Abstract: Short-term daily load forecasting is an important foundation for power system generation scheduling and operation control. To address the problem of insufficient accuracy of traditional forecasting methods in handling complex nonlinear load patterns, this paper proposes a short-term daily load forecasting method based on the TimesFM foundation model for time series, which integrates temperature and holiday covariates. The method first utilizes the pre-trained TimesFM model to learn the temporal features of the load series, and then incorporates temperature data as dynamic numerical covariates and holiday types as dynamic categorical covariates into the forecasting process via the batch contextual exogenous regression (XReg) mechanism. Experimental validation is conducted using historical simulated load and meteorological data from a certain region from January 2022 to June 2026. The results show that the MAPE of the base TimesFM model is 27.14%; after adding temperature covariates, the MAPE decreases to 6.59% with a relative improvement of 77.2% in MAE; after further adding holiday covariates, the MAPE drops to 6.57%. Temperature covariates are a key factor in improving forecasting accuracy, which verifies the effectiveness of the proposed method.