Multi-feature fusion monthly runoff prediction under different climate conditions using APO-optimized CNN-BiGRU-Self-Attention

Monthly runoff sequences exhibit highly nonlinear and nonstationary characteristics that impede traditional single models from capturing long-term dependencies and abrupt changes. This study proposes a hybrid deep learning model, APO-CNN-BiGRU-Self Attention, that integrates the Arctic Puffin Optimization (APO) algorithm, Convolutional Neural Networks (CNN), Bidirectional Gated Recurrent Units (BiGRU), and a Self-Attention mechanism to enhance prediction accuracy. Specifically, a CNN extracts local features, a Self-Attention module assigns weights to higher-order latent features, and the BiGRU detects Bidirectional dependencies within the sequence. At the same time, the APO algorithm optimizes the hyperparameters. Validation was subsequently conducted at the Yingluoxia (YLX) and Manwan (MW) stations. The following metrics were used for evaluation: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Nash-Sutcliffe Efficiency (NSE), correlation coefficient (R), and Kling-Gupta Efficiency (KGE). Experiments demonstrate that, compared to the standalone BiGRU, the proposed model reduces MAPE by 26.681% and improves NSE by 5.498% at the YLX station, while reducing MAPE by 21.874% and enhancing NSE by 3.673% at the MW station. Furthermore, the APO demonstrates superior convergence characteristics compared with the WOA and GWO meta-heuristic optimization algorithms, enabling the hybrid model to achieve robust monthly runoff prediction across diverse climatic zones.