Transportation Research Part C 85 (2017) 591–608

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Short-term forecasting of passenger demand under on-demand ride MARK

services: A spatio-temporal deep learning approach
⁎
Jintao Kea, Hongyu Zhengb, Hai Yanga, Xiqun (Michael) Chenb,
a
Department of Civil and Environmental Engineering, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, China
b
College of Civil Engineering and Architecture, Zhejiang University, Hangzhou, China

AR TI CLE I NF O AB S T R A CT

Keywords: Short-term passenger demand forecasting is of great importance to the on-demand ride service
On-demand ride services platform, which can incentivize vacant cars moving from over-supply regions to over-demand
Short-term demand forecasting regions. The spatial dependencies, temporal dependencies, and exogenous dependencies need to
Deep learning (DL) be considered simultaneously, however, which makes short-term passenger demand forecasting
Fusion convolutional long short-term memory
challenging. We propose a novel deep learning (DL) approach, named the fusion convolutional
network (FCL-Net)
Long short-term memory (LSTM)
long short-term memory network (FCL-Net), to address these three dependencies within one end-
Convolutional neural network (CNN) 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. The experimental results, validated on the real-world data provided by DiDi
Chuxing, show that the FCL-Net achieves the better predictive performance than traditional
approaches including both classical time-series prediction models and state-of-art machine
learning algorithms (e.g., artificial neural network, XGBoost, LSTM and CNN). Furthermore, the
consideration of exogenous variables in addition to the passenger demand itself, such as the
travel time rate, time-of-day, day-of-week, and weather conditions, is proven to be promising,
since they reduce the root mean squared error (RMSE) by 48.3%. It is also interesting to find that
the feature selection reduces 24.4% in the training time and leads to only the 1.8% loss in the
forecasting accuracy measured by RMSE in the proposed model. 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

The on-demand ride service platform, e.g., Urber, Lyft, DiDi Chuxing, is an emerging technology with the boom of the mobile
internet. Ride-sourcing or transportation network companies (TNCs) refer to an emerging urban mobility service mode that private
car owners drive their own vehicles to provide for-hire rides (Chen et al., 2017). On-demand ride-sourcing services can be completed
via smart phone applications. The platform serves as a coordinator who matches requesting orders from passengers (demand) and

⁎
Corresponding author.
E-mail address: [email protected] (X.M. Chen).

http://dx.doi.org/10.1016/j.trc.2017.10.016
Received 20 June 2017; Received in revised form 15 October 2017; Accepted 15 October 2017
Available online 24 October 2017
0968-090X/ © 2017 Elsevier Ltd. All rights reserved.
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

vacant registered cars (supply). There exists an abundance of leverages to influence drivers’ and passengers’ preference and behavior,
and thus affect both the demand and supply, to maximize profits of the platform or achieve the maximum social welfare. Having a
better understanding of the short-term passenger demand over different spatial zones is of great importance to the platform or the
operator, who can incentivize drivers to the zones with more potential passenger demands, and improve the utilization rate of the
registered cars.
Although limited research efforts have been implemented on forecasting short-term passenger demand under the emerging on-
demand ride service platform in most recent years mainly due to the real-world data unavailability, the fruitful studies on the taxi
market can provide valuable insights since there exist strong similarities between the taxi market and the on-demand ride service
market. A series of mathematical models were developed to spell out endogenous relationships among variables in the taxi market (;
Yang et al., 2002, 2005, 2010b; Yang and Yang, 2011) under the two-sided market equilibrium. On the demand side, the accurate
passenger demand was affected by passengers’ waiting time and taxi fare; while on the supply side, drivers’ behavior, i.e., how to find
a passenger, was mainly affected by the expected searching time and taxi fare. The passenger demand was endogenously determined
when the taxi operator decided the taxi fare structure and the number of released licenses of taxis (entry limitation).
In theory, the equilibrium between the demand and supply will eventually be reached when the arrival rate of passengers equals
to the arrival rate of vacant taxis and equals to the meeting rate. However, heterogeneous and exogenous factors in reality, e.g., the
asymmetric information, and short-term fluctuations, may make it difficult to guarantee the spatial distribution of taxis matching the
passenger demand all the time (Moreira-Matias et al., 2013). Hence, disequilibrium states can result from the following two scenarios:
oversupply (an excess in the number of vacant taxis may decrease the taxi utilization) and overfull demand (excessively waiting
passengers may lower the degree of satisfaction). Both scenarios are harmful to the taxi operator as well as the on-demand ride
service platform, raising a strong need for a precise forecasting of short-term passenger demand. It helps the operator/platform
implement proactive incentive mechanism, such as surge pricing and cash/point awards, to attract drivers from regions of oversupply
to regions with overfull demand. These strategies not only shorten the process of reaching equilibrium under a dynamic environment
but also help improve the taxi/car utilization rate and reduce passengers’ waiting time.
However, short-term forecasting of passenger demand or on-demand ride services in each region is of great challenge mainly due
to the three kinds of dependencies (Zhang et al., 2017b):

(1) Time dependencies: the passenger demand has a strong periodicity (for example, the passenger demand is expected to be high
during morning and evening peaks and to be low during sleeping hours); furthermore, the short-term passenger demand is
dependent on the trend of the nearest historical passenger demand.
(2) Spatial dependencies: Yang et al. (2010b) revealed that the passenger demand in one specific zone was not merely determined by
the variables of this zone, but endogenously dependent on all the zonal variables in the whole network. Generally, the variables of
the nearby zones have stronger influences than distant zones, which inspires the need for an advanced model that can capture
local spatial dependencies.
(3) Exogenous dependencies: some exogenous variables, such as the travel time rate and weather conditions, may have strong
influences on the short-term passenger demand. The exogenous variables also demonstrate time dependencies and spatial de-
pendencies.

Although little direct experience suggests solutions to these three dependencies in short-term passenger demand forecasting,
studies on traffic speed/volume prediction and rainfall nowcasting provide valuable insights (Ghosh et al., 2009; Huang and Sadek,
2009; Guo et al., 2014; Wang et al., 2014). Recently, deep learning (DL) approaches have been successfully used for traffic flow
prediction. For example, Ma et al. (2015a) employed the long short-term memory (LSTM) neural network to capture the long-term
dependencies and nonlinear traffic dynamics for short-term traffic speed prediction. Zhang et al. (2017b) presented a deep spatio-
temporal residual network to predict the inflow and outflow in each region of a city simultaneously. Shi et al. (2015) innovatively
integrated CNN and LSTM in one end-to-end DL structure, named the convolutional LSTM (conv-LSTM), which provided a brand-new
idea for solving spatio-temporal sequence forecasting problems. In that research, the numerical experiments showed that the conv-
LSTM outperformed fully connected LSTM in two datasets.
In this paper, we propose a novel DL structure, named the fusion convolutional LSTM network (FCL-Net), to consider the three
dependencies simultaneously in the short-term passenger demand forecasting for the on-demand ride service platform. Different from
the aforementioned studies, this structure coordinates the spatio-temporal variables and non-spatial time-series variables in one end-
to-end trainable model. Before feeding these explanatory variables into the DL structure, a tailored spatial aggregated random forest
is designed to evaluate the feature importance with different categories, look-back time intervals, and spatial locations.
To the best knowledge of the authors, this paper is one of the first attempts to employ spatio-temporal DL approaches in short-
term passenger demand forecasting under the on-demand ride service platform. The main contributions of this paper are within three
folds:

(1) The novel FCL-Net approach characterizes the spatio-temporal properties of the spatio-temporal predictors, captures the temporal
features of non-spatial time-series variables simultaneously, and coordinates them in one end-to-end learning structure for the
short-term passenger demand forecasting.
(2) We propose a spatial aggregated random forest to extract the potential predictors affecting short-term passenger demand and
assess the feature importance of them.
(3) Validated by the real-world on-demand ride services data provided by DiDi Chuxing in a large-scale urban network, the proposed

592
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

DL structure outperforms both the traditional and state-of-the-art benchmark algorithms, including three conventional time-
series models and several efficient machine learning/deep learning approaches.

The rest of the paper is organized as follows. Section 2 first reviews the existing research on the taxi market modeling and taxi-
passenger demand forecasting, and then summarizes the state-of-the-art DL approaches utilized in related problems. Section 3 for-
mulates the short-term traffic flow forecasting problem, and explicitly explains the explanatory variables. Section 4 describes the
structure and mathematical formulation of the proposed FCL-Net, as well as the proposed spatial aggregated random forest algorithm
for feature selection. Section 5 compares the predictive performance between the proposed approach and the benchmark models
based on the real-world dataset extracted from DiDi Chuxing. Finally, Section 6 concludes the paper and outlooks future research.

