Temporal Decomposition Transformer for Probabilistic Energy Forecasting

To ensure the balance of power supply and demand, probabilistic energy forecasting is significant to determine the power generation and dispatch strategies. In order to solve the probabilistic forecasting problem of energy time series, we develop a novel transformer-based decomposition framework, i.e., the temporal decomposition transformer (TDT) to estimate the probability distribution of future time series. TDT achieves accurate and reliable probabilistic forecasting by predicting the mean and standard deviation of time series successively through the decomposition framework. TDT uses the tranformer decoder to capture the temporal feature in the historical time series, and then two tranformer decoders predict the mean and standard deviation of the time series at each future time instant respectively, where the standard deviation is forecasted based on the forecasting of the mean. Finally, the time series probabilistic distribution is modeled by log-likelihood regression. By accomplishing extensive experiments on two real-world energy time series datasets, we conclude that TDT achieves better forecasting accuracy in both point forecasting and probability energy forecasting than compared methods.