A snowmelt runoff hybrid forecasting method and system

Abstract

The present invention discloses a hybrid forecasting method and system for snowmelt runoff, which relates to the technical field of hydrological forecasting and solves the technical problem of insufficient reliability of snowmelt runoff prediction in complex terrain and meteorological conditions in the prior art. The present invention comprises constructing a long short-term memory network model based on a "dynamic-static" hybrid drive, namely a DS-LSTM model, and training and verifying the DS-LSTM model using a data set generated by study area data. The DS-LSTM model uses long-term memory characteristics to capture the influence of previous meteorological factors on snow accumulation characteristics, and depicts the effect of local static underlying surfaces on snow accumulation characteristics through multi-dimensional learning capabilities. The snowmelt calculation module in the snowmelt runoff model is replaced with a verified DS-LSTM model to obtain a snowmelt runoff hybrid forecasting model, and the snowmelt runoff hybrid forecasting model is used to realize snowmelt runoff hybrid forecasting. The present invention can provide snowmelt calculation data with multi-factor dynamic correction for the snowmelt runoff model.

Description

Technical Field

[0001] The present invention relates to the technical field of hydrological forecasting, and in particular to a snowmelt runoff hybrid forecasting method and system. Background Art

[0002] In recent years, climate change has led to increased snow and glacier melt, shifting the spatiotemporal distribution of snowmelt runoff and further increasing the uncertainty of snowmelt runoff simulation and forecasting. Improving the accuracy of snowmelt runoff simulation and forecasting is of great significance for reservoir flood control, power generation, and integrated water resources management. Currently, most snowmelt runoff models are physically based and fall into two main categories: conceptual models based on degree-day factors and energy balance models. Energy balance-based models comprehensively account for the input and output of various energy sources during snowmelt and can more accurately describe the snowmelt process. However, they require high-quality observational data and are complex parameterization processes. This makes them computationally expensive, especially for long-term simulations over large-scale regions, and difficult to implement in real time. Conceptual models based on degree-day factors assume a linear correlation between the snowmelt rate and temperature and use empirical formulas to determine the relationship between average daily temperature and snowmelt volume. These models are simple to use, computationally efficient, and are widely used for snowmelt runoff forecasting in alpine regions. However, the model performs poorly under complex terrain and meteorological conditions. It uses only temperature as the main driving factor and cannot accurately capture the nonlinear and dynamic changes of snow accumulation, resulting in insufficient reliability of snowmelt runoff predictions. Summary of the Invention

[0003] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a hybrid forecasting method and system for snowmelt runoff, which solves the technical problem that the prior art has insufficient reliability in predicting snowmelt runoff under complex terrain and meteorological conditions.

[0004] A hybrid forecasting method for snowmelt runoff includes constructing a long short-term memory network model (DS-LSTM model) based on a "dynamic-static" hybrid drive, and training and verifying the DS-LSTM model using a data set generated by study area data. The DS-LSTM model uses long-term memory characteristics to capture the influence of previous meteorological factors on snow accumulation characteristics, and depicts the effect of local static underlying surfaces on snow accumulation characteristics through multidimensional learning capabilities. The snowmelt calculation module in the snowmelt runoff model is replaced with the verified DS-LSTM model to obtain a hybrid forecasting model for snowmelt runoff, and the hybrid forecasting model for snowmelt runoff is used to implement hybrid forecasting of snowmelt runoff.

[0005] Furthermore, the study area data includes dynamic meteorological data, static underlying surface characteristic data, snow characteristic site measured data and snow remote sensing data.

[0006] Furthermore, the data acquisition process of the study area includes:

[0007] Dynamic meteorological data: Collect and integrate precipitation, temperature, radiation, etc. in the basin. These data are refined in space and time through interpolation and reanalysis techniques to ensure data integrity and consistency.

[0008] Static underlying surface characteristic data: obtain DEM, aspect, slope, land use type, soil, etc., perform spatial rasterization on the collected data (for example, unify to 0.25° resolution), and use interpolation methods to fill missing values ​​and process abnormal data to ensure data integrity and consistency.

