Contributions to optimizing control – Inspired by challenges in oil and gas production

Abstract

This dissertation treats optimization and control with emphasis on oil and gas applications, and consists of an introduction followed by five main contributions. In the first contribution, we focus on allocating a limited available resource between a set of units which is a problem that arises in several application areas. We propose an online derivative-free trust-region model-based method to tackle a fairly general version of the resource allocation problem where units may be turned on or off. The units are considered as black boxes which may only be evaluated given that all the other units are evaluated simultaneously, and no gradient information is available. This method was inspired by an industrial problem and emphasis is put on both providing feasible points during the optimization and on not incurring additional increase in cost while searching for the optimum. The latter cannot be guaranteed, but the algorithm allows for automatic or manual ranking of the different units to attempt to reduce negative impact on the cost. In the second contribution, we treat systems with fast and slow dynamics which give rise to objectives in different time scales which may not be aligned. The existing dynamic optimal control methods might become computationally infeasible due to the fine discretization required to capture the fast dynamics. On the other hand, a Real-Time Optimization (RTO) method based on steadystate models, which is computationally efficient, can greedily drive the plant towards optimal operation. The drawback of the RTO approach is that it may yield actions that only focus on near future goals and the objectives involving the slower dynamics are neglected. We propose to extend RTO with a lookahead strategy by introducing a predictor to capture the effect of changing the current controls on the long-term objective. In this way, we introduce the long-term objectives in RTO while maintaining its computational efficiency and not losing focus of short-term objectives. In the third contribution, we tackle the Daily Production Optimization (DPO) problem which is the task of maximizing production of hydrocarbons subject to operational constraints. Handling of uncertainty in model structure and parameters is of high importance to the usefulness of the solution. Ignoring these challenges will, most likely, render the solution either infeasible or the solution will not be an optimum of the plant. We apply a Reinforcement Learning based Economic Nonlinear Model Predictive Control (RL-based ENMPC) which uses state- and output-measurements from the plant to iteratively update the controller. Thereafter, in the fourth contribution, we investigate similarities and differences between RL-based ENMPC and Modifier Adaptation. Both methods aim to correct for plant-model mismatch in order to improve economic performance. We suggest specific parametrizations of the RL-based ENMPC which allow for selecting, or learning of, a closed-loop steady state in addition to the local policy around it. The proposed parametrizations are theoretically justified, in addition to being illustrated through examples. Finally, in the fifth contribution, the focus is moved away from production optimization to linear regression which is concerned about fitting a model to a set of data. The weighted least squares method is a standard tool for performing linear regression. In this paper, we focus on the case when some of the samples are given priority over others. The residuals for these samples should be given an infinite weighting compared to other samples. However, due to numerical limitations, a weight which is finite but sufficiently large must be chosen instead. We suggest an alternative approach that in practice allows infinite weighting. This is achieved by reformulating the regression optimization problem as a bilevel program.