Nanogrids: A Smart Way to Integrate Public Transportation Electric

Vehicles into Smart Grids

Emanuele Ferrandino

a

, Antonino Capillo

b

,

Fabio Massimo Frattale Mascioli

c

and Antonello Rizzi

d

Department of Information Engineering, Electronics and Telecommunications, University of Rome “La Sapienza”,

Via Eudossiana 18, 00184 Rome, Italy

Keywords:

Smart Grid, Electric Vehicles, Bidirectional Fast Charge, Renewable Energy Source, Vehicle-to-Grid,

Grid-to-Vehicle, Microgrid, Nanogrid, Energy Management System, Fuzzy Logic, Evolutionary Computing,

Hierarchical Genetic Algorithm.

Abstract:

The need for efﬁcient integration of an Electric Vehicles (EVs) public transportation system into Smart Grids

(SGs), has sparked the idea to equip them with Renewable Energy Systems (RESs), in order to reduce their

impact on the SG. As a consequence, an EV can be seen as a Nanogrid (NG) whose energy ﬂows are optimized

by an Energy Management System (EMS). In this work, an EMS for an electric boat is synthesized by a Fuzzy

Inference System-Hierarchical Genetic Algorithm (FIS-HGA). The electric boat follows cyclic routes day by

day. Thus, single day training and test sets with a very short time step are chosen, with the aim of reducing

the computational cost, without affecting accuracy. A convex optimization algorithm is applied for benchmark

tests. Results show that the EMS succesfully performs the EV energy ﬂows optimization. It is remarkable that

the EMS achieves good performances when tested on different days than the one it has been trained on, further

reducing the computational cost.

1 INTRODUCTION

A SG consists of an energy distribution network

that allows to optimize energy exchanges between

the internal nodes and the main energy distribution

network, also considering EVs charging infrastruc-

tures and related Charging Stations (CS). The internal

nodes of a typical SG include residential buildings,

factories and industries, energy production plants

(from renewable sources or from fossil fuels) and

storage systems (Commission, 2006).

In an advanced SG, the internal nodes would also

include a public transportation system made up of

one or more ﬂeets of EVs. Recharging energy infras-

tructures must be orchestrated in real time in order

to ensure continuity of service, thus featuring Grid-

to-Vehicle (G2V) services. On the other hand, the

Vehicle-to-Grid (V2G) paradigm allows to use the

storage system of the EV to power external devices or

to feed energy back into the SG (Deng et al., 2015).

a

https://orcid.org/0000-0001-6472-6597

b

https://orcid.org/0000-0002-6360-7737

c

https://orcid.org/0000-0002-3748-5019

d

https://orcid.org/0000-0001-8244-0015

With the hypothesis that each EV in the ﬂeet is

capable of producing renewable energy on board in-

dependently, as well as consuming it, they potentially

become an active element of the SG with a func-

tion similar to that of a Microgrid (MG). In other

words, the EV described above can implement the

V2G paradigm with a beneﬁcial energy impact on the

SG.

The implementation of the V2G and G2V

paradigms, i.e. of bidirectional exchanges of energy,

together with the independent RES on board the EVs

of a ﬂeet, represents an update of the SG, thanks to the

introduction of new prosumers (or active consumers)

(Deng et al., 2015; Commission, 2007), which can be

summarized as: an increase of the storage energy ca-

pacity (the energy storage system of each EV can be

considered an extension of the main grid energy stor-

age capacity); an increase of the renewable energy

generation capacity (Commission, 2006); an expan-

sion of the SG energy distribution network through

mobile agents; a potential increase in SG ﬂexibility

and resilience at the cost of an increase in the com-

plexity of the SG EMS, or ’tertiary control’ (Olivares

and al, 2014).

512

Ferrandino, E., Capillo, A., Mascioli, F. and Rizzi, A.

Nanogrids: A Smart Way to Integrate Public Transportation Electric Vehicles into Smart Grids.

DOI: 10.5220/0010110005120520

In Proceedings of the 12th International Joint Conference on Computational Intelligence (IJCCI 2020), pages 512-520

ISBN: 978-989-758-475-6

Copyright

c

2020 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

Up to now these aspects have been addressed in

the literature thanks to the integrations of MGs into

the SG (Deng et al., 2015; Areﬁfar et al., 2012), im-

plemented in single houses (Adika and Wang, 2014)

or in small residential agglomerations (Kumar et al.,

2017; Hijjo et al., 2016). The purpose of this pa-

per is to propose a solution to extend this approach

also to EVs. (Mahmud and al, 2020). In the follow-

ing, we focus on a speciﬁc class of EVs that is in-

troduced alongside the term Nanogrid-vehicle (NG-

vehicle). This paper deals with energy management

optimization of a single NG-vehicle and validate the

hypothesis that it is possible to obtain an adequate op-

timization of energy management for the case study

that will be presented and for similar cases even with

short but sufﬁciently dense data sets.

