Vous êtes sur la page 1sur 5

Optimal Pump Scheduling Strategies Accounting for

Background Leakages and Energy Cost


Luigi Berardi
Daniele B. Laucelli
Antonietta Simone
Orazio Giustolisi
Department of Civil Engineering Sciences and Architecture (DICAR)
Technical University of Bari
Bari, Italy
luigi.berardi@poliba.it, danielebiagio.laucelli@poliba.it, antonietta.simone@poliba.it, orazio.giustolisi@poliba.it

Abstract— Pumping optimization in water distribution networks minimizing the energy for pumping would result into water
means to schedule the ON/OFF status of pumps over a typical resource saving and asset deterioration delay. For this reason,
operating cycle, pursuing the minimum energy cost for pumping the cost of each cubic meter of water lost should include the
and matching technical requirements, e.g. sufficient pressure to “environmental and indirect cost” value besides the mere tariff
satisfy customers’ water requests, null water deficit at tanks, of water supply service [5]. Optimal pump scheduling can be
avoid complete emptying and overflows of tanks. In aged water performed using two possible strategies: (i) by programming
distribution networks, pump scheduling should also minimize the ON/OFF status of pumps over time or (ii) by setting water
pressure-dependent leakages, entailing loss of water and energy, level in tanks that control pump status. Actually, tank levels
as well as asset deterioration. Although pumps are usually
mirror the water mass balance through the WDN, thus being
scheduled over time, controlling pump status by water level in
tanks provides a more robust strategy in face of uncertain water
more robust in face of uncertainties on WDN model boundary
demands. This work demonstrates that advanced hydraulic conditions, actual customers’ water requests and spatial
modelling, enabling the simulation of pressure-dependent distribution of pressure-dependent background leakages.
background leakages and variable level tanks, allows optimizing This contribution shows that advanced optimization
levels in tanks that control pumps reducing energy costs for strategies allow to support optimal pump scheduling
pumping and leakages, in a complex water distribution network. accounting for all the above mentioned technical objectives
and constraints that would be difficult to achieve using
Index Terms - Background Leakages, Energy consumption,
Pump scheduling, Water distribution networks.
heuristic approaches. The C-Town WDN including
background leakages model [5] is used since is encompasses
11 pumps, whose status is controlled by water levels in 6 out
I. INTRODUCTION of 7 tanks. The advanced hydraulic model accounting for
Energy saving is a crucial element of water distribution pressure-dependent background leakages and robust
networks (WDNs) operation. Even a small increase of simulation of water levels in tanks is used to analyze WDN
operational efficiency in pumping water over an operating status at each time step of one-week long operating cycle.
cycles can result into significant energy and cost savings for
water companies on annual basis [1][2]. Pumping optimization II. WDN HYDRAULIC MODEL
is usually driven by electricity tariffs and hydraulic Pump scheduling needs the analysis of the hydraulic
requirements to satisfy customers’ water requests, both system behavior over all time-steps of an operating cycle. This
varying over time during the operating cycle. As such, pumps analysis is performed through the extended period simulation
are typically switched on during low energy tariff hours (EPS) of the WDN, which is intended as a sequence of steady-
(usually overnight) to fill-up tanks, while limiting their state WDN model runs. During each time step ΔT the model
working during peak energy cost hours. In addition, pumping represents the emergent behavior of the WDN and the
optimization has to account for service requirements like abstraction of steady-state conditions assumes slow time-
preserving the minimum level and avoid water deficit at tanks varying boundary conditions such as demands, level of tanks,
over an operating cycle. For this reason, heuristic approaches, status of valves, working points of pumps, etc., such that the
that normally drive WDN operations, pursue conservative inertial and dynamic effects are considered negligible and not
solutions where pressure at nodes is always guaranteed while included in the WDN model. The EPS allows to model the
water level in tanks is always maximized to preserve WDN changes over time of the required demands (volumes delivered
reliability [3]. Actually, it was demonstrated [4] that such in ΔT), levels of tanks, status of pumps, status of valves, etc.,
pumping strategy exacerbates pressure dependent background which are obtained from the simulation at previous time step.
leakages because it causes pressure increase (due to pumping) The time variable t is here assumed as the starting time of each
during low water requests (i.e. higher pressure) hours. Vice “snapshot” while its ending time is t+ΔT, which is the time
versa, pursuing the minimization of water loss volume besides variable for the successive snapshot. The mathematical
978-1-5386-3943-6/17/$31.00 ©2017 IEEE
formulation of the hydraulic model at time t, for a WDN where: ΔH = dynamic head varying over time t; ωk = speed
composed of np pipes, nn nodes and n0 reservoirs, is based on factor, not varying over time t in this study; Hkpump = static
(np+nn) energy and mass balance equations as in (1) head (elements of Hppump in (1)); rk and ck = parameters of
internal head loss. Moreover, the model account for efficiency
varying pump efficiency as in (3)
  H n (t ) 
  
