When we use regression learning to predict some state variables in artificial systems, it is important to consider nonstationarity, sparsity, interpretability, reliability, and so on. This paper propose a method that takes them into account, called online linear Nonstationary Sparse Bayesian Learning (OL-NSBL), and considered its application to oil plant data as an example of an artificial system. We first investigated the performance and properties of the proposed method on an artificially generated data set, and demonstrated its effectiveness. Then we applied it to the plant data and it showed high performance.