A novel water quality prediction model based on BiMKANsDformer

Abstract

Water quality prediction is crucial for protecting aquatic ecosystems and ensuring human health. However, the water quality time series exhibits characteristics such as nonlinearity and nonstationarity, making efficient feature extraction crucial for improving prediction accuracy. To achieve more accurate and efficient prediction tasks, this study improves the traditional Transformer and proposes a novel water quality prediction framework based on a Transformer called BiMKANsDformer. Secondly, this study improves the interactive convolution block (ICB) by integrating dilated convolution, developing the D-ICB module suitable for extracting complex time series features. Finally, by combining the long-term dependency capturing capability of D-ICB with the feature extraction advantages of BiMamba+ and KANs, this study integrates these components with a Transformer to enhance its processing ability for time series data. Comparative experiments indicate that BiMKANsDformer shows significant advantages in NSE, MAE, RSR, and MAPE, demonstrating stronger robustness and predictive accuracy.