Hyperparameter-optimized deep learning for next-day pH forecasting in distributed water quality monitoring systems.

Accurate short-horizon water-quality forecasting is important for environmental monitoring systems that aim to move from passive observation toward timely, data-driven intervention. Next-day pH prediction is particularly relevant because pH reflects aquatic chemical balance, pollutant behavior, and ecosystem health. This study addresses next-day pH prediction in a spatio-temporal monitoring context using the public Water Quality Prediction dataset from Georgia, USA, where distributed monitoring records reflect temporal continuity and spatial dependency. Here, spatio-temporal refers to the multi-site and time-indexed dataset structure rather than to an explicit graph-based spatial neural architecture. A two-stage learning framework is used: baseline models are first benchmarked under a unified protocol, and the strongest baseline is then refined through metaheuristic hyperparameter optimization. The framework combines preprocessing, comparative benchmarking, and optimizer-guided refinement, with RegNet as the focal predictive model and the Ocotillo Optimization Algorithm (OCOA) as the hyperparameter search mechanism. OCOA-RegNet is compared with optimizer-assisted RegNet variants based on PSO, GA, WOA DE, MVO, BA, SFS, BBO, and APO. In the baseline stage, RegNet achieved the strongest untuned performance, with an RMSE of 0.0285. In the optimization stage, OCOA + RegNet achieved the lowest RMSE, [Formula: see text] across independent runs, under the adopted protocol. These findings indicate that next-day pH forecasting benefits from both model selection and disciplined hyperparameter optimization, supporting the proposed workflow as a decision-support component for distributed water-quality monitoring.