Applied Thermal Engineering 90 (2015) 1172e1183

Contents lists available at ScienceDirect

Applied Thermal Engineering

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

Research paper

A comparison of the economic benets of centralized and distributed

model predictive control strategies for optimal and sub-optimal mine
dewatering system designs
 M. Maestre e,
Alberto Romero a, b, c, *, Dean Millar a, b, Monica Carvalho d, Jose
e
Eduardo F. Camacho
a

Mining Innovation, Rehabilitation & Applied Research Corporation (MIRARCO), Sudbury, Ontario, Canada
Bharti School of Engineering, Laurentian University, Sudbury, Ontario, Canada
BESTECH Ltd, Sudbury, Ontario, Canada
d
Alternative and Renewable Energy Center, Department of Renewable Energy Engineering, Federal University of Paraba, Brazil
e
tica, Universidad de Sevilla, Sevilla, Spain
Dept. Ingeniera de Sistemas y Automa
b
c

h i g h l i g h t s

Optimal Mine Site Energy Supply (OMSES) applied to underground mine dewatering system.

Systems design and pump schedule, based on past operating conditions, were optimized.

Optimal system's design was simulated based on Model Predictive Control (MPC).

Centralized, distributed and decentralized controllers were compared.

Only centralized MPC showed robust and veried OMSES optimal design.

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 16 June 2014
Accepted 12 January 2015
Available online 30 January 2015

Mine dewatering can represent up to 5% of the total energy demand of a mine, and is one of the mine systems
that aim to guarantee safe operating conditions. As mines go deeper, dewatering pumping heads become
bigger, potentially involving several lift stages. Greater depth does not only mean greater dewatering cost,
but more complex systems that require more sophisticated control systems, especially if mine operators
wish to gain benets from demand response incentives that are becoming a routine part of electricity tariffs.
This work explores a two stage economic optimization procedure of an underground mine dewatering
system, comprising two lifting stages, each one including a pump station and a water reservoir. First, the
system design is optimized considering hourly characteristic dewatering demands for twelve days, one
day representing each month of the year to account for seasonal dewatering demand variations. This
design optimization minimizes the annualized cost of the system, and therefore includes the investment
costs in underground reservoirs. Reservoir size, as well as an hourly pumping operation plan are
calculated for specic operating environments, dened by characteristic hourly electricity prices and
water inows (seepage and water use from production activities), at best known through historical
observations for the previous year. There is no guarantee that the system design will remain optimal
when it faces the water inows and market determined electricity prices of the year ahead, or subsequent years ahead, because these remain unknown at design time. Consequently, the dewatering optimized system design is adopted subsequently as part of a Model Predictive Control (MPC) strategy that
adaptively maintains optimality during the operations phase.
Centralized, distributed and non-centralized MPC strategies are explored. Results show that the system can be reliably controlled using any of these control strategies proposed. Under the operating
conditions considered, total annualized cost savings in the range of 50%, in comparison with nonoptimized designs, can be achieved through design optimization and the use of predictive control.
2015 Elsevier Ltd. All rights reserved.

Keywords:
Optimal mine site energy supply
Dewatering system
Model predictive control
Load shifting
Demand charges

* Corresponding author. Mining Innovation, Rehabilitation & Applied Research Corporation (MIRARCO), Sudbury, Ontario, Canada. Tel./fax: 1 705 675 1151x5054.
E-mail address: aromero@mirarco.org (A. Romero).
http://dx.doi.org/10.1016/j.applthermaleng.2015.01.031
1359-4311/ 2015 Elsevier Ltd. All rights reserved.

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

1. Introduction
The increasing price of energy resources as well as increasing
complexity (lower grade deposits, remoteness of mine locations,
higher depth, etc) of mineral extraction processes motivates mining
companies to continuously develop and apply strategies that allow
them to execute their operations in more energy efcient ways.
Lower grades and remote location negatively affects mine site
energy demand. The lower the mineral concentration of a deposit,
the greater is the amount of waste rock that is extracted and
therefore more energy per unit of desired product is required.
Remote mines imply additional constraints on energy availability,
e.g. electric grid capacity limitation or supply restrictions on fuel
available on site for power generation.
Depth of extraction affects mines energy consumption signicantly. Whether it is an open pit or an underground operation, a
mine consumes energy for processes that increase the potential
energy (height) of material ows. Examples include rock hoisted
(both ore and waste rock) and water. Mine water sources are mainly
groundwater seepage, meteorological precipitation and process
water used in mining operations for cutting, drilling, washing, etc.;
this water must be removed from the mine, otherwise it will ood.
The system responsible for proper water evacuation from mines is
known as the dewatering system.
Although critical to the production process, the dewatering
system is not completely constrained to the pace of remaining
mining processes. This is in contrast to, for example, hoisting
services and ventilation of underground mines, or rock loading
and hauling in open pit mines. In other words, there exists some
exibility to decide when and how to operate the dewatering
system, but nevertheless within its own design limits, e.g. water
reservoir dimensions and for the pumping ow and head
available.
This freedom to select the operating mode permits the application of load shifting (LS) and demand side management (DSM)
strategies. The application of these strategies is especially
important in systems or processes whose operating cost depends
upon an energy market with tariffs that feature variable prices in
different time intervals, for example, so-called time of use prices
(ToU), as well as demand charges (DC). The latter depend upon the
maximum power consumed within a predened time interval.
Electricity markets incorporate this divided tariff structure to
encourage consumers to modify their consumption patterns,
incentivized by reduced rates. LS entail changing consumption
patterns to shift power loads to times when the ToU price is lower.
DSM mitigates the negative consequences of high electric demands that the grid and the generation capacity cannot handle
safely e consumers are economically incentivized to reduce power or, alternatively, are penalized for demand at times when the
system is most stressed.
Optimization of pumping systems is not a new eld of research,
either for design or for operation purposes, on exploring and validating new control strategies during the last 20 years considerable
effort has been expended. This paper introduces a methodology to
assess the problem of optimal design and operation of a water
system subject to operational uncertainties, regarding the case of a
mine dewatering system.
Today, water system design does not receive as much attention
from the academic community [1e4] as operation does [5e14].
When considered, design optimization focuses on pump arrangement and capacity [1e3] and, less frequently, on system components such as piping, valve selection or reservoir capacity [4]. One
of the reasons for this is that there are many robust and precise
design tools available, such as [15,16] that render this design task
straightforward.

