extract the temporal features of the precipitation forecast image and reconstruct the spatial distribution of the multidimensional meteorological data based on the temporal features of the precipitation forecast image; The adaptive weighted gradient loss function includes a mean absolute error function and a gradient difference loss function based on adaptive weights; The adaptive weighted gradient loss function is: ； ； ； ； in, is the adaptive weighted gradient loss function, is the mean absolute error function, is the gradient difference loss function based on adaptive weights, is the pixel value of the pixel in the i-th row and j-th column in the precipitation prediction image output by the forward-inverse dual model based on the spatiotemporal information conversion equation, is the total number of rows, is the total number of columns, is the gradient change rate of the precipitation prediction image in the horizontal direction, is the gradient change rate of the precipitation prediction image in the vertical direction, is the gradient change rate of the real precipitation image in the horizontal direction, is the gradient change rate of the real precipitation image in the vertical direction, is the pixel value of the pixel in the i-1th row and jth column in the precipitation prediction image, is the pixel value of the pixel in the i-th row and j-1-th column in the precipitation prediction image, is the adaptive weight.

2. The method for predicting precipitation using multiple meteorological elements based on a spatiotemporal information conversion equation according to claim 1, wherein: The data preprocessing module preprocesses the multi-source meteorological element data and rainfall data to generate a delay embedding matrix, including: Performing spatiotemporal alignment on multi-source meteorological element data and rainfall data to generate spatiotemporal aligned multi-source meteorological element data and rainfall data; Normalizing the multi-source meteorological element data and rainfall data after time and space alignment to generate normalized multi-source meteorological element data and rainfall data; The normalized multi-source meteorological element data is subjected to delay embedding transformation to generate a delay embedding matrix.

3. The method for predicting precipitation using multiple meteorological elements based on a spatiotemporal information conversion equation according to claim 1, wherein: The space-time converter includes a first encoder and a first decoder, wherein the first encoder is used to perform convolution feature extraction on the delay embedding matrix to generate a spatial feature map, and the first decoder is used to generate a precipitation prediction image based on the spatial feature map.

4. The method for predicting precipitation using multiple meteorological elements based on a spatiotemporal information conversion equation according to claim 3, wherein: The first encoder includes multiple convolutional layers and an encoding ConvLSTM layer, wherein the multiple convolutional layers are used to extract spatial information of the delay embedding matrix, and the encoding ConvLSTM layer is used to generate a spatial feature map based on the spatial information of the delay embedding matrix.

5. The method for predicting precipitation using multiple meteorological elements based on spatiotemporal information conversion equations according to claim 3, characterized in that: The first decoder includes multiple deconvolution layers and a decoding ConvLSTM layer, wherein the multiple deconvolution layers are used to generate precipitation distribution at future moments based on spatial feature maps, and the decoding ConvLSTM layer is used to generate a precipitation prediction image based on the precipitation distribution at future moments.

6. The method for predicting precipitation using multiple meteorological elements based on a spatiotemporal information conversion equation according to any one of claims 2 to 5, wherein: The spatiotemporal converter includes a second encoder and a second decoder, wherein the second encoder is used to extract the temporal features of the precipitation prediction image, and the second decoder is used to reconstruct the spatial distribution of multidimensional meteorological data based on the temporal features of the precipitation prediction image.

7. The method for predicting precipitation using multiple meteorological elements based on spatiotemporal information conversion equations according to claim 1, characterized in that: The calculation formula of the adaptive weight is: ； in, is the precipitation prediction image of the i-th training sample, is the real precipitation image of the i-th training sample, is the number of training samples.

Images