Comparative Analysis of Generative and Deterministic Deep Learning Models for Fast Power Flow Prediction

The reliable operation of modern electrical power grids depends on timely power flow analysis. This task involves calculating steady-state voltages and power flows throughout a network under specific operating conditions. While established iterative numerical methods offer high accuracy in practice, their computational complexity poses challenges for large-scale networks. Furthermore, the rising demand and structural complexity of the network, alongside new consumers such as data centers and other computational loads, motivate the investigation of faster, alternative computational frameworks such as learning-based models. This study compares deterministic architectures, in the sense that during the inference phase the mapping from input to output is entirely deterministic, and generative machine learning architectures to establish the most effective statistical approach for this prediction task. The investigation evaluates multilayer perceptrons, conditional generative adversarial networks, and conditional variational autoencoders using topology-aware grid representations. Quantitative results on unseen test configurations demonstrate that the multilayer perceptron achieves a root mean square error of 0.076875 and a Kolmogorov–Smirnov distance of 0.115846. The conditional generative adversarial network yields a mean absolute error of 0.031281 and a mean absolute percentage error of 314.573. In contrast, the conditional variational autoencoder results in a mean absolute error of 0.141746 and a root mean square error of 0.234046. Ultimately, this comparison reveals that deterministic architectures excel at minimizing large pointwise errors through mean-reverting estimations, whereas generative adversarial networks are necessary to preserve the high-frequency statistical variance and true distributions of power grid dynamics. Additional external validation using real-world PJM operational measurements further demonstrated stable predictive behavior under unseen operational dynamics despite training solely on synthetic GridOPT scenarios.

M.Sc. student under the supervision of Prof. Yoash Levron.