Optimization of Tourism Consumption Prediction Model Integrating LSTM-Attention

This paper addresses seasonal structural drift and window-selection limitations in tourism consumption forecasting by proposing a BiLSTM-Att+PE model to capture dynamic seasonal patterns and improve consumption-pattern representation. The model combines BiLSTM for bidirectional temporal dependency learning, Bahdanau attention for dynamic focus on key time steps such as holidays and emergencies, and learnable positional encoding for seasonal information modeling. It is trained end-to-end on provincial tourism economic data from 2001Q1 to 2024Q4, including consumption, visitor numbers, GDP, income, holidays, and epidemic markers. Four sample sets are constructed through sliding windows with T ∈{1, 2, 3, 4} quarters. The model is optimized using MSE loss and AdamW, while Bayesian search determines the best hyperparameters, including a learning rate of 3.4 × 10−4, 128 LSTM units, and 64 attention dimensions. The optimal performance is achieved with a two-quarter window, reaching RMSE of 21.3 billion and R2 of 0.962, outperforming seven benchmarks. Attention visualization confirms adaptive focus on holiday and epidemic events, providing an interpretable and computationally efficient framework for nonlinear time-series forecasting.