Hybrid metaheuristic machine learning approach for water level prediction: A case study in Dongting Lake

Abstract

A reliable water level prediction in a lake system is crucial for water resources management, flood control, etc. The objective of this study is to propose a machine learning model which is able to achieve a considerably high level of accuracy in terms of water level prediction. Dongting Lake, which is the second-largest freshwater lake system in China, was selected as the study area. The hourly water level, flow rate, rainfall and temperature of the upstream water stations and rainfall of the downstream water stations were used as the input features, to predict the water level at the downstream stations. Multilayer perceptron neural network (MLP-NN), Elman neural network (ENN), and integration of particle swarm optimisation algorithm to Elman neural network (PSO-ENN) were selected as the model development techniques. The PSO-ENN model appears as the best performed model, as it records NSE of 0.929–0.988, RMSE of 0.129–0.322 and MAE of 0.151–0.359 at the downstream stations in Dongting Lake. The PSO-ENN model also shows its ability to provide better performance for the water level prediction of 36 h in advance. In terms of input variables sensitivity, the developed model is most sensitive to flow rate, followed by rainfall.