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