Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Energy Storage and Renewables in New Jersey: Complementary Technologies for Reducing
Our Carbon Footprint
ACEE E-filliates workshop November 14, 2014
Warren B. Powell Daniel Steingart Harvey Cheng Greg Davies
© 2014 Warren B. Powell, Princeton University
Frequency regulation PJM sends charge/discharge signals to generators every 2 seconds to smooth out frequency/voltage variations
1 hour
MW
Solar from PSE&G solar farms Solar from a single solar farm
Solar energy from a single solar farm 1 week
Energy from wind
1 year
Wind power from all PJM wind farms
Jan Feb March April May June July Aug Sept Oct Nov Dec
Energy from wind
30 days
Wind from all PJM wind farms
Wind energy in PJM Total load vs. total current wind (January)
Winter load and solar Total PJM load plus factored solar (January)
Wind energy in PJM Total PJM load plus actual wind (July)
Summer load and wind Total PJM load plus actual wind (July)
99.9 percent from renewables!
Fossil backup
Battery storage
Wind & solar
}20 GW
Lead-acid or lithium-ion
Ultracapacitors
Research challenges How do we control a battery storage system? Challenges include: » Managing a single storage device to handle multiple revenue
streams, over multiple time scales » Controlling a storage system in the presence of a multidimensional
“state of the world” » Controlling dozens to hundreds of storage devices spread around
the grid.
How do we design storage systems? » What type of storage device(s)? » How many are needed? » How should they be distributed across the grid?
How does storage change the economics of renewables?
Revenue streams Frequency regulation Power quality management Battery arbitrage Energy shifting Demand peak management - Many utilities impose charges based on peak usage over a month, quarter or even a year. Peak management for avoiding capacity expansion Backup power for outages
seconds
minutes
hours
days-weeks
weeks-months
months-years
Research goals To design an algorithm that produces near-optimal policies that handle the following problem characteristics: » Responds to predictable time-dependent structural patterns over
hourly, daily and weekly cycles in generation and loads. » Able to simultaneously optimize over multiple revenue streams,
balanced against maximizing the lifetime of the battery. » Able to handle time scales ranging from seconds to minutes, hours
and days. » Handles uncertainty in energy generation, prices and loads. » Handles “state of the world” variables such as weather conditions,
network conditions and prices. » For some applications, we need to scale to large numbers (tens to
hundreds, but perhaps thousands) of grid-level storage devices. » Ability to incorporate forecasts of wind or solar energy, loads, and
weather. » Needs to be computationally very fast.
A storage problem Energy storage with stochastic prices, supplies and demands.
windtE
gridtP
tD
batterytR
1 1
1 1
1 1
1
ˆ
ˆ
ˆ
wind wind windt t t
grid grid gridt t t
load load loadt t t
battery batteryt t t
E E E
P P P
D D D
R R x
+ +
+ +
+ +
+
= +
= +
= +
= +
1 Exogenous inputstW + =State variabletS =
Controllable inputstx =
A storage problem Bellman’s optimality equation
( )1 1( ) min ( , ) ( ( , , )tt x t t t t t tV S C S x V S S x Wγ∈ + += +X E
windtgrid
tloadtbatteryt
E
P
D
R
wind batterytwind loadtgrid batterytgrid loadtbattery loadt
x
x
x
x
x
−
−
−
−
−
1
1
1
ˆ
ˆ
ˆ
windt
gridt
loadt
E
P
D
+
+
+
Managing a water reservoir Backward dynamic programming in one dimension
1 1 1Step 0: Initialize ( ) 0 for 0,1,...,100Step 1: Step backward , 1, 2,... Step 2: Loop over 0,1,...,100 Step 3: Loop over all decisions 0 Step 4:
T T T
t
t t
V R Rt T T T
Rx R
+ + += == − −=
≤ ≤
{ }100
max1
0
Take the expectation over all rainfall levels (also discretized):
Compute ( , ) ( , ) (min , ) ( )
End step 4; End Step 3;
Wt t t t t t
wQ R x C R x V R R x w P w+
=
= + − +∑
*
* Find ( ) max ( , )
Store ( ) arg max ( , ). (This is our policy)
End Step 2;End Step 1;
t
t
t t x t t
t t x t t
V R Q R x
X R Q R xπ
=
=
Managing cash in a mutual fund
( )
1 1Step 0: Initialize ( ) 0 for all states.Step 1: Step backward , 1, 2,... Step 2: Loop over = , , , (four loops) Step 3: Loop over all decisions (a problem if i
T T
t t t t t
t t
V St T T T
S R D p Ex x
+ + == − −
( )
1 1 1
s a vector)ˆ ˆˆ Step 4: Take the expectation over each random dimension , ,
Compute ( , ) ( , )
, , ( ,
t t t
t t t t
Mt t t t
D p E
Q S x C S x
V S S x W w+ +
= +
=( )( )1 2 3
*
100 100 100
2 3 1 2 30 0 0
*
, ) ( , , )
End step 4; End Step 3; Find ( ) max ( , )
Store ( ) arg max ( , ). (This is our policy)
End St
t
t
W
w w w
t t x t t
t t x t t
w w P w w w
V S Q S x
X S Q S xπ
= = =
=
=
∑ ∑ ∑
ep 2;End Step 1;
Dynamic programming in multiple dimensions
Approximate dynamic programming Bellman’s optimality equation » We approximate the value of energy in storage:
( )( )( ) min ( , ) ( , )t
x xt t x t t t t t tV S C S x V S S xγ∈= +X
Inventory held over from previous time period
Approximate dynamic programming We update the piecewise linear value functions by computing estimates of slopes using a backward pass: » The cost along the marginal path is the derivative of the simulation
with respect to the flow perturbation.
