Solving security patrol route planning in large-scale realistic scenarios is a very challenging problem. Here we propose a route planning algorithm PatrolGo based on Monte Carlo Tree Search (MCTS). In this algorithm, the strategy network selects the actions of security resources, and the situation network evaluates the current situation. These deep neural networks are training from the confrontation between PatrolGo and the attacker model. We assume that the attacker and the defender act alternately, and the attacker model is fixed. Specifically, we first apply MCTS to the grid security game scenario. Furthermore, we added the deep neural network to the MCTS algorithm to enhance its efficiency in selecting and evaluating stages. Finally, the PatrolGo algorithm tree search process can provide prescriptive security resources action. We provide a new scheme for security patrol route planning. The simulation results show that the scheme has state-of-the-art performance and effective strategic prescript.