Buildings are one of the major sources of electricity and greenhouse gas emission (GHG) in urban areas all around the world. Since a large integration of solar energy is observed in the form of rooftop photovoltaic (PV) units, electricity use of buildings is highly uncertain due to intermittent nature of solar output power. This leads to poor energy management for both network operators and building owners. In addition, uncertain metrological conditions, diversity and complexity of buildings are big hurdles to accurate prediction of the demand. To improve accuracy of load demand forecast of PV integrated smart building, a hybrid ensemble framework is proposed in this paper. This is based on a combination of five different predictors named as backpropagation neural network (BPNN), Elman neural network (EN), Autoregressive Integrated Moving Average (ARIMA), feed forward neural network (FNN), radial basis function (RBF) and their wavelet transform (WT) models. WT is applied to historical load data to remove the spikes and fluctuations. FNN and RBF network were trained with particle swarm optimization (PSO) for higher forecast accuracy. The output of each predictor in the ensemble network is combined using Bayesian model averaging (BMA). The proposed framework is tested using real data of two practical PV integrated smart buildings in a big university environment. The results indicate that the proposed framework show improvement in average forecast normalized root mean square error (nRMSE) around 17% and 20% in seasonal daily and seasonal weekly case studies, respectively. In addition, proposed framework also produces lowest of nRMSE about 3.88% in seasonal monthly forecast of smart buildings with rooftop PV as compared to benchmark model. The proposed forecast framework provides consistent forecast results for global change institute (GO) and advance engineering building (AEB) during seasonal daily and weekly comparison.