Short Term Energy Consumption Prediction Method For Complex Manufacturing Systems Based On Spatio-temporal Deep Learning

Complex manufacturing systems involve a wide range of areas and consume abnormally large amounts of energy.Short-term energy consumption prediction is the basis for anomaly detection,real-time scheduling and energy-saving management Therefore,the prediction of manufacturing systems is of significant research value.The flexible process path and dynamic production rhythm lead to complex spatio-temporal relationships between production nodes,posing challenges to short-term energy consumption prediction.Existing prediction methods do not fully consider the relationship between energy consumption in manufacturing systems and time and space,resulting in difficulties in meeting application requirements in terms of efficiency and accuracy.This article is based on the energy consumption data of an aluminum profile factory in Guangdong as the research background,and uses a deep learning model to build a spatio-temporal prediction algorithm to achieve short-term energy consumption prediction in the manufacturing system.The specific research content is as follows:A spatio-temporal deep learning network(STDLN)prediction method was proposed from the perspectives of time and space.The model uses gated recurrent neural unit to recursively map the timing sequence containing time-varying information into the prediction model,and then constructs a graph model through graph convolution neural network model and adjacency matrix to mine the spatial relationship of energy consumption data.This model constructs a deep spatio-temporal relationship network from the perspectives of time and space,achieving short-term accurate prediction of energy consumption in manufacturing systems.For STDLN can only predict one attribute energy consumption,a Sparse perceptual graph symmetric spatio-temporal learning network was proposed.SSTLN-SMG)prediction method.This model provides detailed definitions for manufacturing system energy consumption nodes,instrument collectors,and workshops.Build a sparse instrument graph model for energy consumption nodes and energy consumption collection tables,and then mine the relationships between energy consumption nodes along the time axis using gated linear units.The spatial relationships are processed using a graph convolutional neural network with Chebyshev.Compared with the deep spatio-temporal relationship network method,this model not only improves prediction accuracy,but also achieves simultaneous prediction of multiple attribute nodes.Implementing simultaneous prediction of multiple nodes may result in poor prediction of a small number of energy consuming nodes.In order to compensate for the shortcomings in spatio-temporal prediction,a dynamic spatial prediction(ARSDM)model was constructed using convolutional neural networks,gated recurrent neural units,and Attention mechanisms.Achieve accurate prediction of target nodes.Through experimental verification,this model can compensate for the shortcomings of spatio-temporal prediction models.