Deep Learning-Based Predictive Framework for Groundwater Level Forecast in Arid Irrigated Areas

An accurate groundwater level (GWL) forecast at multi timescales is vital for agricultural management and water resource scheduling in arid irrigated areas such as the Hexi Corridor, China. However, of GWL these remains a challenging task owing to deficient hydrogeological data highly nonlinear, non-stationary complex system. The development reliable simulation models necessary profound. In this study, novel ensemble deep learning predictive framework integrating pro-processing, feature selection, uncertainty analysis was constructed. Under framework, hybrid model equipped with currently most effective algorithms, including complete empirical mode decomposition adaptive noise (CEEMDAN) decomposition, genetic algorithm (GA) belief network (DBN) model, quantile regression (QR) evaluation, denoted CEEMDAN-GA-DBN, proposed 1-, 2-, 3-month ahead three observation wells Jiuquan basin, northwest capability CEEMDAN-GA-DBN compared CEEMDAN-DBN standalone DBN terms performance metrics R, MAE, RMSE, NSE, RSR, AIC Legates McCabe’s Index well criterion MPI PICP. results demonstrated higher degree accuracy better objective than all lead times wells. Overall, reduced RMSE testing period by about 9.16 17.63%, while it improved their NSE 6.38 15.32%, respectively. also affirmed slightly reliability method 2- horizons. derived proved ability time steps forecasting, thus, can be used an tool forecasting areas.