Learning Fast Approximations For Nonconvex Optimization Problems Via Deep Learning With Applications To Power Systems