Forecasting online adaptation methods for energy domain

Abstract

Time series forecasting is crucial for addressing renewable energy optimization challenges, such as minimizing financial costs. Factors like seasonality and human activity induce concept drift in energy-related forecasting tasks. Despite this, the majority of studies concentrate on offline forecasting models, which are unable to accommodate concept drift. In response, we propose a replacement learning-based online adaptation framework for multivariate multi-step time series forecasting in the energy domain. This method tackles concept drift by retraining models from scratch, leveraging historical data via clustering-based sampling and dimensionality reduction. Furthermore, the forecast model’s predictions are refined using the proposed correcting factor, which serves to adjust for any inherent bias present in the forecasting model. Additionally, we develop an incremental learning model that adapts to concept drift by building on the previous model, offering a comparative analysis and investigation of incremental learning. Evaluations using synthetic building electricity demand and cooling system electricity demand datasets demonstrate the stability and effectiveness of our proposed replacement learning adaptation, outperforming offline models by over 50% across various dataset characteristics.