Multi-source meteorological element rainfall prediction method based on spatio-temporal information conversion equation

Abstract

The invention provides a multi-source meteorological element rainfall prediction method based on a spatio-temporal information conversion equation, and relates to the field of rainfall prediction, and the method comprises the steps: building a positive-inverse dual model based on the spatio-temporal information conversion equation; multiple groups of training samples are obtained, the training samples comprise multi-source meteorological element data, and labels of the training samples are rainfall data; based on the adaptive weighted gradient loss function and the multiple groups of training samples, training a forward-inverse dual model based on a space-time information conversion equation; generating a multi-source meteorological element prediction rainfall model based on a forward-inverse dual model of the trained spatio-temporal information conversion equation; obtaining to-be-predicted multi-source meteorological element data; the rainfall prediction image is generated based on the to-be-predicted multi-source meteorological element data through the multi-source meteorological element prediction rainfall model, and the method has the advantage of improving the rainfall prediction accuracy.

Description

Technical Field

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

[0002] Precipitation Forecasting refers to quantitatively or qualitatively predicting the precipitation type (such as rain, snow, hail, etc.), intensity (such as light rain, heavy rain) and spatio-temporal distribution in a specific area within a certain future 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 resource management and public life in advance.

[0003] In the prior art, numerical weather prediction uses sub-grid scale parameterization to approximate unresolved physical processes. Although reducing the grid size can improve the spatial resolution, this requires a multiple increase in computing resources to solve the fine-scale atmospheric interactions. In addition, the sensitivity of numerical weather prediction to initial condition perturbations causes the prediction trajectory to quickly deviate, and these limitations hinder its application to short-term precipitation forecasting. And for precipitation prediction methods based on radar echo extrapolation, under highly non-linear and complex atmospheric conditions, iterative extrapolation will amplify the position and intensity errors over time. This error propagation will lead to a significant decrease in prediction accuracy, especially for local heavy precipitation events. Also, due to the rapid rate of climate change, long-term time series data is usually not sufficient to accurately solve the precipitation prediction of heavy precipitation weather events. Currently, most methods that rely on short-term data for prediction perform poorly in predicting heavy precipitation events due to the problem of insufficient short-term data information.

[0004] Therefore, a method for predicting precipitation of multi-source meteorological elements based on a spatio-temporal information conversion equation is needed to solve the problem of inaccurate forecasting of heavy precipitation areas caused by insufficient short-term data information. Summary of the Invention

[0005] The present invention provides a method for predicting precipitation of multi-source meteorological elements based on a spatio-temporal information conversion equation, including: establishing a forward-inverse dual model based on the spatio-temporal information conversion equation; obtaining multiple groups of training samples, where the training samples include multi-source meteorological element data, and the label of the training samples is rainfall data; training the forward-inverse dual model based on the spatio-temporal 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 spatio-temporal information conversion equation; obtaining multi-source meteorological element data to be predicted; generating a precipitation prediction image through 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 spatio-temporal information conversion equation includes a data preprocessing module, a spatio-temporal 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 delayed embedding matrix; the spatio-temporal converter is used to generate a precipitation prediction image based on the delayed embedding matrix; the space-time converter is used to extract the temporal features of the precipitation prediction image and reconstruct the spatial distribution of multi-dimensional meteorological data according to the temporal features of the precipitation prediction image.

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

[0008] Furthermore, the spatio-temporal converter includes a first encoder and a second encoder. Among them, the first encoder is used to perform convolutional feature extraction on the delayed 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. Among them, the multiple convolutional layers are used to extract the spatial information of the delayed embedding matrix, and the encoding ConvLSTM layer is used to generate a spatial feature map based on the spatial information of the delayed embedding matrix.

[0010] Furthermore, the first decoder includes multiple deconvolutional layers and a decoding ConvLSTM layer. Among them, the multiple deconvolutional layers are used to generate the precipitation distribution at future times 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 times.

[0011] Furthermore, the space-time converter includes a second encoder and a second decoder. Among them, 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 multi-dimensional meteorological data according to the temporal features 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: ; ; ; ; wherein, is an adaptive weighted gradient loss function, is a mean absolute error function, is a gradient difference loss function based on adaptive weights, is the pixel value of the pixel at the i-th row and j-th column in the precipitation prediction image output by the forward-inverse dual model based on the spatio-temporal 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 true precipitation image in the horizontal direction, is the gradient change rate of the true precipitation image in the vertical direction, is the pixel value of the pixel at the (i - 1)-th row and j-th column in the precipitation prediction image, is the pixel value of the pixel at the i-th row and (j - 1)-th column in the precipitation prediction image, is the adaptive weight.

[0014] Furthermore, the calculation formula of the adaptive weight is: ; wherein, is the precipitation prediction image of the i-th training sample, is the true precipitation image of the i-th training sample, is the number of training samples.

