The growing penetration of renewable energy sources (RES) in modern grids may result in severe voltage violation problems due to high stochastic features. Conventional centralized approaches could provide optimal solutions for voltage regulation while with great communication burdens. Control methods based on local information usually have non-optimal results and cannot always guarantee voltage security. This paper proposes a neural network-based decentralized strategy for volt/var control using inverter reactive power capacity. Learning from optimal power flow (OPF) results of historical data, the developed controller can provide optimal results approximate to centralized solutions and outperform local control methods in minimizing the power loss. The proposed method is tested on the IEEE 33-bus system and simulation results illustrate the effectiveness in voltage regulation and loss minimization.