A demand for occupancy estimates using low time frequency such as 30 minutes interval energy data from smart meters is increasing, since it is beneficial to society in many ways. Knowing whether a household is occupied by its residents in real time contributes to removing absent delivery, route optimization for work activities, and balancing the supply and demand of energy. A Previous work demonstrated machine learning models in conventional methods such as a random forest (RF) and a support vector machine (SVM) can detect occupancy status with over 80 percent accuracy. In this paper, we develop a new model, bidirectional long short term memory (LSTM) combined with an attention mechanism (BiLSTM-Attention). The two major differences from the previous study are learning long-term time dependency which a LSTM enables and effective use of information at other times that are highly similar to a given time by an attention. We demonstrated that our proposing BiLSTM-Attention can lead to detect occupancy status equally or more accurately compared with a previous study in the most large publicly available data.