[0015] Compared with the prior art, the method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation provided by the present invention has at least the following beneficial effects: Traditional precipitation prediction methods usually rely on short-term historical data and are difficult to fully capture the complex spatio-temporal evolution law of heavy precipitation. By means of the spatio-temporal information conversion equation and the delay embedding conversion mechanism, the present invention maps multi-dimensional spatial information into the time information of future precipitation, effectively integrating the long-term spatio-temporal dependence relationship and solving the problem of insufficient short-term data information. This enables the model to more comprehensively understand the formation mechanism of heavy precipitation, thereby generating more accurate prediction results.

[0016] In long-term precipitation forecasting, errors tend to accumulate over time steps, causing the prediction results to gradually deviate from the actual situation. Through the design of a forward-inverse dual model, the present invention forms a closed-loop feedback mechanism: the spatio-temporal converter is responsible for predicting future precipitation information based on historical data. The space-time converter then attempts to reconstruct the input features from the predicted precipitation information. This two-way constraint effectively suppresses the propagation of errors, enabling the model to maintain a low cumulative error in long-term sequence prediction and improving the stability of long-term precipitation forecasting. The design of the forward-inverse dual model forms a closed-loop constraint, avoiding the overfitting risk of a single prediction model and enhancing the generalization ability of the model 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 stronger practical value.

[0017] Strong precipitation regions usually have a high degree of spatio-temporal heterogeneity, and traditional methods are prone to underestimating precipitation intensity due to the loss of local features. The present invention comprehensively captures the characteristics of strong precipitation regions by integrating multi-source meteorological element data. By constraining the difference between the predicted precipitation image and the target image in the gradient domain, the model's capture of the gradients at the edges and within strong precipitation regions is strengthened, avoiding the underestimation of extreme precipitation.

[0018] Through the delay embedding conversion mechanism, the historical meteorological element sequence is converted into a delay embedding matrix, explicitly encoding the spatio-temporal dependence relationship. The collaborative action of the spatio-temporal converter and the space-time converter realizes the two-way conversion from spatial features to time series and establishes a stable mapping relationship. This enables the model to learn a more robust spatio-temporal feature representation and enhances the generalization ability to complex precipitation patterns.

[0019] By optimizing the model with an adaptive weighted gradient loss function (such as combining MAE and GDLoss), the present invention balances the global error and local details, while enhancing the interpretability and robustness of the model using the dual training framework. Finally, the model can generate high-resolution and high-precision precipitation prediction images, accurately depicting the spatio-temporal distribution of precipitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] This specification will be further described by way of exemplary embodiments, which will be described in detail through the accompanying drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where: Figure 1 is a schematic flowchart of a method for predicting precipitation of multi-source meteorological elements based on a spatio-temporal information conversion equation shown in some embodiments of this specification; Figure 2 is a schematic structural diagram of a forward-inverse dual model based on a spatio-temporal information conversion equation shown in some embodiments of this specification; Figure 3Schematic diagram of the structure of the first encoder shown in some embodiments of this specification; Figure 4 (a) Schematic diagram of CSI scores of all comparison methods shown in some embodiments of this specification under different precipitation threshold divisions; Figure 4 (b) Schematic diagram of HSS scores of all comparison methods shown in some embodiments of this specification under different precipitation threshold divisions; Figure 5 (a) Schematic diagram of CSI scores at different prediction times of all comparison methods shown in some embodiments of this specification under high precipitation thresholds; Figure 5 (b) Schematic diagram of HSS scores at different prediction times of all comparison methods shown in some embodiments of this specification under high precipitation thresholds; Figure 5 (c) Schematic diagram of FAR scores at different prediction times of all comparison methods shown in some embodiments of this specification under high precipitation thresholds; Figure 6 Visual comparison diagram of the overall effects of all comparison methods shown in some embodiments of this specification in precipitation cases; Figure 7 (a) Schematic diagram of CSI scores after eliminating the spatio-temporal converter and adding the spatio-temporal converter by other methods shown in some embodiments of this specification; Figure 7 (b) Schematic diagram of HSS scores after eliminating the spatio-temporal converter and adding the spatio-temporal converter by other methods shown in some embodiments of this specification; Figure 8 (a) Schematic diagram of CSI scores after ablation of the loss function shown in some embodiments of this specification; Figure 8 (b) Schematic diagram of HSS scores after ablation of the loss function shown in some embodiments of this specification; Figure 8 (c) Schematic diagram of BSS scores after ablation of the loss function shown in some embodiments of this specification. Detailed implementation

[0021] To more clearly illustrate the technical solutions of the embodiments of this specification, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some examples or embodiments of this specification. For those of ordinary skill in the art, without creative efforts, this specification can also be applied to other similar scenarios based on these drawings. Unless obvious from the language context or otherwise stated, the same reference numerals in the figures represent the same structures or operations.