2. Literature review

The fast-growing technology of mobile internet enables on-demand ride service platforms for providing efficient connections
between waiting passengers and vacant registered cars. Like the taxi market, the on-demand ride service market is essentially a two-
sided market where both the consumers (passengers) and providers (drivers of vacant cars) are independent and have individual
mode choices. The passengers make mode choice decisions between taxi/on-demand ride service and public transportation according
to the waiting time and trip fare, while the drivers make service decisions by considering the searching time and trip fare. In light of
the fact that the vacant taxis and waiting passengers are unable to be matched simultaneously in a specific zone, Yang et al. (2002,
2010b) and Yang and Yang (2011) proposed a meeting function to characterize the search frictions between drivers of vacant taxis
and waiting passengers. The meeting function pointed out that the meeting rate in one specific zone was determined by the density of
the waiting passengers and vacant taxis at that moment, which indicated that passengers’ waiting time, drivers’ searching time, and
passengers’ arrival rate (demand) were endogenously correlated. The equilibrium state was reached when the arrival rate of waiting
passengers exactly matched the arrival rate of vacant taxis. This equilibrium state along with the endogenous variables were in-
fluenced by the exogenous variables, such as the taxi fleet size and taxi trip fare. The taxi operator might coordinate supply and
demand via the on-demand service platform and thus influence the equilibrium state by regulating the entry of taxi and determining
the taxi fare structure, such as non-linear pricing (Yang et al., 2010a). Apart from the traditional taxi market, some emerging market
structures, like the ride-sourcing market (Zha et al., 2016), e-hailing taxi market (He and Shen, 2015; X. Wang et al., 2016), were
examined under the same equilibrium modeling framework. Recently, the on-demand ride-sharing market and the optimal assign-
ment strategies have also attracted researchers’ attentions(Alonso-Mora et al., 2017).
However, researchers found that a regional disequilibrium occurred when there was an excess in vacant taxis or waiting pas-
sengers in that region (Moreira-Matias et al., 2012). This disequilibrium might lead to a resource mismatch between supply and
demand, which resulted in the low taxi utilization in some regions while low taxi availability in other regions. Therefore, a short-term
passenger demand forecasting model is of great importance to the taxi operator, which can implement efficient taxi dispatching and
time-saving route finding to achieve an equilibrium across urban regions (Zhang et al., 2017a). To attain the accurate and robust
short-term passenger demand forecasting, both parametric (e.g., ARIMA) and non-parametric models (e.g., neural network) have
been examined. For instance, Zhao et al. (2016) implemented and compared three models, i.e., the Markov algorithm, Lempel-Ziv-
Welch algorithm, and neural network. In that research, the results showed that neural network performed better with the lower
theoretical maximum predictability while the Markov predictor had better performance with the higher theoretical maximum pre-
dictability. Moreira-Matias et al. (2013) proposed a data stream ensemble framework which incorporated time varying passion model
and ARIMA, to predict the spatial distribution of taxi passenger demand. Deng and Ji (2011) employed the global and local Moran’s I
values to evaluate the intensity of taxi services in Shanghai. Some socio-demographical and built-environment variables have also
been in use for predicting taxi passenger demand (Qian and Ukkusuri, 2015).
There are a broad range of problems in the domain of transportation, which are similar to short-term passenger demand fore-
casting. These problems include the traffic speed estimation (Bachmann et al., 2013; Soriguera and Robusté, 2011; Wang and Shi,
2013), traffic volume prediction (Boto-Giralda et al., 2010), real-time crash likelihood estimation (Ahmed and Abdel-Aty, 2013; Yu
et al., 2014), car-following behavior prediction(Wang et al., 2017), original-destination matrices forecasting (Toqué et al., 2016), bus
arrival time prediction (Yu et al., 2011), short-term forecasting of high speed rail demand (Jiang et al., 2014), and etc., the solutions
to which offer meritorious inspirations to our problem. To solve these spatio-temporal forecasting problems, a broad range of ap-
proaches have been proposed, including the ARIMA family (Zhang et al., 2011; Khashei et al., 2012), local regression model
(Antoniou et al., 2013), neural network based algorithms (Chan et al., 2012), and Bayesian inferring approaches (Fei et al., 2011).
Vlahogianni et al. (2014) reviewed the existing literature on short-term traffic forecasting, and observed that researchers were
moving from classical statistical models to neural network based approaches with the explosive growth of data accessibility and
computing power.
Recently, more and more DL algorithms have been utilized in traffic prediction due to their capability of capturing complex
relationship from a huge amount of data. Cheng et al. (2016) proposed a DL based approach to forecast day-to-day travel demand
variations in a large-scale traffic network. Huang et al. (2014) predicted short-term traffic flow via a two-layer DL structure with a
deep belief network (DBN) at the bottom and a multitask regression model (MTL) at the top. Polson and Sokolov (2017) found that
the sharp nonlinearities of traffic flow, as a result of transitions between the free flow, breakdown, recovery and congestion, could be
captured by a DL architecture. Combining the empirical mode decomposition (EMD) and back-propagation neural network (BPN),
Wei and Chen (2012) presented a hybrid EMD-BPN method for short-term passenger flow forecasting. Graphical LASSO was also
combined in the neural network, showing its potential in network-scale traffic flow forecasting (Sun et al., 2012). Lv et al. (2015)

593
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

stated that a stacked autoencoder model helped to capture generic traffic flow features and characterize spatial temporal correlations
in traffic flow prediction. Ma et al. (2015b) extended the deep learning theory for the large-scale traffic network analysis, and
predicted the evolution of traffic congestion with the help of taxi GPS data.
One of the obstacles in traffic forecasting is how to capture spatio-temporal correlations. It was found that the vehicle accu-
mulation and dissipation had impacts on the travel volume of adjacent links or intersections, which indicated the spatial correlations
should be considered in forecasting (Zhu et al., 2014). In terms of the spatial correlations, CNN developed by (LeCun et al., 1999) was
used to learn the local and global spatial correlations in large-scale, network-wide traffic forecasting (Chen et al., 2016). To address
temporal correlations (another inherit property in real-time traffic forecasting), the family of recurrent neural networks (RNN)
(Williams and Zipser, 1989) was widely viewed as one of the most suitable structures (Zhao et al., 2017). In the RNN architecture, the
dependent variable in one timestamp was not only dependent on the explanatory variables in this timestamp, but also correlated with
the explanatory variables in the previous timestamps (Sutskever et al., 2009). However, the traditional RNN suffered from a “van-
ishing gradience” effect which made it impossible to store long-term information (Hochreiter, 1991). To address this issue, Hochreiter
and Schmidhuber (1997) presented the long short-term memory (LSTM) which employed a series of memory cells to store in-
formation for exploring long-range dependencies in the data.
However, neither CNN nor LSTM are perfect models for spatio-temporal forecasting problems. CNN fails to capture the temporal
dependencies while LSTM is incapable of characterizing local spatial correlations. To capture spatial and temporal dependencies
simultaneously in one end-to-end training model, researchers have made numerous attempts in recent years. J. Wang et al. (2016)
proposed a novel error-feedback recurrent convolutional neural network (eRCNN) architecture which was comprised of the input
layer, the convolutional layer, and the error-feedback recurrent layer. Zhang et al. (2017b) modeled the temporal closeness, period,
and trend properties of the inflow/outflow of human mobilities with serval separate convolutional layers and then fused these layers
in one end-to-end DL structure. Ma et al. (2017) proposed a deep convolutional neural network for large-scale traffic network speed
prediction, where the space-time matrix was converted to an image as the input of the CNN. Yu et al. (2017) designed a spatio-
temporal recurrent convolutional network for predicting network-wide traffic speeds. This model was comprised of two components:
the lower one modeled the spatial features while the upper one learned the temporal features. Shi et al. (2015) proposed a conv-LSTM
network, which combined CNN and LSTM in one sequence to sequence learning framework, for precipitation nowcasting that was a
typical spatio-temporal forecasting problem. This network structure was different from the previous attempts to combine CNN and
LSTM, as it redefined the mathematical formulations of LSTM by transferring the vector-based LSTM cell to a tensor-based LSTM cell
(which enabled convolutional operations), instead of merely stacking CNN and LSTM. In that research, the results showed that the
conv-LSTM outperformed the fully-connected LSTM, since some complicated spatio-temporal characteristics could be learned by the
convolutional and recurrent structure of the model.
However, short-term passenger demand is not only dependent on its own spatio-temporal properties but also dependent on other
explanatory variables (some with spatio-temporal properties and some only with temporal properties). Therefore, we extend the
conv-LSTM to a more generalized framework which is comprised of two categories of structures, one of which characterizes spatio-
temporal features while the other captures non-spatial temporal features. In this way, the aforementioned three dependencies
(spatial, temporal, and exogenous) can be addressed at the same time.

3. Preliminaries

The short-term passenger demand forecasting is essentially a time-series prediction problem, which implies that the nearest
historical passenger demand can provide valuable information for predicting the future demand. We also observe that the travel time
rate influences the short-term passenger demand, since it reflects the congestion level of trips and zones. For example, passengers will
potentially transfer to the metro transit if they find the trips to their destinations are congested. Furthermore, the attributes of time-
of-day, day-of-week, and weather conditions also have impacts on the short-term passenger demand.
In this section, we first interpret the notations of the variables used in this paper, then give an explicit definition of the short-term
passenger demand forecasting problem.
Definition 1 (Region and time partition). The urban area is partitioned into I × J grids uniformly where each grid refers to a zone. On
the other hand, we consider variables aggregated in a one-hour time interval in this paper.Based on Definition 1, we explicitly define
several categories of variables as follows:

(1) Demand intensity

The intensity of demand at the tth time slot (e.g., hour) lying in grid (i,j ) is defined as the number of orders during this time
interval within the grid, which is denoted by dti,j . The intensity of demand in all I × J grids at the tth time slot is defined as the
matrix Dt ∈ RI×J (R refers to the real set), where the (i,j ) th element is (Dt )i,j = dti,j .
(2) Average travel time rate
The travel time rate represents the travel time per unit travel distance (Chen et al., 2017). In this paper, the travel time rate of the
mth order originating from grid (i,j ) at the tth time slot, is defined as the ratio of its travel time to its travel distance, τti,,mj . The
average travel time rate in grid (i,j ) during the tth time slot, τti,j , is defined as the average of τti,,mj over m. The average travel time
rate in all I × J grids at the tth time slot is defined as the matrix Γt ∈ RI × J , where the (i,j ) th element is (Γt )i,j = τti,j .
(3) Time-of-day and day-of-week
By empirically examining the distribution of demand intensity with respect to time in the training set, 24 h in each day can be

594
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

intuitively divided into 3 periods: peak hours, off-peak hours, and sleep hours. We simply rank the hours based on the empirical
demand intensity, and define the top 8 h, middle 8 h, bottom 8 h, as the peak hours, off-peak hours and sleep hours. We further
introduce the dummy variable ht to characterize this attribute of time-of-day, given by

⎧ 2, if t belongs to peak hours

ht = 1, if t belongs to off-peak hours
⎨
⎩ 0, if t belongs to sleep hours

We also denote another dummy variable wt to be the day-of-week, which catches up the distinguished properties between
weekdays and weekends.

