Order Demand Prediction and Anomaly-point Identification for Online Car-hailing Orders Based on Hybrid Machine Learning Framework

The demand for urban ride-hailing services holds significant potential for understanding residents'travel behaviors, patterns and intrinsic characteristics. Accurately identifying anomalies and optimizing scheduling from the complex and dynamic spatio-temporal data of ride-hailing usage can contribute to extending a platform's capacity. Time series graph of ride-hailing order data is established to analyze its dynamic characteristics. Therefore, a hybrid prediction model that predicts ride-hailing order demand based on machine learning methods, called ARIMA-BPNN-DSR (ABD), is proposed by integrating the auto regressive integrated moving average model (ARIMA) and the back propagation neural network (BPNN) modules. To achieve the hybrid prediction model, the dynamic selection of regression (DSR) method is applied to fuse these two modules. The DSR method takes advantage of the robustness of statistical methods and the efficiency of machine learning methods, and considers the performance of independent models within the local data space. Extensive experiments and analyses are conducted on the time series data from Didi's ride-hailing order demand in Xiamen City, including data from 2019 (without epidemic) and data from 2020 (with epidemic). Experimental results show that: ①The ABD model outperforms baseline models, providing accurate predictions for peak demand. Therefore, incorporating ensemble learning strategies significantly improves the prediction accuracy of the proposed model. ②Ablation experiments reveal that the BPNN significantly enhances the predictive performance of the fusion model in standard sequences. Compared to individual ARIMA and BPNN models, the mean absolute error (MAE) of ABD model is reduced by 22.77% and 13.50%, and the mean absolute percentage error (MAPE) is reduced by 21.71% and 12.37%, respectively. Considering the external interference in 2020, the stability provided by ARIMA is essential. ③By comparing the error between historical data and predicted results with the 3-sigma anomaly detection criteria, ABD model accurately identifies anomalies in the order data, thereby increasing the efficiency of traffic management. In conclusion, the proposed ABD model has a better performance in both accuracy and robustness.