Hybrid Modelling of Water Quality Dynamics: Data Assimilation With Machine Learning for Enhanced Predictions

Abstract

Predicting Escherichia coli concentrations in recreational waters is essential for safeguarding public health and ensuring water quality compliance. This study applies time series analysis to forecast E. coli levels at six sites in New Zealand using historical data from 2005 to 2020. The goal is to develop a reliable predictive model that helps in proactive water management and early contamination warnings. Initially, an autoregressive integrated moving average (ARIMA) model was applied with parameters selected through a stepwise fitting approach. However, ARIMA demonstrated limitations in accurately capturing E. coli variability due to external environmental factors. Then the seasonal autoregressive integrated moving average with exogenous regressors (SARIMAX) model was applied for better predictive performance using water quality parameters and climate variables as input predictors. Results showed that no single water quality parameter consistently predicted E. coli across all sites, though total phosphorus emerged as a key predictor in five locations. The four-year forecasts showed patterns aligned with historical trends, suggesting reasonable predictive capability. However, forecast accuracy varied across sites, likely due to site-specific hydrological conditions. This study highlights the importance of site-specific modelling, real-time environmental data integration, and advanced machine learning techniques to improve water quality predictions. A refined forecasting approach can support early warning systems and risk-based decision-making, ultimately reducing health risks associated with microbial contamination in recreational waters.