1 XYZ: An Efficient Deep Learning Architecture for Spatio-temporal Data for Ride Sourcing

2 Demand Prediction
3
4 Academic Author Name
5 Department of XXX
6 Institution Name, City, State or Country, and Postcode
7 Email: [email protected]
8
9 Public Sector Author Name
10 Position
11 XXStateXX Department of Transportation
12 Department (if applicable)
13 City, State or Country, Postcode
14 Email: [email protected]
15
16 Private Practitioner Author Name
17 Position
18 Company
19 City, State or Country, Postcode
20 Email: [email protected]
21
22 Word Count: 2465 words + 1 table (250 words per table) = 2,715 words
23
24
25 Submitted [Submission Date]
26
 Author, Author, and Author

1 ABSTRACT
2 Deep learning is being widely implemented in several fields in recent years, and the prediction of spatial-
3 temporal demand is one of them. Modern ride hailing services train these models to improve their services
4 by using a huge amount of trip data they generate on a daily basis. Nevertheless, developing a model to
5 achieve a very high degree of precision is still a daunting task. To address this problem, we propose a novel
6 deep learning model by combining a 1D Convolutional Neural Network (CNN) and Long Short-Term
7 Memory (LSTM) into an end-to-end structure (1DConv-LSTM). This model was compared to several
8 baseline standard forecasting models using DiDi's real-world data sets for demand forecasting in two cities
9 in China and found to be outperforming them in terms of precision as well as computing resources and
10 time.
11 Keywords: Convolutional Long short-term memory (Conv-LSTM), Deep Learning (DL),

2
 Author, Author, and Author

1 INTRODUCTION
2 Transportation resource management is very critical to a well-functioning community. Typically,
3 people choose a mode of transport based on their convenience and the price of their trip. Nowadays various
4 disruptive companies are entering the transportation sector, which are known as ride hailing services. For
5 these businesses, the projection of travel demand is a real-life challenge. Such businesses need to align their
6 limited resources and deploy them in the best way possible. Precise travel prediction would allow them to
7 advise their fellow drivers on the next probable trip origin, even before the actual ride request is made. It
8 will cut the driver's idle waiting period significantly, resulting in even improved resource utilization. Yet
9 these models require a lot of data to start training in the first place. Here, ride hailing companies are in luck,
10 as these companies act as the agent between the customer and the driver. In the meantime, they keep log of
11 every trip made and thus they generate a huge amount of trip data every day.
12 Trip data is a typical example of spatio-temporal data. The geographical data related to travel like
13 co-ordinates of a trip, origin point, destination point etc. are spatial, on the other hand the time dependent
14 factors like trip summon time, trip duration etc. are temporal. Trip demand meets intermittent averages,
15 depending on certain parameters. And, if a model is based on historical data, the future demand for travel
16 can be anticipated. The more information there are, the better the forecast will be.
17 No matter what how much is feeded into a prediction model, if the algorithm is not efficient enough
18 then the prediction is going to suffer. Furthermore, people can quickly check on the phone the situation on
19 the road in real time and the weather forecast, and these factors significantly influence their travel decisions.
20 To make a good travel demand prediction, model should take these factors into account too.
21
22 PREVIOUS WORKS
23 There has been a long tradition of travel demand forecasting research, and several models have
24 been developed to date. Travel demand data is spatio-temporal but many early researchers have failed to
25 incorporate both spatial and temporal characteristics into their models. There are a large number of
26 traditional methods in traditional time-series data mining, such as AR, MA, and ARMA, etc. Moreover,
27 some modified models such as ARIMA and SARIMA are introduced. Li et al. (1) created an advanced
28 autoregressive integrated moving average (ARIMA) in an attempt to capture the spatial-temporal variation
29 of passengers in a hotspot from the univariate time series data of historical taxi demand. Chang et al. (2)
30 performed historical data mining of weather, time and taxi car location using clustering algorithms to
31 forecast taxi demand distribution. Moreira-Matias et al. (3, 4) used time series techniques to make a model
32 performing short-term taxi demand forecasting.
33 Later on, some literatures can be found based on artificial neural networks for traffic demand
34 forecasting problems. Xu et al. (5) tested a recurrent neural network model to predict taxi demand in New
35 York. Wei et al. (6) examined spatially related features through an Artificial Neural Network (ANN) and
36 combined the temporal features of demand fluctuations for forecasting the ride-hailing demand using
37 Gradient Boosting Decision Tree (GBDT). Qian et al. (7) combined boosting methods with gaussian
38 conditional random field (GCRF), which was used to predict ride-hailing demand based on calculating
39 probable demand distributions in the future.
40 Only recently the trend emerged to introduce deep neural fusion models to examine both spatial
41 and temporal features of travel demand data. Where Long Short-Term Memory (LSTM) is well known for
42 time series interpretation, Convolutional Neural Network performs well in spatial pattern recognition. Ke
43 et al. (8) came with the Fusion Convolutional Long Short Term Memory (LSTM) Network (FCL-Net) to
44 forecast travel demand for ride hailing services in Hangzhou, China. Liao et al. (9) used two Deep Neural
45 Networks (DNN): (ST-ResNet) and FLC-Net, for taxi demand prediction using New York City taxi data.
46 Wang et al. (10) introduced a Deep Spatio-Temporal ConvLSTM (DeepSTCL) model where three
47 components (closeness, period and trend) were used for three prediction fusion model to outrun traditional
48 models.
49 All of the deep learning fusion models described above were focused on the 2D Convolutional
50 Network. The issue is that it simply implies that all the neighboring places are equally relevant in terms of
51 the spatial trend of travel demand, but this may not be the case all the time. This article presents the concept

