Intra-day Electricity Price Forecasting Based on a Time2Vec-LSTM Model

This paper presents a novel deep neural network architecture, combining stacked LSTM and Time2Vec layers, to predict electricity prices up to eight hours ahead-an essential input for future decision-making tools. We assess this model using hourly wholesale electricity price data from Colombia, comparing it to state-of-the-art time series and machine learning forecasting methods, including SARIMA, Holt-Winters, XGBoost, and MLP. The results demonstrate that our model excels in accurately modeling non-linearity and explicitly characterizing data behavior, yielding more precise price predictions. Particularly, the Time2Vec layer significantly aids in capturing temporal relationships between input and output data. This framework shows promise in enhancing the precision of electric price forecasts, offering valuable insights for the energy sector’s decision-making.