A novel deep learning-based method for theoretical power fitting of photovoltaic generation

Studying the effects of dust accumulation on photovoltaic (PV) power generation are crucial for improving PV system efficiency. However, evaluating power losses resulting from dust accumulation on PV modules is highly influenced by the precision of the theoretical power calculation. This paper proposes a novel deep learning-based method for theoretical power fitting of photovoltaic generation: First, the power generation data from both clean and dust-accumulated panels of the same model are collected under the same environment from an existing campus PV power generation system, eliminating the influence of factors other than dust accumulation. Then, a time-series generative adversarial network based on gated recurrent neural network is used to enhance the original data. Subsequently, this paper innovatively proposes using a convolutional neural network autoencoder to fit the theoretical power. Leveraging the advantages of convolutional structures in short-term local cross-feature extraction, the encoder compresses high-dimensional features into low-dimensional abstract features. The decoder maps these low-dimensional features back to high-dimensional outputs for accurate power fitting. Experiments under four different weather conditions indicate that, compared to the formula-based method, the proposed approach improves MAE by 16.4 %–52.6 %, 36.5 %–69.5 %, and 41.3 %–75.3 %, respectively. This provides a new perspective for applying deep learning in theoretical PV power fitting.