AC electromagnetic interference (EMI) between railway and nearby power lines causes serious concerns about railway personnel and equipment safety. The AC EMI analysis is usually conducted by complex computer simulation software to estimate induced voltages on rail tracks, and induced voltages can be further used for safety evaluation. However, such analysis becomes especially difficult at the transmission line routing stage when only limited information of railway and power lines is available. To overcome this challenge, a novel regression model-based toolbox is developed in this paper to predict induced voltages on rail tracks. To develop this toolbox, the dataset acquisition is critical due to very limited research done in this area. The novel dataset is produced by our newly developed AC EMI study method, where variations of several essential factors are considered, including the power line's current, the separation distance between power lines and railway, the ballast resistance, and the length of rail tracks. Two models are eventually chosen to predict induced voltages: “Gaussian process regression” with “matern 3/2” kernel function; and a tri-layered “neural network” model with “sigmoid” activation function. To improve the accuracy, hyperparameters of the regression algorithms are optimized by Bayesian optimization. The toolbox is accurate and easy-to-use, and is currently in use by Manitoba Hydro, Canada.
Read full abstract