Neural Network Models as Decision-Support Tools for Road Traffic Noise Management in Port Cities: A Comparison with Standard Prediction Methods

Road traffic is the main source of exposure to environmental noise in Europe and Directive 2002/49/EC requires the competent authorities to evaluate and manage it. In port cities, a significant portion of this exposure is generated by road traffic induced by ferry embarkation and disembarkation operations: a pulsating and highly non-stationary source, which standard forecasting methods, designed to return long-term average indicators, describe only in aggregate form. This work compares two forecasting approaches applied to the same waterfront of Olbia (Sardinia, Italy) and addresses a question of an operational rather than metrological nature: which model provides the information that a specific noise management decision actually needs. A physical model based on ISO 9613 and the CNOSSOS-EU methodology was implemented in CadnaA and calibrated on continuous monitoring data conducted in three receptor locations during a low and a high season. A non-linear autoregressive artificial neural network, developed and experimentally validated previously on the same waterfront, was used to predict the equivalent sound pressure level starting from vehicle flow data alone. The Lday, Levening and Lnight indicators returned by the two models are compared, together with the required input data, the spatial and temporal resolution, the latency and the installation and update costs over the five-year cycle of the strategic acoustic mapping. The differences between the two models reach 14 dB in low-flow night periods, while the calculation times vary between 3 and 36 h for the physical model versus an almost real-time inference for the neural one. The two approaches are complementary and not alternative: the physical model remains necessary for planning and for scenarios not yet implemented, while the operational and intraday management of noise in a ferry port is better served by data-driven prediction. The two models agree within 1 dB on average during the day and evening periods, whereas at night the neural model returns systematically higher levels, by 5.4 dB on average and by up to 14 dB, at all receiver positions and in both seasons.