当前位置:
X-MOL 学术
›
Energy Build.
›
论文详情
Our official English website, www.x-mol.net, welcomes your
feedback! (Note: you will need to create a separate account there.)
Enhancing non-intrusive load monitoring through transfer learning with transformer models
Energy and Buildings ( IF 6.6 ) Pub Date : 2025-01-17 , DOI: 10.1016/j.enbuild.2025.115334
Jing Rong, Cong Wang, Qiuzhan Zhou, Yunxue He, Huinan Wu
Energy and Buildings ( IF 6.6 ) Pub Date : 2025-01-17 , DOI: 10.1016/j.enbuild.2025.115334
Jing Rong, Cong Wang, Qiuzhan Zhou, Yunxue He, Huinan Wu
As global energy demands continue to rise, the importance of efficient energy management systems becomes increasingly clear. Non-invasive load monitoring (NILM) technologies, which identify the energy consumption of individual loads through the analysis of aggregate mains data without physical alterations to the electrical system, are gaining widespread attention. To address persistent challenges of low prediction accuracy and weak model generalization in NILM, this paper introduces TransDisNILM—an optimized transformer based NILM model enhanced by transfer learning. First, sinusoidal encoding and improved multi-head transformer encoder layers are employed to capture richer temporal features, thereby improving prediction accuracy in complex multi-load scenarios. Second, transfer learning strategies are applied to systematically select source tasks and fine-tune the model, enabling robust generalization across diverse environments and load types. Evaluation results on multiple public datasets demonstrate that TransDisNILM significantly reduces mean absolute error and normalized signal aggregate error, outperforming state-of-the-art methods. Moreover, TransDisNILM's transfer learning strategies allow effective training across different load types without starting from scratch, thus reducing the reliance on large-scale labeled datasets. Overall, TransDisNILM not only achieves higher accuracy but also exhibits stronger generalization capabilities, advancing the practical deployment of NILM technologies.
中文翻译:
通过使用 transformer 模型进行迁移学习来增强非侵入式负载监控
随着全球能源需求的持续增长,高效能源管理体系的重要性日益凸显。非侵入式负载监控 (NILM) 技术通过分析总电源数据来识别单个负载的能耗,而无需对电气系统进行物理更改,该技术正受到广泛关注。为了解决 NILM 中预测精度低和模型泛化弱的持续挑战,本文介绍了 TransDisNILM——一种通过迁移学习增强的基于变压器的优化 NILM 模型。首先,采用正弦编码和改进的多头变压器编码器层来捕获更丰富的时间特征,从而提高复杂多负载场景下的预测精度。其次,应用迁移学习策略系统地选择源任务并微调模型,从而在不同环境和负载类型之间实现稳健的泛化。多个公共数据集的评估结果表明,TransDisNILM 显着降低了平均绝对误差和归一化信号聚合误差,优于最先进的方法。此外,TransDisNILM 的迁移学习策略允许跨不同负载类型进行有效训练,而无需从头开始,从而减少对大规模标记数据集的依赖。总体而言,TransDisNILM 不仅实现了更高的准确性,而且表现出更强的泛化能力,推动了 NILM 技术的实际部署。
更新日期:2025-01-17
中文翻译:

通过使用 transformer 模型进行迁移学习来增强非侵入式负载监控
随着全球能源需求的持续增长,高效能源管理体系的重要性日益凸显。非侵入式负载监控 (NILM) 技术通过分析总电源数据来识别单个负载的能耗,而无需对电气系统进行物理更改,该技术正受到广泛关注。为了解决 NILM 中预测精度低和模型泛化弱的持续挑战,本文介绍了 TransDisNILM——一种通过迁移学习增强的基于变压器的优化 NILM 模型。首先,采用正弦编码和改进的多头变压器编码器层来捕获更丰富的时间特征,从而提高复杂多负载场景下的预测精度。其次,应用迁移学习策略系统地选择源任务并微调模型,从而在不同环境和负载类型之间实现稳健的泛化。多个公共数据集的评估结果表明,TransDisNILM 显着降低了平均绝对误差和归一化信号聚合误差,优于最先进的方法。此外,TransDisNILM 的迁移学习策略允许跨不同负载类型进行有效训练,而无需从头开始,从而减少对大规模标记数据集的依赖。总体而言,TransDisNILM 不仅实现了更高的准确性,而且表现出更强的泛化能力,推动了 NILM 技术的实际部署。