Efficient Multi-Step Prediction Model That Considers the Influence of Spatial and Temporal Factors on Ride-Hailing Demand

With the rapid development of the information technology era, accurate ride-hailing demand forecasting can facilitate the pre-scheduling of car routes, optimize vehicle utilization, and alleviate urban peak congestion as information technology advances quickly. Even though many specialists have undertaken research in this area, traditional prediction models seldom use historical data and rarely use multi-step prediction since traffic forecasting involves the influence of many variables, including time and geography. Therefore, this paper proposes an efficient multi-step prediction model (PSA-DM, ProbSparse Self-Attention Distribute Model) for accurate multi-step prediction based on historical request data. In the model, which uses an encoder–decoder framework with long-sequence output prediction. The temporal correlation of the input data is first extracted through the embedding layer, and the correlation between regions is extracted using graph attention networks. To achieve the multi-step prediction effect, the decoder then passes the prediction results back to the encoder for weight distribution and fusion. This allows for further exploration of the complicated spatiotemporal correlation between historical and future information as well as the realization of efficient multi-step prediction. We employed three actual ride-hailing datasets, applied the model to determine the spatial similarity between the data, and tested the projected data for use as the subsequent prediction data. The model performed better than the majority of the baseline models and correctly represented the ride-hailing demand issue according to the data. As a result, the transportation sector has a strong theoretical basis for vehicle scheduling.