基于SSA-GRU深度学习优化模型的短时公交客流预测

Short-Term Bus Passenger Flow Forecasting Based on SSA-GRU Deep Learning

摘要: 短时公交客流预测是实时动态调整公交发车频率，实现公交精准动态调度的重要决策基础之一。为挖掘公交客流的时序特征，提升短时公交客流预测精度，建立了一种基于麻雀搜索算法(Sparrow Search Algorithm, SSA)优化门控循环单元(Gated Recurrent Unit, GRU)的短时公交客流预测模型。该模型基于历史客流数据的时序分布特征，利用SSA寻优算法对GRU中的隐含层节点个数、学习率和训练次数进行寻优，然后依据参数寻优结果重构模型超参数，利用门控循环神经网络基进行短时客流预测。为验证优化模型的预测性能，选取某城市1号线站点客流数据进行实验；结果表明，相比于GRU门控循环神经网络，SSA-GRU模型的平均相对百分误差(MAPE)降低了37.9%、平均绝对误差(MAE)降低了42.1%，均方根误差(RMSE)降低了40.7%，组合模型的预测精度高于GRU模型且能够有效进行短时公交客流预测。

Abstract: Short-term bus passenger flow prediction is one of the important decision bases for real-time dy-namic bus frequency adjustment and accurate dynamic bus scheduling. In order to mine the time series characteristics of bus passenger flow and improve the prediction accuracy of short-term bus passenger flow, a short-term bus passenger flow prediction model based on Gated Recurrent Unit (GRU) optimized by Sparrow Search Algorithm (SSA) was established. This model was based on the historical passenger flow data of time series distribution. Firstly, SSA optimization algorithm was used to optimize the hyper parameters of GRU, such as hidden layer nodes, learning rate and train-ing times. Then, based on the parameter optimization results, short-term passenger flow prediction was carried out by using gated recurrent neural network. To verify the predictive performance of the optimized model, the passenger flow data of Line 1 station in a city were selected for experi-ment. The results show that compared with the GRU gated recurrent neural network model, the Mean Absolute Error (MAE) of the SSA-GRU model is reduced by 42.1%, the Root Mean Square Error (RMSE) is reduced by 40.7%, and the Mean Relative Percentage Error (MAPE) is reduced by 37.9%. The combined model can effectively predict short-term bus passenger flow. The results show that compared with the GRU gated recurrent neural network model, the Mean Absolute Error (MAE) of the SSA-GRU model is reduced by 42.1%, the Root Mean Square Error (RMSE) is reduced by 40.7%, and the Mean Relative Percentage Error (MAPE) is reduced by 37.9%. The combined model can ef-fectively predict short-term bus ridership, and the prediction accuracy is higher than that of the GRU model.