Monthly runoff time series prediction based on the hybridized CNN-LSTM-attention model with different bio-inspired optimization algorithms

ABSTRACT Accurate monthly runoff prediction is important for water resource management. This study develops a hybrid CNN-LSTM-Attention model optimized by six bio-inspired algorithms (GA, ACO, CSA, FPA, GWO, and WOA) for monthly runoff forecasting at the Guandi (GD) and Ertan (ET) stations in the Yalong River Basin. The model employs CNN to extract local temporal patterns, LSTM to capture temporal dependencies, and attention to assign adaptive weights to informative LSTM outputs. Under a single-run experimental setting, the six hybrid models were compared with standalone and fused benchmark models. At the GD station, GWO-CNN-LSTM-Attention achieved the best overall performance (MAPE = 24.20%, RMSE = 264.86 m3/s, NSE = 0.9497, R2 = 0.9584), marginally outperforming WOA-CNN-LSTM-Attention. At the ET station, WOA-CNN-LSTM-Attention achieved lower MAPE and RMSE and higher NSE relative to GWO-CNN-LSTM-Attention, whereas GWO-CNN-LSTM-Attention obtained a slightly higher R2. At both stations, all six hybrid models outperformed the benchmarks during the validation period. These results indicate that bio-inspired optimization, particularly GWO and WOA, can improve CNN-LSTM-Attention performance under the evaluated single-run setting. The framework may offer potential value for reservoir operation and water allocation, but these applications were not directly assessed. Future repeated runs will quantify stochastic variability and evaluate model robustness.