Single-Sensor Machine-Learning-Based Lightning Localization using Lightning-Induced Voltages

In this paper, we employ the XGBoost method for lightning strike localization, utilizing the induced voltage waveform measured at one point of an overhead power distribution line. To identify the optimal location for sensor installation, 10 different sensors were deployed to record induced voltages. This dataset is essential for training the machine learning model. It is important to emphasize that only one sensor is utilized for each lightning strike localization. The dataset comprises 5000 samples, calculated through the Lightning-Induced Overvoltage (LIOV) module in the EMTR-RV software. Unlike the previous work that relied on Rusck's formula (infinite single-wire line assumption), our approach takes into account realistic network topologies through the use of Agrawal et al's field-to-transmission line coupling model. The XGBoost algorithm is utilized to train 70% of the dataset, reserving the remaining 30% for testing the model's performance. To improve training efficiency, we implement the principal component analysis (PCA) on the induced voltage. This reduces the time domain features to 65 samples, resulting in a 15-fold acceleration. Numerical analysis indicates that the trained model attains an accuracy exceeding 99%. The prediction error is about 145 m along the x-axis and 152 m along the y-axis within the computational area of 10×10 km2. These results underscore the effectiveness of our proposed methodology for lightning strike localization on overhead power lines.