Accurate building electricity load prediction is critical for Intelligent buildings operation and control. However, the imbalance in the number of workday and holiday samples in historical data often leads to lower prediction accuracy for holidays. Furthermore, existing research typically evaluates performance using global accuracy, which can obscure deficiencies in holiday-specific predictions. To address these issues, this study proposes a local data augmentation method based on time-series generative adversarial network (TimeGAN). By integrating TimeGAN with K-means, the method selectively generates virtual holiday samples to effectively mitigate samples imbalance. The convolutional neural network-long short-term memory (CNN-LSTM) algorithm optimized by genetic algorithm is used to construct the prediction model to improve the prediction performance. The proposed method examines the impact of data augmentation on both global and local accuracy and explores the relationship between sample expansion ratios and augmentation effectiveness. The case results demonstrate that TimeGAN can generate high-quality virtual samples. With local data augmentation, the holiday prediction accuracy improves by 16.14 %, while the global R2 increases by 1.42 %. Compared to global data augmentation methods, the proposed approach achieves superior performance. Further analysis reveals that the optimal sample augmentation ratio for local data augmentation is 1. This research offers a novel solution to address samples imbalance in building load prediction and provides a practical tool for Intelligent buildings operation and control.

中文翻译：

---

缓解不平衡分布：时间序列生成对抗网络中构建电力负荷的局部小样本数据增强


准确的建筑用电负荷预测对于智能建筑的运行和控制至关重要。但是，历史数据中工作日和假日样本数量的不平衡通常会导致假日的预测准确性降低。此外，现有研究通常使用全局准确性来评估绩效，这可能会掩盖特定假期预测的缺陷。为了解决这些问题，本研究提出了一种基于时间序列生成对抗网络 （TimeGAN） 的局部数据增强方法。通过将 TimeGAN 与 K-means 集成，该方法选择性地生成虚拟假期样本，以有效缓解样本不平衡。采用遗传算法优化的卷积神经网络-长短期记忆 （CNN-LSTM） 算法构建预测模型，以提高预测性能。所提出的方法检查了数据增强对全局和局部准确性的影响，并探讨了样本扩展率与增强效果之间的关系。案例结果表明，TimeGAN 可以生成高质量的虚拟样本。通过本地数据增强，假期预测准确性提高了 16.14 %，而全球 R2 增加了 1.42 %。与全局数据增强方法相比，所提出的方法实现了卓越的性能。进一步分析表明，局部数据增强的最佳样本增强比率为 1。本研究为解决建筑负载预测中的样本不平衡问题提供了一种新颖的解决方案，并为智能建筑运行和控制提供了实用工具。