Movement generation and trip distribution

for freight demand modelling applied to city logistics

Guido Gentile

Università degli Studi di Roma “La Sapienza”, Dipartimento di Idraulica Trasporti e Strade

Via Eudossiana 18 - 00184 Roma, Italy, guido.gentile@uniroma1.it

Daniele Vigo

Università di Bologna, Dipartimento di Elettronica Informatica e Sistemistica

Via Venezia 52- 47023 Cesena (FC), Italy, dvigo@deis.unibo.it

1
 ABSTRACT

The quantification of the freight movements occurring in each traffic zone disaggregated by

supply chain and of the corresponding Origin-Destination trip tables connected with the urban

delivery of goods, is essential to evaluate the effects of any city logistic policy in terms of vehicle

congestion and polluting emissions, based on the assignment of truck flows to the road network. To

this end in this paper we propose two innovative demand models: the first one for movement

generation, the second one for trip distribution.

The movement generation is addressed through an extension of category index models, which

takes into account the hierarchy within the classification system of the economic activities, thus

avoiding to aggregate the many classification codes into pre-specified groups. The trip distribution

is addressed through an adaptation of gravity models, which takes into account that deliveries are

organized into tours.

This approach has been successfully applied to the case of Emilia-Romagna Region in Italy,

where an extensive campaign of surveys has permitted to calibrate both models for several towns

and cities with various dimensions and vocations.

Keywords: urban delivery of goods, estimation of freight O-D matrices, category index models

with hierarchy, gravity models with tours.

2
 INTRODUCTION

A relevant issue in city logistics is concerned with the efficient transport of products, from

industrial and stocking sites to retail local units, and then to offices and houses. To better

understand this problem it is worth examining the physical itinerary of important items such as

grocery and household articles, which generally make up more than one third of the total freight

traffic in urban areas (see, e g. Strauss-Wieder et al, 1989). The goods are transported to the shops

in the city either by private carriers (i.e., by the manufacturing companies themselves), or by

logistic agencies, each owning a fleet of vehicles to provide the delivery service, as well as by

autonomous transporters, often owning a single vehicle. In many cases along this journey

transhipment occurs at logistic centres, located near or within the city, where load rearrangement is

performed, i.e., the goods are dropped off from heavy trucks, possibly stocked, and then transferred

to smaller vehicles for the distribution in town. These flows of goods are organized into different

supply chains depending on their characteristics and service requirements (e.g., fresh foods, frozen

foods, dry foods, household articles, …). In the last step, which is not addressed in this paper, the

goods are purchased by final consumers who personally bring them home.

City logistics may be analyzed by two different points of view, each one rising different

problems. For logistic agencies the most relevant aspect is that of minimizing the shipment

industrial costs. Therefore, one of the main problems is that of finding the optimal routes to be used

by the vehicles that shall provide the delivery service. This family of problems, know as Vehicle

Routing Problems (VRPs), has been extensively studied in the context of Operations Research.

Several algorithms for the solution of the main variants of VRP have been proposed in the literature

(see Toth and Vigo, 2002; and Cordeau et al., 2007; for recent surveys) and some of them are

implemented within commercial packages for pickup and delivery planning. Vehicle routing models

are highly disaggregate and require fine-grained data on demand distribution and on the service

requirements. They are mainly suitable for supporting the operational management of a delivery
3
service, or analyzing mid-term problems such as fleet sizing.

On the other hand, for the Public Administration the most important aspect is that of moderating

the social costs generated by freight mobility in the urban area. These are primarily connected with

the impact on traffic congestion of the circulating vehicles for delivery tours, as well as with the

road capacity reductions caused by the stationary vehicles for loading or unloading operations, also

called movements in the following. Then the problem consists either in increasing the number of

logistic centres or in establishing effective policies for the regulation of freight traffic. This latter

option may be implemented by imposing temporal and/or spatial limitations to circulation and/or

parking of trucks and by creating dedicated stopping areas to let freight vehicles perform the

movements.

The quantification of freight demand in terms of movements and flows on each link of the road

network due to the urban delivery of goods is essential for evaluating the effects of any city logistic

policy. In the following we briefly review the main approaches that were proposed in the literature

to model the transport of goods in the urban case. For a detailed survey the reader is referred to

Russo and Comi (2004) and to the extensive bibliography compiled by the Oak Ridge National

Laboratory (2004) for the Freight Model Improvement Program.

In general, the demand models used in the context of freight transport can be classified into three

main types: gravity (Hutchinson, 1974; Ogden, 1992; Taylor, 1997; He and Crainic, 1998; Gorys

and Hausmanis, 1999), spatial price equilibrium (Oppenheim, 1995; Nagurney, 1999), and input-

output (Harris and Liu, 1998; Marzano and Papola, 2004). In any case, the resulting O-D matrices

can be adjusted to match truck link flows. However, some of the above approaches seem to be more

suitable for regional and national planning, than to the context of urban freight shipment, since

many factors (e.g. historical traditions, urban constraints) beside accessibility influence the location

of shops.

Demand models may be further divided into truck-based and commodity-based approaches.

4
Truck-based models (see Zavettero and Waseman, 1981; Habib, 1985; Ogden, 1977; Slavin, 1976)

yield a direct estimation of the trip rates, but suffer transferability problems. Although the average

quantities of consumed goods per inhabitant usually do not vary too much from one city to another

in a same country, the organization of logistics in terms of shop and truck dimensions may indeed

differ considerably.

Commodity-based models (see Taniguchi and Thompson, 1999, 2003 and 2006 Holguin-Veras

and Thorson, 2000) estimate the quantity of goods sold in each traffic zone, which is strictly

correlated to the number of employees in the commerce sector, under the assumption that consumer

demand, which is instead related to the number of residents, must be satisfied, thus reproducing the

process of acquisition and consumption of goods. As a counterpart, commodity-based models

require a loading model to convert the tons of goods into the number of vehicle trips (Noortman,

1984, for example, uses a fixed share of the truck capacity, which may vary according to the

commodity type, although such a simplification is not acceptable for most of the goods travelling in

the urban context, as observed by Button and Pearman, 1981), as well as a complementary empty

trip model to take into account this crucial aspect of logistics (as described, for example, in:

Hautzinger, 1984; Holguín-Veras and Thorson, 2003a and 2003b; Holguín-Veras et al., 2005).

Input-output and spatial price equilibrium models are typically commodity-based or even monetary-

based; in the latter case a value to quantity model is required to convert the currency amounts into

tons of goods.

