Deep Spatiotemporal Learning for Multivariate Water Quality Prediction: Temporal Dynamics–Aware CNN–GRU Hybrid Model

Proper forecasting of water quality indicators is one of the most important factors in sustainable environmental management, protecting ecosystems, and early detection of potential pollution risks. Among them, pH is essential for regulating chemical and biological processes in water and can be significantly altered by natural and man-made factors. Water quality parameters are also nonlinear, multivariate, and time-varying, which makes it challenging to estimate pH in modeling. In this paper, we develop a deep learning-based approach to predicting next-day pH using multivariate time-series water-quality measurements from different monitoring stations. The hybrid CNNGRU architecture is developed in this work in which the short-term temporal patterns of the multivariate water-quality sequences can be obtained through the convolutional layers, whereas the long-term temporal dependencies are learned by the GRU units. Several machine learning models and deep learning models, such as classical ensembles, recurrent neural networks, and hybrids, are designed to learn the complex temporal dynamics of the data. To be more precise, a spatiotemporal modeling framework based on a hybrid Convolutional Neural Network and Gated Recurrent Unit design is developed to reveal the local-temporal dynamics and long-term associations of the variables of water quality under consideration. Many experiments are conducted to assess the predictive accuracy of all models using standard error and agreement measures under normal conditions. According to the results of the experiments, the proposed CNN-GRU always performs better in comparison to classical machine learning (as well as stand-alone deep learning) models in terms of RMSE, MAE, R2, and WI. Experimental studies indicate that hybrid deep learning models achieve substantial gains over classical pH estimation methods. The findings indicate that convolutional feature extraction combined with recurrent temporal modeling can be essential for modeling the dynamic nature of water quality systems.