A hydrological forecasting method for areas lacking data based on multi-source precipitation fusion

Abstract

A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion is disclosed, belonging to the field of hydrological forecasting. First, hydrological data is collected and preprocessed. Second, dimensionless processing is performed. Then, a watershed CNN-TCN hydrological model is constructed, and spatial processing is performed. Remote sensing precipitation and rain gauge station inverse distance weighted interpolated grid precipitation are input into the model. A convolutional neural network is used to fuse the spatial features of multi-source precipitation to obtain accurate precipitation data. The output fused precipitation features, along with potential evapotranspiration and temperature, are used in a temporal convolutional neural network for hydrological forecasting. Finally, the model is trained and validated. This invention uses remote sensing grid precipitation and rain gauge station interpolated grid precipitation as model inputs, and performs precipitation fusion within the model. This enables accurate characterization of the precipitation-runoff nonlinear response, significantly improving the accuracy of hydrological forecasting and providing efficient and reliable technical support for flood control, disaster reduction, and water resource management in data-scarce watersheds.

Description

Technical Field

[0001] This invention belongs to the field of hydrological forecasting and relates to a hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion. Background Technology

[0002] Hydrological forecasting, a crucial measure in water resource management, provides vital decision-making support for watershed flood and drought disaster prevention, safe operation of reservoir systems, and scientific allocation of water resources. Precipitation, as a primary driving factor, directly impacts the accuracy and lead time of hydrological forecasts. Commonly used precipitation data includes rain gauge data, which provides accurate point precipitation information and is therefore the most widely applied. However, in data-scarce areas, the density of rain gauge stations is insufficient, resulting in limited spatial representativeness and difficulty in effectively capturing the spatiotemporal variability of precipitation, thus affecting the reliability of hydrological forecasts. Therefore, it is necessary to study how to fully exploit spatial information about precipitation to obtain more accurate precipitation data and improve the accuracy of hydrological forecasts in data-scarce areas.

[0003] Existing methods primarily acquire continuously distributed precipitation data through interpolation techniques or satellite remote sensing to directly drive hydrological models for hydrological forecasting in data-scarce areas. For example, Chinese invention patent (application number 202110154718.6) proposes a comprehensive interpolation method for daily precipitation data. This method uses external drift kriging and thin-plate spline interpolation methods to interpolate precipitation data from rain gauge stations daily, and obtains high-precision precipitation gridded data through weighted averaging and error correction. Chinese invention patent (application number 202410391992.9) discloses a method for correcting the bias of satellite precipitation data in data-scarce areas. It utilizes a hybrid geographic weighted regression coefficient library established in data-rich areas, selects similar rain gauge stations using the K-nearest neighbor method, and combines kriging interpolation to estimate the precipitation bias in data-scarce areas, obtaining satellite precipitation data with improved accuracy. However, interpolation accuracy heavily relies on the spatial representativeness of limited rainfall stations. In data-scarce areas, interpolation struggles to capture the true spatial characteristics of precipitation, leading to significant uncertainty in precipitation input. On the other hand, while satellite remote sensing can provide large-scale, continuous precipitation distribution, it exhibits systematic biases across different climatic zones and topographic conditions. When directly driving hydrological models, this can easily trigger cumulative errors in runoff simulation. These shortcomings limit further improvements in the accuracy of hydrological forecasts.

[0004] Therefore, this invention proposes a hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion. It leverages the advantages of integrating precipitation data from rain gauge stations and satellite remote sensing using a CNN-TCN model to obtain more accurate precipitation data and improve the accuracy and reliability of hydrological forecasting in data-scarce areas. Summary of the Invention

[0005] To address the above problems, this invention proposes a hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion. This invention constructs a CNN-TCN model, using satellite remote sensing precipitation and interpolated grid precipitation from rain gauge stations as model inputs. Precipitation fusion is performed within the model, fully integrating the advantages of multi-source precipitation to obtain more accurate precipitation data. This data is then used in conjunction with temporal hydrometeorological elements such as potential evapotranspiration and temperature for runoff forecasting. The improved runoff forecasting accuracy reflects the model's full integration of the advantages of multi-source precipitation, thereby enhancing the accuracy and reliability of hydrological forecasting in data-scarce areas.

[0006] To achieve the above objectives, the technical solution provided by the present invention is as follows: A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion includes the following steps: Step 1: Collect precipitation data from all available rain gauge stations, publicly available satellite remote sensing precipitation data, potential evapotranspiration data, temperature data, and measured runoff data within the study basin. Preprocess some of the collected data to obtain raw data, including IDW interpolated grid precipitation data from rain gauge stations, satellite remote sensing precipitation data, average potential evapotranspiration data, and average temperature data, which are then divided into training and validation sets. Specifically: Step 1.1: Collect precipitation data from rain gauge stations and perform preprocessing.

[0007] Precipitation data from all available rain gauge stations within the study basin were collected. The inverse distance weighting (IDW) method was used to interpolate the selected rain gauge station precipitation data into gridded precipitation data with the same spatial resolution as satellite remote sensing precipitation products. The preprocessed IDW interpolated gridded precipitation data for the rain gauge stations was obtained, and its calculation formula is as follows: (1) (2) In the formula, For the first The distance from each rainfall station to the target grid point; The coordinates of the target grid point, The longitude of the target grid point. The latitude of the target grid point; , ) is the first Coordinates of a rainfall station For the first Longitude of each rain gauge station For the first The latitude of each rainfall station; This refers to the distance weighting coefficient. q This represents the total number of rain gauge stations. Number the rain gauge stations; ZFor the target grid point rain gauge station IDW interpolation grid precipitation data, For the first Rainfall data from various rain gauge stations.

