Simulation-Based Evaluation of Electromagnetic Time-Reversal Methods for Partial Discharge Localization in Power Transformers: Critical Analysis of Advanced Techniques

Abstract

We tested three electromagnetic time-reversal methods to find partial discharges in power transformers using computer simulations. Our goal was simple: figure out if newer, fancier methods actually work better than the basic approach. First, we checked the baseline PDVDM method that was published in IEEE journals. Using a standard 24-disk transformer model, we got pretty good results—about 1.9 mm average error with 80% of our tests landing within 2 mm of the actual discharge location at 20 dB signal-to-noise ratio. This matched what other researchers reported, so we knew our simulation setup was working correctly. Then we tried three supposedly "advanced" methods, keeping everything else the same so we could compare fairly. The HO-AMFO hybrid optimization used a bigger 60-disk model with 10 MHz sampling and 12 sensors, but it performed terribly—1885 mm average error, only 15.18% accuracy. The second method was an IEEE-inspired AI-enhanced approach with three windings and 50 MHz sampling. It did better than the first one with 895.3 mm error, but still couldn't get a single test within 2 mm. The third method threw everything at the problem—over 15 AI techniques combined, 100 MHz sampling—and got 1417 mm error with zero tests under 1 mm accuracy. These results surprised us, honestly. Adding more sensors, faster sampling, and complex algorithms didn't help at all. In fact, they made things worse—about 1000 times worse than the simple method. After digging into why this happened, we found three main problems. First, the optimization algorithms weren't actually following electromagnetic physics properly—they were just hunting for mathematical answers that didn't match reality. Second, the AI methods had basically memorized our simplified simulation setup instead of learning real principles, so they failed when conditions changed slightly. Third, nobody had tested how sensitive these methods were to realistic transformer shapes and noise levels that you'd find in actual substations. We wrote this paper to honestly show that these advanced methods failed our tests. Too many papers claim their new method is amazing without showing the full picture. We measured exactly how much worse the complex methods performed compared to simple approaches, and we explain why the failures happened. Before utilities spend money installing these systems in real transformers, somebody needs to test them on actual hardware—not just simulations—using standardized tests that everyone agrees on. The physics has to guide the AI, not the other way around.