Wind energy is growing to be an essential part of the transition towards sustainable energy sources. To facilitate this, it is crucial that wind farm operators can offer competitive energy prices compared to fossil fuel sources; effective and efficient wind farm operations are therefore imperative. However, many full-scale wind farms deal with wake effects, e.g. the disturbed air that travels downstream of a turbine and potentially ends up in other neighbouring turbines. The result of this lower-velocity, higher-turbulence flow field is both decreased power production and increased fatigue loads. One proposed solution for this problem comes in the form of 'wake steering': the yawing (rotating) of upstream turbines' nacelles to facilitate a degree of control over the deflection of wakes. By doing so, the problematic disturbed flow fields can be strategically guided between downstream turbines to maximise collective power production. However, the yawing action itself brings some adverse effects: the upstream turbines, now no longer directly facing the wind, feel decreased power production and increased fatigue loads themselves. These fatigue effects, accumulating over time, can eventually cause increased maintenance costs and nullification of any revenue gains through power optimisation. The problem thus becomes a farm-wide collective revenue optimisation task. This thesis investigates how long-term revenue in wind farms can be maximised, considering both profits through power optimisation and maintenance costs through fatigue-induced component failures due to wake steering control. First, a realistic wind farm simulation environment is constructed, based on a Graph Neural Network (GNN) surrogate wind farm simulation model, to facilitate efficient reinforcement learning training. Next, the environment is used to train fully centralised reinforcement learning agents based on a GNN architecture, resulting in agents that can generalise across all wind conditions and unseen wind farm layouts. Ultimately, the results show that an 'informed' agent that considers all profits and costs involved manages to significantly reduce the cost of energy compared to 'greedy' (power optimisation only), 'risk-averse' (damage minimisation only) and 'baseline' (zero-yaw) policies, furthermore considerably maximising long-term wind farm revenue by as much as 20%. Altogether, this thesis shows promising results in using graph-based reinforcement learning to train maintenance-conditioned, inflow-agnostic, and layout-agnostic wake steering controllers for wind farm revenue optimisation.