3
 Author, Author, and Author

1 that the 1D Convolutional Network will do better in terms of pattern detection than the 2D transport demand
2 prediction, along with better output in the context of time and computing resources.
3
4 PRELIMINARIES
5 In this section data description, related definitions and demand forecast problem is going to be
6 discussed.
7
8 Data
9 The data used in this study are the ride-hailing request data provided by DiDi Chuxing in Chengdu and
10 Xian, China. Under their GAIA (11) program, they agreed to share with the academic community a full
11 summary of the route and travel request details. The dataset was collected from 1 October to 30 November
12 2016. The Chengdu dataset contains around 12 million travel request records, while the Xian dataset logs
13 about 7 million trip requests. The dataset provides origin-destination (OD) points, start-and end-time and
14 route geolocation offering GPS trajectory of the trip for simulating the clear picture of the mobility whole
15 throughout the cities at that time.
16 Climate data is vital for forecasting demand considering the broad influence of weather conditions on the
17 ride demand. We gathered open source weather data from the Dark Sky API (12), including temperature,
18 humidity, wind, weather condition etc.
19 Origin and destination of trips throughout different parts of the city is closely related with the distribution
20 of the important visiting places, commonly known as Point of Interest (POI). For demand prediction
21 modeling, POI data of these two cities was obtained from Peking University Open Research Data Platform
22 (13) which includes 8 basic information, including POI name, type, address, place, province name, city
23 name, POI area name, and POI area code.
24
25 TABLE 1 Summary of Datasets

Description Chengdu Xian

Zones 100 100
Time span 61 days (October 1, 2016 to November 30,2016)
Time step 15 minutes (4 time-slots per hr)
61(day)*24(hr/day)*4(slot/hr) = 5856 slots
Time slot (Last 8 hrs was omitted due to lacking of weather data,
resulting in 5824 useful time slots)
Orders
Temperature (°C), PM2.5, and weather condition (9
Weather
categories)
POI Number of facilities in 25 POI classes
26
27 Definations
28 Definition 1 (Region-time partitioning)
29 For variable aggregation, the study area is divided into 𝑁𝑁×𝑁𝑁 non-overlapping uniformly sized zones:
30 {1,2,3,..., 𝑁𝑁×𝑁𝑁}
31 and total days are divided into T time-slots:
32 {1,2,3,..., T} of M minutes interval.
33 The rest of the variables are defined based on the space-time partitioning.
34

