A cooperative control algorithm of multi-intersection variable-direction lanes based on multi-agent reinforcement learning was proposed to alleviate the congestion of multi-intersection, in order to solve the problem that traditional variable-direction lane control method can't adapt to the complex traffic flow problem under multiple intersections scenarios. In this method, the deep multi-agent reinforcement learning (QMIX ) algorithm was improved. The global reward under variable-direction lane scenarios was composed of basic reward and performance reward, which improved the decision-making accuracy of lane turn control in congestion scenarios. The priority experience playback algorithm was introduced to improve the utilization efficiency of the transfer sequence in the experience playback pool and accelerate the algorithm convergence. Experimental results show that the algorithm has better performance than other control methods in case of queue length, delay times and waiting times, which can effectively coordinate the policy switch of the variable-direction lanes and improve the road network capacity in the multi-intersection scenarios.