PCA-based multivariate LSTM model for predicting natural groundwater level variations in a time-series record affected by anthropogenic factors
| Autor: | Chan-Ik Hwang, Myoung-Rak Choi, Gyoo-Bum Kim |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Global and Planetary Change
geography Multivariate statistics geography.geographical_feature_category Soil Science Geology Soil science Groundwater recharge Pollution Natural (archaeology) Water level Principal component analysis Environmental Chemistry Environmental science Surface water Groundwater Earth-Surface Processes Water Science and Technology Water well |
| Zdroj: | Environmental Earth Sciences. 80 |
| ISSN: | 1866-6299 1866-6280 |
| Popis: | Time series of natural groundwater level considering rainfall effects are usually used for the estimation of groundwater recharge, long-term trend analysis, and assessment of interactions between surface water and groundwater along rivers. However, anthropogenic activities, such as groundwater pumping, land excavation, and barrier construction, may induce abnormal changes in water levels. This study aimed to develop a universal long short-term memory (LSTM) model for predicting natural water level variations in a time-series record that has been affected by groundwater abstraction or other anthropogenic factors. This model uses past and present groundwater levels, rainfall and representative principal components of groundwater level time series as input variables. For this purpose, 17 cases of the developed LSTM model were tested using 13 monitoring wells, of which the case with the highest prediction performance was selected. Among the test cases, case 6 was found to achieve the highest performance, with average RMSE and MAE values of 0.061 and 0.027, respectively. The case 6 model used rainfall, groundwater level of monitoring wells, and four main principal components (1–4) as input variables. Also, its optimum window size was found to be 5. The accuracy of the LSTM model was found to be more strongly affected by window size than by input variables. Although the case 6 LSTM model may have errors for some monitoring wells, it has high potential as a universal model that can consistently determine natural groundwater levels in South Korea. This LSTM model makes it universally available for water level prediction, even with long missing periods in groundwater level time series or in the absence of adjacent observations for model input. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full text from SpringerLink