[0022] Figure 1 is a schematic flowchart of a method for predicting precipitation of multi-source meteorological elements based on a spatio-temporal information conversion equation as shown in some embodiments of this specification. As Figure 1 shown, the method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation may include the following steps.

[0023] Step 110, establish a forward-inverse dual model based on the spatio-temporal information conversion equation.

[0024] Figure 2 is a schematic structural diagram of a forward-inverse dual model based on the spatio-temporal information conversion equation as shown in some embodiments of this specification. As Figure 2 shown, in some embodiments, the forward-inverse dual model based on the spatio-temporal information conversion equation includes a data preprocessing module, a spatio-temporal converter, and a time-space converter.

[0025] The data preprocessing module is used to preprocess the multi-source meteorological element data and rainfall data to generate a delay embedding matrix, specifically including: Perform spatio-temporal alignment on the multi-source meteorological element data and rainfall data to generate spatio-temporally aligned multi-source meteorological element data and rainfall data; Normalize the spatio-temporally aligned multi-source meteorological element data and rainfall data to generate normalized multi-source meteorological element data and rainfall data; Perform delay embedding transformation on the normalized multi-source meteorological element data to generate a delay embedding matrix.

[0026] Specifically, prepare a multi-source meteorological element data highly correlated with precipitation events and a precipitation data with high resolution. Jump-sample each meteorological data and precipitation data at 10-minute time intervals, then adjust their spatial resolutions, and use bicubic interpolation to adjust the block size of all data to 256×256 to avoid increased complexity and unnecessary errors during the training process of the forward-inverse dual model based on the spatio-temporal information conversion equation.

[0027] Due to the differences in the attributes of different meteorological elements, there are huge differences in the numerical range. Directly inputting into the spatio-temporal converter will cause unstable training or difficult convergence. Therefore, each piece of meteorological element data and precipitation data is respectively normalized to the range of 0-1 by maximum-minimum normalization; in order to capture short-term and medium- and long-term dependencies in the time series, the multi-source meteorological data is then transformed through delay embedding into a delay embedding matrix that conforms to the spatio-temporal information conversion equation.

[0028] Performing delay embedding transformation on the normalized multi-source meteorological element data to generate a delay embedding matrix may include the following steps: S11. Express the multi-dimensional meteorological data input into the model in a rectangular form, and its expression is as follows: ; Among them, represents the radar echo image frames of different meteorological factors recorded by the radar sensor in the atmospheric system at time. is the number of variables in the atmospheric space system, is the observation length of the historical observation image sequence.

[0029] S12. Delay embed the meteorological variable matrix Z into a delay embedding matrix that conforms to the spatio-temporal information conversion equation. The formula for the transformed matrix Y is as follows: ; Among them, is the delay embedding matrix with a dimension of , is the number of future image frames of the precipitation variable to be predicted. At the lower right corner of the delay embedding matrix , all the image frames in the sequence after the time point of the target variable are the future image frames to be predicted; The spatio-temporal information conversion equation is: ; Among them, is the forward conversion, which maps the historical observation sequence to the delay embedding matrix , is the reverse conversion, which is used to recover the estimated value of the historical observation sequence from the delay embedding matrix .

[0030] The spatio-temporal converter is used to generate a precipitation prediction image based on the delay embedding matrix.

[0031] In some embodiments, the spatio-temporal converter includes a first encoder and a second encoder. The first encoder is used to perform convolutional 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.

[0032] Figure 3 is a schematic structural diagram of the first encoder shown in some embodiments of this specification. As Figure 3 shown, in some embodiments, the first encoder includes a plurality of convolutional layers and an encoding ConvLSTM layer. The plurality of 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. By capturing the temporal information of the time series, the encoding ConvLSTM layer can adapt to small-batch or unevenly distributed meteorological features and improve the model's ability to capture the connections between different feature factors.

[0033] As Figure 3 shown, in some embodiments, the first decoder includes a plurality of deconvolutional layers and a decoding ConvLSTM layer. The plurality of deconvolutional layers are used to generate the precipitation distribution at future times 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 times. The decoding ConvLSTM layer can ensure the temporal continuity of the future prediction results.

[0034] The spatio-temporal converter is used to extract the temporal features of the precipitation prediction image and reconstruct the spatial distribution of multi-dimensional meteorological data according to the temporal features of the precipitation prediction image.

[0035] In some embodiments, the spatio-temporal converter includes a second encoder and a second decoder. 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 multi-dimensional meteorological data according to the temporal features of the precipitation prediction 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, which will not be elaborated here. The reverse reconstruction process of the spatio-temporal converter increases the prediction robustness of the forward prediction process.

[0036] Step 120, obtain multiple groups of training samples.

