Unlocking Energy from Waste with Machine Learning

Municipal solid waste (MSW) holds significant untapped thermal potential. Traditional calorimetric measurements, while accurate, are costly, time-consuming and limited by sample heterogeneity. Researchers have turned to machine learning (ML) to predict the heating value (HV) of MSW using readily available composition data. A recent study published in Scientific Reports evaluates multiple ML algorithms—Multiple Linear Regression (MLR), XGBoost, CatBoost and Extra Trees—to model HV based on dry sample weight and ultimate analysis (carbon, hydrogen, oxygen, nitrogen, sulfur, ash).

Data Collection and Preprocessing

The study draws on data from 24 counties in Iran’s Khorasan Razavi Province, as reported by the Mashhad municipality solid waste management organization. For each county, waste composition was documented in spring and summer, covering categories like food waste, plastics, paper, metals and textiles. Ultimate analysis provided elemental weights. Seasonal averages produced a final dataset. Heating values were calculated via the modified Dulong formula:

Heating Value (kJ/kg) = 337C + 1420(H + O/8) + 93S + 23N

where C, H, O, S and N are percentages by weight. The dataset was split 80% for training and 20% for testing.

Machine Learning Models and Hyperparameter Tuning

• Multiple Linear Regression (MLR): Traditional baseline, yielded R²_test=0.709 and MSE_test=941,272.5 kJ²/kg².
• XGBoost: Tuned with learning_rate=0.1, max_depth=3, gamma=0.001, achieved R²_test=0.975 and MSE_test=148,696.9 kJ²/kg².
• CatBoost: Tuned with learning_rate=0.1, max_depth=5, iterations=150, attained R²_test=0.951, MSE_test=547,119.4 kJ²/kg².
• Extra Trees: Tuned with n_estimators=300, max_depth=10, max_features=3, delivered top performance with R²_test=0.979, MSE_test=77,455.9 kJ²/kg² and MAE_test=245.9 kJ/kg.

Key Findings and Feature Importance

The ensemble methods XGBoost and Extra Trees significantly outperformed MLR, showcasing the advantage of non-linear, tree-based approaches. Extra Trees stood out for its balance of accuracy and robustness, minimizing both bias and variance. Feature importance analysis revealed that nitrogen content contributed 27.5%, sulfur content 26%, ash 15%, and dry sample weight 10% toward predictive power. This insight can inform data collection priorities and process controls in waste-to-energy facilities.

Implications for Waste-to-Energy Planning

Accurate, rapid predictions of MSW heating values enable better design and operation of thermal conversion systems such as combustion, gasification and pyrolysis. By integrating the Extra Trees model into resource management workflows, municipalities can cut calibration costs, reduce sampling frequency and optimize feedstock mixing in real time. This data-driven approach supports sustainable energy recovery, lowers greenhouse gas emissions and advances circular economy goals.

Conclusion

This research demonstrates that tree-based ML models, particularly Extra Trees, can effectively predict the heating value of diverse MSW streams using elemental composition data. With R²_test near 0.98 and low error metrics, practitioners now have a proven method to replace or supplement laborious calorimetric assays. Future work should expand datasets across regions and integrate sensor networks for real-time waste characterization, further boosting the impact of AI in sustainable energy management.

Key points

• Extra Trees model achieved R²_test=0.979 and MSE=77,455.92 for heating value prediction.
• Machine learning outperformed multiple linear regression, with ensemble methods showing highest accuracy.
• Nitrogen and sulfur contents emerged as the most influential features for energy forecasting.