Application of Nonlinear Time Series and Machine Learning Algorithms for Forecasting Groundwater Flooding in a Lowland Karst Area

Abstract

AbstractIn karst limestone areas interactions between ground and surface waters can be frequent, particularly in low lying areas, linked to the unique hydrogeological dynamics of that bedrock aquifer. In extreme hydrological conditions, however, this can lead to wide‐spread, long‐duration flooding, resulting in significant cost and disruption. This study develops and compares a nonlinear time‐series analysis based nonlinear autoregressive model with exogenous variables (NARX), machine learning based near support vector regression as well as a linear time‐series ARX model in terms of their performance to predict groundwater flooding in a lowland karst area of Ireland. The models have been developed upon the results of several years of field data collected in the area, as well as the outputs of a highly calibrated semi‐distributed hydraulic/hydrological model of the karst network. The prediction of total flooding volume indicates that the performances of all the models are similarly accurate up to 10 days into the future. A NARX model taking inputs of the past 5 days' flood volume; rainfall data and tidal amplitude data across the past 4 days, showed the best flood forecasting performance up to 30 days into the future. Existing real‐time telemetric monitoring of water level data at two points in the catchment can be fed into the model to provide an early warning flood warning tool. The model also predicts freshwater discharge from the inter‐tidal spring into the Atlantic Ocean which hitherto had not been possible to monitor.