Optimizing Renewable Energy Prediction Features

Summary & Methodology Analysis

Predicting wind and solar output often involves processing high-dimensional input data, which creates significant computational overhead when selecting the optimal feature set. The paper addresses this by proposing Cluster-based Sequential Feature Selection (CSFS), a wrapper-based approach that manages feature selection efficiency. Instead of evaluating every feature individually, CSFS groups features into clusters using techniques like correlation analysis, Random Forest feature importance (a method for ranking feature relevance by measuring how much they reduce uncertainty in tree-based models), or random selection. By performing this coarse-grained clustering first, the system can evaluate feature groups rather than individual variables, significantly lowering the number of iterations required to find an optimal set.

The algorithm integrates a non-inferiority test, which is a statistical process designed to determine if removing a feature or cluster causes an unacceptable decrease in model performance compared to a threshold. If a group of features is deemed non-inferior, it is pruned early. If the method fails to identify discardable clusters, it proceeds to a fine-grained analysis at the individual feature level. To ensure the process completes, the pipeline includes a conditional force-remove mechanism that eliminates the worst-performing feature as a final fallback. This approach was tested using several models, including MLP (Multi-Layer Perceptron, a type of feedforward neural network), LightGBM, and XGBoost, alongside Random Forest.

Evaluation was conducted on specific wind and solar power data, including the EDP and PVOD datasets, using ARW Version 3.9.1. While the method provides a 21 percent reduction in computational cost versus SFS, the study acknowledges specific limitations. The results are tied to data from a single wind turbine and one photovoltaic station, and the paper notes that outcomes may shift based on alternative hyperparameter search spaces, different random seeds, or different primary evaluation metrics. As such, the generalization of these computational gains to other hardware or scale environments is not explicitly covered by the facts provided.