Deep learning-based image recognition method for on-demand defrosting control to save energy in commercial energy systems

[Display omitted] •An image recognition-based defrosting control method is proposed to save energy.•Time-consuming experiments and laborious labeling workload are greatly reduced.•Recognition accuracy of proposed model is 5.5% higher than conventional CNN model.•Defrosting frequency and energy consumption reduce by 31.68% and 42.92%.•The defrosting control method can be extended to homotypic or heterotypic devices. Periodical defrosting is essential to restore the initial capability of heat exchangers and improve the operating efficiency of commercial energy systems. The defrosting control method applying image recognition technology is considered a low-cost and easy-operating approach to implement demand-based defrosting cycles. However, with regard to different operation environments, the establishment of high-accuracy image recognition model will consume plenty of labor and cost, which leads to the low extensibility between homotypic or heterotypic devices and severely limits its practical application in commercial energy systems. To this end, a novel deep learning-based image recognition method was presented for the extensible implementation of on-demand defrosting control. In order to improve the recognition accuracy, a convolutional neural network (CNN) model was proposed to extract in-depth and complicated features for frosty state detection. By integrating deep clustering and image augmentation, the time-consuming experiments and labor-intensive labeling workload were greatly reduced. The recognition accuracy of proposed CNN model was on average 5.50% higher than that of conventional CNN model, and the recognition accuracy was further increased to 97.57% through the hyperparameters optimization. Based on the trained CNN model, a defrosting control method was proposed for the on-demand defrosting control according to the real-time frosty state recognition. Compared with original time-based control method (defrosting per device after a fixed interval), the field experiment testified that the defrosting frequency, accumulated time and energy consumption were decreased by 31.68%, 65.83% and 42.92% respectively by adopting the proposed control method. The economic and environmental analysis indicated that the payback time was approximately half a year and annual reduction in CO2 emission is 28.44 t, which signified the great application potential of the proposed method. This study will shed light on the further application of image recognition technolog