[0008] Step 1.2: Collect satellite remote sensing precipitation data and perform preprocessing.

[0009] Based on the selected watershed, appropriate satellite remote sensing precipitation data is chosen. If the spatial resolution of the selected satellite remote sensing precipitation data differs from the target spatial resolution, bilinear interpolation is used to process the satellite remote sensing precipitation data to the target spatial resolution. If the spatial resolution of the selected satellite remote sensing precipitation data is the same as the target spatial resolution, the satellite remote sensing precipitation data at the target spatial resolution is directly downloaded. The time reference of the satellite remote sensing precipitation data is converted from UTC to China Standard Time to ensure consistency with the observation time of the rainfall stations, thus obtaining preprocessed satellite remote sensing precipitation data.

[0010] The formula for the bilinear interpolation method is as follows: (3) In the formula, ( , ), ( , ), ( , ), ( , The coordinates of the points surrounding the target grid point are shown below. The longitudes of the first and fourth points surrounding the target grid point are given. The longitudes of the second and third points surrounding the target grid point. Let the latitudes of the first and second points surrounding the target grid point be given. The latitudes of the third and fourth points surrounding the target grid point; The coordinates of the target grid point, The longitude of the target grid point. The latitude of the target grid point; Z 1. Z 2. Z 3. Z 4 represents the remote sensing grid precipitation data of points surrounding the target grid point. Z 1 represents the remote sensing grid precipitation data of the first point surrounding the target grid point. Z 2 represents the remote sensing grid precipitation data for the second point surrounding the target grid point. Z 3 represents the remote sensing grid precipitation data for the third point surrounding the target grid point. Z 4 represents the remote sensing grid precipitation data of the fourth point surrounding the target grid point; Bilinear interpolation of remote sensing grid precipitation data for target grid points.

[0011] Step 1.3: Collect and preprocess potential evapotranspiration and temperature data. Bilinear interpolation is used, and the time base is converted from UTC to China Standard Time. The data is then multiplied by the proportion of the watershed to each grid cell, summed, and averaged to obtain the average potential evapotranspiration and temperature surface data. Specifically: Step 1.3.1: Using the potential evapotranspiration data and temperature data from the ERA5-Land dataset, if the spatial resolution of the selected potential evapotranspiration data and temperature data is different from the target spatial resolution, bilinear interpolation is used to process the potential evapotranspiration data and temperature data to the target spatial resolution. If the spatial resolution of the selected potential evapotranspiration data and temperature data is the same as the target spatial resolution, the potential evapotranspiration data and temperature data at the target spatial resolution are directly downloaded. At the same time, the time base of the potential evapotranspiration data and temperature data is converted from UTC to China Standard Time.

[0012] Step 1.3.2: Since ERA5-Land specifies that the flux from the atmosphere to the surface is positive, while the flux from the surface to the atmosphere is negative, the potential evapotranspiration data is negative. It is necessary to use absolute value conversion to ensure that the potential evapotranspiration data is positive.

[0013] Step 1.3.3: Finally, calculate the proportion of the watershed in each grid, multiply it by the corresponding evapotranspiration grid data and temperature grid data respectively, sum them up and then take the average to obtain the average data of potential evapotranspiration surface and temperature surface.

[0014] Step 1.4: The preprocessed IDW interpolated grid precipitation data from rain gauge stations, satellite remote sensing precipitation data, average potential evapotranspiration data, and average temperature surface data are used as the raw data, and they are divided into training set and validation set.

[0015] Step 2: In order to eliminate the dimensional differences between different hydrological and meteorological data and avoid the risk of unbalanced gradient updates and overfitting during model training, the present invention performs dimensionless processing on the original data obtained in Step 1.

[0016] The Z-score standardization method is used to subtract the mean from the original data and then divide by its standard deviation, so that the processed data conforms to a standard normal distribution with a mean of 0 and a variance of 1. This yields dimensionless precipitation data from IDW interpolation grids at rain gauge stations, satellite remote sensing precipitation data, mean data of potential evapotranspiration, and mean data of temperature surface. The calculation formulas are as follows: (4) In the formula, v This is the original data. The mean of the original data. The standard deviation of the original data. This is the standardized data.

[0017] Step 3: Construct the watershed CNN-TCN hydrological model using the fusion module and the prediction module.

[0018] The fusion module takes dimensionless rainfall station IDW interpolated grid precipitation data and satellite remote sensing precipitation data processed in step 2 as input, and adopts an adaptive fusion weight-based approach. An improved convolutional neural network (CNN) extracts and fuses multi-source precipitation information (i.e., IDW interpolated grid precipitation from rain gauge stations and satellite remote sensing precipitation) to obtain more accurate fused precipitation. The prediction module takes the fused precipitation obtained from the fusion module, along with the dimensionless potential evapotranspiration average data and temperature surface average data processed in step 2, as inputs. These are used together in a temporal convolutional neural network (TCN) to simulate runoff, obtaining simulated runoff. By calculating the loss between simulated and measured runoff, the parameters of the fusion and prediction modules are continuously updated using the backpropagation algorithm, ultimately yielding a CNN-TCN hydrological model for the watershed. Specifically: Step 3.1, the fusion module adopts an adaptive fusion weight-based approach. The improved CNN performs two-dimensional convolution and pooling operations on the dimensionless rain gauge IDW interpolated grid precipitation data and satellite remote sensing precipitation data processed in step 2 to extract multi-source precipitation information and obtain more accurate fused precipitation data. Specifically: Step 3.1.1: Stack the dimensionless rain station IDW interpolated grid precipitation data and satellite remote sensing precipitation data processed in Step 2 in the channel dimension to form a three-dimensional input tensor, with the three dimensions being the number of channels, height, and width, and organize it into a five-dimensional tensor according to the time step, with the five dimensions being batch, time step, channel, height, and width.

