Abstract: Smart grid construction has put forward higher requirements for low-voltage distribution network load analysis. Taking the 10 kV Taiping Road network structure optimization project of State Grid Beijing Fengtai Power Supply Company as the background, a dynamic load analysis framework based on real-time data collection and intelligent algorithms was constructed. By deploying 25 pole-mounted vacuum circuit breakers and 48-core overhead optical cables, a multi-dimensional load feature recognition model was established. A hybrid prediction strategy combining LSTM deep learning network and traditional statistical methods was adopted to achieve accurate load forecasting in commercial building intensive areas. System operation verification shows that the MAPE of load forecasting reaches 7.2%, which is 17.8% higher than the traditional method, saving 8.3% of annual energy loss, with an investment payback period of 3.2 years, providing important technical support for the optimal operation of distribution network in smart grid environment.