Determination of maceral distributions using proximate and ultimate analysis of some Turkish lignites

Coal is one of the most significant energy resources worldwide. In our country, lignite coal reserves are known to hold particular importance. Research indicates that fossil fuel reserves of oil and natural gas will deplete sooner than coal on a global scale. Despite its critical role, coal characterization involves costly and expertise-demanding processes. Therefore, employing artificial intelligence-based methods to streamline analysis processes and achieve high-accuracy results at lower costs is of great importance. This study aims to introduce artificial neural networks as an alternative method to laboratory analyses. The research focuses on determining the maceral distributions of Turkish coals based on proximate and ultimate analysis values using artificial neural networks. Identifying maceral distributions plays a crucial role in assessing coal quality parameters and depositional environment characteristics. In this context, the study examines the physical and chemical properties of certain Turkish lignite coals, utilizing proximate and ultimate analysis data available in the literature. The obtained data were analyzed using artificial neural networks and optimization techniques. A total of 232 lignite coal data points were used to train the artificial neural network. These data were processed using MATLAB software for feedforward multilayer perceptron and cascade forward neural networks, while Bayesian hyperparameter optimization techniques were implemented in Python. The predictive performance of artificial neural networks was evaluated through regression analysis, error metrics, and other statistical methods. The data were tested with different numbers of hidden neurons, and the optimal neuron layer configuration yielding the best results was selected. As a result, the feedforward multilayer perceptron demonstrated superior performance for complex datasets, whereas the cascade forward neural network was more effective in cases requiring simpler architectures with fewer layers. Bayesian optimization, on the other hand, proved to be efficient in hyperparameter tuning, even with limited data, by yielding more precise results. Among the three algorithms, Bayesian optimization produced the most accurate predictions, followed by the feedforward multilayer perceptron. To maximize efficiency, increasing the dataset size when using literature-based data is suggested as a means to enhance accuracy. This approach ensures that both input and predicted values align closely with the "x=y" reference line, thereby minimizing errors to an optimal level.