[0019] Step 3.1.2, utilizing adaptive fusion weights An improved CNN convolutional layer extracts local spatial features to obtain a convolutional feature map. The pooling layer of the convolutional neural network downsamples the convolutional feature map, outputs a pooling feature map through the pooling layer, flattens the pooling feature map through a fully connected layer, and maps it to the final output value, thereby obtaining the fused precipitation.

[0020] The convolutional layer employs adaptive fusion weights. This is used to adjust the contribution of multi-source precipitation to feature extraction. The specific calculation process is as follows: (5) In the formula, Indicates the first Each channel has an adaptive fusion weight, and exp(·) represents the natural exponential function. Indicates the introduction of the first i The parameters corresponding to each channel are initialized before model training and continuously updated through backpropagation during training. A two-dimensional convolutional kernel is used to slide across multi-source precipitation to extract local spatial features, resulting in a convolutional feature map. The specific calculation process is as follows: (6) In the formula, Indicates the first Layer Each convolutional feature map Represents a non-linear activation function. Indicates the first Layer The input feature matrix for each channel contains two dimensions: height and width. The spatial region for sliding sampling of the convolution kernel. and This represents the weights and biases of the convolution kernel. Indicates the first Adaptive fusion weights for each channel.

[0021] The pooling layer downsamples the convolutional feature map output by the convolutional layer, reducing the spatial dimension, the number of parameters, and the model complexity, thereby preventing overfitting to some extent and obtaining a pooled feature map.

[0022] The fully connected layer flattens the pooling feature map output by the pooling layer and maps it to the final output value, thus obtaining the fused precipitation.

[0023] Step 3.2: The fused precipitation output from Step 3.1, along with the dimensionless average potential evapotranspiration and average temperature surface data processed in Step 2, are used as input data to the prediction module. The output is simulated runoff. The prediction module can adaptively learn the nonlinear time-varying mapping relationship between precipitation and runoff, and continuously optimize the parameters of the fusion module and the prediction module through the backpropagation algorithm, thereby obtaining the watershed CNN-TCN hydrological model and improving forecast accuracy. Specifically: Step 3.2.1: The fused precipitation output from Step 3.1 is concatenated with the dimensionless potential evapotranspiration average data and temperature surface average data processed in Step 2 to form a three-dimensional time input tensor. Then, a one-dimensional causal convolution is used to perform sliding calculations along the time axis to ensure that the current prediction relies only on historical information, conforms to the physical causal order of the precipitation and runoff process, and avoids forecast bias caused by time leakage, thereby obtaining the time convolution output features.

[0024] Step 3.2.2: The temporal convolution output features obtained in step 3.2.1 are processed by dilated convolution. By increasing the dilation coefficient layer by layer, the temporal receptive field is expanded to obtain multi-scale temporal features, enabling the model to capture the long-term impact of precipitation input on runoff response with fewer convolution layers.

[0025] Step 3.2.3: The multi-scale time features obtained in step 3.2.2 are processed using the residual connection method to make the prediction module more likely to learn complex nonlinear mappings and obtain residual output features.

[0026] Step 3.3: Integrate the residual output features obtained in Step 3.2.3 and map them to the final output value, which is the simulated runoff. By calculating the loss between the simulated runoff and the measured runoff, the parameters of the fusion module and the prediction module are continuously updated using the backpropagation algorithm, which is to continuously update the parameters of the watershed CNN-TCN hydrological model, and finally obtain the watershed CNN-TCN hydrological model.

[0027] Step 4: Based on the watershed CNN-TCN hydrological model constructed in Step 3, the gradient descent algorithm is used for training. The model parameters are updated with the goal of minimizing the loss function of the training set, and the model parameters corresponding to the minimum value of the loss function of the validation set are taken as the optimal model parameters, thus obtaining the trained watershed CNN-TCN hydrological model. At the same time, a learning rate decay scheduling strategy is adopted during the training process of the watershed CNN-TCN hydrological model, so that the learning rate gradually decreases with the training rounds to avoid oscillations in the optimization process.

[0028] Step 5, Result Evaluation.

[0029] The watershed CNN-TCN hydrological model obtained in step 4 was evaluated using the Nash efficiency coefficient (NSE), root mean square error (RMSE), and Pearson correlation coefficient (Corr).

[0030] (7) (8) (9) In the formula, Indicates the first Simulated values ​​for each time period, Indicates the first Measured values ​​for each time period Indicates the time period number. Indicates the number of simulation periods. This represents the mean of all simulated data. This represents the mean of all measured data.

[0031] In summary, this invention constructs a watershed CNN-TCN hydrological model comprising a fusion module and a prediction module, utilizing adaptive fusion weights. The improved CNN achieves the integration of the advantages of IDW interpolation grid precipitation from rain gauge stations and satellite remote sensing precipitation, resulting in more accurate fused precipitation. This fused precipitation is then used in conjunction with average data of potential evapotranspiration and average data of temperature surface for runoff simulation. This can effectively alleviate the problem of limited spatial representativeness of rain gauge stations and the difficulty in effectively capturing the spatiotemporal variability of precipitation in areas with scarce data, thereby improving the accuracy of hydrological forecasts in areas with scarce data.

[0032] The beneficial effects of this invention are as follows: (1) This invention constructs a watershed CNN-TCN hydrological model that includes a fusion module and a prediction module, integrates IDW interpolation grid precipitation from rain gauge stations with satellite remote sensing precipitation, fully integrates the advantages of multi-source precipitation, obtains more accurate fused precipitation, and thus uses it for hydrological forecasting, effectively improving the accuracy of hydrological forecasting in areas lacking data.