In practice, most urban models are based on the same “four-steps” approach of sequential models

(generation, distribution, modal split, and assignment) used for passenger transport. For example,

the demand generation and attraction of every supply chain is first determined for each traffic zone

through a regression model, then a gravity model for distribution is applied to obtain an Origin-

Destination (O-D) matrix to be assigned on the road network taking into account access restrictions,

while mode choice usually reduces to the vehicle dimension which depends strongly on the supply

5
chain.

This classical approach involves two open questions. The first one is intrinsic in the

approximation introduced by arbitrarily aggregating the many different economic activities into a

few groups of categories. The second one lays in the fact that many deliveries are performed by a

commercial vehicle within the same tour, so that before assigning the demand flows to the network

one has to transform the pickups and deliveries into an O-D matrix of direct truck trips. Both issues

will be addressed in this paper and a possible solution will be proposed.

The paper is organized as follows. In the next section we outline the overall approach underlying

the proposed models. In Section 2 we describe the generation model, whereas in Section 3 we

describe the distribution model. Finally, Section 4 presents the application of the overall

methodology to the city of Bologna, capital of the Emilia-Romagna Region, and Section 5 draws

some conclusions.

1. THE OVERALL APPROACH

The quantification of the external effects produced by freight mobility in urban areas is currently

an important field of research but, in contrast to what has happened with people mobility, no

consolidated methodology of analysis is yet available. In order to fill this gap some attempts were

recently carried out in Europe. The information collected from some German towns within the

Bestufs Project, financed by the European Union, led to the Wiver model (Meimbresse and

Sonntag, 2000), now integrated within the Viseva software package (Friedrich et al., 2003). The

Good Trip model (Boerkamps and Van Binsbergen, 1999) has been developed in The Netherlands,

while the French Government promoted the development at the Laboratoire d’Economie des

Transportes in Lyon of the Freturb model (Routhier and Aubert, 1998; Ambrosini and Routhier,

2004). In this latter case, the statistical analysis of the data collected through surveys performed in

three important cities (Marseilles, Bordeaux and Dijon) produced an aggregated model that allows

for evaluating a set of impact indicators, based on the average number of movements produced by
6
each category of economic activity as a function of the employees. For example, to evaluate the

occupation of parking slots for freight movements it is assumed that the total number of stopping

hours spent within a given traffic zone is the sum, over all categories of shops and vehicle types, of

the products among: a) the number of shops of that type present in the zone; b) the number of

movements each shop generates for that type of vehicle; c) the average duration of such

movements; and d) an equivalence factor typical of the vehicle.

In this paper we present a new model for the overall analysis of urban freight demand and the

evaluation of its external effects which originates from the intensive collaboration between the

authors and the Council on Transportation of the Emilia-Romagna Region in Italy, that in the recent

years promoted an extensive survey, partially financed by European projects CityPorts and Merope,

on the logistic activities in nine out of its thirteen towns with more than 50,000 inhabitants. Almost

all surveys were performed according to the methodology developed within the CityPorts Project, in

which the phenomenon of urban freight shipment is faced from different points of view, thus

including specific interviews to commercial local units, to logistic operators and classical on-road

interviews to truck drivers. The result is an extensive and homogeneous data source that represents

an almost unique modelling opportunity.

The CityPorts Project proposes a methodological framework for the planning of initiatives to

support city logistics, such as new transhipment infrastructures or regulatory policies, referred to as

logistic actions, whose relative impacts are measured through the so called traffic Zones-Supply

chains (Z-S) matrix associated with the town under study (Gentile and Vigo, 2006). Each element

of the Z-S matrix yields the total number of, say annual, movements for a given supply chain in a

certain traffic zone. Indeed, any specific logistic action affects only a subset of traffic zones and of

supply chains. For example, access restrictions to freight vehicles may be applied to the city centre

and not to other zones, while a logistic centre may serve only a part of the whole town and some

supply chains. Therefore, the overall impact of a logistic action may be easily determined by

7
computing the total number of movements that it “covers” on the Z-S matrix.

The determination of the Z-S matrix as well as its use as a starting point to obtain the Origin-

Destination matrices of truck flows, that can be assigned on the road network, will be addressed in

the remainder of this paper by proposing specific generation and distribution models. These two

models aim, respectively, to find the movements generated by each traffic zone for each supply

chain, and the direct truck trips from each origin zone to each destination zone between two

consecutive vehicle stops, taking into account the structure of the delivery tours, that in general

depends on the supply chain at hand. To this end it should be considered that: a) each local unit

(e.g., shop, office …) generates movements belonging to several supply chains; b) local units are

classified in the available databases according to a hierarchic code that identifies with successive

specifications the sector of economic activity (e.g., the NACE system adopted in Europe, and the

NAICS system adopted in North America); c) the generated movements are organized into truck

tours that begin and end at given portals (logistic centers, or access roads to town) and may serve

local units in several traffic zones.

The main difficulty that prevents the use of a direct statistical analysis to determine the number

of movements for each supply chain generated by every activity code is represented by the fact that

generally the classification systems include hundreds of different codes (e.g., in the NACE system

more than 1500). Therefore, the universe of all possible individuals (in our case all possible activity

codes) is typically partitioned into groups, or categories, and an average for each group is retrieved

from a sample of individuals. However, when many categories are defined one cannot expect the

sample to cover uniformly and satisfactorily each group.

Category index models are an evolution of the above approach based on simple aggregation. The

main idea is to take into consideration the correlation among categories, if present, so as to provide

a better estimate for groups that are not adequately covered by the sample. The classic case is when

the categories are defined on the basis of several attributes, where the correlation among categories

8
arises naturally from sharing the same value for some attribute. In our case, however, there is only

one attribute available, namely the activity code.

In this paper we introduce a new demand generation model for the determination of urban freight

movements that extends the category index approach avoiding any a-priori aggregation of activity

codes into a predefined limited number of groups and assuming as correlated those categories that

have some relationship in the tree of activity codes, thus exploiting the hierarchical structure of the

only explanatory attribute we use. Our approach is indeed based on the explicit consideration of the

tree-based structure adopted by most classification systems for local units, that typically include a

progressive levels of disaggregation represented by the subsequent digits in the code associated

with each economic activity. The main assumption here is that the number of movements generated

by a given activity code is correlated with those generated by activity codes that are associated with

its descendant and ascendant nodes in the classification tree, i.e., those representing its

disaggregations or aggregations.

