<![CDATA[Electrical energy consumption in the smart building sector continues to rise with the growth of urbanization and IoT adoption. Accurate electrical load forecasting is crucial for efficient and sustainable energy management. This study aims to analyze and compar the performance of two modern gradient boosting algorithms, LightGBM and CatBoost, in predicting hourly electricity consumption at smart building network scale. The dataset used is the PJME Hourly Energy Consumption dataset, comprising 145,198 hourly observations recorded over a 16-year period from 2002 to 2018. The methodology involves temporal feature engineering through the extraction of calendar components (hour, day_of_week, month, quarter, day_of_year, is_weekend) and lag features (lag_1h, lag_24h, lag_1week) to equip the models with historical pattern memory. Model validation was conducted using a chronological 80:20 data split and Time-Series Cross-Validation (k=5) to ensure robust generalization. Experimental results demonstrate that LightGBM outperforms across all evaluation metrics with RMSE = 417.11, MAE = 312.72, MAPE = 0.99%, and R² = 0.9959, while also being computationally more efficient with a training time of 2.00 seconds compared to CatBoost at 4.97 seconds. Both models surpassed the R² accuracy threshold of 99%, validating the effectiveness of the proposed time-series-based feature engineering framework. This research contributes a reliable and efficient energy prediction framework for the implementation of energy management systems in smart buildings. ]]>
Copyrights © 2026