1173

Regarding design, [1] optimizes a shfarm pumping station

design and operation schedule for one typical year of water demand to get the least total annual cost, including the corresponding
investment and operation costs (electricity). Ref. [2] optimizes
design of pump stations in a water system that has xed reservoir
size: [3] does not optimize reservoirs size but points out the
advantage of bigger reservoir for greater cost savings through load
shifting strategies. In Ref. [4] it is applied a multi-objective optimization to a water distribution system so that selects the optimal
reservoir size to obtain the least operation (pumping schedule) and
investment cost while assuring water quality. On these papers
there is a lack of discussion on whether or not the optimal design
will remain optimal when the reference, operating conditions are
to change.
There are many examples of pump operation scheduling, and
these are generally cast as mathematical programming problems
solved by means of one of many solver algorithms available (Genetic Algorithms, Particle Swarm Optimization, Branch and Bound
Simplex, etc.) Examples of different approaches and applications
can be found in Refs. [5e14]. Most of the approaches for these
scheduling optimizations minimize operating costs based on ToU
tariffs. Some also minimize pump switch number in a given interval
of time (repeated startups can cause in xed speed pumps overheating and excessive stress, that lead to premature failure). Demand Charges, which strictly depend on the power consumed by
the system (traditionally deterministic calculation based upon a
dened tariff), can also be minimized. Fixed and variable speed
pumps have been investigated in such studies and there appears to
be that variable speed pumps improve energy efciency while load
shifting can be either achieved with xed or variable speed.
Nevertheless, the optimality of the schedules calculated lays on
existing, perhaps not, optimal designs and, once again, they do not
consider the effect of uncertain operating conditions.
In recent years water system control has received attention, and
more specically optimal control. Model Predictive Control [17] is
the preferred approach, for its ability to deal with system's constraints (pump capacity, reservoir limits, etc.) and to accommodate
cost functions for either tracking purposes (maintain water level or
owrate) or simply operating costs. MPC also provides a control
framework that is robust in systems with environment uncertainties. MPC uses system models that normally do not require
great precision and complexity to achieve successful control
results.
Mine dewatering systems are generally composed by several
pumps associated with respective reservoirs. When it comes to the
control of these systems, the centralized approach could be costly
computationally, and in case of a control outage the whole system
will be affected. Given that dewatering systems are physically
distributed, and taking into account that distributed control systems offer a superior reliability, one may also consider noncentralized MPC implementations. This way, the pumping of water will not stop completely even if there is a fault in the control
system. Likewise, non-centralized apporaches may be easier to
implement physically given the difculty of deploying underground communication networks.
The idea behind of non-centralized approaches is to divide the
overall control problem into smaller pieces that are assigned to
local controllers, also known as agents. The natural partition of the
centralized control problem in this context is to consider that there
is an agent for each tank (Fig. 1), which controls the corresponding
water level and governs the pumping of water out of the tank. In
order to coordinate the control tasks of the different controllers, it
is considered in this work that the agents may exchange some
practical information related to the system and to other agents
operating state or intentions by means of a communication

1174

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

Fig. 1. Simple dewatering scheme; Arrows express direction of the ow. Each rectangle
represents a reservoir and its pump station.

network. The type and amount of information exchanged, if any,

has to be dened according to practical criteriaepumps that
discharge into other pumps reservoirs may communicate e and
taking into account the type of scheme used for the distribution of
the centralized control problem. The action of the later may
consider the actions and intentions of the former. Fig. 1 illustrates
the proposed simplied scheme, with several pump levels
exchanging water and information.
If complex systems are split into different control entities
operating individually while including some sort of communication, distributed and decentralized can improve reliability and
robustness of the overall system, reaching optimal or near optimal
control. For further description of these approaches the reader is
referred to [18,19]. Examples of MPC applied to water systems
include [2,3,20]. A case study distributed control in water systems
of different collaboration strategies among agents is presented in
Ref. [14].
Energy systems optimization has been broadly applied in the
industry. However, following the dual shortcoming of the control
perspective, it is difcult to prove that optimal energy system's
design will remain optimal, as mentioned, when performing in real
operating conditions different from the nominal ones for which
design was optimized. [21,22], for example, consider typical energy
demands and prices to calculate optimal design and operation of
energy systems, but there is a lack of discussion about how this
optimal system will perform when those demand and prices were
to change. The discussion is even more necessary because, when
considered, energy storage and renewable energy will follow
different patterns form day to day (far from typical day patterns) as
a result of change in weather conditions, price of energy or users'
behavior.
In light of the reviewed literature, it has been found a lack of
analysis of the complete optimization of energy systems, especially
those dealing with water systems, from design to control. Moreover, design for control is still a eld to explore in the case of
optimal control, especially model predictive control. Attempts of
link design and operation are limited; a good example can be found
in Ref. [23], although the methodology proposed does not accommodate time-variant controllers, such as MPC, which is the
approach selected of the present work.
In this work, MPC is applied to a mine dewatering system, which
has been initially optimized following the methodology used by
Ref. [22]. MPC deals with the existence of physical constraints and
uncertain operation conditions, which are common phenomena in
mining operations. The optimized design takes in to account

