Devising a long-term maintenance plan for a system of large infrastructure assets is an exacting task. Any maintenance activity that induces system downtime can incur a massive production or service loss. This problem becomes increasingly challenging for a system of which the performance is based on the collective output of assets. Current approaches that optimise each asset in isolation or consider a binary performance relationship insufficiently address this issue because the negligence of performance interactions among assets results in an inaccurate cost estimation. To overcome these hurdles, we formulate a mathematical model that explicitly demonstrates dynamic risk of production loss according to the system aggregate output. Further, we propose an integrated solution method that couples a finite loop search with a Genetic Algorithm. Application of our model to a real-world case study has proved to simultaneously strike the balance between cost and risk. Validated by Monte Carlo simulation, the proposed model has shown to outperform existing approaches. By systematically scheduling maintenance actions over the planning horizon, the resultant strategy has demonstrated to offer considerable maintenance cost savings and significantly prolong the average asset life. Sensitivity analyses also evince the robustness of the proposed model under the volatility in key parameters.