A novel method for lake level prediction: deep echo state network

Abstract

Accurately prediction of lake level fluctuations is essential for water resources planning and management. In the present study, the potential of a novel method, deep echo state network (Deep ESN), is investigated for monthly lake level prediction and its results are compared with three data-driven methods, artificial neural networks (ANNs), extreme learning machine (ELM), and regression tree (Reg. Tree). The methods are validated using root mean square errors (RMSE), determination coefficient (R2), and Nash-Sutcliffe efficiency (NSE) criteria. The investigated method (Deep ESN) outperforms the ELM, ANNs, and Reg. Tree by improving accuracies by 61–62–96%, 10–14–84%, and 8–23–80% in prediction 1 month, 2 months, and 3 months ahead lake level fluctuations in terms of RMSE criteria, respectively. Also, accuracy of ELM, ANNs, and Reg. Tree was significantly increased using Deep ESN model by 1.1–1.1–443%, 1.1–1.6–250%, and 1.6–6.5–184% in terms of NSE indicator for different lead-time horizons. Among the ELM, ANNs, and Reg. Tree, the third method provides the worst predictions while the first method performs superior to the second one in all tree time horizons.