Seminar – Network dual reoptimization policies and bounds for managing energy real options

Energy operations and investments are managed with real option models that embed timing and/or switching decisions, where the former and latter types of decisions change irreversibly and reversibly, respectively, the status of the real asset. Optimizing the flexibility in these models requires solving a finite-horizon Markov decision process with finite endogenous states but a high-dimensional and continuous exogenous state space, which is typically intractable and requires heuristic approaches. Least squares Monte Carlo (LSM) is a popular approximate dynamic programming method for this purpose while a known forecast-based reoptimization heuristic (FRH) is not well suited to handle timing decisions. We develop a network dual reoptimization framework that leverages network flow algorithms and employs a novel class of partial information relaxations to overcome this shortcoming. Our framework provides both a new policy and a dual bound and we establish that: (i) the decisions taken by our policy are equivalent to solving a scenario-based two-stage stochastic program but computationally more efficient since scenarios are solved independently; and (ii) our dual bound is provably tighter than the known penalized perfect information bound. Numerical experiments on two energy real options dealing with commodity production and vehicle fleet upgrade show that our policy outperforms FRH by 10-20% and LSM by to 2-5%, while our dual bound improves by less than 1%.

 ET-14