Improving short-term active power prediction through optimization of the categorical boosting model with meta-heuristic algorithms

Anticipating short-term active power is widely recognized as a complex nonlinear issue due to the intricate nature of the underlying processes and variables involved in the prediction task. Addressing this complexity requires the application of advanced algorithms and optimization techniques to boost the precision and efficiency of the predictive frameworks. The presented study employs the categorical boosting (CatBoost) algorithm as a foundational model for estimating considerably short-term active power. To optimize the hyperparameters of CatBoost, various meta-heuristic algorithms are utilized during network training for optimal performance. The K-fold cross-validation approach is utilized to mitigate overfitting during the training process. The efficacy of the recommended hybrid frameworks is examined through a case study utilizing actual data. Ultimately, the performance of all algorithms is compared using six statistical investigation indexes. The outcomes reveal that the combination of CatBoost and the arithmetic optimization algorithm achieves the best values for R², NMSE, RMSE, MAE, and RAE, approximately at 0.999, 0.0028, 13.5490, 7.2820, and 0.0650, respectively. This combination also exhibits the best convergence of about 118.997 in 500 iterations. Additionally, employing meta-heuristic algorithms improves short-term active power forecasting accuracy, improving R² and RMSE indexes by approximately 1.11% and 15.04%, respectively.