[0033] (2) This invention utilizes adaptive fusion weights The improved fusion module automatically adjusts the contribution of IDW interpolation grid precipitation from rain gauge stations and satellite remote sensing precipitation in feature extraction, and outputs fused precipitation that combines the advantages of accurate precipitation from rain gauge stations and spatial continuity of satellite remote sensing precipitation. This solves the problem of hydrological forecast deviation caused by the large uncertainty of precipitation input in traditional methods. (3) This invention integrates the average data of precipitation and potential evapotranspiration and the average data of temperature surface into the prediction module for hydrological forecasting. It adopts a causal convolution structure to avoid forecasting bias caused by time leakage. Through dilatational convolution and residual connection structure, it enhances the ability to learn complex nonlinear mappings, realizes accurate characterization of precipitation-runoff nonlinear response, significantly improves the accuracy of hydrological forecasting, and provides efficient and reliable technical support for flood control and disaster reduction and water resource management in watersheds with insufficient data. Attached Figure Description

[0034] Figure 1 This is a flowchart of the present invention.

[0035] Figure 2 This is a structural diagram of the hydrological model based on multi-source precipitation fusion in this invention; where Q represents measured runoff, Q _sim This represents simulated runoff, where T represents temperature and PET represents potential evapotranspiration. Indicates the merging of precipitation. This represents the Nth type of satellite remote sensing precipitation. This represents the IDW interpolation grid precipitation at rainfall stations, and loss represents the loss function between simulated and measured runoff. CNN IMPROVE Indicates adaptive fusion weights Improved convolutional neural network, TCN stands for Temporal Convolutional Neural Network.

[0036] Figure 3 This is a diagram showing the runoff forecast results of this invention. Detailed Implementation

[0037] The present invention will be further described below with reference to specific implementation examples.

[0038] This invention proposes a hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion, such as... Figure 1 As shown.

[0039] The Donghe River basin controlled by the Qingquan Township station was selected as the study area. The Donghe River is a first-order tributary on the left bank of the Jialing River, a tributary of the Yangtze River, with a total length of 294 km and a total drainage area of ​​5181 km². 2 The Donghe River basin is located on the northern edge of the Sichuan Basin and the southern foothills of the Daba Mountains, in a low-to-medium mountainous and hilly area. The terrain slopes from northeast to southwest, and it belongs to the subtropical humid monsoon climate zone. The average annual precipitation is 1164.6 mm, and the average annual flow at the river mouth is 106 m³ / s. 3 / s. The Qingquan Township Hydrological Station is located on the main stream of the Donghe River, controlling a drainage area of ​​4999 km². 2 This is an important hydrological control station within the basin. There are a total of nine rainfall stations in the basin, but their distribution is uneven, with some areas having sparse rainfall stations, making it a typical data-scarce region. Based on this, a hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion is adopted to improve the accuracy of runoff forecasting in this basin.

[0040] Step 1: Collect daily-scale precipitation data from rain gauge stations, satellite remote sensing precipitation data, potential evapotranspiration data, temperature data, and measured runoff data within the Qingquan Township watershed. Preprocess some of the collected data to obtain raw data, including IDW interpolated grid precipitation data from rain gauge stations, satellite remote sensing precipitation data, average potential evapotranspiration data, and average temperature data. This data is divided into training and validation sets. The training set covers the period from October 1, 2017 to December 31, 2021, and the validation set covers the period from January 2, 2015 to September 30, 2017. Specifically: Step 1.1: Collect precipitation data from rain gauge stations and perform preprocessing.

[0041] Rainfall data from all available rain gauge stations within the Qingquan Township watershed were collected. The inverse distance weighting (IDW) method was used to interpolate the selected rain gauge station rainfall data into a gridded rainfall dataset with a spatial resolution of 0.1°, resulting in preprocessed IDW interpolated gridded rainfall data. The calculation formula is as follows: (1) (2) In the formula, For the first The distance from each rainfall station to the target grid point The coordinates of the target grid point, The longitude of the target grid point. The latitude of the target grid point; , ) is the first Coordinates of a rainfall station For the first Longitude of each rain gauge station For the first The latitude of each rainfall station; This refers to the distance weighting coefficient. q This represents the total number of rain gauge stations. Number the rain gauge stations; Z For the target grid point rain gauge station IDW interpolation grid precipitation data, For the first Rainfall data from various rain gauge stations.

[0042] Step 1.2: Collect satellite remote sensing precipitation data and perform preprocessing.

[0043] Select IMERG-Late and GSMAP-Gauge-NRT satellite remote sensing precipitation data, directly download products with a spatial resolution of 0.1°, convert the time reference of the satellite remote sensing precipitation data from UTC to China Standard Time to ensure consistency with the observation time of the rain gauge stations, and obtain preprocessed satellite remote sensing precipitation data.

[0044] Step 1.3: Collect and preprocess potential evapotranspiration and temperature data. Using the potential evapotranspiration and temperature data from the ERA5-Land dataset, the time base is converted from UTC to China Standard Time. The data is then multiplied by the proportion of the watershed to each grid cell, summed, and averaged to obtain the average potential evapotranspiration surface data and the average temperature surface data. Specifically: Step 1.3.1: Using the potential evapotranspiration data and temperature data from the ERA5-Land dataset, directly download the product with a spatial resolution of 0.1°. At the same time, convert the time base of the potential evapotranspiration data and temperature data from UTC to China Standard Time.

[0045] Step 1.3.2: Since ERA5-Land specifies that the flux from the atmosphere to the surface is positive, while the flux from the surface to the atmosphere is negative, the potential evapotranspiration data is negative. It is necessary to use absolute value conversion to ensure that the potential evapotranspiration data is positive.

