Accurately forecasting electricity system outcomes, such as power flow schedules and nodal prices, is crucial to integrating renewable generation and battery storage technologies into the grid. These outcomes are the result of a large-scale constrained optimization problem and must satisfy various hard constraints. We propose combining a neural network with a differentiable approximate proximal message passing solver to produce an end-to-end model for grid forecasting. Our method does not require expensive implicit differentiation steps and generalizes to new system topologies unobserved during training. Initial experiments on a Western U.S. grid dataset suggest our method can improve upon traditional architectures in both constraint satisfaction and generalization to unseen problem data.