Explainable Dual-Path Deep Learning Framework Using DyRoPE-DPRLT with Cuckoo-Catfish Optimization for Accurate PV Power Prediction

For efficient energy management, precise PV power forecasting is essential. This study proposes a unique hybrid Deep Learning (DL) architecture with bio-inspired hyperparameter optimization for PV power forecasting. The proposed process starts with extensive preparation, which includes exception mining, data transformation, discretization using binning, and EDA to ensure data quality. After autocorrelation-based feature extraction, feature engineering normalizes the input space via MinMax scaling. For efficient forecasting, Dynamic Rotary Positional Encoding–Enhanced Dual-Path Retention–Linear Attention Transformer with Adaptive Cross-Path Gated Fusion (DyRoPE-DPRLT) is employed. The proposed DyRoPE-DPRLT combines two parallel pathways: a Retention Network (RetNet) road that employs a retention mechanism to capture long-term relationships and a Linear Attention Transformer path for efficient local pattern extraction. An Adaptive Cross-Path Gated Fusion (ACPG) method that dynamically balances inputs from both directions is used to merge these two routes. Model hyperparameters are optimized using the Cuckoo-Catfish Optimizer (CCO), a bio-inspired method that mimics catfish behavior with Levy flights of cuckoo search for better exploration and convergence. Continuous power output forecasts are produced by the regression head and the last dense layers. The model becomes interpretable by SHAP analysis, which enables the identification of significant meteorological and temporal features influencing forecasts. Experimental evaluation on combined solar power generation datasets reveals the usefulness of the proposed framework. The proposed model had the highest prediction accuracy with an R2 value of 0.9875, outperforming the Temporal Fusion Transformer (0.9759) and LSTM (0.9559) models. Furthermore, with MAE = 0.0199, MSE = 0.000913, RMSE = 0.0302 and NRMSE = 0.03034, it achieved the lowest forecasting errors, demonstrating its outstanding prediction ability.