Abstract: Accurate prediction of bus turnaround time is the foundation and prerequisite for intelligent bus scheduling, which is the key to formulate driving schedules. In order to improve the prediction accuracy of bus turnaround time, a prediction model for bus turnaround time based on hunter-prey optimization for long short-term memory (HPO-LSTM) neural network was proposed. The hyperparameters of the long short-term memory (LSTM) neural network, including the number of hidden layer nodes, the number of iteration cycles and initial learning rate, were mapped to the population positions of the hunter-prey optimization (HPO) algorithm. The root mean square error generated (ERMS) by the predicted value and the real value of the LSTM model was taken as the population fitness function to optimize the population position, achieving LSTM neural network hyperparameter optimization. The LSTM neural network was established by the optimal hyperparameters to predict bus turnaround time. The proposed model was validated and analyzed by the data of bus Line 1 in a certain city. The results show that compared to the BP, LSTM, FA-BP, and HPO-BP models, the mean absolute percentage error (EMAP) of the HPO-LSTM model is decreased by 10.44%, 4.00%, 3.61% and 2.04%, respectively.

摘要： 公交周转时间的准确预测是公交智能排班的基础和前提，是制定行车时刻表的关键。为提高公交周转时间的预测精度，提出了基于猎人猎物优化长短时记忆神经网络（HPO-LSTM）的公交周转时间预测模型，将长短时记忆神经网络（LSTM）中的超参数（隐含层节点数、迭代循环数以及初始学习率）映射为猎人猎物优化算法（HPO）种群的位置；以LSTM模型预测值与真实值产生的均方根误差ERMS作为种群适应度函数，优化种群位置，实现LSTM神经网络超参数寻优；用最优超参数构建LSTM神经网络，进行公交周转时间预测。采用某市公交1号线数据对模型进行验证分析，结果表明：相比于BP、LSTM、FA-BP、HPO-BP模型，HPO-LSTM模型平均绝对百分比误差EMAP分别降低10.44%、4.00%、3.61%、2.04%。