Career · 2017
Risk Optimized Microgrid Resource Management
Explore the four cases ↓My master’s thesis studied how a distribution network could move from centrally dispatched resources toward distributed participation. Local controllers scheduled dispatchable generation, storage and flexible demand; a master controller coordinated those schedules through nodal power flow and prices.
I built the model in GAMS, with MATLAB for input preparation and analysis, and studied an IEEE 33-bus network at 0%, 33%, 66% and 100% local participation. The model combined network and device constraints with conditional value at risk to account for unreliable delivery.
The interesting result was a limit as well as an opportunity: local flexibility changed prices and dispatch across the network, but the fully participating case alternated between solutions instead of converging. The explorer below uses the archived results to make that behavior, and the architecture behind it, easier to inspect.
Research explorer · saved GAMS results
A network of decisions.
What changes when the edge of the grid starts making its own dispatch decisions? Follow four stages of participation through the same 33-bus network.
One coordinator, no participating local resources.
The master controller still dispatches generation, storage and flexible load. Local controllers contribute their base demand, but none have participating distributed resources. Zero participation does not mean an empty grid.
IEEE 33-bus distribution feeder
Final saved schedule · hour 17- Line 30 · buses 30–31
- 500 kW
- Bus 31 voltage
- 0.992 p.u.
- Local dispatch price
- 0.095 $/kWh
Toward the substation · 100% of line limit
1.000 is nominal voltage
Price received by this local controller
Even this baseline has congestion. At the evening peak, the outer branches carry power back toward the feeder. Read the case ↗
Two parts of the feedback loop
Bus 31, across the day.
The local controller receives a price and returns a schedule. The master then solves the network again and produces new nodal prices. These are different signals, at different steps of the same round.
Hour 17 · the resulting nodal price at bus 31 is $0.095/kWh.
How the model works
Coordinate the network. Keep device decisions local.
Local controllers
Each controller schedules its own devices against the prices it receives: dispatchable generation, flexible load, battery and EV constraints. The master sees net demand or injection, rather than every device behind the meter.
Published results, model boundaries and source
The transition is between levels of distributed participation, not an electrical transient. The study uses day-ahead schedules and an iterative optimization loop; it is not a time-domain stability simulation.
Case inputs also change the master controller’s resource fleet to retain feasibility and congestion. These four cases are not a controlled sweep with every other variable held constant.
| Participation | MC, risk | LC, risk | MC, no risk | LC, no risk |
|---|---|---|---|---|
| 0% | 0.0279 | 0.0276 | 0.0280 | 0.0276 |
| 33% | 0.0208 | 0.0339 | 0.0208 | 0.0268 |
| 66% | 0.0241 | 0.0384 | 0.0239 | 0.0263 |
| 100% | 0.0213 | 0.0435 | 0.0211 | 0.0258 |
Risk adjustment keeps the master-controller results close to the runs without risk, while changing local-controller costs. At 100% participation, termination at the iteration limit must not be confused with convergence.
The interactive figures replay the archived risk-adjusted GAMS runs at their saved precision. Lines show final real-power flow; node voltage comes from the linearized network solution. The local-price trace is the price received at the start of the final round; the nodal-price trace is the master’s resulting price. Arrows encode direction; motion is illustrative, not electricity’s propagation speed. No optimizer runs in your browser.
The original model also includes reactive power, device energy and operating constraints, and CVaR-based reliability adjustments. Its linearized power-flow approximation and stopping behavior limit what can be concluded from these runs.
Read the full thesis (PDF, 5.2 MB) ↗ · University of Denver record ↗