一种基于时频特征融合和极限学习机的非侵入式负荷识别方法

Nonintrusive load identification method based on time-frequency feature fusion and extreme learning machine

以往的负荷识别方法在提高识别精度和实际落地部署方面遇到了一些挑战,如传统经典方法的识别准确度相对较低,而先进深度学习方法又因其较复杂的模型而很难部署到嵌入式设备上。为解决上述问题,本文提出一种基于高频样本数据识别方法。首先采用快速傅里叶变换(FFT)和希尔伯特变换(HT)对样本进行时域和频域特征提取。然后基于极限学习机(ELM)提出了一种多特征融合学习策略来获取特征与负荷类型之间的映射关系,设计了一种窗口调整方法,以获得最优模型和最合适的窗口长度。最后在两个高频公共数据集BLUED和PLAID上对该方法进行了实验评估。实验结果表明,所提方法具有识别性能较好、易于在嵌入式设备上落地部署的优点。

Prior load identification methods have encountered a challenge in improving identification accuracy and being deployed. For example, the identification accuracy of traditional classical methods is relatively low, and advanced deep learning methods are difficult to deploy on embedded devices due to their complex models. In this paper, a novel load identification method based on high-frequency samples is proposed. The fast Fourier transform and Hilbert transform are applied to extract the features of samples both in time and frequency domain. Then, based on extreme learning machine, a feature fusion learning scheme is proposed to obtain mapping relationship between extracted features and load types. Furthermore, a window adjustment method is designed to obtain the optimal model and the most proper sampling length. In experiment, the proposed method is evaluated on two high-frequency public datasets, BLUED and PLAID. Based on the experiment results, it can be proved that the proposed method has a faster speed, better performance, and easier to be deployed on embedded devices to the previous.