Effectiveness of forecasters based on Neural Networks for Energy Management in Zero Energy Buildings

Abstract

Energy management is an important challenge in Zero Energy Buildings (ZEB) with Photovoltaic (PV) generation systems. Measuring, Forecasting and Controlling Energy are important elements for energy management in buildings of all types, therefore, this research work focuses on the forecast stage of photovoltaic generation and energy consumption of the building. For this purpose, an analysis and comparison of five neural network techniques, namely Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM) single, dual and bidirectional modality and Gated Recurrent Units (GRU), were carried out. The results show that the MLP forecaster for photovoltaic energy generation turned out to be faster, however the dual-layer LSTM forecaster was more reliable and accurate with errors less than 1% between measured and forecast data. For energy consumption forecasting, the bidirectional LSTM forecaster turned out to be better. The evaluated forecasters turned out to be effective under homogeneous conditions of radiant sun, but they have effectiveness of 62% under conditions of cloudiness or solar intermittency, consequently failing to optimize the energy management of the ZEB system. This research work contributes to determining the best neural network technique as a forecaster for the energy management of ZEB as well as determining the impact and effectiveness of the forecaster in the face of solar intermittence. In addition, it provides an overview for decision making for those interested in developing smart electrical management systems for ZEB.