Solar events originate from the sun and normally manifest within its solar system. Solar flares release lethal radio-active energy into space and have been theorized to trigger coronal mass ejection (CME) with lethal terrestrial impacts. The National Aeronautics and Space Administration (NASA)’s Solar Dynamics Observatory(SDO) has captured high-quality images of solar flares that can be used for solar flare studies. Expert physicists struggle with the classification of the images due to their large data size. Convolutional Neural Networks (CNNs) have proven effective with large datasets. This study improves the classification accuracy of solar flares observed at wavelength 1,600  \({\mathop {A}\limits ^{\circ }}\) from 94% to 100% by utilizing an ensemble of diverse CNN models that are deliberately set up to be more effective than the best base learners. Data from the Atmospheric Imaging Assembly(AIA) containing three classes of solar flares: limb, two ribbon, and compact, was used. A ‘background’ class was included for images with no solar flare. The CNN models used are AlexNet, LeNet5, XceptionNet, NASNetLarge, and ResNet50. The best CNNs were AlexNet, LeNet5, and NASNetLarge with f1-scores and accuracies of 99%. The Soft and Hard voting ensembles outperformed the best individual CNNs with an f1-score of 100% and an accuracy of 100%. The conclusion which supports previous literature from other domains, is that practically demonstrable base learner diversity improves an ensemble’s performance.