In the process industry, uncertain factors, such as yield, can be quantified by analyzing industrial data generated from continuous sources. Traditional data-driven robust optimization models are mostly built on estimated probability distributions and convex uncertainty sets. As a result, the scheduling solution is only applicable to the limited sample of stochastic scenarios. We developed a rolling-horizon optimization approach to adapt the robust model to the changing environmental and operational conditions. First, a novel uncertainty set is defined by the probability density contours, covering scenarios with high possibility of occurrence. Then, we propose using new robust formulations induced by the outer-approximations of nonconvex uncertainty set. By implementing the raised model on a real-world ethylene production process using the available data, the fluctuation in fuel gas consumption can be controlled within 2%. Additionally, in agreement with our proof, the system's total profit and consumption of fuel gas stabilize in finite steps. (C) 2018 Elsevier Ltd. All rights reserved.