[0037] Among them, the training samples include multi-source meteorological element data, and the label of the training samples is rainfall data.

[0038] Step 130, train a forward-inverse dual model based on the spatio-temporal information conversion equation based on the adaptive weighted gradient loss function and multiple groups of training samples.

[0039] Specifically, the adaptive weighted gradient loss function includes a mean absolute error function and a gradient difference loss function based on adaptive weights.

[0040] The adaptive weighted gradient loss function is as follows: ; ; ; ; Among them, is the adaptive weighted gradient loss function, is the mean absolute error function, is the gradient difference loss function based on the adaptive weight, is the pixel value of the pixel at the i-th row and j-th column in the precipitation prediction image output by the forward-inverse dual model based on the spatio-temporal 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 at the (i - 1)-th row and j-th column in the precipitation prediction image, is the pixel value of the pixel at the i-th row and (j - 1)-th column in the precipitation prediction image, is the adaptive weight, and || is the operation of the absolute value function. , The calculation methods of , are similar to those of

[0041] The calculation formula of the adaptive weight is as follows: ; Among them, 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.

[0042] For the prediction loss of the precipitation prediction task, the mean absolute error (MAE) loss is used to calculate the average gap between the model prediction value and the real value, so that the average gap between the predicted precipitation value and the real value gradually decreases during the progressive training of the model, improving the prediction accuracy. The specific formula is as follows: ; Among them, Represents the model prediction result, represents the true value, and n represents the number of samples.

[0043] To alleviate the problems of underestimated and overestimated precipitation in the weak precipitation and strong precipitation regions by the MAE loss, the Gradient Difference Loss (GDLoss) is introduced to calculate the precipitation change in space between the predicted image and the target image and the difference in precipitation on the gradient, enabling the model to better learn the edge contours and detailed information in the precipitation radar echo map. If the gradient difference is large, it indicates that the intensity change of the predicted precipitation information is inconsistent with that of the target precipitation information, and the GDLoss will penalize it, thus achieving the consistency between the predicted precipitation information and the true value features during the training process. The loss value of the GDLoss is adjusted using the adaptive weight α, and finally the loss functions for the entire forward process and reverse process are obtained.

[0044] Combining the MAE loss function and the GDLoss function can not only use the MAE to ensure the balanced accuracy of the overall prediction result, but also enhance the edge information through the GDLoss and pay attention to the gradient differences in complex precipitation regions; α represents the dynamic parameter for adaptive gradient weighting, which is used to enhance the sensitivity of the model to the prediction in strong precipitation regions. When the model underestimates precipitation in complex precipitation regions, α is greater than 1, which overall amplifies the ability of the GDLoss to capture high-frequency precipitation information, dynamically optimizes the prediction accuracy of the model in strong precipitation regions, and reduces the cumulative error of the model for the precipitation prediction task.

[0045] Calculate the loss between the obtained preliminary precipitation prediction result and the true precipitation value, and calculate the loss between the reconstruction result of the multi-dimensional meteorological data and the true value of the real meteorological data. Optimize the forward process and the reverse process 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.

[0046] Step 140, based on the forward-inverse dual model of the trained spatio-temporal information conversion equation, generate a multi-source meteorological element prediction precipitation model.

[0047] Specifically, the multi-source meteorological element prediction precipitation model includes the delay embedding transformation unit and the spatio-temporal converter in the data preprocessing module of the forward-inverse dual model of the trained spatio-temporal information conversion equation.

[0048] Step 150, obtain the multi-source meteorological element data to be predicted.

[0049] Step 160, based on the multi-source meteorological element data to be predicted, generate a precipitation prediction image through the multi-source meteorological element prediction precipitation model.

[0050] The following is an illustration of the effect of the method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation in combination with experiments.

[0051] The following experiments all use the multi-modal radar dataset SEVIR (Storm Event Imagery), which records approximately 10,000 weather events (including storm events and random meteorological events) that occurred in a certain region between 2017 and 2019. These weather events are respectively high spatio-temporal resolution image data recorded by the GOES-16 geostationary satellite and the NEXRAD weather radar. Each weather event is an image sequence containing 4 hours, where the time resolution of precipitation data is 5 minutes, and the spatial resolution is approximately 0.01°, equivalent to a horizontal resolution of 1 km × 1 km, which can provide relatively detailed regional meteorological information. This dataset combines and aligns different meteorological sensing methods into one dataset, avoiding the problem that traditional radar and meteorological satellite datasets are too large to be processed and calculated in batches. The sensor types of this rich dataset include four satellite sensors and record five different data types, as shown in Table 1.