Rδ
Approximate dynamic programming Testing on a deterministic problem demonstrates that we can precisely capture optimal time-dependent behavior:
Approximate dynamic programming Benchmarking against optimality on a stochastic model
0
20
40
60
80
100
120
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Perc
ent o
f opt
imal
Pe
rcen
t of o
ptim
al
Storage problem
Approximate dynamic programming Benchmarking against optimality on a stochastic model
95
96
97
98
99
100
101
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Perc
ent o
f opt
imal
Pe
rcen
t of o
ptim
al
Storage problem
Slide 25
Grid level storage
Slide 26
Grid level storage control
Slide 27
Grid level storage control Monday
Time :05 :10 :15 :20
Slide 28
Grid level storage control Monday
Time :05 :10 :15 :20
Slide 29
Grid level storage control Monday
Time :05 :10 :15 :20
Grid level storage control Approximate dynamic programming (blue) vs. optimal using linear programming (green)
Heterogeneous fleets of batteries
Heterogeneous fleets of batteries A tale of two batteries » Ultracapacitor – High power, high efficiency, low capacity » Lead acid – Lower power, lower efficiency, high capacity
Lead acid
Ultra capacitor
Heterogeneous fleets of batteries Control algorithm adapts to characteristics of each storage device
Time (hours) energy is held in storage device
Cum
ulat
ive
dist
ribut
ion
Handling multiple time scales
0 . . . 1 2 3 3 4 23 24 Daily (hourly increments)
0 5 10 15 20 55 60 . . . Hourly (5-min. increments) ( )1( ) min ( , ) (
tt x t t tV S C S x V Sγ∈ += +X E
( )1( ) min ( , ) (tt x t t tV S C S x V Sγ∈ += +X E
Handling multiple time scales
0 . . . 1 2 3 3 4 23 24
0 5 10 15 20 55 60 . . .
Daily (hourly increments)
Hourly (5-min. increments)
( )1( ) min ( , ) (tt x t t tV S C S x V Sγ∈ += +X E
( )1( ) min ( , ) (tt x t t tV S C S x V Sγ∈ += +X E
Handling multiple time scales
0 . . . 1 2 3 3 4 23 24
0 5 10 15 20 55 60 . . .
Daily (hourly increments)
Hourly (5-min. increments)
There are 43,200 2-second increments in a day, over 300,000 in a week.
( )1( ) min ( , ) (tt x t t tV S C S x V Sγ∈ += +X E
0 2 4 6 8 298 300 . . . 5 min (2-sec. increments) ( )1( ) min ( , ) (
tt x t t tV S C S x V Sγ∈ += +X E
( )1( ) min ( , ) (tt x t t tV S C S x V Sγ∈ += +X E
Solar energy Princeton solar array
What is the value of storage in managing the variability from renewables?
Our model, “SMART-Storage” simultaneously optimizes the ramping of generators as well as storage.
Solar-storage experiments
Load
Storage level
Solar intensity LMP
Sunny day
Cloudy day
Mixed sun/cloud
Mixed sun/cloud
Solar
Energy from storage
Solar = 2.3GW Storage = 12Gwh/600MW
Solar = 23GW Storage = 12Gwh/600-MW
Nuclear Steam
CC Gas turbine Pumped storage
Sunn
y da
y C
loud
y da
y
Solar-storage experiments
23MW 230MW 2.3GW 11.5GW 23GW
23MW 230MW 2.3GW 11.5GW 23GW
Solar capacity
Solar capacity
Load covered by solar %
% of solar used
Solar-storage experiments Some conclusions: » The model will only put energy in storage when storage is the only
way to meet fast variations in generation and loads. » The reason is the losses that are incurred when converting energy
is stored. It is always better to ramp down a generator during periods of high energy generation from wind or solar, than to store the energy and use it better.
» The idea that the conversion losses do not matter when the energy is free is a myth…. It only applies when the total generation from renewables exceeds the total load (which was never the case in our experiments).
Research goals To design an algorithm that produces near-optimal policies that handle the following problem characteristics: » Responds to predictable time-dependent structural patterns over
hourly, daily and weekly cycles in generation and loads. » Able to simultaneously optimize over multiple revenue streams,
balanced against maximizing the lifetime of the battery. » Able to handle time scales ranging from seconds to minutes, hours
and days. » Handles uncertainty in energy generation, prices and loads. » Handles “state of the world” variables such as weather conditions,
network conditions and prices. » For some applications, we need to scale to large numbers (tens to
hundreds, but perhaps thousands) of grid-level storage devices. » Ability to incorporate forecasts of wind or solar energy, loads, and
weather. » Needs to be computationally very fast.
Thank you!
For more information see:
http://energysystems.princeton.edu