Abstract: Traditional energy-saving operation optimization methods for thermal power units under deep peak shaving conditions have not fully considered the strong nonlinearity and time-varying characteristics of unit energy consumption, resulting in a large root mean square error in load tracking. Design an energy-saving operation optimization model for thermal power units under deep peak shaving conditions based on LSTM-DP algorithm. Build an energy-saving operation optimization model for deep peak shaving conditions. Design and improve long short-term memory networks as dynamic energy consumption proxy models. Embedding the improved long short-term memory network proxy model into the recursive solution process of dynamic programming, and achieving simultaneous prediction and optimization through solving the energy-saving operation optimization model for deep peak shaving conditions. The example test results show that under the design method, the cumulative coal consumption has been reduced by 26.47 tons, and the average coal consumption rate has been reduced by 13.98 g/(kW · h). The RMSLTE of the design model is less than 2MW, indicating that under the design model, the response of the unit to the grid load command is accurate, and the additional energy consumption caused by response deviation is relatively small.