The second contribution of this paper is a distribution model that implements the gravity

paradigm in the case of urban freights, taking into account the average number of deliveries made

along shipping routes of each supply chain. Every route of a vehicle serving a certain supply chain

is assumed to begin at some portal, make a specific number of stops to perform the deliveries (or

the pickups, in the case of reverse logistics), and end at the same portal. Therefore every section of

the tour between two successive stops, included the first and the last leg from and to the portal

respectively, must be associated with a direct trip in the O-D matrix from the traffic zone of the

previous stop to the traffic zone of the following stop. In other words, each movement occurring in

a given traffic zone must be associated to one direct trip that terminates in that zone and with

another direct trip that originates in the same zone. For each supply chain, the total number of tours

is given by the total number of movements divided by the average number of stops per tour. These

tours are associated to the different portals proportionally to their relative importance, measured for

9
example by the truck traffic that they generate. Then, the number of direct trips originating and

terminating at each portal is equal to the number of tours associated with it. The probability that a

tour leaving a given zone will reach another zone through a direct trip is assumed to decrease with

the distance on the network between the two zones, for example measured by the generalized travel

cost, thus implementing the gravity paradigm.

Without loss of generality we will refer to the case where the operations performed by a same

vehicle are homogeneous deliveries of the same supply chain, since from a modelling point of view

the case where the operations are pickups, as in reverse logistics and garbage collection, is

equivalent to the former.

The two demand models, along with a classical assignment model, have been integrated into a

software tool called CityGoods, and were calibrated with the data of the Emilia-Romagna surveys.

2. THE GENERATION MODEL

In this section we describe a demand generation model aimed at the estimation of the number of

trips produced by each local unit for freight shipment and collection activities in urban areas. It

should be noted that in the case of freight urban shipment the movement generators are actually the

destinations of the goods, i.e., the shops, whereas in modeling the passenger transport the traffic

generators are typically the origins of the trips, i.e., the homes, since we usually refer to morning

commuters.

The main characteristic of the model we propose is that it does not require any a-priori

aggregation of the local units into a limited number of categories. Indeed, our model explicitly

considers a fine-grained classification, such as that informing the coding systems adopted by most

countries to identify the economic activity in census statistics. Classification systems have normally

a hierarchic structure which reflects the progressive disaggregation of the economic sectors into

sub-sectors; specifically, each disaggregation level is associated with an additional digit of the code

that identifies the economic activity in the system. For example, the Italian specification of the
10
NACE system, called ATECO, identifies 15 main categories further disaggregated into sub-

categories with up to five hierarchic levels.

The hierarchic relationship between a sub-category and its immediate aggregation in the higher

level is clearly unique, i.e., each sub-category is originated from the disaggregation of a single

category; hence, the portion of a classification system relative to a main category may be suitably

represented as a tree. The union of these category trees, i.e., the whole classification system, can be

then represented as a forest.

Formally, the classification system is here represented as a tree T = (N, A), i.e., an acyclic

oriented graph, where the set N includes a node for every category plus a special root node r , which

is the only node with no entering arcs, while all the other nodes have a single entering arc. The final

nodes of the arcs exiting from the root represent the main categories, which can be considered

independent from each other in terms of any relevant feature. The hierarchic relationships among

economic activities within the classification system are represented by the set of oriented arcs, A,

connecting the nodes. More precisely, each node i∈N\{r} is associated with another node, called

father, which represents its immediate higher level in the hierarchic structure and is denoted by

f(i)∈N. Therefore, the set of oriented arcs constituting the tree is A = {(i, j): j∈N\{r}, i = f( j)}. The

predecessors P(i) of node i∈N\{r}, are the nodes included in the unique path in T connecting the

root node r with node i, except the root itself, i.e., P(i) = {…, f(f(i)), f(i), i}. Similarly, the

successors S(i) of node i∈N are all the nodes whose predecessors include i, except node i itself, i.e.,

S(i) = {j∈N: i∈P( j)}\{i}. The children, FS(i) of node i∈N are the nodes whose father is i, i.e., FS(i)

= { j∈N: f( j) = i}. The leaves L of the tree are the nodes having no children, i.e., L = {i∈N: FS(i) =

∅}. This mathematical representation of a classification system is exemplified in Figure 1.

11
 P(11) FS(11)

r 1 11 111

12 112

13 S(2)

2 21 211

22 212

213

3 31
f(31)
32 321

Figure 1. Graph representation of a classification system. The root is black, the leaves are grey.

Given the universe U of local units that are present in the specific area under analysis, each node

i∈N\{r} is associated with the probability βi that it is the child of his father f(i), to be evaluated as

follows. Each local unit u∈U is characterized by a category n(u)∈N. Let ωi be the number of local

units in the universe belonging to the category associated with node i∈N, and Ωi be the number of

local units belonging to any category associated with a successor of node i and with i itself, that is:

Ωi = ωi + ∑ j∈S(i) ωj = ωi + ∑ j∈FS(i) Ωj. (1)

Then, the probability βi is computed as:

βi = Ωi / ∑ j∈FS( f(i)) Ωj = Ωi / (Ωf(i) - ωf(i)). (2)

The number of local units Ωi and the probabilities βi may be computed in linear time by visiting the

tree in reverse topological order, i.e., from the leaves up to the root. We can then define the

relationship ψji of any successor j∈S(i) of node i∈N with i itself as the probability that j is the

descendant of i, expressed as a product of conditional probabilities at nodes:

ψji = ∏ h∈ P(j)\P(i) βh . (3)

12
 The main assumption of the generation model we propose is that the number of movements

associated with a given activity code of the classification system, i.e., with a given node of the tree,

is correlated with the number of movements generated by the codes of its predecessors and

successors nodes. To specify the form of this correlation we assume that a non-negative number of

movements mj is associated with each arc of the tree (f( j), j), j∈N\{r}, so that the number of

movements Mi generated by the generic node i∈N in the reference period (e.g. one year) is obtained

as the sum of two different terms: Hi , given by the sum of the movements associated with the arcs

entering each predecessor j∈P(i) of i, and Wi , given by the sum of the movements associated with

the arcs entering each successor j∈S(i) of i suitably weighted by their relationship ψji with i. In

other words, although the main output of the model is the number of movements Mi produced by

each node of the tree, these variables are correlated among each other due to the hierarchic structure

of the classification system. Therefore, the relevant independent variables to be estimated become

the number of movements mj associated with each arc of the tree, that express the contribution to

the demand generation common to node j and to all and only its successors. Indeed, each arc of the

tree represents an additional specification of economic activity that identifies a category, and thus

contributes with a specific number of movements to its demand generation as well as to those of all

the sub-categories obtained through its further disaggregation. Based on the above assumptions any

portion of the classification system relative to each main category, i.e., every tree of the forest, is

independent from the others in terms of movement generation.