investment costs, ToU and DC. The latter is characterized statistically regarding the calculation methodology of Ontario electricity
Market [24]. MPC is used for the reasons already mentioned, but
also in order to validate DC's inuence in the optimal design. The
aim is to ll the gap between design optimization and optimal
control of water systems, which is considered a necessary and
previous step that enables the generalization of this approach, if
proven successful, to energy systems including different forms of
energy and technologies.
The outline of the rest of the paper is as follows: in Section 2
mine dewatering systems are described in order to familiarize the
reader with these installations. Section 3 explains the traditional
way for dealing with design and operation of dewatering systems.
In section 4 are introduced and developed the subsequent optimization problems: design and control, to which the dewatering
system is exposed in order to minimize total annual costs, including
investment and operation.
Section 5 describes the cases subject to study: the general
design and operating conditions and all the control strategies
applied. These conditions will dene the design and the control
optimization problems. The emphasis is done in the comparison
between centralized, distributed and decentralized control. The
paper follows with a results section (6), discussion (7) and
conclusion (8).
2. Mine dewatering systems
The dewatering system's function is to collect and remove water
from the mine, in order to maintain a safe operating environment.
The water collected comes from several sources, among which the
most important are: meteorological precipitation, ground water
seepage and service water used for drilling, dust control, washing
and even pipeline leaks [25]. The term inow will be used here to
refer to them effectively.
The nature of the inows can result in daily and even seasonally
varying ow proles. Service water for production processes can be
estimated by knowing how much water each productive task requires and, together with the expected seepage, the pumping
operation may be scheduled.
Dewatering facilities include different elements such as pipes,
valves, pumps, water reservoirs, sumps, among others. The system
is designed to meet specied performance requirements, mainly
water owrate and head. Due to the particular characteristics of the
uid pumped-water containing solid particles and dissolved
chemical substances-as well as the uncertainty in inows and
permitted times of operation, the design and operation of a mine
dewatering system can quickly become more complex.
Underground mines, up to thousands of meters deep, collect the
water in sumps from where it is pumped up to the surface or to an
intermediate sump level, where another pumping plant is located.
These plants consist of one or more pumps, normally connected in
parallel, and include the necessary valves, by-passes and lters to
permit proper set up, operation and maintenance. Due to the solid
content of the pumped water, it is necessary at least to provide
proper settling and decantation capacity.
3. Design and control of dewatering systems
Although determining the actual water ow collected in mine
sumps is challenging in advance of construction, it is not impossible
to estimate with enough precision to design safe and reliable
dewatering systems and schedule the pumping.
The rule of thumb for deciding upon underground dewatering
system dam capacity is that they have to be able to store between 8
and 24 h of water inow for each dam without pumping. Pump

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

discharge capacity at each level is designed to match the expected

inow. Normally, a higher capacity is installed to provide exibility
and backup pumps feature to increase safety in case of unexpectedly high inows.
Underground mine pump operation is generally triggered by
maximum and minimum water levels at each water reservoir
(dams or sumps) where pumps are installed. The resulting pumping schedule can only be improved upon with deferred pumping
strategies if enough storage and pumping capacity is available
([3,6]). Regardless the type of mine and pumping strategy, in case of
emergency automatic signals activate pumps so that the mine
environment is always properly dewatered.
4. Optimizing design and operation of mine dewatering
systems
4.1. Optimal design
The optimization process of a dewatering system is divided into
two stages. The rst considers design optimization and the second
concerns operation control.
In the present work design only affects the calculation of dams'
capacity. The number of pumps, their arrangement, as well as the
diameter and material of the pipe work can be capably left to
common practice design. This consideration is based upon the fact
that pump capacity is mainly determined on worst pumping scenarios (high and constant inows).
Optimized water reservoir capacities are calculated with a
variant of the Optimal Mine Site Energy Supply methodology
described in Ref. [22]. Dewatering demand, as well as electricity
tariff, was characterized as though the dewatering system were the
only energy system contemplated in the mine. Dewatering demands are treated as material ows that interact with energy ows
by means of the pumps. Balance equations (constraints) in water
and energy were implemented using the system of [22]. Economical parameters include: the cost of building the dams, expressed in
economic units per unit of volumetric storage capacity; the hourly

1175

electricity price; and the demand charges, calculated for individual

hours of the year, as described in the same reference paper.
For illustrative purposes a complete dewatering system is split
into two similar subsystems each comprising a pump and a reservoir (Fig. 2). Pumps operate over the whole range of speeds admissible, leading to a corresponding range of discharge (m3/h),
following [3]. Two pumps connected to their respective dams or
water reservoirs effect mine dewatering. The inows are assumed
known, as is the performance of the pumps (their relationship
between owrate and power consumed). Similarly, the price of the
electricity is also assumed known in advance. Consequently, the
results will indicate a design conditioned by the expected environment in which is going to operate.
The difference with previous models in the literature (e.g.
Ref. [3]) is the use of a linear approximation of the power-owrate
function for variable speed pumps. In the case of the design problem, the optimization problem remains linear; this is a desirable
characteristic in terms of computational burden, and because the
same model is used in the operation optimization problem, the
coherency of the models permits direct comparison between both
design and the operation problem's results. It is also considered
that there exist minimum operating ows given by, for instance, a
requirement for a minimum rotational speed or minimum ow
within the pipe work in order to avoid solids deposition. Maximum
ow may be limited by the frequency converter or the requirement
to avoid cavitation. Pumps can be activated or not, so the ow limits
only apply in the prior condition.
The model therefore denes a mixed integer linear programming problem, where, in addition to reservoirs size, hourly pump
states and owrates are the decision variables. Thus, the solution of
the design problem includes an optimal schedule of pump operation (that may or may not remain optimal depending on the real
operating conditions: inows and electricity prices), given the
schedules of inows and electricity prices.
When one considers the information needs of this formulation,
it quickly becomes clear that the optimal design solution can only
be really regarded as optimal for the given schedules of inows and

Fig. 2. Dewatering system with two lifting stages in an underground mine.

1176

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

prices, which can only be known certainly historically, even if

probabilistic approaches can be taken to characterize the inows
and pricing time series. In more simplistic terms, the problem of
design optimization produces and optimal design for last years'
(day, month) environment. The question remains: how will this
optimum design fare on the coming year's inows and prices? Will
it remain optimum?
4.2. Extensions for robust optimality in operations
This work extends the determination of an optimal design that
incorporates determination of an optimal pumping plan (albeit for
static, possibly historical statements of the price environment and
the water inow environment) using OMSES technique [22] with a
subsequent phase of operational optimization e pump control e
solved using MPC. The system behavior, assumed linear (principally
for reasons of computational efcacy), is discretized over a specied time horizon. The model used here is based on similar MPC
works found in the literature [3], and considers cases of centralized,
distributed, and decentralized control.
4.2.1. Centralized model predictive control (CMPC)
A centralized control assumes a complete knowledge of the
whole system's state and centralized decision making. The solution
of the problem, if feasible, is the optimal pumping plan. This
translates to achieving minimum operating costs, while respecting
the constraints of maximum and minimum reservoir water levels
and maximum and minimum discharge of each pump, if activated.
As mentioned in previous sections, centralized control compromises the whole system in case of communication outage. Noncentralized approaches can enhance system control robustness or
the ability of the whole system to safely operate even when each
pump is behaving without knowledge of its neighbor's control
objectives.
4.2.2. Distributed model predictive control (DMPC)
The distributed model predictive control (DMPC) problem assumes perfect communication between the subsystems (agents),
the pump-reservoir couples of each level and selsh decision
making by each agent. The constraints of each subsystem are
reservoir level and ow rate limits. The main consideration relates
to the sequence of the optimization, the information shared and the
possibility of recalculating the respective pumping plans as part of a
type of bargain between the agents. For DMPC the present work
assumes serial progressive optimization in which in the bottoming
subsystem optimizes its plan rst based on its predicted inows
and electricity prices. Then, the upper subsystem optimizes its plan
making use of the solution (pumping plan, including current
pumping state) of the former, as well as its expected water inows
and environmental electricity price. The overall solution obtained
may not be equivalent to the optimal reservoir capacities and
pumping plan of the centralized solution, but the DMPC subproblems embody a practical simplication in sub-surface communications and a control robustness that is desirable in a mining
context.
4.2.3. Decentralized model predictive control (DcMPC)
In absence of communication between controllers a so-called
decentralized model predictive control (DcMPC) case arises which
can be solved through adoption of one of a variety of control policy
augmentations that are described in the literature [26]. Here it is
solved by assuring that there is sufcient remaining storage capacity of an upper reservoir to accommodate ow discharged by a
lower pump, sustained for a specied period of time. An MPC
system following a lower water level set point is implemented in

