The intermittent and fluctuating solar irradiance makes photovoltaic (PV) power generation unstable, which brings great challenges to the power grid system. Existing deep learning-based PV power generation prediction models usually employ Long Short-Term Memory (LSTM) or Convolutional Neural Networks (CNNs) to predict the future PV power generation. However, the recurrent structure-based methods usually failed to model the long-term dependence due to the accumulated error caused by backpropagation through time. To overcome these issues, an improved Transformer-based prediction model, which is called GLformer, is proposed in this work. Specifically, the proposed GLformer decomposes an input series into global pattern subsequences and local pattern subsequences, where the global pattern subsequences represent the trends and seasonality and the local pattern subsequences are used to describe the fluctuations. The global and local features are extracted and fused using a Transformer-based global-local feature extraction module, which enhances the model’s capability of modeling long-term dependence, and obtaining complete time features for prediction. Extensive experimental results show that the proposed method is superior to most SOTA forecasting models.