Rainfall runoff simulation method for physical mechanism guided deep learning

Abstract

The invention provides a rainfall runoff simulation method based on physical mechanism guided deep learning. A hydrological physical process and hydrometeorological big data are combined; a penalty term of a preferred function is set; a deep learning simulation scheme under the guidance of a physical mechanism is realized. According to the runoff simulation method provided by the invention, an extreme event sample is adopted as punishment, a rainstorm event and a long-term rainless event are considered, a brand-new calculation formula is provided, and compared with a traditional deep learningmethod, the runoff simulation method provided by the invention solves the problem that deep learning lacks a physical mechanism, and avoids non-monotonicity in the simulation process. The method can be widely applied to rainfall runoff simulation, runoff simulation and prediction can be completed systematically and completely, and a basis is provided for scientific decision making.

Description

technical field [0001] The invention belongs to the technical field of hydrological models, and in particular relates to a rainfall runoff simulation method guided by a physical mechanism for deep learning. Background technique [0002] Runoff prediction is the basis for water resources management, allocation and efficient use. The hydrological model obtains meteorological data such as precipitation to calculate the runoff, so as to achieve the purpose of forecasting and forecasting. [0003] Theory-based models are well suited for representing processes where known scientific principles are conceptually well understood. However, because of the inherent complexity of hydrological processes, traditional hydrological models have difficulty expressing incomprehensible processes outside the current body of knowledge. In this case, theory-based models are often forced to make some simplifying assumptions about the physical process and set some conceptual parameters, which not o...

AI Technical Summary

Problems solved by technology

This method ignores the complex underlying surface conditions and hydrological processes, and learns and stores a large number of input-output model mapping relationships to understand and analyze the change law of runoff. However, it does not have a physical mechanism and can only capture the correlation between variables. relationship and thus cannot fully satisfy the goal of understanding causality in scientific problems

[0005] The data-driven rainfall runoff simulation mainly adopts the deep learning method to solve time series problems, but the existing deep learning rainfall runoff simulation methods have the following problems: (1) The probability of extreme rainfall in historical data is very low, which leads to the model's accuracy of extreme rainfall. The flood peak flow simulation error of heavy rain is relatively large, and the phenomenon of water imbalance is very easy to occur; (2) it often leads to the appearance of negative flow, and the model will also collapse when faced with the extreme situation of no rain for a long time; (3) the application is based on When the hydrological model of the physical mechanism simulates runoff, the relationship between rainfall and runoff is monotonous. The greater the rainfall, the greater the runoff, and the greater the evaporation, the smaller the runoff. However, the traditional deep learning rainfall-runoff simulation method cannot guarantee the monotonicity of the simulation.

Images