Data-based yearly forecasting of beach volumes along the Dutch North Sea coast

Beach and nearshore levels have been measured yearly along the entire Dutch North Sea coast since the mid 1960s (the ‘Jarkus’ data set). This data set has been processed to create separate time series of beach volumes at longshore intervals of about 250 m, giving over 2000 time series in total. These time series typically show a high annual variability with weak long-term trends. The present Dutch national coastal management strategy involves making year-ahead forecasts of beach volumes by extrapolating a linear least squares trend through the previous ten years' data separately for each longshore location. In this paper, these forecasts are shown to be worse than the trivial forecast in which the most recently measured beach volume persists unchanged into the future, with a mean square error (MSE) about 13.5% worse (equivalent to a root mean square error (RMSE) 6.5% worse). Improvements to these forecasts are sought by testing six different univariate forecasting methods. The two best methods improve on the persistence of the most recently measured beach volume by about 15% MSE (8% RMSE), and on the presently used linear least squares trend method by about 25% MSE (13.5% RMSE). Further comparisons are made between the forecasting methods to investigate several factors. These include varying the amount of fitting data for the forecasting methods, smoothing of the fitting data, different methods for interpolating gaps in the data, the longshore aggregation of data, making forecasts for coastal profiles with and without nourishments, and making forecasts up to five years ahead. These forecasting methods are designed as a coastal management tool to provide yearly forecasts quickly and routinely for the whole Dutch North Sea coast.