基于S-Catboost算法的短时公交客流预测及影响因子分析

Short-termbus passenger flow forecast based on Stacking-Catboost algorithm

针对城市公交实时客流数据多样化和特征复杂的状况,提出一种基于S-Catboost模型的客流特征提取及短时客流预测方法和影响因子分析流程。首先,通过爬虫技术获取公交客流数据的环境和时变特征,扩充客流数据的特征维度。其次,通过对客流数据进行时间和精度加权并采用LSTM和随机森林2种基模型对客流数据进行堆叠(Stacking),提取强特征并加入第二层子模型的特征矩阵。最后,利用Catboost算法对类别特征进行数值化处理,得到预测结果。实验结果表明该模型比传统LSTM、随机森林、GBDT以及SVM在预测准确度和计算时间上都有明显的优势,并给出了不同影响因子对客流量的相对贡献度和各自的偏效应,该模型对于公交系统进行实时线网优化调度和线路拥挤度信息发布具有比较高的实用价值。

Aiming at the diversification of real-time passenger flow data and complex features of urban public transport,this paper proposes a method of passenger flow feature extraction and short-term passenger flow prediction and influence factors analysis based on the weighted Stacking-Catboost model.First of all,the environment and time-varying characteristics of bus passenger flow data were acquired through crawler technology,and the characteristic dimension of passenger flow data was expanded.Secondly,by weighting the time and accuracy of the passenger flow data,stacking the passenger flow data using two basic models of LSTM and random forest,we extracted strong features and added the feature matrix of the second layer of sub-models.Finally,we used the Catboost algorithm to calculate the category features and obtained the predicted results.The experimental results show that compared to traditional LSTM,random forest,GBDT and SVM,the model has obvious advantages in prediction accuracy and calculation time.The study also gives the relative contribution and partial effect of different influence factors on passenger flow.Therefore,this model has high practical value for real-time line network optimization dispatching and line congestion information release of public transportation system.