Artificial bee colony (ABC) algorithm is applied to the medium and long term power load forecasting. Combined with the combination forecasting model, it optimizes the weights of combination prediction in objective function. A disturbing term and worst honey substitution are introduced to overcome the problems of slow convergence speed in the early stage and easy falling to local optimum in the late stage of the existing ABC algorithm. Case analysis shows that the improved method has rapid convergence and strong global optimization. Compared with the forecasting result of single model, the combination forecasting value by using the improved method is more accurate, which shows it is feasible in the medium and long term power load forecasting.