Enhancing River Water Quality Prediction: A Hybrid Approach of Deep Autoregression and Bidirectional LSTM Networks

Abstract

Time series forecasting models are essential for the prediction of river water quality, which is an important part of maintaining environmental sustainability and protecting public health. However, traditional prediction methods often fail to capture the complex temporal dependencies and non-linear relationships that are inherent to river water quality information. An innovative approach is proposed here, to improve river water quality prediction by combining deep autoregression and bidirectional LSTM networks. The deep autoregression model takes advantage of the inherent temporal correlations that are present in the river water quality data and integrates autoregressive features that encapsulate historical information about the target variable. The BiLSTM architecture lets the model to understand from the historical and future contexts, allowing a holistic view of the temporal dynamics that affect water quality in the river basin. We also introduce feature augmentation technique that improves the proposed model’s ability to capture the complex relationships and patterns in the data. By enriching the model's input features with domain specific information, including meteorological data and hydrological parameters as well as land use characteristics that are relevant to the river basin. By incorporating domain-specific information through feature augmentation, the model is poised to provide more accurate and insightful predictions, thereby aiding in effective water resource management and environmental conservation efforts. The error rate has been approximately reduced by 42.6% by using the DAFA-BiLSTM model, when compared to the conventional LSTM models.