Water Quality Index Forecasting via Transformers: A Comparative Experimental Study

Water quality index (WQI) forecasting plays a crucial role in water resource management and water pollution control. It can provide early warnings and facilitate proactive prevention measures. This paper focuses on the task of long-term WQI forecasting utilizing transformer models. We first provide a brief overview of the basic Transformer and its seven improved models, namely Reformer, Informer, Autoformer, Pyraformer, FEDformer, Crossformer, and PatchTST. Based on this survey, we conduct a comparative experimental study by applying these eight transformer models to forecast WQI for two national monitoring river sections. The experimental results reveal that the PatchTST and Crossformer models outperform other models across almost all future forecasting window sizes, including very long-term settings. Furthermore, PatchTST and Crossformer exhibit higher computational efficiency due to their patching or segmentation strategy. Therefore, PatchTST and Crossformer are effective and efficient for WQI forecasting. To the best of our knowledge, this paper is currently the most systematic study on the application of transformers in the field of WQI forecasting. Through our study, we have observed the tremendous potential of transformers for WQI forecasting.