A novel method overcomeing overfitting of artificial neural network for accurate prediction: Application on thermophysical property of natural gas

As a powerful tool to solve nonlinear problems, artificial neural network method (ANN) gets a wide range of applications in data regression. However, the overfitting often occurs during the ANN training process, which results in high accuracy of correlating the training data but poor prediction performance. At the same time, the principle of k-Nearest Neighbor method (kNN) makes it impossible to make an accurate prediction exceeding the range of the training data, but it can confine the overfitting of ANN. In this work, combining ANN and kNN, a new machine learning method called ANN-kNN combination method (AKC) for thermophysical property prediction of material is proposed. To evaluate the performance of AKC, we take the thermophysical properties of natural gas as an example. The inputs of AKC are temperature, pressure and the components of natural gas, the outputs are the compressibility factor, speed of sound and viscosity. AKC not only overcomes the overfitting problem but also needs less training data than ANN. The average absolute relative deviation of AKC for prediction are 2.5%, which are better than ANN (5.9%) and kNN (19.2%).