Abstract: Power-system time-series forecasting increasingly involves diverse forecasting entities, heterogeneous data sources, and complex operational constraints. Large language models (LLMs) and time-series foundation models (TSFMs) offer new approaches to cross-entity reuse, zero- and few-shot adaptation, multimodal fusion, and forecasting workflow coordination, yet the sources of their gains and their applicability remain unclear. This paper classifies existing methods by their knowledge sources and roles in the forecasting pipeline into text-pre-trained LLMs, LLM-based time-series adaptation methods, native TSFMs, multimodal forecasting models, and tool-enabled agents. These methods are reviewed in terms of capability requirements, key techniques, application scenarios, and evidence of effectiveness. Current evidence suggests that their main value lies in cross-entity reuse, rapid adaptation, the use of external information, and probabilistic forecasting. The independent contribution of language pre-training to purely numerical tasks still requires controlled validation; zero-shot TSFM performance depends on pre-training data and target-domain shifts; and multimodal gains depend on whether external information is genuinely available at forecasting time. Future work should strengthen cross-region and cross-entity validation and jointly evaluate probabilistic calibration, physical and hierarchical consistency, resource costs, and operational value under fixed downstream decision models.