Traditional cooling load prediction methods often neglect temporal factors, leading to poor prediction performance and low generalization ability. To address these issues, a long short-term memory (LSTM) neural network prediction model optimized by the weight decay adaptive moment estimation algorithm (WAdam-LSTM) was proposed for high-precision building cooling load prediction. The cross-correlation function was employed to identify optimal feature variables as model inputs. WAdam-LSTM was constructed by introducing a weight decay term during the update of LSTM variable parameters. Hourly cooling load data from two representative large commercial buildings were used to evaluate and compare model prediction performance. Results show that the incorporation of the decoupled weight decay term enhances the stability and convergence of the optimization algorithm, making it suitable for LSTM network parameter optimization. WAdam-LSTM demonstrates superior prediction accuracy compared to SVR, SCOA-LSTM, and Adam-LSTM, with mean square error reductions of 83%, 66%, and 30%, respectively. WAdam-LSTM exhibits stronger generalization ability than single models (LSTM, SVR, and BPNN) and hybrid prediction models (SCOA-LSTM and Adam-LSTM), enabling precise cooling load predictions for different commercial buildings across varying months.