Energy Consumption Prediction for Manufacturing in Industrial IoT Based on Heterogeneous GNN

Energy consumption prediction plays a crucial role in green manufacturing for industry, providing a theoretical foundation for energy-saving planning, workshop scheduling, and process optimization. However, owing to the processing complexity of mechanical products and the multiplicity of production information, previous methods can only achieve prediction for individual devices, which are quite restrictive to collaborative manufacturing planning. To deal with product-level energy management, this paper first proposes a heterogeneous manufacturing association graph (HetMG) which contains various manufacturing elements including product projects, decomposed orders and processing devices. Then, with HetMG, an energy consumption prediction method, HetEP, is introduced based on heterogeneous graph neural network (GNN), which incorporates Relational Graph Convolutional Networks (R-GCNs), and Long Short-Term Memory (LSTM) to mine the spatial-temporal relation of heterogeneous elements during the manufacturing process. Moreover, Graph Attention Network (GAT) is utilized in HetEP to adaptively learn the graph-structured data so that informative representations of each element can be obtained. The experimental results based on a manufacturing dataset from a generator factory demonstrate that the proposed HetEP yields superior prediction performance compared to the device-oriented prediction methods in terms of prediction accuracy.