Cycleift: A deep transfer learning model based on informer with cycle fine-tuning for water quality prediction
摘要
Long-term water quality (WQ) prediction is crucial for water environment management as it enables the assessment of water ecosystem quality and risk in advance, reflecting its environmental health status. WQ prediction aims to estimate future changes in WQ according to historical data. Due to their significant advantages in mapping nonlinear relationships, deep learning methods have achieved excellent performance in WQ prediction tasks. However, these methods are susceptible to data perturbations in short-term prediction, while long-term prediction tasks require large amounts of data to train the corresponding model effectively. Collecting such large datasets can hinder the progress of deploying models at WQ monitoring stations. Transfer learning methods can address the issue of limited data collected by WQ monitoring stations. Nonetheless, such methods are prone to overfitting the source domain data during the pre-training stage, and the long-term dependencies of the WQ data are often lost during transfer learning processes, reducing prediction accuracy. In this paper, we develop a transfer learning model based on Informer with cycle fine-tuning (cycleIFT). cycleIFT utilizes the self-attention mechanism and encoder–decoder structure of the Informer model to learn long-term time dependencies. It innovatively combines with a cyclic fine-tuning transfer learning framework to alleviate the overfitting problem. Experimental data are collected from 120 automatic monitoring stations across major rivers and lakes through the China National Environmental Monitoring Center. The model performance is evaluated using these data. The results indicate that cycleIFT can effectively improve the long-term prediction accuracy for the four WQ indicators (pH, DO,