DeepTL4SE: Deep Transfer Learning for Power System State Estimation via Physics-Informed Data Generation

Abstract

Power system state estimation (PSSE) requires accurate and timely inference from noisy measurements, but large labeled operational datasets are often unavailable and deployment data may differ from offline training data. This paper presents Deep Transfer Learning for State Estimation (DeepTL4SE), which combines Physics-Informed Data Generation (PIDG) with supervised transfer learning. PIDG evaluates the AC power-flow relations for selected voltage states and constructs paired measurements and state labels directly from the network equations. DeepTL4SE produces a point estimate of the current electrical state. A prox-linear network is pretrained on the physics-derived pairs, selected early layers are frozen, and the remaining layers are fine-tuned on target-domain samples. Tests on the IEEE 14- and 118-bus systems show the lowest average RMSE among the reported prox-linear and feed-forward baselines. The best average RMSE values are 0.01266% for the IEEE 14-bus system with MSE loss and 0.00452% for the IEEE 118-bus system with Huber loss.