Predicting Coal Stock Levels In India’s Thermal Plants: An Ensemble Machine Learning Approach And Policy Implications

Abstract

Accurate forecasting of coal stock levels at thermal power plants is crucial for energy security and supply planning. This study uses daily coal stock data (2018–2023) from India’s Ministry of Power to develop models for the prediction of next-year stock levels. We have used an ensemble of machine learning models—including Linear Regression, Support Vector Regression (SVR), XGBoost, and Neural Networks—and combined them in a stacking framework. The models are evaluated on regression metrics (R², MAE) and a derived classification task (flagging critical stock conditions). The stacked ensemble gives output as an R² of ~0.73, which significantly outperforms individual models, which achieve a value between 0.29–0.70 for the coefficient of determination. Figures summarize the pipeline and model performance. We discuss how such forecasts can inform policy: enabling planners to avoid shortages, optimise coal dispatch and maintain normal stock levels. Our research work provides a suggestive strategy that data-driven stock predictions can enhance energy resilience by guiding procurement strategies, coordinating rail and mining schedules, and mitigating supply-chain risks.