Enhanced Long Short-Term Memory Model for Runoff Prediction

Abstract

Runoff prediction plays a crucial role in the scheduling and management of water resources. A novel enhanced long short-term memory (LSTM) model called LN-LSTM-PSO is proposed by integrating layer normalization (LN), LSTM network, and particle swarm optimization (PSO) to improve prediction accuracy. The model in general is a data-driven model, and compared with the traditional mechanism model, its most notable advantage is that it has a flexible structure. In the enhanced LSTM model, LN is added to the hidden layer of the LSTM model, and PSO is used to determine the optimal parameters. The application of the proposed enhanced model is illustrated using hydrological and meteorological data from Jiulong River Basin in Fujian Province, China, to test model accuracy. Results indicate that the proposed model has high accuracy. Specifically, the enhanced LSTM model has a root mean square error of 0.142, mean absolute percentage error of 0.032, Nash–Sutcliffe efficiency of 0.968, and determination coefficient of 0.968. The enhanced runoff prediction model exhibits improved performance compared with a support vector regression model, an artificial neural network, a recurrent neural network, and a long short-term memory network. LN can accelerate the convergence speed of the LSTM network, and PSO substantially increases model performance by automating the hyperparameter selection. The virtual runoff input of the current time step can also remarkably improve model accuracy.