In the context of accelerating the construction of a new power system, improving the accuracy of load forecasting is an important measure to ensure the economic, safe and stable operation of the power system, and it is also the key to promote the development of smart grid. In order to enhance the ability of regional load forecasting, a short-term load forecasting method based on dynamic adaptive meteorological impact factors is proposed. Firstly, a load/meteorological information fusion module based on parallel multi-scale temporal convolutional neural networks is employed to mine the multi-time period change model of historical load and regional weather forecast. Then, a dynamic identification module of meteorological factors based on depth-gated residual neural network is proposed. By dynamically adjusting the feature contribution and optimizing the feature selection, the fusion of feature weights of different spatio-temporal scales is enhanced, and the ability of the model to extract key features is improved. Finally, the load data of a region in Beijing, Tianjin and Hebei are used as an example to prove that the proposed regional load forecasting method has higher forecasting accuracy and better tracking effect on regional load trend changes.