Pattern Classification-based Electricity Price Interval Forecasting with Discrete-Continuous Inputs

Abstract

The worldwide restructuring of electricity markets toward deregulation over the past few decades has attracted significant research attention to electricity price forecasting due to its importance for market participants. In recent years, market participants are more concerned with forecasting the general price level than the exact price value, as this better supports bidding strategies and investment decision making. However, current interval forecasting approaches face key challenges in efficiency, robustness against volatility, and ineffective periodicity extraction technique. To address the limitations of existing electricity market price forecasting methods and to develop more practical strategies for real-world market conditions, we propose a novel model that combines TimesBlock —a state-of-the-art frequency-domain method capable of extracting multi-periodic patterns from data— with LSTM, characterized by strength in capturing long-term temporal dependencies. This integration enables accurate pattern classification-based interval forecasting of electricity prices. A discrete-continuous hybrid data input strategy is innovatively designed to enhance the model’s ability to handle volatile data while preserving essential periodic information. The experimental results demonstrate that our model provides reliable electricity price interval forecasts and outperforms other deep learning and machine learning models on multiple metrics. Additionally, the Discrete-Continuous hybrid data input improves computational efficiency, achieving several times faster running speed compared to traditional continuous-input only methods.