A method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion

Abstract

This invention discloses a method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion. First, multi-source data of the study area are acquired, and the regional extreme precipitation is calculated. An extreme gradient boosting model is used to select the key atmospheric circulation indices that have the most significant impact on extreme precipitation in the target area. The multi-source data undergoes unified preprocessing. Then, a spatiotemporal prediction model fusing convolutional neural networks and long short-term memory networks is constructed, and a deep learning prediction model is generated through joint training of multiple features. Finally, the preprocessed multi-source data and the deep learning model are used to perform spatiotemporal prediction of extreme precipitation. This invention integrates extreme precipitation data with atmospheric circulation indices and topographic DEM data into a unified input framework, enabling the model to simultaneously utilize spatiotemporal precipitation distribution characteristics, atmospheric circulation patterns, and topographic information to jointly predict precipitation, resulting in high prediction and simulation accuracy.

Description

Technical Field

[0001] This invention belongs to the field of hydrometeorological technology, and in particular relates to a method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion. Background Technology

[0002] Extreme precipitation, as a type of extreme climate event, often leads to disasters such as floods, mudslides, landslides, and urban flooding, causing significant harm to the natural environment and human society. Atmospheric circulation is one of the important factors influencing the occurrence of extreme precipitation events. Atmospheric circulation directly or indirectly affects the occurrence of extreme precipitation by influencing water vapor transport and the evolution of local climate dynamics. For example, during the South Asian monsoon and East Asian monsoon, large amounts of water vapor are transported from the ocean to the land, easily triggering torrential rains and floods. In addition to the dynamic characteristics of atmospheric circulation, regional topographic features (such as altitude, elevation gradient, and slope aspect) also have a significant impact on the distribution, intensity, and frequency of extreme precipitation events. Changes in altitude, differences in slope and aspect, and the interaction between complex topography and the atmosphere determine the uneven spatial distribution of precipitation.

[0003] Traditional methods for predicting extreme precipitation mainly include statistical analysis methods (such as regression analysis, time series models, and extreme value theory) and physical process-based modeling methods (such as the Global Climate Model (GCM) and the Regional Climate Model (RCM). These methods have significant limitations in comprehensively considering the combined effects of atmospheric circulation and topographic features. In the formation of extreme precipitation, global-scale atmospheric circulation and local-scale topographic forcing are closely related. Traditional methods struggle to fully capture this multi-scale, multi-factor nonlinear mechanism and achieve high accuracy simultaneously in both time and space. Statistical methods are biased towards time series modeling and struggle to capture spatial distribution characteristics; while physical models can describe spatial patterns, their temporal resolution is insufficient, and their handling of local features is not refined enough. Summary of the Invention

[0004] The purpose of this invention is to provide a method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion, which solves the technical problem that existing technologies only focus on a single feature type, ignore the synergistic effect of topographic features and key meteorological factors on extreme precipitation, and fail to fully explore the effective information in the spatial and temporal dimensions.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following solution:

[0006] The method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion includes the following steps:

[0007] Step S1: Obtain multi-source data for the target area, including historical precipitation raster data, atmospheric circulation index time series data, and topographic DEM data.

[0008] Step S2: Calculate the extreme precipitation index Rx5day for the target area and save it as a raster file with the same coordinate system as the source file. The calculation formula is as follows:

[0009] (1)

[0010] (2)

[0011] In the above formula, S i It is the total rainfall over 5 consecutive days starting from day n. P i For the first i Daily rainfall Rx5day This represents the maximum rainfall over five consecutive days each month.

[0012] Step S3: Use the Extreme Gradient Boosting (XGBoost) algorithm to screen the key atmospheric circulation indices that have the most significant impact on extreme precipitation in the target area.

[0013] Step S4: Resample the terrain DEM data to align it with the grid of the precipitation raster data, while ensuring that the terrain DEM data and the precipitation raster data have the same coordinate system.