In formal terms, we assume that for each i∈N it is:

Mi = Hi + Wi , (4)

where

Hi = ∑ j∈P(i) mj , (5)

Wi = 0, if i∈L, and Wi = ∑ j∈S(i) ψji ⋅ mj , otherwise. (6)

To clarify the physical interpretation of the model we can observe that: a) the number of

13
movements generated by a leaf is given by the sum of the movements associated with the arcs

entering its predecessors; b) the correlation among two leaves is given by the sum of the movements

associated with the overlapping arcs in the two paths connecting the root to them; c) a node which is

not a leaf synthesizes all the leaves in its successors, indeed it generates the average number of

movements relative to the leaves in its successors weighted by the relationship of each leaf with it.

Assertions a) and c) are formally expressed by the following equations:

Mi = ∑ j∈P(i) mj , i∈L, (7)

Mi = ∑ j∈S(i)∩L ψji ⋅ Mj , i∈N\L. (8)

While (7) derives directly from the model (4)-(6), (8) could be easily proved through a recursive

argumentation.

For the two contributions, respectively, we can state the following local and recursive formula:

Hr = 0 , Hi = mi + Hf(i) i∈ N\{r} , (9)

Wi = 0, i∈L ; Wi = ∑ j∈FS(i) βj ⋅ (mj + Wj), i∈N\L . (10)

Indeed, it is:

Hi = mi + ∑ j∈P(f(i)) mj , (11)

Wi = ∑ k∈FS(i) βk ⋅ ∑ j∈S(k)∩L (∏ h∈P(j)\P(k) βh) ⋅ (mk + ∑ h∈P(j)\P(k) mh). (12)

The triangular system (9) can be solved by processing the nodes in topological order, while the

triangular system (10) can be solved by processing the nodes in reverse topological order. On this

basis the model can be solved in linear time.

The system (4)-(6) expressing the dependent variables Mi in terms of the independent variables

mj is linear. Therefore we can also write:

Mi = ∑ j∈N πij ⋅ mj , i∈N , in compact form M = π ⋅ m , (13)

where πij represents the unit contribution of mj to Mi , which can be determined as follows.

Three different cases may occur:

a) j∉ P(i)∪S(i) , then πij = 0;

14
 b) j∈P(i) , from (5) we have that πij = 1;

c) j∈S(i) , from (6) we have that πij = ψji.

The matrix π can be computed directly from the probabilities βi without introducing explicitly the

relationships ψji as shown in the algorithm presented below.

π=0
for each i∈N
ψ=1
j=i
loop until j = r
ψ = ψ ⋅ βj
πij = 1
j = f( j)
πji = ψ

The formulation as a linear system turns out to be useful for the calibration of the model. Let zq

be the movements generated by the generic element q∈Q of the sample Q⊆U. The calibration

process can be addressed by solving the following least square problem through the algorithm

NNLS - NonNegative Least Squares of (Lawson and Hanson, 1995):

min{||z - Π ⋅ m||2: m ≥ 0} , (14)

where Π = [πn(1)T , … , πn(q)T, … , πn(|Q|)T]T, πi is the generic i-th row of matrix π, with i∈N, and

z = (z1 , … , zq , … , z|Q|)T. The proposed model introduces many parameters to be determined.

However, in practice the number of movements associated with an arc is left to zero by the

calibration process whenever no element of the sample requires to introduce that specific

disaggregation for distinguishing its category from those in the rest of the sample. In such a case,

the number of movements generated by that category is equal to the number of movements

generated by its father.

The proposed model can be considered a particular application of the category index approach,

since for each economic activity we deduce the rate of freight movements produced for every

supply chain by averaging the sample data through a procedure that introduces a correlation among

the category codes, taking into account the hierarchic structure of the classification system.
15
 This model does not exploit any attribute of the local unit other than its code to explain the

magnitude of the generation phenomenon. A major improvement can then be achieved by

considering some attribute expressing the dimension of the local unit, such as the number of

employees or the available selling surface; in general, E can be a size function (see Ben Akiva and

Lerman, 1985).

The proposed approach is easily extended to obtain a linear model with respect to a given

dimension attribute E as follows:

Mi + E ⋅ Xi = ∑ j∈N πij ⋅ (mj + E ⋅ xj) , i∈N , in compact form M + E ⋅ X = π ⋅ (m + E ⋅ x) . (15)

In this version of the model the variable Mi expresses the minimum number of movements that local

units of the category associated to node i∈N generate in any case regardless of their dimension,

whereas Xi expresses the additional number of movements generated per each dimension unit, while

the variables mi and xi referred to the arc entering node i allow to express the correlation among M

and X, respectively, in terms of independent variables.

The calibration problem becomes:

min{||z - [Π, diag(E) ⋅ Π] ⋅ [mT, xT]T||2: m ≥ 0, x ≥ 0}. (16)

where Eu is the dimension attribute of the generic local unit u∈U, and E = (E1 , … , Eq , … , E|Q|)T

are the dimension attributes of the sample Q.

Once we have calibrated our disaggregated model which allows to associate a number of

movements to each local unit given its economic activity and dimension, the application to a study

area typically requires to aggregate all the movements generated by each local unit present in a

single traffic zone z∈Z, where Z denotes the set of zones in which the study area is partitioned. Let

z(u)∈Z be the zone where the generic unit u∈U is located. Since each local unit generates a yearly

number of movements Mn(u) + Eu ⋅ Xn(u) , the total number of movements Dz generated in the generic

traffic zone z∈Z is:

Dz = ∑ u∈U: z(u) = z Mn(u) + Eu ⋅ Xn(u) . (17)

16
 For the sake of simplicity we referred our analysis to generic movements, although in practice

we shall build a specific model for each supply chain, given that every category may produce

movements of several supply chains. To this end we simply have to add the required index to all

movement variables in the formulas reported above.

The proposed model has been validated on a dataset of 315 surveys carried out in the framework

of the CityPort European Project to assess the movements produced by retail shops (i.e., the sub-

tree rooted at code 52 of the NACE tree) in the city centre of Bologna, Italy. The validation has

been performed by randomly choosing a subset of the surveys, including 5%, 10%, 20%, or 50%, of

the whole sample. Our model without size function factors (e.g., employees), is calibrated on the

complementary fraction of the sample (i.e., 95%, 90%, 80%, 50%), and the output is compared with

the actual values. This process is repeated 1000 times for each size of the sample subset and only

total movements are considered for the comparison; indeed a model based solely on the activity

code may not provide an accurate estimation for the single shop. This is consistent with the actual

need of proving the quality of the method at a rather aggregated level, since in practice it will be

applied to estimate the movements generated by all shops of a given traffic zone.

