Power Load Forecasting: A Time-Series Multi-Step Ahead and Multi-Model Analysis

Distribution System Operators and Aggregators can derive benefits from innovative approaches in Power or Energy Load Forecasting (PLF-ELF). Enhanced accuracy in PLF-ELF can support the management of energy imbalances between production and consumption or operations like Demand Response. This research aims to assess a wide range of models as potential solutions for multi-step PLF-ELF, utilizing time-series machine learning models for power consumption. The experimentation encompassed multi-step PLF-ELF across various resolutions, ranging from 15-minute intervals up to one day. Results indicate that while Long Short-Term Memory recurrent Neural Networks and Lasso Regression with Cross-Validation perform better for very short-term PLF-ELF, tree-based regressors like Gradient Boosting, Histogram Gradient Boosting, Light Gradient Boosting Machine and CatBoost outperformed other models in the upcoming steps. The optimal PLF-ELF technique proposes different algorithms per forecasting horizon.