In order to solve the problem of uneven distribution of toll lane resources in expressway toll stations, a method of toll lane opening and closing configuration based on traffic flow prediction is proposed by combining Bayesian optimization algorithm(BOA) and long short term memory(LSTM) neural network. First, the toll data is preprocessed to obtain basic data such as traffic volume, vehicle type proportion, toll method proportion and service time, which are used to build the model training dataset. Secondly, a M/G/K queueing model based on multiple charging method is proposed to achieve the theoretical description of toll station passing process and the calculation of key indicators such as average queue length, average stay time and traffic capacity. Thirdly, aiming at the difficult problem of hyperparameter setting, a combination model of BOA-LSTM is constructed with Bayesian optimization algorithm to achieve traffic flow prediction. Then, with the traffic flow prediction results as input and the number of lane openings as output, a lane opening and closing configuration model with the goal of minimum comprehensive cost is constructed. Finally, taking Hebei Xinyuan Expressway airport toll station as an example, the empirical analysis shows that the BOA-LSTM combination model can achieve good prediction results, in which the root mean square error of traffic volume and vehicle type proportion are 16.24 and 0.03 respectively, and the mean absolute percentage error is 13.32% and 1.77% respectively. Compared with the current scheme, the average daily comprehensive cost of working days is reduced by 2.30% and the average daily comprehensive cost of rest days is reduced by 5.14%.