To address the problem of low prediction accuracy due to the influence of multidimensional factors such as weather factors and calendar information on air conditioning load and the difficulty in sufficiently extracting the time series characteristics of load data, an air conditioning load prediction model for public buildings using long short-term memory（LSTM）recurrent neural networks based on an adaptive sliding window is proposed. The model first analyzes the influencing factors of air conditioning load in public buildings. Considering that traditional time series prediction models often perform poorly when dealing with non-stationary data, an adaptive sliding window mechanism is innovatively introduced. This mechanism can dynamically adjust the window size to better capture the variations in temperature and historical air conditioning load data, thereby improving the effectiveness of data preprocessing. Furthermore, given the complexity and long-term and short-term dependencies of air conditioning load variations, a multi-layer LSTM network architecture is designed to achieve accurate prediction of air conditioning load in public buildings. Taking the load data of a specific region as an example, proposed model achieves higher fitting ability and better prediction results when an appropriate sliding window size is selected.