[0046] Step 1.3.3: Finally, calculate the proportion of the Qingquan Township watershed in each grid, multiply it by the corresponding evapotranspiration grid data and temperature grid data respectively, sum them up and then take the average to obtain the average data of potential evapotranspiration surface and temperature surface.

[0047] Step 1.4: The preprocessed IDW interpolated grid precipitation data from rain gauge stations, satellite remote sensing precipitation data, average potential evapotranspiration data, and average temperature surface data are used as the original data, and they are divided into training set and validation set. The training set is from 2017-10-01 to 2021-12-31, and the validation set is from 2015-01-02 to 2017-09-30.

[0048] Step 2: In order to eliminate the dimensional differences between different hydrological and meteorological data and avoid the risk of unbalanced gradient updates and overfitting during model training, the present invention performs dimensionless processing on the original data obtained in Step 1.

[0049] The Z-score standardization method is used to subtract the mean from the original data and then divide by its standard deviation, so that the processed data conforms to a standard normal distribution with a mean of 0 and a variance of 1. This yields dimensionless precipitation data from IDW interpolation grids at rain gauge stations, satellite remote sensing precipitation data, mean data of potential evapotranspiration, and mean data of temperature surface. The calculation formulas are as follows: (4) In the formula, v This is the original data. The mean of the original data. The standard deviation of the original data. This is the standardized data.

[0050] Step 3: Construct the CNN-TCN hydrological model for the Qingquan Township watershed using the fusion module and the prediction module.

[0051] The fusion module takes dimensionless rainfall station IDW interpolated grid precipitation data, IMERG-Late satellite remote sensing grid precipitation data, and GSMAP-Gauge-NRT satellite remote sensing grid precipitation data processed in step 2 as inputs, and employs an adaptive fusion weighting approach. An improved convolutional neural network (CNN) extracts and fuses multi-source precipitation information (i.e., IDW interpolated grid precipitation from rain gauge stations, IMERG-Late satellite remote sensing grid precipitation, and GSMAP-Gauge-NRT satellite remote sensing grid precipitation) to obtain more accurate fused precipitation. The prediction module takes the fused precipitation obtained from the fusion module, along with the dimensionless potential evapotranspiration average data and temperature surface average data processed in step 2, as inputs. These are used together in a temporal convolutional neural network (TCN) to simulate runoff, obtaining simulated runoff. By calculating the loss between simulated and measured runoff, the parameters of the fusion and prediction modules are continuously updated using the backpropagation algorithm, ultimately yielding the CNN-TCN hydrological model for the Qingquan Township watershed. Specifically: Step 3.1, the fusion module adopts an adaptive fusion weight-based approach. The improved CNN performs two-dimensional convolution and pooling operations on the dimensionless rain gauge IDW interpolated grid precipitation data and satellite remote sensing precipitation data processed in step 2 to extract multi-source precipitation information and obtain more accurate fused precipitation data. Specifically: Step 3.1.1: Stack the dimensionless rainfall station IDW interpolated grid precipitation data, IMERG-Late satellite remote sensing grid precipitation data, and GSMAP-Gauge-NRT satellite remote sensing grid precipitation data processed in Step 2 in the channel dimension to form a three-dimensional input tensor, with the three dimensions being the number of channels, height, and width, respectively. Then, organize it into a five-dimensional tensor according to the time step, with the five dimensions being batch, time step, channel, height, and width, respectively.

[0052] Step 3.1.2, utilizing adaptive fusion weights An improved CNN convolutional layer extracts local spatial features to obtain a convolutional feature map. The pooling layer of the convolutional neural network downsamples the convolutional feature map, outputs a pooling feature map through the pooling layer, flattens the pooling feature map through a fully connected layer, and maps it to the final output value, thereby obtaining the fused precipitation.

[0053] The convolutional layer employs adaptive fusion weights. This is used to adjust the contribution of multi-source precipitation to feature extraction. The specific calculation process is as follows: (5) In the formula, Indicates the first Each channel has an adaptive fusion weight, and exp(·) represents the natural exponential function. Indicates the introduction of the first i The parameters corresponding to each channel are initialized before model training and continuously updated through backpropagation during training. A two-dimensional convolutional kernel is used to slide across multi-source precipitation to extract local spatial features, resulting in a convolutional feature map. The specific calculation process is as follows: (6) In the formula, Indicates the first Layer Each convolutional feature map Represents a non-linear activation function. Indicates the first Layer The input feature matrix for each channel contains two dimensions: height and width. The spatial region for sliding sampling of the convolution kernel. and This represents the weights and biases of the convolution kernel. Indicates the first Adaptive fusion weights for each channel.

[0054] The pooling layer downsamples the convolutional feature map output by the convolutional layer, reducing the spatial dimension, the number of parameters, and the model complexity, thereby preventing overfitting to some extent and obtaining a pooled feature map.

[0055] The fully connected layer flattens the pooling feature map output by the pooling layer and maps it to the final output value, thus obtaining the fused precipitation.

[0056] Step 3.2: The fused precipitation output from Step 3.1, along with the dimensionless average potential evapotranspiration and average temperature surface data processed in Step 2, are used as input data to the prediction module. The output is simulated runoff. The prediction module can adaptively learn the nonlinear time-varying mapping relationship between precipitation and runoff, and continuously optimize the parameters of the fusion module and the prediction module through the backpropagation algorithm, thereby obtaining the CNN-TCN hydrological model for the Qingquan Township watershed and improving forecast accuracy. Specifically: Step 3.2.1: The fused precipitation output from Step 3.1 is concatenated with the dimensionless potential evapotranspiration average data and temperature surface average data processed in Step 2 to form a three-dimensional time input tensor. Then, a one-dimensional causal convolution is used to perform sliding calculations along the time axis to ensure that the current prediction relies only on historical information, conforms to the physical causal order of the precipitation and runoff process, and avoids forecast bias caused by time leakage, thereby obtaining the time convolution output features.

