Development of evolutionarily optimized random forest models to accurately estimate coke strength after reaction

Accurate prediction of coke strength after reaction (CSR) is essential for optimizing coke quality in the metallurgical industry, thereby minimizing production costs and maximizing resource utilization. In this research, the hyperparameters of the powerful random forest machine learning approach is tuned distinctively via nine evolutionary optimization methodologies namely particle swarm optimization, genetic algorithm, grey wolf optimization, whale optimization, ant colony optimization, cuckoo search algorithm, artificial bee colony, dragonfly algorithm and memetic algorithm utilizing a dataset collected from a coke plant to forecast CSR. The K-fold cross-validation technique is utilized in the training phase to mitigate the overfitting problem. The efficiency of each algorithm is evaluated using optimization time and standard performance metrics. The results show that moisture content, ash content, plastic layer thickness, mean maximum reflectance are all directly correlated with CSR while it is inversely related with volatile matter content, sulfur content, maximum fluidity and basicity index. It was observed that particle swarm optimization and dragonfly algorithms are the best algorithms, leading to the most accurate random forest predictive models as demonstrated by evaluation metrics of average absolute relative error, R-squared, and mean square error. Also, the results indicated that memetic algorithm needs the longest time while particle swarm optimization algorithm requires the shortest optimization time when processing the training phase.