Water Quality Forecasting Using Time Series Techniques

Water quality measurement is critical for assessing the health of water samples; however, it lacks predictive capabilities to support proactive decision-making and precautionary measures. Predicting future water quality levels requires analyzing key water components, known as quality parameters, which are measured using sensors installed in water sources. However, these sensors are costly and prone to malfunctions, causing the production of missing data. In this paper, we employ SARIMA and Prophet models to forecast the levels of individual quality parameters over a 15-day period. We also investigate the impact of missing values on models' performance and explore causal relationships between different quality parameters. Our results demonstrate that Prophet outperforms SARIMA in forecasting most parameters, requiring less computational power while achieving superior accuracy. Additionally, Prophet exhibits greater robustness to missing values compared to SARIMA. Though causal relationships between parameters were identified, not all of them had a positive impact on prediction accuracy. The best Mean Absolute Percentage Error (MAPE) values obtained for pH, temperature, ORP, turbidity, conductivity, and combined chlorine were 0.7%, 3.93%, 0.57%, 3.72%, 5.2%, and 8.2%, respectively.