[en] Growing energy needs, the possible depletion of fossil resources, and the ambition to protect the environment have motivated the development of renewable energy sources, particularly solar and wind power, in power generation systems. The intermittency of these sources makes it more complex to control power grids, which are based on the production-demand balance at all times. Storage systems, such as batteries, make it possible to compensate for this variability, but they require control to be considered over a longer future time horizon, which is difficult to reconcile with the short-term dynamics of the system, and over which renewable production forecasts are imperfect. In the context of micro-grids, many control strategies have been developed, based on different choices in the representation of the physical system and its dynamics, in the integration of the forecast data, or in the optimization algorithm. The most sophisticated strategies are the least widespread in practice, and their added value is poorly known due to the lack of various comparative studies, which are difficult to conceive on real systems. This thesis aims to shed light on the practical performance of control strategies, with the implementation of different optimization methods in association with a real-time controller, within a single simulation platform, on data from real instances of off- or on-grid micro-grids. Our experiments show a significant interest in accurately considering the variability of production on a daily scale, through scheduling approaches, and highlight the ability of the predictive control loop to absorb the system dynamics modeling flaws, which are inherent to these approaches. We then observe the superiority of a mathematical scheduling approach based on a highly simplified analytical model to accelerate the resolution to the optimum, compared to a metaheuristic approach associated with a precise numerical simulation model, but without any guarantee of optimality. This thesis also focuses on the model of a grid-connected photovoltaic-based micro-grid with storage, and independently driven through the notion of engagement. In the context of calls for tenders by the French Energy Regulatory Commission for non-interconnected areas, the independent producer commits to the grid operator on its production capacity in a day-ahead fashion, and is required to pay penalties for deviations from the commitment observed the next day. In this context, the minimisation of penalties is added to the profit objective of the real-time control problem, and a new decision problem appears upstream, at the tactical level, for the calculation of the optimal engagement. We propose to extend the different control methods to the calculation of the engagement for comparison purposes, and we observe again, empirically, an advantage of the mathematical programming approach, both in terms of profit and runtime. Finally, we examine in this thesis the impact of taking into account the uncertainties in the solar production forecast. As for modeling errors, these uncertainties can be partly compensated by frequent recalculation within the control loop. On the other hand, the uncertainties being higher at the time of the engagement calculation, errors in the engagement decision can be the source of significant penalties. The mathematical programming approach lends itself to the integration of uncertainties through stochastic or robust optimization. Two extensions of the deterministic engagement calculation model have been realized for comparison purposes: a stochastic model after generation and simulation of solar production scenarios, and a robust optimization approach by bi-level programming, with progressive generation of Bender cuts. The experimental evaluation shows the importance of taking into account uncertainties in the engagement calculation, and a slight superiority of the robust approach in terms of profit. (author)