Dynamic XGBoost-based Quantile Predictor for Real-time Electricity Price Forecasting

Electricity price forecasting helps traders and market participants manage risks by determining the best strategy for bidding. Given the strong correlation between electricity prices and load consumption, there has been a growing interest in developing frameworks to forecast both electricity price and load. This work proposes a real-time electricity price forecasting model built upon an existing real-time XGBoost-based electricity load forecasting framework. Our proposed model comprises three key components: Feature Selector, XGBoost-based Quantile Predictor, and Range Controller. To train and test the prediction model, the Feature Selector preprocesses and selects the appropriate features. A machine learning enabled quantile predictor is then applied to obtain one step ahead predictions (OSAP) of electricity prices. Lastly, the Range Controller offers a training-efficient approach to adjust the prediction interval and quantile values, as well as to classify spike and normal cases using an unsupervised classification method. Upon analysis using R
²
, RMSE and MAE, it has been consistently demonstrated that our adaptive forecasting model exhibits more than two times improvement over a year.