A Two-Stage Ultra-Short-Term Photovoltaic Power Forecasting Framework Using I-GPT and IP-U-Net = I-GPT와 IP-U-Net을 이용한 2단계 초단기 태양광 발전량 예측 프레임워크

다국어 초록 (Multilingual Abstract)

Solar forecasting using ground-based sky image offers a promising approach to reduce uncertainty in photovoltaic (PV) power generation. However, existing methods often rely on deterministic predictions that lack diversity, making it difficult to capture the inherently stochastic nature of cloud movement. To address this limitation, we propose a new two-stage probabilistic forecasting framework. In the first stage, we introduce I-GPT, a multiscale physics-constrained generative model for stochastic sky image prediction. Given a sequence of past sky images, I-GPT employs a Transformer-based vector quantized latent architecture, multi-scale physics-informed recurrent units (Multi-scale PhyCell), and dynamic weighted fusion of physical and appearance features to generate multiple plausible future sky images with realistic, coherent cloud motion. In the second stage, these predicted sky images are fed into an Image-to-Power U-Net (IP-U-Net) to produce 15 minute ahead probabilistic PV power forecasts. In experiments using our dataset, the proposed approach significantly outperforms deterministic, other stochastic, multimodal, and smart persistence baselines models, achieving a superior reliability sharpness trade-off. It attains a Continuous Ranked Probability Score (CRPS) of 2.912 kW and a Winkler Score (WS) of 33.103 kW on the test set and CRPS of 2.073 kW and WS of 22.202 kW on the validation set. Translating to 35.9% and 42.78% improvement in predictive skill over the smart persistence model. Notably, our method excels during rapidly changing cloud cover conditions. By enhancing both the accuracy and robustness of short-term PV forecasting, the framework provides tangible benefits for Virtual Power Plant (VPP) operation, supporting more reliable scheduling, grid stability, and risk-aware energy management.

목차 (Table of Contents)

• Content
• List of Figures ⅳ
• List of Tables vi
• Abstract vii
• 1. Introduction 1
• 2. Related Works 8
• 3. Data and preprocessing 12
• 3.1 Data 12
• 3.1.1 Sky images 12
• 3.1.2 PV power 13
• 3.2 Data processing 14
• 3.3 Data partition 16
• 4. Methodology 18
• 4.1 Sky image prediction 19
• 4.2 PV output prediction 25
• 4.3 Loss Function 27
• 4.3.1 Loss function of sky image prediction 27
• 4.3.2 Loss function of PV output prediction 29
• 5. Experimental Setup 31
• 5.1 Evaluation metrics 31
• 5.1.1 Evaluation metrics of sky image prediction 31
• 5.1.2 Evaluation metrics of PV output prediction 32
• 5.2 Implementation Details 35
• 5.2.1 Implementation details of sky image prediction 35
• 5.2.2 Implementation details of PV output prediction 36
• 6. Results 38
• 6.1 Phase 1: proposed sky image prediction model 38
• 6.1.1 Proposed sky image prediction model 38
• 6.1.2 Comparison between the proposed sky image prediction modelwith the baseline models 39
• 6.1.3 Generation diversity of the proposed sky image prediction model 43
• 6.1.4 Ablation experiments for the proposed sky image prediction model 46
• 6.2 Phase 2: proposed PV output prediction model 47
• 6.2.1 Proposed PV output prediction model 47
• 6.2.2 Comparison between the proposed PV output prediction model with the baseline models 50
• 6.2.3 Ablation Experiments for the proposed PV output prediction model 57
• 7 Conclusion and Future Plans 60
• 7.1 Conclusion 60
• 7.2 Future Plans 60
• References 62
• Abstract in Korean 69
• Acknowledgement 71