Artificial intelligence for smart solar energy monitoring: Genetic attention-based hybrid deep–handcrafted fusion for faulty solar panel image classification

• A novel Genetic Attention-Based Hybrid Deep–Handcrafted Fusion (GA-HyFusionNet) framework is proposed for intelligent solar panel fault classification. • The model fuses multi-resolution DWT textures, class-aware BoVW histograms, and fine-tuned ResNet50 embeddings within a unified PCA-aligned latent space. • A Genetic Algorithm optimizes attention weights across modalities, achieving adaptive, interpretable, and computationally efficient fusion. • GA-HyFusionNet attains 94.35 % accuracy with only 19–21 ms inference time and a lightweight deployment size (∼87–107 MB), ensuring real-time edge feasibility. • Early detection of dust, cracks, and surface defects enables up to 5 % annual energy-loss recovery, supporting SDG 7: Affordable and Clean Energy. With global photovoltaic (PV) capacity exceeding the terawatt scale, timely and accurate fault detection is pivotal to minimizing energy losses and accelerating the transition to net-zero energy systems. In this context, we introduce GA-HyFusionNet—a novel, interpretable, and lightweight hybrid learning framework for solar panel fault classification that combines the strengths of handcrafted and deep features with an optimization-driven fusion strategy. The proposed method begins by extracting low-frequency structural textures via a four-level Discrete Wavelet Transform (DWT), capturing multiscale spatial patterns from grayscale PV imagery. Concurrently, class-specific Bag-of-Visual-Words (BoVW) histograms are generated from Speeded-Up Robust Features (SURF) through modular k-means clustering, producing compact, mid-level descriptors. In parallel, high-level semantic embeddings are obtained from the avg_pool layer of a fine-tuned ResNet50 model. Each feature stream is dimensionally aligned using Principal Component Analysis (PCA), projecting all representations into a unified 300-dimensional latent space. A Genetic Algorithm (GA) then learns optimal attention weights by minimizing the five-fold cross-validation error of a Linear Discriminant Analysis (LDA) classifier, enabling a discriminatively guided and adaptive fusion process. Furthermore, by linking early and accurate detection of dust, cracks, and surface defects to timely cleaning or maintenance of PV modules, GA-HyFusionNet directly minimizes energy yield degradation and reduces downtime, thereby improving operational efficiency and energy sustainability. The framework achieves a peak accuracy of 94.35 % with both LDA and Linear SVM, outperforming deep-only baselines while maintaining low model complexity. Importantly, it requires only 19–21 ms per image at inference and a compact deployment size of ∼87–107 MB, ensuring real-time feasibility on edge devices. By recovering up to 5 % of annual PV energy losses through early and reliable fault diagnosis, GA-HyFusionNet not only reduces the Levelized Cost of Electricity (LCOE) but also advances the practical realization of Sustainable Development Goal 7: Affordable and Clean Energy. Its interpretable architecture, energy-aware design, and real-time feasibility position it as a valuable contribution to next-generation solar diagnostics.