Net-zero energy building (NZEB) is an emerging active carbon reduction solution. It achieves on-site renewable generations and zero balance between supply and demand. In NZEB, the photovoltaic (PV) accounts for the majority of distributed generations. However, its random and stochastic nature aggravates the accuracy of look-ahead PV output forecasting. Meanwhile, the spatial-temporal uncertainty of customers' electricity consumption can also result in improper load predictions. Besides, missing data in the late-model NZEB energy system enhances the difficulty of accurate forecasting. In this paper, a LSTM and Transformer-based forecasting models with K-nearest Neighbors (KNN) data interpolation and Generative Adversarial Network (GAN) data augmentation are demonstrated for load consumption and PV output forecasting of NZEB. The proposed framework obtains significant improvements in terms of root-mean-square error (RMSE) by 7.168 kW and 4.603 kW, in load and PV power forecasting respectively.