In order to improve the operational security of power system, the risk assessment theory is introduced in classic reactive power optimization (RPO). So a new mathematical model of risk-based multi-objective reactive power optimization (RMRPO) is established. Besides, a novel transfer tribe reinforcement learning (TTRL) is proposed to solve the RMRPO problem, which effectively combines the tribal search collaboration and the trial-error process of reinforcement learning. The Q-matrix is adopted as the knowledge matrix to store the optimization information. Then the knowledge transfer can be executed to accelerate the rate of the proposed algorithm. The simulation on IEEE 118-bus system demonstrates that the rate of TTRL can be 2 to 10 times faster than other existing artificial intelligence algorithm. Meanwhile, the quality of the optimal solutions of TTRL can be guaranteed.