Knowledge Extraction From PV Power Generation With Deep Learning Autoencoder and Clustering-Based Algorithms

The unpredictable nature of photovoltaic solar power generation, caused by changing weather conditions, creates challenges for grid operators as they work to balance supply and demand. As solar power continues to become a larger part of the energy mix, managing this intermittency will be increasingly important. This paper focuses on identifying daily photovoltaic power production patterns to gain new knowledge of the generation patterns throughout the year based on unsupervised learning algorithms. The proposed data-driven model aims to extract typical daily photovoltaic power generation patterns by transforming the high dimensional temporal features of the daily PV power output into a lower latent feature space, which is learned by a deep learning autoencoder. Subsequently, the Partitioning Around Medoids (PAM) clustering algorithm is employed to identify the six distinct dominant patterns. Finally, a new algorithm is proposed to reconstruct these patterns in their original subspace. The proposed model is applied to two distinct datasets for further analysis. The results indicate that four out of the identified patterns in both datasets exhibit high correlation (over 95%) and temporal trends. These patterns correspond to distinct weather conditions, such as entirely sunny, mostly sunny, cloudy, and negligible power generation days, which were observed approximately 61% of the analyzed period. These typical patterns can be expected to be observed in other locations as well. Identified PV power generation patterns can improve forecasting models, optimize energy management systems, and aid in implementing energy storage or demand response programs and scheduling efficiently.