Day-Ahead and Intra-Day Building Load Forecast With Uncertainty Bounds Using Small Data Batches

An approach to provide day-ahead and intra-day load forecasts of buildings, such as electrical or thermal power consumption, is presented. The method aims obtain a nominal forecast associated error bounds with small data batches two weeks for the training phase, resulting in ready-to-go algorithm that can be employed whenever large datasets months years are not available manageable. These cases include new renovated constructions, buildings subject changes purpose occupants’ behavior, applications on local devices memory limits. relies so-called “fictitious input” signal capture prior information seasonal periodic trends consumption. Then, linear multistep predictors different horizon lengths trained periodically batch most recent data, worst case derived, using set membership (SM) methods. Finally, computed, each time step, by intersecting predictions taking central value obtained interval. Such applied here first real-world consumption medium-size building cooling complex. In both cases, results indicate tightening between 15% 25% average respect those standard SM approach.