[0014] Step S5: Perform Min-Max normalization on the above multi-source data, and construct a spatiotemporal prediction model that integrates a convolutional neural network (CNN) and a long short-term memory network (LSTM). Use the multi-source data to make spatiotemporal predictions of extreme precipitation. The process can be represented as follows:

[0015] (3)

[0016] In the above formula, P extreme These are simulated values ​​for extreme precipitation. X P , X Cn , X T The inputs are historical extreme precipitation data, key atmospheric circulation index data, and topographic DEM data, respectively. f This is a spatiotemporal prediction model driven by multi-source data, which fuses convolutional neural networks (CNN) and long short-term memory networks (LSTM).

[0017] Step S6: Select the first 80% of the time series of extreme precipitation data and atmospheric circulation index data as the training set, construct and train the model, use the mean square error (MSE) as the optimization objective, calculate the square error between the predicted value and the actual value, and through multiple rounds of training, make full use of the deep learning model driven by multi-source data to complete the simulation and prediction of the complex spatiotemporal pattern of extreme precipitation.

[0018] Step S7: Calculate the correlation coefficient R for each pixel prediction value to evaluate the accuracy of the simulation results.

[0019] Deep learning, as a core technology in the field of artificial intelligence, has demonstrated powerful data-driven capabilities and nonlinear modeling advantages in the field of hydrology and meteorology, enabling it to efficiently capture complex spatiotemporal features and multi-factor interactions. This invention, based on Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), overcomes the shortcomings of traditional methods in describing multi-scale, multi-dimensional interactions in spatial feature extraction and time series modeling. Furthermore, deep learning can handle large-scale, multi-source, heterogeneous data, thereby improving prediction accuracy and generalization ability.

[0020] Further optimization, step S3 specifically includes:

[0021] Step S3.1: Using ArcGIS pixel statistics methods, calculate the average time series of extreme precipitation in the target area. The calculation formula is as follows:

[0022] (4)

[0023] In the above formula, For the target area in time period T The average extreme precipitation value within the region, N The total number of pixels, d T,j For raster files in time Section T medium pixel j The value of .

[0024] Step S3.2: Align the time series of atmospheric circulation indices and extreme precipitation in time, merge them into a single set of data as input data, and input this data into the Extreme Gradient Boosting (XGBoost) algorithm. Use the first 80% of the time series as the training set and the remainder as the test set. XGBoost learns the impact of each indice on extreme precipitation. Decision trees are the basic building blocks of XGBoost. During the construction of each tree, XGBoost selects the optimal split point based on different feature values ​​to maximize the model's predictive ability for the target variable. In this step, the algorithm calculates the importance of each indice based on its contribution to the prediction of extreme precipitation time series at each split node. Finally, it sorts the indices by importance and selects the four that contribute the most as key atmospheric circulation indices. The importance can be calculated using the following formula:

[0025] (5)

[0026] In the above formula, I k For index k The importance of Tr The set of all decision trees in the model. St For trees t The set of all split nodes in the set. f tr,s Number tr At the split node s The selected features; α ( ) is an indicator function, when f tr,s = k hour, α ( f tr,s = k =1, otherwise 0.

[0027] Further optimization, step S5 specifically includes:

[0028] Step S5.1: Copy the atmospheric circulation index data in the spatial dimension and align it with the pixels. Merge the extreme precipitation data, atmospheric circulation index data, and topographic DEM data in the channel dimension to form a multi-channel input feature tensor. Use a one-dimensional convolutional neural network (CNN) to extract the local spatial feature map of each pixel in the spatial dimension, and fuse the multi-channel information of extreme precipitation, atmospheric circulation index, and topographic DEM data to generate a discriminative spatial feature map. The convolution process can be represented by the following formula:

[0029] (6)

[0030] In the above formula, FThis is the output feature map of the convolutional layer. X For the input feature map, W For convolution kernel weights, b G is the bias, and G is the activation function.

[0031] Step S5.2: Input the spatial feature sequence output by the Convolutional Neural Network (CNN) into the Long Short-Term Memory (LSTM) neural network for time series modeling. The LSTM selectively memorizes important information and forgets irrelevant information through a gating mechanism, thereby capturing dynamic trends in past time and outputting the hidden state vector at the last moment. This vector condenses the comprehensive information of the entire time series and serves as the basis for subsequent predictions. This process is represented by the following formula:

[0032] (7)

[0033] In the above formula, h t The hidden state at the current time step. x t The input spatial feature map for the current time step. h t-1 This is the hidden state of the previous time step.

