Multimodal Learning Techniques for Time Series Forecasting in Renewable Energy Systems: A Comprehensive Survey

Renewable energy systems, such as solar, wind, and hybrid sources, present complex forecasting challenges due to their stochastic and weather-dependent nature. The growing availability of heterogeneous and complementary data modalities including meteorological forecasts, satellite imagery, numerical sensor streams, and grid interaction logs has motivated the application of multimodal learning techniques to improve time series forecasting accuracy. Unlike previous surveys that either focused narrowly on single-modality forecasting or provided only high-level overviews, this paper offers a focused and technically grounded review of multimodal learning strategies specific to renewable energy forecasting. We categorize and compare key fusion strategies (early, late, hybrid), analyze deep model architectures (e.g., co-attention and transformer-based fusion), and discuss deployment-related challenges such as data alignment, modality missingness, and interpretability. Then, this paper reviews multimodal learning techniques based on four main aspects, i.e., classify fusion techniques, examine model architectures, showcase practical application scenarios, and summarize benchmark datasets. Finally, the paper identifies existing research gaps and outlines promising future directions in this rapidly evolving field, such as self-supervised multimodal learning and adaptive fusion under uncertainty.