Conventional AGC systems are based on classical linear control theory. Most load frequency controls are primarily composed of an integral controller. Because of the inherent non-linearity and variable structure of the system， integrator gain is very difficult to set. This often results in generator overshoot or insufficient regulation. In this paper， the effect of non-linearity， such as dead-band of speed governor， transmission delay， generation rate control（GRC） and unit ramp control（URC） for AGC is considered comprehensively. AGC system is discretized as a Markov chain， and Q -learning algorithm as controller replaces PI controller. Area control error as state variable of the system is learned by Q -learning， one of reinforcement learning method. Furthermore， to improve study efficiency， the information of AGC environment using fuzzy integrated decision-making is translated into prior knowledge of Q -learning. Simulation result of this AGC system shows the improved Q -learning method has a higher learning efficiency and convergence speed. In order to test feasibility of this method， the results of simulation are calculated to test if it meets the CPS standard. The result is satisfactory in this regard.