An interpretable and adaptable data-driven model for performance prediction in thermal plants

To safely operate complex industrial systems such as thermal power plants, establishing reliable monitoring tools is paramount for better understanding the underlying processes. Data-driven models are a useful aid for monitoring and control of thermal power plants, but they require an effective feature selection to allow for an accurate, computationally efficient, and interpretable model. In this study, we systematically compared three different modes of feature selection for predicting the live steam flow in a thermal plant: purely expert-based, purely data-driven, and a hybrid combining both. While a fully data-driven approach yields the highest accuracy, a hybrid approach, refined from more than 3,000 features, achieves nearly equivalent precision (NMAE = 1.14%) while using only 44 physical sensor signals, significantly improving the computational efficiency and enabling interpretability. The model is dynamically retrained using a sliding window approach to effectively handle load variations and plant shutdowns, which allows for the real-time tracking of deviations from the expected performance. We further validated our approach on a second thermal plant, achieving an NMAE of 2.49% despite substantial operational differences. By balancing predictive accuracy, interpretability, and transferability across plants, this work provides a practical framework for robust, data-driven monitoring and decision support in complex industrial power systems. • A prediction model is trained using on-site thermal plant data (>3,000 features). • Feature selection is performed by expert knowledge and/or data-driven algorithms. • A hybrid model (expert + data-driven) accurately predicts a performance parameter. • The feature importances are evaluated and the model physics can be interpreted. • The method is applied to a similar plant proving the adaptability of this approach.