Section 2 illustrates in detail what we mean by

NG-vehicle, the network architecture, its EMS and

the objective function chosen for the optimization of

the energy ﬂows involving the NG-vehicle. Section

3 illustrates the case study chosen to conduct the ex-

periments, the creation of the data set and the formal-

ization of a mathematical model of the NG-vehicle

useful for conducting the experiments in a simulation

environment (MATLAB). Section 4 illustrates the op-

timization method chosen to optimize the NG-vehicle

EMS, according to the FIS-HGA paradigm. Section 5

reports an alternative optimization method useful for

comparison with that is described in section 4 and

forming part of the synthesis procedure of an ade-

quate EMS. In section 6 the synthesis procedure is

illustrated and in sections 7 and 8 are reported the re-

sults of the experimentation and the conclusions, re-

spectively.

2 NANOGRID-VEHICLE: AN

ELECTRIC VEHICLE AS A

MICROGRID

We want to highlight the similarities and dissimilari-

ties between a MG, the NG-vehicle and a traditional

EV. Unlike a traditional EV, the NG-vehicle can pro-

duce energy independently and exchange energy from

and to the outside. This factor makes the NG-vehicle

a grid, which is composed of the four main nodes

characteristic of each MG: generation, load, storage

and a link with an external grid.

Unlike a MG, the NG-vehicle does not have a per-

manent link to the external grid. The Stand-Alone

conﬁguration of an MG ﬁnds correspondence in the

NG-vehicle at all times when it is moving. Con-

versely, while it is stationary and connected to a CS,

it is in the corresponding On-Line conﬁguration of an

MG. This peculiarity translates, as we will see below,

in the fact that the equation of the inner energy bal-

ance depends on a binary variable, which represents

the working mode (’0’: Stand-Alone, ’1’: On-Line).

The proposed EV is classiﬁed as NG, due to its

smaller physical and energy dimensions than a MG.

The NG-vehicle requires a energy ﬂow management,

or ’secondary control’ (Olivares and al, 2014; Kumar

et al., 2017; Sabzehgar, 2015). The EMS is responsi-

ble for deciding how much energy to exchange with

the external grid, represented by a Bidirectional Fast

Charge Station (BFCS), as well as the energy ﬂow di-

rection in order to meet the following requirements:

• Be consistent with the problem, i.e. there must be

no exchange of energy between the NG-vehicle

and the SG when the former is not On-Line;

• Ensure the completion of each route, i.e. do not

consume all the energy of the storage system be-

fore completing the route;

• Keep the storage system close to the Safety Op-

eration Area (SOA) and make sure that each day

starts and ends with a good energy level;

• When the NG-vehicle is On-Line, transfer any

surplus energy produced by the on board RES to

the SG without compromising the storage system.

We assume that any energy request from the SG is

immediately available and that the amount of energy

to be delivered can always be accepted by the SG, i.e.

the storage energy capacity of the SG is assumed inﬁ-

nite. It is also assumed that both the energy generated

by on board renewable source system and the energy

required by the propulsion system of NG-vehicle are

hard to be predicted. For this reason we have cho-

sen to implement the EMS of the NG-vehicle as a FIS

(Santis et al., 2013; Leonori et al., 2017; Gaoua et al.,

2013; Leonori et al., 2016a; Ansari et al., 2014).

2.1 NG-vehicle Architecture

Figure 1 illustrates the grid architecture of the NG-

vehicle including the EMS and the BFCS.

The square nodes (N and S) can exchange bidirec-

tional energy ﬂows, while the circular nodes (G and

L) can exchange unidirectional energy ﬂows (in par-

ticular G can only produce energy and L can just con-

sume it). Each node is associated with a variable of

the type E

X

k

which represents the fraction of energy

exchanged by node X in the time-slot k that goes from

the discrete instant k to the discrete instant k + 1. It

is assumed that the fraction of energy E

X

k

evaluated

at the discrete instant k is constant during the entire

time-slot. It is also assumed that if at the time-slot

Nanogrids: A Smart Way to Integrate Public Transportation Electric Vehicles into Smart Grids

513

Figure 1: NG-vehicle Architecture and EMS with BFCS

link. The EMS, the input variables and the output variable

are colored red.

k the node X is producing energy then E

X

k

> 0, oth-

erwise E

X

k

< 0. It follows that E

L

k

≤ 0, E

G

k

≥ 0 and

that E

N

k

and E

S

k

can be positive, negative or null at any

time-slot k.

Aggregating the unidirectional nodes G ad L as a

unique bidirectional node allows treating it as a pro-

sumer. So, we will represent with E

GL

k

the fraction

of energy exchanged by GL node. It is possible to

readapt the grid architecture of a generic MG to this

application (Kumar et al., 2017; Hijjo et al., 2016;

Moore and Lopes, 2014; Cheddadi et al., 2015).

