Simulation-based deep learning for lightning localization using convolutional autoencoder and multilayer perceptron

Lightning-induced overvoltages are a major cause of faults and outages in overhead transmission systems, significantly affecting the reliability and operational continuity of power networks. Accurate and timely localization of lightning strikes is therefore essential for improving system monitoring, fault diagnosis, and post-event analysis. In this paper, a simulation-based data-driven approach for two-dimensional lightning localization is proposed using voltage signals induced on transmission lines and evaluated at limited sensor locations. The method combines a convolutional autoencoder (CAE) for feature extraction with a multilayer perceptron (MLP) for regression, forming a hybrid deep learning architecture (CAE-MLP) tailored to the nonlinear nature of the localization problem. The framework uses Rusck’s analytical formulation to generate waveforms for six line configurations, enabling assessment of topology effects on localization accuracy. No field-measured transmission-line records were used. Compared with CNN, LSTM, and standalone MLP, CAE-MLP achieves the lowest coordinate RMSE in 10 of 12 configuration/axis cases. Across six configurations, its RMSE ranges from 121 to 758 m in the x -coordinate and from 123 to 610 m in the y -coordinate; median two-dimensional error ranges from 150 to 427 m. The results demonstrate in-distribution feasibility using two sensor waveforms under assumptions of an infinite, lossless, single-conductor line above perfectly conducting ground. Archived SNR tests indicate lower sensitivity than comparison models over 5–40 dB. However, because training and testing data were generated by the same idealized model, the results do not establish practical deployment readiness. Validation with measured records or an independent, higher-fidelity electromagnetic model remains necessary.