Short-term thermal and electric load forecasting in buildings
Increasing environmental awareness and energy costs encourage the increase of the contribution of renewable energy sources (RES) to the energy supply of buildings. However, the integration of RES and energy storage systems introduces significant challenges for the energy management system (EMS) of complex building energy systems. An energy management strategy based on fixed control rules may fail to efficiently operate such systems. These circumstances raise the need to apply advanced control strategies. A promising approach is model predictive control (MPC), which allows the consideration of the expected dynamic system behavior as well as of forecasts of the loads and of the renewable energy generated. Obviously, the performance of an MPC-based EMS crucially depends on the accuracy of the load forecasts. The goal of this paper is to compare the capabilities of neural networks (NNs) and of the least squares support vector machine (LS-SVM) in forecasting the hourly thermal and electric load of buildings. Two short-term load forecasting algorithms are evaluated which treat every hour of the day separately by an individual forecasting model. Additionally, the algorithms also distinguish between working days, weekends and holidays. In order to adapt to changing load patterns, the algorithms use the sliding window training approach. Both algorithms are tested using the measured thermal and electric load data of a large office building and of a small building which houses a kindergarten. In the tests conducted, in general, the forecasting algorithm based on the LS-SVM shows a better performance than the forecasting algorithm based on NNs. In addition, the LS-SVM involves fewer free parameters to be determined than a NN, which makes the former easier to apply. The results reported further indicate that the accurate forecasting of the load of a small building is the more challenging task compared to the load forecasting of a large office building. Furthermore, using a training window size of more than 20 days does not significantly improve the performance of the algorithms examined.
Record created on 2015-09-23, modified on 2016-08-09