Comparing High Accurate Regression Models For Short-term Load Forecasting In Smart Buildings

Load prediction is at the core of many smart grid topics because its accuracy affects the resilience and optimal operation of the power grid. Short term load forecasting (STLF) specific to buildings is needed for both demand and production sides management, to help reducing peak electricity demand, optimizing energy planning, management, and conservation among other. In this paper, we analyze and compare five state of art regressive STLF algorithms relying on deep learning techniques. Algorithms were chosen based on their performances in the context of load prediction. These methods are compared in the context of smart buildings with real (large) data-set, leading to a comparison on quality metrics and a precise evaluation. The load forecasting is deployed with and without taking into account environmental features to better evaluate the benefits of reaching for external information. The idea is to help determine which of the five forecasting algorithm could be considered in the optimization loop of an energy management systems, either at the level of a district (local energy communities, or smart districts) or directly at the level of individual households (smart residential buildings) using quality metrics and need of exogenous data. Results show that the algorithms will perform better if we accept a larger training time and also present a high sensitivity on the choice of influencing parameters.