In the context of smart tourism, accurate prediction of educational tourism demand is crucial for the sustainable development of tourism industry. Therefore, this paper constructs an educational tourism demand prediction model (SAE-LSTM) based on self-encoder and long-short-term memory network. The model is pre-trained with layer-by-layer feature extraction and pre-training of tourism education data by stacked self-encoding (SAE) to enhance the model’s ability to learn complex data patterns. Afterwards, the extracted features are fed into the LSTM network and its powerful time series modeling capability is utilized for demand prediction. In addition, an attention mechanism layer is embedded in the model to enhance the model’s attention to important features, thus improving the model’s prediction accuracy. The empirical results show that the SAE-LSTM model proposed in this paper has the best performance, and the MAE, RMSE and MAPE of the SAE-LSTM model on the 2 datasets are reduced by 133.6032-950.1945, 197.1081, 197.1081, -1168.4574 and -1168.4574, compared with the LSTM model, the SAE-BPNN model and the BPNN model, respectively. -1168.4574 and 3.5582-9.3684, and its R2 is improved by 0.0389-0.0989. This shows that the model performs well in educational tourism demand forecasting, with higher accuracy and robustness compared to traditional methods.