关键词: 非侵入式负荷辨识, 多尺度卷积神经网络, 双向门控循环单元, 多层感知机, 公共建筑, 电压-电流（V-I） 轨迹特征, 能源管理

Abstract:

In the energy management of public buildings， load identification plays a critical role in optimizing energy utilization and reducing energy consumption. Traditional load monitoring methods are primarily intrusive， relying on hardware equipment or macro-level load characteristics， which fail to meet the refined management requirements of modern intelligent buildings and smart cities. To address the challenges posed by the diversity and uncertainty of public building loads， a non-intrusive load identification method was proposed based on multi-scale convolutional neural network （MSCNN）， bidirectional gated recurrent unit（BiGRU）， and multilayer perceptron（MLP）. The model integrated voltage-current（V-I） trajectory features， power features， and harmonic features to achieve classification and identification of typical socket-based loads in public buildings. MSCNN was employed to extract V-I trajectory features， capturing stable and "fingerprint-like" characteristics of equipment during operation. BiGRU was utilized for time-series modeling of power and harmonic features， revealing the dynamic characteristics of load signals. MLP was then applied to classify the fused features. Experiments on various common public building loads validated the effectiveness of the proposed model. The results showed that the MSCNN-BiGRU-MLP model achieved a load identification accuracy of 0.917 1， accurately identifying load types and maintaining high robustness under dynamic feature changes and high-frequency noise conditions.

Key words: non-intrusive load identification, multi-scale convolutional neural network, bidirectional gated recurrent unit, multilayer perceptron, public buildings, voltage-current（V-I） trajectory features, energy management