Temporal water quality forecasting in Peruća Reservoir, Croatia, using a hybrid graph neural network and transformer model under climate variability.

The Peruća Reservoir, a critical karstic water resource in Dalmatia, Croatia, experiences significant water quality fluctuations influenced by rapid surface and groundwater recharge and climate change. This study develops a predictive framework utilizing a hybrid Graph Neural Network (GNN) and Transformer model to forecast key water quality parameters. Based on a dataset of monthly surface water quality measurements from 2019 to 2021, including water level, dissolved oxygen (DO), electrical conductivity (EC), pH, hardness, temperature, and precipitation, the model accurately predicts the Water Quality Index (WQI), DO, and EC (R2 > 0.92 for 5-day forecasts). The analysis focuses on temporal water quality variations at a single monitoring station due to data limitations, with spatial resolution limited to inflow influences. Climate projections for 2050, assuming a 5% decrease in precipitation and a 1.7 °C temperature rise, indicate a potential decline in reservoir water levels by 0.75–1.17 meters and a summer WQI reduction to 67–70. The proposed model consistently outperforms benchmark Random Forest and Long Short-Term Memory (LSTM) models, demonstrating its robust predictive capability and providing crucial decision-support for the sustainable management of karstic reservoirs. The Peruća Reservoir in Croatia faces water quality fluctuations due to rapid recharge and climate change. A hybrid GNN-Transformer model accurately predicts water quality parameters (R2 > 0.92), outperforming other models. Climate projections for 2050 suggest declining reservoir levels and water quality, aiding sustainable management. [ABSTRACT FROM AUTHOR]