A robust electricity price forecasting framework based on heteroscedastic temporal Convolutional Network

• Published in: International journal of electrical power & energy systems Vol. 161; p. 110177
• Main Authors: Shi, Wei, Feng Wang, Yu
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.10.2024; Elsevier
• Subjects: Deep learning; Electricity price forecasting; Encoder-decoder framework; Feature selection; Maximum Likelihood Estimation; Temporal Convolutional Network; Deep learning; Electricity price forecasting; Feature selection; Temporal Convolutional Network; Encoder-decoder framework; Maximum Likelihood Estimation
• ISSN: 0142-0615
• DOI: 10.1016/j.ijepes.2024.110177

•Introduces the Heteroscedastic Temporal Convolutional Network for day-ahead electricity price forecasting.•Employs a heteroscedastic output layer to represent variable uncertainty.•Utilizes a maximum likelihood estimation-based loss function to handle heteroscedasticity.•Integrates a multi-view feature selection algorithm to improve forecast precision.•Demonstrates state-of-the-art performance on multiple electricity datasets. Electricity price forecasting (EPF) is a complex task due to market volatility and nonlinearity, which cause rapid and unpredictable fluctuations and introduce heteroscedasticity in forecasting. These factors result in varying prediction errors over time, making it difficult for models to capture stable patterns and leading to poor performance. This study introduces the Heteroscedastic Temporal Convolutional Network (HeTCN), a novel Encoder-Decoder framework designed for day-ahead EPF. HeTCN utilizes a Temporal Convolutional Network (TCN) to capture long-term dependencies and cyclical patterns in electricity prices. A key innovation is the heteroscedastic output layer, which directly represents variable uncertainty, enhancing performance under fluctuating market conditions. Additionally, a multi-view feature selection algorithm identifies crucial factors for specific periods, improving forecast precision. The framework employs an improved loss function based on maximum likelihood estimation (MLE), which adjusts for the heteroscedastic nature of electricity prices by predicting both the mean and variance of the price distribution. This approach mitigates the impact of extreme price spikes and reduces overfitting, resulting in robust and reliable predictions. Comprehensive evaluations demonstrate HeTCN’s superiority over existing solutions such as DeepAR and the Temporal Fusion Transformer (TFT), with average improvements of 25.3%, 24.9%, and 17.4% in the mean absolute error (MAE), symmetric mean absolute percentage error (sMAPE), and the root of mean squared error (RMSE) compared to DeepAR, and 17.6%, 14.4%, and 13.6% relative to TFT across five distinct electricity markets. These results underscore HeTCN’s effectiveness in managing volatility and heteroscedasticity, marking a significant advancement in electricity price forecasting.