0, if t belongs to weekdays
wt = ⎧
⎨
⎩ 1, if t belongs to weekends

(4) Weather
We consider 5 categories of weather variables, including temperature (measured by Celsius degree), relative humidity (measured
by percentage), weather state, wind speed (measured by mile per hour), and visibility (measured by kilometers). The weather
state variable contains 5 categories: sunny (5), cloudy (4), light rain (3), moderate rain (2), and heavy rain (1). The weather state
takes one value for one hour, while temperature, humidity, weather state, wind speed, and visibility are taken average in each
hour. In this paper, the temperature, humidity, weather state, wind speed, and visibility during the tth time interval are denoted
as att ,aht ,ast ,awt ,avt , respectively.
All of the aforementioned variables demonstrate the time-varying attributes, but the demand intensity and average travel time
rate show the zonal-based attributes, which mean they have different values across grids. The variables with time-varying at-
tributes only have temporal dependencies, while the variables with time-varying and zonal-based attributes have both spatial and
temporal dependencies, which implies that they should be treated in different ways. Thus, we give the definition of the spatio-
temporal variables and non-spatial time-series variables in Definition 2.
Definition 2 (Spatio-temporal variables). Refer to the variables showing distinction across time and across space, which imply there
exist spatio-temporal correlations, e.g., the demand intensity, and travel time rate. Other variables, including the time-of-day, day-of-
week, and weather variables, are denoted as non-spatial time-series variables, which vary across time instead of space.With the
aforementioned definition of the explanatory variables, we can formulate the short-term passenger demand forecasting as Problem 1.
Problem 1. Given the historical observations and pre-known information {Ds ,Γs |s =
0,…,t −1;hs,ws |s = 0,…,t;ats,ahs,ass,aws,avs |s = 0,…,t −1} , predict Dt. It is noteworthy that the time-of-day and day-of-week of the tth
time slot (ht and wt ) are pre-known at t.

4. Methodology

In this paper, we propose a novel DL architecture, i.e., FCL-Net, to capture the spatial dependencies, temporal dependencies, and
exogenous dependencies, in short-term passenger demand forecasting. To reduce the computation complexity, we also present a
spatial aggregated random forest algorithm to rank the importance of explanatory variables and select the important ones. In this
section, we first present a brief review of the traditional LSTM and conv-LSTM, then introduce the proposed architecture and training
algorithm of FCL-Net, and finally illustrates the proposed spatial aggregated random forest.

4.1. LSTM and Conv-LSTM

The traditional artificial neural network (ANN) lacks the ability to catch up time-series characteristics since it does not take the
temporal dependencies into consideration. To overcome this shortcoming, RNN is proposed, where the connection between units is
organized by timestamps. The inner structure of an RNN layer is illustrated in Fig. 1, where the input is a T time-stamp vector
sequence x = (x1,x2,…,xT ) and the output is a hidden vector sequence h = (h1,h2,…,hT ) . It is noteworthy that xt can be a one-di-
mensional vector or scalar, while ht does not necessarily have the same dimension as xt . The hidden unit value in timestamp t, i.e., ht ,

Fig. 1. Illustration of the inner structure of an RNN layer.

595
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

stores the information, including hidden values (h1,h2,…,ht − 1) and input values (x1,x2,…,xt − 1) , of the previous timestamps. Together
with the input in t, i.e., xt , it is passed to the next timestamp t + 1 at each iteration. In this way, RNN can memorize the information
from multiple previous timestamps. Although RNN exhibits strong ability in catching temporal characteristics, it fails to store in-
formation for a long-term memory.
LSTM, as a special RNN structure, overcomes RNN’s weakness on the long-term memory. Like the standard RNN, each
LSTM cell maps the input vector sequence x to a hidden vector sequence h by T iterations. As demonstrated in Eqs. (1)–(5),
it ,f t,ot ,ct ,(t = 1,2,…,T ) represent the input gate, forget gate, output gate, and memory cell vectors, respectively, sharing the same
dimension with ht .
it = σ (Wxi xt + Whi ht − 1 + Wci ∘ct − 1 + bi ) (1)

f t = σ (Wxf xt + Whf ht − 1 + Wcf ∘ct − 1 + bf ) (2)

ct = f t ∘ct − 1 + it ∘tanh(Wxc xt + Whc ht − 1 + bc ) (3)

ot = σ (Wxo xt + Who ht − 1 + Wco ∘ct + bo) (4)

ht = ot ∘tanh(ct ) (5)
The operator ‘∘’ refers to Hadamard product, which calculates the element-wise products of two vectors, matrices, or tensors with
the same dimensions. σ and tanh are the two non-linear activation functions given by
1
σ (x ) =
1 + e−x (6)

e x −e−x
tanh(x ) =
e x + e−x (7)
Wcf ,Wci,Wco,Wxi,Whi,Wxf ,Whf ,Wxc,Whc,Wxo,Who are the weighted parameter matrices which conduct a linear transformation from the
vector of the first subscript to the second subscript, while bi,bf ,bc,bo are the intercept parameters.
Multiple LSTM cells can be stacked to form a deeper and more complicated neural network, which can better discover the
complex relationships between the inputs and outputs. In this paper, each LSTM cell is denoted as a function F L: RT × L → RT × L′, where
T is the length of time sequences, L is the length of one input vector, and L′ is the length of one output vector (see Fig. 2).
However, LSTM is not an ideal model for the passenger demand forecasting with spatial and temporal characteristics in this paper,
because it fails to capture the spatial dependencies. To overcome this shortcoming, the conv-LSTM network, which combines CNN
and LSTM in one end-to-end DL architecture, is proposed.
The core idea of conv-LSTM is to transform all the inputs, memory cell values, hidden states, and various gates in Eqs. (1)–(5) to
3D tensors (shown in Eqs. (8)–(12)).
It = σ (Wxi ∗Xt + Whi ∗Ht − 1 + Wci ∘Ct − 1 + bi ) (8)

Ft = σ (Wxf ∗Xt + Whf ∗Ht − 1 + Wcf ∘Ct − 1 + bf ) (9)

Ct = Ft ∘Ct − 1 + It ∘tanh(Wxc ∗Xt + Whc ∗Ht − 1 + bc ) (10)

Ot = σ (Wxo ∗Xt + Who ∗Ht − 1 + Wco ∘Ct + bo) (11)

Ht = Ot ∘tanh(Ct ) (12)
The input tensors, hidden tensors, memory cell tensors, input gate tensors, output gate tensors, and forget gate tensors are denoted

Fig. 2. Illustration of the inner structure of an LSTM layer.

596
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Fig. 3. Illustration of the transformation function.

as Xt ,Ht ,Ct ,It ,Ot ,Ft ∈ RM × N × L , respectively, where M ,N are spatial dimensions (M rows and N columns of the grids), and L refers to
the length of one input feature vector in each grid. The operator ’∗’ stands for the convolutional operator. Here,
Wxf ,Whf ,Wxc,Whc,Wxo,Who serve as convolutional flitters, which are replicated across the tensors with shared weights, and thus explore
spatially local correlations. bi,bf ,bc,bo are tensors of intercept parameters, the dimension of which is consistent with the left-hand-side
(for example, the dimension of bi equals to that of It ). The boundary grids of the map may not have neighbor grids for convolutional
operation in some directions, thus we use zero padding techniques which creates some virtual neighbor grids and fills them with zero
value. In this way, the spatial dimensions (rows and columns) of the input and output of a convolutional unit are consistent.
Through these T iterations, each conv-LSTM layer can map a sequence of input tensors X = (X1,X2,…,XT ) to a sequence of hidden
tensors H = (H1,H2,…,HT ) . In this paper, each conv-LSTM cell is denoted as a function F CL: RT × M × N × L → RT × M × N × L′, where T is the
length of time sequences, M ,N refer to dimensions of rows and columns, L′ represents the length of an output feature vector in each
grid respectively. Similar to LSTM, multiple conv-LSTM layers can be stacked to build up a deep conv-LSTM neural network.
However, the spatio-temporal variables, such as demand intensity and travel time rate during one time interval, are 2D matrices
(see Definition 1 and 2), thus a transformation function F T : RM × N → RM × N × 1 is employed to transfer the initial input matrices into 3D
tensors by simply adding one dimension (see Fig. 3).

