Novel Data-Driven Models Applied to Short-Term Electric Load Forecasting

This work brings together and applies a large representation of the most novel forecasting techniques, with origins applications in other fields, to short-term electric load problem. We present comparison study between different classic machine learning deep techniques recent methods for data-driven analysis dynamical models (dynamic mode decomposition) ensemble applied forecasting. explores influence critical parameters when performing time-series forecasting, such as rolling window length, k-step ahead forecast number/nature features used characterize information predictors. The architectures considered include 1D/2D convolutional recurrent neural networks their combination, Seq2seq without attention mechanisms, based on gradient boosting principles. Three groups stand out from rest according scenario: (a) average results, (b) simple linear regression very forecasts, (c) combinations convolutional/recurrent longer-term forecasts.