Download as pdf or txt
Download as pdf or txt
You are on page 1of 32

ECSE 6140 Power Generation and Control

Unit Commitment

Prof. Joe H. Chow

Rensselaer Polytechnic Institute


ECSE Department

September 8, 2021
Operation of Large Systems of Supply and Demand
Power systems operate very different from many other systems
Telephone – past: analog wired connections limited by trunk
capacity; present: digital, spectrum/bandwidth, fiber optics,
wireless transmission
Gasoline stations – delivery trucks, gas tanks
Natural gas – pipelines and storage tanks
Electric power
Available 24/7
Cannot be stored economically in medium or large quantities;
pumped hydro units are large storage systems, but impossible to
build now; advances in battery technology: fuel cells, Lithium ion, ...
Electric energy supply mostly comes from large central generating
stations through transmission systems at various voltage levels;
renewable resources, however, are growing at a rapid pace;
distributed generation resources (DERs) expensive but valued for
providing resilience after blackouts
J.
c H. Chow UC September 8, 2021 2 / 32
Load Cycles
Large steam generators cannot be turned on or off arbitrarily and
instantaneously; requires hours to generate enough steam to start
generation (startup cost)
Daily load cycle: loads vary throughout the day, but in some
predictable manner; increasingly impacted by renewable resources
NYISO Load - Sept 22, 2011
24

23 Forecast
Real Time
22

21

20
Load (GW)

19

18

17

16

15

14
0 5 10 15 20 25
Time (hrs)

J.
c H. Chow UC September 8, 2021 3 / 32
CAISO daily load profile

J.
c H. Chow UC September 8, 2021 4 / 32
Bid Load and Unit Commitment Process
Forecasted load: load day types: M–F, Sa, Su, holidays; historical data;
neural-network-based tools
Bid load: amount of loads bid into the day-ahead market (DAM) by
the load serving entities (LSEs); NY has 11 zones, each zone has many
buses (nodes); load bids are for each nodes; LSE bids in the loads on
behalf of you and me.
Process of matching load and generation:
Each morning by 5 am (like Monday), LSEs enter hourly load bids
on the NYISO website for the nodes for the 24 hours of the next
day (like midnight Monday to midnight Tuesday)
Generators also provide hourly supply bids on the NYISO website,
with prices (including start up), generation limits, and ramp limits
At 5 am, the NYISO DAM engineer will assemble the load and
generation bids and execute a security-constrained
unit-commitment (SCUC) program. By 11 am, the DAM is cleared
and the hourly generation setpoints issued to the generators being
dispatched.
J.
c H. Chow
A generator only UC
knows its own setpoints.
September 8, 2021 5 / 32
Unit Commitment Problem
1 Given a collection of generators (various types: steam, gas, hydros;
various sizes) and the forecasted load profile for the next day, UC
determines which of these generators should be turned on and for
how much at each hour.
2 We assume that the total bid-in generation > peak load. Some
control regions make sure it is true using an installed capacity
market (ICAP). ICAP is an auction. Example: Gen 1, rated 200
MVA, bids 2$/MW for the first 6 months of the year to supply 190
MW to the system. If accepted, Gen 1 will get this much money
to bid in everyday at least 190 MW into the market.
3 In SCUC, line loading limits from off-line transient and voltage
stability N − 1 contingency analysis are enforced (hence the name
“security constrained”)
4 Unit commitment also needs to include reserve considerations

J.
c H. Chow UC September 8, 2021 6 / 32
Operating Reserves (in NYISO)
From time to time, a power system will experience a generation
unit trip. Thus a system needs to have some generating capacity
on standby to make up for this loss.
Operating reserve is a one of the ancillary services that is market
based, i.e., reserve bids are part of the generator bids at 5 am.
The reserve market is set to 1.5 times the capacity of the largest
single unit (1200 MW) in NY
The 1800 MW is separated into 4 products (NYISO Ancillary
Service Manual) – the amount of reserve is seasonal and varies by
hours
10-minute spinning reserve
10-minute non-synchronous reserve
30-minute spinning reserve
30-minute non-synchronous reserve
Other special reserves: Beacon Flywheel in Stephentown
When called onto service, reserves are paid for energy (positive or
negative) and for movement
J.
c H. Chow UC September 8, 2021 7 / 32
Location Reserves
Suppose the flow P12 is at its limits. If Area 2 loses a gen unit of 1000
MW, reserve in Area 1 would not be helpful – the line will be
overloaded. Thus one needs reserve in both Areas 1 and 2.

