Forecasting Spain's electricity load: a comparative analysis of classical time series, neural networks, and deep learning models.

Abstract

The objective of this research is to implement and analyze various forecasting models, including statistical, machine learning, and deep learning approaches, to forecast the electricity load in Spain from 2015 to 2020. Utilizing time-series data sourced from the Open Power System Data (OPSD) project, the study leverages the historical forecast feature of the Darts library, highlighting its retrain functionality for enhanced accuracy. The study involved setting up and testing different models in four distinct configurations to investigate the role of past covariates and encoders on the accuracy of the forecasts. The research focused on the MSTL and AutoARIMA models for statistical analysis, while exploring the capabilities of XGBoost, LightGBM, and RandomForest models in the machine learning segment. In the realm of deep learning, NHiTS and NBEATS models were used. A detailed evaluation process was carried out, mainly using the RMSE metric to assess the performance of the various models. The results showed that deep learning models performed the best, followed by certain machine learning models, especially the LightGBM, and then the AutoARIMA model. Notably, the non-retrained versions of the models performed better than their retrained counterparts, showcasing a subtle trend in model performance. The effectiveness of including past covariates and encoders varied greatly, depending on the specific model being analyzed. The study highlights the complex nature of electricity load forecasting and emv phasizes the need for sophisticated methods in selecting and setting up models. It also suggests potential directions for future research, focusing on a deeper understanding of the retraining process, incorporating different covariates and encoders, and exploring the effects of hyperparameter tuning on model performance.