[Paper Review] Towards Understanding the Unreasonable Effectiveness of Learning AC-OPF Solutions

My H. Dinh, Ferdinando Fioretto+2 more|arXiv (Cornell University)|Nov 22, 2021

Optimal Power Flow Distribution4 citations

TL;DR

This paper investigates why deep neural networks (DNNs) can accurately approximate AC-Optimal Power Flow (AC-OPF) solutions and proposes a novel recurrent neural network (RNN)-based model that leverages solution trajectories from iterative solvers to improve prediction robustness and reduce memory usage. The RNN model achieves up to a 100x reduction in load flow errors and optimality gaps compared to prior DNN methods, while scaling efficiently to large systems.

ABSTRACT

Optimal Power Flow (OPF) is a fundamental problem in power systems. It is computationally challenging and a recent line of research has proposed the use of Deep Neural Networks (DNNs) to find OPF approximations at vastly reduced runtimes when compared to those obtained by classical optimization methods. While these works show encouraging results in terms of accuracy and runtime, little is known on why these models can predict OPF solutions accurately, as well as about their robustness. This paper provides a step forward to address this knowledge gap. The paper connects the volatility of the outputs of the generators to the ability of a learning model to approximate them, it sheds light on the characteristics affecting the DNN models to learn good predictors, and it proposes a new model that exploits the observations made by this paper to produce accurate and robust OPF predictions.

Motivation & Objective

• To understand why DNNs achieve high accuracy in approximating AC-OPF solutions despite the problem's non-convex and nonlinear nature.
• To investigate the factors affecting prediction robustness, especially the volatility of generator output trajectories.
• To identify the role of physical and engineering constraints in enabling accurate DNN predictions.
• To develop a scalable, robust DNN model for AC-OPF prediction that reduces memory footprint and improves constraint satisfaction.

Proposed method

• The method uses deep autoregressive Long Short-Term Memory (LSTM) recurrent neural networks (RNNs) to model sequential generator output predictions across solver iterations.
• Each RNN unit conditions its prediction on the previous unit’s output and the system’s load demands, enabling progressive refinement of the solution.
• A weighted loss function is applied, with √t scaling to emphasize later time steps, and a Lagrangian constraint loss is used to promote feasibility of the final prediction.
• The model is trained on solution trajectories generated by iterative nonlinear solvers like IPOPT, which provide intermediate solution states during convergence.
• The final prediction is the output of the last RNN unit, which is optimized to minimize both prediction error and constraint violation.
• The approach avoids large fully-connected layers, enabling scalability to large systems like RTE-2868 with 2868 buses.

Experimental results

Research questions

• RQ1 Why are DNNs able to approximate AC-OPF solutions with low error, despite the problem’s non-convexity?
• RQ2 How does the volatility of generator output trajectories affect the ability of a DNN to predict them accurately?
• RQ3 What role do physical and engineering constraints of the OPF problem play in enabling or limiting DNN prediction performance?
• RQ4 Can a sequence-based deep learning model improve robustness and scalability over existing DNN-based OPF predictors?

Key findings

• The paper finds that many generators exhibit linear-like behavior with respect to load, making them easy to predict and contributing to overall low DNN error rates.
• Generators with highly volatile output trajectories are significantly harder to predict, and standard DNNs struggle with such cases, leading to robustness issues.
• The proposed RNN model reduces load flow errors and optimality gaps by one order of magnitude compared to the best existing DNN method (FCC), achieving 0.007 on IEEE-118 and 0.04 on IEEE-300.
• The RNN model achieves superior constraint satisfaction, particularly for KCL (Kirchhoff's Current Law), which is notoriously difficult for DNNs to enforce.
• The RNN model has a significantly smaller memory footprint than fully-connected DNNs, enabling it to scale to large systems like RTE-2868 (2868 buses), which are infeasible for prior DNNs.
• The RNN model reduces the number of parameters by orders of magnitude compared to FCC models, avoiding memory overflow issues seen in large systems.