Boiler operation predictions by integrating thermo-fluid principles within an artificial neural network framework

Optimising boiler operations is challenging due to fluctuating conditions in complex thermo-fluid systems. This study introduces a novel approach to improve efficiency in coal-fired boilers by developing and validating an artificial neural network (ANN) model that provides both statistically accurate and scientifically feasible predictions. Three multi-layer perceptron (MLP) feedforward ANN models were developed, with variable selection supported by principal component analysis (PCA) and hyperparameter optimisation performed using Latin hypercube sampling (LHS). The best ANN achieved test root mean square errors (RMSEs) of 2.11 t/h for steam flow, 2.11 t/h for blowdown, 4.98 °C for superheated steam temperature, 0.69 bar for steam pressure, and 0.86 % for efficiency. The mean absolute percentage error (MAPE) for efficiency remained below 1.25 %, with deviations constrained within ±4.25 %. Statistical and thermodynamic validations were applied, including bootstrap aggregation of prediction variance and mass and energy balance checks. Results showed that 96.76 % of samples achieved water mass balance deviations of less than 0.01 %. Furthermore, 100 % of predictions for efficiency and energy output fell within a 5 % absolute error range. The novelty of this work lies in integrating ANN predictions with thermo-fluid validation. Theoretically, it advances current literature by bridging the gap between statistical accuracy and physical feasibility. Practically, it provides a reliable framework for evaluating efficiency in operational settings and lays the foundation for a machine learning (ML)–aided decision-support framework (DSF) for energy efficiency optimisation in coal-fired boilers. • ANN predicts scientifically feasible and accurate boiler operating conditions • ANN achieved RMSE of 0.86 % and MAPE below 1.25 % for boiler efficiency • All efficiency predictions had less than 5 % error; 70 % of test data within 95 % CI • Mass and energy balance errors stayed within ±1.06 % and ±4.5 %, respectively • This work enables future ML-aided decision systems for in-time boiler optimisation