Awesome

Copyright © 2023, Georgia Institute of Technology. All rights reserved.

ml4acopf benchmark

Machine Learning for AC Optimal Power Flow(ACOPF) benchmark for VNNCOMP

Environment Setup

To set up the environment, follow these steps:

1. Create the environment using the command: conda env create -f env.yaml
2. Activate the environment using the command: conda activate onnx-vnnlib-env

The benchmark files are located in the onnx and vnnlib folders:

To reproduce vnnlib files, run: python generate_properties.py

Vnnlib description

Input:

+- a% perturbation of the reference active and reactive load + random noise between -b% and b%, where a and b are self-defined values.

Output:

Check the properties of the NN output

For example, we are interested in if power balance violation is within some threshold:

for each bus i

• |p_balance[i]| <= max(10^(-3), 10^(-2)*pd_i)
• |q_balance[i]| <= max(10^(-3), 10^(-2)*qd_i)

where pd_i and qd_i refer to the active and reactive load at each bus respectively.

Inference

The code to run inference is presented in the main.py file.

Example

Take 14-bus system as an example:

• onnx file: 14_ieee_ml4acopf.onnx
• vnnlib file: 14_ieee_prop1.vnnlib

Parameters

• N = 14 : number of buses
• G = 5 : number of generators
• L = 11 : number of loads
• E = 20 : number of lines

Onnx model description

Input: pd/qd

dim: (2L) = 22

• pd: Real power demand. (L)
• qd: Reactive power demand. (L)

Output: pg/qg/vm/va/pf/pt/qf/qt/thrm_1/thrm_2/p_balance/q_balance

dim: (2G + 4N + 6E) = 186

• pg: Real power generation. (G)
• qg: Reactive power generation. (G)
• vm: Voltage magnitude. (N)
• va: Voltage angle. (N)
• pf: Real power flow from. (E)
• pt: Real power flow to. (E)
• qf: Reactive power flow from. (E)
• qt: Reactive power flow to. (E)
• thrm_1: Thermal limit from residual. (E)
• thrm_2: Thermal limit to residual. (E)
• p_balance: Real power balance residual. (N)
• q_balance: Reactive power balance residual. (N)

Compute graph