Unsupervised domain adaptation framework for photovoltaic power forecasting using variational auto-encoders

The global transition towards renewable energy sources necessitates accurate forecasts of such energy sources for efficient grid management. While deep learning models offer effective solutions for intermittent renewable energy forecasts, they face challenges due to their inherent data intensity. Transfer learning methods have emerged as valuable tools to address such challenges. However, existing transfer learning frameworks used in renewable energy forecasting, require a significant amount of labelled training data for fine-tuning and knowledge transfer, limiting their applicability to scenarios where abundant data are available. This paper introduces a domain adaptation framework that enables seamless knowledge transfer from forecasting models trained with abundant data to models that need to be trained without labelled data. The proposed domain adaptation framework, leverages variational inference techniques to align feature spaces between source and target domains, utilizing a generative variational auto-encoder architecture. Experimental validation across solar parks with varying configurations demonstrates the replicability and adaptability of the proposed method. This research underscores the enduring potential of domain adaptation in advancing photovoltaic power forecasting while providing valuable insights into overcoming challenges in transfer learning-based renewable energy forecasting. • Transfer Learning framework based on unsupervised domain adaptation. • Stochastic latent feature space generation using Variational Auto Encoder (VAE). • Domain adaptation framework based on feature space regularization. • Effectiveness of proposed framework explored through various sets of comparative experiments.