[0057] Step 3.2.2: The temporal convolution output features obtained in step 3.2.1 are processed by dilated convolution. By increasing the dilation coefficient layer by layer, the temporal receptive field is expanded to obtain multi-scale temporal features, enabling the model to capture the long-term impact of precipitation input on runoff response with fewer convolution layers.

[0058] Step 3.2.3: The multi-scale time features obtained in step 3.2.2 are processed using the residual connection method to make the prediction module more likely to learn complex nonlinear mappings and obtain residual output features.

[0059] Step 3.3: Integrate the residual output features obtained in Step 3.2.3 and map them to the final output value, which is the simulated runoff. By calculating the loss between the simulated runoff and the measured runoff, the parameters of the fusion module and the prediction module are continuously updated using the backpropagation algorithm, which is to continuously update the parameters of the CNN-TCN hydrological model of Qingquan Township watershed, and finally obtain the CNN-TCN hydrological model of Qingquan Township watershed.

[0060] Step 4: Based on the CNN-TCN hydrological model of Qingquan Township watershed constructed in Step 3, the gradient descent algorithm is used for training. Mean squared error (MSE) is selected as the loss function. The model parameters are updated with the goal of minimizing the loss function on the training set. The model parameters corresponding to the minimum loss function on the validation set are taken as the optimal model parameters, resulting in the trained CNN-TCN hydrological model of Qingquan Township watershed. Specifically, the number of training epochs is 43, the batch size is 16, the batch duration (seq_length) is 15, the learning rate (lr) is 0.001, the first convolutional layer has a kernel size of 4, 2 kernels, a stride of 1, and uses max pooling with a pooling kernel size of 2, the second convolutional layer has a kernel size of 3, 2 kernels, a stride of 1, and uses max pooling with a pooling kernel size of 2, the fully connected layer has an input dimension of 10, an output dimension of 1, a dilation factor of 4, and a residual block size of 2. Meanwhile, during the training of the CNN-TCN hydrological model in the Qingquan Township watershed, an exponential decay learning rate scheduling strategy (ExponentialLR, γ=0.95) was adopted to gradually reduce the learning rate with each training round in order to avoid oscillations in the optimization process.

[0061] Step 5, Result Evaluation.

[0062] The watershed CNN-TCN hydrological model obtained in step 4 was evaluated using the Nash efficiency coefficient (NSE), root mean square error (RMSE), and Pearson correlation coefficient (Corr).

[0063] (7) (8) (9) In the formula, Indicates the first Simulated values ​​for each time period, Indicates the first Measured values ​​for each time period Indicates the time period number. Indicates the number of simulation periods. This represents the mean of all simulated data. This represents the mean of all measured data.

[0064] In summary, this invention constructs a CNN-TCN hydrological model for the Qingquan Township watershed, comprising a fusion module and a prediction module, and utilizes an adaptive fusion weight-based approach. The improved CNN achieves the integration of the advantages of IDW interpolation grid precipitation from rain gauge stations and satellite remote sensing precipitation, resulting in more accurate fused precipitation. This fused precipitation is then used in conjunction with average data of potential evapotranspiration and average data of temperature surface for runoff simulation. This can effectively alleviate the problem of limited spatial representativeness of rain gauge stations and the difficulty in effectively capturing the spatiotemporal variability of precipitation in areas with scarce data, thereby improving the accuracy of hydrological forecasts in areas with scarce data.

[0065] The daily runoff simulation results of the CNN-TCN hydrological model for the Qingquan Township watershed during the validation period are as follows: Table 1: Simulation Results of Daily Runoff from the CNN-TCN Hydrological Model in Qingquan Township Watershed

[0066] It can be seen that in the daily runoff simulation during the validation period of the CNN-TCN hydrological model in the Qingquan Township watershed, the Nash efficiency coefficient (NSE) reached 0.856, and the root mean square error (RMSE) was 67.555m. 3 / s, Pearson correlation coefficient (Corr) is 0.938. Both NSE and Corr are close to 1, indicating that the model fits the overall runoff process very well, and the simulated values ​​are highly consistent with the measured values; the RMSE value is small, indicating that the simulation error is controlled at a low level.

[0067] Figure 3 This is a comparison chart of simulated and measured daily runoff results during the validation period of the CNN-TCN hydrological model in the Qingquan Township watershed. Figure 3As can be seen, the flow process curves simulated by the CNN-TCN hydrological model generally match the measured flow process curves well. During the dry season, the simulated flow and the measured flow are almost identical, indicating that the model has high accuracy in simulating low flow processes. During the flood season or flood events, the model can accurately capture the rise, peak, and recession of the flow, and the simulated peak magnitude and peak occurrence time are close to the measured values, indicating that the model has a good response capability to extreme hydrological events. This is likely because the IMERG-Late passive microwave (PMW) can accurately capture precipitation intensity. Through the synergistic effect of dual-frequency precipitation radar (DPR) and multi-channel microwave imager (GMI), it has advantages in capturing extreme precipitation. The GSMAP-Gauge-NRT uses cloud motion vector morphology algorithm to fully identify the duration of precipitation events. This invention integrates the precipitation products of the above two satellite remote sensing precipitation products with the IDW interpolated grid precipitation from rain gauge stations into the CNN-TCN hydrological model, thereby combining their respective advantages to obtain more accurate precipitation. It is also used together with potential evapotranspiration, temperature and other time-series hydrometeorological elements for runoff simulation, thus significantly improving the accuracy of runoff simulation.

