Identifying price index classes for electricity consumers via dynamic gradient boosting

Abstract

Electricity retailers buy electricity at spot prices and resell energy to their customers at fixed retail prices. However, the electricity market is complex with highly volatile spot prices, and high price events might happen during peak time periods when energy demand significantly increases, leading to the decision of the retail price a challenging task. Understanding consumer price index, a price indicator that is associated with electricity consumption of customers helps energy retailers make critical decisions on pricing strategy. In this work, we apply dynamic gradient boosting model, namely CatBoost, to classify customers into different groups according to their price indices. To benchmark our results, we compare the performance of CatBoost with other baselines, including Random Forest, AdaBoost, XGBoost, and LightGBM. Our experimental results proved that CatBoost outperformed other algorithms due to its effective overfitting detector and categorical encoding techniques. Besides, the area under the curve of the Receiver Operating Characteristics (ROC), often known as AUC, is used as a standard measure metric to evaluate and compare between classifiers. Hence, CatBoost gained the lowest difference score of 0.02 between train AUC and test AUC scores that successfully competed other models.