1 of 19

NSF IRES INNOVATOR Summer 2024 Final Presentation

Tyler May, Rehadyan Utomo

2 of 19

Overview

  • Project Problem and Solution

​

  • Methodology and Design

​

  • Experimentation

​

  • Results

3 of 19

Fault Ride Through

  • Fault Ride Through (FRT) is a crucial capability required from most power systems connected to the grid
  • Discourages false tripping for inconsequential disturbances
  • IEEE 1547 provides standards on Distributed Energy Resources (DERs) grid connections and FRT

4 of 19

Inverter DC Voltage PI Controller

  • Using a simulated PV system with grid connection
  • Inverter controller uses Proportional-Integrator (PI) for control
  • Static terms are inefficient for enforced-energization cases
    • IEEE 1547 can require system remain online even at 70% nominal voltage

5 of 19

Solution Proposal

Use an Artificial Neural Network (ANN) to dynamically adjust the PI gains for optimal continuous power flow during long-term disturbance situations

​

6 of 19

AI-Assisted PI Control for FRT

  • The idea is to use machine learning and have the controller predict best gain settings outside of nominal conditions

​

  • An automatic-simulation program was created to generate specific datasets for training and testing

​

  • The ANN will be provided with the grid Per Unit voltage, and the output will simply be the PI gain changes from the nominal
    • Nominal: P = 7; I = 800

7 of 19

Simulation Preparation

  • Fault values were studied with the system to determine the voltage drop cases for testing

​

Variable P with 16e6 Q

Variable Q with zero P

90% ratio between P and Q

​

P

QL

P

QL

P

QL

95%

23 MW

16 MVAR

0 MW

32 MVAR

36 MW

4 MVAR

90%

73.5 MW

16 MVAR

0 MW

68 MVAR

78.75 MW

8.75 MVAR

85%

135.5 MW

16 MVAR

0 MW

113.5 MVAR

135 MW

15 MVAR

80%

222 MW

16 MVAR

0 MW

175 MVAR

220.5 MW

24.5 MVAR

75%

361 MW

16 MVAR

0 MW

270 MVAR

355.5 MW

39.5 MVAR

70%

641 MW

16 MVAR

0 MW

450 MVAR

648 MW

72 MVAR

65%

1615 MW

16 MVAR

0 MW

1025 MVAR

1620 MW

180 MVAR

8 of 19

Optimal Gains

The optimal gains were found by rerunning the simulation with various configurations to limit the output current and stabilize the input voltage to the nominal 500v

9 of 19

Proportional Testing

  • The proportional term of the PI controller influences system response. The best terms found were reducing the gain for the sake of the current

10 of 19

Integrator Testing

  • The integrator term of the PI controller stabilizes the system’s steady state
  • The system influence was minimal compared to the Proportional term, and therefore can be safely kept at 800 for both nominal and reasonable abnormal conditions
    • Values tested only as low as 70% Per Unit according to IEEE 1547

11 of 19

Optimal Gain Data

  • Initially created changes in both Kp and Ki, but simplifying our NN meant keeping a constant Ki and variable Kp.
  • Additionally, we observed the overshoot error in voltage, steady state in voltage, but prioritized peak 3-phase current in order to preserve equipment health
  • Limited voltage to 5% change in steady-state, grid peak current to 10% of 320 A.

*If there was more time, I’d have liked to write a script that can minimize a cost function instead of qualitatively finding optimal gains

12 of 19

Automatic Dataset Generation

  • No publicly available datasets fit the project’s needs, and so a MATLAB script was created to make a CSV file of these Per Unit values, both from a genuine simulation and random values for training
  • The label CSV files were created similarly

13 of 19

Google Collab

  • Used Collab to create a simple sequential Keras network from TensorFlow
  • The input are 0-1 Per Unit values, and the output is a factor by which to multiply into the nominal gains

14 of 19

Results

  • The trained Neural Network was provided with multiple 3 second sequences of voltage Per Unit matrices and appropriately gave dynamic gain changes according to how deep the sag is
  • The output is a factor to reduce the Proportional term in accordance with the optimal gains found
  • As shown, the nominal case remains untouched

15 of 19

EXPERIENCES IN AAU

  • APES Research Seminar
  • PhD Defense, Yuan Li
  • Energy Storage Systems Course
  • Danfoss Drive Tour

16 of 19

Exposure to the Field

The state of machine learning applications in power electronics. Such as it’s benefits and drawbacks compared to traditional control systems. EVs, batteries, grid-connections, and renewable energy.

Research currently being done in the advancements of controllers. “Tailored Advancements in Model Predictive Control of Power Converters” Yuan Li’s PhD defense.

17 of 19

Appropriate applications for different renewable energies, and energy storage systems.

18 of 19

Research Seminar

Motor Drive Manufacturing

Walked us through the process from assembling components onto a PCB, using large machinery to physically and electronically test.

Many concepts out-of-scope but a great walkthrough of today’s advancements in technology. How to best apply model predictive control; inner-workings of a multiport DC-DC converter.

19 of 19

THANK YOU

To everyone AAU for their never-ending hospitality

A special thanks to Dr. Mateja and Dr. Shuai for the mentorship and guidance during this project.

Further thanks to the National Science Foundation for the funding of this experience and for the chance to participate.