[en] This manuscript deals with large-scale non-smooth optimization that may typically arise when performing Lagrangian relaxation of difficult problems. This technique is commonly used to tackle mixed-integer linear programming - or large-scale convex problems. For example, a classical approach when dealing with power generation planning problems in a stochastic environment is to perform a Lagrangian relaxation of the coupling constraints of demand. In this approach, a master problem coordinates local subproblems, specific to each generation unit. The master problem deals with a separable non-smooth dual function which can be maximized with, for example, bundle algorithms. In chapter 2, we introduce basic tools of non-smooth analysis and some recent results regarding incremental or inexact instances of non-smooth algorithms. However, in some situations, the dual problem may still be very hard to solve. For instance, when the number of dualized constraints is very large (exponential in the dimension of the primal problem), explicit dualization may no longer be possible or the update of dual variables may fail. In order to reduce the dual dimension, different heuristics were proposed. They involve a separation procedure to dynamically select a restricted set of constraints to be dualized along the iterations. This relax-and-cut type approach has shown its numerical efficiency in many combinatorial problems. In chapter 3, we show Primal-dual convergence of such strategy when using an adapted sub-gradient method for the dual step and under minimal assumptions on the separation procedure. Another limit of Lagrangian relaxation may appear when the dual function is separable in highly numerous or complex sub-functions. In such situation, the computational burden of solving all local subproblems may be preponderant in the whole iterative process. A natural strategy would be here to take full advantage of the dual separable structure, performing a dual iteration after having evaluated only a subset of the dual local sub-functions, instead of all of them. This type of incremental approach has already been applied for sub-gradient algorithms and the reported computational results are promising. In chapter 4, we use instead a specialized variant of bundle methods known by their stability and precision. We give numerical illustrations for the EDF's mid-term generation planning problem. In chapter 5, we focus on this specific problem. The French power mix includes hydraulic, nuclear, and classical thermal power plants, with each generation unit having its own technical constraints. For an horizon of one to three years, the generation planning problem is a large scale stochastic program, un-tractable with a direct frontal approach. We will present a software using Lagrangian decomposition on a scenario tree to tackle this problem. The dual function is maximized using a proximal bundle method. As a byproduct of our work we proposed some improvements of the algorithms to fasten and stabilize the resolution process and extract useful informations regarding the primal solutions. We also developed a heuristical valuation of water and nuclear resources. (author)