[0034] Step S5.3: The hidden state vector output by the Long Short-Term Memory (LSTM) neural network is mapped back to spatial features through a fully connected layer and associated with each pixel location. The mapping of the fully connected layer can be represented as follows:

[0035] (8)

[0036] In the formula, z The feature vector output by the fully connected layer. h T This represents the final hidden state of the LSTM. W fc and b fc These are the weights and biases for the fully connected layer.

[0037] Step S5.4: The mapped features are further transformed into extreme precipitation Rx5day prediction values ​​consistent with the pixel distribution through the final convolutional layer, thereby generating spatial distribution predictions for future time points. This process is represented by the following formula:

[0038] (9)

[0039] In the formula, Y Extreme precipitation forecast for each pixel; z The multidimensional feature vector obtained in step 5.3; Conv This is a one-dimensional convolution process;b Conv For bias.

[0040] Further optimization is achieved by calculating the correlation coefficient R in step S7 using the following formula:

[0041] (10)

[0042] In the formula, R The correlation coefficient for each pixel; M The number of samples; X m and Y m For each cell m The true and predicted values ​​of each sample; and The mean of the true value and the predicted value for each cell.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. This invention constructs a spatiotemporal prediction method for extreme precipitation driven by deep learning based on multi-source data. This method overcomes the limitations of traditional methods in utilizing topographic features and key atmospheric circulation factors, and further mines the effective information in multi-source data. It fully utilizes the synergistic effect of topographic features and key meteorological factors on extreme precipitation, providing an innovative solution for the prediction of extreme precipitation.

[0045] 2. Compared with traditional methods, the extreme precipitation spatiotemporal prediction provided by this invention integrates extreme precipitation data with atmospheric circulation indices and topographic DEM data into a unified input framework. This allows the model to simultaneously utilize spatiotemporal precipitation distribution characteristics, atmospheric circulation patterns, and topographic information to jointly predict precipitation, resulting in high prediction and simulation accuracy. This addresses the problem of existing technologies focusing only on a single feature type.

[0046] 3. The method for simulating extreme precipitation based on multi-source data fusion and using a CNN-LSTM model, as described in this invention, provides an innovative solution for the spatiotemporal prediction of extreme precipitation. It can be widely applied in the fields of water resource management and disaster early warning, providing strong support for scientific decision-making. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method for simulating extreme precipitation based on multi-source data fusion and using a CNN-LSTM model, as described in this invention.

[0048] Figure 2 This is a schematic diagram of the spatiotemporal prediction model that integrates convolutional neural networks (CNN) and long short-term memory networks (LSTM) as described in this invention.

[0049] Figure 3 A spatial distribution map of the correlation of the prediction results for the Yarlung Tsangpo River Basin from 2011 to 2022 was obtained through simulation.

[0050] Figure 4 This is a comparison chart of the predicted and actual values ​​of a certain pixel in the simulation results of the Yarlung Tsangpo River Basin.

[0051] Figure 5 This is a spatial distribution map showing the correlation of the prediction results for 2011-2022 obtained using only a single feature (atmospheric circulation index). Detailed Implementation

[0052] The technical solution of the present invention will be described in detail below with reference to the embodiments, but the scope of protection of the present invention is not limited to the embodiments described.

[0053] The Yarlung Tsangpo River basin is located in the southern part of the Qinghai-Tibet Plateau and the northern foothills of the Himalayas. Within my country, its average elevation exceeds 4,000 meters, making it one of the world's highest international rivers. It is characterized by high altitude, heavy rainfall, high heterogeneity, and a steep gradient. The Yarlung Tsangpo River's drainage area within my country is approximately 240,000 km². 2 The main stream of the river is over 2,000 km long and has an annual runoff of over 160 billion cubic meters. 3 The river has a drop of 2725m in its lower reaches. Under the influence of global warming, precipitation in the Yarlung Tsangpo River Basin has shown an overall upward trend at all scales over the past 40 years. Therefore, the spatiotemporal prediction of extreme precipitation in this basin has a profound impact on disaster prevention and mitigation, comprehensive basin management, and sustainable development.