4.2. Fusion convolutional LSTM (FCL-Net)

In this section, we propose a novel fusion convolutional LSTM network (FCL-Net), which integrates spatio-temporary variables and
non-spatial time-series variables into one DL architecture for short-term passenger demand forecasting under the on-demand ride
service platform. The structure of the proposed FCL-Net is illustrated in Fig. 4. Conv-LSTM layers and convolutional operators are
employed to capture characteristics of spatio-temporary variables, while LSTM layers are implemented for non-spatial time-series
variables. To fuse these two categories of variables, techniques including repeating and transformation functions, are utilized in the
structure. The repeating function is denoted as F R (·;M ,N ): R → RM × N × 1, where (F R (x ;M ,N ))m,n,1 = x for any
m ∈ (1,2,…,M ),n ∈ (1,2,…,N ) . It is also worth mentioning that the transformation function F T should be applied to transfer 2D
matrices Dt ,Γt to 3D tensors Dt ,Γt ∈ RI × J × 1 to meet the consistent requirement of convolutional operators.
As mentioned in Problem 1, the forecasting target is the demand intensity during the tth time interval, which is denoted as
Xt = Dt .

Fig. 4. Framework of the proposed FCL-Net approach.

597
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

4.2.1. Structure for spatio-temporary variables

Among the variables utilized in this paper, the historical demand intensity and travel time rate are spatio-temporary variables, as
denoted in Definition 2. By considering that the historical demand intensity and travel time rate influence the future demand
intensity in different ways, these two kinds of variables are fed into two separate architectures, each of which consists of a series of
stacked conv-LSTM layers and convolutional operators. Suppose K d,K τ are the look-back time windows, Ld ,L τ are the number of
stacked conv-LSTM layers, of the demand intensity and travel time rate, respectively, the formulations of the architecture for spatio-
temporal variables are given as Eq. (13)–(16). The look-back time window refers to the number of previous hours taken as features for
each sample, and equals to the input time step for each sample in a LSTM-based structure.

(Ut(−LK) ,Ut(−LK) +1,…,Ut(−L1) ) = F LCL···FlCL···F1CL (Dt−K ,Dt−K +1,…,Dt−1)

d
d
d
d
d
d d d (13)
u
X t ̂ = σ (W ux ∗Ut(−Ld1) + bu ) (14)

(Vt(−LK) ,Vt(−LK) +1,…,Vt(−L1) ) = F LCL···FlCL···F1CL (Γt−K ,Γt−K +1,…,Γt−1)

τ
τ
τ
τ
τ
τ τ τ (15)
v
X t ̂ = σ (W vx ∗Vt(−L1τ ) + bv ) (16)
where Ut(−Ldk),k
= 1,2,…,K d,Vt(−Ldτ ),k
= 1,2,…,K τ are the output hidden tensors in the highest-level layers of the architectures of demand
and travel time rate, respectively. W ux,W vx are convolutional operators utilized to further capture the spatial correlations of the
highest-level output tensors, while bu,b v are the intercept parameters. Through these two structures, two high-level components
u v
X t ̂ ,X t ̂ can be obtained, which will be further substituted into the fusion layer.

4.2.2. Structure for non-spatial time-series variables

Time variables (including the time-of-day and day-of-week) and weather variables (temperature, humidity, weather state, wind
speed, and visibility) are the two classes of non-spatial time-series variables. Considering that time variables and weather variables
affect the future demand intensity in different ways, we define two sequences of vectors:
es = (hs,ws ),as = (ats,ahs,ass,aws,avs ),s = 1,2,…,t . These two sequences of vectors are fed into two separate stacked LSTM architectures,
p q
which produce the two high-level components X t ̂ and X t ̂ .

(pt(−LK) +1 ,…,pt(L ) ) = FLL ···FlL···F1L (et−K +1,…,et−1,et)

e
e
e
e e (17)
p
X t ̂ = F T (F R (σ (wp pt(Le) + bp))) (18)

(qt(−LK) +1,…,qt(−L1) ) = FLL ···FlL···F1L (at−K ,…,at−1)

a
a
a
a a (19)
q
X t̂ = F T (F R (σ (wq qt(−L1a) + bq))) (20)
where pt(−Lke) ,k
= 1,2,…,K e,qt(−LKa),k
= 1,2,…,K a are the output hidden vectors in the highest LSTM layers Le ,La . K e,K a are the look-back time
window of time features and weather features (non-spatial time-series features), while Le ,La are the number of designed LSTM layers
for time and weather features respectively. wp,wq,bp,bq are the weighted and intercept parameters.

4.2.3. Fusion
Inspired by the fact that the high-level components have different contributions to the prediction, we employ Hadamard product
‘∘’ to multiply these components by the four parameter matrices W u,W v,W p and W q , which can be learnt to evaluate the importance
of the components during the training process. Therefore, the estimated demand intensity during the tth time interval is given by
u v p q
Xt ̂ = W u∘X t ̂ + W v ∘X t ̂ + W p∘X t ̂ + W q∘X t ̂ (21)

4.2.4. Objective function

During the training process of the FCL-Net, the object is to minimize the mean squared error between the estimated and real
demand intensity, through which the weighted and intercept parameters can be learnt. The objective function of the architecture is
shown in Eq. 22.

min ‖Xt −Xt ‖̂ 22 + α ‖W ‖22

w,b (22)
The second term of the objective function represents an L2-norm regularization term, which helps avoid overfitting issues. W
stands for all the weighted parameters in Xt ,̂ and α refers to a regularization parameter which balances the bias-variance tradeoff.
The training steps of the FCL-Net is illustrated in Algorithm 1.
Algorithm 1. FCL-Net Training

Input Observations of demand intensity {D1,…,Dn} in training set

Observations of demand intensity {Γ1,…,Γn} in training set

598
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Observations of time-of-day {h1,…,hn} , day-of-week {w1,…,wn} in training set

Observations of weather variables {at1,…,atn},{ah1,…,ahn},{as1,…,asn},{aw1,…,awn},{av1,…,avn}
lookback-windows: K d,K τ ,K e,K a
Output FCL-Net with learnt parameters
1:procedure FCL-Net Training
2: Initialize a null set: L ← ∅
3: for all available time intervals t (1 ⩽ t ⩽ n) do
4: Std ← [Dt − K d,Dt − K d + 1,…,Dt − 1]
5: Stτ ← [Γt − Kτ ,Γt − Kτ + 1,…,Γt − 1]
6: Ste ← [et − K e,…,et − 1,et ], where es = (hs,ws )
7: Sta ← [at − K a,…,at − 1,at ], where as = (ats,ahs,ass,aws,avs ) ▷where Std,Stτ ,Ste,Sta are the sets of different categories of
explanatory variables in one observation.
8: A training observation ({Std,Stτ ,Ste,Sta},Dt ) is put into L
9: end for
10: Initialize all the weighted and intercept parameters
11: repeat
12: Randomly extract a batch of samples Lb from L
13: Estimate the parameters by the minimizing the objective function shown in Eq. (22) within Lb
14: until convergence criterion met
15:end procedure

4.3. Spatial aggregated random forest for feature selection

Random forest, first introduced by Breiman (2001), is one of the most powerful ensemble learning algorithms for regression
problems. Consider a training set with m observations, L = {(X (1),y (1) ),…,(X (m),y (m) )} , where X (i) ∈ Rp is the ith observation of fea-
tures, and y (i) ∈ R is the ith observation of label. The kth decision tree can be represented as f k : Rp → R . Firstly, K sub-sample sets,
L1,…,LK , are randomly generated from the whole training set L using the bootstrap method (the same sample can be selected by
several sub-sample sets) and used for training models. Secondly, the kth out-of-bag sample set, Ok , defined as the difference of set L
and set Lk , is used for testing. Finally, the out-of-bag error of the kth tree, errOk , is denoted as the average testing error in the set Ok
(shown in Eq. (23)).
1
errOk = ∑ (y (i) −y (̂ i) )2
n i ∈ Ok (23)

where y (̂ i) = fk (X (i)) is the estimated value of the ith label based on tree k . The out-of-bag error can be utilized to calculate the feature
importance through the following steps (Genuer et al., 2015): (1) permute the jth variable of X in each Ok to get a new out-of-bag
∼j
samples Ok′; (2) calculate the out-of-bag error, errOk in the new sets of samples Ok′; (3) the importance of the jth variable, VI (X j) , is
∼j
equal to the average difference between errOk and errOk of all trees (shown in Eq. (24)).
1 ∼j
VI (X j) =
K
∑ (errOk −errOk )
k (24)

Considering that the dependent variables in the passenger demand forecasting, Dt ∈ RI×J, form an I × J matrix, instead of a
continuous value in the standard random forest, we develop a spatial aggregated random forest which consists of I × J standard
random forests, to examine the aggregated variable importance partitioned by category and look-back time window. To illustrate the
spatial aggregated random forest, we extend Problem 1 to Problem 2, given by
Problem 2. Given the historical observations and known information {dsi,j,τsi,j |s =
i′,j′
t −K ,…,t −1,i ∈ {1,…,I },j ∈ {1,…,J };hs,ws |s = t −K + 1,…,t;ats,ahs,ass,aws,avs |s = t −K ,…,t −1} , predict ds via the standard random forest
f i′,j′, for all i′ ∈ {1,…,I },j′ ∈ {1,…,J } . The length of the look-back window is denoted as K.
VI i′,j′ is denoted as the function to calculate variable importance in random forest f i′,j′. Two tensors, d , τ ∈ RI × J × I × J × K , where
( )i,j,i′,j′ ,k = VI i′,j′ (dti−,j k ),( τ )i,j,i′,j′ ,k = VI i′,j′ (τti−,j k ), are denoted to store the variable importance of the two categories of spatial-temporal
d

variables. (d )i,j,i′,j′ ,k refers to the variable importance of the passenger demand in {i,j} during time interval t −k in the problem of
forecasting passenger demand of {i′,j′ } during time slot t. As for non-spatial time-series variables, we define
h, w ,at ,ah,as ,aw ,av ∈ RI × J × K , where (h)i′,j′ ,k = VI i ′ ,j′ (ht − k ) , and the same expression for w,at ,ah,as,aw,av . All the variable
importance in d , τ ,h, w ,at ,ah,as ,aw ,av is normalized to percentage via dividing each variable importance by the sum of all
variable importance.
Firstly, we examine the variable importance partitioned by category, i.e. ∑i ∑j ∑i′ ∑ j′ ∑k d ,∑i′ ∑ j′ ∑k h , etc., to select the
important variables in terms of category. Secondly, we investigate the variable importance partitioned by category and look-back
time window k (k ∈ {1,2,…,K }) , to select a suitable look-back window for each variable category.

