Designing robust adaptive ensemble deep learning based decomposition technique for sea level variability prediction

Designing robust adaptive ensemble deep learning based decomposition technique for sea level variability prediction

Sea level variability is an urgent climate risk, threatening to vanish islands and coastal areas. Forecasting future sea level rise accurately is fundamental to support experts for flooding and erosion control. In this paper, a novel sea level variability forecast model (CG-CEEMDANsingle bondHFS-AEM) is proposed integrating correlation graph (CG), a complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), Hilbert feature selection approach (HFS), adaptive ensemble model (AEM), and oppositional learning sparrow search algorithm (OLSSA). The AEM is a novel ensemble model that combines the strengths of GRU, self-attention LSTM and XGBoost models based on dynamic weights assignment strategy, adjusting to real-time changes in sea level rise by updating the weights according to the error and performance of the models. Firstly, the input data is pre-processed using correlation graph to remove lower correlated variables and fill the missing values in the data. After that the CEEMDAN technique is employed to decompose the data, followed by HFS to select the most efficient features. The selected features are then employed into the AEM model where the OLSSA is adopted to select the optimal hyper-parameters of the proposed model. To verify the efficiency of the proposed CG-CEEMDANsingle bondHFS-AEM against comparing models, extensive experiments were conducted to forecast sea level variability for Hillary and Burnie stations in Australia. The results shows that the proposed model obtained the highest accuracy in terms of goodness-of-fit metrics against the state-of-the-art benchmark comparing models. The proposed model can offer a valuable tool for coastal planning and policy making under the recent climate change.