a novel hierarchical frequency based load control architecture
TRANSCRIPT
A Novel Hierarchical Frequency-Based Load Control Architecture
Takashi NishikawaNorthwestern University (lead)
Network Optimized Distributed Energy Systems (NODES)Annual Review Meeting
February 12-13, 2019
Project Summary
1
‣ Overall goal of the project: To develop and validate a new, hierarchically coordinated, frequency-based load control architecture that provides a reliable complement to generator inertia and governor response so as to enable a high penetration of renewable generation.
Project Summary
2
‣ Challenge: To control a large amount of appliances distributed in the whole system in a coordinated manner while simultaneously achieving the objectives of guaranteeing system stability, ensuring fast frequency regulation, and maintaining customers’ QoS.
‣ Deliverables:
1. Scalable algorithms for frequency-load control.2. Validation of the algorithms using real power-grid network
models and datasets via large-scale real-time simulations and HiL testing.
‣ Expected outcome: Demonstration that the developed control architecture can provide additional synthetic frequency response reserve to the existing grids, which would enable higher penetration of renewable generation.
Team
3
‣ Partners– Northwestern University (NU): Leader in research on complex network control
and frequency-synchronization phenomena.– Argonne National Laboratory (ANL) has extensive expertise/experience in
modeling, analysis, and control of power systems.– OPAL-RT is a leader in providing real-time simulation/HiL environment.– Schneider Electric (SE) has experience/expertise in hardware/software
development, providing microgrid solutions, and commercialization.– CPS Energy is a utility serving San Antonio area.– Lawrence Berkeley National Laboratory (LBNL) has facility & experience for
HiL testing.‣ NU designed control algorithms with help from ANL on power system modeling. ‣ OPAL is helping to perform real-time simulations and HiL validation.‣ CPS provided distribution network data.‣ SE is conducting a commercialization review.‣ LBNL is helping with HiL testing of the developed control algorithms.
Project Progress
4
1.2: network decomposition method
Year 1 Year 2 Year 3Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12
2.1: ext. depth-L method
1.1: performance test on optimization algorithms
2.2: distribution network control
3.1: appliance model development
3.2: design decentralized control
3.3: design building EMC control
Task 1
Task 3
Task 2
Project Progress
5
4.2: power syst. topo. & control modeling
Year 1 Year 2 Year 3Q1 Q2 Q3 Q4 Q5 Q6 Q7 Q8 Q9 Q10 Q11 Q12
5.1: Definition of HiL test plan
4.1: large-scale simulation scenarios and tools
5.2: HiL modeling & interfaces
Task 4
Task 5
4.3: validation
5.3: testing
Image removed to comply with NDAImage removed to comply with NDA
Hierarchical control architecture
6
Transmission gridcontroller
Distribution gridcontroller Building controller
Image removed to comply with NDA
Transmission Grid Optimization Controller
7
‣ Task 1 (completed): Load assignment algorithms for transmission system stability
ERCOT system
Reformulation of λmax optimization into an easier-to-solve OPF problem
Highly optimized commercial optimization software (Knitro)
+
Optimization of the ERCOT demo completes in 5–10 min and yields 10–15% improvement in λmax.
Sequential Nonlinear Programming (SNLP) Method:
Linearized power system DAE model:
Optimization problem: minimize the spectrum abscissa
where
Smooth NLP problemNon-smooth problem
Transmission Grid Optimization Controller
8
Sequential Nonlinear Programming (SNLP) Method:
Linearized power system DAE model:
Optimization problem: minimize the spectrum abscissa
where
Smooth NLP problemNon-smooth problem
Sequential Nonlinear Programming (SNLP) Method:
Linearized power system DAE model:
Optimization problem: minimize the spectrum abscissa
where
Smooth NLP problemNon-smooth problem
Sequential Nonlinear Programming (SNLP) Method:
Linearized power system DAE model:
Optimization problem: minimize the spectrum abscissa
where
Smooth NLP problemNon-smooth problem
(swing eq. + power flow)
Sequential Nonlinear Programming (SNLP) Method:
Linearized power system DAE model:
Optimization problem: minimize the spectrum abscissa
where
Smooth NLP problemNon-smooth problem
Transmission Grid Optimization Controller
9
Run power flow, initialize 𝑥𝑥 = 𝑥𝑥0;Set upper and lower limits of controllable load;
Set 𝜏𝜏 = 𝜏𝜏0, 𝑡𝑡𝑡𝑡𝑡𝑡 = 10−4, 𝑘𝑘 = 1.