599
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Fig. 5. The investigated region partitioned into 7 × 7 grids.

5. Experiments and results

5.1. On-demand ride service platform data

The datasets utilized in this paper are extracted from DiDi Chuxing, the largest on-demand ride service platform in China, during
one-year period between November 1, 2015 and November 1, 2016. We randomly obtain 1,000,000 requesting orders from the
platform, each of which consists of the requesting time, travel distance, travel time, longitude and latitude. Although these 1,000,000
samples are randomly selected with a sampled ratio of 10% due to the limitation of data availability, the random sampling could
largely ensure the transfer from this dataset to the field application. In this dataset, the demand intensity can be calculated by
summing up the orders starting from one region in one time interval; the average travel time rate can also be obtained from the record
of travel distance and travel time with the definition in Section 3. The studied site is located in the downtown of Hangzhou, China,
starting from 120.00° E to 120.35° E in longitude, and from 30.15° N to 30.45° N in latitude. The dataset is partitioned into 1-h time
intervals, and the investigated region is partitioned into 7 × 7 grids, as shown in Fig. 5. Each grid is roughly a rectangle with length
equaled to 4.77 km and width equaled to 4.81 km. We also collect one-hour aggregated weather variables, including temperature,
humidity, weather state, wind speed, and visibility, during the same period from the China Meteorological Administration.
To avoid using future information, the dataset is divided into 70% training set comprised of observations between November 1,
2015 and July 14, 2016, and the 30% test set consisting of the remaining observations between July 15, 2016 and November 1, 2016.
Fig. 6(a) shows the total demand of all girds in different days within the investigated training (before the red dash line) and testing
(after the red dash line) period. It can be observed that both the training and test sets have some periods when the total demand is

Fig. 6. Temporal variations of the on-demand ride services.

600
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

abnormally low. The first low-demand period is due to the Chinese Lunar New Year in the beginning of February (in the training set),
while the second one is affected by the G20 summit held in the beginning of September (in the test set). These abnormal periods raise
challenges for the short-term demand forecasting.
It can be observed from Fig. 6(b), which shows the mean and variance of passenger demand in different hours of a day based on
the training set, that the passenger demand in weekdays demonstrates a double-peak nature while the passenger demand in weekends
shows a single-peak property. Therefore, the peak hours, off-peak hours, and sleep hours are separately defined for weekdays and
weekends.

5.1.1. Exploring the spatio-temporal correlations

The reason for utilizing a tailored spatio-temporal DL architecture is that there exist spatio-temporal correlations among the
spatio-temporal variables, i.e., the demand intensity and travel time rate. To validate this assumption, we examine the correlations
between the demand intensity at the tth time interval and spatio-temporal variables ahead of the tth time interval by employing the
Pearson correlation, given by

E [(Y −E (Y ))′ (Z −E (Z ))]

Corr (Y ,Z ) =
E [(Y −E (Y ))2] E [(Z −E (Z ))2] (25)

where Y ,Z are two random variables with the same number of observations.
Firstly, we calculate the Pearson correlations between the demand intensity at time t in grid (i′,j′ ) and demand intensity, travel
time rate at t −k time interval in grid (i,j ) , for all i,i′ ∈ {1,…,I },j,j′ ∈ {1,…,J },k ∈ {1,2,3,4} . Secondly, we average these correlations
partitioned by spatial distances and look-back time intervals. The spatial distance of grid (i,j ) and (i′,j′ ) is denoted as the Euclidean
distance between the central points of the two grids.
Fig. 7 shows the average correlations between the dependent variables (the demand intensity at time t in grid (i′,j′ ) ) and the
explanatory variables (the demand intensity and travel time rate at time t −k in grid (i,j )) . It can be observed that the average
correlations drop gradually but not sharply, with the increase of the spatial distance, indicating that there exit strong spatial cor-
relations between each grid and its neighbors. On the other hand, it is not surprising that variables with shorter look-back time
intervals have higher correlations, but the variables with large look-back time intervals are also correlated with the to-be-predicted
demand intensity to some extent. This correlation analysis of the dataset provides evidence that the spatial and temporal de-
pendencies exist among the spatio-temporal variables.

5.2. Feature selection

Firstly, Fig. 8 shows the variable importance partitioned by category, based on our proposed spatial aggregated random forest
algorithm. It can be observed that the two categories of spatial-temporary variables, travel time rate, and demand intensity, are the
dominating factors, followed by time-of-day and temperature. However, other variables, such as day-of-week, humidity, etc., have
few contributions (less than 5%) to the prediction.
Secondly, Table 1 presents the relative importance of the variables sorted by category and look-back time interval (“-” represents
not applicable). Fig. 9 displays the top 20 important variables. We set the look-back time window K to be 8 for all category of
variables. It can be found that the time-of-day during time interval t is the most important variable, followed by the demand intensity
and travel time rate during t −1 time interval. The importance of all kinds of variables decreases with the look-back time window, but
different categories of variables show different descent speeds. The travel time rate far before t still has considerable variable im-
portance, while the time-of-day prior to time t makes little contribution.
Selecting appropriate categories of variables and the suitable look-back window helps to improve computation efficiency with

Fig. 7. Average correlations partitioned by distance and time.

601
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

0.5

Explanatory variables
0.4 travel time rate

Relative importance
demand intensity
time−of−day
0.3
temperature
wind speed
0.2
humidity
visibility
0.1 weather state
day−of−week
0.0

Fig. 8. Variable importance ranking by the random forest.

Table 1
The relative importance of variables partitioned by category and look-back time interval (% ).

s Ds Γs ats ahs ass aws avs hs ws

t −8 2.628 5.005 0.589 0.074 0.036 0.080 0.089 – –

t −7 2.040 4.432 0.275 0.084 0.035 0.086 0.076 0.137 0.041
t −6 1.978 4.790 0.251 0.088 0.032 0.083 0.072 0.039 0.018
t −5 2.280 4.852 0.223 0.071 0.032 0.074 0.091 0.013 0.013
t −4 2.895 4.984 0.205 0.085 0.043 0.079 0.089 0.021 0.022
t −3 3.140 6.038 0.333 0.097 0.033 0.069 0.071 0.514 0.013
t −2 3.518 7.008 0.450 0.110 0.039 0.172 0.070 0.589 0.022
t −1 11.999 10.658 1.105 0.116 0.039 0.083 0.076 0.671 0.015
t – – – – – – – 13.808 0.012

time−of−day (t)
demand intensity (t−1)
travel time rate (t−1)
travel time rate (t−2)
travel time rate (t−3)
travel time rate (t−8)
travel time rate (t−4)
Important variables

travel time rate (t−5)

travel time rate (t−6)
travel time rate (t−7)
demand intensity (t−2)
demand intensity (t−3)
demand intensity (t−4)
demand intensity (t−8)
demand intensity (t−5)
demand intensity (t−7)
demand intensity (t−6)
temperature (t−1)
time−of−day (t−1)
temperature (t−8)
0.00

0.05

0.10

Relative importance

Fig. 9. Top 20 important variables partitioned by category and time.

little loss of predictive performance. In this paper, by considering the trade-off between computation efficiency and predictive
performance, we select 4 categories of variables: demand intensity, travel time rate, time-of-day, and temperature, with 4,8,2,2 look-
back time windows, respectively. This feature selection reduces the number of variables in each observation from
8 × 7 × 7 × 2 + 8 × 5 + 8 × 2 = 840 to 8 × 7 × 7 + 4 × 7 × 7 + 2 + 2 = 592.

