To efficiently balance traffic demand and capacity, optimization of air traffic management relies on accurate predictions of future capacities, which are inherently uncertain due to weather forecast. This dissertation presents a novel computational efficient approach to address the uncertainties in air traffic system by using chance constrained optimization model.

First, a chance constrained model for a single airport ground holding problem is proposed with the concept of service level, which provides a event-oriented performance criterion for uncertainty. With the validated advantage on robust optimal planning under uncertainty, the chance constrained model is developed for joint planning for multiple related airports. The probabilistic capacity constraints of airspace resources provide a quantized way to balance the solution's robustness and potential cost, which is well validated against the classic stochastic scenario tree-based method.

Following the similar idea, the chance constrained model is extended to formulate a traffic flow management problem under probabilistic sector capacities, which is derived from a previous deterministic linear model. The nonlinearity from the chance constraint makes this problem difficult to solve, especially for a large scale case. To address the computational efficiency problem, a novel convex approximation based approach is proposed based on the numerical properties of the Bernstein polynomial. By effectively controlling the approximation error for both the function value and gradient, a first-order algorithm can be adopted to obtain a satisfactory solution which is expected to be optimal. The convex approximation approach is evaluated to be reliable by comparing with a brute-force method.

Finally, the specially designed architecture of the convex approximation provides massive independent internal approximation processes, which makes parallel computing to be suitable. A distributed computing framework is designed based on Spark, a big data cluster computing system, to further improve the computational efficiency. By taking the advantage of Spark, the distributed framework enables concurrent executions for the convex approximation processes. Evolved from a basic cloud computing package, Hadoop MapReduce, Spark provides advanced features on in-memory computing and dynamical task allocation. Performed on a small cluster of six workstations, these features are well demonstrated by comparing with MapReduce in solving the chance constrained model.