Flood prediction method and device based on long-short-term memory network and transfer learning

Abstract

The invention provides a flood prediction method and device based on a long-short-term memory network and transfer learning, and the method comprises the steps: screening a watershed with high similarity as a source watershed for data transfer through the cosine similarity calculation of the natural environment of the watershed after flood data information of a resource watershed is collected; calculating the flood similarity between a source watershed and a target watershed by using a DTW algorithm, forming an input matrix by using the calculated flood similarity and hydrological data as theinput of the model, and calculating weighted data by using an Attention mechanism, so that the model can better mine the information of important characteristic factors; and finally, inputting the weighted data matrix into an LSTM network to establish a flood prediction model, and predicting the future flood flow of the target watershed. According to the method, the characteristic of data shortagein a flood prediction problem of a specific watershed is overcome, the superior performance of the LSTM network for processing the time sequence problem is fully utilized, and flood prediction is more accurate.

Description

technical field [0001] The invention relates to the field of flood prediction, in particular to a flood prediction method and device based on long short-term memory network and transfer learning. Background technique [0002] Flood disaster is a kind of common natural disaster, it is because the oversized surface runoff river channel can't accommodate, the water in the depression can't be removed in time, or the sea surface rises suddenly and the seawater invades the land and floods. Flood disaster is one of the most serious natural disasters in the world. It not only occurs frequently and widely, but also has far-reaching impact. China has a vast territory, complex terrain, and significant monsoon climate. It is a country with frequent floods and a wide range of impact in the world. one. As one of the most common natural disasters, flood is destructive and poses a huge threat to society. Therefore, practical and effective flood forecasting methods are conducive to rapid r...

AI Technical Summary

Problems solved by technology

[0004] However, the structure of the hydrological conceptual model is limited because it is not yet possible to rigorously describe these sub-processes using mathematical equations derived from considering the physical properties of the watershed

In addition, their optimization methods determine the dependence of model parameters on measured rainfall flow data, which is limited in practical applications

Traditional data-driven models do not consider the physical process of rainfall flow patterns, which leads to trying to utilize all available hydrological data, which greatly increases the feature size and requires a large amount of training data

Moreover, when dealing with the flood prediction problem of a designated watershed, due to the suddenness, timeliness, and short-term nature of floods, the hydrological data of the designated watershed are often very scarce, which cannot provide enough training data for the model

Images