the upper subsystem, which also considers that the expected ow

from the lower tank is equal to that received from it during the time
interval immediately prior to the current time interval of consideration. This assumption is reasonable due to the trend of the lower
pump to maintain discharge ows, given that the ToU prices are
broadly similar from 1 h to the next.
The selected tariff structure consists of a payment for the
consumed energy (ToU term) and an additional charge for the
power demand charge (DC term) applicable for a certain interval of
time. The value of the DC is adopted from work reported in Ref. [22]
e a cost is added monthly to the energy bill based on the power
demanded by the dewatering system from the utility during
particular hours of the year when the grid (at a national or provincial scale) suffers from a peak in power demand. The hours
where peak grid demand is expected can be to a certain extent
forecasted e in the case of Ontario the peak is reached during
extreme summer days [22].
4.3. Problem formulation
4.3.1. Design problem
The objective function of the design optimization problem can
be expressed as follows

Min Ctot Cfix Cvar

(1)

Ctot is the total annual cost, while Cx is the annuitized xed cost
(derived from dam's construction) and Cvar is the annualized variable cost (pump electricity consumption and maintenance costs).
The annuitized xed cost is expressed by the Eq. (2). The variable cost integrates the operation along the 24 h of each representative day of the year (here, a representative day applies for each
month).

Cfix crf $1 fic $

X
CDW;p $Rp

(2)

Cvar

XXXh
p

i
X
pij $Fpij $tij
GAi

(3)

where crf is the capital recovery factor, also known as annuitization

factor, and fic accounts for engineering and supervision expenses,
etc; CDW,p is the cost of water storage per unit of volume and Rp is
the respective reservoir capacity for p 1,2; pij the cost of electricity
in each hour of each representative day; and GAi is the monthly
payment for demand charges. tij is the duration of the interval of
time for the typical day j 1,,12 (twelve typical days, one for each
month) at the hour i 1,,24, and F pij is the energy consumed by
the pump p 1,2 at day j within the hour i. For further interpretation of Eq. (2) and Eq. (3) the reader is referred to [22].
4.3.2. Control problem
The operational optimization problem is dened with the same
balance equations of the design problem, and solved every time
interval in a rolling horizon of prices and inows. The control
problem objective function is therefore:

Min Ctot Cvar

Cvar

XXh
p

(4)

i
2

X
pi Dpi F0pi $ti
CN;P $ Lref  LN;p

(5)

The objective function dened by (4) and (5) do not include

xed costs, but instead, (5) includes a quadratic cost function of
the level of the reservoir at a certain time interval, LN,p, which is

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

taken to be the last time interval of the rolling horizon. This term is
used only in DMPC and DcMPC cases where partial or no information sharing between individual controllers apply. The term
implements a localised control policy that aims to ensure the current reservoir water level is attracted to a reference level, Lref e
driven by a virtual cost. The co-efcient CN,p is tuned to the specic
case. This term makes no contribution to the actual annual cost of
operating the system, it merely points the system toward a given
attracting state; the actual annual costs are represented by the right
hand term only, in (5), with actual electricity prices, pi, and energy,
F0 pi. In optimal control, Cvar is calculated as the sum of every term
within the control horizon (H), (H: i 1,, N). N will be generally
24 h, or less if system stability requires. In comparison to (3), in (5)
the DC term is removed andDpi is included. The latter is an additional electricity cost that will affect only to the hours where expected demand charges have the inuence (i.e. Summer afternoon
hours, from 12pm to 6pm). This price increment is set as high as
needed to discourage pumping in those hours, normally considering that Dpi >> pi.
Water balance equations, equivalent to energy balance equations in Ref. [22], remain similar but are subject to two further
considerations. First, water storage efciencies are equal to 1 (for
both design and control problem). Second, and only for the operational control extension, the 24 h balance expressing that the
water level at the beginning and the end of that time span is the
same, is removed from the equations. This is clearly an unnecessary
over constraint for MPC.
Equation (5) is formulated for p 1, 2 in the case of the
centralized problem. In the distributed and decentralized cases this
subscript is unnecessary, because the problems are solved for each
subsystem.
5. Framework for assessment
The case study adopted to illustrate the approach to optimal
design is the dewatering system of an underground (UG) mine in
Ontario. Firstly, the optimal design is calculated using mathematical programming technique subject to the expected electricity
prices and inows and technical constraints. Subsequently the
resulting system is simulated under MPC, seeking minimum cost
operation. All outcomes of optimal design and optimal control are
referenced to a Base Case that faithfully represents traditional
design and control practice of dewatering systems in mines.
The ensuing discussion sets out by describing the characterization of the economic environment and the tariff environment
faced by the system and continues by describing how pump duty is
modulated within the model. Next the various cases and scenarios
explored are described. These aim to elicit the merits or relative
demerits of each case explored, where the principal variable is the
type of MPC assumed adopted. This section sets out the framework
for assessment of these mine dewatering system optimization
schemes. Results of each case and scenario and represented in
section 6.