The average of the relative gap, defined as |sample - model output| / sample, in terms of total

generated movements (depicted by white bars in Figure 2) decreases from 16.4% for 5% of the

sample (16 shops) down to 7.1% for 50% of the sample (157 shops). As expected, the average

number of generated movements tends to become more stable for larger populations of shops. In

any case, the goodness to fit shown by the model can be considered satisfactory, since a relative gap

of 10% in demand modelling is suitable for most applications. Moreover, the distribution of the

relative gap (without taking the absolute value) among the random samples (depicted by white

diamonds in Figure 4), shows that the for most of the experiments the result lays within the interval

[-20%, +20%], which also assesses good robustness of the model.

To investigate the convenience of the proposed hierarchic approach with respect to a classical

17
category index model, we extended the validation analysis to a simple aggregation of local units

based on the third digit of the activity code and to the trivial model where all shops are aggregated

into one group: this latter yields the mean of the calibration sample, and is taken as the basis for the

comparison. The main results are reported in Figure 3, where the ρ2 (rho square) indicator is defined

as the complement to one of the ratio between the relative gap of a model and the relative gap of the

basis: ρ2 = 1, identifies a perfect model, whereas ρ2 = 0, identifies a useless model which is not

better than the basis.

As far as the relative gap is concerned, all three models that we examined proved to be suitable

for many engineering applications, especially at high aggregation levels. This result tells us that

sophisticated generation models (e.g. consumer acquisition based models), which typically require

lots of expensive data (either in terms of modelling efforts or in terms of survey costs), have a hard

life proving to be really better than simple models, that already have sufficient quality. However, as

expected, the aggregation is better than the mean, and the proposed model is better than the

aggregation. This is true regardless the sample, although the differences among the three model

decrease with its dimension. In conclusion, the proposed model is better than simple ones, but does

not require additional data since there is no explanatory variable other than the activity code.

18
 30.0% mean
simple aggregation
25.0% hierarchic model

20.0%

15.0%

10.0%

5.0%

0.0%
5% 10% 20% 50%

Figure 2. Fitting of the hierarchic model, compared to the simple aggregation and the mean, in

terms of the average relative gap for increasing shares of the sample.

0.40 simple aggregation

hierarchic model
0.35

0.30

0.25

0.20

0.15

0.10

0.05

0.00
5% 10% 20% 50%

Figure 3. Improvement of the fitting obtained by the hierarchic model and by the simple

aggregation with respect to the mean, in terms of the rho square for increasing shares of the sample.

19
 5% of the sample mean 10% of the sample mean
simple aggregation simple aggregation
hierarchic model hierarchic model

-100% -80% -60% -40% -20% 0% 20% 40% 60% 80% 100% -100% -80% -60% -40% -20% 0% 20% 40% 60% 80% 100%

20% of the sample mean 50% of the sample mean

simple aggregation simple aggregation
hierarchic model hierarchic model

-100% -80% -60% -40% -20% 0% 20% 40% 60% 80% 100% -100% -80% -60% -40% -20% 0% 20% 40% 60% 80% 100%

Figure 4. Statistical distributions of the relative gap obtained with the three models, for

increasing shares of the sample.

3. THE DISTRIBUTION MODEL

The distribution model we propose in this paper repeatedly applies a gravity model so as to

better approximate the main characteristic of urban freight shipment (or collection, that is its

reverse). In this case the direct trips that vehicles perform from one point of the study area to

another without intermediate stops do not satisfy a specific mobility demand to travel from one

origin to a destination, as in the transport of passengers, but are merely the legs of a delivery tour

between two consecutive movements. Moreover, the delivery movements occurring at local units,

i.e., the intermediate stops of the tours, are densely located within the urban environment, whereas

the pickup movements occurring at logistic centres are much more dispersed in the territory, being

often out of town or even further away in other towns close to the production centres. Therefore it is

preferable to define a limited number of tour terminals, called logistic portals, located at the town

boundaries along the main routes followed by the trucks entering the city or at logistic centres
20
within the study area. Each portal is associated with a weight to represent its relative importance,

given for example by the average number of commercial vehicles passing that road section or

leaving that logistic centre.

Also in this case we will discuss the model for generic movements, without distinguishing them

by supply chain. However, specific distribution models that allow for the O-D matrix estimation of

a particular supply chain may be easily obtained by using the related generation data and adding the

required index to all formulas reported below.

Given a set C of centroids that act as generators and attractors of freight traffic, the gravity

model allows to determine a demand matrix of, say yearly, truck flows Fod from each origin o∈C to

each destination d∈C, by using a vector of trips Go generated by each origin o∈C, a vector of trips

Ad attracted by each destination d∈C, and a LoS (Level of Service) matrix of generalized travel cost

Tod from each origin o∈C to each destination d∈C, which may in turn be expressed by a linear

combination of suitable attributes weighted by coefficients to be calibrated. By definition, we have:

∑ d∈C Fod = Go , ∀o∈C , (18)

∑ o∈C Fod = Ad , ∀d∈C , (19)

which also imply that the total generation and attraction are at equilibrium, i.e.:

∑ o∈C Go = ∑ d∈C Ad = ∑ o∈C ∑ d∈C Fod . (20)

The functional form of the gravity model is:

Fod = ao ⋅ bd ⋅ Go ⋅ Ad / λ(Tod) , ∀od∈C×C , (21)

where Go and Ad play the role of the masses, while the monotonic function λ(Tod) of generalized

travel costs plays the role of the distance square. In our analysis we have considered the following

distance function:

λ(Tod) = exp(γ⋅ Tod) , (22)

where γ is a parameter to be calibrated, and the generalized travel costs is that of the shortest path

on the road network computed with respect to link costs at the user equilibrium.
21
 In a origin-constrained model the structural coefficients ao have to be consistent with equation

(18), while bd is set to 1, so that we have:

ao = 1 / [∑ d∈C Ad / λ(Tod)] , ∀o∈C . (23)

In a doubly-constrained model the structural coefficients ao and bd have to be consistent with both

equations (18) and (19), so that we have, respectively:

ao = 1 / [∑ d∈C bd ⋅ Ad / λ(Tod)] , ∀o∈C , (24)

bd = 1 / [∑ o∈C ao ⋅ Go / λ(Tod)] , ∀d∈C . (25)

The non-linear system defined by (24)-(25) can be conveniently seen as a fixed point problem that

turns out to be a contraction, i.e., it can be easily solved by initially setting the coefficients to

arbitrary values, e.g. ao = bo = 1 for each o∈C, and then calculating (24)-(25) iteratively until

convergence is achieved. The above solution approach is also adopted in Bablu and Sanat Kumar

(2006), where the doubly-constrained gravity model is formulated as an equivalent minimization

problem.

In our case, the set of centroids is partitioned into two subsets, namely the set Z of traffic zones

and the set P of logistic portals, so that formally we have: C = Z ∪ P, Z ∩ P = ∅. The urban freight

shipping from the portals to the local units is performed by vehicles that leave from a specific

portal, visit some units possibly located in more than one zone, and return to the same portal.

