Load Forecasting Method for Building Energy Systems Based On Modified Two-Layer LSTM

Accurate load forecasting is the foundation of the building energy system to participate in smart grid scheduling. However, as diverse appliances are connected to the building, the load profile contains more different patterns, which brings in challenges for load forecasting research. What's more, since the active-reactive power coordination scheduling in smart grids has become more important, the reactive load is also needed so that additional attention must be paid to the reactive load forecasting. In this paper, a load forecasting method for building energy systems based on modified two-layer long short-term memory (LSTM) is proposed to deal with such problems. In the structure of the designed two-layer LSTM deep learning neural network, the lower layer LSTM network is trained to capture the temporal characteristic between active load and its influencing factors. The upper layer LSTM network is trained to learn the characteristic of reactive load, by feeding into the historical reactive loads in addition to the hidden information from the lower layer network, based on the physical concept that reactive power of each appliance is coupled with its active power. As a result, the joint forecasting of active and reactive loads can be achieved by the parallel training of the lower layer and upper layer LSTM networks. The simulation results verify that the proposed method show better accuracy compared to the single LSTM-based approach.