Predicting Beach Profile Changes using Neural Networks with Recursive Add and Repeat Simulation

Abstract

Accurate prediction of long-term beach profile changes is critical for sustainable coastal management, particularly in the face of climate change, sea level rise, and shifting wave conditions. This study evaluates the performance of two artificial neural network architectures Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) in forecasting 15 years of annual beach profile evolution at Narrabeen-Collaroy Beach, Australia. Both models were trained using the Add and Repeat (AdRpt) method, an iterative forecasting approach that extends prediction horizons by incorporating previous outputs as new inputs. Key environmental variables included sea level trends, significant wave height, and wave period. Model performance was assessed across five profiles (PF1, PF2, PF4, PF6, and PF8). Results show that the LSTM consistently outperformed the MLP, achieving RMSE as low as 0.45 m and R² values up to 0.97. While LSTM captured temporal patterns effectively, both models struggled with abrupt morphological changes, such as the severe erosion observed at PF2 in 2001. Profiles near the intertidal zone also exhibited greater prediction variability. Furthermore, the study highlights that relying solely on R² can be misleading, as high R² may coincide with substantial RMSE and MAE values. A multi-metric evaluation approach is essential to ensure reliable model interpretation. These findings support the application of LSTM-based models for data-driven, long-term coastal planning and adaptive nourishment strategies.