Automatic Multiclass Classification of Unlabeled Ground-Based Sky Images for Minute-Scale PV Energy Yield Forecasting

Ground-based Sky Images (GSIs) are widely recognized as one of the most promising inputs for intra-hour minute-scale photovoltaic (PV) generation forecasting. Given the ongoing challenges in capturing short-term solar power variability, GSI classification has emerged as a data-centric approach with the potential to create more informative inputs for PV generation forecasting models and to increase their robustness against imbalanced GSI datasets. However, preassigning ground-truth labels to GSIs is a very challenging task that limits the effectiveness and scalability of existing classification methods. This paper introduces a novel unsupervised learning method for multiclass classification of unlabeled GSIs. The proposed method combines enhanced handcrafted feature extraction with hybrid auto-encoders to create feature sets that fully represent GSIs and capture PV energy yield variations. GSI clusters are generated through a novel forecast-driven k-means clustering approach that integrates downstream evaluation with co-optimization. Classification performance is evaluated based on its impact on PV energy yield forecasts generated by a best-in-class minute-scale forecasting model. Experimental results from two case studies demonstrate the feasibility of automatic multiclass GSI classification that improves minute-scale PV energy yield forecasting without requiring preassigned ground-truth labels or computationally intensive methods.