Enhancing Model Adaptability Using Concept Drift Detection for Short-Term Load Forecast

Concept drift refers to the relation between input and target envolves over time in an online supervised learning scenario. With the development of smart grid and smart meter, mass data accessibility poses a huge challenge to learning model adaptability. To address such issue, this paper proposed a model adaptability enhancement approaches based on concept drift detection for short-term forecast model. It exploits canonical correlation analysis to measure mapping relation between input and output of the forecast model. Then the correlation coefficient vector sequence will be monitored over an adaptive window to detect concept drift. Model is updated on data over the sliding window only when a concept drift happens. Experimental results shows this method reduces memory assumption and computing resources remarkably meanwhile guarantees the forecast accuracy.