A Q
 pp p ( t ) + A pn    = − A p 0 H 0 ( t − ΔT ) + H pump
p (t ) 1

  ΔH 0 ( t )  4η 4η  H pump  c
η = − 2max Q 2 + max Q with Qmax =   (3)
 Qmax Qmax  r 
  Vn ( H n ( t ) , t )   
    0n  (1)
  Δ T    where ηmax = maximum pump efficiency. The hydraulic
    
 A Q (t ) −  model herein also allows any kind of devices usually installed
 =  
 np p    into a WDN, such as directional or control valves, parallel

  V H ( t ) , t + Ω ΔH ( t )   V ext ( t )  pipes, flow control valves, etc.
  0( n ) 0 0   0 
  ΔT  ΔT 
    C. Pressure-driven analysis
It is worth noting that the model in (1) entails the pressure-
where, Qp = [np,1] column vector of unknown pipe flow rates driven analysis (PDA), where water outflows Vn(Hn(t),t)
varying over time t; Hn = [nn,1] column vector of unknown represents the superimposition of various pressure-dependent
nodal heads varying over time t; H0 = [n0,1] column vector of water demand components under normal working conditions
known nodal heads of reservoirs varying over time t; Hppump = including [7][8]: orifices controlled by customers;
[np,1] column vector of static heads of pump systems installed free/uncontrolled orifices (e.g. sprinkler working for a given
along pipes (if any) varying over time t when pumps are time interval); and uncontrolled diffuse background leakages
controlled (by states or by tank levels); Vn = [nn,1] column depending on pipe material and pressure. In more details, the
vector of volumes of water withdrawals in the nodes varying background leakage model adopted in this work is that
over time t; Apn = ATnp and Ap0 = topological incidence sub- proposed in Germanopoulos [9], as in (4), although other
matrices of size [np, nn] and [np, n0], respectively, derived from models can be easily introduced
the general topological matrix Āpn=[Apn ¦ Ap0] of size [np,
nn+n0]; AppQp = [np,1] column vector of pipe head losses
containing the terms of internal head losses of pump systems, Vk leaks ( t )  β L Pαk ( t ) Pk , mean ( t ) > 0
=  k k k , mean (4)
minor head losses and evenly distributed head losses. ΔT  0 Pk , mean ( t ) < 0
A. Generalized WDN model for variable level tanks
where Vkleaks = volume of the background leakages along the
In classical WDN hydraulic models, the analysis of tank kth pipe varying over time with mean pipe pressure Pk,mean; αk
level variations is performed in two steps, i.e., the mass
and βk are model parameters mainly depending on pipe
balance at tanks is performed to update the tank levels outside
deterioration and material, whose assessment is part of the
the simulation runs. This might cause instability troubles when
calibration of the model, which is out of the scopes of this
the tanks are close to each other and/or the flow paths among
work. Pressure heads drive the solution (Qp(t), Hn(t), ΔH0(t))
them are characterized by low hydraulic resistances. The
of the system in (1), providing also the outlet water volumes
generalized WDN model [6] was proved to overcome such
(Vn(t)). This modelling approach is different from the classic
instabilities integrating the mass balance at tank nodes and
demand-driven analysis, implemented in few WDN models
their level variations within the hydraulic model formulation,
(e.g. EPANET2), which is based on the assumption of fixed
including also possible external feeding pipes. Each snapshot
demands, i.e. fixed outlet volumes. Indeed, such assumption
at time t in (1) is arranged to have a mass balance equation for
would not allow reliable analyses for WDN operational
each tank node among the unknowns, ΔH0(t) = H0(t) - H0ini(t)
purposes like pump scheduling, where pressure are likely to
= H0(t) - H0(t-ΔT), representing the variation of tank levels
affect each water demand components and mass balance at
during ΔT. Ω0 is a column vector of the cross sectional areas
tanks driving pumps.
of cylindrical tanks. Then, the generalized WDN model
represents a steady-state model of a WDN with varying tank III. OPTIMAL PUMP SCHEDULING
volumes Ω0ΔH0(t) in ΔT.
Pumping optimization can be performed by setting optimal
B. Pumps level in tanks controlling the pump status or by scheduling
The curve of a pump installed on the kth pipe is optimal ON/OFF status over time. The control by time is the
implemented in the model using equation (2) mostly adopted because it is easily referable to the time
pattern of energy tariffs, thus facilitating the assessment of
pumping costs. Nonetheless pump scheduling by time are
  Qk ( t ) 
ck
 unavoidably affected by the assumptions about time varying
Δ H ( t ) = ±ωk2 ( t )  H kpump − rk    (2)
demands that are adopted in the WDN model and are not

