Short-Term Water Quality Index Forecasting Using Bayesian-Optimized Machine Learning and Hybrid Time-Series Features

Freshwater aquaculture environments depend on strict water quality standards, but traditional monitoring is point-in-time and existing literature treats water quality largely as static classification, leaving high-frequency temporal forecasting a significant gap. This study investigates high-resolution forecasting of the Weighted Arithmetic Water Quality Index (WAWQI) using hybrid time-series features and Bayesian-optimized machine learning on data recorded at a nominal 15-s interval in an aquaculture lake, across five horizons from 1 to 60 min. Seven algorithms are compared against persistence, AutoRegressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) baselines on identical rows under a leakage-controlled protocol (chronological split, embargo, single test-set reading), and every comparison is tested with the Diebold–Mariano test and corrected for multiple testing. Persistence alone attains a coefficient of determination of 0.950 at 1 min, falling to −0.214 at 60 min, so goodness-of-fit statistics reported without a naive reference are uninformative. Of 112 model–baseline comparisons, one improves significantly on persistence. Rolling-origin evaluation shows forecasting skill is conditional on target variability, reaching a skill score of 0.503 at 60 min in the most variable test window and −0.065 in the least variable (zero indicates no improvement on the last observation). Permutation importance on held-out data ranks dissolved oxygen (DO), not electrical conductivity, first among sensor channels, and variance decomposition attributes 67.7 % of index variance to DO and 24.5 % to turbidity. These results establish the conditions under which high-frequency forecasting adds information beyond persistence.