HFNet: Forecasting Real-Time Electricity Price via Novel GRU Architectures

Abstract

Electricity price forecasting is critical to numerous tasks in the power system such as strategic bidding, generation scheduling, optimal scheduling of storage reserves and system analysis. Most existing price forecasting models focus on hourly prediction for the day ahead market. This work focuses on the real-time, 5-minute market, with the goal of developing a model able to capture both the long- and short-term temporal distribution of the data. Extending the recent advances in deep learning models of time series forecasting, the proposed model - named HFnet - is a novel multi-branch Gated Recurrent Unit (GRU) architecture for electricity price forecasting. Extensive empirical analyses using real-time data from the New York Independent System Operator (NYISO) illustrate the value of the proposed model when compared to state-of-art prediction models, with an average reduction in error of 10%.