A novel PoI temperature prediction method for heat source system based on graph convolutional networks

The ever-increasing functional density and complexity of the heat source system, the harsh chip cooling environment, as well as the cost reduction measures that require less sensor involvement are increasingly driving the need to develop new approaches for temperature monitoring and predicting. Most current researches investigate the temperature field reconstruction of the whole system. However, the reconstruction is not necessary or practical in some cases for computing consumption and unprocurable structure. The temperature prediction of the points of interest (PoIs) like the heat-sensitive area in the electronics is essentially important for function maintenance. Thus, a complete solution is proposed based on the graph convolutional networks (GCN) in this paper including the dimensional alignment, graph modeling and corresponding GCN construction. Moreover, various methods have been explored for edge modeling and node embedding obtained by Node2Vec has been integrated for better graph representation. After model training, the real-time temperature prediction of PoIs can be realized according to the corresponding temperature of monitoring points (MoPs) by the GCN. The results of experiments show that this method approach well prediction that the mean absolute error is less than 0.01K under the condition possessing diverse MoPs and PoIs. Moreover, the comparison experiments with the baseline methods further verify the validity of this GCN-based solution