[0054] In this embodiment, as Figure 1 As shown, the method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion specifically includes the following steps:

[0055] Step 1: Obtain historical precipitation raster data, topographic DEM data, and atmospheric circulation index data for the Yarlung Tsangpo River Basin.

[0056] In this embodiment, precipitation raster data spanning from January 1, 1961 to December 31, 2022, with a spatial resolution of 0.1°, totaling 2402 pixels, was used. The topographic DEM data consisted of SRTM elevation data with a spatial resolution of 90m. The atmospheric circulation index data comprised 88 different monthly time series spanning from January 1961 to December 2022.

[0057] Step 2: Calculate the Yarlung Tsangpo River Basin extreme precipitation index Rx5day. Based on the daily-scale historical precipitation raster data obtained in Step 1, taking the extreme precipitation index Rx5day (maximum consecutive 5-day precipitation, with subsequent extreme precipitation representing Rx5day) as an example, the monthly-scale extreme precipitation raster data of the Yarlung Tsangpo River Basin is calculated. This data maintains the same coordinate system as the source file. The formula for calculating Rx5day is as follows:

[0058] (1)

[0059] (2)

[0060] In the above formula, S i It is the total rainfall over 5 consecutive days starting from day n. P i This refers to daily rainfall. Rx5day This represents the maximum rainfall over five consecutive days each month. All extreme rainfall data mentioned in subsequent steps refer to the data calculated in this step.

[0061] Step 3: Using the Extreme Gradient Boosting (XGBoost) algorithm, the four key atmospheric circulation indices that have the most significant impact on extreme precipitation in the Yarlung Tsangpo River Basin were selected. The specific steps are as follows:

[0062] Step 3.1: Using ArcGIS pixel statistics, calculate the average time series of extreme precipitation in the watershed. Partial data obtained are shown in Table 1. The calculation formula is as follows:

[0063] (3)

[0064] In the above formula, For the Yarlung Tsangpo River Basin in a certain time period T The average extreme precipitation value within the region, N The total number of pixels, d T,j For raster files in time periods T medium pixel j The value of .

[0065] In this embodiment, the time period T is each month from 1961 to 2022, and the average extreme precipitation in the basin during this period is shown in Table 1.

[0066] Table 1. Average extreme precipitation in the basin

[0067]

[0068] Step 3.2: Align the atmospheric circulation index time series and the average extreme precipitation time series calculated in Step 3.1 temporally, merge them into a single set of data as input, and feed them into the Extreme Gradient Boosting (XGBoost) algorithm. Use the first 80% of the time series as the training set and the remainder as the test set. XGBoost learns the impact of each index on extreme precipitation. Decision trees are the basic building blocks of XGBoost. During the construction of each tree, XGBoost selects the optimal split point based on different feature values ​​to maximize the model's predictive ability for the target variable. In this step, XGBoost calculates the importance of each index based on its contribution to the prediction of the extreme precipitation time series at each split node. Finally, it sorts the indices by importance and selects the four contributing indices as key atmospheric circulation indices. The importance can be calculated using the following formula:

[0069] (4)

[0070] In the above formula, I k For index k The importance of Tr The set of all decision trees in the model. St For trees t The set of all split nodes in the set. f tr,s Number tr At the split node s The selected features; α ( ) is an indicator function, when f tr,s = k hour, α ( f tr,s = k =1, otherwise 0.

[0071] In this embodiment, the key atmospheric circulation indices selected are: East Asian trough intensity index, Northern Hemisphere polar vortex center intensity index, North Atlantic subtropical high area index, and Northern Hemisphere polar vortex intensity index. Some data for each index are shown in Table 2.

[0072] Table 2 Key Atmospheric Circulation Indices

[0073]

[0074] Step 4: Using the watershed precipitation raster data as the target template, resample the topographic DEM data to align it with the grid of the precipitation raster data, while ensuring that the topographic DEM data and the precipitation raster data have the same coordinate system.

