Torrefaction is a promising thermochemical pretreatment technology. It enhances the fuel quality, energy density, and storage characteristics of lignocellulosic biomass, thus improving its suitability for sustainable bioenergy applications. However, the complex interactions among biomass properties and torrefaction operating conditions make accurate prediction and optimization of torrefaction performance challenging. In this research, an explainable artificial intelligence (XAI)-based framework was developed to predict, interpret, and optimize torrefaction performance. A comprehensive database collected from published literature was used to train and evaluate three machine learning models: Random Forest (RF), Extreme Gradient Boosting (XGB), and CatBoost (CAT). Among them, the CatBoost model achieved the best performance with an R2 of 0.861, RMSE of 0.049, MAE of 0.036, and MAPE of 2.97%. The XAI analysis identified HHV, FC, and VM as key variables influencing mass yield, while sulfur, ash, and FC strongly affected energy yield. HHV and FC were the main positive contributors to energy density factor, providing interpretable insights into biomass retention, energy recovery, and fuel upgrading. The optimized CatBoost model results predicted a maximum energy yield of 75.67%, with a mass yield of 68.87% and an energy density factor of 1.0988. The predicted optimum condition (240 °C and 10 min) was experimentally validated using Prosopis juliflora wood waste, showing good agreement between predicted and experimental torrefaction performance. Overall, the developed explainable machine learning framework integrates prediction, interpretation, optimization, and experimental verification, providing an effective approach for intelligent upgrading of lignocellulosic biomass into high-quality solid biofuels.