Abstract:

A method was proposed based on TCN-GAT and hybrid neural networks for identifying anomalies in energy usage for drying systems. First， a multi-scale temporal convolutional network （TCN） and a multi-head graph attention network （GAT） were introduced to capture the temporal and spatial properties of temperature， pressure， and other variables， respectively. An anomaly detection model was built upon a combination of back propagation neural network （BPNN） and variational autoencoder （VAE）. Furthermore， an energy consumption anomaly index was formulated based on prediction errors and reconstruction probability. The peak over threshold （POT） model was utilized to fit the Pareto distribution and establish an anomaly threshold. Finally， a case study was carried out at the painting workshop of a Chongqing automobile manufacturer， where Internet of Things （IoT） devices were used to gather real-world data. Data analysis was implemented to verify the effectiveness and superiority of the proposed method.

Key words: drying system, spatio-temporal feature extraction, energy consumption anomaly detection, energy consumption anomaly index

摘要：

提出了一种基于时间卷积网络-图注意力网络（TCN-GAT）与混合神经网络的烘干系统能耗异常检测方法。首先引入多尺度TCN和多头GAT分别捕获温度、压力等数据的时间特征与空间特征；然后联合反向传播神经网络（BPNN）与变分自编码器（VAE）搭建异常检测模型；再次基于预测误差与重构概率构建能耗异常指标，并引入超阈值模型（POT）拟合Pareto分布建立异常阈值；最后在重庆某汽车工厂涂装车间开展案例验证，利用物联网设备（IoT）采集烘干系统能耗等数据，通过数据分析验证了所提方法的有效性和优越性。

关键词: 烘干系统, 时空特征提取, 能耗异常检测, 能耗异常指标