Preprint: Predicting Infiltration Rates in Soil Aquifer Treatment: A Comparative Study of Machine Learning Models Using Operational and Weather Data

Abstract

We present a comparative benchmark of machine learning models with static and temporal sequence inputs for predicting infiltration rates in Soil Aquifer Treatment systems, trained on ten years of operational and weather records. Using expanding-window cross-validation, nonlinear models achieved accurate basin-level predictions without explicit water-quality or soil-property measurements. Gradient Boosting performed best among cycle-aggregated tabular input models using current-cycle information and basin identity. LSTM achieved comparable performance for temporal 365-day inputs. Feature ablation shows that adding weather variables to models using basin identity and operational conditions did not yield a clear improvement in our setting, with LSTM achieving 0.654 cm/h MAE without weather information. Integrated Gradients showed strongest per-day attributions for recent inputs, with weaker, model-dependent long-range contributions. Our findings suggest that data-driven models based on recent, routinely available operational records can support SAT simulation, basin-level planning, and future control strategies.