TFT Architecture and RNN Variants for Water Quality Prediction of Bharathapuzha River

Abstract

Water quality is a crucial aspect of the health and well-being of communities and ecosystems around the world. This study focuses on water quality prediction for the Bharathapuzha River, which is susceptible to various sources of contamination. The prediction model is built using Recurrent Neural Network (RNN) variants and Temporal Fusion Transformers (TFT) approach.  The dataset developed for building the prediction model consists of 2190 unique instances containing physicochemical and seasonal parameters. Water quality varies over time due to changes in natural and human factors, such as weather conditions, land use, pollution levels, and treatment processes. The objective is to capture the time series patterns in the data and to forecast the water quality index accurately. The performance result demonstrates that the TFT outperforms the RNN variants in prediction. This study highlights the importance of TFT in trend analysis and developing a reliable forecasting model.