Hybrid Approaches for Steam Demand Forecasting: Combining First Principles, Box and Jenkins, and Neural Network Models

The pulp and paper industry relies heavily on batch sulphite digesters for chemical cellulose production, where steam is a critical utility – lack of steam prediction results in venting losses, increased costs, and a negative environmental impact. Accurate prediction of steam demand is therefore essential for optimising digester cooking cycles and resource allocation. This study aims to develop and compare predictive models for steam demand in batch pulp digesters using magnesium bisulphite cooking liquor. Three years of production data were pre-processed to extract digester temperature profiles and batch steam demands. Seven modelling approaches were evaluated: a mechanistic first-principles energy balance model, Box–Jenkins ARIMA, two neural network models (LSTM and CNN), and three hybrid models combining first-principles with ARIMA, LSTM, and CNN. The hybrid frameworks employed dimensionless parameters from the mechanistic model as exogenous variables to compensate for unavailable process data. Model accuracy was assessed using RMSE and MAE metrics. The results show that hybrid models consistently outperformed their standalone counterparts. In particular, the hybrid first-principles–CNN model achieved the highest predictive accuracy, demonstrating the CNN’s ability to extract features and capture nonlinear temporal dependencies in steam demand. The hybrid first-principles–ARIMA model also surpassed both the standalone ARIMA and mechanistic models. Integrating mechanistic insights with data-driven methods significantly enhances prediction accuracy in complex batch processes. The findings highlight the value of hybrid modelling strategies for improving steam demand forecasting, with potential benefits for process optimisation, energy efficiency, and batch scheduling in the pulp and paper industry. Copyright © 2025 by Authors, Published by Universitas Diponegoro and BCREC Publishing Group. This is an open access article under the CC BY-SA License (https://creativecommons.org/licenses/by-sa/4.0).