Groundwater level predictions in the Thames Basin, London over extended horizons using Transformers and advanced machine learning models

Abstract

This study breaks new ground by using the Temporal Fusion Transformer (TFT) method for groundwater level prediction, addressing the complex dynamics of the Thames Basin aquifer in England. Our research combines extensive hydrological data collected from the Thames Basin with advanced machine learning, where a complex network of rivers and streams substantially affects groundwater dynamics. Unlike previous studies, this research focuses on long-term forecasting with deep learning, offering, for the first time, a 60-day prediction horizon based on daily data. To rigorously examine the model performance and robustness on new, unseen data, we applied the walk-forward validation method and other matrices such as RMSE and R2 coupled with the Holdout technique. The models used were Long Short-Term Memory (LSTM), Attention-based LSTM, LSTM with Bayesian optimisation, Attention-based LSTM with Bayesian optimisation and TFT. They were used on the basin's Chalk, Jurassic Limestone, and Lower greensand aquifers. Whilst both LSTM models were optimised using the Bayesian technique, TFT was applied for its inherent capability in complex time series. Our methodology processed historical groundwater and rainfall data from 2001 to 2023, accounting for the potential lag in aquifer response to the proximity of the river system. The dataset served as training, validation, and holdout for each model, focusing on capturing the dynamic temporal fluctuation. The results clearly showed the superiority of the TFT model in all aquifer types compared to other models across all horizons. The Limestone had the greatest result in the 7-day projections, with an RMSE of 0.02 and R2 of 0.98; Whilst the Chalk and Lower greensand, had RMSEs of 0.03 with R2 values of 0.75 and 0.95, respectively. The Limestone aquifer performed best for the 30-day horizon again (RMSE=0.06, R2=0.85), with the Chalk and Lower greensand aquifer yielding RMSE of 0.04 and 0.12 and R2 values of 0.64 and 0.74, respectively. In the 60 days predictions, the best results were observed in the limestone aquifer with RMSE of 0.09 and R2 of 0.65 in holdout validation. However, in chalk and lower greensand aquifers, the TFT showed RMSEs of 0.05 and 0.15 and R2s of 0.45 and 0.58, respectively. Traditional LSTM models demonstrated limited predictive power compared to the main model TFT, while the attention mechanism slightly improved the accuracy. This study not only sets a new benchmark in hydrological modelling accuracy but also highlights the potential of advanced machine learning in managing complex aquifers and predicting the water table.