The widespread global adoption of solar energy requires efficient and intelligent fault detection techniques to guarantee the operational sustainability of the installed infrastructure. Visually inspecting solar cell structures to detect faults is not feasible at a global scale. Typically, state-of-the-art CNN architectures may fail to adequately handle the poor quality of the thermographic images and may not correctly classify minute anomalies. Therefore, to address the said problems, we introduce the concept of an attention-driven multi-model fusion technique for automatic solar cell fault detection from drone-captured thermographic images. This technique uses three heterogeneous CNN architectures with transformers as parallel feature extractors with a 16-head attention mechanism to fuse the extracted features adaptively. This unique fusion technique not only facilitates focused extraction of key spatial and semantic information from diverse CNN architectures but also incorporates their global dependencies for accurate anomaly classification. Performance analysis of this technique was accomplished with the InfraredSolarModules dataset of 20,000 images of solar cells with 12 types of faults. This technique reported a high accuracy of 79.34%, outperforming existing CNN architectures and confirming the efficiency of our unique fusion technique that employs the power of attention for feature fusion. This study validates that by employing complementary learning and attention for feature fusion, accurate fault detection can be achieved by a simple, scalable method for intelligent and instantaneous global solar energy farming system analysis.