[0075] Step 5: Perform Min-Max normalization on the multi-source data, including the extreme precipitation raster data obtained in Step 2, the key atmospheric circulation index obtained in Step 3, and the topographic DEM data obtained in Step 4, and input them into the constructed spatiotemporal prediction model that fuses a convolutional neural network (CNN) and a long short-term memory network (LSTM). Figure 2 As shown. The process is represented as follows:

[0076] (5)

[0077] In the above formula, P extreme These are simulated values ​​for extreme precipitation. X P , X Cn , X T These are, respectively, normalized historical extreme precipitation data, data for n key atmospheric circulation indices, and topographic DEM data. f This is a spatiotemporal prediction model driven by multi-source data, which fuses convolutional neural networks (CNN) and long short-term memory networks (LSTM).

[0078] The detailed steps are as follows:

[0079] Step 5.1: In deep learning, data is typically input into the model as a multidimensional array (tensor). The feature tensor in the model can have multiple channels (dimensions), each representing different features or information. In this model, firstly, atmospheric circulation index data is copied spatially to align with the pixels; topographic DEM data is copied temporally to align with each time step (monthly scale). Secondly, extreme precipitation data, atmospheric circulation index data, and topographic DEM data are superimposed as different channels to form a multi-channel input feature tensor. Subsequently, a one-dimensional convolutional neural network (CNN) is used to extract the local spatial feature map of each pixel in the spatial dimension, and the multi-channel information of precipitation, atmospheric circulation index, and topography is fused to generate a discriminative spatial feature map. The convolution process is represented by the following formula:

[0080] (6)

[0081] In the above formula, F This is the output feature map of the convolutional layer; X In this model, the input feature map is used as the basis for input feature map. X It consists of multi-channel information composed of precipitation data, atmospheric circulation index data, and topographic DEM data; W The convolutional kernel weights are randomly initialized before training and automatically adjusted based on the results after each training session. bThe bias is initially set to 0, and will be automatically adjusted based on the results after each training session; G is the activation function, and the ReLU function is selected in this embodiment.

[0082] Step 5.2: Input the spatial feature sequence output by the Convolutional Neural Network (CNN) into the Long Short-Term Memory (LSTM) neural network for time series modeling. LSTM selectively memorizes important information and forgets irrelevant information through a gating mechanism, thereby capturing dynamic trends in past time and outputting the hidden state vector at the last moment. This vector encapsulates the comprehensive information of the entire time series and serves as the basis for subsequent predictions. This process is represented by the following formula:

[0083] (7)

[0084] In the formula, h t The hidden state at the current time step. x t For the input features at the current time step, when t =0 x 0 The result calculated by formula (6) F , h t-1 This is the hidden state of the previous time step.

[0085] Step 5.3: The hidden state vector output by the Long Short-Term Memory (LSTM) neural network is mapped back to spatial features through a fully connected layer and associated with each pixel location. The mapping of the fully connected layer can be represented as follows:

[0086] (8)

[0087] In the formula, z The feature vector output by the fully connected layer. h T This represents the final hidden state of the LSTM. W fc For the weight matrix of the fully connected layer, this case uses the Kaiming Uniform method for default initialization. b fc This is the sum and bias vector of the fully connected layer. In this case, it is initialized to 0 and will be automatically adjusted based on the results after each training session.

[0088] Step 5.4: Through the final convolutional layer, the information from multiple feature channels of the multidimensional feature vector is integrated into one output channel to generate the extreme precipitation Rx5day prediction value for each pixel, thereby generating the spatial distribution prediction for future time points. This process can be represented by the following formula:

[0089] (9)

[0090] In the above formula, Y Extreme precipitation forecast for each pixel; z The multidimensional feature vector obtained in step 5.3; Conv This is a one-dimensional convolution process; b Conv This is the bias, initially set to 0 by default, and will be automatically adjusted based on the results after each training session.

[0091] Step 6: Select the first 80% of the time series of extreme precipitation data and atmospheric circulation index data from Step 5.1 as the training set, construct and train the model, and use the mean square error (MSE) as the optimization objective to reflect the square error between the predicted value and the true value.