J.
c H. Chow UC September 8, 2021 8 / 32
NYISO locational Reserve Requirement Table

J.
c H. Chow UC September 8, 2021 9 / 32
Co-optimization of Dispatch and Reserve
Two components:
For a reliable power market, needs to calculate dispatch and
reserve simultaneously in DAM – called co-optimization (mention
CAISO market fiasco)
A generator is allowed to bid into the energy and reserve market
simultaneously. For example, a hydro unit rated at 300 MW can
simultaneously bid 300 MW into the energy market and 30 MW in
spinning reserve.
Nature of problem as defined by generator characteristics

J.
c H. Chow UC September 8, 2021 10 / 32
Daily Load Curves – Generator Stacking
An ideal UC situation without network congestion consideration

(load curves in Japan)


Hot summer days – all base units on line; gas turbines on during peak
load (high energy prices)
Spring or fall days – not all base units online (some may be on
maintenance for two to four weeks); gas turbines are mostly offline

J.
c H. Chow UC September 8, 2021 11 / 32
Constraints in UC
Thermal unit constraints: minimum up time, minimum down
time; ramp rate, startup cost, minimum generation cost
Limited energy units - finite amount of fuel available
Must-run units: (1) for voltage support in an urban load center
and (2) co-generation unit supplying steam to city population
(e.g., NYC Linden co-gen); must-run units do not set price (more
later on deregulated markets)
Hydraulic units: (1) amount of water in reservoir, (2) seasonal
variations – more water in spring, less in summer, (3) pumped
hydro (Blenheim-Gilboa, NY, Northfield Mountain, MA) – limited
capacity, (4) run of water – succession of dams on the same river
(Columbia River, WA, Tennessee River, TVA) – has to generate,
otherwise water will stop flowing (like for irrigation, salmon
season)
Flow constraints on transfer paths – contingency based
J.
c H. Chow UC September 8, 2021 12 / 32
UC Methods

1 Priority list
2 Dynamic programming
3 Lagrange relaxation (LR)
4 Mixed integer programming (MIP) – not covered in this course,
requiring too much background stuff

J.
c H. Chow UC September 8, 2021 13 / 32
Priority List

See earlier daily load cycle figure


Select the least expensive steam units for base load
Select the least expensive cycling units to meet the morning load
rise
Dispatch the gas turbines if peak load is very high
Make sure all ramp rate, energy constraints, etc., are satified

J.
c H. Chow UC September 8, 2021 14 / 32
Dynamic Programming (DP)

An improvement of priority list by limiting feasible cases to study


Use forward DP – go forward on hourly basis; enumeration
forward takes into account ramp rate, up time, and down time
Even with DP efficiency, still a cumbersome amount of work
Sometimes constraints on going forward may affect solutions in
previous stages; thus keep some less optimal solutions as options,
and not thrown away until later

J.
c H. Chow UC September 8, 2021 15 / 32
Lagrange Relaxation (LR)
Used by several EMS software vendors (ABB and GE Grid Solution
(Alstom/Areva/ESCA)) before being gradually replaced by MIP.
D. P. Bertsekas, G. S. Lauer, N. R. Sandell, and T. A. Posbergh, “Optimal
Short-Term Scheduling of Large-Scale Power System,” IEEE Trans.
Automatic Control, vol. AC-28, no. 1 pp. 1–11, Jan. 1983.
Primal Problem:
J = min f (x) subject to g(x) ≤ 0
x
where x is a vector of unknown variables to be optimized, f is a scalar
function, and g is a vector function.
Define the Lagrangian: L(x, λ) = f (x) + λT g(x)
Optimality conditions:
∂L ∂L
= 0, =0
∂x ∂λ
The second equation is needed if any of the g(x) = 0 (binding
constraint).
J.
c H. Chow UC September 8, 2021 16 / 32
Dual Problem
Define a dual function q(λ) as
q(λ) = min L(x, λ)
x

which reads as “for each fixed λ, find x to minimize q.”


Dual problem:
q ∗ = max q(λ)
λ≥0
Note that some λs may be zero as some constraints may not be binding.
Let the optimal value of the primal problem be J ∗ = minx∗ f (x∗ ). We
can show that J ∗ ≥ q ∗ (because g ≤ 0).
Duality gap: J ∗ − q ∗ ; relative duality gap: (J ∗ − q ∗ )/q ∗
For convex problems (if both f and g are convex, J ∗ − q ∗ = 0)
In general, the dual problem is easier to solve. As the end of the
optimization process, one can compute the relative duality gap:
(J − q)/q
J.
c H. Chow UC September 8, 2021 17 / 32
UC Problem Formulation
Given N generation units, each with power output Pi and cost function
Fi (Pi ), i = 1, 2, ..., N
Uit = 0 if unit i is off-line during period t, and 1 if unit i is on-line
during period t
The optimal UC objective function is
T X
X N
F (P, U ) = (Fi (Pit ) + Cit )Uit
t=i i=1

where Citis the start-up cost of unit i at time t. Cit is zero if unit i is
online in period t − 1. That is, it is only paid when it is first started.
System operating constraints
1. Generation and load balance (assuming no losses)
N
X
t
Pload − Uit Pit = 0, t = 1, ..., T
i=1