4
 Author, Author, and Author

1
2 Figure 1 Chegdu and Xian study area
3
4 Definition 2 (Spatio-temporal variable)
5 The total number of orders from a zone in a time interval is referred as the travel demand. Travel demand
6 is defined as follows:
7
8 𝑫𝑫𝑘𝑘 𝑚𝑚,𝑛𝑛 =Σ 𝑂𝑂𝑂𝑂𝑘𝑘∈𝑅𝑅 |{𝑅𝑅(x,y)}|
9
10 Where, 𝑅𝑅(x,y) means the geographical rectangle (x,y); 𝑂𝑂𝑂𝑂𝑘𝑘∈𝑅𝑅 represents the kth timestamp belonging to R;
11 𝑇𝑇𝑇𝑇𝑡𝑡 𝑥𝑥,𝑦𝑦 means the order count at kth timestamp in R.
12 The speed of a zone means the average speed of all vehicles at a certain time where 𝑂𝑂M𝑘𝑘∈𝑅𝑅. Simialrly the
13 mobility of a zone means the number of all vehicles passing through a zone during a given time, and 𝑂𝑂S𝑘𝑘∈𝑅𝑅.
14
15 𝑺𝑺𝑘𝑘 𝑚𝑚,𝑛𝑛 =Σ 𝑂𝑂S𝑘𝑘∈𝑅𝑅 |{𝑅𝑅(x,y)}|
16 𝑴𝑴𝑘𝑘 𝑚𝑚,𝑛𝑛 =Σ 𝑂𝑂M𝑘𝑘∈𝑅𝑅 |{𝑅𝑅(x,y)}|
17
18 The travel demand at a moment is regarded as a snapshot. This snapshot is an M×N matrix. Snapshots over
19 a period of time form a snapshot stream. It can be represented by a tensor X∈ℝ𝑡𝑡×𝑁𝑁×𝑁𝑁.
20
21 Definition 3 (Temporal variables)
22 The temporal variables are the features that differ randomly across time, but not among the grid zones. Such
23 variables are not evenly periodic, but may show some level of periodicity. Weather is such a temporal
24 variable. We have included many weather features in our model like cloud cover, dew point, humidity,
25 weather, temperature, visibility and wind speed. It will be called as 𝑤𝑤𝑡𝑡 𝑚𝑚,𝑛𝑛 , which will generate the weather
26 matrix of zone (𝑚𝑚,𝑛𝑛) of time t, 𝑾𝑾𝑘𝑘 𝑥𝑥,𝑦𝑦
27
28 Definition 4 (Context variables)
29 The context variables are either periodic or fixed across time. For spatio-temporal forecasting,
30 these variables are either repeated across the zones or repeated across the time-slots.
31
32 (1) Temporal context: Through empirically analyzing the distribution of demand rate with respect to
33 training time, 24 hours a day was intuitively separated by Ke et al. (8) into 3 periods: peak hours, off-peak

