In the future, we intend to include further refinements like on-demand accuracy and cut generation. The setup can also be extended to abandon the perfect foresight assumption within a year, study long-term dynamics of long-duration storage or to represent endogenous learning. 10/10

Using stabilization and parallelization, the algorithm scales well with the number of weather years. In our tests, the reference configurations consistently outperform Gurobi from 8 weather years, for some cherry-picked configurations already at 2 weather years. 9/10

Second, the structure of the decomposed problem allows solving the scenarios in parallel. Some overhead remains but the speed-up usually amounts to half the number of scenarios, e.g. four times faster for eight weather years relative to solving the scenarios sequentially. 8/10

By robust and implementable, we not only mean the complexity of coding the algorithm but also a low sensitivity to hyperparameter tuning, which can prove quite tricky and have a major effect on performance. 7/10

First, stabilization limits the step size of the algorithm preventing oscillation. We find that the most robust and implementable approaches for this problem are linear/quadratic trust regions (ex. shown below for 3 dimensions) improving performance by a factor of 10 to 20. 6/10

Yet, the standard Benders converges slowly, and commercial solvers can still solve the original linear program faster by a factor of 100. Hence, we scoured the convex optimization literature for refinements to speed up the algorithm. I focus on the two most effective ones: 5/10

We split this linear optimization problem into an expansion part and independent operational parts, one for each weather year to avoid a single unsolvable big problem. Benders decomposition then iteratively solves the individual problems, converging to the optimal solution. 4/10

The algorithm solves a two-stage stochastic problem. The first stage covers capacity expansion; the second operation for different scenarios representing weather years. The setup includes short- and long-duration storage — critical for systems with high shares of renewables. 3/10

Our new European Journal of Operational Research paper adopts and compares various refinements to a Benders Decomposition algorithm to find a fast and implementable way of accounting for climate uncertainty by including many weather years. 2/10