Investigating the applicability of Bayesian networks to demand forecasting during the final phase of support operations

A challenge faced by businesses that provide logistical support to systems is when the provision of those support services is no longer required. A typical example of such a situation is when military operations come to an end. In such cases, those companies that have a contract with the Armed Forces to provide maintenance support for the deployed systems, need to maintain those systems at minimum cost during that final phase, that is from the time the decision to stop the operations is announced until their very end. During the final phase, a challenging problem is forecasting the demand for spare parts, corresponding to equipment failures within the system. This is because the support context, the number of supported systems, the support equipment or even the operational demand can change during that period, and also because there can be very limited opportunities to place orders to cover demand. This thesis suggests that these types of problems can take advantage of the data that have been collected during the support operations prior to the initiation of the closing down process. Moreover, the thesis investigates the exploitation of these data by the use of Bayesian Networks to forecast the demand for spares that will be required for the provision of maintenance during the final phase. The research uses stochastically simulated Support Chain scenarios to generate data and also to evaluate different methods of constructing Bayesian Networks. The simulated scenarios differ in the demand context as well as in the complexity of the Equipment Breakdown Structure of the supported systems. The Bayesian Networks’ structure development methods that are tested include unsupervised machine learning, eliciting the structure from Subject Matter Experts, and two hybrid approaches that combine expert elicitation and machine learning. These models are compared to respective logistic regression models, as well as subject matter experts-adjusted single exponential smoothing forecasts. The comparison of the models is made using both accuracy metrics and accuracy implication metrics. These forecast models’ comparison methods are analysed in order to evaluate their appropriateness. The analyses have provided a number of novel outputs. The algebraic analysis of the accuracy metrics theoretically proves empirical problems that have been discussed in the literature but also reveals others. Regarding the accuracy implication metrics, the analysis shows that for the particular type of problems examined in this thesis –final phase problems – the accuracy implication metrics commonly applied are not enough to inform decision making, and a number of additional ones are required.The research shows that for the scenarios examined, the Bayesian Networks that had their structure learned using an unsupervised algorithm performed better in the accuracy metric than any of the other models. However, even though these Bayesian Networks also did well with the accuracy implication metrics, neither they, nor any of the others was consistently dominant. The reason for the discrepancy in the results between the accuracy and the accuracy implication metrics is that the latter are not only driven by how accurate the forecast model’s prediction is, but also by the model of the residual error and the bias.