We propose a multi-fidelity neural network (MFNN) tailored for rapid high-dimensional grid power flow simulations and contingency analysis with scarce high-fidelity contingency data. The proposed model comprises two networks -- the first one trained on DC approximation as low-fidelity data and coupled to a high-fidelity neural net trained on both low- and high-fidelity power flow data. Each network features a latent module which parametrizes the model by a discrete grid topology vector for generalization (e.g., $n$ power lines with $k$ disconnections or contingencies, if any), and the targeted high-fidelity output is a weighted sum of linear and nonlinear functions. We tested the model on 14- and 118-bus test cases and evaluated its performance based on the $n-k$ power flow prediction accuracy with respect to imbalanced contingency data and high-to-low-fidelity sample ratio. The results presented herein demonstrate MFNN's potential and its limits with up to two orders of magnitude faster and more accurate power flow solutions than DC approximation.