5
 Author, Author, and Author

1 hours and sleep hours. We have denoted the hours based on the empiric demand strength and classify the
2 top 8 h, middle 8 h, bottom 8 h as peak hours, off-peak hours and sleep hours. We also add the dummy
3 variable 𝑐𝑐𝑐𝑐𝑡𝑡 to describe this time-of-day parameter., given by
4
2, 𝑖𝑖𝑖𝑖 𝑡𝑡 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝑜𝑜𝑜𝑜𝑜𝑜 𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 (4𝑝𝑝𝑝𝑝 − 12𝑎𝑎𝑎𝑎)
5 𝑐𝑐𝑑𝑑𝑡𝑡 = �1, 𝑖𝑖𝑖𝑖 𝑡𝑡 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 (8𝑎𝑎𝑎𝑎 − 4𝑝𝑝𝑝𝑝)
0, 𝑖𝑖𝑖𝑖 𝑡𝑡 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑠𝑠 (12𝑎𝑎𝑚𝑚 − 8𝑎𝑎𝑎𝑎)
6
7 We also denote that another dummy variable 𝑐𝑐𝑐𝑐𝑡𝑡 is the day-of-week, which captures the differentiating
8 properties between weekdays and weekends.
9
0, 𝑖𝑖𝑖𝑖 𝑡𝑡 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤
10 𝑐𝑐𝑤𝑤𝑡𝑡 = �
1, 𝑖𝑖𝑖𝑖 𝑡𝑡 𝑖𝑖𝑖𝑖 𝑖𝑖𝑖𝑖 𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑤𝑒𝑒𝑒𝑒𝑑𝑑𝑑𝑑
11
12 The context variables are repeated throughout the zones to create daytime matrix 𝑪𝑪𝑫𝑫𝑘𝑘 𝑥𝑥,𝑦𝑦 and weekday
13 matrix 𝑪𝑪𝑾𝑾𝑘𝑘 𝑥𝑥,𝑦𝑦 , which can be merged and resulted in temporal context matrix 𝑪𝑪𝑻𝑻𝑘𝑘 𝑥𝑥,𝑦𝑦
14
15 (2) Spatial context: The spatial context refers to the number of POIs across the zones. It is going to be
16 denoted as 𝑐𝑐𝑐𝑐𝑚𝑚,𝑛𝑛 , which is fixed across whole timeline. POI has many categoiers, and the categoires are
17 going to be one-hot encoded. Here, it is going to create spatial context matrix, represented as 𝑪𝑪𝑺𝑺𝑘𝑘 𝑥𝑥,𝑦𝑦 .
18
19 Problem formulation
20 The goal of this analysis is to enhance the predictive method's functionality in terms of precision,
21 computational efficiency, temporal granularity, and the scalability of space. We will split the city region
22 into smaller areas for predicting demand while also providing adequate data points for model training in
23 each field. We are going to divide two cities into 10×10 grid. Moreover, highly compact dataset also made
24 it possible to take smaller time segments than standard hourly partition, resulting in more data points. We
25 took time interval of 15 minutes.
26 Another significant thing about our model is to provide temporal and context details for the learning model,
27 which will improve the precision of the forecast. For example, environmental conditions have a significant
28 effect on riding-hailing demands, e.g. there might be more or less trips when it's cloudy or rainy.
29 Additionally, we are including POI data and week-day data in our modelling, which definitely impacts on
30 the demand of the trip requests. For instants, there might be more people around parks and other leisure
31 activity areas on the weekend, where on week days peak hours will see huge traffic around office and
32 educational institutions.
33 In this article, we are trying to predict the travel demand based on the available historical spario-temporal
34 variables of the two cities for every moment in different regions.
35 Based on the above considerations, we defined the problem that needs to be solved as:
36 Given the historical observations of travel demand 𝑫𝑫𝑘𝑘 𝑥𝑥,𝑦𝑦 , speed 𝑺𝑺𝑘𝑘 𝑥𝑥,𝑦𝑦 , mobility 𝑴𝑴𝑘𝑘 𝑥𝑥,𝑦𝑦 , weather data
37 𝑾𝑾𝑘𝑘 𝑥𝑥,𝑦𝑦 , context variable matrix 𝑪𝑪𝑻𝑻𝑘𝑘 𝑥𝑥,𝑦𝑦 , 𝑪𝑪𝑺𝑺𝑘𝑘 𝑥𝑥,𝑦𝑦 as input, to create a prediction model that outputs 𝑫𝑫𝑡𝑡 𝑥𝑥,𝑦𝑦 .
38 Here, k ≥ 1 is the number of past time steps used for prediction. In this study, k ∈ [1,9] , N is 10, so the area
39 is divided into 10×10 grid and the duration for each time slot, M is 15 min. Therefore, the past 120 min
40 ride-hailing demands and other variable records are used to predict ride-hailing demand at time t.
41
42 METHODOLOGY
43 A novel deep learning architecture named XYZ is developed in this experiment for travel demand
44 forecasting. Our data includes spatio-temporal variable, temporal variables and context variables, and
45 different available deep learning architectures are better at analyzing different variables. This is why we
46 designed the whole architechture merging three different deep learning algorithms (CNN, LSTM and
47 ConvLSTM) to tackle the varity of data.

6
 Author, Author, and Author

1
2 Convolutional Neural Network (CNN)
3 CNN is a specially designed neural network for spatial pattern recognition. Here, we are going to use 1D
4 CNN for our calculations. This system works by sliding different filters over the data, and different filters
5 can detect different patterns from the output. A filter on a portion of the data multiplies with the underlying
6 values and thus gets the desired output. If there are K filters in a one-dimentional CNN layer, then the
7 output can be denoted as
8 Z=
k σ ( ( X t ∗Wk ) + bk )
9 Here ∗ is convolytional operator, X t is the input matrix convolved with the filter k ∈ K , Z k refers to the
10 output feature vector for the filter, Wk serves as the shared weight matrix of the filter, and bk is the bias
11 vector of the filter.

12
13 Figure 2 Example of one-dementional CNN
14
15 For illustration, the inputs and outputs of a one-dimensional CNN with two Convolutional layers are shown
16 in Figure 2. The first layer has two filters, where the later layer has one. The putputs of the layers are the
17 columns of features. To keep the output always the same size as the input, zero padding is used. It is a
18 common practice to use pooling layer to make the output smaller and faster calculation, but it tends to lose
19 detail of the pattern, thus we decided to ignore the norm.
20 As the filters operates throughout a centre cell and some adjacent cells at a time, it can recognize the
21 interrelation among the nearby cells, in short, CNN is efficient for spatical pattern recognition.
22
23 Long-Short Term Memory (LSTM)
24 LSTM or Long-Short Term Memory is a modified Recurrent Neural Network (RNN) structure to address
25 the classical problem related with backpropagation of RNN: vanishing gradient and exploding gradient.
26 LSTM has some logic gates in the structure, which prevents the gradient issues happening in the first place.
27 Moreover, it has two kind of activation function: tanh and sigmoid ( σ ) activation.
28 Where tanh regulates all outgoing values from -1 to 1, σ converts values in the range of 0 to 1.
29 An LSTM cell contains the following components:
30 • Forget Gate F (a neural network with σ )
31 • Candidate layer C (a neural network with tanh )

