Purpose: Although Long Short-Term Memory (LSTM) models have been widely used for predicting time-series data in building energy applications, studies focusing on cooling load prediction for office buildings remain limited. In particular, the influence of model configuration and data preprocessing strategies on prediction performance has not been systematically investigated. This study aims to evaluate the impact of stepwise model improvement techniques on the prediction accuracy of LSTM-based cooling load forecasting models. Method: Weather data from the Korea Meteorological Administration for 2023~2025 were collected and converted into EPW format. Cooling load data were generated using EnergyPlus for a medium office reference building. An LSTM-based prediction model was developed, and six cases were constructed by sequentially introducing lagged input variables, data normalization, low-load data filtering, and hyperparameter tuning. Prediction performance was evaluated using the coefficient of determination (R² ), coefficient of variation of the root mean square error (CvRMSE), and mean bias error (MBE). Result: Prediction performance improved significantly as the model improvement steps were applied. The initial model showed low accuracy with an R 2 of 0.3955 and a CvRMSE of 139.7%. After stepwise improvements, the final model achieved an R2 of 0.9318 and a CvRMSE of 11.5%, with an MBE of -2.7%. The final model satisfied the ASHRAE Guideline 14 criteria for hourly prediction models, demonstrating the effectiveness of the proposed stepwise model improvement strategy.
Seong et al. (Thu,) studied this question.