Evaluation of photovoltaic panel dust deposition based on CNN-BiLSTM-Attention method

Accurate assessment of the dust deposition state on photovoltaic panels is an essential prerequisite for mitigating power generation losses. To address the inherent limitations of high hardware costs and poor generalization capabilities in existing image-based or manual inspection methods, this paper proposes a data-driven evaluation approach for dust deposition density that integrates physical model correction with power prediction. Under controlled operating conditions, an empirical linear correlation was established between dust deposition density and the degradation of photovoltaic output power. This was accomplished by integrating dust deposition density into an idealized photovoltaic array model, which allowed for the adjustment of both the effective irradiance and the panel surface temperature. The Fuzzy C-Means clustering algorithm was employed to classify historical operational data into three representative weather regimes: sunny, cloudy, and rainy conditions. On this basis, a hybrid CNN-BiLSTM-Attention neural network was trained using weather parameters and power output data as input features to develop predictive models for clean photovoltaic power output under each weather regime. The severity of dust accumulation was then quantitatively evaluated by calculating the deviation between the real power and the predicted clean power baseline. The results indicate that the Mean Absolute Percentage Error (MAPE) of the baseline prediction model is 3.87%. Under sunny and cloudy conditions, the MAPE of the proposed dust assessment method is 6.28% and 11.78% respectively, demonstrating high accuracy within a hardware-free framework. Relying solely on routine operational data, this method enables effective evaluation of the dust deposition state, providing practical guidance for the low-cost operation and maintenance of photovoltaic power plants.