Source-linked AI summary
Short-Term Forecasting of Passenger Demand under On-Demand Ride Services: A Spatio-Temporal Deep Learning Approach
Jintao Ke, Hongyu Zheng, Hai Yang, Xiqun, Chen
TL;DR
Short-term passenger-demand forecasting remains challenging across spatial and temporal settings, while real-world data have been unavailable. The paper proposes FCL-Net and reports that its two variants significantly outperform benchmark algorithms, including approximately 50.9% lower RMSE than conv-LSTM.
Problem
Short-term passenger-demand forecasting across different spatial settings remains challenging, with limited real-world data availability constraining prior work.
Method
The paper proposes the fusion convolutional long short-term memory network (FCL-Net), a deep-learning structure combining convolutional and LSTM-based components, with spatially aggregated random-forest feature selection.
Results
The two FCL-Nets significantly outperform benchmark algorithms on RMSE and R-square, achieving approximately 50.9% lower RMSE than conv-LSTM.
Takeaways & Limitations
The results support FCL-Net as a deep-learning approach for forecasting passenger demand while examining spatio-temporal correlations.
Takeaways & Limitations
The range of economic variables in the on-demand ride-service market is insufficient for predicting passenger demand, motivating the use of more economic variables in future work.
Abstract
from arXiv · showhide
Short-term passenger demand forecasting is of great importance to the on-demand ride service platform, which can incentivize vacant cars moving from over-supply regions to over-demand regions. The spatial dependences, temporal dependences, and exogenous dependences need to be considered simultaneously, however, which makes short-term passenger demand forecasting challenging. We propose a novel deep learning (DL) approach, named the fusion convolutional long short-term memory network (FCL-Net), to address these three dependences within one end-to-end learning architecture. The model is stacked and fused by multiple convolutional long short-term memory (LSTM) layers, standard LSTM layers, and convolutional layers. The fusion of convolutional techniques and the LSTM network enables the proposed DL approach to better capture the spatio-temporal characteristics and correlations of explanatory variables. A tailored spatially aggregated random forest is employed to rank the importance of the explanatory variables. The ranking is then used for feature selection. The proposed DL approach is applied to the short-term forecasting of passenger demand under an on-demand ride service platform in Hangzhou, China. Experimental results, validated on real-world data provided by DiDi Chuxing, show that the FCL-Net achieves better predictive performance than traditional approaches including both classical time-series prediction models and neural network based algorithms (e.g., artificial neural network and LSTM). This paper is one of the first DL studies to forecast the short-term passenger demand of an on-demand ride service platform by examining the spatio-temporal correlations.
1. Introduction
Short-term passenger-demand forecasting supports regional driver incentives and improved vehicle utilization, but must jointly address temporal, spatial, and exogenous dependences. The paper proposes FCL-Net with feature selection and evaluates it on DiDi Chuxing data.
- Forecasting demand across spatial zones helps platforms incentivize drivers toward high-demand regions and improve registered-car utilization.
- Regional supply-demand mismatches create oversupply or overfull demand, reducing vehicle utilization or passenger satisfaction.
- Short-term forecasting is challenging because demand exhibits temporal, spatial, and exogenous dependences, including periodicity, neighboring-zone influence, and weather effects.
- FCL-Net coordinates spatio-temporal and non-spatial time-series variables in one end-to-end trainable architecture.
- A spatially aggregated random forest evaluates explanatory-variable importance before feature selection.
- Using DiDi Chuxing data, the proposed deep-learning structure outperforms five benchmark algorithms spanning conventional time-series and classical deep-learning methods.
2. Literature Review
Prior work models taxi-market equilibrium, demand forecasting, and transportation prediction with statistical, neural, and deep-learning methods. The review highlights persistent challenges in capturing spatial and temporal correlations simultaneously.
- Regional disequilibrium can produce excess vacant taxis or waiting passengers, causing resource mismatch and uneven taxi availability.
- Passenger-demand forecasting research includes ARIMA, Markov, Lempel-Ziv-Welch, neural-network, ensemble, spatial-statistical, and socioeconomic approaches.
- Traffic-forecasting literature has shifted from classical statistical models toward neural networks as data accessibility and computing power increased.
- CNNs capture local and global spatial correlations, while recurrent networks model temporal correlations but traditional RNNs suffer vanishing gradients.
3. Preliminaries
The paper formulates demand forecasting over spatial grids and hourly intervals using demand, travel-time, calendar, and weather variables. It distinguishes variables with spatial-temporal variation from purely temporal covariates.
- The urban area is partitioned into I × J uniform grids, with variables aggregated over one-hour intervals.
- Demand intensity is the number of orders in a grid during a time interval and is represented across all grids by matrix D_t ∈ R^I×J.
- Average travel-time rate is defined as travel time per unit distance and aggregated across grids in matrix Γ_t ∈ R^I×J.
- Time-of-day is categorized into peak, off-peak, and sleep hours, while day-of-week distinguishes weekdays from weekends.
- Weather inputs include temperature, humidity, weather state, wind speed, and visibility; weather state has five ordered categories.
- The forecasting problem uses historical demand, travel-time, and weather observations plus known current time-of-day and day-of-week information to predict D_t.
4. Methodology
The methodology combines convolutional and recurrent components to model spatial, temporal, and exogenous dependences in short-term passenger demand forecasting. FCL-Net integrates separate processing paths for spatio-temporal and non-spatial time-series variables within one architecture.
- FCL-Net is proposed to capture spatial, temporal, and exogenous dependences in short-term passenger demand forecasting.
- The approach uses a spatially aggregated random forest to rank explanatory-variable importance and select relevant features.