In Figure 1 E

GL

k−1

is the fraction of energy mea-

sured on the GL node, SOE

k−1

is the State of Energy

measured on the S node and E

N

k

is the fraction of en-

ergy that controls the N node.

Figure 1 also shows a boolean variable that selects

the NG-vehicle mode, m

k

, which acts as a gauge of

the presence of the link with N. The energy fraction

exchanged by S is determined by the energy balance

equation (1) and is limited by the actual SOE status

and maximum charge/discharge current:

E

S

k

+ E

N

k

× m

k

+ E

GL

= 0 (1)

E

S

k

≤ E

S,MaxDischarge

k

(2)

E

S

k

≥ −E

S,MaxCharge

k

(3)

E

S,MaxDischarge

k

= SOE

k

× E

S

max

(4)

E

S,MaxCharge

k

= (1 − SOE

k

) × E

S

max

(5)

E

S

max

= V

S

nom

× I

S

max

× dt (6)

where V

S

nom

is the nominal voltage of the storage sys-

tem, I

S

max

is the maximum charge/discharge current

(determined by the nominal capacity and the maxi-

mum C-Rate) and dt is the length of a time-slot ex-

pressed in hours (all energy quantities are expressed

in kWh).

The SOE is updated through the following formu-

las:

C

S

k

= C

S

k−1

− E

S

k

(7)

SOE

k

=

C

S

k

C

S

nom.

(8)

where C

S

nom.

is the nominal capacity of the storage sys-

tem.

2.2 Objective Function

The objective function is calculated as summation

of the NG-vehicle performance for each time-slot of

the simulation, for a given EMS. In order to extend

the storage system life in terms of charge/discharge

cycles, the performance is a measure of the storage

stress. The stress of the storage system is due to deep

charges and discharges, i.e. charge over 0.8 p.u. and

discharge over 0.2 p.u. These values of the SOE de-

ﬁne the SOA. Thus, optimization translates into min-

imizing the objective function. The stress of the stor-

age system is a personalization of the ’penality func-

tion’ (Leonori et al., 2017; Storti et al., 2015; ?), il-

lustrated in Figure 2 and whose expression is set out

below:

P

S

k

= f

P

(SOE

k

) = ((SOE

k

−SOE

opt

)/SOE

opt

)

12

(9)

where SOE

opt

is 0.5 p.u. is the value in which the

SOA is centered and in which the penalty function

assumes a null value.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

SOE [p.u.]

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Penality [p.u.]

Penality function

Figure 2: Penality function.

The objective function is calculated as follows:

OF =

1

n

n

∑

k=1

P

S

k

(10)

where n represents the length of the experiment in

terms of time-slots.

CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage

514

3 CASE STUDY: THE ELECTRIC

BOAT VALENTINO III

In the context of the European project Life for Silver

Coast - LFSC (LIFE16 ENV/IT/000337), an exam-

ple of an intermodal sustainable mobility system, our

department, in collaboration with the Laboratories of

Cisterna di Latina Polo per la Mobilit

`

a Sostenibile

(POMOS), are focused on the design and engineering

of a small ﬂeet of electric autonomously driven boats

dedicated to public transportation and environmental

monitoring (Lisena et al., 2016) in inland waters (the

Western and Levant Lagoons of Orbetello).

The boats are called Valentino III (see Figure 3),

as they represent the third generation of a class of

electric trimaran ferry boats. Each boat integrates:

an energy generation system based on a photovoltaic

plant, which forms the roof, also providing protection

to passengers from the sun’s rays; an energy storage

system based on a battery pack consisting of lithium-

ion phosphate cells; a propulsion system, which con-

sists of four paddle wheels independently driven by as

many motors located on both the sides of the boat; a

BFCS compatible interface.

Figure 3: Rendering of the Valentino III.

Every day the boat must follow four scheduled

routes: at 8:00, 10:00, 12:00 and 14:00. The dura-

tion of each route is one hour and consists of an out-

ward and a return. For both the lagoons, the piers of

Orbetello represent the starting and arrival points of

each route, as they are equipped with BFCSs. Both

routes cover approximately 6km. This implies that

the average speed that the boat must hold is 6km/h or

approximately 4kn.

During navigation, the propulsion system must be

supplied with adequate power to maintain the speed of

navigation and to face winds below ten meters of alti-

tude and surface water currents produced by the wind

itself (drift water currents). In addition, the propul-

sion effort is inﬂuenced by the weight on board, which

represents mostly passengers and crew members.

3.1 Data Set

A data set was created containing the mooring pro-

ﬁles, photovoltaic generation and consumption of the

propulsion system. Project speciﬁcations were taken

into consideration for the ﬁrst proﬁle: duration and

daily frequency of the routes. For the generation and

consumption proﬁles, it was necessary to obtain in

advance the speed proﬁles of the boat, the use of

the public transportation system by people, the pho-

tovoltaic generation, winds and surface water cur-

rents. With regard to the use of public transporta-

tion, it was considered that the weight of the boat has

a Gaussian distribution between the dry weight and

the fully loaded weight and centered in the middle of

the year (high season). An online simulation platform

was used for photovoltaic generation and wind data

(https://www.renewables.ninja/). The surface water

currents proﬁle was created based on the wind pro-

ﬁle.

