Ultra-Short-Term Photovoltaic Power Forecasting Approach Based on Cloud Features and ViT+LSTM Neural Network

[Objective] To achieve accurate ultra-short-term photovoltaic power forecasting and address the insufficient extraction of cloud-information from ground-based sky images in traditional neural networks， this paper proposes an ultra-short-term photovoltaic power forecasting approach based on cloud features and a vision transformer+long short-term memory （ViT+LSTM） neural network. [Methods] First， an adaptive cloud recognition algorithm using Otsu’s method （OTSU） is adopted to generate high-accuracy binary images of cloud distribution. Second， a hybrid cloud‑motion‑vector algorithm is proposed， combining a similarity‑weighted cloud‑motion approach with the Farneback optical flow method to generate pixel‑level cloud‑displacement matrices. Ground‑based sky images， cloud distribution images and cloud motion matrices are concatenated to generate fused images. Finally， the ViT+LSTM neural network architecture is constructed for photovoltaic power forecasting. The ViT neural network extracts global spatial features from the fused images， and then global spatial features concatenated with historical photovoltaic power and temporal feature data are fed into LSTM neural network to capture temporal dynamic features. [Results] Case studies demonstrate that the approach effectively reduces cloud motion calculation error. The proposed approach achieves a 16.75% reduction in RMSE relative to the baseline model for ultra-short-term forecasting tasks. [Conclusions] The proposed cloud-feature extraction approach successfully extracts explicit cloud features， the proposed neural network architecture significantly outperforms existing models in forecasting performance； the proposed approach validates its accuracy in forecasting photovoltaic power fluctuations under different weather conditions.