RESUMEN
Accurate prediction of regional terrestrial water storage change (TWSA) is of great significance for water resources planning and management, and early warning of extreme climate disasters. Aiming at the problem that the conventional methods on prediction of TWSA time series are difficult to be accurate, the six typical regions are selected in China as examples, including the upper reaches of the Yangtze River (UYR), the southwest region (SWR), the Liaohe River Basin (LRB), the North China Plain (NCP), the Qinghai-Tibet Plateau (QTP), and the Pearl River Basin (PRB). The mascon product from GRACE/GRACE-FO provided by CSR is used to extract TWSA time series in six typical areas. The improved Back Propagation (BP) neural network, Long Short-Term Memory (LSTM) neural network and the latest Bidirectional LSTM (BiLSTM-attention) neural network model based on attention mechanism are proposed to predict and analyze the regional TWSA. In the experiment, the selection of the optimal model parameters such as the number of hidden layer nodes and the number of hidden units of the neural network model is tested and analyzed in detail. Meanwhile, the model prediction results are compared with the traditional least squares method and random forest (RF) prediction method. The root mean square error (RMSE), determination coefficient (R2), Nash-Sutcliffe efficiency coefficient (NSE) and mean absolute percentage error (MAPE) were used to evaluate the accuracy of the predicted results. The results show that the improved BP, LSTM and Bi-LSTM-attention neural network models all achieve higher prediction accuracy in UYR and SWR areas. RMSE is less than 2.641 cm, R2 is as high as 0.8 or more, NSE is above 0.6, and MAPE is within 0.1. Compared with the least square method, the RMSE of the predicted results from three neural network decreased by 0.998 cm, 0.700 cm and 0.7563 on average, and the R2 increased by 81.75%, 69.89% and 72% on average. Compared with RFML method, the RMSE from three neural network is reduced by 0.601 cm, 0.316 cm and 0.360, and R2 is increased by 38.20%, 24.60% and 27.06% on average. NSE and RMSE are improved to varying degrees in the above regions. It shows that the improved BP, LSTM and BiLSTM-attention model used can effectively predict TWSA. The research methods and results in this paper can provide important reference for the rational utilization of regional water resources and disaster risk assessment.
RESUMEN
Monitoring and predicting the regional groundwater storage (GWS) fluctuation is an essential support for effectively managing water resources. Therefore, taking Shandong Province as an example, the data from Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On (GRACE-FO) is used to invert GWS fluctuation from January 2003 to December 2022 together with Watergap Global Hydrological Model (WGHM), in-situ groundwater volume and level data. The spatio-temporal characteristics are decomposed using Independent Components Analysis (ICA), and the impact factors, such as precipitation and human activities, which are also analyzed. To predict the short-time changes of GWS, the Support Vector Machines (SVM) is adopted together with three commonly used methods Long Short-Term Memory (LSTM), Singular Spectrum Analysis (SSA), Auto-Regressive Moving Average Model (ARMA), as the comparison. The results show that: (1) The loss intensity of western GWS is significantly greater than those in coastal areas. From 2003 to 2006, GWS increased sharply; during 2007 to 2014, there exists a loss rate - 5.80 ± 2.28 mm/a of GWS; the linear trend of GWS change is - 5.39 ± 3.65 mm/a from 2015 to 2022, may be mainly due to the effect of South-to-North Water Diversion Project. The correlation coefficient between GRACE and WGHM is 0.67, which is consistent with in-situ groundwater volume and level. (2) The GWS has higher positive correlation with monthly Global Precipitation Climatology Project (GPCP) considering time delay after moving average, which has the similar energy spectrum depending on Continuous Wavelet Transform (CWT) method. In addition, the influencing facotrs on annual GWS fluctuation are analyzed, the correlation coefficient between GWS and in-situ data including the consumption of groundwater mining, farmland irrigation is 0.80, 0.71, respectively. (3) For the GWS prediction, SVM method is adopted to analyze, three training samples with 180, 204 and 228 months are established with the goodness-of-fit all higher than 0.97. The correlation coefficients are 0.56, 0.75, 0.68; RMSE is 5.26, 4.42, 5.65 mm; NSE is 0.28, 0.43, 0.36, respectively. The performance of SVM model is better than the other methods for the short-term prediction.