The dynamic equation of motion of the boat was

used to calculate the energy consumption data, which

uses the speed proﬁle of the boat, the weight proﬁle,

the wind proﬁle and the water current proﬁle .

It should be noted that in practice both the energy

consumed and the energy produced are unpredictable

as it is not known a prior the intensity of the wind and

water currents, the number of people on board and the

photovoltaic generation.

Finally the data set consists of two sequences

{(E

GL

k

,m

k

)}. Given the cyclical nature of the routes, a

time-slot dt equal to 1 minute was chosen, in order to

investigate deeply every operational condition of the

boat. This results is a very dense data set.

3.2 Adopted Vehicle Model

This section illustrates the formulation of a NG-

vehicle model (which also includes a simplifying

model of the BFCS). The model is used to simulate

the energy ﬂows determined by its EMS and simu-

lation data. The NG-vehicle model also includes the

calculation of the objective function which summa-

rizes the performance of the EMS.

The EMS output is the quantity q

N

k

deﬁned in the

real interval [0,1] which must be translated into the

energy fraction E

N

k

, as follows:

E

N

k

= f

N

(q

N

k

) = E

N

max

×(2×q

N

k

−1),q

N

k

∈ [0,1] (11)

E

N

max

= P

N

nom.

× dt (12)

where P

N

nom.

is the nominal power of the BFCS.

Nanogrids: A Smart Way to Integrate Public Transportation Electric Vehicles into Smart Grids

515

In the same way, the input variable E

GL

k

is nor-

malized in a quantity q

GL

k

∈ [0,1] deﬁning the inverse

model of GL:

q

GL

k

= f

−1

GL

(E

GL

k

) =

(

0.49

E

L

max

× (E

L

max

− E

GL

k

),E

GL

k

< 0

(E

GL

k

+ E

G

max

) ×

0.5

E

G

max

,E

GL

k

≥ 0

(13)

E

L

max

= 4 × P

M

nom.

× dt (14)

E

G

max

= P

G

nom.

× dt (15)

where P

M

nom.

and P

G

nom.

are the rated powers of each

motor and the photovoltaic system, respectively.

The input variable SOE

k

on the other hand is al-

ready normalized as it is calculated in parts per unit.

4 FIS-HGA DESIGN

EMS synthesis and optimization was carried out by

the FIS-HGA paradigm (Santis et al., 2013; Leonori

et al., 2017; Siddique and Adeli, 2013; Shi et al.,

1999; Capillo et al., 2018; Leonori et al., 2016b), ac-

cording to which a Mamdani FIS is optimized by a

Hierarchical Genetic Algorithm (HGA) (Santis et al.,

2017; Delgado et al., 2001; Tang et al., 1998).

4.1 FIS Structure

The FIS consists of two inputs (SOE and E

GL

) and

one output (E

N

). The Term Set of each input and of

the output counts ﬁve Membership Functions (MFs),

which partially overlap (Fig. 4).

0.10

0

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.2

0.4

0.6

0.8

1

“Very Low” “Low” “Medium” “High” “Very High”

Degree of membership

Input

Figure 4: The Term Set for both the two inputs and the out-

put.

The rule base is set to contain all the possible rules

(25), which come from the combination between the

input MFs, as follows:

n

rules

= n

n

inputs

MFs

(16)

where n

rules

, n

MFs

and n

inputs

are the numbers of rules,

MFs and inputs, respectively. The only logical opera-

tor between antecedents is AND.

4.2 FIS Genetic Encoding

The FIS is encoded into the genes of the HGA generic

individual, as follows:

• The ﬁrst 10 genes are binary and represent the

presence/absence of a MF in the two input Term

Sets (not by chance, the size of this set of genes is

the product of n

MFs

by n

inputs

);

• The next 39 genes are real and encode the vertices

abscissas of all the input and output MFs. all the

abscissas can be tuned, except the ones outside the

Universe of Discourse (Fig. 4). The number of

these genes is the product of the tunable abscissas

of each input and output Term Sets (13) by the

total number of inputs and outputs of the FIS (3);

• The next 25 real genes represent the rule weights;

• The last 25 genes are integers selecting MFs from

the output Term Sets (each rule must have only

one MF in the consequent part).

