As the conventional optimization algorithms of power flow cannot meet the requirements of real-time scheduling of power system with complex and nonlinear descriptional multi-objective optimal power flow（OPF）,this paper presents a multi-step Q（λ） learning algorithm based on the semi-Markov decision process.This algorithm,independent of any accurate model,converts the constraints,actions and targets of the optimal power flow to the status,actions and rewards of the algorithm,and dynamically finds the optimal action by continuous fault testing,retrospecting and iteration.By comparing comparison of the proposed algorithm with other algorithms in several IEEE standard examples,it is found that the Q（λ） learning algorithm is feasible and effective in dealing with multi-objective OPF problems.