Comparative Analysis of Filter, Wrapper, and Embedded Feature Selection Methods to Predict Load Progression in Combined Cycle Power Plant (CCPP) Startup Across Diverse Startup Classes

In the realm of big data, the utilization of advanced analytics techniques, particularly artificial intelligence (AI), holds paramount importance in extracting valuable insights from data generated by combined cycle power plants (CCPPs). This study meticulously compares three sophisticated feature selection methods: Pearson Correlation Coefficient (PCC) filtering, Random Forest-based Recursive Feature Elimination (RF-RFE) wrapper-based selection, and LASSO regularization for embedded selection, aimed at predicting load progression (Generator Watts) during startup using LSTM models. Leveraging a comprehensive dataset consisting of 240 startup processes, each comprising 14,000 observations and categorized into four startup classes (HOT, WARM 1, WARM 2, COLD), this research carefully selects features for each class independently through the aforementioned techniques. Subsequently, distinct LSTM models are constructed for each startup class based on the selected features, and their predictive capabilities are rigorously evaluated using performance metrics such as Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Squared Error (MSE). The comparison highlights variations in the performance of feature selection techniques across different startup classes. Specifically, RF-RFE demonstrates superior efficacy in the Hot and Warm 1 classes, while LASSO regularization outperforms other methods in the Warm 2 and Cold classes. This study offers valuable insights into the importance of experimentation and empirical validation in selecting the optimal feature selection technique for predictive modeling tasks. While some research suggests that certain methods may outperform others, this study emphasizes the importance of tailoring feature selection approaches to dataset characteristics, guiding practitioners and researchers toward more accurate predictive models and improved real-world outcomes.