Therefore, the same vehicle may perform during one tour not only a trip from a portal, origin of the

freight, to a single zone, destination of the freight, and backward, but also trips between different

zones and within such zones.

We are interested here in estimating the number of direct trips between centroids, but not in

deriving the actual structure of vehicle routes which, being the result of VRP algorithms, is out of

the scope of the demand oriented methodology proposed in this paper. However, the gravity model

is not conceived to yield O-D flows that satisfy the topological constraints of tours. To force such a

structure we will consider a sequence of gravity models, each one representing implicitly a further

22
stop of the delivery tour.

From the generation model we know the total number of movements Dz occurring in each zone

z∈Z. Moreover, we assume to known the average number μ of movements per route, for example as

a result of a specific survey. The total number of performed tours will then be:

R = ∑ z∈Z Dz / μ . (26)

Each portal p∈P is associated with a weight wp to represent its relative importance. We can

therefore compute the number of tours Rp heading to the portal p as:

Rp = R ⋅ wp / ∑ s∈P ws . (27)

Clearly, the average number μ of movements per route and weight wp of each portal p∈P are

specific characteristics of each supply chain.

In the proposed model M1, the flow of vehicles in terms of direct O-D trips results from the

contribution of the successive legs of delivery tours, which are implicitly represented by each loop

of the following iterative algorithm.

for each od∈C×C

Fod = 0
for each o∈C
if o∈P then Go = Ro else Go = 0
if o∈Z then Ao = Do else Ao = 0
do until Ad = 0 ∀d∈C
for each od∈C×C
Xod = Go ⋅ [Ad / λ(Tod)] / ∑ j∈C [Aj / λ(Toj)] (28)
for each d∈C
Qd = ∑ o∈C Xod
if Qd > Ad then
for each o∈C
Xod = Xod ⋅ Ad / Qd (29)
for each od∈C×C
Fod = Fod + Xod
Ad = Ad - Xod
Go = Go - Xod
Gd = Gd + Xod
for each o∈C
if o∈P then Ao = Ro else Ao = 0
for each od∈C×C
Xod = ao ⋅ bd ⋅ Go ⋅ Ad / λ(Tod) (30)
Fod = Fod + Xod
23
As far as the first leg is concerned, the generation Go is equal to the number of heading tours Ro of

each portal o∈P and is null for each zone o∈Z, while the attraction Ao is equal to the total number

of movements Do of each zone o∈Z and is null for each portal o∈P. Then, the probability that the

next stop of any delivery tour carries out a movement in zone d∈Z is given by the origin-

constrained model expressed by equation (28), where the attraction is the number of delivery

movements to be still covered in d at each step of the algorithm. But if the resulting movements Qd

exceed the deliveries to be performed in zone d∈Z, then only a share of the vehicles are moved

there, as implied by (29). To take into account the last legs of the delivery tours needed for going

back to the portal, i.e., to represent the empty trips, we consider the doubly-constrained model (30),

where the attraction Ao is equal to the number of heading tours Ro of each portal o∈P and is null for

each zone o∈Z, while the generation Go of each centroid o∈C is that resulting at the end of the main

cycle.

It is interesting to observe that, for μ → 1, the number of tours tends to the total number of

movements so that the proposed model collapses into a classical gravity model between origin

portals and destination zones (and back), without direct trips among zones. Instead, for γ → ∞, the

model can be regarded as an approximate solution to the corresponding VRP with the relaxation

that the number of stops per route is satisfied only on average; indeed in this case the gravity model

is a solution to the corresponding transportation problem (Bablu and Sanat Kumar, 2006).

In the following alternative model M2 the latter relaxation is removed, under the simplifying

assumption that each movement occurring at zones has the same probability to be the k-th stop of

the delivery tour that performs it, for k = 1 to μ:

for each od∈C×C

Fod = 0
for each o∈C
if o∈P then Go = Ro else Go = 0
if o∈Z then Ao = Do / μ else Ao = 0
for each od∈C×C
Xod = ao ⋅ bd ⋅ Go ⋅ Ad / λ(Tod) (31)
24
 Fod = Fod + Xod
Fdo = Fdo + Xod
for each o∈C
if o∈Z then Go = Do / μ else Go = 0
for each od∈C×C
Xod = ao ⋅ bd ⋅ Go ⋅ Ad / λ(Tod) (32)
Fod = Fod + Xod ⋅ μ

The probability that the first stop (and symmetrically the last stop) of any delivery tour carries out a

movement in zone d∈Z is given by the doubly-constrained model(31),where the generation Go is

equal to the number of heading tours Ro of each portal o∈P and is null for each zone o∈Z, while the

attraction Ao is equal to the number of first (or last) stops Do / μ of each zone o∈Z and is null for

each portal o∈P. The probability that an intermediate k-th stop of any delivery tour carries out a

movement in zone d∈Z is given by the doubly-constrained model (32), with symmetric generation

and attraction equal to the number of k-th stop Do / μ of each zone o∈Z and null for each portal

o∈P. Since these μ gravity models are identical to each other, we compute one of them and

multiply the resulting O-D flows by μ times.

The model calibration, i.e., the estimation of parameter γ, is performed by maximizing the

likelihood, which is defined to be the joint probability that the model associates to the independent

events revealed in the sample. In this case, the event is a vehicle performing a particular tour, so that

the generic element k∈K of the sample is a sequence of zones {z(1, k), z(2, k), … , z(μ, k)} and a

portal p(k). The probability ηk(β) that the model associates to such a tour is:

ηk(γ) = Fp(k) z(1, k) / Dp(k) ⋅ Fz(1, k) z(2, k) / Dz(1, k) ⋅ … ⋅ Fz(μ-1, k) z(μ, k) / Dz(μ-1, k) ⋅ Fz(μ, k) p(k) / Dz(μ, k) , (33)

where the dependency from γ is implicit in the flows resulting from (21) through (22). The

parameter shall then be determined by solving the following unconstrained nonlinear optimization

problem with one variable:

max{∏ k∈K ρk(γ):γ∈ℜ} , (34)

which we suggest to address simply by determining the objective function values for a suitable

25
interval of γ values through fixed steps.

We illustrate the behaviour of the two models M1 and M2 by means of a simple numerical

example, depicted in Figure 5, that includes two logistic portals (nodes 1 and 5) and three nodes.

The figure reports the results obtained with two different levels of the coefficient γ (0.01 and 1),

which expresses the relevance of travel costs in the trip distribution. By its nature, model M2 tends

to spread more the flows between centroids. In the case γ = 1 travel impedances force the flow

distribution towards a sort of vehicle routing solution with minimal logistic costs.