[0052] Table 1

[0053] Four types of data other than lightning events are selected for experiments because the spatio-temporal resolution of lightning data is too different from that of the remaining data, and large errors will occur during the standardization process with other data. Moreover, the uncertainty of lightning regarding precipitation is too large, and it has the least improvement on the learning ability of the model. In addition, in the VIL precipitation data, the numerical range of each image is 0 - 254, and 255 represents a missing value. The pixel values of each image are converted to the true unit of VIL according to the following conversion formula, and it can be seen from the formula that the pixel values of VIL are positively correlated with the true unit of precipitation. The formula is as follows: ; where x is the pixel value of each precipitation radar echo image in the VIL data.

[0054] Finally, divide the dataset of 10,000 weather events according to the allocation ratio of 8:1:1. Use the data of the first 8,000 weather events after randomly shuffling the order as the training dataset, the data of 1,000 weather events from 8,000 to 9,000 for validation, and the data of the last 1,000 weather events to evaluate the performance of the model. In addition, due to the 5-minute resolution of precipitation data resulting in little difference between two consecutive images, in order to make the precipitation maps of each frame have greater differences, jump sampling (sampling every 10 minutes) is performed on the four sensor types, and then 60-minute meteorological data is used to predict the precipitation information in the next 60 - 180 minutes.

[0055] To quantitatively evaluate the forward-inverse dual model based on the spatio-temporal information conversion equation, evaluation metrics in the image field of computer vision and those in meteorological science are used, as shown in Table 2. Specifically, the Mean Square Error (MSE) is used as the evaluation metric, and the average mean error between the predicted value and the true value is used to measure the overall pixel value deviation between the predicted image and the true image. The Peak Signal to Noise Ratio (PSNR) analyzes the ratio of signal to noise between images and evaluates the image quality of the predicted image. The Continuous Ranked Probability Score (CRPS) can obtain the cumulative error between the distribution of the predicted image and the distribution of the true image, and is used to evaluate the model's ability to model the uncertainty of precipitation prediction. In addition, in meteorological science, different levels of precipitation are required, and it is of great value to accurately predict different levels of precipitation. Therefore, according to the pixel values of VIL, the index discrimination thresholds are set as [16, 74, 133, 160, 181, 219], and the corresponding true precipitation values are [0.15, 0.78, 3.53, 7.07, 12.14, 32.23], with the unit of kg / m². At the same time, the Critical Success Index (CSI), False Alarm Rate (FAR), Heidke Skill Score (HSS), and Brier Skill Score (BSS) are selected to measure the false alarm rate and accuracy of all methods for precipitation prediction under different threshold conditions.

[0056] Table 2

[0057] Among them, is the total number of samples, represents the true value of the radar echo image, represents the corresponding predicted value; represents the maximum value of all pixel values; represents the probability distribution of the real-time image; is the unit step function; is the number of hits, indicating that the model predicted precipitation and precipitation actually occurred. The specific mathematical representation is prediction = 1, true value = 1; is the number of false precipitation events, indicating that the model predicted precipitation but precipitation did not actually occur. The specific mathematical representation is prediction = 1, true value = 0; is the number of missed precipitation events, and the mathematical representation is prediction = 0, true value = 1; is a true negative, indicating that the prediction is no precipitation and there is actually no precipitation. The mathematical representation is prediction = 0, true value = 0; is the predicted probability on the i-th day, is the actual result on the i-th day, is the baseline Brier score, is the Brier score of the prediction model.

[0058] Compare the precipitation prediction effects with other methods. Use multi-source meteorological data in the previous hour as input to predict precipitation in the next 60 minutes. A total of 6 frames of predicted images are output, and the time interval between each frame of predicted image is 10 minutes. To more fully verify the quality and performance of the forward-inverse dual model (STI-DRNN) based on the spatio-temporal information conversion equation for high-resolution short-term precipitation forecasting, and measure the applicability of the STI-DRNN model to perform short-term forecasting of different levels of precipitation, select currently advanced methods in the spatio-temporal prediction field for quantitative comparison, including ConvLSTM, PredRNN, PhyNet, Rainformer, Earthformer, SimVP, TAU, and PastNet, which are mainly basic spatio-temporal prediction methods based on CNN, RNN, and Transformer. To ensure the accuracy of the experimental results of this study, all methods select the model with the minimum loss function value after training and evaluate and test on the same test set. Table 3 shows the 1-hour forecast results of the forward-inverse dual model based on the spatio-temporal information conversion equation and the other methods on the SEVIR dataset.

