To address the problem that the existing short-term power load forecasting methods have difficulty in effectively modeling non-linear structures and lack cross-scale and cross-variable interaction capabilities， a short-term power load forecasting model based on Dynamic Convolution Decomposition and Multi-Scale Graph （DCDMSG） was proposed. First，to cope with the complex non-linear structure， a dynamic convolution decomposition method was adopted to decompose the trend and seasonal items in the load sequence. Second， the trend item was forecasted using a linear layer directly， and for the seasonal item， a multi-scale sequence was constructed for deep modeling using an adaptive multi-scale sequence construction method. Third， to model the complex dependencies within and outside the multi-scale sequence， a multi-scale fusion graph was used to capture dependencies within sequences at different scales， a multi-variable correlation graph was utilized to model correlations between variables， and a multi-scale mix-hop propagation was employed to aggregate features in the graph. Finally， the seasonal items were forecasted， and the obtained forecast results were weighted and fused with the trend items' forecast results to obtain the final forecast values. Experimental results show that， on the Australian dataset， the Mean Absolute Error （MAE） and Root Mean Square Error （RMSE） of DCDMSG reached 0.379 MW and 0.529 MW， respectively， which were reduced by 7.11%-21.37% and 9.88%-26.32% compared with all comparison models. On the Cele dataset， DCDMSG achieved an MAE of 0.437 MW and an RMSE of 0.708 MW， representing decreases of 5.61%-24.66% and 5.22%-24.12% over all comparison models. In the short-term power load forecasting task， DCDMSG effectively improves the forecast accuracy through seasonal-trend decomposition as well as cross-scale and cross-variable modeling.