Learning-based Solar Power and Load Forecasting in DC Net-zero Energy Building with Incomplete Data

To cope with global climate change and carbon emission reduction, net-zero energy building (NZEB) was advocated in the building industries in these years, which aims to minimize energy needs and achieve zero-energy balance through renewables generation. Practically, because of the increasing energy demand and uncertainty in renewables generation, the building microgrid mostly relies on grid supply for daily usage instead of renewables generation, which violates the intention of zero energy. Thus, accurate load and PV power forecasting are imperative for NZEB power system planning. However, data-driven forecasting requires high volume of complete dataset. This paper compares load and PV power forecasting performance by learning-based models with incomplete data based on a demonstration NZEB project. The result highlights that Gated Recurrent Unit (GRU) and Bidirectional Long Short Term Memory (BiLSTM) with K-nearest neighbors (KNN) interpolation on improvement of weekly load and PV power forecasting RMSE by 10.132 kW and 6.650 kW respectively.