[0068] The above-described embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion, characterized in that, The hydrological forecasting method for areas lacking data includes the following steps: Step 1: Collect precipitation data from all available rain gauge stations, publicly available satellite remote sensing precipitation data, potential evapotranspiration data, temperature data, and measured runoff data within the study basin. Preprocess some of the collected data to obtain raw data, including IDW interpolated grid precipitation data from rain gauge stations, satellite remote sensing precipitation data, average potential evapotranspiration data, and average temperature data, and divide them into training and validation sets. Step 2: Perform dimensionless processing on the raw data obtained in Step 1; Step 3: Construct a watershed CNN-TCN hydrological model using the fusion module and the prediction module; The fusion module takes dimensionless precipitation data from IDW interpolated grids of rain gauge stations (processed in step 2) and satellite remote sensing precipitation data as input, and employs an adaptive fusion weighting approach. An improved convolutional neural network (CNN) extracts and fuses multi-source precipitation information to obtain more accurate fused precipitation. The multi-source precipitation includes IDW interpolation grid precipitation from rain gauge stations and satellite remote sensing precipitation. The prediction module takes the fused precipitation obtained by the fusion module and the dimensionless potential evapotranspiration average data and temperature surface average data processed in step 2 as inputs, and uses them together with a temporal convolutional neural network (TCN) to simulate runoff, obtaining simulated runoff. By calculating the loss between simulated runoff and measured runoff, the parameters of the fusion module and the prediction module are continuously updated using the backpropagation algorithm, and finally the watershed CNN-TCN hydrological model is obtained. Step 4: Based on the watershed CNN-TCN hydrological model, use the gradient descent algorithm to train it, and obtain the trained watershed CNN-TCN hydrological model. The model parameters are updated with the goal of minimizing the loss function of the training set, and the model parameters corresponding to the minimum value of the loss function of the validation set are taken as the optimal model parameters to obtain the trained watershed CNN-TCN hydrological model. Step 5, Result Evaluation.

2. The hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 1, characterized in that, Specifically, step 1 is as follows: Step 1.1: Collect precipitation data from rain gauge stations and perform preprocessing; Precipitation data from all available rain gauge stations within the study basin were collected. The inverse distance weighting (IDW) method was used to interpolate the selected rain gauge station precipitation data into gridded precipitation data with the same spatial resolution as satellite remote sensing precipitation products. The preprocessed rain gauge station IDW interpolated gridded precipitation data was obtained, and its calculation formula is as follows: （1） （2） In the formula, For the first The distance from each rainfall station to the target grid point; The coordinates of the target grid point, The longitude of the target grid point. The latitude of the target grid point; , ) is the first Coordinates of a rainfall station For the first The longitude of each rain gauge station, For the first The latitude of each rainfall station; This refers to the distance weighting coefficient. q This represents the total number of rain gauge stations. Number the rain gauge stations; Z For the target grid point rain gauge station IDW interpolation grid precipitation data, For the first Rainfall data from individual rain gauge stations; Step 1.2: Collect satellite remote sensing precipitation data and perform preprocessing; Select appropriate satellite remote sensing precipitation data based on the research watershed. If the spatial resolution of the selected satellite remote sensing precipitation data is different from the target spatial resolution, use bilinear interpolation to process the satellite remote sensing precipitation data to the target spatial resolution. If the spatial resolution of the selected satellite remote sensing precipitation data is the same as the target spatial resolution, directly download the satellite remote sensing precipitation data at the target spatial resolution. The time reference of the satellite remote sensing precipitation data was converted from UTC to China Standard Time to ensure consistency with the observation time of the rain gauge stations, thus obtaining the preprocessed satellite remote sensing precipitation data; The formula for the bilinear interpolation method is as follows: （3） In the formula, ( , ), ( , ), ( , ), ( , The coordinates of the points surrounding the target grid point are shown below. The longitudes of the first and fourth points surrounding the target grid point are given. The longitudes of the second and third points surrounding the target grid point. Let the latitudes of the first and second points surrounding the target grid point be given. The latitudes of the third and fourth points surrounding the target grid point; The coordinates of the target grid point, The longitude of the target grid point. The latitude of the target grid point; Z 1. Z 2. Z 3. Z 4 represents the remote sensing grid precipitation data of points surrounding the target grid point. Z 1 represents the remote sensing grid precipitation data of the first point surrounding the target grid point. Z 2 represents the remote sensing grid precipitation data for the second point surrounding the target grid point. Z 3 represents the remote sensing grid precipitation data for the third point surrounding the target grid point. Z 4 represents the remote sensing grid precipitation data of the fourth point surrounding the target grid point; Bilinear interpolation of remote sensing grid precipitation data for target grid points; Step 1.3: Collect potential evapotranspiration data and temperature data and preprocess them; use bilinear interpolation to interpolate, and convert the time base from UTC to China Standard Time. Multiply the data of each grid by the proportion of the watershed to the corresponding grid data, sum them up and then take the average to obtain the average data of potential evapotranspiration surface and average data of temperature surface. Step 1.4: The preprocessed IDW interpolated grid precipitation data from rain gauge stations, satellite remote sensing precipitation data, average potential evapotranspiration data, and average temperature surface data are used as the raw data, and they are divided into training set and validation set.