5.3. Model comparisons

The proposed FCL-Net with full variables and selected variables (selected by the spatial aggregated random forest) are trained on

602
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

the training set and validated on the test set, respectively. Meanwhile, the conv-LSTM network, which is only fed with the historical
demand intensity, is trained and tested in the same way. The definition of the aforementioned three models are shown as follows:

(1) Conv-LSTM with only historical demand intensity: This model utilizes historical observations of demand intensity
{Ds |s = t −K d,…,t −1} to predict future demand intensity Dt, where K d = 8 in this paper. The architecture of Conv-LSTM is in-
troduced in Section 4.2.
(2) FCL-Net with full variables: Historical observations of all variables {Ds ,Γs |s = t −
K ,…,t −1;hs,ws |s = t −K + 1,…,t;ats,ahs,ass,aws,avs |s = t −K ,…,t −1} , where K = 8, are utilized to predict future demand intensity Dt.
The training process of this model is illustrated in Algorithm 1.
(3) FCL-Net with selected variables: This model utilizes the historical observations of the selected variables
{Ds |s = t −K d,…,t −1,Γs |s = t −K τ ,…,t −1;hs |s = t −Kh + 1,…,t;ats |s = t −K at ,…,t −1} where K d = 4,K τ = 8,Kh = 2,K at = 2 , to forecast
future demand intensity Dt. The training process of this model is the same as Algorithm 1 except that the inputs are replaced by
the selected variables.

In this section, the structure of FCL-Net is comprised of 4 Conv-LSTM layers for spatio-temporal features and 4 LSTM layers for
non-spatial time-series features. Each Conv-LSTM cell consists of 40 filters/channels to fully catch up spatial information. The number
of training epochs is set to be 100, while the batch size is set to be 10.
Apart from the proposed three models, several benchmark algorithms are also tested. The benchmark algorithms include three
traditional time-series forecasting models (i.e., HA, MA, and ARIMA) and several state-of-the-art machine learning/deep learning
approaches (i.e., XGBoost, ANN, CNN, and LSTM).

(1) HA: The historical average model predicts the future demand intensity in the test set based on the empirical statistics in the
training set. For example, the average demand intensity during 8–9 AM in grid (i,j ) is estimated by the mean of all historical
demand intensity during 8–9 AM in grid (i,j ) .
(2) MA: The moving average model is widely-used in time-series analysis, which predicts future value by the mean of serval nearest
historical values. In this paper, the average of 8 previous demand intensities in grid (i,j ) are used to predict the future demand
intensity in grid (i,j ) .
(3) ARIMA: The autoregressive integrated moving average model (Box and Pierce, 1970) integrates the autoregressive (AR), in-
tegrated (I), and MA parts, and considers trends, cycles, and non-stationary characteristics of a dataset simultaneously.
(4) ANN: The artificial neural network (Rumelhart et al., 1985, 1988) employs all the variables, including the historical demand
intensity, travel time rate, hour and week state, and weather variables, with look-back time window K = 8, of a specific grid (i,j ) ,
to predict the future demand intensity in grid (i,j ) . ANN does not differentiate variables across time and thus fails to capture time
dependencies.
(5) LSTM: In LSTM (Hochreiter and Schmidhuber, 1997), all the variables in grid (i,j ) are reshaped to a matrix with one axis as time
step (the size of which equals look-back time window K = 8) and another axis as the feature category. In this way, all the features
utilized in FCL-Net are fed into LSTM for training. LSTM considers temporal dependencies, but does not capture spatial de-
pendencies.
(6) CNN: In the convolutional neural network(Krizhevsky et al., 2012), the demand intensity and travel time rate of all grids in each
time step (e.g., t −1 time step) is viewed as a picture, thus 8 × 2 = 16 pictures are obtained for each sample. We build 16
convolutional neural networks for spatio-temporal features and a fully-connected neural network for the non-spatial time-series
features, which are then fused together in the top layer.
(7) XGBoost: XGBoost(Chen and Guestrin, 2016) is a scalable end-to-end tree boosting system, which is proven to achieve out-
standing performance in a board range of machine learning challenges. All the variables utilized in FCL-Net (including spatio-
temporal features and non-spatial time-series features) are reshaped to a vector and fed into XGBoost for training.

To ensure fairness, the aforementioned benchmark machine learning algorithms have the same input features (same category and
look-back time windows) as the FCL-Net, while the three traditional time-series models make use of the whole time-series of the
historical demand intensities. The deep learning approaches (ANN, LSTM, and CNN) are trained with 100 epochs, which is consistent
with FCL-Net. All the benchmarks are under fine-tuned. Before model training and validation, the demand intensity, travel time rate,
hour-of-day, day-of-week, and weather variables, are standardized to the range [0,1], through the max-min standardization, re-
spectively.
Our experiment platform is a server with 20 CPU cores (Intel(R) Xeon(R) CPU E5-2630 v4 @ 2.20 GHz), 251 GB RAM, and one
GPU (NVIDIA UNIX x86-64 Kernel Module 375.66). We use python 2.7.5 with scikit-learn (Pedregosa et al., 2011), tensorflow (Abadi
et al., 2015), keras (Chollet, 2015), and XGBoost (Chen and Guestrin, 2016) on CentOS Linux release 7.2.1511 (Core) for comparing
the models.
We evaluate the models via the four measures of effectiveness: root mean squared error (RMSE), coefficient of determination (R2) ,
mean absolute error (MAE), and mean absolute percentage error (MAPE), given by

1 n
RMSE =
n
∑i =1 (y (i) −y (̂ i) )2 (26)

603
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Table 2
Predictive performance comparison.

Model RMSE R2 MAE MAPE@10a Time (min)

HA 0.0378 0.736 0.0192 28.06% 0.01

MA 0.0511 0.518 0.0260 43.63% 0.01
ARIMA 0.0345 0.780 0.0178 30.40% 2.27
XGBoost 0.0322 0.801 0.0176 27.34% 2.13
ANN 0.0331 0.797 0.0198 27.95% 17.38
LSTM 0.0332 0.798 0.0172 31.27% 193.39
CNN 0.0175 0.773 0.0106 27.52% 85.42
b
Conv-LSTM (demand) 0.0315 0.806 0.0175 26.18% 99.93
FCL-Net (selected) c 0.0163 0.803 0.0096 20.46% 169.90
FCL-Net (full) d 0.0160 0.812 0.0094 21.79% 224.71

a
MAPE in the testing samples with demand intensity greater or equal to 10.
b
Conv-LSTM with only demand intensity as input features.
c
FCL-Net with selected variables as input.
d
FCL-Net with full variables as input.

n
∑i = 1 (y (i) −y (̂ i) )2
R2 = 1− n
∑i = 1 (y (i) −y )2 (27)
n
1
MAE = ∑ |y (i) −y (̂ i) |
n i=1 (28)
n
1 |y (i) −y (̂ i) |
MAPE = ∑
n i=1
y (i) (29)

where y (i) ,y (̂ i)
are the ith ground truth and estimated value of the demand intensity, respectively. y is the mean of all y (i) ,
and n is the
size of the test set. We use RMSE, (R2) , and MAE to measure the total predictive accuracy/degree of fitting in the whole test data, and
use MAPE@10 (MAPE in the samples with demand intensity greater or equal to 10) to measure models’ predictive performance in the
highly demanded regions and time periods. The platform would pay more attention to the predicted highly demanded regions in peak
hours, with which it can dispatch vacant drivers to those regions. In this paper, MAPE@10 covers the top 4.45% largest samples in the
test set.
Table 2 shows the predictive performance of the proposed models and benchmark models on the test set. It can be found that the
proposed FCL-Nets outperform the benchmarks in the four measurements of predictive performance. The FCL-Net with full variables
achieves the best predictive performance measured by RMSE (0.016), which is 8.6% lower than the best benchmark CNN (0.0175).
With the feature selection, FCL-Net achieves a 24.4% drop on the training time, while bearing only a 1.8% decrease on predictive
performance measured by RMSE. The FCL-Net with selected variables (with MAPE@10 equaled to 20.46%) even outperforms the
FCL-Net with full variables (with MAPE@10 equaled to 20.46%) in the samples with a large demand intensity. It indicates that
feature selection is valuable to FCL-Net since it takes into account the trade-off between the computation complexity and predictive
performance.
It is also interesting to find that both FCL-Nets have relatively 48.3% lower RMSE than Conv-LSTM with only the historical
demand intensity, which indicates that the exogenous variables make great contribution to the short-term passenger demand fore-
casting. From the comparisons among the benchmark models, we can see that the machine learning approaches have better pre-
dictive performance but longer computing time than the traditional time-series models. CNN which captures spatial correlations
outperforms LSTM which only models the temporal characteristics, providing another verification that the consideration of spatial
correlations is of great importance to the city-wide demand forecasting.
Fig. 10 shows some samples of heat maps of the ground truth passenger demand and predicted results by FCL-Net, where the
deeper color implies a larger demand intensity. It is obvious that the demand intensity in peak hours (e.g., 9–10 AM and 6–7 PM) is
much higher than that in sleep hours (e.g., 0–1 AM). The demand intensity is unbalanced across space: the central grids have much a
higher demand intensity than other grids. The trend of the demand intensity over time is even different in different grids and different
days, which makes it hard to forecast short-term passenger demand. From the samples of visualization, we can find that the FCL-Net
primarily captures the spatio-temporal characteristics of the demand intensity and makes accurate forecasting. The combination of
short-term passenger demand forecasting and visualization helps traffic operators of the platform/government to detect and forecast
grids with oversupply and overfull demand and design proactive strategies to avoid these imbalanced conditions.

5.4. Sensitivity analysis

In this section, we conduct the sensitivity analysis and parameter tuning on FCL-Net, where four kinds of parameters are in-
vestigated, including the number of training epochs, look-back time windows, layers of the structure, and filters of convolutional
units. Firstly, the validation error after each training epoch is recorded for FCL-Net and the benchmark models. From Fig. 11(a), it can

604
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Fig. 10. Comparison of the ground truth (G) and predicted passenger demand by FCL-Net (P).

