UCLA

This study reviews machine learning based time series forecasting approaches and applies them to the day-ahead electricity price forecasting problem. We review the literature on transformer models for time series forecasting, in addition to two MLP-based methods, NBEATSx and NHiTS, and XGBoost. Four models are chosen and implemented: NBEATSx, NHiTS, Temporal Fusion Transformer, XGBoost. They are tested on electricity price data from the PJM market. We product prediction intervals using adaptive conformal inference. We find that NBEATSx and NHiTS perform best on this dataset. We find that by averaging as few as two models, we are able to significantly improve prediction accuracy.