Harmful Cyanobacterial Blooms forecasting based on improved CNN-Transformer and Temporal Fusion Transformer

Harmful Cyanobacteria have the potential to produce toxins and odors, not only in drinking water but also in public waters where recreational activities take place. Thus, predicting the number of Harmful Cyanobacteria cells is crucial for managing their growth. However, predicting changes in aquatic environments is highly challenging, and predicting aquatic ecosystems is particularly difficult due to uncertainty and complexity in unknown areas. In the past, research has focused on mechanism-based prediction techniques, such as EFDC and Delft3D models. However, with the recent rise of artificial intelligence-based deep learning techniques, it has become imperative to leverage these methods. Additionally, it is crucial to develop a method that can directly predict the number of Harmful Cyanobacteria cells, rather than chlorophyll-a, which only indicates the total generation of algae. In this study, a technique based on artificial intelligence deep learning was proposed to directly predict the number of Harmful Cyanobacteria cells. Advanced analysis was conducted to achieve this goal, by combining the CNN and Transformer algorithms and comparing the results with the Temporal Fusion Transformer (TFT) technique. The models were trained using water quality and algae data collected between 2012 and 2021 and validated the predictions using data from 2022. Employing the proposed method for short-term prediction of Harmful Cyanobacteria is anticipated to assist in operating the algae warning system in Korea. [Display omitted] •We developed and evaluated deep learning-based models for harmful cyanobacterial blooms (HCBs) prediction.•We used Bayesian optimization techniques for hyper-parameter tuning.•We have directly predicted the number of Harmful Cyanobacteria cells.