Multi-step tap-water quality forecasting in South Korea with transformer-based deep learning model

The prediction of tap water quality serves as a pivotal component in enhancing water resource management. The intricate nonlinearity and inherent instability in water quality data make this task challenging. In this paper, we present a Tap-Water Quality Temporal Prediction Network (TWQ-TPN) to accurately predict tap-water quality by focusing on the impact of temporal nonlinear patterns and long-term seasonal fluctuations. To achieve this, we design two modules, namely the Temporal Feature Extraction Module (TFEM) and the Feature Transformation and Prediction Module (FTPM). The TFEM learns complex dynamic nonlinear features in the temporal domain. The FTPM is to realize feature transformation in the high-dimensional features for long-term seasonal fluctuations. Thus, our TWQ-TPN can accurately predict tap water quality trends to help improve water management. We validate TWQ-TPN's superiority using 5 years' data from 33 major water facilities in South Korea, demonstrating excellence in pH, turbidity, and residual chlorine. Ablation experiments support TWQ-TPN's rationale.