Calculate eigenvalues and their derivatives of 𝐽𝐽(𝑥𝑥0)Set 𝑓𝑓0 = max
𝑖𝑖𝑅𝑅𝑅𝑅{𝜆𝜆𝑖𝑖(𝐽𝐽(𝑥𝑥0))}
𝜏𝜏 < 𝑡𝑡𝑡𝑡𝑡𝑡 ?
Solve the NLP problem with 𝑥𝑥∗ = 𝑥𝑥𝑘𝑘−1 with KNITRO to obtain 𝑦𝑦∗, perform eigen decomposition on 𝐽𝐽(𝑦𝑦∗)
max𝑖𝑖𝑅𝑅𝑅𝑅{𝜆𝜆𝑖𝑖(𝐽𝐽(𝑦𝑦∗))} < 𝑓𝑓𝑘𝑘−1?
Sequential Nonlinear Programming (SNLP) Method:
output 𝑥𝑥∗ = 𝑥𝑥𝑘𝑘−1 and stopYes
No
set 𝑥𝑥𝑘𝑘 = 𝑦𝑦∗, 𝑓𝑓𝑘𝑘 = max𝑖𝑖𝑅𝑅𝑅𝑅{𝜆𝜆𝑖𝑖(𝐽𝐽(𝑦𝑦∗))},
𝜏𝜏 ← 𝜏𝜏 × 2, 𝑘𝑘 ← 𝑘𝑘 + 1
Find the least integer 𝑛𝑛 such that max𝑖𝑖𝑅𝑅𝑅𝑅{𝜆𝜆𝑖𝑖(𝐽𝐽(𝑥𝑥∗ +
12𝑛𝑛 (𝑦𝑦
∗ − 𝑥𝑥∗)))} < 𝑓𝑓𝑘𝑘−1. Set 𝜏𝜏 ← 𝜏𝜏 1
2𝑛𝑛
No
Yes
• Maintain a trust region with radius 𝝉𝝉;• Use a gradient bundle;• Guarantee local optimality.
Distribution Grid Controller
10
‣ Task 2 (completed): Depth-L observability controlDistributed control of each node using signals from only the depth-L state information neighborhood (SIN) determined by the participation matrix.
Walzem distribution network used for testingData provided by CPS Energy
, ,
, ,1 1
trace( ).
trace( )c p o q
pq Nm Nnc p o qp q
W W
W Wj
= =
=å åParticipation matrix:
Image removed to comply with NDA
Distribution Grid Controller
11
Distribution Control
Transmission control gives rise to P ⇤0 , Q
⇤0 at the point of common
coupling
Distribution side active and reactive powers: P0, Q0
Error signals dp = P ⇤0 � P0 and dq = Q⇤
0 �Q0 sent to each buildingwith controllable loads
New active power target for the building:
P ⇤l
= Pl
+ wl
(P ⇤0 � P0) (9)
where wl
=
|Pl.dp|PLi=1|Pl.dp|
NW ARPAE Project Evanston, January 25, 2018
Current implementation
Control performance for Walzem network
Distribution Control
Transmission control gives rise to P ⇤0 , Q
⇤0 at the point of common
coupling
Distribution side active and reactive powers: P0, Q0
Error signals dp = P ⇤0 � P0 and dq = Q⇤
0 �Q0 sent to each buildingwith controllable loads
New active power target for the building:
P ⇤l
= Pl
+ wl
(P ⇤0 � P0) (9)
where wl
=
|Pl.dp|PLi=1|Pl.dp|
NW ARPAE Project Evanston, January 25, 2018
Distribution Control
Transmission control gives rise to P ⇤0 , Q
⇤0 at the point of common
coupling
Distribution side active and reactive powers: P0, Q0
Error signals dp = P ⇤0 � P0 and dq = Q⇤
0 �Q0 sent to each buildingwith controllable loads
New active power target for the building:
P ⇤l
= Pl
+ wl
(P ⇤0 � P0) (9)
where wl
=
|Pl.dp|PLi=1|Pl.dp|
NW ARPAE Project Evanston, January 25, 2018
(PCC), after PI controller loop for fast frequency regulation
Building Level Controller
12
Type I: On/off loads
Type II: Continuously adjustable loads
Development of Control Algorithms for Flexible
Appliances in Buildings
§ Building Level Energy Management Control (EMC) Design– Level II: Optimal Dispatch of Controllable Loads in Each Individual Building – Based on the desired demand generated in Level I, the following optimization
problem is formulated to dispatch Type I and Type II loads in each building
6
,, ,
,ˆmax
j i ji j i j i j
x p j j
q x P pt
t t t+ × +å å, ,ˆs.t. i j i j i j i
j j
q x P pt t t tq+ × + £å å,min ,max
, , , , ,i j i j i j i j i jp p pt t tx x× £ £ ×
{ }0,1jx Î
{ }, 0,1i jx Î
Maximize load power in each building
Total load power should not exceed the desired demand
Type II loads should not violate their power limits
On/off status of Type I loads
On/off status of Type II loads
‣ Task 3 (>90% complete): Building Level Energy Management Control (EMC) Design
