Optimization of Leachate Recirculation for Enhanced Landfill Energy Recovery: A Bayesian Framework with Neural Network Surrogates

This study presents a Bayesian neural network (BNN)-based optimization framework for leachate injection operations in municipal solid waste (MSW) landfills aimed at maximizing biogas and heat energy recovery. While landfill gas recovery for energy production is a well-established practice, recovery of the substantial heat generated during organic waste degradation remains largely untapped. However, the methane (CH4) generation and temperature evolution within the landfills is dependent on the underlying complex coupled processes, and predicting these trends to maximize energy recovery from landfills is challenging. Validated coupled-process numerical models provide generalizable predictions and can be useful for analyzing complex landfill dynamics and evaluating the long-term performance of landfills. However, their computational expense limits their applicability in broader optimization applications involving numerous large-scale simulations. To address this, we use a data-driven approach where BNN surrogate models are trained on data generated using the coupled thermo-hydro-bio-mechanical (CTHBM) model to alleviate the associated computational expense. BNNs also quantify the epistemic uncertainty associated with surrogate model predictions, enhancing the reliability of predictions. Two separate BNNs were trained on CTHBM model data to predict the cumulative CH4 generation and maximum temperature evolution in the landfill. The BNN surrogate models depicted excellent accuracy in predicting the CTHBM model data (R2 ∼ 0.99 in most test cases). Further, partial dependence plots depicting the relationship between the input variables and surrogate model predictions showed that surrogate models effectively captured the underlying process relationships. Finally, an optimization framework was developed to leverage predictions from the BNNs for uncertainty-aware decision-making in optimizing leachate injection operations for maximum energy recovery. The effectiveness of this framework in identifying optimal leachate injection parameters through exhaustive search is demonstrated through multiple example case studies. The identified injection parameters using the proposed framework showed potential for improved energy recovery compared to both no-injection and continuous injection scenarios while maintaining safe landfill temperatures (<65 °C). Overall, the findings of this study underscore the potential of Bayesian machine learning-based surrogate modeling in optimizing landfill performance for enhanced energy recovery.