4.3 The Optimization Process

Two subsequent optimization processes are per-

formed. During the ﬁrst optimization, the HGA tunes

the FIS rule base. Only the ﬁrst 10 binary genes of the

generic individual evolve through generations, while

the OF is minimized. If the i-th gene is set to 0, this

means that the i-th MF in the set of the whole inputs

MFs (from the ﬁrst to the last input) is deleted. As a

consequence, each rule enclosing the i-th MF as an-

tecedent is deleted from the rule set. Thus, the ﬁrst

optimization aims at reducing the number of rules

by selecting the most relevant ones for the problem

at hand. It is worth mentioning that, when a MF is

deleted, the vertices abscissas of the remaining ad-

jacent MFs are modiﬁed in order to cover the entire

Universe of Discourse (Fig. 5). During the second

optimization, the rest of genes of the generic indi-

vidual evolve over generations, meaning that the FIS

parameters (MFs vertices abscissas and rule weights)

and the consequent are tuned, while the OF is mini-

mized. Only the genes related to the rule base coming

from the ﬁrst optimization are tuned. In this sense, the

ﬁrst binary genes are at an higher hierarchical level

than the others, making the GA hierarchical.

4.4 HGA Operators and Settings

For the ﬁrst optimization, a one-point crossover is

adopted, together with a bit string mutation. For the

second optimization, the choice of the crossover op-

erator depends on the type of genes: for the genes

of the vertices abscissas, a convex crossover; for the

CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage

516

0.10

0

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

0.2

0.4

0.6

0.8

1

“Very Low” “Low” “Medium” “High” “Very High”

Degree of membership

Input

Figure 5: Universe of Discourse coverage after deleting a

MF (in this case, the central one).

weights and consequent genes, a uniform crossover.

A uniform mutation is performed during the second

optimization. A tournament selection with tourna-

ment size 2 is chosen for both the ﬁrst and the second

optimization.

The population consists of 100 individuals for

both the optimization processes, as well as for the

crossover fraction, set to 0.8. The stopping condition

is always the reaching of the maximum stall gener-

ations while the maximum number of generations is

set to 50 for the ﬁrst optimization and to 300 for the

second optimization.

5 CONVEX OPTIMIZATION

PROBLEM

In order to know how good is the solution found by

optimization through FIS-HGA, it is necessary to ﬁnd

the optimal solution. This can be exactly found by

solving a convex optimization problem.

Note that the penalty function expressed in Equa-

tion (9) is a convex function of the SOE, which is a

linear combination of the energy fractions exchanged

over time between node N and the NG-vehicle. So

the penalty function is a convex function of the EMS

output E

N

k

, that becomes the variable of the problem.

The problem is a constrained problem.

In particular: SOE

k

must belong to the real

range [0,1]; E

S

k

must belong to the real range

[−E

SMaxCharge

k

,E

SMaxDischarge

k

], in which the extremes

depend on the SOE

k−1

; the fraction of energy E

N

k

must belong to the real interval [−E

N

max

,E

N

max

].

While it is not even easy to understand whether

the problem is convex or not, it is possible in many

cases to rewrite a non-convex problem in such a form

that it is convex. Some tools can solve optimization

problems, even non-convex ones, for example solving

the dual problem if it turns out to be convex (Boyd and

Vandenberghe, 2004).

The problem can be written as follows:

min

E

N

k

OF

s.t 0 ≤ SOE

k

≤ 1

E

S

k

≤ E

SMaxDischarge

k

E

S

k

≥ −E

SMaxCharge

k

kE

N

k

k ≥ kE

N

max

k

E

S

k

= −(E

N

k

× m

k

+ E

GL

k

)

(17)

where OF is described in Equation (10).

6 SYNTHESIS PROCEDURE

The FIS synthesis is the result of the following proce-

dure:

1. Train the FIS by the HGA and solve the convex

optimization problem (17) on the Training set;

evaluation of the training errors as the Root Mean

Square Error Percent - RMSEP on SOE and on

E

N

. If the training errors are too high (greater than

10%) different solutions must be considered:

• Increase the number of iterations of the HGA;

• Change the HGA optimization process.

2. Test the HGA optimized FIS and solve the convex

optimization problem (17) on sets distinct from

the training set; evaluation of the test errors; if

the test errors are too high (greater than 10%) it

implies data overﬁtting. It is necessary to repeat

steps 1. and 2. until the conditions on RMSEP are

met.

The training set is sufﬁciently rich even lasting

only 1 day as a time-slot of 1 minute was chosen, un-

like the typical time-slot of 15 minutes, as in (Deng

et al., 2015; Olivares and al, 2014; Leonori et al.,

2016b; Santis et al., 2017), for data on MGs and SGs.

In fact, 1 day contains 1440 samples which is a good

amount of data. We are particularly interested in the

high season. In particular, for the training set we con-

sidered the day in which the highest peak of energy

generation by the photovoltaic plant was recorded (in-

dicative of the high season).

7 RESULTS

The signiﬁcant reduction in the Fuzzy Rule Base, due

to the ﬁrst optimization process (from 25 to 6 Fuzzy

Rules), has lead to a lower computational cost for the

FIS-EMS synthesis.

