Short-term Load Forecasting and Regime Switching Detection Based on Normalized Causal Entropy Boosting

Abstract

Abstract This paper introduces a method based on normalized causal entropy boosting (NCEBoosting) to identify regime switches in load patterns and improve the adaptability of prediction models. Load time-series exhibit multidimensional, nonlinear changes influenced by factors such as human behavior and meteorological environments. To detect regime switches, we calculate the sparsified causal entropy matrix using a small batch of load flow data, which corrects the original prediction model. Subsequently, the corrected model is employed to predict the load requirements after the switch. By comparing the causal entropy across different time periods, we effectively detect conceptual drift, phase changes, and unexpected events in the data, enabling the determination of load mode switches. Given that load timing data is typically generated in batches, calculating the causal entropy of successive data batches provides a robust metric. In this paper, we determine the load mode change by calculating the integrated causal entropy of the batch data streams and calibrating the prediction model accordingly. Extensive computational experiments on real building load datasets are conducted to validate the proposed method. The results demonstrate that the normalized causal entropy boosting effectively detects load mode switches compared to traditional approaches. Furthermore, the method demonstrates efficient adaptive learning based on online data streams, resulting in improved adaptability and prediction accuracy.