Multistep-ahead forecasting of chlorophyll a using a wavelet nonlinear autoregressive network

Abstract

Multistep-ahead forecasting is essential to many practical problems, such as the early warning of disasters. However, existing studies mainly focus on current-time or one-step-ahead prediction since forecasting multiple steps continuously presents difficulties, such as accumulated errors and long-term time series modeling. In this paper, an effective multistep-ahead forecasting model wavelet nonlinear autoregressive network (WNARNet), which integrates the wavelet transform and a nonlinear autoregressive neural network (NAR), is proposed for the forecast of chlorophyll a concentration. The wavelet transform decreases the accumulative errors by dividing complicated time series into simpler ones. Simultaneously, the NAR maintains the dependencies between the time series. The buoy monitoring data of the Wenzhou coastal area obtained in 2014-2015 is used to verify the feasibility and effectiveness of WNARNet. The model performs well in predicting the dynamics of chlorophyll a and it is able to predict different horizons flexibly and accurately without training new models. Furthermore, experimental results demonstrate that WNARNet significantly outperforms other benchmark methods of multistep-ahead forecasting. When forecasting 20 steps ahead, the r of WNARNet is 0.08 higher and the RMSE is 0.04 lower than the values of the benchmark models. Therefore, the newly proposed approach represents a promising and effective method for the future prediction of chlorophyll a.