Feasibility Study of AI‑Driven Short‑Term Forecasting of Train Noise Levels in the Semarang–Cepu Corridor

Abstract

Environmental noise from railway systems is an increasingly relevant concern in urban settings due to its effects on human health and comfort. This study explores the feasibility of applying machine learning techniques to forecast short-term train noise levels, using data collected from six monitoring sites along the Semarang–Cepu railway corridor in Indonesia. The dataset includes noise levels in decibels, ambient temperature, time features, and location information. After preprocessing, including one-hot encoding and the construction of lag features, a Random Forest Regressor is trained to predict one-step-ahead noise levels. The model is evaluated alongside K-Nearest Neighbors, XGBoost, and Linear Regression for comparison. Random Forest achieves the best overall performance with a mean absolute error (MAE) of 0.71 dB, a root mean squared error (RMSE) of 1.56 dB, and an R2 score of 0.9475. Feature importance analysis highlights the significance of recent noise history, with lag variables providing the strongest predictive power. The results suggest that machine learning can support the development of real-time railway noise forecasting systems, providing a valuable tool for proactive environmental management and urban planning.