Machine Learning Surrogated Coal Blend Optimisation for Inventory-Aware Coking

The disruptions triggered by the SARS COVID-19 pandemic, followed by the Russia-Ukraine conflict and American sanctions resulted in reduced accessibility of Russian coal to Indian steelmakers. This decline in availability forced producers to undertake extensive trials and experiments hunting viable alternatives. This study argues that optimized blend compositions formulated exclusively from the available coal inventory can effectively address such disruptions thereby ensuring production and scheduling remain independent of external supply fluctuations. To achieve this, multiple machine learning models including multivariate regression, decision trees, partial least squares regression, random forests and neural networks are developed from data collected from an integrated steel plant to predict coke quality from blend data. The most accurate model is adopted as a surrogate objective function for a genetic algorithm that reallocates blend proportions under inventory constraints. Results demonstrate that coke quality can be maintained without introducing new coal sources, enabling resilient, data-driven adaptation to uncertain supply chains.