- LSTM and Conv-LSTM: Conv-LSTM combines CNN and LSTM in an end-to-end architecture, using convolutions to explore spatial local correlations.Convolutional filters are shared across tensors, and zero padding maintains spatial dimensions.
- Fusion convolutional LSTM: FCL-Net integrates spatio-temporal variables with non-spatial time-series variables, using Conv-LSTM for the former and LSTM for the latter.Repeating and transformation functions fuse the two categories of variables.
- Spatio-temporal variables: Historical demand intensity and travel time rate are processed through separate stacked Conv-LSTM architectures because they affect future demand differently.Separate high-level components are produced before fusion.
- Non-spatial time-series variables: Time-of-day, day-of-week, and weather variables are represented as sequences and processed by separate stacked LSTM architectures.Weather variables include temperature, humidity, weather state, wind speed, and visibility.
15: end procedure
The paper develops a spatially aggregated random forest to rank explanatory-variable importance for passenger-demand forecasting and select categories and look-back windows.
- Importance calculation: Random-forest feature importance is calculated from the average out-of-bag error difference after permuting a variable.The procedure permutes the jth variable, recalculates out-of-bag error, and averages the difference across trees.
- Spatial aggregated random forest: Spatially aggregated random forest uses I × J standard random forests because demand is represented as an I × J matrix.It examines aggregated variable importance across spatial locations.
- Feature selection: The selected spatial-temporal variables include demand intensity and travel time rate, while non-spatial variables include time-of-day and weather-related categories.The formulation distinguishes spatial-temporal variables from non-spatial time-series variables and stores their importance separately.
- Feature selection: Variable importance is analyzed by category and look-back time window to select important variables and suitable windows for each category.Importance values are normalized to percentages before comparison.
5. Experiments and Results
Experiments on DiDi Chuxing data from Hangzhou examine demand patterns, spatial-temporal correlations, and variable importance before comparing forecasting configurations.
- Data and demand patterns: Weekday demand has a double-peak pattern, whereas weekend demand has a single peak.The study defines peak, off-peak, and sleep hours separately for weekdays and weekends.
- Spatio-temporal correlations: Average correlations decrease with spatial distance but remain strong, while shorter look-back intervals generally have higher correlations.Longer look-back variables remain correlated with the demand intensity to some extent.
- Feature importance: Travel time rate and demand intensity are the dominant variable categories, followed by time-of-day and temperature; day-of-week and humidity contribute less than 5%.The time-of-day at time t is the most important individual variable, followed by demand intensity and travel time rate at t − 1.
- Feature selection: Selecting variable categories and look-back windows improves computation efficiency with little loss of predictive performance.The selected configuration uses four categories with look-back windows of 4, 8, 2, and 2, respectively, and reduces computational complexity by 29.5%.
(1) Conv-LSTM with only historical demand intensity: this model utilizes his-
The experiments compare Conv-LSTM, FCL-Net variants, traditional time-series models, and neural-network baselines for grid-level demand forecasting. FCL-Net performs best while feature selection substantially reduces complexity with small metric changes.
- Evaluation: The evaluation uses RMSE, R2, and MAE to compare forecasting models on the test set.Ground-truth and estimated demand intensities are used in the metric formulations.
- Model comparison: FCL-Net outperforms the other tested methods, including traditional time-series models and neural-network benchmarks.The comparison includes HA, MA, ARIMA, ANN, LSTM, Conv-LSTM, and FCL-Net variants.
- Model comparison: 50.9% lower RMSE is achieved by both FCL-Net variants than by Conv-LSTM using only historical demand intensity.The result indicates that exogenous variables contribute to short-term passenger-demand forecasting.
- Feature selection: 29.5% lower computational complexity is obtained after feature selection, reducing variables per observation from 840 to 592.The selected-variable FCL-Net shows a 0.6% RMSE, 0.1% R2, or 1.1% MAE decrease in predictive performance.
- Forecast visualization: FCL-Net visualization primarily captures spatio-temporal demand characteristics and produces more accurate forecasts in sampled heat maps.The visualizations show higher demand during peak hours and stronger demand in central grids.
6. Conclusions
The paper proposes FCL-Net for short-term passenger-demand forecasting, combining convolutional and LSTM components with explanatory-variable selection. On real-world DiDi Chuxing data, the FCL-Nets outperform benchmark methods, while feature selection reduces dimensionality with little predictive loss.
- Proposed approach: FCL-Net fuses conv-LSTM layers, LSTM layers, and convolutional operators to model historical demand and exogenous explanatory variables.The inputs include travel time rate, time-of-day, day-of-week, and weather conditions.
- Feature selection: A spatially aggregated random forest ranks explanatory-variable importance, and the ranking supports feature selection for two FCL-Net variants.The variants use either full variables or selected variables.
- Predictive performance: The two FCL-Nets significantly outperform five traditional and neural-network benchmarks on DiDi Chuxing data, measured by RMSE, R-square, and MAE.The benchmarks include HA, MA, ARIMA, ANN, and LSTM.
- Predictive performance: 50.9% lower RMSE than conv-LSTM is achieved by both FCL-Nets, which use more than historical passenger demand as explanatory variables.The comparison is against conv-LSTM with only passenger demand as variables.
- Feature selection: 0.6% RMSE loss accompanies nearly 30% lower variable dimension after feature selection, indicating reduced computational complexity with little accuracy loss.The paper concludes that appropriate feature selection balances computation complexity and predictive performance.
- Scope and future work: The study addresses short-term demand forecasting for on-demand ride services and identifies a need for further analysis of endogenous and exogenous-variable relationships.The application is framed around understanding complex interactions in the on-demand ride service market.