A method based on Long Short-Term Memory (LSTM) networks is proposed to forecast hourly energy consumption. Using an office building in Shanghai as a case study, hourly data on occupancy, weather, and energy consumption were collected. Daily energy consumption was analyzed using single-link clustering, and days were classified into three types. The key input variables significantly influencing energy consumption, solar radiation, occupancy, and outdoor dry bulb temperature are identified by the Pearson correlation coefficient. By comparing five algorithms, it was found that the LSTM model performed the best. After considering the occupancy, the hourly MAPE was reduced from 11% to 9%. Accuracy improvements for each day type were noted as 1% for weekdays, 4% for Saturday, and 7% for Sunday. Further analysis indicated that the model started to predict the time (1:00) and commute time (7:00 and 17:00) with large errors. The model was optimized by varying the time step. For the times 1:00, 7:00, and 17:00, the best optimization of the model was achieved when the time step values were set to 6 h, 24 h, and 18 h with an MAPE of 3%, 6%, and 5%, respectively. As the model time step increased (≤2 weeks), the accuracy of the model decreased to 6%.

Forecasting and intelligent operation of HVAC systems in public buildings are crucial for energy conservation, carbon reduction, and enhancing occupant comfort. This area has seen significant progress. However, for intermittent HVAC systems, noise reduction techniques applied to periodic load sequences may distort the load during start–stop transitions. To tackle this challenge, an innovative hybrid prediction method, GCMWSG-HOA-LightGBM, is proposed in this study. The GCMWSG filter utilizes period division and boundary protection mechanisms to avoid cross-period coupling errors in traditional filters, preserving the authenticity of cooling load start–stop transitions. Additionally, the HOA is employed to automatically optimize the hyperparameters of the LightGBM model, enhancing both parameter tuning efficiency and prediction accuracy. The proposed model is validated using operational data from an office building in Foshan, Guangdong Province, and its performance is compared with several hybrid models: HOA-LightGBM, BWO-LightGBM, AO-LightGBM, TPE-LightGBM, HOA-RF, BWO-RF, AO-RF and TPE-RF. The results show that the GCMWSG-HOA-LightGBM model outperforms all comparison models across all metrics, achieving R², MAPE, and CVRMSE values of 92.22%, 7.05%, and 9.46%, respectively, highlighting its accuracy and stability. The findings provide an efficient solution for intelligent load prediction of HVAC systems with intermittent and periodic characteristics, offering valuable engineering insights and theoretical guidance.