IEEE Power & Energy Magazine - July/August 2021 - 46

The approach we developed and demonstrated in this project
explores a key feature of a distribution-level market: how to
optimize the use of participating DERs for network support.
server every 5 min. Network load patterns are forecast
and trigger a DER network support negotiation if they
are outside the network's capacity.
2) The NAC server solves the network subproblem covering
the next 24 hours using load forecasts as input for
nonparticipating customers. From this solution, networkfeasible
real and reactive powers are sent to each customer
as a request alongside the relevant standing LMPs.
3) EMSs solve the customer subproblems by optimizing
their DER use in response to the LMPs and support
power requested by the NAC cloud server. Their best
response is sent back to the NAC server, where the LMPs
are then updated to reflect the changes to the network
use. The negotiation iterates between stages 2 and
3 until the algorithm identifies convergence to a level of
support where the network's constraints are satisfied and
customers are happy with the offered LMPs.
In case convergence cannot be reached for whatever reason,
the system is designed to fall back on the solution from
the most recent successful negotiation. Alternatively, such
situations can be flagged with the network operator for more
manual intervention, e.g., in circumstances where a network
fault has led to an infeasibility. Currently, there is no known
theoretical convergence guarantee when using the alternating
direction method of multipliers to solve a nonconvex
problem, such as OPF. However, in practice, convergence
has not been an issue, and we still expect to get solutions
near to maximizing the social welfare of the customers and
network. This is backed up by experiments that use the alternating
direction method of multipliers for solving OPF in
similar settings.
NAC on Bruny Island
Within the context of Bruny Island, NAC was responsible
for negotiating the dispatch of the 34 battery systems to
ensure the undersea cable stayed below its current limit. The
OPF subproblem takes as input a forecast for the island load
(generated from recloser readings) and a model of the 11-kV
unbalanced three-phase medium voltage (MV) network.
The network OPF then iteratively negotiates further supply
from the batteries or, alternatively, the diesel generator, until
the undersea cable is safely below its limit.
DNSPs typically have poor-quality data on network configuration
and state, and Bruny Island was no exception. The
NAC model used in the trial had an unbalanced three-phase
model of the network and required the following:
46
ieee power & energy magazine
✔ the network configuration, such as which phase a customer
is connected to, conductor type, and configuration
of transformers
✔ the current state of the network, including the voltage
at each bus, current of each line, and load at each
customer site.
Due to the unbalanced nature of the network and loads
connected to it, it was particularly important to have good
information about the phasing of customers. What phases
a participant connects to would alter how much his or her
DER could contribute to managing the network constraint,
depending on the imbalance of the network during the peak
event. The three-phase model we used as part of NAC could
account for these differences when determining the optimal
dispatch, and it also had implications for how participants
should be rewarded. Phasing data had to be manually collected
as part of the trial by TasNetworks operators.
One limitation of the available data was the lack of an
accurate low-voltage (LV) network model, so, instead, customers
were approximated as being connected directly to the
MV network after accounting for the customer and distribution
transformer phasing. The thermal limit of the undersea
cable was the key operational constraint in our demonstration.
We did some simulations where MV network voltages
were limited as a proxy for LV voltages, but those results
could not be properly evaluated due to the missing LV model.
Forecasting the island load is critical to ensuring enough
forward preparation of the batteries, i.e., that they charge in
a period of lower demand in the lead-up to a peak. However,
load forecasting for small loads, such as Bruny Island,
is challenging, and it is compounded by the impact of public
holidays and the weather. It is also of key importance to predict
a peak with enough forewarning so that the batteries can
precharge in anticipation. A neural network-based load-forecasting
engine was implemented to significantly improve the
performance of the overall approach.
Network Benefits
A series of trials was conducted in 2018 to demonstrate
and evaluate the NAC approach. In total, NAC was run for
65 days, which covered 16 peak periods that required network
support. As discussed earlier, the traditional approach
to managing peak loads on the undersea cable connecting
Bruny Island to the rest of the NEM was dispatching a local
diesel generator, so a key metric would be the savings in diesel
consumption. We also utilized counterfactual simulations
july/august 2021
IEEE Power & Energy Magazine - July/August 2021

Table of Contents for the Digital Edition of IEEE Power & Energy Magazine - July/August 2021
