AC power system state estimation process aims to produce a real-time 'snapshot' model for the network. Therefore, a grand challenge to the newly built smart grid is how to 'optimally' estimate the state with increasing uncertainties, such as intermittent wind power generation or inconsecutive vehicle charging. Mathematically, such estimation problems are usually formulated as Weighted Least Square (WLS) problems in literature. As the problems are nonconvex, current solvers, for instance the ones implementing the Newton's method, for these problems often achieve local optimum, rather than the much desired global optimum. Due to this local optimum issue, current estimators may lead to incorrect user power cut-offs or even costly blackouts in the volatile smart grid. To initialize the iterative solver, in this paper, we propose utilizing historical data as well as fast-growing computational power of Energy Management System, to efficiently obtain a good initial state. Specifically, kernel ridge regression is proposed in a Bayesian framework based on Nearest Neighbors search. Simulation results of the proposed method show that the new method produces an initial guess excelling current industrial approach.