• The LSTM-based machine learning soiling model is an extensively approach for technical capability to detect, quantify, and predict signs of PV soiling with high sensitivity. • The sophisticated technique is to identify the integral distribution of accumulated dust mass over time ranging from 0.1 to 0.68 g/m² per week due to soiling, • The soiling index of PV cells revealed that the decrease in light transmission around 0.43% per week due to the accumulated mass dust has led to a power loss of 156 W, corresponding up to 25% of energy production. • The data-driven ML approach provides recommendations tailored to maintenance activities and optimize cleaning frequency in order to minimize overall economic losses, • The model is applicable in any tropical climate zone and is suitable for certain semi-arid regions with high-reliability, critical PV systems. Photovoltaic soiling poses a significant threat to photovoltaic installations, massively increasing sunlight transmission losses and degrading photovoltaic energy production. This article develops a comprehensive machine learning-based approach to extensively model losses due to soiling in order to improve PV electrical performance characteristics. The core features of the algorithms, including deep learning neural network, feedforward neural network, long short-term memory network, and autoencoders, enabled to capture all integrated data while identifying the best-performing model. The LSTM-based machine learning soiling model primarily aims for technical capability to detect, quantify, and predict signs of PV soiling with high sensitivity. It can be applied in any tropical climate zone and is suitable for certain semi-arid regions with high-reliability, critical photovoltaic systems. Furthermore, this model offers impeccable efficiency and accuracy for controlling and monitoring even the smallest, unacceptable losses due to dust accumulation. The enhanced LSTM-based predictive model incorporates meteorological data such as temperature, solar radiation, precipitation, relative humidity, and wind speed. Furthermore, environmental variables such as dust particle diameter and size, fine particle concentration and sedimentation rate, as well as particle interactions and movements, were also considered. Photovoltaic parameters such as current, voltage, power, tilt angle, module fill factor, and temperature were accurately considered during the modeling development. The PV soiling model is developed using genuine operational data covering large period from January 1, 2019, to July 31, 2024, collected from the Malaysian Meteorological Department facilities, at 24-hour intervals to improve the training accuracy. The proposed approach identifies the integral distribution of accumulated dust mass over time, while estimating the quantity of light transmission lost due to soiling rate of the PV panels. Based on predictions of drop in light transmission, the degradation of electrical characteristics—specifically short-circuit current, open-circuit voltage, and maximum power—was evaluated in relation to its impact on the PV system's performance. As a result, the model of dust deposition impact by particle layers estimated an accumulated dust mass density ranging from 0.1 to 0.68 g/m² per week. The soiling index of PV cells revealed that the decrease in light transmission predicted around 0.43% per week due to the accumulated mass dust, has led to a power loss of 156 W, corresponding up to 25% of energy production. The hybrid algorithm configuration in this approach provides a solid technical foundation for completely addressing soiling pollution in photovoltaic installations. The immediate implementation of this technology through machine learning integration significantly contributes to the efficient improvement of sustainable PV energy production.
Said et al. (Sun,) studied this question.