1177

averaging process involved in establishing the price proles reduces variance in the prices for the representative day. When
considering real time hourly prices, as is the case with the operational optimization problem that follows, the enhanced potential
savings will be recognized under MPC. To illustrate this point,
Fig. 3 compares the average price and the real time price for three
consecutive days.
For the operational optimization, or control problem, the predicted price of electricity is assumed to be perfectly known. This is
generally false, but what is available 24 h (on a rolling basis) in
advance is a predicted price. These prices are known to be volatile
during already high price periods, exemplied by the Ontario
network [24], and such circumstances clearly identify the need for
adaptive responses in the pumping plan of a system already optimized for design.
Water reservoirs capacity may be constrained by an upper limit
during optimal design of the dewatering system. There will always
be space limitations for this kind of infrastructure (perhaps geotechnically constrained at depth), and here it is considered that it is
possible to excavate and build a proper dam or sump in the mine up
to 3000 m3. This assumption does not affect an estimated cost of
construction of the reservoirs, which is assumed to be proportionally dependent on the size (volume).

5.2. Water inow

For each reservoir, the inow can be approximated by considering a water ow term dependent on production (drilling,
washing, etc) and a term dependent on inltration from the water
table, aquifers, rain, snow, etc. The former may have a daily pattern;
the later could be relatively constant throughout days representative of each month, but with seasonal variations month to month
(or representative day to representative day).
The inows adopted in Fig. 4, which are used for both the design
and the control optimization problems. Inltration is considered as
a constant ow rate while production associated ows follow the
same hourly pattern for every representative day of the year. The
sum of seepage and production associated inow yields an inow
for each hour of the day for each reservoir. The upper sump does
not receive water other than from seepage and the lower reservoir.

5.1. Economic environment

For the design stage, hourly price proles have been established
for representative days of each of the twelve months, using 2011
electricity prices in Ontario, Canada. Demand charges, also representative of the Ontario system are calculated based upon the
consumption for the representative day of July, where peak grid
demand is expected. For more detail concerning the electric tariff,
consult [22], as well as [24].
This averaged ToU prole may underestimate savings from
load shifting actions in design optimization, because the

Fig. 3. HOEP June 2011; Actual (dotted line) Vs average (continuous line) hourly
prole.

1178

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

Fig. 4. Water inows; Left: Monthly inltration inows (constant ow) for each reservoir; Right: water inow associated with mine production activities into the lower reservoir.

While the design problem uses the typical daily inow pattern, the
control problem repeats the same daily inow pattern for each day
of the considered month. In the case of more accurate knowledge of
mining operations or inltration behavior being available, the
predicted inows for the control horizon will be improved with
improved estimates of savings.
5.3. Variable speed pump implementation
The performance of a pump is dened by the ow that is able to
deliver when connected to a pipework system, as it consumes a
certain amount of power. This eventually denes the ToU costs. The
methodology chosen for both problems (design and control) is
based on the characterization of the couple formed by a pump and
the system in which it is installed. Given a pump with variable
rotational speed and a xed system arrangement (xed static head,
valves, etc) the power that a variable speed drive will consume as a
function of the owrate delivered by the pump can be analytically
characterized by a series of interpolations and substitutions of the
actual curves of both the pump and the system, as well as the afnity laws [3].
Here the resulting cubic expression for the power-owrate
curve following [3] is replaced by a linear approximation, as

Fig. 5. Example of power-ow curves for a variable speed pump: cubic (continuous
line) and linear (dashed line) approximations.

illustrated in Fig. 5. The cubic expression, drawn in Fig. 5 as a

continuous line, is a cubic interpolation across data from an actual
pump and a known system. This precise denition may not be
necessary if we are interested in the performance of the pump in a
narrower range of operation, and a lower order function (i.e.
linear, dashed line in Fig. 5) could reduce simulation and optimization time of both the design and control problem. If this is the
case, the emphasis when linearizing the power-ow function has
to be placed in establishing the range where the system will be
practically operated.
5.4. Design and control cases explored
Fig. 6 presents a structured listing of the individual situations
that are considered in this case study example that aims to elucidate the various considerations articulated. The individual cases are
given abbreviated names, and for each case, the effects of two DC
scenarios are simulated: i) maintain an unavoidable and thus xed
DC and ii) minimize the DC. The benchmark case (TD) adopts the
traditional design method for determination of reservoir capacity
and considers maximum and minimum level switch control with
constant speed pumps (when they are in operation). For a given
reservoir, its volume is calculated by summing eight (8) hours mine
inows, including any ow rate arriving from other pumps. Practically, this amounts to 8 times the mean inows for each representative day. This type of control offers no sensitivity to ToU
electricity prices or DC.
The remaining four cases (OD, CMPC, DMPC and DcMPC) aim to
reveal benets or detriments of MPC options, illustrated in Fig. 7,
each adopting the plant portfolio of the OMSES design stage. Simulations are conducted to reect the two scenarios for DC.
For these cases, having obtained the optimal design, the MPC
simulations aim to permit assessment of the robustness of optimality in the face of unexpected changes in the environment,
which in this case could be for either the tariff environment or the
physical environment. Consequently, upon the tariff time series for
electricity price and DC used in design optimization, an additional
electricity price hike is superimposed to reect times of peak
system demand; the hourly price that the controller senses is
deliberately increased (Dpi 600 CAD/MWh for these trials, for
i 12,,18 h). Therefore, although this perceived price only is used
by the controller, the desired outcome of reducing DC is possible
even when ToU charges signicantly increase.
Communications between controllers for the varying MPC cases
(CMPC, DMPC & DcMPC) are indicated in Fig. 7. CMPC involves ow
of complete system state, environmental and planning information

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

1179

Fig. 6. Dewatering system optimization & control simulation hierarchical summary, illustrating design, optimization and simulation cases considered (cases abbreviated with
emboldened letters).