[0092] In this embodiment, a time series from 1961 to 2010 was selected, the model was set to undergo 100 training rounds, and an early stopping mechanism was embedded. Training automatically terminates when the MSE (Mean Segregated Estimate) does not decrease within 10 consecutive training rounds. Through these steps, a deep learning model driven by multi-source data is fully utilized to simulate and predict complex spatiotemporal patterns of extreme precipitation.

[0093] Step 7: Calculate the correlation coefficient R of the predicted values ​​for each pixel during the validation period (i.e., 2011-2022) to assess the accuracy of the simulation results. The correlation coefficient can be calculated using formula (10):

[0094] (10)

[0095] In the formula, R The correlation coefficient for each pixel; M This is the sample size, which is 120 in this case. X m and Y m For each cell m The true and predicted values ​​of each sample; and The mean of the true value and the predicted value for each cell.

[0096] The model's spatiotemporal prediction accuracy in the Yarlung Tsangpo River basin is as follows: Figure 3 As shown, the average correlation coefficient of all pixels is 0.82, indicating good overall model accuracy. Pixels in row 20 and column 115 were selected. The actual and predicted values ​​for each time step during the validation period were read, and the actual and predicted values ​​were compared during the prediction period, with Pearson correlation calculated. Figure 4 As shown, the time-series correlation coefficient reaches 0.84. Furthermore, the spatial distribution of correlation coefficients obtained by models using only atmospheric circulation indices for prediction, under the same model structure and parameters, is as follows: Figure 5The average correlation coefficient of all pixels was 0.75, which is less accurate than the scheme that uses both atmospheric circulation index and terrain features.

[0097] This demonstrates that the prediction model based on the fusion of atmospheric circulation index and topographic DEM data can accurately capture extreme precipitation events and has a good predictive ability for extreme precipitation amounts.

[0098] The above description is merely an example of the implementation of this invention in a specific watershed, but the scope of protection of this invention is not limited thereto. If those skilled in the art are inspired by this description and design methods and embodiments similar to this technical solution without departing from the spirit of this invention, such methods and embodiments should fall within the scope of protection of this invention.

