关键词: 光伏发电, 功率预测, 卷积神经网络, 深度学习, 组合模型

Abstract:

With the global energy structure transitioning towards cleaner and low-carbon sources， photovoltaic （PV） power generation， as an important form of renewable energy， has witnessed a continuous and rapid increase in installed capacity. However， PV power output is significantly influenced by meteorological and environmental factors， exhibiting strong volatility， randomness， and non-stationarity. These characteristics pose severe challenges to the stable scheduling and safe operation of power grids. Accurate PV power prediction is a key technology for addressing these issues and achieving efficient integration of PV power. To systematically review the research progress and identify future directions in this field， deep learning-based PV power prediction techniques are reviewed. The core meteorological factors affecting PV power output are analyzed， including solar irradiance， ambient temperature， wind speed， and weather types. The classification system of PV prediction based on time scales， spatial scales， and modeling approaches is described. Special attention is given to various deep learning models applied in PV prediction. Among them， convolutional neural networks （CNN） and their temporal variants， temporal convolutional networks （TCN）， are capable of extracting local features and handling long-term dependencies. Recurrent neural networks （RNNs） and their improved variants， such as long short-term memory networks （LSTM）， gated recurrent units （GRU）， and bidirectional LSTM （BiLSTM）， exhibit excellent performance in capturing dynamic patterns in time series data. Additionally， Transformer models based on self-attention mechanisms and generative adversarial networks （GAN） used for data augmentation demonstrate great application potential. To address the limitations of single models， hybrid prediction models and optimization strategies are further reviewed. Techniques such as variational mode decomposition （VMD） for signal preprocessing or intelligent algorithms like shuffled frog leaping algorithm （SFLA） and particle swarm optimization （PSO） for optimizing model hyperparameters can significantly improve prediction accuracy and model robustness. Future efforts should focus on developing specialized deep learning architectures tailored to the characteristics of PV data and exploring efficient data preprocessing techniques to mitigate the interference of missing and anomalous data in practical engineering. This will promote the high-quality development of PV power generation in new power systems.

Key words: photovoltaic power generation, power prediction, convolutional neural network, deep learning, combinatorial model