A Novel Approach to Electricity Price Forecasting in Market Economics: Three-stage Machine Learning Ensemble Utilizing Feed Forward Neural Network, XGBoost, and LightGBM
Lake Zhou
ABSTRACT
Electricity Price Forecasting (EPF) is a critical and complex task, necessary for many electrical market participants. This paper proposes a novel approach to EPF utilizing a three-stage machine learning model, FXL3, including base learners, a stacking model, and a final residual correcting model. Using the EPF Toolbox Python library, we benchmark our approach against ten state-of-the-art EPF models, showing a significant improvement in accuracy. Then, to evaluate the efficiency of the proposed model, we conduct a series of variation experiments. Furthermore, in conjunction with the pro- posed model, this paper also introduces a novel and comprehensive dataset that includes five exogenous factors: hourly temperature, wind speed, natural gas price, shift factor, and location. Finally, using experimental results, this paper demonstrates the effectiveness of the three-stage approach and the high correlation between Shift Factor and electrical price, which has been commonly neglected in EPF research.
