AI-powered machine learning models for monitoring and optimization of biodrying process

Biodrying is a promising waste treatment technology that utilizes microbial heat generation to reduce moisture content, enhancing the production of refuse-derived fuel (RDF). However, the nonlinear and dynamic interactions between influencing factors governing biodrying remain challenging to predict using traditional statistical and kinetic models. This study integrates machine learning and deep learning techniques to develop predictive models to capture these complex relationships and improving biodrying efficiency. Experimental data were collected from four distinct waste types: fluff RDF, coarse RDF, fresh municipal solid waste, and market waste. The models were trained and tested using a comprehensive dataset comprising aeration rate, initial waste characteristics, daily gas emissions, mass changes, and hourly temperature profiles. Among all models tested, deep learning methods demonstrated superior accuracy, with Long Short-Term Memory (LSTM) achieving the highest predictive performance (R² = 0.988 training, R² = 0.987 testing). The feature importance analysis identified sequencing influence on the biodrying process by waste mass (IG = 19.563) and organic content (IG = 12.532) as the most influential parameters. Granger causality analysis further revealed oxygen concentration as the strongest causal predictor of leachate formation (ΔLoss = 158.9 × 10⁻⁴). A novel biodrying operating index (BOI) (total = 12.41) was introduced to determine its optimum condition. These findings establish LSTM as a reliable AI-driven tool for optimizing biodrying, with significant implications for sustainable waste management and RDF production.