UDA-MTVIB: An unsupervised domain-adaptive multi-site temporal variational information bottleneck for generalizable wind power forecasting

Abstract

Distributional shifts across wind farms and non-stationary dynamics pose significant challenges to accurate and generalizable wind power forecasting. This paper proposes UDA-MTVIB, an unsupervised domain-adaptive multi-site temporal variational information bottleneck framework. The model integrates a dynamic-gated LSTM encoder with uncertainty-aware gating to regulate temporal memory updates under domain shift. A dual variational information bottleneck is employed to (i) compress task-relevant temporal representations and (ii) regularize domain-related latent factors, which are aligned via adversarial training for robust domain alignment. Unsupervised domain adaptation is achieved by adversarial alignment between labeled source data and unlabeled target-farm inputs, without target power labels. Experiments on WIND Toolkit data across seven U.S. wind farms demonstrate that UDA-MTVIB achieves strong robustness and generalization under cross-farm and cross-year shifts, improving predictive accuracy and stability across multiple metrics. This work provides an information-theoretic perspective for domain-adaptive temporal forecasting in renewable energy systems.