  ωk ( t )  
 flexible with respect to abnormal increase/reduction of water
outflows. For this reason this paper account for pump
scheduling by setting the optimal water levels in tanks including the constraints as a conflictual objective function
controlling each pump. makes the pump scheduling a multi-objective optimization
problem with the twofold advantage of promoting the
A. Pumping optimization problem exploration of the search space without neglecting potentially
Following the discussion in the introduction section, the good solutions, and the possibility to get multiple solutions
pumping optimization should pursue the minimization of representing different trade-offs between total operational cost
energy costs for pumping as well as the minimization of and technically viable alternatives [4].
background leakage volume over an operating cycle, which
generally spans over one-day or one-week. In addition, the B. Optimization strategy
optimization problem should match technical constraints From an optimization standpoint the problem of
entailing supply service requirements (i.e. minimum pressures optimizing the level of tanks controlling pumps requires a
for a correct service at demand nodes), reliability (i.e. significant computational burden since the hydraulic system
minimum level to be preserved at tanks), water overflows (i.e. behavior needs to be predicted for all the states of an operating
maximum level of tanks) and deficit in global mass balance in cycle and the number of possible operational solutions is large
each tank at the end of the operating cycle. The pumping even for few pumps and tanks. For this reason, various
optimization problem is formulated herein as in (5): optimization techniques have been applied to the operational
optimization problem, including linear programming,
nonlinear programming, dynamic programming, fuzzy logic,
 PDA model (EPS) ∀t = 0, ΔT , 2 × ΔT ,  , T × ΔT nonlinear heuristic optimization, flexible constraint

min { f1 (Tariff , Energy ,Water Losses )} satisfaction and genetic algorithms. Genetic algorithms (GA)
 was proved to not require simplifications in the problem
min { f 2 ( Constraints )} formulation, having a significant advantage in finding a near
with (5) global optimal solution compared to other optimization
T × ΔT m np T
methods [4].
f1 =   E ( t ) × ΔT × T ( t ) + C V
pu f w k
leaks
(P
k , mean (t ) , t )
t = 0 pu =1 k =1 t = 0
TABLE I. CODING OF DECISION VARIABLES IN GA
f 2 = (1 + Vdeficit )(1 + nPdeficit ) (1 + nVmin )(1 + nVmax )
Device Pump 1 … Pump m
Controlling
where PDA model (EPS) stands for the generalized WDN Tank x … Tank s
tank
model in (1) for each t = 0, ΔT, …, T×ΔT with T = number of Water level
switch ON switch OFF

switch ON switch OFF
EPS model snapshots of the operating cycle; ΔT = time below above below above
GA string 2 4 … 3 5.5
interval of the real hydraulic system snapshot in hours. Indeed,
water losses are generally larger during night as the leakages TABLE II. EXAMPLE OF CANDIDATE VALUES OF DECISION VARIABLES
increase with pressure, while the pumping cost is generally
lower during the night as the tariff is lower. Thus, the main Pump Candidate water level in controlling tanks
driver of both phenomena are the lower requests of water and 1 0 0.25 0.5 0.75 1 1.25 1.50 1.75 2

