Improving the prediction accuracy of cyanobacterial blooms using a hybrid model integrating dimensionality reduction and deep learning

Cyanobacteria Harmful Blooms have negatively impacted human socio-economic activities and the natural ecological environment. Predicting cyanobacterial blooms requires the application of advanced analytical models powered by artificial intelligence to accurately model bloom development. In this study, we developed a hybrid model that integrates the Temporal Convolution Network (TCN) with the Long Short-Term Memory Network (LSTM) to achieve highly accurate predictions of cyanobacterial bloom events. We used Principal Component Analysis (PCA) during the preprocessing phase to reduce data dimensionality of model input variables and enhance computational efficiency. We used a data set from the Darling River in Australia to conduct a comparative analysis with traditional deep learning models and comprehensively assess the proposed model performance. We also investigated the application of the correlation coefficient matrix method in variable selection. The results demonstrate that the proposed model outperforms conventional methods in predicting cyanobacterial blooms. When benchmarked against the Recurrent Neural Network (RNN) model, our model demonstrates a superior ability to capture the intricacies of time series data. Specifically, it achieves a 79.5 ​% reduction in Mean Square Error (MSE) and a 197 ​% increase in the Coefficient of Determination (R 2 ). Furthermore, comparative experiments were carried out across several prediction time steps, indicating that the model achieves optimal performance in minimizing both the MSE and Mean Absolute Error (MAE) when predicting one day ahead. Notably, the error distribution is most condensed at this time step, outperforming models with prediction horizons ranging from two to seven days. We also undertook a thorough analysis of the significance of each principal component in model predictions. The results highlight that principal component 5 and principal component 2 exhibit the most substantial contribution to the prediction outcomes, collectively accounting for 82 ​% of the total contribution. We delved deeper into the interdependence of principal components within the model and elucidated their collaborative role in the cyanobacterial bloom prediction process. Finally, the applicability and generalization capability of the model has been validated using data from Caohai Lake in China. In conclusion, the model demonstrates considerable potential in predicting the development of blooms, offering novel insights and solutions to pertinent domains. • Applying PCA reduced data complexity. • The hybrid PTL model significantly improved prediction accuracy. • Prediction for 1 day outperforms that for 2–7 days. • SHAP analysis revealed that PC5 and PC2 contribute the most to the prediction output. • The model's applicability and generalization ability are confirmed using a new dataset.