This paper presents a dynamic programming (DP) model of optimal operation of a nuclear power plant (NPP). In each period the operator decides whether to run the reactor at a given capacity level, to shut it down for preventive maintenance or refueling, or permanently close the plant for decommissioning. The maximal lifespan of a NPP is determined by the length of the operating license issued by the Nuclear Regulatory Commission. The optimal lifespan of a NPP (from the private perspective of the plant owner as opposed to the social perspective of the regulator) is the solution to a generalized optimal stopping problem: one closes a plant as soon as the expected discounted value of future operating profits (losses) falls below the costs of decommissioning. This model extends the DP model of NPP operations introduced in Rust and Rothwell (1995) by allowing for the occurrence of ``major problem spells''. The DP model predicts that under ordinary operating conditions it is highly unlikely that an NPP will be closed, but that the probability of decommissioning increases by orders of magnitude during a major problem spell. We compare the actual evolution of the nuclear power industry over the period 1984 to 1994 to stochastic simulations of our estimated DP model and show that our model provides accurate out-of-sample predictions of total nuclear power generation and early retirements of NPPs. We use the estimated DP model to forecast nuclear power generation and plant closures under two policy scenarios: 1) the current 40 year license span with no possibility of extension, and 2) a ``costless'' extension in operating licenses to 60 years. Our simulations show that an immediate, costless extension in operating licenses to 60 years would extend the life of the US nuclear power industry from 2030 to 2050 and double the expected present discounted value of profits of NPPs.
