Online decoupling feature framework for optimal probabilistic load forecasting in concept drift environments

Probabilistic load forecasting (PLF) is crucial for optimizing power production and distribution in energy management systems (EMS), enhancing grid stability. However, the issue of concept drift has become increasingly prevalent due to the high sensitivity of electric loads to external features, such as weather and holidays, which cause shifts in the distribution characteristics of load data over time. The current study suffers from the following limitations: (1) Current probabilistic models that handle concept drift often overlook the coupling between external features. (2) There is a notable lack of research exploring the impact of concept drift on quantile and interval predictions, particularly concerning quantile crossing issues in a concept drift setting. To address these challenges, we propose an online probabilistic decoupling feature (OPDF) framework. It captures the coupling relationships among high-impact factors using a decoupling feature structure model based on least absolute shrinkage and selection operator. In the framework, a quantile reconstruction strategy is developed to address the quantile crossover problem in concept drift environments. The quantile reconstruction coefficients are adaptively determined based on the degree of concept drift impact on the model, obtaining optimal probabilistic predictions in terms of sharpness and resolution. Furthermore, the framework employs online caching and adapting schemes to track elusive data patterns in real time and adjust the model learning strategy to accommodate various data distributions in concept drift environments. The proposed framework is validated using real-world load data from three regions in the United States with varying concept drift frequencies (high, moderate, and low) and further demonstrated on the public building load dataset from Suzhou, China, encompassing over 700 users. The analysis demonstrates that our method yields higher quality probabilistic forecasts compared to other online learning approaches and exhibits greater robustness to concept drift than offline probabilistic models.