Reliability prediction of oil wells by support vector machine with particle swarm optimization for variable selection and hyperparameter tuning

In oil industry, sudden failures of wells are not desirable since they bring non-planned costs, e.g. loss of assets, costumer dissatisfaction, environmental damages, among others. In order to prevent these situations, reliability prediction can be performed so as to enable proper maintenance planning. Different factors, such as operational and aging conditions, impact the reliability behavior of production systems and a comprehensive analytical treatment may be prohibitive due to the increased complexity of the problem. Alternatively, for situations in which failure data exist, empirical regression modeling may be effective. In this paper, Support Vector Regression (SVR) for reliability prediction is used combined with a particle swarm optimization for simultaneous tuning of SVR hyperparameters and variable selection. The proposed PSO+SVR methodology is applied to the reliability prediction of oil wells located in the Brazilian Northeast region. The results show that the hyperparameter tuning combined with variable selection improves the prediction ability of SVR.