Explainable Machine Learning Framework for Forecasting Household Organic Waste Generation to Support Sustainable Bioenergy Systems

The increase in the volume of organic household waste and the transition to a circular economy require the development of reliable forecasting tools capable of ensuring effective planning of waste recycling and bioenergy production systems. Despite significant advances in machine learning methods, most current research focuses on improving forecasting accuracy, while paying insufficient attention to model interpretation and the use of the results to support management decisions. The aim of this study is to develop and validate, using empirical municipal data, an explainable machine learning framework for forecasting the generation of organic household waste to support decision-making regarding the development of biogas and biomethane systems. The proposed framework combines data preprocessing, feature engineering, ensemble machine learning algorithms, prediction quality assessment, Explainable Artificial Intelligence (XAI) methods, and bioenergy potential assessment into a unified decision support system. The study was conducted using an empirical dataset comprising 10,800 observations for 20 local communities, obtained from the monitoring and accounting records of LKP “Green City” (Lviv, Ukraine) and covering the period from 1 January 2024 to 23 June 2025. The dataset includes demographic, socioeconomic, territorial, tourism-related, natural and climatic, infrastructural, logistical, organizational and economic, and temporal characteristics. To forecast the normalized index of organic household waste generation, we developed and compared Random Forest, Extra Trees, Gradient Boosting, XGBoost, LightGBM, and CatBoost models. A Linear Regression baseline fitted to the overlapping index-related predictors achieved MAE = 0.02339, RMSE = 0.02886, and R2 = 0.96412 on the chronological holdout set, indicating that a substantial part of the constructed target is linearly reconstructible. Among the six ensemble models, CatBoost yielded the most favorable point estimates, with MAE = 0.02933, RMSE = 0.03720, and R2 = 0.84958. Strict leave-one-municipality-out validation showed lower zero-shot spatial transferability, with Extra Trees achieving the highest pooled performance (MAE = 0.05198, RMSE = 0.06464, R2 = 0.58159). Extra Trees-based local adaptation substantially improved transfer to previously unseen communities, reaching pooled MAE = 0.03234, RMSE = 0.04121, and R2 = 0.83424 after 120 days of community-specific observations, although performance remained heterogeneous across municipalities. SHAP analysis identified tourism activity, air temperature, household income, the proportion of the urban population, weekends, and distance to water bodies as important contributors to the model predictions and revealed nonlinear model-specific associations for tourism activity, temperature, and income. Based on the predicted organic waste generation, a scenario-based assessment of the annual bioenergy potential of the investigated communities was performed, demonstrating substantial differences in potential electricity generation among communities. The scientific contribution of this study lies in the integration of ensemble forecasting, explainable artificial intelligence methods, and scenario-based bioenergy potential assessment within a single decision-support framework. The practical value of the proposed approach lies in its potential use for preliminary forecasting of organic waste resource availability and scenario-based assessment of bioenergy potential. Further operational, infrastructure, or investment applications require external validation using independent municipal datasets and verification under real operating conditions.