available at the centralized controller. The partial information ows

for DMPC and DcMPC are illustrated with dotted circles and arrows
linking the variables.
For DMPC the pump in the lower mine level rst optimizes its
control actions for the given planning horizon; a 24 h long planning
horizon is adopted. Information of the plan is sent to the controller
at the upper level, which optimizes its plan in accordance with this
information and its forecasted mine inows and electricity price. In
addition to the linear cost, the quadratic function term in (5) is
activated to maintain the water level around a set point that assures
sufcient reservoir volume to deal with high owrate from the
pump of the lower level. The DMPC implementation adopted has
unidirectional information ow (dotted circles and lines in Fig. 7)
which can lead to shortcomings of DMPC as some controllers may
not be able to sense plans for other parts of the system that may be
material. The upper subsystem has a control horizon and prediction
of 5 h that, as it will be discussed, is enough in this problem. The
referred set-point is selected once the design volumes have been
obtained.
In DcMPC, there is no information exchange between controllers. The upper agent deals with this situation by assuming that, for
a given interval, the current discharge from the lower pump is to be
sustained during several hours. Practically, rather than in these
exploratory simulations, this duration may be established heuristically based upon historical pump operation. DcMPC systems
generally achieve inferior results than DMPC systems, especially
when the latter may exchange information bi-directionally and
iteratively.
The problem of overow becomes more complicated in DcMPC,
that is, pumping plans may violate constraints that were in fact

embedded in the design problem. Some of the actions that render

the feasibility of the top system (upper reservoir) are:
- Limit a lower subsystem's pump speed. The immediate consequence is a lower owrate that provides more time to react,
especially in the case we consider, where system measure and
control actions are applied each hour.
- Lower the reference level for the quadratic term in the cost
function. If one tunes the terminal cost to have a low reference
level (Lref) and/or a higher penalty factor (CN,p) then at every
control step there will be more room for inow water. However,
this reduces actual cost optimization (CAD).
- Increase upper reservoir capacity. This action is related to the
initial design, not to the controller itself. Increase capacity helps
when the subsystem has not enough pump capacity in case of a
high ow rates during several hours,
For the example problem considered, it was necessary to adopt
all of these actions in combination in order to return solutions that
were feasible. The approach followed is to maintain upper reservoir's water level as low as possible. June is the month with the
highest inow conditions, and therefore the controllers are more
stressed. The values used for the optimization problem are given in
Table 1. A simplied scheme is shown in Fig. 7.
6. Results
Table 2 shows the results of system design using traditional
design methods and OMSES methodology, expressed as the discounted annual cost of dewatering, with discount rate of 10% and

1180

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

Fig. 7. Dewatering system schematic and MPC options evaluated in simulations. Distributed MPC and Decentralised MPC have partial information ows intra and inter-MPC
controller (dots).

20 year system's life. R1 refers to the bottom reservoir, while R2

does for the top level reservoir. The results for the operating costs
for the traditional design, both ToU and DC, have been obtained
from simulation of the high and low level triggers control.
Table 2 shows that traditional design involves higher investment in reservoirs. Total investment cost is lower in both OD1
(optimal design without DC minimization) and OD2 (optimal
design with DC minimization) compared with TD case. The optimized designs assign more capacity to the lower reservoir, R1>R2
in comparison with TD, where R2>R1. The savings obtained are

clearly favourable to optimized solutions, but the control problem

will reveal if these idealized amounts are reasonable or not.
Table 3 shows the results of the operational problem. DMPC and
DcMPC results are not shown, for they produced infeasible control
problems with the same simulation conditions as CMPC case (24 h
control and prediction horizon, OD2 design). CMPC1, DMPC1 and
DcMPC1 refer to control without DC minimization, while CMPC2,
DMPC2 and DcMPC2 do minimize DC. Robust non centralized
controllers were not possible to realize unless the upper reservoir
capacity or upper pump maximum owrate were increased.

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

Table 1
Inputs for design problem.

Table 3
Summary of optimized operation using CMPC.

Pump performance
Pump capacity (Q)
Range (Q)
Max
Min
Nominal head
Nominal power
Bottom level (L1)
Pumps number
Max size reservoir
Top level (L2)
Pumps number
Max size reservoir
Other inputs
Interest rate (i)
Life (y)
Capital Recovery Factor (crf)
Reservoir construction costs (CDW)
Installation cost factor (fic)

Case
[m3/h]

400

[-]
[-]
[m]
[kW]

120%
80%
440
800

[-]
[m3]

1
3000

[-]
[m3]

1
3000

[-]
[year]
[-]
[$/m3]
[-]

0.1
20
0.12
100
1.15

In order to check optimality of OMSES approach, as well as for

assuring feasibility of non-centralized approaches, upper reservoir
capacity was increased 2.5 arbitrarily, but such that the DMPC and
DcMPC solutions became feasible and so the investment costs were
of the same order as those for the TD case. To permit comparison
with the distributed and decentralized control cases, the feasible
solutions of CMPC1 and CMPC2 of Table 3 were recalculated with
increased reservoir capacities to produce the results presented as
CMPC10 and CMPC2' of Table 4.
For completeness for any reader who may wish to conrm the
results of the simulations, a summary of the control setup for
distributed and decentralized control, used to obtain the feasible
solutions, is shown in Table 5.
The results supercially indicate that:
1. MPC results are superior to TD for feasible solutions
2. CMPC results are superior to DMPC and DcMPC
3. DMPC and DcMPC became feasible only when the optimum
design using OMSES is ignored and an intervention is exercised.
7. Discussion
The work presented represents a novel approach that puts
together two currently prevalent approaches to energy systems
optimization. The rst of these is the energy system design optimization, based on expected operating conditions, which has been
shown to be successful in identifying energy savings. The second
approach is Model Predictive Control applied to energy systems
operational optimization. The link that makes it possible to obtain
results for the design optimization that are comparable with the

Table 2
Results Design optimization.

Case
Reservoir size
R1
[m3]
R2
[m3]
Costs
[CAD/year]
ToU Cost
Demand Charges
Reservoirs (Annuitized)
Discounted annual cost of dewatering
Difference cost relative to traditional

1181

Traditional

Optimized design

TD

OD1

OD2

1560
3560

1222
556

2469
1050

258516
382677
69160
710353

234080
432790
24017
690886
3%

226921
77866
47534
352321
50%

R1
R2
ToU
Demand charges
Peaks:

Average powera
Invest.
Total
Savingsb

[m3]
[m3]
[CAD/yr]
[CAD/yr]
July 4pm
July 4pm
July 5pm
July 2pm
[MW]
[CAD/yr]

TD

CMPC1

CMPC2

1560
3560
258516
382677
1.92
1.92
1.92
0.96
1.68
69160

2469
1050
188631
145782
0
0.96
0.64
0.96
0.64
47534

2469
1050
190883
0
0
0
0
0
0
47534

710362

381950

238417

46%

66%

Power demanded in the reference hours; DC is calculated as the average of ve

reference hours (2011 peak demands in Ontario). In order to simplify the calculations, only reference hours of July were used for the average, which is consistent
with the approach of the design problem.
b
Difference cost relative to traditional.

