Traffic Noise Modelling Using Artificial Neural Network

Abstract

Highway traffic noise has been a Federal, State, and local problem. Emanating from vehicle engines, exhaust systems, and tires interacting with pavement, traffic noise affects the quality of life for nearby residents and businesses by drowning out conversations, disrupting sleep, and discouraging outdoor activities. Over the years, community and motorist concerns have fueled the push to improve noise measurement and modeling tools that help transportation agencies address the highway traffic noise problem. The heterogeneous physiognomy of traffic noise, together with the characteristics of environmental noise, with their great spatial, temporal and spectral variability, makes thematter of modeling and prediction a very complex and non-linear problem, therefore a need is being felt to develop a traffic noise prediction model suitable for the Indian condition. The present work represents a traffic noise prediction model taking Patiala-Sangrur highway as a representative/demonstrative site. All the measurements of noise levels were measured at different selected points around the highway at different time intervals on number of days in a random/staggered manner in order to account for statistical temporal variations in traffic flow conditions. The noise measurement parameters recorded are Equivalent Noise Level (Leq), Percentile Noise Level (L10), Maximum Equivalent Noise Level (LMAX), Minimum Equivalent Noise Level (LMIN). Artificial Neural Network (ANN) approach has been applied for traffic noise modeling in the present study. The measured parameters were divided into two classes i.e. output parameters (L10, Leq)) and input parameters (vehicle volume/hr., percentage of heavy vehicles and average vehicle speed). The input parameters are further divided randomly into three kinds of samples • Trainingset • Validation set • Testing set After training and testing of the ANN, it was founded that the values of correlation coefficient(R) were 0.9434, 0.9644 & 0.92863 for the training, validation and testing samples respectively, and the percentage error varied from -0.19 to 0.64 and 0.54 to 0.99 for Leq and L10, therefore a good correlation coefficient and less percentage error between experimental and predicted output is an indication of better prediction capability of neural network.