[0059] Table 3

[0060] From the results of the quantitative evaluation indicators in the above table, it can be seen that compared with other methods, the forward-inverse dual model based on the spatio-temporal information conversion equation has better performance in short-term forecasting and higher accuracy of the predicted images. In the basic 1-hour precipitation forecast, the ConvLSTM model based on RNN and the SimVP model based on CNN have relatively large MSE scores and relatively small PSNR scores, indicating that the models based on CNN and RNN have weak capabilities in extracting precipitation information. At the same time, the MSE values of the Rainformer model and the Earthformer model are both greater than those of STI-DRNN, and their PSNR values are both less than those of the STI-DRNN model, indicating that the forward-inverse dual model based on the spatio-temporal information conversion equation performs better than the Transformer method. In addition, to further verify whether the forward-inverse dual model based on the spatio-temporal information conversion equation has efficient uncertainty modeling performance, the CRPS index was used to measure the error between the probability distributions of the predicted images and the real images of different methods. It can be seen from the table that compared with the worst model, the STI-DRNN model improved the CRPS by 9.39%, and compared with the sub-optimal PredRNN model, the STI-DRNN model improved the CRPS by 4.06%, indicating that the results of using the STI-DRNN model for precipitation prediction have the probability distribution most similar to the real image, that is, the uncertainty modeling ability of the model is higher. Secondly, the average scores of FAR, CSI, and HSS at a 60-minute lead time for all methods were calculated. Its FAR-M index is the lowest, indicating that the false alarm rate of the model is the lowest, and its CSI-M and HSS-M indexes are the highest, indicating that the overall forecasting performance of STI-DRNN is the best and the prediction is more accurate. Specifically, in short-term forecasting, compared with the best-performing other method, PredRNN, STI-DRNN increased by 10.03% in CSI-M, 11.07% in HSS-M, and 9.71% in FAR-M.

[0061] Since the meteorological field usually focuses on the model performance under different precipitation thresholds and it is necessary to measure the performance of all methods under different levels of precipitation, the CSI scores and HSS scores of each method under different thresholds were measured respectively, as Figure 4 (a) and Figure 4 (b) shown. Taking the Rainformer method as the baseline, the percentage increase in the CSI scores and HSS scores of all the other methods compared with the Rainformer method was measured. As the precipitation threshold gradually increases, the difficulty of reconstructing high-intensity radar echoes for all methods also increases. From Figure 4It can be seen that the improvement of STI-DRNN under all thresholds is a positive improvement compared to Rainformer, and the improvement value at each threshold is at the maximum level. In particular, as the precipitation threshold gradually increases, the difficulty of precipitation forecasting gradually increases. However, the percentage of improvement of the STI-DRNN model shows a positive correlation with the precipitation threshold, indicating that the STI-DRNN model is more applicable to medium-intensity precipitation forecasting.

[0062] From the above experimental analysis, it can be known that STI-DRNN has better overall prediction performance for medium and heavy precipitation levels. Therefore, an index time-effect analysis diagram with precipitation thresholds of 181 and 219 is drawn to further verify the prediction details of STI-DRNN under medium and heavy precipitation levels, such as Figure 5 from (a) to Figure 5 (c) as shown. From Figure 5 from (a) to Figure 5 (c), it can be seen that as the prediction time gradually increases, the uncertainty of the model about the future also gradually increases, resulting in the prediction performance of all methods gradually decreasing over time. However, the overall performance of the STI-DRNN model on all thresholds is better than that of the other methods, and the downward trend of the prediction index is smaller. Relatively, as the prediction time gradually increases, the performance gap between STI-DRNN and other methods also gradually increases. To be exact, as shown in Table 4, compared with the PredRNN, a model with better comprehensive prediction ability among other models, the CSI score has increased by 27.77% at a threshold of 181 and 82.64% at a threshold of 219, and the HSS score has increased by 25.75% at a threshold of 181 and 84.06% at a threshold of 219. This reflects that STI-DRNN has a stronger capture ability for high-intensity precipitation levels, further verifying the superiority of this model.

[0063] Table 4

[0064] Among them, 181 and 219 correspond to precipitation thresholds of 12.14 kg / m 2 and 32.23 kg / m 2 respectively.

[0065] To verify the early warning ability of STI-DRNN for real-world precipitation, randomly selected extreme precipitation events that are rare in daily life were tested. In addition, in order to clearly see whether the prediction results of the model have overestimated and underestimated problems for real images, an error heat map between the sixth frame of the predicted image and the sixth frame of the real image was additionally added to the visual image. The red area represents the underestimation phenomenon, and the blue area represents the overestimation phenomenon.

[0066] FromFigure 6 It can be seen that as the prediction time gradually increases, all methods are gradually losing the detailed information of precipitation prediction. In particular, the PhydNet method loses the most information about precipitation events. At the same time, compared with other models, although other methods can also achieve basic prediction effects in terms of prediction visual effects, they all have more problems of underestimation and overestimation of real precipitation images, and there are greater errors between the specific positions and general shapes of these methods in the precipitation area and the real images, that is, problems such as blurred prediction and inaccurate position in the dark areas of yellow and above. In contrast, although the STI-DRNN still inevitably has problems of underestimation and overestimation, its prediction effect is closer to the real value, and both the problems of overestimation and underestimation are reduced. This shows that the STI-DRNN method efficiently transforms the detailed information of spatio-temporal data, making the spatio-temporal modeling ability of the model better than other methods in the case of moderate to heavy precipitation.