operational optimization is that the simulation model is the same

system model, the denition of the problem is as a mathematical
programming problem and a sufcient knowledge of the operating
conditions exist, at least for OMSES and CMPC combined.
The average price used to calculate the optimal design may not
show the real potential of load shifting strategies but it will give
enough information to achieve the optimal design. For CMPC,
comparison between Tables 3 and 4 reveals how increasing reservoir size (R2 in this case) gives lower annual cost savings. Any
operational cost reduction is countered by the increase of investment costs, thus demonstrating that, for the conditions simulated,
design remains optimal even subject to price variability and
uncertainty.
Given the choice, lower savings for the DMPC or DcMPC, mean
that they are undesirable in comparison to a centralized controller.
This is a result of the reduced level of communication that exists
with the non-centralized systems. How would this reduced level of
communication occur? It would not be designed in through selection of, say, DMPC over CMPC, rather the communication system
for CMPC could fail. Table 4 illustrates the effect one of many types
of intervention that could mitigate the effect of partial communications failure with the decentralized MPC system reecting a more
serious failure.
Such interventions override optimal design parameters and
potentially call the usefulness of OMSES into question. However
using either OMSES or CMPC on their own for design (for the interventions imply a return to design), or using them together, leads
to dramatic savings improvements, but using them together increases savings so one would always opt for this combined
approach.
DMPC and DcMPC approaches, even with articially increased
upper reservoir size for reliability reasons, substantially improved
upon TD. Better performance of DcMPC compared with CMPC and
DMPC, for non-minimized DC, is due to a random casual
improvement regarding power consumption in July which affects
DC. As expected, if DC is minimized, then the CMPC approach is the
best, followed by DMPC and nally DcMPC.
If the demand charge is not-deterministically dened (as is the
case explored herein because Ontario demand charges cannot be
determined deterministically in advance of the occurrence of the
peaks in a given year) design optimization can be improved using
previous year's data for the occurrence of the peaks, to estimate
occurrence in the current year, statistically. Operations optimization through improved control can be applied in the same fashion,

1182

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

Table 4
MPC comparison with increased systems design robustness.
Case
R1
R2
ToU
Demand charges
Peaks:

Average power
Invest.

[m3]
[m3]
[CAD/yr]
[CAD/yr]
July 4pm
July 4pm
July 5pm
July 2pm
[MW]
[CAD/yr]

Total
Savings

CMPC10

CMPC20

DMPC1

DMPC2

DcMPC1

DcMPC2

2469
2625
198885
145782
0
0.96
0.64
0.96
0.64
68809

2469
2625
201918
0
0
0
0
0
0
68809

2469
2625
221340
118698
0
0.7188
0.64
0.7256
0.5211
68809

2469
2625
222526
0
0
0
0
0
0
68809

2469
2625
232046
100225
0.80
0.96
0
0
0.44
68809

2469
2625
230164
44976
0.79
0
0
0
0.20
68809

413476

270727

408850

291335

401082

343950

42%

62%

42%

59%

44%

52%

in this case penalizing energy consumption by articially increasing

electricity price, which creates a perceived higher price during the
undesirable hours.
It is clear that if a high unpredicted inow happens, any
controller designed here (re-optimizing every hour) may not be
able to pump out the water fast enough to prevent overow,
especially when the reservoir is near capacity and for noncentralized control. In mining, safety is always a priority, and will
be prioritized before any energy cost saving strategy. The choice of
whether one adopts TD or CMPC depends upon the relative resilience of each design and control system to extreme environmental
events. While we have considered how resilient a CMPC system is,
through its degradation to DMPC and subsequently DcMPC, we
have not established how well a TD system would behave in the
same circumstances. The reason for this is that such a determination is strongly dependent on the state of the system when the
extreme event occurs. The results of such analyses, which are
through necessity, computationally comprehensive, will be reported elsewhere, in future.
Even without a more precise simulation that will reveal model
inaccuracies regarding real power consumption, costs savings
remain of the same order. For example: it has been determined that
the dewatering simulation model would consume up to 10% more
energy using the cubic expression, compared with the linear one,
(Fig. 5) when the ow rate is at 120% the nominal value. A 10%
increase in the energy cost is still absorbed by load shifting, the
investment savings and the DC minimization, depending on the
MPC case we focus on.
If the number of pumps is included as a decision variable, an
increase of the number of pumps is expected. The optimal number
of pumps will depend on several variables: specic cost of the
pumps (CAD per kW or per m3/h), cost of the reservoirs, inows
and electricity prices (ToU and DC). If more pumps are installed,
more economic advantage can be taken of the load shifting
opportunities.

8. Conclusions
Cost minimization strategies such as load shifting, whether it is
related to time of use or demand charges, requires understanding of
Table 5
Non centralized controllers setup.
Parameter

Units

DMPC

DcMPC

CN
Lref
N
Q1,max

[CAD]
[m3]
[-]
[%]

0.01
500
5
120

1
300
2
100

the system dynamics, the environment in which it performs, and

the consequences of applying the resulting control actions.
In this paper, an optimization strategy has been applied to a
mine dewatering system in order to minimize its total annual cost,
dened by the investment costs and the operation costs.
Using expected owrate and electric tariff data, an optimized
design was obtained by means of mathematical programming. This
design is signicantly different from the traditional approach
where, at each mine level, the size of water dams or reservoirs
depend upon its own expected inow.
In order to evaluate the effectiveness of this optimal design
solution, model predictive control was applied to a model based
simulation. Centralized and non centralized control were tested
and compared. For the environmental conditions used (inows and
prices), the centralized MPC was the only control approach that
ensured feasibility, i.e. it perfectly avoided overow. CMPC was
used to test the optimality of the optimal design. Increased size of
the upper reservoir didn't lead to higher savings, which proves the
optimality of the design.
In a later analysis with improved reliability of the controllers
and the system (upper reservoir size), the three of them got similar
performance when DC was not subject to minimization, otherwise
centralized MPC got better results than the distributed and
decentralized. These also largely improved the annual costs of the
traditional design and its usual mode of operation based on
switching triggers (High level/Low level).
The distributed problem, as well as the decentralized, revealed
the shortcomings of having two constrained, coupled optimization
problems with insufcient communication. Upper reservoir overow was overcome by introducing a quadratic terminal cost to
maintain the water level close to a feasible and safe operating point,
as well as by system's design modication. For the decentralized
problem it was necessary also to reduce lower pump maximum
owrate.
The demand charges minimization strategy applied to MPC
demonstrated its validity but it had to be adapted to achieve the
objective of the optimal design.
Further analyses should be executed regarding inow perturbations, as well as model inaccuracies impact. Different communication options and bargain strategies between the controllers
should also be explored in order to improve distributed solution.
The present work is the preliminary step towards the generalization of MPC application to energy systems design optimization,
especially for OMSES.
Acknowledgements
This research was supported by the Ontario Research Funded
Research Excellence Round 5 research project Smart Underground

