Optimizing dust accumulation quantification on photovoltaic panels using deep learning visual models with hyperparameter optimization

The increasing integration of solar photovoltaic (PV) systems is driven by their cost-effectiveness and sustainability. Nonetheless, dust accumulation significantly reduces PV performance, especially in arid regions like the UAE. This study investigates multiple deep learning architectures, deep residual neural network (DRNN), vision transformer, ResNet-50, and EfficientNet-B7 for accurate dust quantification on PV panels. A diverse indoor imaging dataset is generated with varying zoom lengths (18–200 mm) and dust concentrations (1g–400g). Preprocessing techniques, such as silver line removal, enhance image quality, while model performance is evaluated using mean absolute error (MAE), mean squared error (MSE), and loop error coefficient. DRNN demonstrates superior accuracy in the indoor imaging dataset and is subsequently evaluated on two additional datasets: outdoor drone-captured images (from 4 to 30m heights) and a combined indoor-outdoor dataset. Its performance is further improved through hyperparameter optimization techniques such as Bayesian optimization, particle swarm optimization, genetic algorithm and hyperband. Bayesian optimization excels in indoor and outdoor datasets, while hyperband efficiently balances resources for the combined dataset, enhancing dust estimation and PV maintenance planning.