Comparisons between physics-based, engineering, and statistical learning models for outdoor sound propagation

Abstract

Many outdoor sound propagation models exist, ranging from highly complex physics-based simulations to simplified engineering calculations, and more recently, highly flexible statistical learning methods. Several engineering and statistical learning models are evaluated by using a particular physics-based model, namely, a Crank-Nicholson parabolic equation (CNPE), as a benchmark. Narrowband transmission loss values predicted with the CNPE, based upon a simulated data set of meteorological, boundary, and source conditions, act as simulated observations. In the simulated data set sound propagation conditions span from downward refracting to upward refracting, for acoustically hard and soft boundaries, and low frequencies. Engineering models used in the comparisons include the ISO 9613-2 method, Harmonoise, and Nord2000 propagation models. Statistical learning methods used in the comparisons include bagged decision tree regression, random forest regression, boosting regression, and artificial neural network models. Computed skill scores are relative to sound propagation in a homogeneous atmosphere over a rigid ground. Overall skill scores for the engineering noise models are 0.6%, -7.1%, and 83.8% for the ISO 9613-2, Harmonoise, and Nord2000 models, respectively. Overall skill scores for the statistical learning models are 99.5%, 99.5%, 99.6%, and 99.6% for bagged decision tree, random forest, boosting, and artificial neural network regression models, respectively.

Signal propagation modeling in complex, three-dimensional environments

“…EASEE v3 incorporates a novel computational approach for near-ground sound propagation, namely machine-learning algorithms (Hart et al 2016). Specifically, the Java class MachineLearningPropagator provides an interface that can call four different pre-trained machine-learning algorithms: an artificial neural network, bagged decision tree regression, random forest regression, and boosting regression.…”

“…The machine-learning algorithms were trained using a dataset of 27,000 CNPE calculations with randomized source frequency, source height, atmospheric parameters (wind direction, friction velocity, roughness height, sensible heat flux, boundary-layer depth), and ground parameters (static flow resistivity and ground porosity) (Hart et al 2016). This parameterization is intended for near-ground sound propagation only.…”

Statistical prediction of far-field wind-turbine noise, with probabilistic characterization of atmospheric stability

“…31 Improvements on the ISO-9613-2, such as the Nord2000 32 and HARMONOISE 33 models, have been made by building upon ray-tracing approaches. But these do not address some important factors, such as the Nord2000 model's lack of range-dependent sound-speed profiles or HARMONOISE model's neglect of multiple ground reflections; 34 the former is relevant due to wind-turbine wakes, and the latter is an issue in winter and nighttime (stable) conditions. Further, these models do not readily facilitate the prediction of variability per observable stability statistics (again, we are interested here in finding climatologically representative noise statistics).…”

A physics-informed neural network for sound propagation in the atmospheric boundary layer

“…Probabilistic machine learning (ML) and statistical learning (SL) methods [1][2][3][4] are helping to establish viable and testable ways of accounting for the many sources of uncertainty. Pettit and Wilson 2 used proper orthogonal decomposition (POD) for compact representations of the process's variability from an ensemble of transmission loss (TL) realizations, and cluster-weighted models of the joint probability density function of each POD coefficient and the governing parameters.…”

“…Hart, et al 3 expanded this line of work using the same outdoor sound propagation model as the benchmark for testing the predictive skill of several engineering, machine learning, and statistical learning models. The ML and SL models were trained on samples from the CNPE-MOST model computed for various combinations of atmospheric refraction and ground impedance conditions, i.e., at various points in the parameter space.…”