Data-based Forecasting Of Beach Volumes On Monthly To Yearly Timescales

摘要

Data-based methods for forecasting beach volumes are tested using ground-measured bathymetry from Duck, North Carolina, comprising 26 profiles, 20 year duration and one-month resolution. Derived beach volume time series show weak seasonal and strong event signals. The forecasting methods used are: Holt-Winters (standard and modified), three types of linear regression, and a default forecast in which the latest measurement persists unchanged into the future. Improved forecast accuracies are obtained by two modifications to Holt-Winters, involving an autocorrelation correction and long-term trend-damping, and by smoothing the fitting data using running medians or wavelet approximations. Beach volume forecasts are tested mainly at monthly intervals up to 12 months ahead, with further tests at up to 36 months ahead. Overall, modified Holt-Winters performs best and the default forecast second-best. With an added artificial seasonal signal, modified Holt-Winters outperforms the other methods more substantially.