Comparison of Single-Stage and Two-Stage LSTM-Based Methods for Forecasting Electricity Demands and Prices

• Published in: 2022 IEEE PES Innovative Smart Grid Technologies Conference Europe (ISGT-Europe) pp. 1 - 5
• Main Authors: Zhu, S., Zou, M., Djokic, S. Z., Gunda, J., Papic, I.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.10.2022
• Subjects: Electricity demand and price; forecasting; Hybrid power systems; long short-term memory method; Power supplies; Predictive models; regression method; Smart grids; Time series analysis; Training; Uncertainty
• DOI: 10.1109/ISGT-Europe54678.2022.9960681

This paper compares performance of long short-term memory (LSTM) based methods for single-stage and two-stage forecasting of electricity demands and prices. A reference LSTM single-stage forecasting model was built first, which forecasts electricity demands or prices using their original historical time series. Afterwards, original demand and price time series are processed to obtain two new time series: a) time series of average values, and b) time series of differences from the average values. These are then used to formulate additional single-stage and two-stage forecasting models, where either forecasted average values (first stage) are combined with forecasted differences from average values (second stage), or regression-based and LSTM-based methods are combined in a hybrid single-stage forecasting method. Actual demand and price time series recorded in the UK power supply systems are used for evaluating all methods, and obtained results show improvement of overall forecasting performance of two-stage models over conventional single-stage model, but demonstrate that a hybrid single-stage model has the best performance. After analysing model uncertainty through multiple trainings, Bayesian Optimisation is adopted to reduce the uncertainty.