Claims

1. A method for simulating extreme precipitation using a CNN-LSTM model based on multi-source data fusion, characterized in that, Includes the following steps: Step S1: Acquire multi-source data for the target area, including historical precipitation raster data, atmospheric circulation index time series data, and topographic DEM data; Step S2: Calculate the extreme precipitation index Rx5day for the target area and save it as a raster file with the same coordinate system as the source file. The calculation formula is as follows: S i =P i +P i+1 +P i+2 +P i+3 +P i+4 (1) Rx5day=max(S i ) (2) In the above formula, S i P is the total rainfall over 5 consecutive days starting from day i. i Let Rx5day be the daily precipitation on day i, and Rx5day be the maximum precipitation over 5 consecutive days in a month. Step S3: Using the Extreme Gradient Boosting (XGBoost) algorithm, select the key atmospheric circulation indices that have the most significant impact on extreme precipitation in the target area, specifically including: Step S3.1: Using ArcGIS pixel statistics methods, calculate the average time series of extreme precipitation in the target area. The calculation formula is as follows: In the above formula, Let N be the average extreme precipitation value of the target area within time period T, and N be the total number of pixels. T，j The value of cell j in the raster file within the time period T; Step S3.2: Merge the time series of atmospheric circulation indices and extreme precipitation, using the first 80% of the time series as the training set and the remainder as the test set. Use the Extreme Gradient Boosting (XGBoost) algorithm to learn the impact of each index on extreme precipitation. This algorithm calculates the importance of each index based on its contribution to the prediction results at the split nodes. Finally, the four indices with the largest contributions are selected as the key atmospheric circulation indices, ranked by importance. The importance is calculated using the following formula: In the above formula, I k Let be the importance of the exponent k, Tr be the set of all decision trees in the extreme gradient boosting algorithm XGBoost, St be the set of all split nodes in tree t, and f be the importance of the exponent k. tr,s The characteristic chosen by the number tr at the split node s; α() is the indicator function, when f tr,s When =k, α(f tr,s =k) ​​= 1, otherwise 0; Step S4: Resample the topographic DEM data to align it with the grid of the precipitation raster data, while ensuring that the topographic DEM data and the precipitation raster data have the same coordinate system; Step S5: Perform Min-Max normalization on the above multi-source data, and construct a spatiotemporal prediction model that integrates a convolutional neural network (CNN) and a long short-term memory network (LSTM). Use the multi-source data to make spatiotemporal predictions of extreme precipitation. The process can be represented as follows: P extreme =f[X P ,X Cn ,X T ] (3) In the above formula, P extreme X represents the simulated value of extreme precipitation. P X Cn X T The inputs are historical extreme precipitation data, key atmospheric circulation index data, and topographic DEM data, respectively. f is a spatiotemporal prediction model driven by multi-source data, which integrates a convolutional neural network (CNN) and a long short-term memory network (LSTM). Building a CNN-LSTM model specifically includes: Step S5.1: Copy the atmospheric circulation index data in the spatial dimension to align it with the pixels. Merge the extreme precipitation data, atmospheric circulation index data, and topographic DEM data in the channel dimension to form a multi-channel input feature tensor. Use a one-dimensional convolutional neural network (CNN) to extract the local spatial feature map of each pixel in the spatial dimension, and fuse the multi-channel information of extreme precipitation, atmospheric circulation index, and topographic DEM data to generate a discriminative spatial feature map. The convolution process is represented by the following formula: F = G(W*X + b) (6) In the formula, F is the output feature map of the convolutional layer, X is the input feature map, W is the weight of the convolutional kernel, b is the bias, and G is the activation function; Step S5.2: Input the spatial feature sequence output by the Convolutional Neural Network (CNN) into the Long Short-Term Memory (LSTM) neural network for time series modeling. The LSTM selectively memorizes important information and forgets irrelevant information through a gating mechanism, thereby capturing the dynamic trends in past time and outputting the hidden state vector at the last moment. This vector condenses the comprehensive information of the entire time series and serves as the basis for subsequent predictions. This process is represented by the following formula: h t =LSTM(x t ,h t-1 ) (7) In the above formula, h t The hidden state at the current time step, x t h is the input spatial feature map for the current time step. t-1 This is the hidden state from the previous time step; Step S5.3: The hidden state vector output by the Long Short-Term Memory (LSTM) neural network is mapped back to spatial features through a fully connected layer and associated with each pixel location. The mapping of the fully connected layer is expressed by the following formula: z=W fc ·h T +b fc (8) In the above formula, z is the feature vector output by the fully connected layer, and h T For the final hidden state of the LSTM, W fc and b fc The weights and biases of the fully connected layer; Step S5.4: The mapped features are further transformed into extreme precipitation Rx5day prediction values ​​consistent with the pixel distribution through the final convolutional layer, thereby generating spatial distribution predictions for future time points. This process is represented by the following formula: Y=Conv(z)+b Conv (9) In the formula, Y is the predicted extreme precipitation value for each pixel; z is the multidimensional feature vector obtained in step 5.3; Conv is the one-dimensional convolution process; b Conv For bias; Step S6: Select the first 80% of the time series of extreme precipitation data and atmospheric circulation index data as the training set, train the model, use the mean square error (MSE) as the optimization objective, calculate the square error between the predicted value and the actual value, and through multiple rounds of training, make full use of the deep learning model driven by multi-source data to complete the simulation and prediction of the complex spatiotemporal pattern of extreme precipitation. Step S7: Calculate the correlation coefficient R for each pixel prediction value to evaluate the accuracy of the simulation results.

2. The method for simulating extreme precipitation based on multi-source data fusion and using a CNN-LSTM model according to claim 1, characterized in that, In step S7, the correlation coefficient R is calculated using the following formula: In the formula, R is the correlation coefficient of each pixel; M is the sample size; X m and Y m For each pixel, the true value and predicted value of the m-th sample; and The mean of the true value and the predicted value for each cell.

Images