7
 Author, Author, and Author

1 • Input Gate I (a neural network with σ )

2 • Output Gate O (a neural network with σ )
3 • Hidden state H (a vector)
4 • Memory state C (a vector)
5 The main equations for LSTM are like this, where  represents the Hadamard product:
6 Ft = σ ( X tU f + H t −1W f + b f )
7 I t = σ ( X tU i + H t −1Wi + bi )
8 Ot = σ ( X tU o + H t −1Wo + bo )
9 =
C t tanh ( X tU c + H t −1Wc + bc )
10=Ct ft  Ct −1 + I t  C t
11 H t = Ot  tanh ( Ct )
12 Here is the diagram for LSTM cell at the time step t

13
14 Figure 3 An LSTM cell
15 First, the previous hidden state and the current input get concatenated and served with weights ( U , W ) and
16 biases ( b ), which gets used by many later components of LSTM cell. This data gets fed into the forget
17 layer F . This σ activated layer removes non-relevant data. A candidate layer C is created using the
18 concatenated data with tanh activation. This layer generates values to add to the cell state. Concatenated
19 data also gets fed into the input layer. The input gate I decides what data from the candidate should be
20 added to the new cell state. After computing the forget layer, candidate layer, and the input layer, the cell
21 state is calculated using those vectors and the previous cell state. The output is then computed through
22 output gate O . Pointwise multiplying the output and the new cell state gives us the new hidden state.
23 An LSTM cell has 12 parametes: b f , W f , U f , bi , Wi , U i , bo , Wo , U o , bc , Wc , U c

8
 Author, Author, and Author

1 One benefit of using the memory cell and gates to the information flow is that the gradient would be stuck
2 in the cell (also known as constant error carousels) and stopped from disappearing too easily, which is a
3 crucial downside for the basic RNN model.
4 As LSTM contains a memory state, this neural network can attain pattern related with repeatation, which
5 means LSTM is good with temporal pattern recognation.
6
7 Convolutional LSTM (ConvLSTM)
8 ConvLSTM is a joint architecture of CNN and LSTM, which means it can find out both spatial and temporal
9 features, which means it is suitable for analyzing travel deamand data.
10
11 The main equations for ConvLSTM are mostly like LSTM equations previously described with some
12 modifications due to the addition of convolution. Here ∗ is Convolutional operator, and  as before
13 denotes the Hadamard product:
14 F=
t σ ( X t ∗ U f + H t −1 ∗ W f + b f )
15 I=
t σ ( X t ∗U i + H t −1 ∗Wi + bi )
16 O=
t σ ( X t ∗U o + H t −1 ∗Wo + bo )
17 C t tanh ( X t ∗ U c + H t −1 ∗ Wc + bc )
=
18=Ct ft  Ct −1 + I t  C t
19 H t = Ot  tanh ( Ct )
20 And a ConvLSTM cell looks like this:

21
22 Figure 4 1DConvLSTM cell
23
24 In 1DConvLSTM, inputs X 1 , X 2 ,..., X t , cell outputs C1 , C2 ,..., Ct , hidden states H1 , H 2 ,..., H t and gates
25 I t , Ft , Ot of the ConvLSTM are 3D tensors whose last two dimensions are spatial dimensions (rows and

9
 Author, Author, and Author

1 columns). Here, the inputs, outputs and states may be imagined as vectors standing on a spatial grid. The
2 ConvLSTM predicts the upcoming state of a cell by the inputs and past states of its own and the adjacent
3 cells, determined by the kernel size. For this, in the inter-state and input-to-state steps a convolution operator
4 is used (see Fig. 2).

