Doctor of Philosophy (PhD)


The department of Electrical and computer engineering

Document Type



Distributed optimization approaches are gaining more attention for solving power systems energy management functions, such as optimal power flow (OPF). Preserving information privacy of autonomous control entities and being more scalable than centralized approaches are two primary reasons for developing distributed algorithms. Moreover, distributed/ decentralized algorithms potentially increase power systems reliability against failures of components or communication links.

In this dissertation, we propose multiple distributed optimization algorithms and convergence performance enhancement techniques to solve the OPF problem. We present a multi-level optimization algorithm, based on analytical target cascading, to formulate and solve a collaborative transmission and distribution OPF problem. This algorithm enables transmission and distribution system operators to solve their OPF subproblems in a parallel, yet collaborative manner, which would result in a more economical and reliable operating point for the whole power system.

Motivated by observing the sensitivity of most augmented Lagrangian (AR)-based distributed algorithms to the choice of initial values of tuning parameters and the level of importance of each term in objective functions, a technique is proposed to balance convergence speed and accuracy of these types of distributed algorithms. A balancing coefficient is determined and incorporated into AR-based distributed OPF to, potentially, accelerate its speed and enhance its robustness to the choice of initial values.

To further enhance the performance of AR-based distributed OPF algorithms, a prediction-correction based asynchronous alternating direction of multipliers (A-ADMM) is proposed. This algorithm does not need synchronization of subproblems at each iteration, which means a computationally cheap subproblem no longer needs to wait for the most updated information of its computationally expensive neighbors. A second loop, which uses anomaly detection and learning techniques, is added to the proposed A-ADMM to reduce the impact of prediction error propagation, especially if subproblems are computationally heterogeneous with a significant level of asynchrony.



Committee Chair

Kargarian, Amin