[Objective] As the core driving factors for short-term power load forecasting， multivariate meteorological elements still confront prominent constraints in the acquisition and practical application of refined meteorological data.Accordingly， this paper investigates meteorological Agents empowered by large language models （LLMs） and applies them to short‑term load forecasting. [Methods] First， a workflow agent integrated with large language models is constructed via the LLM development platform， which realizes the dynamic acquisition and intelligent parsing of structured meteorological information. This approach effectively improves the real-time response performance of the forecasting model and reduces the development and application complexity of short-term load forecasting systems. On this basis， a multimodal input framework for power system short-term load forecasting is established by integrating meteorological factors， power load time-series data， and calendar features. Combined with the inherent characteristics of time series and the gradient boosting tree algorithm， an optimized short-term power load forecasting model is further developed.Finally， the feasibility and effectiveness of the proposed model and algorithm are verified through experimental validation and case analysis. [Results] The proposed model and algorithm streamline the meteorological information‑acquisition process and improve the forecasting accuracy. [Conclusions] This research provides a novel technical idea for the low-cost acquisition and efficient utilization of meteorological data in power system operation and analysis.