[0067] , as the forecast lead time gradually increases, the accuracy of the STI-DRNN in predicting weather events always remains at the optimal level, and this situation still applies under different precipitation threshold conditions. In addition, the power spectrum curve of the STI-DRNN is almost the same as the real value in the wavelength range from 64KM to 256KM, while other models have deviated from the real value when the wavelength range is 128KM, which indicates that the amount of information predicted by the STI-DRNN in the large-scale spatial range is almost the same as the real value. In the small-scale range, all methods have produced lower PSD scores, which seems to have nothing to do with the prediction lead time, indicating that deep learning methods introduce a smoother and more blurred precipitation field in the small-scale range, but the blurring effect of the STI-DRNN is lower than that of other methods. Generally speaking, the STI-DRNN is more applicable to extreme precipitation events in the large scale.

[0068] The SEVIR dataset contains different types of additional meteorological variables such as infrared brightness temperature and visible light. A total of 4 different meteorological variables were used as the input of the model in the above experiments. To verify the impact of using additional dimensions of meteorological variables on the STI training framework, each type and quantity of the input meteorological variables were systematically eliminated. The experimental results are shown in Table 5. Theoretically speaking from the delay theorem, as the dimensionality of the input data space increases, the number of spatial interaction features that the model can learn will increase, thus expanding the non-linear mapping relationship of the model for future precipitation time information. The gradual decrease in the number of meteorological variables is related to the downward trend of the performance of STI-DRNN in all evaluation metrics. This decrease is particularly obvious in the FAR and CSI metrics, indicating that using more meteorological variables helps the model effectively map spatial information to the time information of future precipitation variables, and also enables the model to capture the complex interactions and dependencies in the precipitation atmospheric process, improving the accuracy of precipitation prediction.

[0069] Table 5

[0070] To clearly verify the optimization ability of the spatio-temporal transformer for the auxiliary training of precipitation forecasting, the spatio-temporal transformer was ablated and the spatio-temporal transformer was added to other methods with better prediction performance to verify the effect of the spatio-temporal transformer. Although current deep learning methods can already achieve good forecasting results, due to the rapid pace of climate change, past meteorological information may not accurately reflect the current atmospheric conditions, and existing precipitation forecasting methods have poor prediction effects on heavy precipitation using short-term time data. Introducing the spatio-temporal conversion equation into the precipitation forecasting method can convert multi-dimensional spatial data into future time information, solve the problem of insufficient short-term data information, and ultimately improve the performance of short-term and nowcasting precipitation forecasting. The experimental results are shown in Table 6.

[0071] Table 6

[0072] Among them, ConvLSTM-STI is ConvLSTM with the spatio-temporal transformer added, PredRNN-STI is PredRNN with the spatio-temporal transformer added, PastNet-STI is PastNet with the spatio-temporal transformer added, and STI-DRNN-NoSTI is STI-DRNN with the spatio-temporal transformer ablated.

[0073] According to the experimental results, after integrating the spatio-temporal transformer into all methods, there is a significant improvement in each quantization index. Especially for the ConvLSTM method, the MSE score and CSI score are improved by 13.05% and 9.20% respectively, and the CRPS score is improved by 3.58% after combining with STI. These results demonstrate the effectiveness and generality of the spatio-temporal transformer in spatio-temporal dynamic modeling. The model integrated with the spatio-temporal transformer can significantly improve the prediction accuracy and robustness, thus generating more accurate precipitation forecast results.

[0074] In addition, to verify the improvement effect of the spatio-temporal transformer on precipitation events of different intensities, scatter plots of the performance improvement of the HSS index and CSI index of several methods are drawn under all threshold conditions, as Figure 7 shown in Figure 7 Fig. (a) and Figure 7 Fig. Figure 7 It can be seen from (b) that after integrating the spatio-temporal transformer, all methods have a positive improvement in CSI and HSS. Especially in the case of moderate and heavy precipitation, the improvement effect is the most significant. This proves the effectiveness of the STI dual training structure, which can make up for the lack of short-term data information and improve the overall prediction performance of the model.

[0075] Since the climate changes rapidly in the real world, common regression loss functions often ignore sudden heavy rain events, resulting in more underestimation of weather events with medium to strong precipitation levels. Therefore, an adaptive weight gradient loss function (ADGLoss) is proposed, which can effectively optimize the sensitivity of the model to the occurrence of heavy precipitation events and improve the prediction accuracy. To verify the effectiveness of ADGLoss, L1Loss, L2Loss, and ADGLoss without weights are used as loss functions to train the model to evaluate its effect. The experimental results are shown in Table 7 and Figure 8 Figs. (a) to Figure 8 (c).