[0009] Snow characteristic station measured data: collect existing snow water equivalent measurement station data and select stations with long-term observation data to ensure that the selected stations have less missing data and higher observation frequency;

[0010] Snow remote sensing data: Select snow remote sensing products with a long series and stable update frequency. After completing the data download, perform preliminary data processing.

[0011] Furthermore, the generated data set includes snow observation sites screened according to the measured data of snow characteristic sites, determining the grid position of each site after screening, and extracting dynamic meteorological data, static underlying surface characteristic data and snow remote sensing data from the study area data based on the grid position.

[0012] Furthermore, the DS-LSTM model includes a dynamic input module, a static feature integration module and a feature fusion module. The dynamic input module uses the memory unit of LSTM to capture the dynamic change characteristics of meteorological elements in the time series as dynamic features, thereby accurately predicting future snow characteristics; the static feature integration module combines the multi-dimensional data mining capabilities of the model to capture the influence of local underground surface characteristics on snow distribution and characteristics as static features; the feature fusion module introduces an embedding layer to map the static features to the same feature space of the dynamic features, splices the dynamic features output by the dynamic input module with the static features output by the static feature integration module to form a mixed feature vector, and then introduces an attention mechanism to perform weighted processing on the mixed feature vector to enhance the contribution of key features to the prediction of snow characteristics.

[0013] Furthermore, the training verification includes setting a loss function, a hyperparameter optimization algorithm, an early stopping strategy, and a Patience parameter. If the verification performance does not improve within multiple consecutive training cycles specified by the Patience parameter, the training will be terminated. After the training is terminated, any snow observation station is extracted as a verification station. The snow water equivalent remote sensing data, dynamic meteorological data, and static underlying surface data at the station location are used as input data and fed into the trained DS-LSTM model for verification. The RMSE is calculated based on the verification data to perform a business evaluation of the corrected snow water equivalent data. If the evaluation is qualified, the final DS-LSTM model is obtained. Furthermore, the snowmelt runoff model adopts the HBV model, which includes a snowmelt calculation module, a soil module, and a runoff module. The snowmelt calculation module in the HBV model is replaced with the trained and verified DS-LSTM to obtain the snowmelt runoff hybrid forecast model DS-LSTM-HBV, which realizes dynamic correction of snow accumulation.

[0014] Furthermore, after the construction is completed, the DS-LSTM-HBV is calibrated and verified to improve the accuracy of snowmelt runoff mixing prediction.

[0015] A snowmelt runoff hybrid forecasting system includes a data input module, a prediction module and an output module. The prediction module includes a snowmelt runoff hybrid forecasting model constructed using a snowmelt runoff hybrid forecasting method. The data input module is used to input measured data from snow observation stations. The prediction module is used to predict the snow water equivalent of the snow observation stations based on the measured data input by the data input module. The output module is used to output the snow water equivalent predicted by the prediction module.

[0016] Furthermore, it also includes a dynamic database of meteorological elements, a static database of underlying surface characteristics, a snow characteristic site measurement database and a snow remote sensing database, which are used to store various measured data of snow observation sites.

[0017] The beneficial effects of the present invention include:

[0018] 1. It can provide snowmelt calculation data with multi-factor dynamic correction for snowmelt runoff models.

[0019] 2. Dynamic snow data correction based on snow observation station data can achieve accurate correction and spatial interpolation of snow remote sensing data under different terrain and climatic conditions, providing more reliable input data for watershed management and hydrological forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of a snowmelt runoff hybrid forecasting method involved in an embodiment of the present application.

[0021] Figure 2This is a verification effect diagram of the snow water equivalent correction results of a first site involved in an embodiment of the present application.

[0022] Figure 3 This is a verification effect diagram of the simulation results of the snowmelt runoff model at a second site involved in an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0024] A hybrid snowmelt runoff forecast method, such as Figure 1 As shown, it includes constructing a long short-term memory network model based on "dynamic-static" hybrid drive, namely the DS-LSTM model, and using the data set generated by the study area data to train and verify the DS-LSTM model. The DS-LSTM model uses the long-term memory characteristics to capture the influence of previous meteorological factors on snow characteristics, and uses multi-dimensional learning capabilities to depict the effect of local static underlying surfaces on snow characteristics. The snowmelt calculation module in the snowmelt runoff model is replaced with the verified DS-LSTM model to obtain a snowmelt runoff hybrid forecast model, and the snowmelt runoff hybrid forecast model is used to realize snowmelt runoff hybrid forecast.

[0025] In another embodiment, the study area data includes dynamic meteorological data, static underlying surface characteristic data, snow characteristic site measured data, and snow remote sensing data. Specifically, the study area data acquisition process includes:

[0026] Dynamic meteorological data: Collect and integrate precipitation, temperature, radiation, etc. in the basin. These data are refined in space and time through interpolation and reanalysis techniques to ensure data integrity and consistency.

[0027] Static underlying surface characteristic data: obtain DEM, aspect, slope, land use type, soil, etc., perform spatial rasterization on the collected data (for example, unify to 0.25° resolution), and use interpolation methods to fill missing values ​​and process abnormal data to ensure data integrity and consistency.

[0028] Measured data of snow characteristics stations: collect existing snow water equivalent measured station data, and screen out stations with long-term observation data, so that the screened stations have less missing data and higher observation frequency; among them, for stations with long-term observation data, considering that the measured station data are mostly short and discontinuous, it is sufficient to screen out relatively long series.

[0029] Snow remote sensing data: Select snow remote sensing products with a long series and stable update frequency. After completing the data download, perform preliminary data processing.

[0030] In this embodiment,

[0031] Dynamic meteorological data: Download a snow depth dataset for a specific station from a data platform; download precipitation and temperature data (CN05) from a meteorological website (spatial resolution 0.25°×0.25, daily) from 2012 to 2019; and download radiation data (1km resolution) from a data cloud platform. After downloading, resample and rasterize the data to a uniform 0.25° resolution.

[0032] Static underlying surface characteristic data: Download a digital elevation map (DEM) of a specific location from a data platform. This elevation map is a 1km-scale data set generated from 1:250,000 contour lines and elevation points. It includes DEM, hill shade, slope, and aspect maps. Download a 30-meter annual land cover dataset and its dynamic changes (1985-2022) from a data center. Download annual 30-meter resolution land use data for a specific location from 1990 to 2022 from a resource cloud. Download a high-resolution soil information grid basic attribute dataset (2010-2018) from a data platform. After downloading, resample and grid the data to a 0.25° resolution.

[0033] Snow characteristic site measured data: from a data center

[0034] Download the snow depth dataset for regular stations in a typical snow-covered area (2017-2019). This dataset contains manual and instrumental observation data from 41 regular meteorological stations in a typical snow-covered area from 2017 to 2019. From this dataset, select 25 stations with long, continuous data series. Interpolation is used to fill missing values ​​and address outliers to ensure data integrity and consistency.

[0035] Snow remote sensing data: Log in to a website to query AMSR data products (spatial resolution: 0.25°, time: July 2, 2012 to present, daily snow water equivalent data), write Python code to automatically download daily snow water equivalent data, and convert the AMSR data from the initial he5 format to the nc format. The AMSR snow remote sensing data information is converted from the EASE-GRID projection to the WGS84 projection.

[0036] In another embodiment, the generated data set includes snow observation sites screened based on measured data of snow characteristic sites, determining the grid position of each site after screening, and extracting dynamic meteorological data, static underlying surface characteristic data and snow remote sensing data from the study area data based on the grid position.

[0037] In this embodiment, the grid positions of 25 meteorological observation stations in typical snow-covered areas were screened from the measured data of snow characteristic stations, and daily-scale dynamic meteorological data (precipitation, temperature and radiation), static underlying surface data (elevation, slope aspect, land use and soil type), and AMSR snow remote sensing data from 2017 to 2019 were extracted to ensure that the model input covers the key dynamic and static factors affecting snow cover changes, providing a solid data foundation for the accuracy and applicability of the model.

[0038] In another embodiment, the DS-LSTM model includes a dynamic input module, a static feature integration module and a feature fusion module. The dynamic input module uses the memory unit of LSTM to capture the dynamic change characteristics of meteorological elements in the time series as dynamic features, thereby accurately predicting future snow characteristics; the static feature integration module combines the multi-dimensional data mining capabilities of the model to capture the influence of local underground surface characteristics on snow distribution and characteristics as static features; the feature fusion module introduces an embedding layer to map the static features to the same feature space of the dynamic features, that is, to map the static features to the same grid position of the dynamic features, and splices the dynamic features output by the dynamic input module with the static features output by the static feature integration module to form a mixed feature vector, and then introduces an attention mechanism to perform weighted processing on the mixed feature vector to enhance the contribution of key features to the prediction of snow characteristics.

[0039] In another embodiment, the training verification includes setting a loss function, a hyperparameter optimization algorithm, an early stopping strategy, and a Patience parameter. If the verification performance does not improve within multiple consecutive training cycles specified by the Patience parameter, the training will be terminated; after the training is terminated, any snow observation site is extracted as a verification site, and the snow water equivalent remote sensing data, dynamic meteorological data, and static underlying surface data at the site location are used as input data and brought into the trained DS-LSTM model for verification. The RMSE is calculated based on the verification data to perform business evaluation on the corrected snow water equivalent data. If the evaluation is qualified, the final DS-LSTM model is obtained.

[0040] In this example, observation data from 25 common weather stations in typical snow-covered areas are used as true values, with 2017 and 2018 serving as training periods and 2019 serving as validation periods. The parameter ranges are set, including the number of hidden layer neurons (50-100), the time step (10-100), and the learning rate (0.001-0.01). The extracted daily dynamic meteorological data and static geographic feature data from 2017 to 2019 are input into the LSTM model for training. By continuously adjusting the model parameters and optimizing the model performance until RMSE ≤ 0.85, the model has achieved high calibration accuracy and the model is complete.

[0041] Model effectiveness evaluation. To verify the universality of the model, representative stations were selected and their AMSR remote sensing data, dynamic data, and static data were used as input data for the DS-LSTM model to obtain corrected snow water equivalent data. Figure 2 The correction results of AMSR snow data for a certain first station from October 1 to December 31, 2017 are shown. The RMSE value after correction is 0.85, which meets the validation requirements of the model.

[0042] In another embodiment, the snowmelt runoff model adopts the HBV model, which includes a snowmelt accumulation calculation module, a soil module and a confluence module. The snowmelt accumulation calculation module in the HBV model is replaced with the trained and verified DS-LSTM to obtain the snowmelt runoff hybrid forecasting model DS-LSTM-HBV, thereby realizing dynamic correction of snow accumulation.

[0043] Specifically, calculating snow accumulation and melt is crucial. This module uses the degree-day factor method to calculate snow accumulation and melt, whose magnitude is controlled by the basin's temperature and previous snow accumulation. The DS-LSTM remote sensing snow correction model obtained in the above steps replaces the HBV model's snow accumulation and melt calculation module. This constructs the DS-LSTM-HBV hybrid snowmelt-runoff forecast model, which uses intelligent correction of remote sensing snow products to achieve dynamic correction of the model's snow accumulation.

[0044] In another embodiment, after the construction is completed, the DS-LSTM-HBV is calibrated and verified to improve the accuracy of snowmelt runoff mixing prediction.

[0045] In this example, the newly constructed DS-LSTM-HBV model was calibrated and verified to obtain a snowmelt runoff model with high accuracy. Figure 3 The model simulation results of the snowmelt runoff model of a second station before and after replacing the snowmelt calculation module from March 1 to May 31, 2021 are shown. The R2 value of the snowmelt runoff simulation before replacement was 0.76, and the R2 value of the snowmelt runoff simulation after replacement was 0.87. The model accuracy was improved after replacement, indicating that the remote sensing snow intelligent correction method that replaces the snowmelt calculation module can improve the simulation and forecast accuracy of the snowmelt runoff model.

[0046] A snowmelt runoff hybrid forecasting system includes a data input module, a prediction module and an output module. The prediction module includes a snowmelt runoff hybrid forecasting model constructed using a snowmelt runoff hybrid forecasting method. The data input module is used to input measured data from snow observation stations. The prediction module is used to predict the snow water equivalent of the snow observation stations based on the measured data input by the data input module. The output module is used to output the snow water equivalent predicted by the prediction module.

