Reasonable and efficient allocation plan of emergency resources is key to successful disaster response. The Q-learning algorithm may provide a new approach to the resource allocation problem, which can take a variety of factors into consideration and timely respond to subsequent changes. In this paper, we propose a reinforcement learning Q-learning model in which the subsequent changes of disasters are included. Furthermore, in order to obtain a convincing result, three penalty functions are presented to represent three key factors, including efficiency, effectiveness and fairness. We test our proposed method under the background of actual flood disaster and compare results with real-time disaster development. Our empirical results show that the proposed model is in line with the development of disaster, and extremely sensitive to information on subsequent changes in the disaster situations, verifying its effectiveness in allocation of emergency resources.