Dongting Lake algal bloom forecasting: Robustness and accuracy analysis of deep learning models

Harmful algal blooms (HABs) pose a significant threat to aquatic ecosystems, prompting efforts to predict their occurrence for swift action by water management agencies. Despite the potential for high-precision forecasting through machine learning, the effectiveness of these models is often compromised by data quality issues, such as incomplete data sets, inaccuracies in historical records, inconsistencies in sampling methods, and the dynamic nature of environmental factors, leading to temporal and spatial variability. This study develops an early warning system for HABs using water quality data from a freshwater lake prone to such blooms. It employs a deep learning approach that integrates time series analysis with the iTransformer model to enhance prediction accuracy. The methodology utilizes the iTransformer model's robust preprocessing capabilities to address missing values and maintain data continuity, ensuring effectiveness even when with incomplete datasets. Additionally, the study identifies key factors influencing algal density by analyzing the model's attention weights, highlighting the importance of nutrients and temperature. A feature ablation experiment underscores the model's inherent robustness, showcasing its ability to deliver reliable predictions despite incomplete data. The research contributes to water quality management in Dongting Lake and presents a novel application of deep learning in environmental monitoring. Despite the model's current effectiveness, future work should explore additional environmental variables to enhance its predictive power and generalizability.