Abstract: To address the challenges in multi-station residential load forecasting, including parameter-sharing modeling across multiple stations, exogenous variable coupling, and multi-scale fluctuations, this paper proposes a multi-source and multi-scale load forecasting method based on pre-trained large language model (PLLM). A multi-station load data matrix under a unified time axis is constructed, and single-station load forecasting samples are generated via the sliding-window approach. Most parameters of the pre-trained model are frozen so that temporal dependencies of loads can be captured with a small number of trainable parameters. Natural language templates are employed for semantic encoding of timestamps and temperature data, so as to integrate exogenous variable information into the forecasting model. Multi-scale load token representations are introduced to characterize periodic patterns and local fluctuations. The proposed method is validated using the Guangxi residential load dataset (China) and the Ausgrid load dataset (Australia), and compared with representative models including ARIMA, LSTM, PatchTST and AutoTimes. Results show that the proposed method outperforms ARIMA, LSTM, PatchTST, AutoTimes and other typical models in terms of MSE and MAE. Compared with the best-performing baseline model on each dataset, the MSE decreases by 4.8% and 6.6% on the two datasets respectively, which effectively improves the accuracy of multi-station residential load forecasting and the representation capability for complex variation patterns.