Distributionally robust SDDP

We study a version of stochastic dual dynamic programming (SDDP) with a distributionally robust objective. The classical SDDP algorithm uses a finite (nominal) probability distribution for the random outcomes at each stage. We modify this by defining a distributional uncertainty set in each stage to be a Euclidean neighbourhood of the nominal probability distribution. We derive a formula for the worst-case expectation of future costs over this set that can be applied in the backward pass of SDDP. We verify the correctness of this algorithm, show its almost sure convergence under standard assumptions, and illustrate it by applying it to a model of the New Zealand hydrothermal electricity system.