Optimizing biomass combustion and mitigating potassium induced slag formation with deep learning
Udomsak Kaewsiri, Supachate Innet
Abstract
Biomass power plants that rely on potassium-rich agricultural waste experience operational difficulties due to the formation of slag on the walls of the furnace and heat exchangers. This causes thermal efficiency to drop, increases the cost to fix problems and leaves the plant unavailable for longer periods. The main objective is to improve combustion stability and efficiency by studying slag formation and implementing control strategies that use deep learning. The combustion happens in a sealed, adiabatic chamber along with a Waste Heat Recovery Boiler that uses rubberwood fuel with a high moisture and potassium concentration. For better efficiency and lower emissions, combustion is managed by separating it into drying, ignition and burnout zones. Rather than burning directly inside the boiler, flue gases transfer heat more gently which increases component life. This study applies Convolutional Neural Networks (CNNs) for image processing analysis of potassium content during combustion processes. A pre-trained model and a custom CNN was used for analyzing a dataset consisting of 30 images through transfer learning procedures. GoogleNet produced 96.44% accuracy but its enhanced variation obtained greater than 98.74% accuracy levels. The ResNet-50 model achieved 87.19% accuracy before augmentation but attained 88.32% accuracy afterward and the Custom CNN demonstrated 79.62% accuracy. The research demonstrates that augmenting data improves pretrained model output whereby GoogleNet delivers superior results than every other model. The combustion control system uses Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTMs) for deep learning because these techniques optimize operational performance.The developed combustion controller integrated RNN and LSTM connections to read potassium levels while adjusting wood-to-root ratios in order to prevent slagging and corrosion and fouling. The integration of Augmentation with GoogleNet served as the best solution by combining functionality with the real-time LSTM system to monitor parameters and achieve better stability and emissions performance.