A field flood forecasting method considering deep learning coupled with similar inflow rates

Abstract

This invention discloses a flood forecasting method for each flood event considering deep learning coupled with similar inflows, belonging to the field of hydrological forecasting technology. The method acquires a set of target floods and historical floods to be forecasted. Based on the feature vectors of each historical flood in the target flood and historical flood sets, it calculates the mechanistic similarity weight between the target flood and historical floods. It calculates the number of effective samples based on the similarity weights, adaptively filters and calibrates the sample set, and uses a perturbation gating mechanism to eliminate periods of human interference, constructing a weighted objective function to calibrate the model parameters. Using a data-driven model configured with a differentiable time-shifted layer, it corrects the amplitude and aligns the peak time of the preliminary forecast sequence of the physical model. This invention solves the problems of traditional similarity matching neglecting physical mechanisms and deep learning's difficulty in correcting phase errors, thus improving the accuracy and robustness of flood forecasting.

Images