Runoff time series prediction using LSTM dynamic neural network optimized by logistic chaotic mapping chicken swarm algorithm

Abstract

Abstract An algorithm, named long short-term memory (LSTM)-logistic chaos mapping chicken swarm algorithm (LCCSA), is proposed for initializing the weights and thresholds of LSTM neural networks using the Logistic chaotic mapping chicken swarm algorithm (CSA). This algorithm aims to improve mid- to long-term runoff sequence prediction in river basins. In this model, the logistic chaotic mapping method is used to initialize the chicken swarm, and LCCSA is employed to pre-train the weights and thresholds of each layer of LSTM 50 times, using the training results' initial weights of LSTM to enhance convergence accuracy and speed. Taking the Manas river and Kuitun river, two typical basins in northern Xinjiang, China, as the research objects, LSTM-LCCSA was used to forecast the mean monthly runoff in the mid- to long-term under different lag time series by using the runoff evolution data within a certain period. The example using the basin located in northern Xinjiang demonstrates the effectiveness, stability, and generality of the LSTM-LCCSA method in mid- to long-term prediction of average monthly runoff, and the prediction accuracy and universality of LSTM-LCCSA are better than other data-driven models.