Deep learning applications for biomass gasifier monitoring: A dual approach with direct and recursive long short-term memory forecasting

Continuous monitoring of biomass gasifier outputs is vital for maximizing system potential. While it is conventionally achieved via experimental equipment, increasing attention has been directed toward machine learning. Most available publications focus on steady-state predictions, however, this study addresses the inherent limitations of such models through time-series forecasting. In the first part, three Long Short-Term Memory (LSTM) architectures were evaluated for direct one-step ahead forecasting: a single-layer LSTM, a double-layer LSTM, and a hybrid bi-directional/one-directional LSTM. The single-layer LSTM outperformed more complex architectures, achieving average Root Mean Squared Error (RMSE) values of 0.038 for H2 and 0.034 for the High Heating Value (HHV). These results demonstrate that high accuracy is attainable with lower computational costs. In the second part, recursive multi-step forecasting using an encoder-decoder LSTM was employed across three horizons: 5-step, 10-step, and 30-step ahead. While the results satisfactorily conform to experimental trends, increasing the horizon from 5 to 30 steps amplified the RMSE for H2 from 0.085 to 0.18. Unbiased validation was ensured through an experiment-fold approach. The one-step model is identified as effective for real-time monitoring during sensor interruptions, while the multi-step model functions as a comprehensive digital twin. This dual methodology provides a robust framework for dynamic modelling, facilitating improved operational efficiency and process optimization.