A grand challenge for state estimation in newly built smart grid lies in how to deal with the increasing uncertainties. To solve the problem, we propose a data-driven state estimation approach based on recent targeted investment on sensors, data storage, and computing devices. An architecture is proposed to use power system physics and pattern to systematically clean historical data and conduct supervised learning, where historical similar measurements and their states are used to learn the relationship between the current measurement and the state. In order to deal with nonlinearity, kernel trick is used to produce linear mapping in a carefully selected higher dimensional space. To speed up the data-driven approach for online services, we analyze power system data set and discover its clustering property due to the periodic pattern of power systems. This leads to significant dimension reduction and the idea of preorganizing data points in a tree structure for inquiry, leading to 1000 times speedup. Numerical results show that the proposed data-driven approach works well in a smart grid setting with increasing uncertainties and it produces an online state estimate excelling current industrial approach.