Project description

Project Overview and Goals

Given a mix of power plants, the Unit Commitment problem (UC) can be succinctly stated as finding a generation scheduling attending to the following principles:

This easily stated problem is in fact very difficult to model and solve (in particular for EDF), mainly for the following reasons:

In this project, we will focus on two forms of UC. The first, "traditional" one (TUC), determines one day ahead a generation plan with minimal production and start-up costs, and which satisfies the physical customer load, as well as certain reserve and system service constraints. The TUC output is a reference planning for the next day. The second problem is the intra-day UC (IUC), solved once every hour, to determine a generation that deviates the least from the planned generation, taking into account the actual realization of uncertainty in the considered hour. Deviation must be minimal in the sense that only a limited number of (groups of) units can modify their generation output on a short notice.

Both UC variants above are useful, not only in themselves as steps of the actual operational planning, but also as sub-models of the "real" UC problem. This is a stochastic program that can be modelled as a problem with recourse, as follows. Having the reference schedule, corrective decisions (the "recourse") can be taken to adapt the planned generation to last minute changes dictated by a given realization of uncertainty, for example, the actual renewable generation or the actual demand of power in the next hour. In this context, TUC outputs the reference schedule (here-and-now decisions) while IUC provides the recourse decisions for any given scenario.

Current Approach

The solution process is currently subdivided in two phases:

Project Proposal

From the above, we can conclude that the current approach is not always satisfactory. In particular, the method in the Primal Phase is too slow, rather unstable and provides a primal solution that is not consistent with the obtained dual signal.

This project will address the issues above, keeping in mind that there is a trade-off to be found between accuracy and computational speed. Specifically, for optimization problems as complex as the real-life UC problems solved by EdF, it is unreasonable to expect gains on all the fronts. The ultimate goal of this collaboration is indeed to develop a computational tool that produces fast a primal feasible solution together with a consistent dual signal. This is an ambitious objective, difficult to achieve, since it may very well happen that to improve feasibility of the primal solutions, the overall computational time needs to be substantially increased. Also, it is possible that, due to the presence of nonconvex/combinatorial relations, an increase in primal feasibility is accompanied by a deterioration in the quality of the dual signal.

Having said that, the project team counts with a vast know-how and experience in the domain, which made it possible to identify several improvements to the current approach, essentially related to three key topics: