Streamflow Forecast via VIC Model and Deep Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current streamflow forecast methods, both process-driven and data-driven, face challenges such as excessive dependence on meteorological input data accuracy, high computational costs, model complexity, and poor explainability, leading to suboptimal forecast performance and inability to meet high-efficiency and high-accuracy requirements.
Innovation Solution
A long-term streamflow forecast method and system based on process-data synergic drive, which combines a composite model that includes a VIC distributed hydrological model with bias correction and feature variable screening, using a LASSO regression model and deep learning, to reduce computational costs and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If process-driven models are used to assimilate historical meteorological data and forecast streamflow, then the forecast can incorporate physical mechanisms and watershed characteristics, but the accuracy is excessively dependent on meteorological input data accuracy and computational costs are high
Solution Approach 1:
The model segments the watershed into multiple grids with different spatial resolutions, allowing computationally intensive process-driven calculations only in critical areas while using simpler data-driven approaches in other regions, thereby reducing overall computational cost while maintaining forecast accuracy
Solution Approach 2:
The model dynamically adjusts the level of process-driven detail based on watershed characteristics and available computational resources, transitioning between fully process-driven and data-driven modes to optimize the balance between accuracy and computational efficiency
2Measurement precision
If distributed hydrological models are used to improve spatial accuracy, then the model can capture watershed characteristics, but computer storage and running costs increase significantly
Solution Approach 1:
The model applies different grid resolutions and process complexity levels to different spatial locations based on their importance and data availability, using fine-resolution distributed modeling only where necessary while coarser resolutions are used elsewhere, reducing storage requirements while preserving spatial accuracy in critical areas
Solution Approach 2:
The model transforms the spatial dimension by aggregating grid outputs to different scales, allowing distributed model results to be summarized at watershed or sub-watershed levels, thereby reducing the quantity of stored data while maintaining spatial information where needed
3Productivity
If data-driven models are used to establish mapping relationships between historical data and predictors, then computational efficiency improves, but explainability and physical mechanism consideration deteriorate
Solution Approach 1:
The model uses process-driven model outputs as an intermediary layer that translates physical mechanisms into data-driven predictors, allowing the data-driven component to learn from physically meaningful variables rather than raw data, thereby maintaining explainability while improving efficiency
Solution Approach 2:
The model creates a composite forecasting system that combines process-driven and data-driven components, where the process-driven part provides physical mechanism-based predictions and the data-driven part captures empirical patterns, with both working synergistically to achieve both efficiency and explainability
4Duration of action of moving object
If climate model forecasts are used as inputs for hydrological models, then long-term forecasts can be generated, but the spatial accuracy is relatively low and fails to satisfy input accuracy requirements
Solution Approach 1:
The model segments the climate model output domain into multiple regions and applies different downscaling or adjustment techniques to each, improving spatial accuracy by capturing regional variations while maintaining the long-term forecast capability at the watershed scale
Solution Approach 2:
The model replaces the direct use of low-resolution climate model outputs with a hybrid approach that substitutes in observed meteorological data where available and uses process-driven calculations to upscale and refine climate model predictions, thereby improving spatial accuracy without sacrificing forecast lead time
Data Source
AI summary
The present invention provides a long-term streamflow forecast method and system based on process-data synergic drive. The long-term streamflow forecast method includes: step 1, collecting data; step 2, constructing a VIC model; step 3, performing interpolation, bias correction, and disaggregation to obtain daily data; step 4, by using climate model forecasts, driving the VIC model; step 5, constructing an improved VIC model by selecting grid cells from the distributed hydrological model, among which the series of monthly soil moisture forecasts in the third layer in all grid cells are treated as independent variables and the monthly streamflow of the outlet hydrological station of the studied watershed is treated as the dependent variables; step 6, driving the improved VIC model to perform the gridded runoff yield calculation of a full time period; step 7, forming a candidate predictor set; step 8, screening predictors and training a deep learning model to obtain a composite model, and performing long-term streamflow forecast.