Multi-objective intelligent flood control operation rules extraction for reservoirs-lake system based on long and short-term memory neural networks coupled with physical constraints

Abstract

Study region

Chaohu Basin, lower Yangtze River region, China.

Study focus

This study proposes a multi-objective intelligent operation rules extraction method for flood control in reservoirs-lake system, which integrates Long Short-Term Memory (LSTM) networks with physical constraints. The method extends the input factors of the model to realize the prediction of multi-objective non-inferior solutions set and incorporates hydrological physical constraints into the model’s loss function. This coupling method improves both the model’s physical interpretability and the predictive ability in yielding multi-objective non-inferior solutions over reservoirs-lake system, which minimizes flood indices on reservoirs, lake and flood control point.

New hydrological insights for the region

The results show that compared to the conventional LSTM, the CP-LSTM (Physical Constrained Long Short-Term Memory) model demonstrates the following advantages: (1) The CP-LSTM reduces Root Mean Square Error by 0.33 % and increases Nash-Sutcliffe efficiency by 1.12 %, indicating improved accuracy in prediction modeling decision variables. Moreover, it significantly enhances peak flow prediction accuracy, improving the limitation of low peak flow prediction accuracy in the LSTM model; (2) In terms of objective space prediction, the maximum reduction in Objective Values Error reaches 8.00 %, while the Area Overlap Ratio of the predicted objective space to the true objective space improves by 21.50 %. By integrating hydrological physical constraints into the deep learning framework and extending it to multi-objective decision-making problems, this study provides a novel and effective approach for Artificial Intelligence -based flood control rules extraction and its application in multi-objective flood operation strategies.