Probabilistic electricity price forecasting based on penalized temporal fusion transformer

Abstract

In the deregulated electricity market, it is increasingly important to accurately predict the fluctuating, nonlinear, and high-frequent electricity price for market decision-making. However, the uncertainties associated with electricity prices, such as non-stationarity, nonlinearity, and high volatility, pose critical difficulties for electricity price forecasting (EPF). Unlike point forecasting, which provides only a single, deterministic estimate of future prices, probabilistic forecasting gives a more comprehensive and nuanced picture of future price dynamics, which can help market participants make better-informed decisions when facing uncertainty. Therefore, in this paper, we propose a robust deep learning method for multi-step probabilistic forecasting. First, we use the least absolute shrinkage and selection operator (LASSO) in the expert model to generate point forecasts. Second, we introduce the smoothly clipped absolute deviation regularization term, a nonconvex penalty with proven oracle properties in model selection, into temporal fusion transformers. Finally, we employ the proposed model to integrate point forecasts to give probabilistic forecasts. To evaluate the proposed forecasting model, real-data experiments are conducted in the Nord Pool electricity market and the Polish Power Exchange market. Empirical results show that the proposed model has demonstrated superior probabilistic forecasting performances compared with other competitors and has proven its effectiveness in real-world applications.