Real time electricity price time series forecasting models based on deep learning

Yang, Haolin
Thumbnail Image
Other Contributors
Willemain, Thomas, TRW
Wang, Meng, MW
Schell, Kristen, KRS
Pazour, Jennifer, JAP
Issue Date
Decision sciences and engineering systems
Terms of Use
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute (RPI), Troy, NY. Copyright of original work retained by author.
Full Citation
Due to its nature of high-noise, high-nonlinearity and high-uncertainty (3H), predicting and modeling the real-time, spot electricity price is of the utmost difficulty. Poor forecasting brings risks and challenges for multiple power system tasks such as strategic bidding and generation rescheduling. Failure to predict price spikes due to climate change events will result in huge social and property loss, on the scale that was witnessed during the 2021 Texas power outage caused by extreme cold weather and the 2019 California power curtailments caused by wildfires. Recently, deep learning (DL) has been actively attracting researchers' attention as a potential solution to the problems posed by 3H. However, there are still huge research gaps. Specifically, most of current existing models are focused on hourly or larger time intervals and only a few models are based on high-frequency (less than 5 minutes) forecasting. Moreover, traditional prediction models cannot effectively model the temporal dependencies among the 3H time series. For example, the spikes caused by commercial forecast loss or looped power flow are a big challenge to most forecasters. This work focus on the development of electricity price forecasting via deep learning based methods. Firstly, a novel multi-branch Gated Recurrent Unit (GRU) architecture (HFnet) is developed for the real-time prediction task, which uses date index features. The strong performance of this model is the basis for several following model development efforts. Secondly, a new multi-branch model with a novel parallel convolution neural network (CNN) and time series statistical features has been developed to enhance the forecasting accuracy (GHTnet). Thirdly, a temporal transfer learning model (GRU-TL) is proposed to measure feature-reuse among hybrid subzone datasets, which contributes to performance improvement and model robustness. Fourthly, a CNN-based Autoencoder (QCAE) is designed to capture features expressed both temporally and spatially. Finally, an explainable model with attention mechanism (ATTnet) is proposed to improve prediction quality and interpretability in real-time electricity price forecasting.
School of Engineering
Dept. of Industrial and Systems Engineering
Rensselaer Polytechnic Institute, Troy, NY
Rensselaer Theses and Dissertations Online Collection
Restricted to current Rensselaer faculty, staff and students in accordance with the Rensselaer Standard license. Access inquiries may be directed to the Rensselaer Libraries.