Journal of Geo-information Science ›› 2023, Vol. 25 ›› Issue (1): 77-89.doi: 10.12082/dqxxkx.2023.220662

Previous Articles     Next Articles

CLAB Model: A Deep Learning Model for Short-term Prediction of Passenger Rental Travel Demand

ZHOU Yuxin1,2,3(), WU Qunyong1,2,3,*()   

  1. 1. Key Lab of Spatial Data Mining and Information Sharing of Ministry of Education, Fuzhou University, Fuzhou 350108, China
    2. National & Local Joint Engineering Research Center of Satellite Geospatial Information Technology, Fuzhou 350108, China
    3. The Academy of Digital China (Fujian), Fuzhou 350003, China
  • Received:2022-09-06 Revised:2022-10-05 Online:2023-01-25 Published:2023-03-25
  • Contact: WU Qunyong E-mail:531725767@qq.com;qywu@fzu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(41471333);Fujian Science and Technology Plan Guidance Project(2021H0036)

Abstract:

Passenger travel demand prediction is an integral part of intelligent transportation systems, and accurate travel demand prediction is of great significance for vehicle scheduling. However, existing prediction methods are unable to accurately explore its potential spatiotemporal correlation and mostly ignore the impact of historical inflow on travel demand. In order to further exploit the spatiotemporal characteristics of spatiotemporal big data and improve the accuracy of the model in predicting passenger travel demand, this paper proposes a Conv-LSTM Attention BiLSTM (CLAB) model for short-time prediction of passenger rental travel demand. The attention-based Conv-LSTM module extracts spatial features and short-term temporal features of passenger travel demand at the near moment, where the attention mechanism automatically assigns different weights to discriminate the importance of demand sequences at different times. To explore long-term temporal features, two BiLSTM modules are used to extract temporal features of historical inflow sequences and temporal features of daily passenger temporal features of the demand series. Experiments are conducted using the order data of online and cruising taxis on Xiamen Island, and the results show that: (1) the CLAB model is more suitable for predicting the future 5-min short-time passenger travel demand using 30-min historical data; (2) the overall effect error of the CLAB model is lower and has better prediction results compared with the benchmark prediction model. The CLAB model is more effective than the CNN-LSTM, LSTM, BiLSTM, CNN, and Conv-LSTM by 33.179%, 33.153%, 33.204%, 5.401%, and 5.914% in mean absolute error (MAE) and 34.389%, 34.423%, 34.524%, 6.772%, and 6.669% in Root Mean Square Error (RMSE), respectively; (3) the CLAB model performs better for weekday prediction with higher regularity than non-working day prediction, with best prediction for weekday morning peaks.

Key words: traffic big data, travel demand forecasting, deep neural network, attention mechanism, combined forecasting model, spatiotemporal fusion, Xiamen Island, LSTM