Enhancing Electricity Demand Prediction In Mexico: A Comparative Analysis Of Forecasting Models Using Conformal Prediction

Objectives: This study aims to identify a reliable prediction model for Mexico's wholesale electricity market to prevent supply-demand imbalances by analyzing three machine learning models with conformal prediction techniques. Theoretical Framework: Electricity demand forecasting traditionally uses SARIMAX models or artificial neural networks (ANN)/deep learning (DL). However, SARIMAX models are sensitive to data perturbations, while ANN/DL models face interpretability and computational challenges. Decision tree-based models (LGBM and CatBoost) have emerged as alternatives, offering potential advantages for time series prediction. Method: The study analyzed electricity demand data from CENACE (2018-2022), using four years for training and one for validation. Model performance was evaluated using MAPE, MAE, MSE, RMSE, and R² metrics, along with Conformal Prediction metrics for uncertainty quantification. Results and Discussion: CatBoost demonstrated superior performance (MAPE: 1.85%, R²: 0.9) compared to LGBM and SARIMAX models. SARIMAX showed significant limitations in prediction accuracy, particularly during seasonal variations. LGBM achieved intermediate performance with acceptable accuracy (R²: 0.89). Research Implications: The LGBM and CatBoost models achieved MAPE values within the literature's benchmark range (0.4-3%) while requiring fewer computational resources than complex DL models, making them practical alternatives for electricity demand forecasting. Originality/Value: This research pioneers the application of three prediction models with conformal prediction techniques in the Mexican electricity market context, offering probabilistic guarantees for prediction validity instead of traditional point predictions.