A Stacking Framework for Online Locational Marginal Price Prediction Considering Concept Drift

Abstract

Concept drift means the statistical properties of the variable that a predictor is predicting change over time in unforeseen ways. Existing research solves concept drift in the locational marginal price (LMP) prediction process by updating predictors in online approaches. However, new data is indiscriminately utilized to update predictors in these methods. The new property changes can not be accurately captured when concept drift occurs. This paper proposes a stacking framework for online LMP prediction considering the concept drift phenomenon. Long short-term memory networks and graph attention networks are selected as the base predictors to capture the spatio-temporal dependencies in LMPs. When concept drift occurs, data with drift selected by the adaptive windowing algorithm is used to update the stacked predictor. Numerical results based on real data from Australian Energy Market Operator and Midcontinent Independent System Operator validate the effectiveness of the proposed framework. The comparative experiments prove that attempts to change or simplify the proposed framework can undermine prediction accuracy.