The impact of irregular days on machine learning predictions of office building load and corresponding optimization

The application of machine learning methods in building load prediction models has increased significantly, and optimized models have shown good performance in predicting hourly electric loads of office buildings throughout the year. However, machine learning algorithms are often challenged when predicting the load for irregular days, such as holidays and adjusted work schedules.

These irregular days significantly reduce the model’s accuracy. Therefore, this study analyzes the reasons why irregular days affect machine learning prediction models, explains the limitations of common basic optimization methods such as algorithm selection and feature selection, and proposes an optimization method using time-partitioned dataset splitting to address the decline in prediction accuracy caused by irregular days. The results indicate that irregular days lead to a more than twofold decrease in prediction accuracy, and common optimization methods based on the model itself are ineffective in solving this issue.

Furthermore, the underlying reason for the accuracy decline due to irregular days is the failure of features caused by the change at the boundary between workdays and non-workdays. Finally, an optimization method based on time-patitioned dataset splitting, addressing the inherent reasons, can reduce the prediction error by 10.86% during the irregular days. This study provides a more efficient optimization direction for building load prediction and enhances the reliability of load forecasts, making them more applicable to real-world engineering challenges such as supply-demand balance and energy distribution.