Calibration and sensitivity analysis of long-term generation investment models using Bayesian emulation

Abstract

Investments in generation are high risk, and the introduction of renewable technologies exacerbated concern over capacity adequacy in future power systems. Long-term generation investment (LTGI) models are often used by policymakers to provide future projections given different input configurations. To understand both uncertainty around these projections and the ways they relate to the real-world, LTGI models can be calibrated and then used to make predictions or perform a sensitivity analysis (SA). However, LTGI models are generally computationally intensive and so only a limited number of simulations can be carried out. This paper demonstrates that the techniques of Bayesian emulation can be applied to efficiently perform calibration, prediction and SA for such complex LTGI models. A case study relating to GB power system generation planning is presented. Calibration reduces the uncertainty over a subset of model inputs and estimates the discrepancy between the model and the real power system. A plausible range of future projections that is consistent with the available knowledge (both historical observations and expert knowledge) can be predicted. The most important uncertain inputs are identified through a comprehensive SA. The results show that the use of calibration and SA approaches enables better decision making for both investors and policymakers.