Model of the Energy Production Process in Thermal Generation Plants Considering the Operation Profile

This work was developed to predict, by means of a model, the net energy production of a thermoelectric power plant. As a case study, data from the historical operating hours of the Quevedo Thermoelectric Power Plant were analyzed. As primary descriptors, input variables such as heavy fuel oil (HFO), diesel (DO) and lubricant (LO) were identified. Energy predictions were developed by applying multiple machine learning regressors (e. g., Gaussian processes, gradient boosting algorithms, random forests, support vectors, and neural networks). A total of 365 data sets were employed, for which, the data were first randomized and divided into training (80%) and testing (20%) portions. To avoid overfitting, a five-fold cross-validation was applied. A plant generation efficiency of 15.52% ± 3.76 was found with an availability index of about 84% describing normal plant operation. The HFO descriptor represents the main factor for predicting the net energy produced through a feature importance analysis. The obtained energy predictions, when compared, show 97.85% accuracy compared to the measured results. The algorithms showing the highest accuracy were multilayer neural networks, Gaussian processes and the linear support vector regressor.