J.
c H. Chow UC September 8, 2021 18 / 32
2. Unit limits and other constraints, for i = 1, ..., N , and t = 1, ..., T
Uit Pimin ≤ Pit ≤ Uit Pimax
Pimin is min gen, and Pit is 0 for off-line units
Ramp rates
∆Pidown max ≤ Pit+1 − Pit ≤ ∆Piup max

Limited energy resources for unit i


T
X
PiT ≤ Ei
i=1
+ other constraints like thus for pumped hydro units
3. Network constraints
Pij ≤ Pij max
which is a line flow limit, a proxy for some kind of stability limits (from
transient and voltage stability studies)
Voltage constraints: at Bus k (used only in AC power flow calcuation)
Vkmin ≤ Vk ≤ Vkmax
J.
c H. Chow UC September 8, 2021 19 / 32
Decision Variables
Define
   
P11 ··· P1t · · · P1T U11 ··· U1t · · · U1T
 1
P2t · · · P2T
 1
 P2 ···  U2 ··· U2t · · · U2T
 
 
P = . , U = .
   
.. .. .. .. .. .. .. ..
 ..  ..

 . . . . 
  . . . . 

PN1 · · · PNt · · · PNT UN1 t
· · · UN T
· · · UN

Lagrange multiplier λ = [λ1 , ..., λt , ..., λT ]


Lagrangian
T
X N
X
L(P, U, λ) = F (P, U ) + λt (Pload
t
− Uit Pit )
t=1 i=1

T N N
" ! !#
X X X
= (Fi (Pit ) + Cit )Uit +λ t t
Pload − Uit Pit
t=1 i=1 i=1

J.
c H. Chow UC September 8, 2021 20 / 32
Instead of sum over time, rearrange the expression to sum over the
units
N T T
!
X X X
t t t t t t
L(P, U, λ) = (Fi (Pi ) + Ci )Ui − λ Ui Pi + λt Pload
t

i=1 t=1 t=1

N
X T
X
= Li (Pit , Uit , λt ) + λt Pload
t

i=1 t=1

Note that with λt fixed, the individual Lagrangians are, for i = 1, ..., N
T
X
Li (Pit , Uit , λt ) = (Fi (Pit ) + Cit )Uit − λt Uit Pit
t=1

With the generator unit constraints for unit i dependent only on Pi ,


the cost function is separable over units.
These units are “coupled” by the load constraints, i.e., at each period
t, total generation = total load.
J.
c H. Chow UC September 8, 2021 21 / 32
T
X
Li (Pit , Uit , λt ) = ((Fi (Pit ) + Cit ) − λt Pit )Uit
t=1

Thus
dFi (Pit )
 
∂Li
= − λ Uit
t
∂Pit dPit
Note: λt is fixed. Thus no partial derivative term involving λt .
Two cases
1 dFi (Pit )/dPit ≤ λt : feasible Pit solution, such that Uit = 1
if dFi (Pit )/dPit < λt : maximum gen, Pit = Pimax
if dFi (Pit )/dPit = λt : marginal unit
2 dFi (Pit )/dPit > λt for any Pit : Gen i not dispatched, energy too
expensive at Pit , such that Uit = 0

J.
c H. Chow UC September 8, 2021 22 / 32
Lagrange Relaxation

Main idea:
1. make a guess at the values of λ = [λ1 , ..., λt , ..., λT ], i.e., the
incremental cost for each time period (hour).
2. Use this set of λ to dispatch each generator i = 1, ..., N throughout
the duration of T time periods.
For each hour t and Gen i
λt too low: Uti = 0
λt too high (like nuclear unit during peak hour): Pi = Pimax , Uti = 1
Else, solve dFi (Pit )/dPit = λt for Pit .
Still need to check if Pit is feasible under the ramp rate constraint:
∆Pidown max ≤ Pit − Pit−1 ≤ ∆Piup max .

J.
c H. Chow UC September 8, 2021 23 / 32
3. After all the generators have been dispatched, for each period t,
check
a. If N Pit > Pload , then λt is too high; so adjust λt to be lower
P
Pi=1
b. If N t t t
i=1 Pi < Pload , then λ is too low; so adjust λ to be higher

4. In the next iteration, adjust λt up or down accordingly, until the


load-generation balance is within a tight tolerance.

