Analysis of feature selection techniques for prediction of boiler efficiency in case of coal based power plant using real time data

Monitoring and enforcing the performance of equipment in coal-based thermal power plants play a vital role in operational management. As the coal-based power plant is a nonlinear system involving multiple inputs and multiple outputs, the standard and typical identification methods tend to deviate. This can happen due to factors such as strong coupling, multivariable characteristics, time delays, and noise contamination of onsite data. Due to the availability of distributed control system and advances in data collection technology, huge amounts of data is available for analysis. Past works in this area have underutilized the historical data collected from the physical analysis of power plants. This work has analyzed this data using Feature Selection and Machine Learning modeling techniques to derive meaningful variables and make accurate predictions using them. These predictions can help power plant operators improve boiler efficiency and reduce toxic gas emissions from their plants. This paper presents approaches using empirical modeling based on classification algorithms along with feature selection methods (Pearson Correlation, Random Forest) that are used to predict boiler efficiency. In the present research work, the XGBoost algorithm outperforms Gradient Boosting. In addition to that XGBoost algorithm is trained with 4, 8, and 13 features resulted in an accuracy of 91.98%, 91.13%, and 91.3%. The outcome of our research provides recommendations to the power plant operator for improving the boiler efficiency and hence lowering the quantity of toxic flue gases emitted into the atmosphere.