Prediction of Coal Properties Using Neural Networks: Novel Approach Based on Small Training Dataset

Owing to higher accuracy than traditional statistical methods, neural networks are widely being used to predict coal properties. Many researchers have reported that neural networks can be successfully applied to predict several important coal properties such as elemental composition, grindability and gross calorific value using easily obtainable properties, usually, proximate analysis of coal. However, building the neural networks model is not always possible with enough coal properties data. Despite significant impact of coal properties on power plant boiler operation, in many cases, coal properties are not measured regularly and their databases are very limited. On the other hand, as one of data-driven modelling, the predictive performance of the neural networks models is influenced not only by the quality but also by the quantity of the training datasets. Sufficient training dataset, or as much as possible sample set covering wide range of data is crucial to build the neural network based models. The purpose of this paper is to present data expansion technique, mega-trend-diffusion, for small measured training datasets to improve the accuracy of predictive models. Using this approach, additional datasets are generated covering the gaps between available data. The proposed approach is tested in a variety of benchmarks and the results show that the predictive performances of neural networks are significantly improved compared to original training dataset as indicated by reduced mean squared errors ranging from 8 - 80%. The use of this approach will provide significant benefit to not only model coal properties but also to develop other property estimations using small amount of experimental data or actual plant data.