Prediction Model for Railway Noise Emission in Curves

Based on five measurement campaigns in different sections of the Austrian heavy rail network (radii between 226 and 440 m), a prediction model to estimate railway noise emission in curves is developed. Scopes in model building are practical usefulness and general applicability. After data filtering a total of 29,097 train pass bys are available for further evaluation. To consider different boundary conditions in each section, 190 directly measured or calculated predictors are defined. A feature selection process reduces the predictor quantity to 17. The final model is built with the random forest algorithm and is trained on two long-term measurement campaigns. To estimate general applicability, validation is done on data points of three short-term campaigns with completely different curve radii, climatic conditions and train type distributions. Considering benchmarks on the latter, predictive performance is 4 dB RMSE and 0.57 R-2. Predictions averaged among a longer time span deviate by 0.5 dB to -1.1 dB in energetic mean and -0.2 dB to 1.4 dB in median from the original values.