26
 3

20

80 centroid movements weight

1 100
4
2 20
80 3 100
4 100
2 5 150
70

10

1 110

γ = 0.01 γ=1
1 2 3 4 5 1 2 3 4 5
242 8.8 20 100 100 13.2 242 8.8 20 100 100 13.2
1 8.8 0 1.737 3.531 3.531 0 1 8.8 0 8.8 0 0 0
model M1

2 20 2.955 3.988 5.5386 5.703 1.814 2 20 0 11.2 5.486 3.314 0

3 100 3.046 6.794 46.207 38.867 5.083 3 100 8.8 0 78 0 13.2
4 100 2.798 6.535 39.506 44.856 6.302 4 100 0 0 16.514 83.486 0
5 13.2 0 0.943 5.215 7.040 0 5 13.2 0 0 0 13.2 0.

1 2 3 4 5 1 2 3 4 5
242 8.8 20 100 100 13.2 242 8.8 20 100 100 13.2
1 8.8 0 1.307 4.097 3.395 0 1 8.8 0 2 6.8 0 0
model M2

2 20 1.307 4.806 6.444 6.748 0.692 2 20 2 18 0 0 0

3 100 4.097 6.748 44.823 38.428 5.902 3 100 6.8 0 90 0 3.2
4 100 3.395 6.444 38.731 44.823 6.604 4 100 0 0 0 90 10
5 13.2 0 0.692 5.902 6.604 0 5 13.2 0 0 3.2 10 0

Figure 5. O-D matrices resulting from the application of models M1 and M2 to a simple example

for low and high values of the coefficient γ.

4. THE CASE OF EMILIA-ROMAGNA REGION

The models presented in the previous sections have been developed and preliminarily validated

within a study about city logistics for the Emilia-Romagna Region, located in northern Italy.

Emilia-Romagna has about 4 Million inhabitants and a surface of 22,000 square Km. Given its

geographical position, the region plays a central role in the North-South communication system and
27
is characterized by several mid-size towns with various economic and industrial vocations: from the

mainly tourist town of Rimini to the industrial harbour of Ravenna, from the capital Bologna with

one pf the oldest universities in the world to the towns of river Po valley that combine agriculture

with middle-size mechanical and hi-tech industries.

Since a long time, the regional authority of Emilia-Romagna has been very active in the

investigation and planning of mobility in the region and in the more recent years has devoted an

increasing attention to freight mobility. In particular, within two European Research projects of the

InterReg Programme, denominated CityPorts and Merope, and other initiatives directly sponsored

by the regional authority itself, a large and systematic survey regarding the urban freight

movements has been undertaken in 2004 for nine out of the thirteen towns of the region with more

than 50,000 inhabitants. These surveys were performed according to a homogeneous interview

scheme and included specific questions for demand generators (shops and other economic activities

located in the urban area), demand attractors (distribution centres and shipping companies), and on-

road interviews to freight vehicle drivers.

The initial purpose of the survey was to improve the knowledge on freight distribution activities

in Italian towns by obtaining information disaggregated over the main supply chains so as to

support the impact analysis of infrastructures and policies. The richness of the data that have been

collected was indeed very promising, reporting on the freight movements generated by a given

economic activity with respect to different supply chains and with many information on commercial

vehicles, time windows, parking habits, delivery features, etc. However, the available funds have

considerably limited the size of the samples, hence making impossible its direct statistical use, and

motivating the development of an appropriate modelling tool that would allow for the extension of

the observations to the universe in a structured way. In particular, the overall number of interviews

to demand generators was about 1700 and, for example, in the town of Bologna the sample included

some 300 activities out of more than 30,000. In addition, about 80 freight distribution facilities and

28
companies as well as 3500 trucks were interviewed. More details on the whole data set and its

organization can be found in (Regione Emilia Romagna, 2005).

As previously mentioned, the quantity and quality of logistic information associated with the

observed sample is very rich, whereas very little is available for the whole universe, that is census

data reporting the localization of each economic activity, its category code according to the

ATECO/NACE classification system, and its size measured in terms of employees (but in many

cases not highly reliable). An extensive preliminary testing and validation of the demand generation

and distribution models presented in this paper has been performed during the years 2005-06 on

various towns by using these data for the calibration. In particular the analysis has considered 6

supply chains and has allowed to determine the average number of freight movements generated

yearly in each traffic zone of the town and the specific O-D matrices of freight vehicles serving

each such supply chain, which are then assigned to the road network jointly with private cars. As an

example of the output, Figure 6 depicts the total number of movements and their disaggregation

over the different supply chains for the town of Bologna. The figure clearly shows the different

behaviour of the various zones both in terms of overall number of movements and of their

subdivision over the different supply chains. In addition, Figure 7 shows the localization of the

logistic portals for the distribution model and Figure 8 depicts the flow chart of trucks on road links

determined with the distribution and assignment models.

29
 Figure 6. Plot of the Z-S matrix, for the town of Bologna.

Figure 7. Localization of the logistic portals for the application of the distribution model to the

municipal area of Bologna (in white).

30
 Figure 8. Flow chart of trucks, for the town of Bologna.

The calibration of the generation and distribution models using the rich database of Emilia-

Romagna Region has made available a set of models for different city typologies, from small to

middle size with various urban structures, which are representative of most Italian and European

towns. As a consequence, the proposed demand models, along with a standard traffic assignment

model to load the truck flows on the road network, have been incorporated within an overall GIS-

based software tool, called CityGoods, that is currently used by the regional authority.

5. CONCLUSIONS

In this paper we have introduced new models that are able to describe generation and distribution

of freight movements in urban areas. The main characteristic of the generation model is its direct

use of the hierarchical structure informing the classification system of the economic activities to

introduce a correlation between the freight generation behaviours of the various economic activities,

disaggregated by supply chains. The approach does not require an a-priori mapping of the economic

activities into supply chains and fits very naturally with the scarce data available for the universe of

potential freight generators, producing fine-grained information that proved to be reliable when

aggregated for traffic zones. The distribution model incorporates another important issue of urban

31
freight shipment related to the fact that vehicle tours are headed out of the urban area and make

several intermediate stops in town, whose number depends on the specific supply chain.

The proposed models represent a sound and practical tool that fits within well-established

modelling paradigms and requires easily available data, such as census information on the economic

activities and commercial street networks, that favour its usability by practitioners and decision

makers.

ACKNOWLEDGMENTS

The authors thanks Rino Rosini, Paola Bartolini and Monica Zanarini from the Council on

Transportation of the Emilia-Romagna regional authority for their helpful and active support in

producing this research. The authors also gratefully thank the anonymous Referees and the Editors

for the thorough and helpful comments.

