An Improved Hybrid Neural Network Ultra-short-term Photovoltaic Power Forecasting Method Based on Cloud Image Feature Extraction

The photovoltaic power sequence is affected by a variety of characteristic factors, showing a high degree of randomness and volatility. Unlike distributed photovoltaics, centralized photovoltaics have the same geographic location and irradiation level. The occlusion of sporty clouds often leads to minute-level fluctuations in photovoltaic power, which poses a challenge to the accuracy of photovoltaic power prediction. To solve the above problems, an improved hybrid neural network ultra-short-term photovoltaic power prediction method was proposed in this paper, based on cloud image feature extraction. First, by extracting and matching local feature descriptors of color cloud images, a cloud trajectory tracking method based on ground-based cloud images was proposed; Secondly, in order to evaluate the ultra-short-term irradiance changes caused by the moving cloud clusters, an irradiance coefficient prediction model based on cloud trajectory tracking was established. Furthermore, in order to characterize the inherent correlation of each feature sequence, an IAM-CNN-LSTM hybrid neural network was proposed for ultra-short-term photovoltaic power prediction. On this basis, this paper combined weather type and volatility cluster identification, extracted the power fluctuation process and established an error correction model to further improve the prediction accuracy. The data of a centralized photovoltaic power station in Northwest China was used for verification. The results show that the method proposed in this paper can effectively improve the prediction accuracy and has certain engineering practical value. © 2021 Chin. Soc. for Elec. Eng.