Fig. 11. Parameter tuning and sensitivity analysis of FCL-Net.

605
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Fig. 12. Sensitivity analysis on multi-step prediction.

be observed that the RMSE of FCL-Nets drops quickly in the initial 5 epochs and then decreases gradually, which indicates that FCL-
Nets have a high convergence speed. It is reported in Section 5.3 that the FCL-Net with full variables can cost 224.71 min for 100
epochs’ training, which seems to be computationally expensive. However, the number of training epochs can be reduced to 10 or even
5, with the predictive performance slightly sacrificed, under the case where high-efficient computation resources are not reachable.
With the knowledge of the training convergence in mind, we then train the FCL-Net with full variables under different look-back
time windows, layers, and filters/channels, with 30 training epochs (which are enough for convergence). Fig. 11(b) shows that the
RMSE of FCL-Net gradually decreases while the training time linearly increases with the increase of the look-back time window. This
result is intuitive: the more temporal information is utilized, the better predictive performance can be achieved but the longer
training time is needed. Fig. 11(c) demonstrates that the RMSE first drops and then rises with the increase of Conv-LSTM layers in the
structure of FCL-Net, while the computing time monotonously increases with the number of layers. With the increase of the depth of
FCL-Net, it could face overfitting issues, which can explain the rise of its RMSE when the number of layers is greater than 4. From
Fig. 11(d), it is interesting to find that both the training error (RMSE) and training time have no significant relationships with the
number of filters/channels of convolutional units. This phenomenon could be explained as that the DL backend tensorflow is dis-
tributed, thus the extracted features in each filter/channel of FCL-Net’s convolutional units can be calculated in parallel.
The sensitivity analysis on the multi-step prediction is also conducted in this section. The x-axis in Fig. 12 refers to the look-ahead
time window (n) to predict, which indicates predicting the demand intensity of t + n time interval with the information prior to time
interval t, while the two y-axes record the RMSE and MAPE@10 in different n, respectively. The results show that the worst RMSE is
0.0218, which is lower than most of the baselines on predicting the t + 1 demand intensity. The model can also maintain a MAPE@10
lower than 30% within 7 look-ahead time windows. These results show that the predictive accuracy of FCL-Net deteriorates slowly
over the look-ahead time window, which means the proposed method adapts well to the multi-step prediction tasks.

6. Conclusions

In this paper, we propose a DL approach, named the fusion convolutional LSTM (FCL-Net), for the short-term passenger demand
forecasting under an on-demand ride service platform. The proposed architecture is fused by multiple conv-LSTM layers, LSTM
layers, and convolutional operators, and fed with a variety of explanatory variables including the historical passenger demand, travel
time rate, time-of-day, day-of-week, and weather conditions. A tailored spatially aggregated random forest is employed to rank the
importance of the explanatory variables. The ranking is then used for feature selection. We trained two FCL-Nets, i.e., one trained
with full variables, and the other trained with selected variables. In addition, the conv-LSTM which only takes historical passenger
demand as the explanatory variable is also established. These three models are compared with several benchmark algorithms in-
cluding the HA, MA, ARIMA, XGBoost, ANN, LSTM, and CNN. The models are validated on the real-world data provided by DiDi
Chuxing, the results of which show that the two FCL-Nets outperform the benchmark algorithms in the measurements of RMSE, R-
square, MAE, and MAPE in large samples, indicating that the proposed approach performs better at capturing the spatio-temporal
characteristics for the short-term passenger demand forecasting. The FCL-Net utilizing full variables outperforms the best benchmark
CNN by 8.6% in RMSE. The feature selection process helps FCL-Net reduce training time by 24.4% while only suffering from the 1.8%
loss in the predictive performance (measured by RMSE). The experiments results also show that the consideration of exogenous
variables is important since the FCL-Net achieves a 48.3% lower RMSE compared to Conv-LSTM with only the historical demand
intensity. The evidence from benchmark models proves that characterizing spatial correlations in models can greatly improve the
predictive accuracy.
This paper explores the short-term passenger demand forecasting under the on-demand ride service platform via a novel spatio-
temporal DL approach. Accurate real-time passenger demand forecasting can provide suggestions for the platform to rebalance the
spatial distribution of cruising cars to meet passenger demand in each region, which will improve the car utilization rate and
passengers’ degree of satisfaction. One limitation of this paper is that the dataset obtained is randomly sampled with the 10%
sampling ratio, thus there might exist bias between the studied dataset and field application. The proposed model will be tested with
more data in future research. In particular, we anticipate to evaluate the proposed methods in the field application on on-demand ride

606
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

service platforms in the future. Furthermore, to understand the complex interactions among the variables in the on-demand ride
service market is far beyond predicting passenger demand. In the following studies, we expect to utilize more economic analyses and
machine learning techniques to further explore the relationship between the endogenous and exogenous variables under on-demand
ride service platforms.

Acknowledgements

This research is financially supported by Zhejiang Provincial Natural Science Foundation of China [LR17E080002], Key
Laboratory of Road & Traffic Engineering of the Ministry of Education [TJDDZHCX004], National Natural Science Foundation of
China [51508505, 71771198, 51338008], and Fundamental Research Funds for the Central Universities [2017QNA4025]. The au-
thors are grateful to DiDi Chuxing (htpp://www.xiaojukeji.com) for providing us some sample data.

References

Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G.S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G.,
Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Mané, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B.,
Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Viégas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., Zheng, X., 2015.
TensorFlow: Large-scale machine learning on heterogeneous systems. Software available from <http://tensorflow.org/>.
Ahmed, M., Abdel-Aty, M., 2013. A data fusion framework for real-time risk assessment on freeways. Transp. Res. Part C: Emerg. Technol. 26, 203–213.
Alonso-Mora, J., Samaranayake, S., Wallar, A., Frazzoli, E., Rus, D., 2017. On-demand high-capacity ride-sharing via dynamic trip-vehicle assignment. Proc. Nat. Acad.
Sci. 201611675.
Antoniou, C., Koutsopoulos, H.N., Yannis, G., 2013. Dynamic data-driven local traffic state estimation and prediction. Transp. Res. Part C: Emerg. Technol. 34, 89–107.
Bachmann, C., Abdulhai, B., Roorda, M.J., Moshiri, B., 2013. A comparative assessment of multi-sensor data fusion techniques for freeway traffic speed estimation
using microsimulation modeling. Transp. Res. Part C: Emerg. Technol. 26, 33–48.
Boto-Giralda, D., Díaz-Pernas, F.J., González-Ortega, D., Díez-Higuera, J.F., Antón-Rodríguez, M., Martínez-Zarzuela, M., Torre-Díez, I., 2010. Wavelet-based de-
noising for traffic volume time series forecasting with self-organizing neural networks. Comput.-Aided Civil Infrastruct. Eng. 25, 530–545.
Box, G.E., Pierce, D.A., 1970. Distribution of residual autocorrelations in autoregressive-integrated moving average time series models. J. Am. Stat. Assoc. 65,
1509–1526.
Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32.
Chan, K.Y., Dillon, T.S., Singh, J., Chang, E., 2012. Neural-network-based models for short-term traffic flow forecasting using a hybrid exponential smoothing and
levenberg–marquardt algorithm. IEEE Trans. Intell. Transp. Syst. 13, 644–654.
Chen, Q., Song, X., Yamada, H., Shibasaki, R., 2016. Learning deep representation from big and heterogeneous data for traffic accident inference. In: Thirtieth AAAI
Conference on Artificial Intelligence.
Chen, T., Guestrin, C., 2016. Xgboost: A scalable tree boosting system. In: Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and
data mining, ACM. pp. 785–794.
Chen, X., Zahiri, M., Zhang, S., 2017. Understanding ridesplitting behavior of on-demand ride services: an ensemble learning approach. Transp. Res. Part C: Emerg.
Technol. 76, 51–70.
Cheng, Q., Liu, Y., Wei, W., Liu, Z., 2016. Analysis and forecasting of the day-to-day travel demand variations for large-scale transportation networks: a deep learning
approach. doi:http://dx.doi.org/10.13140/RG.2.2.12753.53604.
Chollet, F., et al., 2015. Keras. <https://github.com/fchollet/keras>.
Deng, Z., Ji, M., 2011. Spatiotemporal structure of taxi services in shanghai: Using exploratory spatial data analysis. In: The 19th IEEE International Conference on
Geoinformatics, pp. 1–5.
Fei, X., Lu, C.C., Liu, K., 2011. A Bayesian dynamic linear model approach for real-time short-term freeway travel time prediction. Transp. Res. Part C: Emerg. Technol.
19, 1306–1318.
Genuer, R., Poggi, J.M., Tuleau-Malot, C., 2015. Vsurf: An R package for variable selection using random forests. The R Journal 7, 19–33.
Ghosh, B., Basu, B., O’Mahony, M., 2009. Multivariate short-term traffic flow forecasting using time-series analysis. IEEE Trans. Intell. Transp. Syst. 10, 246–254.
Guo, J., Huang, W., Williams, B.M., 2014. Adaptive kalman filter approach for stochastic short-term traffic flow rate prediction and uncertainty quantification. Transp.
Res. Part C: Emerg. Technol. 43, 50–64.
He, F., Shen, Z.J.M., 2015. Modeling taxi services with smartphone-based e-hailing applications. Transp. Res. Part C: Emerg. Technol. 58, 93–106.
Hochreiter, S., 1991. Untersuchungen zu dynamischen neuronalen Netzen ( Ph.D. thesis). diploma thesis, institut für informatik, lehrstuhl prof. brauer, technische
universität münchen.
Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural Comput. 9, 1735–1780.
Huang, S., Sadek, A.W., 2009. A novel forecasting approach inspired by human memory: the example of short-term traffic volume forecasting. Transp. Res. Part C:
Emerg. Technol. 17, 510–525.
Huang, W., Song, G., Hong, H., Xie, K., 2014. Deep architecture for traffic flow prediction: deep belief networks with multitask learning. IEEE Trans. Intell. Transp.
Syst. 15, 2191–2201.
Jiang, X., Zhang, L., Chen, X., 2014. Short-term forecasting of high-speed rail demand: a hybrid approach combining ensemble empirical mode decomposition and gray
support vector machine with real-world applications in china. Transp. Res. Part C: Emerg. Technol. 44, 110–127.
Khashei, M., Bijari, M., Ardali, G.A.R., 2012. Hybridization of autoregressive integrated moving average (ARIMA) with probabilistic neural networks (PNNs). Comput.
Ind. Eng. 63, 37–45.
Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep convolutional neural networks. In: Advances in Neural Information Processing
Systems, pp. 1097–1105.
LeCun, Y., Haffner, P., Bottou, L., Bengio, Y., 1999. Object recognition with gradient-based learning. Shape Contour Group. Comput. Vision 1681, 823.
Lv, Y., Duan, Y., Kang, W., Li, Z., Wang, F.Y., 2015. Traffic flow prediction with big data: a deep learning approach. IEEE Trans. Intell. Transp. Syst. 16, 865–873.
Ma, X., Dai, Z., He, Z., Ma, J., Wang, Y., Wang, Y., 2017. Learning traffic as images: a deep convolutional neural network for large-scale transportation network speed
prediction. Sensors 17, 818.
Ma, X., Tao, Z., Wang, Y., Yu, H., Wang, Y., 2015a. Long short-term memory neural network for traffic speed prediction using remote microwave sensor data. Transp.
Res. Part C: Emerg. Technol. 54, 187–197.
Ma, X., Yu, H., Wang, Y., Wang, Y., 2015b. Large-scale transportation network congestion evolution prediction using deep learning theory. PloS one 10, e0119044.
Moreira-Matias, L., Gama, J., Ferreira, M., Damas, L., 2012. A predictive model for the passenger demand on a taxi network. In: The 15th IEEE International
Conference on Intelligent Transportation Systems, pp. 1014–1019.
Moreira-Matias, L., Gama, J., Ferreira, M., Mendes-Moreira, J., Damas, L., 2013. Predicting taxi–passenger demand using streaming data. IEEE Trans. Intell. Transp.
Syst. 14, 1393–1402.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E., 2011. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830.

