Electronic International Standard Serial Number (EISSN)
1873-6769
abstract
Power load forecasting plays a critical role in the context of electric supply optimization. The concept ofload characterization and profiling has been used in the past as a valuable approach to improve forecasting performance as well as problem interpretability. This paper proposes a novel, fully fledged theoretical framework for a joint probabilistic clustering andregression model, which is different from existing models that treat both processes independently. The clustering process is enhanced by simultaneously using the input data and the prediction targets during training. The model is thus capable of obtaining better clusters than other methods, leading to more informativedata profiles, while maintaining or improving predictive performance. Experiments have been conducted using aggregated load data from two U.S.A. regional transmission organizations, collected over 8 years. These experiments confirm that the proposed model achieves the goalsset for interpretability and forecasting performance.
Classification
subjects
Telecommunications
keywords
clustering; forecasting; machine learning; power load; probabilistic model; profiling