3. The hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 2, characterized in that, Specifically, step 1.3 is as follows: Step 1.3.1: Using the potential evapotranspiration data and temperature data from the ERA5-Land dataset, if the spatial resolution of the selected potential evapotranspiration data and temperature data is different from the target spatial resolution, then bilinear interpolation is used to process the potential evapotranspiration data and temperature data to the target spatial resolution. If the spatial resolution of the selected potential evapotranspiration data and temperature data is the same as the target spatial resolution, then the potential evapotranspiration data and temperature data at the target spatial resolution are directly downloaded. At the same time, the time base of the potential evapotranspiration data and temperature data is converted from UTC to China Standard Time. Step 1.3.2: If the potential evapotranspiration data is negative, ensure that the potential evapotranspiration data is positive by converting it to an absolute value; Step 1.3.3: Calculate the proportion of the watershed in each grid, multiply it by the corresponding evapotranspiration grid data and temperature grid data respectively, sum them up and then take the average to obtain the average data of potential evapotranspiration surface and temperature surface.

4. The hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 3, characterized in that, Step 2 specifically includes: The Z-score standardization method is used to subtract the mean from the original data and divide it by the standard deviation, so that the processed data meets the standard normal distribution with a mean of 0 and a variance of 1, thereby obtaining dimensionless rain station IDW interpolated grid precipitation data, satellite remote sensing precipitation data, potential evapotranspiration average data and temperature surface average data. The calculation formula is as follows: （4） In the formula, v This is the original data. The mean of the original data. The standard deviation of the original data. This is the standardized data.

5. A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 4, characterized in that, Step 3 specifically includes: Step 3.1, the fusion module adopts an adaptive fusion weight-based approach. The improved CNN performs two-dimensional convolution and pooling operations on the dimensionless rain gauge IDW interpolated grid precipitation data and satellite remote sensing precipitation data processed in step 2 to extract multi-source precipitation information and obtain fused precipitation; specifically: Step 3.1.1: Stack the dimensionless rain station IDW interpolated grid precipitation data and satellite remote sensing precipitation data processed in Step 2 in the channel dimension to form a three-dimensional input tensor, with the three dimensions being the number of channels, height, and width, and organize it into a five-dimensional tensor according to the time step, with the five dimensions being batch, time step, channel, height, and width; Step 3.1.2, utilizing adaptive fusion weights The improved CNN convolutional layer extracts local spatial features to obtain a convolutional feature map. The pooling layer of the convolutional neural network is used to downsample the convolutional feature map. The pooling layer outputs a pooling feature map. The pooling feature map is flattened by a fully connected layer and mapped to the final output value to obtain the fused precipitation. Step 3.2: The fused precipitation output from Step 3.1, along with the dimensionless average potential evapotranspiration and average temperature surface data processed in Step 2, are used as input data to the prediction module. The output is simulated runoff. The prediction module can adaptively learn the nonlinear time-varying mapping relationship between precipitation and runoff, and continuously optimize the parameters of the fusion module and the prediction module through the backpropagation algorithm, thereby obtaining the watershed CNN-TCN hydrological model; specifically: Step 3.2.1: The fused precipitation output from Step 3.1 is concatenated with the dimensionless potential evapotranspiration average data and temperature surface average data processed in Step 2 to form a three-dimensional time input tensor. Then, a one-dimensional causal convolution is used to perform sliding calculation along the time axis to obtain the time convolution output features. Step 3.2.2: The temporal convolution output features obtained in step 3.2.1 are processed by dilated convolution. By increasing the dilation coefficient layer by layer, the temporal receptive field is expanded to obtain multi-scale temporal features, so that the model can capture the long-term impact of precipitation input on runoff response with fewer convolution layers. Step 3.2.3: The multi-scale temporal features obtained in step 3.2.2 are processed using the residual connection method to obtain residual output features; Step 3.3: Integrate the residual output features obtained in Step 3.2.3 and map them to the final output value, which is the simulated runoff. By calculating the loss between the simulated runoff and the measured runoff, the parameters of the fusion module and the prediction module are continuously updated using the backpropagation algorithm, and the parameters of the watershed CNN-TCN hydrological model are continuously updated, finally obtaining the watershed CNN-TCN hydrological model.

6. A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 5, characterized in that, In step 3.1.2: The convolutional layer employs adaptive fusion weights. This is used to adjust the contribution of multi-source precipitation to feature extraction; the specific calculation process is as follows: （5） In the formula, Indicates the first Each channel has an adaptive fusion weight, and exp(·) represents the natural exponential function. Indicates the introduction of the first i The parameters corresponding to each channel are used; a two-dimensional convolution kernel is slid across the multi-source precipitation to extract local spatial features and obtain a convolution feature map; The pooling layer: downsamples the convolutional feature map output by the convolutional layer to reduce the spatial dimension, reduce the number of parameters, reduce the model complexity, and obtain a pooled feature map. The fully connected layer flattens the pooling feature map output by the pooling layer and maps it to the final output value to obtain the fused precipitation.

7. A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 6, characterized in that, In step 4, a learning rate decay scheduling strategy is adopted during the training of the watershed CNN-TCN hydrological model, so that the learning rate gradually decreases with each training round.

8. A hydrological forecasting method for data-scarce areas based on multi-source precipitation fusion according to claim 7, characterized in that, In step 5: The watershed CNN-TCN hydrological model obtained in step 4 was evaluated using the Nash efficiency coefficient, root mean square error, and Pearson correlation coefficient. The trained watershed CNN-TCN hydrological model, which includes a fusion module and a prediction module, utilizes adaptive fusion weights. The improved CNN can combine the advantages of IDW interpolation grid precipitation from rain gauge stations with satellite remote sensing precipitation to obtain accurate fused precipitation. This fused precipitation can then be used together with average potential evapotranspiration data and average temperature surface data for runoff simulation, thereby improving the accuracy of hydrological forecasts in data-scarce areas.

Images