Current cooling load prediction method of overall buildings for large-scale shopping malls cannot provide a reasonable control strategy for demands of various areas of the shopping mall. By studying the characteristics of cooling load in different areas of shopping malls, the key influencing factors of cooling load in different areas of shopping malls were screened by using grey relational degree analysis method. To solve the instability of the influence degree of each input variable on cooling load in actual situation, a short-term zoned cooling load prediction model based on double attention mechanism and LSTM was proposed. LSTM network fully considers the nonlinear relationship between air-conditioning cooling load and related characteristic variables. Feature attention analyzes the relationship between historical information and input variables autonomously to extract important features. Sequential attention selects historical information at critical moments of LSTM network to improve the stability of long-term prediction effects. The experimental results show that compared with LSTM model, CNN-LSTM model and attention-LSTM model, the error indexes MAPE and RMSE of the proposed model decrease significantly, and its R2 increases significantly and remains stable above 0.99, indicating good generalization ability and strong stability.