[0047] In another embodiment, it also includes a dynamic database of meteorological elements, a static database of underlying surface characteristics, a snow characteristic site measurement database and a snow remote sensing database, which are used to store various types of measured data of snow observation sites.

[0048] The above-described embodiments merely represent specific implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of protection of the present application. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the technical concept of the present application, and all such variations and improvements fall within the scope of protection of the present application.

Claims

1. A mixed snowmelt runoff forecasting method, characterized in that: include: A DS-LSTM model, a long-short-term memory network model based on a dynamic-static hybrid drive, was constructed and trained and verified using a dataset generated from data in the study area. The DS-LSTM model includes a dynamic input module, a static feature integration module, and a feature fusion module. The dynamic input module uses the LSTM memory unit to capture the dynamic change characteristics of meteorological elements in the time series as dynamic features. The static feature integration module combines the multidimensional data mining capabilities of the model to capture the impact of local underground surface characteristics on snow distribution and characteristics as static features; The feature fusion module introduces an embedding layer, maps static features to the same feature space of dynamic features, splices the dynamic features output by the dynamic input module with the static features output by the static feature integration module to form a mixed feature vector, and then introduces an attention mechanism to perform weighted processing on the mixed feature vector; the snowmelt calculation module in the snowmelt runoff model HBV is replaced with a trained and verified DS-LSTM model to obtain a snowmelt runoff hybrid forecast model DS-LSTM-HBV, and the snowmelt runoff hybrid forecast model DS-LSTM-HBV is used to realize snowmelt runoff hybrid forecast; the data in the study area include dynamic meteorological data, static underlying surface feature data, snow accumulation feature station measured data and snow accumulation remote sensing data.

2. A mixed snowmelt runoff forecasting method according to claim 1, characterized in that: The data acquisition process of the study area includes: Dynamic meteorological data: collect and integrate precipitation, temperature, and radiation within the basin, and perform spatial and temporal refinement through interpolation and reanalysis techniques; Static underlying surface characteristic data: obtain DEM, aspect, slope, land use type, soil, perform spatial rasterization on the collected data, and use interpolation methods to fill missing values ​​and handle abnormal data; Snow characteristic site measurement data: collect existing snow water equivalent measurement site data and select sites with long-term observation data; Snow remote sensing data: Select snow remote sensing products, complete data download, and perform preliminary data processing.

3. A mixed snowmelt runoff forecasting method according to claim 1, characterized in that: The generated dataset includes snow observation sites screened based on measured data of snow characteristic sites, determines the grid position of each site after screening, and extracts dynamic meteorological data, static underlying surface characteristic data and snow remote sensing data from the study area data based on the grid position.

4. The snowmelt runoff mixed forecasting method according to claim 1, characterized in that: The training verification includes setting a loss function, a hyperparameter optimization algorithm, an early stopping strategy, and a Patience parameter. If the verification performance does not improve within multiple consecutive training cycles specified by the Patience parameter, the training will be terminated. After the training is terminated, any snow observation site is extracted as a verification site, and the snow water equivalent remote sensing data, dynamic meteorological data, and static underlying surface data at the site location are used as input data and brought into the trained DS-LSTM model for verification. The RMSE is calculated based on the verification data to perform business evaluation on the corrected snow water equivalent data. If the evaluation is qualified, a trained and verified DS-LSTM model is obtained.

5. A mixed snowmelt runoff forecasting method according to claim 1, characterized in that: After the construction is completed, DS-LSTM-HBV is calibrated and verified.

6. A snowmelt runoff hybrid forecasting system, characterized in that: It includes a data input module, a prediction module and an output module. The prediction module includes a snowmelt runoff hybrid prediction model DS-LSTM-HBV constructed using a snowmelt runoff hybrid prediction method according to any one of claims 1 to 5. The data input module is used to input measured data of a snow observation station. The prediction module is used to predict the snow water equivalent of the snow observation station based on the measured data input by the data input module. The output module is used to output the snow water equivalent predicted by the prediction module.

7. A snowmelt runoff mixed forecasting system according to claim 6, characterized in that: It also includes a dynamic database of meteorological elements, a static database of underlying surface characteristics, a snow characteristic site measurement database and a snow remote sensing database, which are used to store various measured data from snow observation sites.

Images