REFERENCES

C. Ambrosini and J.-L. Routhier (2004), “Objectives, methods and results of surveys carried out

in the field of urban freight transport: an international comparison”, Transport Reviews 24, 57-77.

S. Bablu and M. Sanat Kumar (2006), “The constrained gravity model with power function as a

cost function”, Journal of Applied Mathematics and Decision Sciences, Article ID 48632.

The Oak Ridge National Laboratory (2004), “Freight Modeling Bibliography”, for the Freight

Model Improvement Program (see http://www.fmip.gov/models/library.htm), U.S. Department of

Transportation (USDOT).

M. Ben Akiva and R. Lerman (1985), Discrete Choice Analysis: Theory and Application to

Travel Demand, MIT Press, Cambridge, Massachusetts.

J. Boerkamps and A. Van Binsbergen (1999), “GoodTrip - A new approach for modelling and

evaluation of urban goods distribution”, in City logistics, E. Taniguci and R.G. Thompson (eds),

Institute of System Science Research, Japan.

32
 K.J. Button and A.D. Pearman (1981), The economics of urban freight transport, Macmillan,

London, and Holmes and Mier, New York.

J.-F. Cordeau, G. Laporte, M. Savelsbergh and D. Vigo (2007), “Vehicle Routing”, in

Transportation, C. Barnhart and G. Laporte (eds), Elsevier.

M. Friedrich, T. Haupt and K. Nökel (2003), “Freight modelling: data issues, survey methods,

demand and network models”, presented at the 10th International Conference on Travel Behaviour

Research, Moving Through Nets: The Physical and Social Dimensions of Travel, Lucerne.

G. Gentile and D. Vigo (2006), “A Demand Model for Freight Movements in City Logistics

Applications”, presented at Odysseus 2006, Altea, Spain.

J. Gorys and I. Hausmanis (1999), “A strategic overview of goods movement in the Great

Toronto Area”, Transportation Quarterly 53.

P.A. Habib (1985), “Urban freight practice: an evaluation of selected examples”, Transportation

Research Record 1038.

R.I. Harris and A. Liu (1998), “Input-output modelling of the urban and regional economy: the

importance of external trade”, Regional Studies 32.

H. Hautzinger (1984), “The prediction of interregional goods vehicle flows: some modeling

concepts”, Ninth International Symposium on Transportation and Traffic Flow Theory, VNU

Science Press, 375-396.

S. He and T.G. Crainic (1998), “Freight transportation in congested urban areas: issues and

methodologies”, presented at the 8th World Conference on Transport Research, Antwerp, Belgium.

J. Holguin-Veras and E. Thorson (2000), “An investigation on the relationships between the trip

length distributions in commodity-based and trip-based freight demand modeling”, presented at the

79th Transportation Research Record, Washington.

J. Holguín-Veras and E. Thorson (2003a), “Modeling commercial vehicle empty trips with a first

order trip chain model”, Transportation Research B 37, 129-148.

33
 J. Holguín-Veras and E. Thorson (2003b), “Practical implications of modeling commercial

vehicle empty trips”, Transportation Research Record 1833, 87-94.

J. Holguín-Veras, J.C. Zorrilla and E. Thorson (2005), “Modeling commercial vehicle empty

trips: theory and application”, XVI International Symposium of Traffic and Transportation Theory

(ISTTT). H. Mahmassani (ed), Elsevier, 585-608.

B.G. Hutchinson (1974), “Estimating urban goods movement demands”, Transportation

Research Record 496.

C.H. Lawson and R.J. Hanson (1995), Solving least squares problems, SIAM.

G.F. List and M.A. Turnquist (1994), “Estimating truck travel patterns in urban areas”,

Transportation Research Record 1430.

V. Marzano and A. Papola (2004), “Modelling freight demand at a national level: theoretical

developments and application to Italy”, in Proceedings of 2004 ETC Conference, Strasbourg.

B. Meimbresse and H. Sonntag (2000), “Modelling urban commercial traffic with the model

Wiver”, presented at the Jacques Cartier Conference, Montreal.

A.Nagurney (1999), Network economics: a variational inequality approach, second revised

edition, Kluwer Academic Publishers, Dordrecht, Netherlands.

H.J. Noortman (1984), “Goods distribution in urban area”, in Goods Distribution Systems in

Urban Area: Report of the 61st Round Table on Transport Economics, European Conference of

Ministers of Transport, Paris.

K.W. Ogden (1977), “Modelling urban freight generation”, Traffic Engineering and Control 18.

K.W. Ogden (1992), Urban goods movement. A guide to policy and planning, Ashgate,

Aldershot, England.

N. Oppenheim (1995), “On the integrability problem in discrete spatial activity systems with site

and network externalities”, Regional Science and Urban Economics 25, 85-108.

Regione Emilia-Romagna (2005), “City Ports Project interim report”, in Quaderni del Servizio

34
Pianificazione dei Trasporti e Logistica 7, Bologna (available at http://www.regione.emilia-

romagna.it/wcm/ERMES/Canali/trasporti/logistica_merci/pubbl_vol7/City_Ports_vol7_ingl.pdf.).

J-L. Routhier and P-L. Aubert (1998), “FRETURB, un modèle de simulation des transports de

marchandises en ville”, presented at the 8th World Conference on Transport Research, Antwerp,

Belgium.

F. Russo and A. Comi (2004), “A state of the art on urban freight distribution at European scale”,

presented at ECOMM 2004, Lion, France.

H.L. Slavin (1976), “Demand for urban good vehicle trip”, Transportation Research Record 591.

A. Strauss-Wieder, K. Kang and M. Yokei (1989), “The truck commodity survey in the New

York-New Jersey metropolitan area”, in Good transportation in urban areas, Proceedings of the 5th

Conference, Santa Barbara, CA, 83-94.

E. Taniguchi and R.G. Thompson (1999), City Logistics, Institute of Systems Science Research,

Japan.

E. Taniguchi and R.G. Thompson (2003), “Logistics systems for sustainable cities”, in

Proceedings of 3trd Conference on City Logistics, Elsevier.

E. Taniguchi and R.G. Thompson (2006), “Recent advances in city logistics, in Proceedings of

the 4th International Conference on City Logistics, Elsevier.

S.Y. Taylor, (1997) “A basis for understanding urban freight and commercial vehicle travel”,

Transport Research Report, ARR300.

P. Toth and D. Vigo (2002), The vehicle routing problem, SIAM Series on Discrete Mathematics

and Applications, SIAM, Philadelpia, PA.

D.A. Zavettero and S.E. Waseman (1981), “Commercial vehicle trip generation in the Chicago

Region”, Transportation Research Record 834.

35