– Based on the desired demand generated by distribution grid controller, the following optimization problem is solved:
Building Level Controller
13
Development of Control Algorithms for Flexible
Appliances in Buildings
§ Building Level Energy Management Control (EMC) Design
8
LLUP: local load update phase LTUP: load target update phaseIUP: intrusion update phaseACP: admissible control phase
LLUP LTUP ACP IUP
Summarize local fixed and flexible loads
Update local load target
Coordinate local controllable loads: formulation I
Coordinate local controllable loads: formulation II
‣ Task 3 (>90% complete): Building Level Energy Management Control (EMC) Design
Building Level Controller
14
Development of Control Algorithms for Flexible
Appliances in Buildings
§ Building Level Energy Management Control (EMC) Design– Test case: EMC covering 20�20 local buildings– Each building includes 36 Type I loads and 4 Type II loads– Use 24-hour data to validate the EMC design
9
Time (24-hour)
Desir
ed an
d Ac
tual
Load
Pow
er (k
W)
Time (24-hour)Re
lativ
e Er
ror (
%)
Limitation: 5%
Development of Control Algorithms for Flexible
Appliances in Buildings
§ Building Level Energy Management Control (EMC) Design– Test case: EMC covering 20�20 local buildings– Each building includes 36 Type I loads and 4 Type II loads– Use 24-hour data to validate the EMC design
9
Time (24-hour)
Desir
ed an
d Ac
tual
Load
Pow
er (k
W)
Time (24-hour)
Rela
tive
Erro
r (%
)
Limitation: 5%Desired load power (kW)Actual load power (kW)
Relative error (%)5% target
Time (24-hour) Time (24-hour)
Image removed to comply with NDA
Image removed to comply with NDA
Validation: Large-Scale Simulations
15
‣ Task 4 (~80% complete): Large-scale simulations– Using OPAL-RT ePHASORsim platform
– Transmission network modeling: Full Texas ERCOT model based on FERC 2017 data
– Distribution network modeling: Based on data from CPS Energy
– Building level modeling: Flexible appliance models developed in Task 3
Transmission Model Validation
16
• Power flow from ePHASORsim was compared with the PSS/E power flow results for ERCOT 2018 Summer Peak data
• Voltages at all the buses and currents on selected transmission lines were compared
• Deviations were within 0.5% for voltage and current magnitudes and 2 degrees for angles
Max. Voltage Magnitude Deviation 0.178% Max. Voltage Angle Deviation 0.039ºMax. Current Magnitude Deviation 0.1784%Max. Current Angle Deviation 0.208º
Distribution Model Validation
17
50% PV Penetration:• PV penetration level of 50% was accomplished by adding 776 ECGs in the
distribution feeder
• PV/ECG nodes were selected such that the substation power demand reduces by 50% during peak PV profile
• Results from ePHASORsim were compared with the CYMEdist load flow results• Bus voltages at all the nodes and currents at selected distribution lines
were compared
• Deviations were within 0.5% for voltage and current magnitudes and 2 degrees for angles
Max. Voltage Magnitude Deviation 0.0562% Max. Voltage Angle Deviation 0.048ºMax. Current Magnitude Deviation 0.1035%Max. Current Angle Deviation 0.735º
Building Load Model Validation
18
‣ Flexible appliance models (HVAC systems and battery-based loads) were validated in Task 3.1.