[0076] Table 7

[0077] It can be seen from the quantitative evaluation results of the above ablation loss function that the adaptive weighted gradient loss function does not achieve the optimal result in terms of MSE loss. This is because the precipitation data has the characteristic of long-tailed distribution, and the optimal solution of L1Loss may show a more robust result in the squared error space of MSE. Although the adaptive weighted gradient loss function fails to reach the optimal in terms of the MSE metric, the adaptive weighted gradient loss function has more advantages in the meteorological precipitation metrics, especially the FAR score, which is improved by 6.41% compared to the sub-optimal loss function. In addition, it can also be seen from the index curve graphs of different time periods that ADGLoss is in the leading level in the case of high-threshold precipitation, and with the increase of the prediction time, the performance degradation of ADGLoss is the least. This indicates that the adaptive weighted gradient loss function can better alleviate the problem of the error aggravation caused by the increase of the prediction time of the model and improve the prediction accuracy of moderate and heavy precipitation.

[0078] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other deformations may also fall within the scope of this specification. Therefore, as an example rather than a limitation, the alternative configurations of the embodiments of this specification can be regarded as consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments clearly introduced and described in this specification.

Claims

1. A method for predicting precipitation of multi-source meteorological elements based on a spatio-temporal information conversion equation, characterized in that, Including: Establishing a forward-inverse dual model based on a spatio-temporal information conversion equation; Obtaining multiple groups of training samples, where the training samples include multi-source meteorological element data, and the label of the training samples is rainfall data; Training the forward-inverse dual model based on the spatio-temporal 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 spatio-temporal information conversion equation; Obtaining multi-source meteorological element data to be predicted; Generating a precipitation prediction image through the multi-source meteorological element precipitation prediction model based on the multi-source meteorological element data to be predicted.

2. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 1, wherein The forward-inverse dual model based on the spatio-temporal information conversion equation includes a data preprocessing module, a spatio-temporal converter, and a time-space converter; The data preprocessing module is used to preprocess the multi-source meteorological element data and rainfall data to generate a delay embedding matrix; The spatio-temporal converter is used to generate a precipitation prediction image based on the delay embedding matrix; The time-space converter is used to extract the time features of the precipitation prediction image and reconstruct the spatial distribution of multi-dimensional meteorological data according to the time features of the precipitation prediction image.

3. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 2, characterized in that, The data preprocessing module preprocesses the multi-source meteorological element data and rainfall data to generate a delay embedding matrix, including: Performing spatio-temporal alignment on the multi-source meteorological element data and rainfall data to generate spatio-temporally aligned multi-source meteorological element data and rainfall data; Normalizing the spatio-temporally aligned multi-source meteorological element data and rainfall data to generate normalized multi-source meteorological element data and rainfall data; Performing a delay embedding transformation on the normalized multi-source meteorological element data to generate a delay embedding matrix.

4. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 2, characterized in that, The spatio-temporal converter includes a first encoder and a first decoder, where the first encoder is used to perform convolutional 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 according to the spatial feature map.

5. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 4, characterized in that, The first encoder includes multiple convolutional layers and an encoding ConvLSTM layer, where 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.

6. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 4, wherein The first decoder includes multiple deconvolutional layers and a decoding ConvLSTM layer, where the multiple deconvolutional layers are used to generate the precipitation distribution at future times 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 times.

7. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to any one of claims 2-6, characterized in that, The time-space converter includes a second encoder and a second decoder, where the second encoder is used to extract the time features of the precipitation prediction image, and the second decoder is used to reconstruct the spatial distribution of multi-dimensional meteorological data according to the time features of the precipitation prediction image.

8. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to any one of claims 1-6, characterized in that The adaptive weighted gradient loss function includes a mean absolute error function and a gradient difference loss function based on an adaptive weight.

9. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 8, wherein The adaptive weighted gradient loss function is: ； ； ； ； Among them, is the adaptive weighted gradient loss function, is the mean absolute error function, is the gradient difference loss function based on the adaptive weight, is the pixel value of the pixel at the i-th row and j-th column in the precipitation prediction image output by the forward-inverse dual model based on the spatio-temporal 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 true precipitation image in the horizontal direction, is the gradient change rate of the true precipitation image in the vertical direction, is the pixel value of the pixel at the (i - 1)-th row and j-th column in the precipitation prediction image, is the pixel value of the pixel at the i-th row and (j - 1)-th column in the precipitation prediction image, is the adaptive weight.

10. The method for predicting precipitation of multi-source meteorological elements based on the spatio-temporal information conversion equation according to claim 9, characterized in that, The calculation formula for the adaptive weight is: ； Among them, is the precipitation prediction image of the i-th training sample, is the true precipitation image of the i-th training sample, is the number of training samples.

Images