Automatic recognition of activities of daily living (ADL) can be applied to realize services to support user life such as elderly monitoring, energy-saving home appliance control, and health support. In particular, 'real-time' ADL recognition is essential to realize such a service that the system needs to know the user's current activity. There are many studies on ADL recognition. However, none of these studies address all of the following problems: (1) privacy intrusion due to the utilization of high privacy-invasive devices such as cameras and microphones; (2) limited number of recognizable activities; (3) low recognition accuracy; (4) high deployment and maintenance costs due to many sensors used; and (5) long recognition time. In our prior work, we proposed a system which solves the problems (1)-(4) to some extent by using user's position data and power consumption data of home electric appliances. In this paper, aiming to solve all the above problems including (5), we propose a new system by extending our prior work. To realize 'real-time' ADL recognition while keeping good recognition accuracy, we developed new power meters with higher sensing frequency and introduced new techniques such as adding new features, selecting the best subset of the features, and selecting the best training dataset used for machine learning. We collected the sensor data in our smart home facility for 11 days, and applied the proposed method to these sensor data. As a result, the proposed method achieved accuracy of 79.393% in recognizing 10 types of ADLs.