Nanogrids: A Smart Way to Integrate Public Transportation Electric Vehicles into Smart Grids

517

The best energy ﬂows returned by the FIS-HGA

EMS for the Training Set and the same ﬁgures re-

turned by the optimal EMS (optimized by convex op-

timization), are shown in Figure 6 and Figure 7, re-

spectively. The OF value of the two solutions and the

training errors are listed in Table 1.

Table 1: Optimal and SubOptimal OF values, E

N

RMSEP

(E

N

err.) and SOE RMSEP (SOE err.) on Training day.

Opt.OF Sub Opt.OF E

N

err. SOE err.

p.u. p.u. % %

3.7e-08 5.8e-08 4.56 4.79

It can be seen that for both solutions the require-

ments of the EMS are met. In particular, in both the

solutions, the energy ﬂows between the BFCS and the

NG-vehicle show a charging phase of the storage sys-

tem alternating with each scheduled route to an ex-

tent that guarantees the completion of the next route

and the containment of the SOE in the SOA limiting

the stress on the storage system. Furthermore, when

recharging does not take place, in the remaining time-

windows between two routes, the surplus energy pro-

duced by the RES is transferred to the SG, as it is not

needed to bring the storage system back to a better

SOE value.

The EMS optimized by the FIS-HGA performs

recharges of the storage system softer and longer than

the optimal one. This reduces the time-windows in

which the EMS can transfer surplus energy to the SG.

In Figure 8 are shown the E

N

RMSEP on the

Training day (in the top plot title), the E

N

proﬁles of

the two solutions (in the top plot) and the difference

between them (in the bottom plot). Figure 9 shows the

SOE RMSEP (in the top plot title), the SOE proﬁles

of the two solutions (in the top plot) and the difference

between them (in the bottom plot). Despite the differ-

ences in terms of E

N

k

, both the RMSEP are lower than

the established threshold of 10%, in particular below

5%, allowing to consider the whole synthesis proce-

dure to be effective.

200 400 600 800 1000 1200 1400

[min]

-0.5

0

0.5

[kWh]

Training on day 146:

OF = 5.8017e-08

Energy Flows

GL node

N node

S node

200 400 600 800 1000 1200 1400

[min]

0

0.5

1

[p.u.]

State of Energy

200 400 600 800 1000 1200 1400

[min]

0

2

4

6

[p.u.]

10

-6

Storage Stress

Figure 6: EMS Optimized by FIS-HGA (Training Set).

In order to have a more precise idea of the differ-

ences just shown between the EMS optimized by the

FIS-HGA and the optimal EMS, the amount of energy

200 400 600 800 1000 1200 1400

[min]

-0.5

0

0.5

[kWh]

Optimal Solution on day 146:

OF = 3.7262e-08

Energy Flows

GL node

N node

S node

200 400 600 800 1000 1200 1400

[min]

0

0.5

1

[p.u.]

State of Energy

200 400 600 800 1000 1200 1400

[min]

0

1

2

3

[p.u.]

10

-6

Storage Stress

Figure 7: Optimal EMS on Training Set.

200 400 600 800 1000 1200 1400

[min]

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

[kWh]

EN RMSEP on day 146 = 4.559696

EN

optimal

EN

FIS

200 400 600 800 1000 1200 1400

[min]

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

[kWh]

EN error

Figure 8: Training error: energy fractions at node N.

200 400 600 800 1000 1200 1400

[min]

0

0.2

0.4

0.6

0.8

1

[p.u.]

SOE RMSEP on day 146 = 4.793316

SOE

optimal

SOE

FIS

200 400 600 800 1000 1200 1400

[min]

0

0.2

0.4

0.6

0.8

1

[p.u.]

SOE error

Figure 9: Training error: State of Energy.

taken from the SG in one day, Q

N

buy

, and the amount

of energy transferred to the SG in one day, Q

N

sell

, for

the two solutions are shown in Table 2.

Table 2: Optimal and Sub-Optimal energy from and to the

SG.

Q

N

sell

Q

N

buy

Opt.Q

N

sell

Opt.Q

N

buy

kWh kWh kWh kWh

2.586 17.472 7.374 21.561

Note that the use of NG-vehicles for the implemen-

tation of a transportation system has a lower energy

impact on the SG of Q

N

sell

, compared to the use of

equivalent EV without RESs (Eq. EVs).

In Table 3 are summarized the results of the test

phase performed by simulating the EMS obtimized by

FIS-HGA and solving the convex optimization prob-

lem (17) for the eight days. The E

N

and SOE RM-

SEP values are below the 10% threshold for all tests:

the optimization of the EMS through FIS-HGA on

the Training Set has generated a solution that is valid

throughout most of the year. The worst result is that

CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage

518

Table 3: Optimal and Sub Optimal OF values on Test days,

E

N

k

RMSEP (E

N

k

err.) and SOE RMSEP (SOE err.) on Test

days.

