关键词: 网约车需求预测, 深度学习, 卷积神经网络, 注意力机制, 双向长短期记忆

Abstract: With the advent of mobile internet and the evolution of intelligent transportation systems, online car-hailing services have become increasingly popular as a primary mode of transportation. These platforms, such as Didi Chuxing and Uber, have revolutionized the way people travel by reducing information asymmetry and enhancing the efficiency of resource allocation. They achieve this by effectively matching the supply and demand between passengers and drivers. However, as the frequency of usage has grown, the problems with unreasonable vehicle scheduling and mismatch between supply and demand arise, including extended waiting times for passengers and high empty load rates for drivers. To address these issues, numerous scholars have proposed various deep learning methods to predict car-hailing demand more accurately. Despite these efforts, current forecasting methods still fall short in fully exploiting the features of time series data.
In light of this gap, our paper introduces a novel hybrid prediction model known as Convolutional Attention mechanism Bi-directional Long Short-Term Memory (CNN-ATTBiLSTM). This model innovatively captures time series features from both macro and micro perspectives, utilizing both forward and reverse directions to enhance the accuracy of car-hailing demand forecasts. This approach aims to mitigate the problems with suboptimal vehicle scheduling and the imbalance between supply and demand.
We validate our model using a dataset of car-hailing orders from HaiKou in June 2017, provided by Didi Chuxing’s Gaia Plan. The process begins with data cleansing to remove invalid entries, followed by resampling the data at ten-minute intervals. We then employ the K-means algorithm to identify passenger hot spots, which are used to define the unit regions for analysis. The model, CNN-ATTBiLSTM, is fed with external factors such as weather conditions and the demand for unit areas and time slots to forecast demand for the next time slot. To demonstrate the superiority of our model, we compare it with LSTM and CNN-LSTM models using evaluation metrics like Coefficient of Determination (R²), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Symmetric Mean Absolute Percentage Error (SMAPE). Additionally, we visually compare the forecasting and actual demand curves to analyze the models’ fitting degrees during peak and off-peak times.
The experimental results indicate that CNN-ATTBiLSTM outperforms both LSTM and CNN-LSTM across all evaluation metrics. Our visual analysis reveals that while the LSTM and CNN-LSTM models exhibit a degree of lag in their predictions, our CNN-ATTBiLSTM model more effectively captures the dynamics of time series data during both peak and off-peak periods. Theoretically, the CNN component extracts local features, the Attention mechanism assigns varying weights to these features to model their global relationships, and the BiLSTM component provides robust temporal sequence feature extraction capabilities. Collectively, these elements enable the CNN-ATTBiLSTM to model time series features comprehensively, enhancing demand prediction accuracy and providing valuable insights for car-hailing platforms to devise policies for efficient vehicle scheduling.
However, there are still some improvements that can be made in the future. Currently, our study focuses solely on passenger travel demand, neglecting the potential for drivers to be assigned to restricted areas so as to subsequently cancel orders, leading to a resource waste. Additionally, our spatial analysis does not account for the geographical locations and semantic significance of Points of Interest (POIs). In future research, we plan to incorporate the influence of passenger destinations and consider spatial similarities and POI semantic information as crucial factors to further enhance forecasting accuracy.

Key words: online car-hailing demand prediction, deep learning, convolutional neural networks, attention mechanism, bi-directional long short-term memory