Predicting Marine Fog on the Grand Banks of Newfoundland & Labrador

Abstract:

Wood Environment & Infrastructure conducted research on marine fog formation, dissipation, and movement on the Grand Banks of Newfoundland & Labrador, Canada during the period 2014 to 2020. This research and development project included a multi-year assessment of marine fog predictability for this location that has dense marine advection fog as frequently as fifty percent of the time in the spring and summer. The project included development and validation of: a suite of deterministic and probabilistic visibility prediction models that were assessed against in-situ measurements and meteorologist-produced real-time forecasts; a satellite-based detection and extrapolation nowcast system; and, a Bayesian neural network (BNN) system used to create a “model of models” that enabled forecasters to interpret a plethora of guidance systems which had varying levels of accuracy. The deterministic models were based on published and Grand Banks-specific algorithms and used inputs from a range of publicly available and Wood-operated numerical weather prediction (NWP) and oceanic models. The probabilistic models were developed using machine learning techniques applied to a 20-year observational data set, and use input from NWP and oceanic models. The satellite-based nowcasting system consisted of fog maps that were created using GOES-16 data, and a sequence of these maps plus NWP data input into a machine learning application extrapolated fog maps several hours into the future. The BNN “meta model” incorporated data from the developed fog prediction models and used real-time observations. The results showed that meteorologists provided value beyond the skill level of model predictions at short-term forecast horizons; that different NWP systems had varying levels of skill at different lead times and meteorological setups; that probabilistic models based on machine learning out-performed fully deterministic systems; that satellite-derived fog data provided short-term fog prediction skill; that visibility systems implemented in publicly available NWP models were not optimized for the Grand Banks, high-wind, advection fog environment; and that Bayesian neural networks showed promise for both improving prediction skill and for assessing predictability.