Multivariate Time-Series Forecasting of Boiler Water Conductivity in Power Plant Using Machine Learning

Abstract

Water quality is a critical parameter for maintaining the reliable operation of power plants, particularly in the water–steam cycle and Heat Recovery Steam Generator (HRSG) systems. This study develops a predictive monitoring approach based on multivariate time-series forecasting to predict HP Drum Water conductivity using 35,064 hourly observations, of which 34,083 valid observations remained after preprocessing. Three models, namely Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), were evaluated at forecasting horizons of 1, 6, and 24 h. The results show that XGBoost achieved the best performance at the 1-h horizon, with an RMSE of 3.3193, an MAE of 1.6358, and an R² of 0.9379. LSTM achieved the best performance at the 6-h and 24-h horizons, with R² values of 0.7256 and 0.5966, respectively. SHapley Additive exPlanations (SHAP) analysis identified current conductivity as the most influential feature. The predictive monitoring system also detected two early-warning events when conductivity exceeded the statistical threshold of approximately 44–45, demonstrating its potential to support predictive water-quality monitoring and earlier operational decision-making.