5
6 Figure 5 Flow of information through ConvLSTM layers
7
8 Architecture of XYZ
9

10
11 Figure 6 Workflow of XYZ
12 A comprehensive overview of the XYZ architecture for the processing of various variable types (i.e., spatio-
13 temporal variables, temporal variables, and context variables) is given in this section. XYZ architecture
14 workflow is shown in Figure 4. At first the spatio-temporal variable, temporal variable and context variables
15 are provided as the inputs. The spatio-temporal featues are extracted by two layers of 1DConvLSTM

10
 Author, Author, and Author

1 stacked together. Then the output is concatenated with the spatial context data, which is then passed through
2 two layers of one-dimentional CNN for feature extraction. At the same time, temporal variables are fed into
3 two layers of LSTM, followed by a repeat vector to distribute the feature all throughout the zones at a given
4 time, as temporal variables vary not with space, but only with time. Therefore, in the next step the output
5 of 1DCNN, repeat vector and temporal context is concatenated into one matrix and passed into a dense
6 layer for demand prediction, thereby compared with the ground truth data while training for assessing the
7 loss and therefore, efficiency of the model.
8
9 Algorithm 1. Training of XYZ
10 Input: Training set observations of spatio-temporal variables D1x , y , D2 x , y ,...DT x , y , ( )
11 (S 1
x, y x, y
, S 2 ,...ST x, y
) and ( M 1
x, y x, y
, M 2 ,...M T x, y
)
12 Training set observations of temporal variables W1x , y , W2 x , y ,...WT x , y( )
13 Training set observations of spatial context CS1 ( x, y x, y
, CS 2 ,...CST x, y
)
14 (
Training set observations of temporal context CT1x , y , CT2 x , y ,...CTT x , y )
15 Past observation limit k ∈ [1,9]
16 Output: Trained XYZ with learned weights and biases
17 1: Procedure XYZ training
18 2: need to edit
19 3: need to edit
20 4: End procedure
21
22
23 EXPERIMENTS AND DISCUSSION
24 This section sheds light on the details of the experiments and the results obtained. Moreover, model
25 structure and sensitivity analysis will be discussed elaborately, followed by the interpretations of the XYZ
26 models.
27
28 TABLE 1 Hyperparameter Settings of XYZ

Need to edit Need to edit Need to edit

29
30

11
 Author, Author, and Author

1
2 Benchmark models
3
4 Evaluation of the model
5 In the experiment the dataset was devided into three catagories: training data, validation data and test data.
6 (timeslot split 4024:900:900). Here, in every epoch the model was trained on training data and the model
7 was evaluated on validation data. When the training was finished, the model was run with test data to test
8 the error in the demand prediction. The error was calculated using these equations:
1 N ×N
( )
2
 x, y
9 =
Mean Absolute Error, MAE ∑
n t =1
Dt x , y − Dt

1 N ×N
( )
2
 x, y
10 =
Root Mean Absolute Error, RMSE ∑
n t =1
Dt x , y − Dt

11 Mean Absolute Percentage Error, MAPE =

100% N × N
∑
(D
t
x, y 
−Dt
x, y
)
n t =1 
D
x, y
t

12  x , y is the true value and N × N is the total geographical zones.

Here Dt x , y is the predicted demand, Dt
13
14 Platform
15
16
17
18 Model comparison
19
20
21

12
 Author, Author, and Author

1
2 CONCLUSION
3
4 ACKNOWLEDGMENTS
5 Nunc eros felis, dapibus eu nisl convallis, pellentesque facilisis sapien. Duis vel magna ac ligula
6 rutrum volutpat. Donec sed malesuada ipsum. Morbi sed metus sem. Etiam bibendum posuere elementum.
7 In a tempor mauris. Mauris congue in diam ac malesuada. Praesent sed dignissim erat. Mauris sagittis lectus
8 nec ornare condimentum. Etiam est ante, tristique vitae tellus id, venenatis tincidunt tellus. In hac habitasse
9 platea dictumst. Interdum et malesuada fames ac ante ipsum primis in faucibus. Donec quis pellentesque
10 augue, ac elementum turpis. Quisque nec tristique nunc, in porttitor magna. Donec eget laoreet tortor,
11 euismod aliquam felis.
12
13 AUTHOR CONTRIBUTIONS
14 The authors confirm contribution to the paper as follows: study conception and design: X. Author, Y.
15 Author; data collection: Y. Author; analysis and interpretation of results: X. Author, Y. Author. Z. Author;
16 draft manuscript preparation: Y. Author. Z. Author. All authors reviewed the results and approved the final
17 version of the manuscript.

13
REFERENCES