-+Pg*
Pg
PI-+
id*
id
PIvd'
PWM
ToDC/AC
Converter++
decoupling
vd*
-+Qg*
Qg
PI-+
id*
id
PIvd'
++
decoupling
vq*
dqtoabc
GridVoltageAngle
vd*vq*
Bi-directionalDC/DC
Converter
Li-ionBattery
DCLink
DC/ACConverter
GRID-
+
iL VdcPg,Qgid,iq
vd*,vq*
fg
fg,Vg
f
P
VgQ
V
Power-Frequencydroop
Volt-VarCurveForVoltageRegulation
Validation: Large-Scale Simulations
19
Replace with Tasks_4_5_validation_procedure.pdf(2 pages)
TG DG BL
Step 1: Test performance criteria
Step 2: Identify the aggregate response model for BL and DG.
Step 3: Integrate the DG and BL model and run the simulation for the whole operation cycle with given loadand renewable generation profile. Assess the system-wide performance.
ercot cps flexlab1). Reserve Magnitude Target (RMT) (> 2%)2). Reserve Magnitude Variability Tolerance (RMVT) (< +/-5%)3). Duration (> 30s)4). Availability (> 95%)5). Initial Response Time (< 2s)6). Ramp time (< 8s)7). Contingency support (loss of renewables)8). Small-signal stability (improve 𝜆𝑚𝑎𝑥)9). Frequency deviation (< +/-0.5Hz) Device transients;
Controller dynamics;
Optimization dynamics in the loop.
BL: Flex LabDG model
Overall Procedure (to be used for HiL tests also)
Validation: Large-Scale Simulations
20
BL Test: Building controller (BC) + LBNL devices (with or without PV)
DG Test: CPS + BC+LBNL
TX Test: ERCOT+ CPS + BC + LBNL
• Apply step change to target power of the whole building, and observe the overall response;• Compute the performance metrics;• Identify the overall response model of the building.
• Apply step change to target power of the distribution system at PCC, and observe the response;• Compute the performance metrics;• Identify the overall response model of the whole distribution system.
• Run the whole hierarchical model for the entire operation cycle;• Assess the performance the of three level controllers;• Apply contingency (sudden loss of renewables), and observe the system response.
Majority of the performance targets were met in 1st round of large-scale simulations.
Validation: HiL Tests
21
‣ Task 5 (~60% complete): HiL testingTesting sites & schedule:
– LBNL FLEXLAB(Phase 1: Jan ’19, Phase 2: Mar/Apr ’19)
• 100 controllable hardware load devices: HVAC fans/pumps (6), blowers (60), heaters (31), PV/inverter/battery systems (3)
– Stone Edge Farm (SEF) Microgrid (May ’19)• ~20 controllable hardware devices: gas turbine (1),
electrolyzer (1), fuel cells (1), batteries (9), PV (2), motors (3), HVAC (1)
Validation: HiL Tests
22
Building 90
PG & E (utility)
Rotational testbedTB90XR-ATB90XR-B
Testbed #1TB90X1-ATB90X1-B
Testbed #3TB90X3-ATB90X3-B
Testbed #2TB90X2-ATB90X2-B
PV system3 inverters3 batteries
Cell A
Cell B
uPMU
uPMU
uPMU
HVAC(2 fans, 4 pumps)
SSR/Heaters(30)
VFD/Motors(10 x 6 racks)
Powermeters
SM_ERCOT
TX_SOLVER(transmission)
TX_CONTROL
Hostcomputer
DG_SOLVER(distribution)
DG_CONTROL
BL_CONTROL(ctrl. BLs/inv.)