A. Romero et al. / Applied Thermal Engineering 90 (2015) 1172e1183

Monitoring and Integrated Technologies (SUMIT), Sub-Project 6:

Sustainable Operations e Energy Management and Ventilation on
demand. Financial support by the FP7-ICT DYMASOS (under grant
agreement No 611281) and by the Junta de Andaluca (by means of

n Optima
the Gestio
de Edicios de Energa Cero with reference P11TEP-8129) is gratefully acknowledged. The rst author gratefully
acknowledges the use of the services and facilities of the Dept.
tica, at the University of Seville, and
Ingeniera de Sistemas y Automa
the generous advice from their research group.
References
[1] I. Pulido-Calvo, J.C. Gutierrez-Estrada, R. Asensio-Fernandez, Optimal design of
pumping stations of inland intensive shfarms, Aquac. Eng. 35 (2006)
283e291.
[2] H. Zhang, X. Xia, J. Zhang, Optimal sizing and operation of pumping systems to
achieve energy efciency and load shifting, Electr. Power Energy Syst. 86
(2012) 41e50.
[3] H. Zhang, Optimal Sizing and Operation of Pumping Systems to Achieve Energy Efciency and Load Shifting (Master thesis), 2011.
[4] Wojciech Kurek, Avi Ostfeld, Multi-objective optimization of water quality,
pumps operation, and storage sizing of water distribution systems, J. Environ.
Manag. 115 (2013) 189e197.
[5] X. Zuhan, X. Xia, Optimal operation scheduling of a pumping station with
multiple pumps, Appl. Energy 104 (2013) 250e257.
[6] J. Vosloo, L. Liebenberg, D. Velleman, Case study: energy savings for a deepmine water reticulation system, Appl. Energy 92 (2012) 328e335.
[7] A.J. van Staden, J. Zhang, X. Xia, A model predictive control strategy for load
shifting in a water pumping scheme with maximum demand charges, Appl.
Energy 88 (2011) 4785e4794.
[8] Y.C. Chen, B. Coulbeck, Optimized operation of water supply systems containing a mixture of xed and variable speed pumps, Int. Conf. Control 32
(1991) 1200e1205 (conference publication 32).
[9] G.E. du Plessis, L. Liebenberg, E.H. Matheus, Case study: the effects of variable
ow energy saving strategy on a deep-mine cooling system, Appl. Energy 102
(2013) 700e709.
[10] J. Zhang, X. Xia, A model predictive control approach to the periodic implementation of the solutions of the optimal dynamic resource allocation problem, Automatica 47 (2011) 358e362.

1183

[11] Z. Yang, H. Borsting, Energy efcient control of a boosting system with multiple variable-speed pumps in parallel, in: Proceedings of the IEEE Conference
on Decision and Control (49th IEEE CDC, 2010), 2010, pp. 2198e2203.
[12] M.M. Al-Gwiaz, M. Noureldin, Pump load management via repeated simulation, in: AIChE Annual Meeting, Conference Proceedings 2008, 2008.
[13] G. McCormick, R.S. Powell, Derivation of near-optimal pump schedules for
water distribution by simulated annealing, J. Oper. Res. Soc. 55 (2004)
728e736.
[14] I. Alvarado, D. Limon, D. Munoz de la Pena, J.M. Maestre, M.A. Ridao, H. Sheu,
W. Marquart, R.R. Negenborn, B. De Schutter, F. Valencia, J. Espinosa,
A comparative analysis of distributed MPC techniques applied to the HD-MPC
four-tank benchmark, J. Process Control 21 (2011) 800e815.
[15] EPANET http://www.epa.gov/nrmrl/wswrd/dw/epanet.html (accessed June
2014).
[16] WaterCAD
http://www.bentley.com/en-US/Products/WaterCAD/(accessed
June 2014).
[17] E.F. Camacho, C. Bordons, Model Predictive Control, Springer Verlag, 2004.
[18] R. Negenborn, Multi-agent Model Predictive Control with Applications to
Power Networks (Doctoral Thesis), Delft University of Technology, 2007.
[19] J.M. Maestre, R.R. Negenborn, Distributed Model Predictive Control Made
Easy, Springer, 2013.
[20] A. Middelberg, J. Xhang, X. Xia, An optimal control model for load shifting e
with application in the energy management of a colliery, Appl. Energy 86
(2009) 1266e1273.
[21] M. Stadler, C. Marnay, A. Siddiqui, J. Lai, B. Coffey, H. Aki, Effect of Heat and
Electricity Storage and Realiability on Microgrid Availability: a Study of
Commercial Buildings in California and New York States, Environmental Energy Technologies Division, Ernest Orlando Lawrence Berkeley National Laboratory, 2009.
[22] M. Carvalho, A. Romero, G. Shields, D.L. Millar, Optimal synthesis of energy
supply systems for remote open pit mines, Appl. Therm. Eng. 64 (2014)
315e330.
[23] I.K. Kookos, J.D. Perkins, An algorithm for simultaneous process design and
control, Ind. Eng. Chem. Res. 40 (2001) 4079e4088.
[24] IESO Independent Electricity System Operator, Changes to the Global
Adjustment,
2014.
http://www.ieso.ca/imoweb/b100/ga_changes.asp
(accessed May 2014).
[25] J. De la Vergne, The Hard Rock Miner's Handbook, McIntosh Engineering,
2003.
[26] P.D. Christodes, R. Scattolini, D. Munoz de la Pena, J. Liu, Distributed model
predictive control: a tutorial review and future research directions, Comput.
Chem. Eng. 51 (2013) 21e41.