Occupancy Forecasting using LSTM Neural Network and Transfer Learning

Abstract

Neural networks can be used as a forecasting tool in several fields such as medicine, agriculture, and entertainment. Accurate forecasting of human habit such as the entry/exit behavior of a person may be exploited to control electrical appliances in order to reduce energy consumption while maintaining comfort. However, the neural network has a problem that is it can be trained to forecast behavior of only one person. If the neural network is used to predict another person, It will decrease accuracy. Although new data will be collected to re-train the neural network, data collection might take long time. This paper proposes to use transfer learning on a Long Short-Term Memory (LSTM) network in order to improve the performance of the network after a specific person uses the room, the person changes his/her behavior, or a new person occupies the room. After a network is trained by a norm dataset, then new batches of sampling data can be applied to update the network, in other words, to transfer the new knowledge on top of the existing one. The results show that transfer learning helps the LSTM network to be able to track the behavior change continually. Its forecast becomes more and more accurate, when compared to that of the norm one.