Power load forecasting is the basis of power system development planning and power generation plan. The load data of power grid is huge

complicated in structure and diverse in statistical scope

and the factors affecting load change are changeable. The large-scale access of new energy further increases the difficulty of power load forecasting. A multi-level load forecasting method of power system based on improved long short term memory（LSTM）is proposed

which establishes the time series relationship of multi-level load at the provincial

municipal and substation levels

takes historical load data

meteorological data and regional economic data of different load levels as input of the forecasting algorithm

and classifies the load factors at each level. Constraints are added from the perspective of planning development

and the three-layer stacked neural networks prediction model based on improved LSTM is used to complete the overall prediction of each level of load. The simulation example is based on the actual power load data and PV output data of S province and Y city in East China. The results show that the proposed method has a good effect on improving the prediction accuracy of multi-level power load.