Meteorological Input Selection for Cooling Load Forecasting in a Large Public Building: A Case Study

Cooling electricity consumption in central air-conditioning systems of large public buildings accounts for a substantial share of urban electricity use and is strongly influenced by outdoor meteorological conditions. Under increasingly frequent extreme summer heat events, accurate cooling-load forecasting is important for HVAC operation, building energy management, urban electricity security, and power-system planning. This study investigates the effects of measured outdoor meteorological inputs on cooling-load forecasting for a large public building in Guangzhou. Consecutive hourly cooling-load data and measured meteorological data, including outdoor air temperature, relative humidity, solar radiation, wind speed, and wind direction, were collected from June to September 2022. The corresponding 2023 dataset was analyzed separately using the same modeling and evaluation procedure to assess cross-year repeatability; data from the two years were not combined. Correlation and univariate linear regression analyses were first used for preliminary candidate-input screening. Nine Long Short-Term Memory sub-models with different meteorological input combinations were then developed and compared using the 2022 dataset, and the selected input configuration was subsequently re-evaluated using the separate 2023 dataset. Solar radiation exhibited the strongest marginal association with cooling load, followed by outdoor air temperature and relative humidity. The negative association of relative humidity reflected its coupled variation with temperature and solar radiation during the investigated summer period. For the 2022 dataset, the model using outdoor air temperature, relative humidity, and solar radiation achieved the lowest MAPE. Compared with the model using all five meteorological variables, it reduced MAE, RMSE, and MAPE by 14.55%, 7.24%, and 19.07%, respectively, while R2 increased from 0.9542 to 0.9601. Evaluation using the 2023 dataset showed corresponding reductions of 20.01%, 18.37%, and 25.81% in MAE, RMSE, and MAPE, respectively, together with an increase in R2 from 0.9592 to 0.9708.