J.
c H. Chow UC September 8, 2021 24 / 32
Dual Problem

q ∗ = max q(λ)
λ

where
q(λ) = min L(P, U, λ)
P,U

is separable into minimization for each generator.


Spelling out the dual problem
T N
!
X X
t t
max q(λ) = max F (P, U ) + λ Pload − Pit Uit
λ λ
t=1 i=1

Each λt canP be determined separately as:


t
if Pload > N P t U t , then increase λt (price is not high enough)
i=1 i i
t < N t t t
P
else Pload i=1 Pi Ui then decrease λ (price is too high)

J.
c H. Chow UC September 8, 2021 25 / 32
Adjusting λt
Gradient approach
T N
!
X X
t t
q(λ) = (λ − independent parts) + λ Pload − Pit Uit
t=1 i=1
then
N
dq(λ) t
X
= P load − Pit Uit
dλt
i=1
which can be either positive or negative.
λ adjustment:
dq(λ)
new λt = old λt + αt
dλt
Suggestion by W-W-S
1 if dq(λ)/dλt > 0, use αt = 0.01 (not enough generation)

2 if dq(λ)/dλt < 0, use αt = 0.002 (too much generation)

αt is related to the quadratic terms of the cost function F (Pi )


J.
c H. Chow UC September 8, 2021 26 / 32
Participation Factors

00
At period t, compute second derivative Fi (Pit ) for all generators that
are dispatched. Let
N N  
t
X X 1
∆Pmismatch = Pload − Pit Uit = ∆λt 00
i=1 i=1
Fi (Pit )

Then
new λt = old λt + ∆λt
Extra class activities: check references proposing the use of
participation factors

J.
c H. Chow UC September 8, 2021 27 / 32
Approaching Optimal Solution
Suppose you have a set of λ, s.t.
N
X
Pit Uit − Pload
t
>0
i=1
i.e., all loads served with the units committed (with some extra).
Compute J = F (P, U ) and q = L(P, U, λ), using
Pit , i = 1, ...N, t = 1, ..., T
If we are not at optimal, then J > J ∗ (optimal) and q < q ∗ (optimal).
As we over generate, J will be high. Compute (relative) duality gap
J −q
q
Stop if this value is small. Practicality: $1000 in a $50 million DAM
Note: Even at optimal UC, q ∗ may not be equal to J ∗ . The duality
gap is zero only if the optimization is convex. See Example 4E in
W-W-S 3rd edition (Example 5E in 2nd edition).
J.
c H. Chow UC September 8, 2021 28 / 32
UC with Ramp Rate Constraints

Even though we can determine the power provided by the generators


for each period, we need to make sure that the generators can reach
their generation levels at each period.
Consider a two-period optimization. Suppose UC is performed w/o
ramp rate constraints, but

P22 − P21 > R2 max

Redo the optimization for Gen 2 with P22 − P21 = R2 max . Form the new
Lagrangian, including the ramp rate as a binding constraint with µ > 0

L2 = F2 (P21 ) + F2 (P22 ) − λ1 P21 − λ2 P22 + µ(P22 − P21 − R2 max )

J.
c H. Chow UC September 8, 2021 29 / 32
Optimality conditions
∂L2 dF2 dF2
1 = 1 − λ1 − µ = − (λ1 + µ) = 0
∂P2 dP2 dP21

∂L2 dF2 dF2


2 = 2 − λ2 + µ = − (λ2 − µ) = 0
∂P2 dP2 dP22
∂L2
= P22 − P21 − R2 max = 0
∂µ
Note: µ acts as a price modifier. In period 1, the price is now higher at
λ + µ and in period 2, the price is lower at λ − µ. So the solution is to
dispatch Gen 2 to a higher value at period 1, and a lower value at
period 2.

J.
c H. Chow UC September 8, 2021 30 / 32
Solution:
1. Combine the first 2 optimality conditions to form
∂F2 ∂F2
1 + − (λ1 + λ2 ) = 0
∂P2 ∂P22

2. Substitute P22 from the third optimality condition into dF2 /dP22 to
obtain a single equation for P21 to solve for P21 .
3. From P21 , calculate P22 = P21 + R2 max
4. Then solve for ∆λ1 and ∆λ2 from
t t
∆Pmismatch = Pload − (P11 + P21 )
t
∆Pmismatch
∆λt =
participation factor of Gen 1
as Gen 2 is constrained.

J.
c H. Chow UC September 8, 2021 31 / 32
5. Then
new ∆t = old ∆t + ∆λt
If F1 (P1 ) is a quadratic function of P1 , solution is obtained in one
iteration.

J.
c H. Chow UC September 8, 2021 32 / 32

You might also like