Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improved Physics-Informed Neural Network based AC Power Flow for Distribution Networks

Victor Eeckhout, Hossein Fani, Md Umar Hashmi, Geert Deconinck

TL;DR

A novel physics-informed loss function is developed that can be used to train (deep) neural networks aimed at power flow simulation and shows strong prediction capabilities when tested on scenarios outside the training sample set, something that is a substantial deficiency of model-free techniques.

Abstract

Power flow analysis plays a critical role in the control and operation of power systems. The high computational burden of traditional solution methods led to a shift towards data-driven approaches, exploiting the availability of digital metering data. However, data-driven approaches, such as deep learning, have not yet won the trust of operators as they are agnostic to the underlying physical model and have poor performances in regimes with limited observability. To address these challenges, this paper proposes a new, physics-informed model. More specifically, a novel physics-informed loss function is developed that can be used to train (deep) neural networks aimed at power flow simulation. The loss function is not only based on the theoretical AC power flow equations that govern the problem but also incorporates real physical line losses, resulting in higher loss accuracy and increased learning potential. The proposed model is used to train a Graph Neural Network (GNN) and is evaluated on a small 3-bus test case both against another physics-informed GNN that does not incorporate physical losses and against a model-free technique. The validation results show that the proposed model outperforms the conventional physics-informed network on all used performance metrics. Even more interesting is that the model shows strong prediction capabilities when tested on scenarios outside the training sample set, something that is a substantial deficiency of model-free techniques.

Improved Physics-Informed Neural Network based AC Power Flow for Distribution Networks

TL;DR

A novel physics-informed loss function is developed that can be used to train (deep) neural networks aimed at power flow simulation and shows strong prediction capabilities when tested on scenarios outside the training sample set, something that is a substantial deficiency of model-free techniques.

Abstract

Power flow analysis plays a critical role in the control and operation of power systems. The high computational burden of traditional solution methods led to a shift towards data-driven approaches, exploiting the availability of digital metering data. However, data-driven approaches, such as deep learning, have not yet won the trust of operators as they are agnostic to the underlying physical model and have poor performances in regimes with limited observability. To address these challenges, this paper proposes a new, physics-informed model. More specifically, a novel physics-informed loss function is developed that can be used to train (deep) neural networks aimed at power flow simulation. The loss function is not only based on the theoretical AC power flow equations that govern the problem but also incorporates real physical line losses, resulting in higher loss accuracy and increased learning potential. The proposed model is used to train a Graph Neural Network (GNN) and is evaluated on a small 3-bus test case both against another physics-informed GNN that does not incorporate physical losses and against a model-free technique. The validation results show that the proposed model outperforms the conventional physics-informed network on all used performance metrics. Even more interesting is that the model shows strong prediction capabilities when tested on scenarios outside the training sample set, something that is a substantial deficiency of model-free techniques.
Paper Structure (12 sections, 9 equations, 4 figures, 5 tables)

This paper contains 12 sections, 9 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Contributions of the paper
  3. Proposed loss function
  4. AC Power Flow
  5. Proposed physics-informed loss function
  6. Performance metrics used
  7. Numerical evaluation
  8. Network under review: 3-bus radial feeder
  9. Scenarios for training and testing ACPF models
  10. Out of sample scenarios for ACPF testing
  11. Numerical results
  12. Conclusion and future work

Figures (4)

  • Figure 1: 3-bus network
  • Figure 2: Approximated power imbalance at bus 1
  • Figure 3: Voltage profiles for 1 day of the test data comparing models: model-free XGB, and physics-informed GNNb and GNNp.
  • Figure 4: Comparing predicted and true voltage values. The green zone shows the voltage levels lower than $V_{\max}$ and not seen in the training dataset, while, the red zone shows the out-of-sample scenarios with overvoltage instances.