A mixture of experts approach to sandy shoreline modelling in storm dominated systems

Abstract

The complexity of sandy shoreline dynamics along storm-dominated coastlines is largely driven by the fundamentally distinct processes that govern individual storm events and post-storm recovery periods. Despite advancements in both physics-based and machine learning methods, accurately predicting both the rapid shift in shoreline response due to storms, and the subsequent recovery periods across multi-annual forecasting horizons remains a significant challenge. In this study, we introduce a Mixture of Experts (or ‘Mixture’) approach to shoreline modelling that augments a Long Short-Term Memory (LSTM) neural network with a specialized linear regression storm model. This state dependent approach, guided by a threshold gating mechanism, generates stable multi-year forecasts that effectively capture both storm impacts and longer-term shoreline trends. We apply the Mixture at two storm-dominated sites along the southeast Australian coastline and observe an improvement in NMSE of 0.26 at Narrabeen and 0.61 at the Gold Coast, relative to a baseline standalone LSTM model. The findings of this work emphasize that context-informed modelling decisions can significantly enhance machine learning methods, leading to more accessible and actionable forecasts while minimizing an increase in model complexity.