Ensemble deep learning modeling for Chlorophyll-a concentration prediction based on two-layer decomposition and attention mechanisms

This study proposes a novel hybrid prediction model for short-term Chl-a concentration prediction in the Chaohu Lake basin. The model adopts a two-layer decomposition, grouping prediction, and summation design, with an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) used for data decomposition and sample entropy (SampEn) for determining the complexity of the decomposed components. The component with the highest SampEn is then further decomposed using variational mode decomposition (VMD). In the grouping prediction part, a hybrid prediction model is utilized, based on bidirectional gate recurrent unit (BiGRU), temporal convolutional network (TCN), and attention mechanism (AM), to predict all components after the second decomposition separately. The final Chl-a concentration prediction is obtained by linearly summing the predicted values of each component. Through comparison with sensitivity analysis, point estimation, and interval estimation, this study finds that the proposed ICEEMDAN-VMD-BiGRU-TCN-AM model performs well in predicting the content of Chl-a concentration in lakeshore. The model has the highest fitting determination coefficient of 98.79%, the smallest root-mean-square error (RMSE) of only 0.0063 (ug/L), and the largest probability interval coverage percentage (PICP) of 97.57%. These results indicate that the IVBTA model has significant advantages in accuracy, precision, and stability compared to other models. This means that our method can more accurately predict the content of Chl-a in reservoirs, providing a new approach for water quality prediction and effective prevention and control of eutrophication in reservoirs.