BL_OPT
TX_OPT solver
NU server
MILP solver
SS_WALZEM
SC_CONSOLE
Opal-RT system
SSHTCP/IP
Power measurements (NI)
Phasor measurement(C37.118)
PQ targets (API, HTTP)
Power targets (NI)
On/off status targets(DO)
On/off status targets(AO)
LBNL/FLEXLAB/FLEXGRID
2 progmmableload banks
OP5600
LBNL_BL_Model
LBNL_BL_Battery
HIL Test at LBNL FLEXLAB
Project Challenges
25
‣ Challenges overcome:– Proposed network decomposition method was not
adequate when scaled up à reformulation of optimization problem + commercial optimization software
– Completely re-planned HiL tests with new sites– Validated communication links to ~100 devices via
multiple protocols at LBNL FLEXLAB‣ Remaining challenges:
– Complete 2nd round large-scale simulations– Minimize communication delays and complete LBNL
FLEXLAB HiL test– Set up & complete SEF Microgrid test
Publications
26
‣ Y. Yang, T. Nishikawa, and A.E. Motter, Vulnerability and Co-susceptibility Determine the Size of Network Cascades, Phys. Rev. Lett. 118, 048301 (2017).
‣ Software for simulation of physical cascade model and identification of cosusceptibile groups, https://github.com/yangyangangela/determine_cascade_sizes
‣ Y. Yang, T. Nishikawa, and A.E. Motter, Small vulnerable sets determine large network cascades in power grids, Science 358, eaan3184 (2017).
‣ Y. Yang and A.E. Motter, Cascading failures as continuous phase-spacetransitions, Phys. Rev. Lett. 119, 248302 (2017).
‣ A. Haber, F. Molnar, and A.E. Motter, Global network control from localinformation, to be published.
‣ Publications being prepared on:– State information neighborhood and nonlinear localized control– LBNL FLEXLAB HiL test configuration and results
Tech-to-Market Path
27
‣ Commercial objectives: License/patent developed control
algorithms
‣ Target markets/segments: Building managers, Microgrid
operators, ISOs
‣ IAB members:– Sunny Elebua, Vice President, Corporate Strategy, Exelon Corporation
– Troy Nergaard, Director of Technical Product Management, Doosan GridTech
– Dave Kaplan, Chief Technology Officer, Doosan GridTech
– Holly Benz, Senior Vice President, Commercial & Industrial, CLEAResult
– Jean Bélanger, Chief Executive Officer & Chief Technology Officer, OPAL-RT
Technologies
– Mark Feasel, Vice President, Electric Utility Segment & Smart Grid, Schneider
Electric
– Shaneshia McNair, Manager, Customer Engineering, CPS Energy
Tech-to-Market Path Update
28
‣ Tasks completed:– Defined process for selecting target market segment (“beachhead”),
and value proposition.– Reviewed technology reports, and identified potential value
propositions, target market segments, and channel access to target market segment.
– Completed team discussions on value propositions for residential, and medium to large commercial building segments, and reviewed competing offers addressing similar value propositions to identify solution differentiations.
– Concluded that demand response type of solution is best positioned to deliver a strongly differentiated solution to load aggregators and large commercial customers.
– Completed team discussion on product definition.
Remaining Tasks
29
‣ Task 3: Customize building level controller for SEF testing‣ Task 4: Complete the 2nd round large-scale simulation‣ Task 5: Complete HiL tests at LBNL FLEXLAB & SEF
Microgrid‣ Task 6 (Tech-to-Market tasks):
– Review demand response programs to confirm value proposition and level of competitive differentiation.
– Define whole product solution– Complete two more rounds of feedback from IAB– Identify and interview commercial building customers participating in
demand response programs to confirm value proposition.