energy by customers. Therefore, PDA as in (1) is mandatory m 0 0.5 1 1.5 2 2.5 3
to correctly consider the volume of water losses due to
leakages in the WDN.
Table 1 exemplifies the coding of decision variables (i.e.
The objective function f2 includes the technical constraints water level in controlling tanks) in the GA. Table 2 reports an
in terms of: Vdeficit = sum of volume deficit at tanks at the end example of candidate values for each decision variable. It is
of the operating cycle; nPdeficit= number of simulation worth noting that continuous coding (real numbers) of the
snapshots showing some nodes with pressure lower that the levels is possible, but the advantage of the discrete coding
pressure for correct service Piser; nVmin number of snapshot the (integer numbers) is the search space reduction and the
water level in tanks falls below the minimum allowed; nVmax robustness of the solution with respect to quick level
number of snapshot the water level in tanks exceed the variations (provided that a good step size is selected) and
maximum allowed (overflow). The objective function f2 measurement errors. Although the selection of the tank
includes the electric power Epu (in kWh) as in (6), depending controlling each pump can be also part of the solution, it is
on pump efficiency η as computed in (3): known a priori based on WDN hydraulic scheme. In addition,
the same tank can control multiple pumps, each with a set of
9.806 ⋅ Δ H pu ( t ) Q pu ( t ) control levels. The same strategy also applies to other control
E pu ( t ) = (6) devices like controlled gate valves.
η
IV. CASE STUDY
pu = subscript of the puth pump; m = number of pumps; and Tf
= tariff cost/kWh varying over time t; Cw = cost of the non- The C-Town WDN, as in the Battle for Background
revenue water per cubic meter, it accounts for the both the Leakage Assessment in Water Networks [5], is used herein to
value of water and the environmental and indirect costs related demonstrate the pumping optimization strategy. Network data
to water losses; k = subscript of the kth pipe. As such, function and modelling details are reported in the referenced work,
f2 is expected to increase as f1 decreases. The approach of
while Fig. 1 shows network layout where the icon of pumps
(“PM”) are filled with the same color of the controlling tanks.

