Predicting the Demand of Ride-Hailing Based on a QRF-SA-ConvLSTM Model

Considering the spatiotemporal features and predictive uncertainty in the demand prediction of ride-hailing are essential for improving the efficiency and robustness of the transportation system. This study proposes a novel prediction framework that integrates quantile regression forests (QRF) and a self-attention convolutional long short-term memory network (SA-ConvLSTM) to jointly optimize spatiotemporal feature learning and probabilistic modeling, which could lead to high-precision prediction and uncertainty estimation. First, the model employs QRF to generate the distributions of predicted demand within specified probability intervals from historical spatiotemporal data. It then establishes a spatiotemporal probability matrix by weighted fusion of original data through the attention mechanism. The matrix is then fed into the SA-ConvLSTM module, where the convolutional structure extracts local spatial patterns, the self-attention mechanism focuses on key spatiotemporal nodes, and LSTM captures long-term temporal dependencies. Finally, a multi-task output layer simultaneously optimizes point predictions and interval estimations. Combining the ability of distributional modeling of QRF and the advantage of extraction of spatiotemporal feature of deep learning, the proposed model improves both prediction accuracy and uncertainty estimation. The study validates the model's effectiveness and accuracy using Didi's open-source dataset of ride-hailing trajectories within the second ring road in Xi'an. The results show that in terms of mean prediction, the proposed model reduces mean absolute error and mean squared error by 21% and 17.2%, respectively, comparing to the SA-ConvLSTM baseline; in terms of uncertainty estimation, the prediction interval coverage probability at 95% confident level improves by 5.3%, 2.5%, and 0.9% comparing to linear quantile regression (LQR), QRF, and SA-ConvLSTM, respectively, while the average width of the prediction results is reduced by 41.5%, 30%, and 18.5%. These results validate the proposed model's goodness of fit in terms of both predictive accuracy and reliability.