Test Opt.OF Sub Opt.OF E

N

err. SOE err.

day p.u. p.u. % %

59 9.7e-10 3.0e-09 2.18 4.63

90 1.0e-09 2.5e-07 2.93 6.85

120 1.8e-09 4.4e-08 3.52 5.48

151 6.0e-07 5.3e-04 6.33 8.78

181 8.4e-07 1.4e-05 5.74 5.37

212 1.3e-07 7.9e-07 5.08 4.71

243 2.2e-09 1.1e-07 3.93 5.10

273 1.1e-09 4.2e-08 2.89 5.63

related to the day 151 (high season). This is explained

by the fact that around this day the maximum weight

at full load is recorded; remember that the weight on

board, representing the people who use the transporta-

tion service offered by the single boat, strongly inﬂu-

ences the consumption of the propulsion system.

8 CONCLUSIONS

The EMS of an NG-Vehicle has been synthesized

through a FIS-HGA algorithm, aiming at manag-

ing the bidirectional exchanges of energy fractions

with an SG through a BFCS according to the time-

windows in which the NG-vehicle is On-Line and the

completion of the scheduled routes. The EMS opti-

mization was performed aiming to minimize the stor-

age stresses possibly due to deep charges/discharges

of the NG-vehicle storage system. The synthesis

of EMS as FIS through the proposed procedure has

proved effective in avoiding data overﬁtting. These

choices led to a synthesis with a satisfactory perfor-

mance value for all the 8 test days, as also reported by

the RMSEP values. The synthesized EMS has satis-

ﬁed all the proposed requirements. In particular, the

NG-vehicle transfers its energy surplus to SG, due to

the continuity of the energy production by the photo-

voltaic roof. Therefore, since the optimization aims at

minimizing the stress on the storage system, the gen-

eration system from renewable sources on board not

only supports the storage system during Stand-Alone

mode (i.e. navigation mode), as commanded by the

NG-vehicle architecture, but it reduces the energy im-

pact that the NG-vehicle has on the SG.

Figure 10 shows the energy impacts on the SG of

the Valentino III with sub-optimal EMS, the Valentino

III with the optimal EMS and the Eq.-EV (Valentino

III without RES), for each of the eight test days. Re-

sults show that the energy impact of the Valentino III

with sub-optimal EMS is closer to the energy impact

1 59 90 120 151 181 212 243 273 365

[day]

0

5

10

15

20

25

30

35

[kWh]

Energy Impact on SG

Equivalent Vehicle

NG-Vehicle

optimal NG-Vehicle

Figure 10: Energy Impact on the SG of the Valentino III

Eq-EV, the Valentino III with sub-optimal EMS and the

Valentino III with optimal EMS.

of the one with optimal EMS than to the energy im-

pact of the Eq.-EV. Future analisys will be performed

in order to increase the difference in terms of energy

impact from the worst case. To achieve this goal, one

way consists in performing an optimization with the

same FIS-HGA algorithm, while minimizing both the

stress on the storage system and the energy impact on

the SG.

We also reserve for the future the wider topic of

optimizing the energy ﬂow management of a ﬂeet of

NG-vehicles.

ACKNOWLEDGMENTS

The POMOS Laboratories and the DIET Department

(University of Rome “La Sapienza”) would like to

thank the EU for ﬁnancial support to environmental

and climate action projects like LIFE for Silver Coast

(LIFE16 ENV/IT/000337). Such a help is crucial to

achieve natural and historical preservation of Italy, es-

pecially of touristic areas.

REFERENCES

Adika, C. and Wang, L. (2014). Autonomous appliance

scheduling for household energy management. IEEE

Transactions on Smart Grid, vol. 5, no. 2, pages 673–

682.

Ansari, M., Al-Awami, A., Abido, M., and Sortomme, E.

(2014). Optimal charging strategies for unidirectional

vehicle-to-grid using fuzzy uncertainties. IEEE PES

TD Conference and Exposition, pages 1–5.

Areﬁfar, S., Mohamed, Y., and El-Fouly, T. (2012). Supply-

adequacy-based optimal construction of microgrids in

smart distribution systems. IEEE Transactions on

Smart Grid, vol. 3, no. 3, pages 1491–1502.

Boyd, S. and Vandenberghe, L. (2004). Convex Optimiza-

tion. Cambridge University Press.

Capillo, A., Luzi, M., Paschero, M., Rizzi, A., and Masci-

oli, F. F. (2018). Energy transduction optimization of

a wave energy converter by evolutionary algorithms.

Nanogrids: A Smart Way to Integrate Public Transportation Electric Vehicles into Smart Grids

519

International Joint Conference on Neural Networks

(IJCNN), pages 1–8.

Cheddadi, Y., Gaga, A., Errahimi, F., and Es-Sbai, N.