607
J. Ke et al. Transportation Research Part C 85 (2017) 591–608

Polson, N.G., Sokolov, V.O., 2017. Deep learning for short-term traffic flow prediction. Transp. Res. Part C: Emerg. Technol. 79, 1–17.
Qian, X., Ukkusuri, S.V., 2015. Spatial variation of the urban taxi ridership using GPS data. Appl. Geography 59, 31–42.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1985. Learning internal representations by error propagation. Technical Report. California Univ San Diego La Jolla Inst
for Cognitive Science.
Rumelhart, D.E., Hinton, G.E., Williams, R.J., et al., 1988. Learning representations by back-propagating errors. Cogn. Model. 5, 1.
Shi, X., Chen, Z., Wang, H., Yeung, D.Y., Wong, W.K., Woo, W.c., 2015. Convolutional lstm network: A machine learning approach for precipitation nowcasting. In:
Advances in Neural Information Processing Systems, pp. 802–810.
Soriguera, F., Robusté, F., 2011. Estimation of traffic stream space mean speed from time aggregations of double loop detector data. Transp. Res. Part C: Emerg.
Technol. 19, 115–129.
Sun, S., Huang, R., Gao, Y., 2012. Network-scale traffic modeling and forecasting with graphical lasso and neural networks. J. Transp. Eng. 138, 1358–1367.
Sutskever, I., Hinton, G.E., Taylor, G.W., 2009. The recurrent temporal restricted boltzmann machine. In: Advances in Neural Information Processing Systems, pp.
1601–1608.
Toqué, F., Côme, E., El Mahrsi, M.K., Oukhellou, L., 2016. Forecasting dynamic public transport origin-destination matrices with long-short term memory recurrent
neural networks. In: The 19th IEEE International Conference on Intelligent Transportation Systems, pp. 1071–1076.
Vlahogianni, E.I., Karlaftis, M.G., Golias, J.C., 2014. Short-term traffic forecasting: where we are and where we’re going. Transp. Res. Part C: Emerg. Technol. 43, 3–19.
Wang, J., Deng, W., Guo, Y., 2014. New Bayesian combination method for short-term traffic flow forecasting. Transp. Res. Part C: Emerg. Technol. 43, 79–94.
Wang, J., Gu, Q., Wu, J., Liu, G., Xiong, Z., 2016a. Traffic speed prediction and congestion source exploration: a deep learning method. In: The 16th IEEE International
Conference on Data Mining, pp. 499–508.
Wang, J., Shi, Q., 2013. Short-term traffic speed forecasting hybrid model based on chaos–wavelet analysis-support vector machine theory. Transp. Res. Part C: Emerg.
Technol. 27, 219–232.
Wang, X., He, F., Yang, H., Gao, H.O., 2016. Pricing strategies for a taxi-hailing platform. Transp. Res. Part E: Logist. Transp. Rev. 93, 212–231.
Wang, X., Jiang, R., Li, L., Lin, Y., Zheng, X., Wang, F.Y., 2017. Capturing car-following behaviors by deep learning. IEEE Trans. Intell. Transp. Syst., in press, http://
dx.doi.org/10.1109/TITS.2017.2706963
Wei, Y., Chen, M.C., 2012. Forecasting the short-term metro passenger flow with empirical mode decomposition and neural networks. Transp. Res. Part C: Emerg.
Technol. 21, 148–162.
Williams, R.J., Zipser, D., 1989. A learning algorithm for continually running fully recurrent neural networks. Neural Comput. 1, 270–280.
Yang, H., Fung, C., Wong, K., Wong, S.C., 2010a. Nonlinear pricing of taxi services. Transp. Res. Part A: Policy Pract. 44, 337–348.
Yang, H., Leung, C.W., Wong, S.C., Bell, M.G., 2010b. Equilibria of bilateral taxi–customer searching and meeting on networks. Transp. Res. Part B: Methodol. 44,
1067–1083.
Yang, H., Wong, S.C., Wong, K., 2002. Demand–supply equilibrium of taxi services in a network under competition and regulation. Transp. Res. Part B: Methodol. 36,
799–819.
Yang, H., Yang, T., 2011. Equilibrium properties of taxi markets with search frictions. Transp. Res. Part B: Methodol. 45, 696–713.
Yang, H., Ye, M., Tang, W.H.C., Wong, S.C., 2005. A multiperiod dynamic model of taxi services with endogenous service intensity. Oper. Res. 53, 501–515.
Yu, B., Lam, W.H., Tam, M.L., 2011. Bus arrival time prediction at bus stop with multiple routes. Transp. Res. Part C: Emerg. Technol. 19, 1157–1170.
Yu, H., Wu, Z., Wang, S., Wang, Y., Ma, X., 2017. Spatiotemporal recurrent convolutional networks for traffic prediction in transportation networks. Sensors 17, 1501.
Yu, R., Abdel-Aty, M.A., Ahmed, M.M., Wang, X., 2014. Utilizing microscopic traffic and weather data to analyze real-time crash patterns in the context of active traffic
management. IEEE Trans. Intell. Transp. Syst. 15, 205–213.
Zha, L., Yin, Y., Yang, H., 2016. Economic analysis of ride-sourcing markets. Transp. Res. Part C: Emerg. Technol. 71, 249–266.
Zhang, D., He, T., Lin, S., Munir, S., Stankovic, J.A., 2017a. Taxi-passenger-demand modeling based on big data from a roving sensor network. IEEE Trans. Big Data 3
(3), 362–374.
Zhang, J., Zheng, Y., Qi, D., 2017b. Deep spatio-temporal residual networks for citywide crowd flows prediction. In: AAAI, pp. 1655–1661.
Zhang, N., Zhang, Y., Lu, H., 2011. Seasonal autoregressive integrated moving average and support vector machine models: prediction of short-term traffic flow on
freeways. Transp. Res. Rec.: J. Transp. Res. Board 2215, 85–92.
Zhao, K., Khryashchev, D., Freire, J., Silva, C., Vo, H., 2016. Predicting taxi demand at high spatial resolution: approaching the limit of predictability. In: IEEE
International Conference on Big Data.
Zhao, Z., Chen, W., Wu, X., Chen, P.C., Liu, J., 2017. Lstm network: a deep learning approach for short-term traffic forecast. IET Intel. Transp. Syst. 11, 68–75.
Zhu, J.Z., Cao, J.X., Zhu, Y., 2014. Traffic volume forecasting based on radial basis function neural network with the consideration of traffic flows at the adjacent
intersections. Transp. Res. Part C: Emerg. Technol. 47, 139–154.

608