Figure 1. C-Town layout

Figure 2. Energy cost pattern

According to the original problem statement, it is assumed


that water losses cost 2 €/m3, which includes the value of
water as well as the environmental and indirect costs due to
leakages. Fig. 2 shows the pattern of energy cost over a one-
week (i.e. 168 hours) long operating cycle. From hydraulic
standpoint optimal pumping should guarantee at least 20m of
pressure at nodes with non-null customers’ demand, none of
the tanks should empty during the operating cycle and the
water levels in tanks at the end of the operating cycle should
be larger than the initial one. Differently from the original
problem statement, it is assumed herein that no pipes are
replaced/closed, no pressure reduction valves are installed and
Figure 3. Performances of pumping strategies: (A-B-C-D) on annual
the original volumes of tanks is not changed. Fig. 3 shows the basis; (E) on one-week operating cycle
performance of some solutions of the optimization problem in
(5), sorted in descending order by total annual cost (Fig.1. A). Indeed, moving from solution n.1 to solution n.24 is
Although energy cost (Fig.1. B) and pumping energy (Fig.1. possible to save more than 30,000 m3 on annual basis, saving
C) show the same decreasing trend as the total cost, solution also money for its abstraction, treatment and pumping, before
with similar costs (e.g. 11, 12 and 13) might result into reaching the WDN. Nonetheless, not all the solutions reported
different pumping energy consumption. This is due to the in Fig.1 are actually feasible. In fact, although all of them
variability of energy costs during the one-week operating match the constraint on pressure for a correct service and do
cycle, which results into similar solutions entailing quite not violate the constraint on minimum and maximum level in
different pumping policies. Fig.1 (D) shows that pumping the tanks, they show some volume deficit at the end of the
optimization, besides saving energy (and money) to run operating cycle (Fig.1 E). In more details, the black bars
pumps, allows to save large water volumes. represents solutions with pressure deficit exceeding 35m3 on
weekly basis, while all the others solutions results into an ACKNOWLEDGMENTS
average deficit lower than 5m3 per day, which is acceptable This research was partly funded by “Fondo di Sviluppo e
from technical point of view, considering the conservative Coesione 2007-2013 – APQ Ricerca Regione Puglia
assumptions usually adopted on customers’ demands and “Programma regionale a sostegno della specializzazione
pressure for correct service. Based on the analysis of results, intelligente e della sostenibilità sociale ed ambientale -
solution n. 19 (in red in Fig. 3) could be recommended since it FutureInResearch”.
entails the minimum cost for energy and the minimum volume
of non-revenue water with technically negligible volume
deficit at tanks. Fig. 4 reports the pattern of pump status and REFERENCES
relevant power over 168 hour of operating cycle. It is worth [1] P.W. Jowitt, and G. Germanopoulos, “Optimal pump scheduling in
noting that higher pumping energy are not necessarily spent at water-supply networks,” J. Water Res. Plan. Manage., vol. 118, pp.
minimum energy cost hours due to the simultaneous 406-422, 1992.
[2] K.E. Lansey, and K. Awumah, “Optimal pump operations considering
minimization of water losses. All the test are performed using pump switches,” J. Water Res. Plan. Manage., vol. 120, pp. 17-35,
the WDNetXL [10] Management module, which integrate the 1994.
pressure-driven simulation of all water demand components, [3] J.M. Wagner, U. Shamir, and D.H. Marks, “Water distribution
permitting also the explicit modelling of background leakages reliability: simulation methods,” J. Water Resour. Plan. Manage., vol.
as diffuse outflows from pipes and the generalized WDN 114, pp. 276-294, 1988
[4] O. Giustolisi, D. Laucelli, L. Berardi, “Operational optimization: Water
model. In particular, this module allows the optimal pumping Losses vs. Energy Costs”, J. of Hydraul. Eng., ASCE. 139(4), pp. 410–
based on either scheduling pump status at each ΔT or 423, 2013.
identifying tank levels controlling pumps. [5] O. Giustolisi, L. Berardi, D. Laucelli, D. Savic, T. Walski, B. Brunone,
“Battle of Background Leakage Assessment for Water Networks
V. CONCLUDING REMARKS (BBLAWN) at WDSA Conference 2014”, Procedia Eng., vol. 89, pp.
4-12, 2014.
The complexity of pumping optimization problem stems [6] O. Giustolisi, L. Berardi, and D.B. Laucelli, “Generalizing WDN
from the need of matching correct service requirements while simulation models to variable tank levels,” J. Hydroinf., vol. 12, pp.
minimizing the overall operational cost, which should account 562-573, March 2012.
for both costs of energy and indirect costs of water losses. [7] O. Giustolisi, and T.M. Walski, “A Demand Components in Water
Distribution Network Analysis,” J. Water Resour. Plan. Manage., vol.
This contribution demonstrate the application of a pumping 138, pp. 356 -367, April 2012.
optimization strategy where the decision variables are the [8] O. Giustolisi, D.A. Savic, and Z. Kapelan, “Pressure-driven demand
water level in tanks controlling pump status in a complex and leakage simulation for water distribution networks,” J. Hydraul.
WDN. The proposed approach permits to analyze and Eng., vol. 134, pp. 626-635, 2008.
compare various technically feasible solutions entailing [9] G. Germanopoulos, “A technical note on the inclusion of pressure
dependent demand and leakage terms in water supply network models,”
different energy costs for pumping and volumes of water Civil Eng. Syst., vol. 2, pp. 171–179, 1985.
losses. The integration of an advanced WDN hydraulic model [10] O. Giustolisi, D. Savic, L. Berardi, D. Laucelli, “An Excel-based
including pressure-driven modeling of customers’ water solution to bring water distribution network analysis closer to users”, in
demand, background leakages and robust simulation of water Proc. 2011 Computer and Control in Water Industry (CCWI) Conf.,
level at tank is of key importance to get reliable decision September 5-7, Exeter, UK, D.A..Savic, Z. Kapelan, D. Butler (Eds)
vol. 3, pp. 805-810.
support for water utilities.

Figure 4. Pumping pattern for Solution n.19

Vous aimerez peut-être aussi