(2015). Design of an energy management system for

an autonomous hybrid micro-grid based on labview

ide. 3rd International Renewable and Sustainable En-

ergy Conference (IRSEC), pages 1–6.

Commission, E. (2006). The SmartGrids European Tech-

nology Platform — SmartGrids. Luxembourg.

Commission, E. (2007). Strategic Research Agenda for Eu-

rope’s Electricity Networks of the Future. Brussel.

Delgado, M. R., Zuben, F. V., and Gomide, F. (2001). Hier-

archical genetic fuzzy systems. Information Sciences,

Vol.136, page 29 – 53.

Deng, R., Yang, Z., Chow, M., and Chen, J. (2015). A sur-

vey on demand response in smart grids: Mathematical

models and approaches. IEEE Transactions on Indus-

trial Informatics, vol. 11, no. 3, pages 570–582.

Ehsani, M., Gao, Y., and Emadi, A. (2009). Modern elec-

tric, hybrid electric, and fuel cell vehicles: fundamen-

tals, theory, and design. CRC Press.

Gaoua, Y., Caux, S., Lopez, P., and Salvany, J. (2013). On-

line hev energy management using a fuzzy logic. 12th

International Conference on Environment and Electri-

cal Engineering, EEEIC, pages 46–51.

Hijjo, M., Felgner, F., Meiers, J., and Frey, G. (2016). En-

ergy management for islanded buildings integrating

renewables and diesel generators. IEEEPES Power-

Africa, Livingstone, pages 62–66.

Kumar, A., Deng, Y., He, X., Kumar, P., and Bansal, R.

(2017). Energy management system controller for

a rural microgrid. The Journal of Engineering, vol.

2017, no. 13, pages 834–839.

Leonori, S., Paschero, M., Rizzi, A., and Mascioli, F. F.

(2016a). Optimization of a microgrid energy man-

agement system based on a fuzzy logic controller.

IECON.

Leonori, S., Paschero, M., Rizzi, A., and Mascioli, F. F.

(2017). An optimized microgrid energy management

system based on ﬁs-mo-ga paradigm. Fuzzy Systems

(FUZZ-IEEE), pages 1–6.

Leonori, S., Santis, E. D., Rizzi, A., and Mascioli, F. F.

(2016b). Multi objective optimization of a fuzzy

logic controller for energy management in microgrids.

IEEE Congress on Evolutionary Computation (CEC),

pages 319–326.

Lisena, V., Paschero, M., Gentile, V., Amicucci, P., Rizzi,

A., and Mascioli, F. F. (2016). A new method to re-

store the water quality level through the use of electric

boats. IEEE International Smart Cities Conference

(ISC2), pages 1–4.

Mahmud, K. and al (2020). Real-time load and ancillary

support for a remote island power system using elec-

tric boats. IEEE Transactions on Industrial Informat-

ics, vol. 16, no. 3, pages 1516–1528.

Moore, R. and Lopes, J. (2014). The energy management

and optimized operation of electric vehicles based on

microgrid. IEEE Transactions on Power Delivery, vol.

29, no. 3, pages 1427–1435.

Olivares, D. and al (2014). Trends in microgrid control.

IEEE Transactions on Smart Grid, vol. 5, no. 4, pages

1905–1919.

Sabzehgar, R. (2015). A review of ac/dc microgrid-

developments, technologies, and challenges. IEEE

Green Energy and Systems Conference (IGESC),

pages 11–17.

Santis, E. D., Rizzi, A., and Sadeghian, A. (2017). Hierar-

chical genetic optimization of a fuzzy logic system for

energy ﬂows management in microgrids. Applied Soft

Computing, vol. 60, pages 135 –149.

Santis, E. D., Rizzi, A., Sadeghiany, A., and Mascioli, F.

(2013). Genetic optimization of a fuzzy control sys-

tem for energy ﬂow management in micro-grids. Joint

IFSA World Congress and NAFIPS Annual Meeting

(IFSA/NAFIPS), pages 418–423.

Shi, Y., Eberhart, R., and Chen, Y. (1999). Implementation

of evolutionary fuzzy systems. IEEE Transactions on

Fuzzy Systems, vol. 7, no. 2, pages 109–119.

Siddique, N. and Adeli, H. (2013). Computational Intel-

ligence: Synergies of Fuzzy Logic, Neural Networks

and Evolutionary Computing. Wiley.

Storti, G., Paschero, M., Rizzi, A., and Mascioli, F.

(2015). Comparison between time-constrained and

time-unconstrained optimization for power losses

minimization in smart grids using genetic algorithms.

Neurocomputing, vol. 170, page 353 – 367.

Tang, K., Man, K., Liu, Z., and Kwong, S. (1998). Min-

imal fuzzy memberships and rules using hierarchical

genetic algorithms. IEE Transactions on Industrial

Electronics, vol. 45, no.1, pages 162–169.

CI4EMS 2020 - Special Session on Computational Intelligence for Energy Management and Storage

520