Nonlinear System Identification Using Temporal Convolutional Networks: A Silverbox Study

摘要

Identification of nonlinear systems is presented using a neural network variant known as the temporal convolutional network (TCN). The identification capabilities of TCNs and standard feedforward neural networks (FNNs) are benchmarked and compared using the Silverbox dataset: a publicly available dataset from a circuit equivalent to a nonlinear spring-mass damper. The TCN is found to have superior performance in simulation of the test portion of the dataset. In addition, published benchmark results are surveyed and compared to the TCN results. Analysis of existing results reveals testing variances that effect model performance, so guidelines for fair comparison of models on the Silverbox benchmark are presented.