Smart grids are automated, distributed energy exchange networks that, in contrast to traditional electricity grids, feature reconfigurable network topologies. Network planning is an essential function of smart grids that connects customers to energy sources using available physical links in the network. We model this as the problem of creating a spanning forest with a capacity constraint on each tree bounding its total weight. Each tree of the forest corresponds to a set of customers, rooted at a source. We call this the Capacitated Spanning Forest (CSF) problem. CSF is NP-complete even on unweighted graphs with two sources. We present a solution to this problem using a Local Search heuristic and demonstrate its performance on square grids and on a real-world sample grid topology.