Streamflow Forecasting in the Eastern Black Sea Basin Using a Contrastive PSO-Optimized Transformer-LSTM Framework

Reliable streamflow forecasts are essential for effective water resources management, providing vital information for flood and drought mitigation, and sustainable water allocation. In this study, a novel hybrid CA-PSO-LSTM-Transformer model was developed for streamflow forecasting by integrating LSTM, attention, and transformer fragments to daily data from three flow measurement stations (FMSs) in the Eastern Black Sea Basin. The accurate R2 and RMSE improvement observed in Ulucami FMS by the value of 0.942 and 57% (from 4.780 to 2.048), respectively. In this station, streamflow data had a more uniform distribution and the average flow rate was reached the highest value. Additionally, modelling analysis indicated that the model’s error distribution was close to a normal distribution and contained no systematic deviation. The proposed CA-PSO-LSTM-Transformer model achieved the lowest errors and highest accuracy across all FMSs. The hybrid model was demonstrating strong generalization for accurate water resource management and flood forecasting applications.