Flash flood early warning method based on deep learning

Abstract

The invention discloses a flash flood early warning method based on deep learning. According to the method, an RNN (recurrent neural network) model with an LTSM (long short-term memory) function is adopted. Problems of gradient disappearance and gradient explosion in the RNN can be effectively solved by adding the LTSM to a model algorithm, prediction of small watersheds or flash flood areas can be performed more effectively, reliable flash flood disaster early warning and predictionare achieved, and prediction time is greatly shortened. The model is only required to focus on input data, so that the model has the characteristics of being convenient to operate and high in operation speed, and facilitating value shifting to other areas. Moreover, with the practical application of the model,data is accumulated continuously, no secondary parameter adjustment is required, parameters can be optimized and adjusted actively by the model, and accordingly, the model is simpler and more convenient to use.

Description

technical field [0001] The invention relates to the field of disaster early warning, in particular to a deep learning-based mountain torrent early warning method. Background technique [0002] The dynamic early warning of mountain torrents is realized by issuing early warnings to areas in danger of torrents of torrents. As of 2018, my country has delineated flash flood danger areas nationwide, and Sichuan Province has more than 28,000 flash flood danger areas. These flash flood warnings are currently carried out by setting rainfall or water level warning indicators. The specific method is to set up rainfall or water level stations in the area, link the flash flood danger area (can be multiple) to the corresponding measuring station, and give early warning through the actual measured rainfall (water level). In order to prolong the warning time and improve the warning accuracy, some areas have begun to use hydrological models for flash flood warning, but this warning method ...

AI Technical Summary

Problems solved by technology

First, the theoretical structure is not perfect. Both models are highly generalized hydrological phenomena, and there are theoretical flaws in physical processes such as actual precipitation and evaporation, precipitation and runoff, surface interception and infiltration.

The second is that there are too many model parameters. In the actual adjustment process, in order to better fit the actual situation, some parameters will lose their physical meaning after adjustment.

The third is poor fault tolerance, and it is difficult to build models for areas with poor data or no data

Fourth, the forecast time is limited by the confluence time of the river basin, and the forecast time for small watersheds (mountain flood areas) is relatively short. Even if the forecast is made, it is difficult for the disaster reduction department to respond in time

Fifth, the calculation process is complex, slow, and difficult to operate. Since the construction of the model is to simulate the flood process in the river basin, it is very complicated. Although the model has been generalized, it still has many parameters and a huge amount of calculation. In some cases, different parameters are used, resulting in slow generation results, difficult operation, and difficult promotion of applications

Images