InterGCNet: An Interpolation Geometric Constructive Neural Network for Industrial Data Modeling

For industrial data modeling, how to build data-driven models on resource-constrained industrial devices is a research hotspot. However, existing literature is either high resource consumption or poor prediction performance. To bridge this gap, we propose an interpolation geometric constructive neural network (InterGCNet) with lightweight and good prediction performance. Specifically, we first use the maximum likelihood function to analyze the reasons for the poor prediction of the existing method. Subsequently, we propose an interpolated label with a decay factor to replace the raw label based on the interpolation theory and the simulated annealing. Finally, we prove the universal approximation property (UAP) of InterGCNet by induction. Experimental results, including five benchmark datasets and a PV dataset collected from PV power stations, demonstrate that InterGCNet is performing exceptionally well in terms of prediction performance.