Exploring Dynamic Factor Graphs for Forecasting Building Electrical Load

We present a detailed study of Dynamic Factor Graphs (DFG) for addressing time-series forecasting problem. Some modifications to the basic DFG network have been proposed in order to improve stability of the model and accuracy of the forecasts. We experimented with the real electrical load data from Office Buildings to study the various proposed versions of DFG. The main novelty of this work is that an efficient network structure is derived and implemented based on the domain knowledge for superior performance without increasing the complexity of the network. The various experimental models of DFG are compared with the basic version and all the suggested modifications proved to perform better than the basic one. The model with exogenous variable directly connected to output variable gives least errors on an average for Load Forecasting.