Biblio

We present a privacy-preserving distributed reinforcement learning-based control scheme to address the problem of frequency control and economic dispatch in power generation systems. The proposed control approach requires neither a priori system model knowledge nor the mathematical formulation of the generation cost functions. Due to not requiring the generation cost models, the control scheme is capable of dealing with scenarios in which the cost functions are hard to formulate and/or non-convex. Furthermore, it is privacy-preserving, i.e. none of the units in the network needs to communicate its cost function and/or control policy to its neighbors. To realize this, we propose an actor-critic algorithm with function approximation in which the actor step is performed individually by each unit with no need to infer the policies of others. Moreover, in the critic step each generation unit shares its estimate of the local measurements and the estimate of its cost function with the neighbors, and via performing a consensus algorithm, a consensual estimate is achieved. The performance of our proposed control scheme, in terms of minimizing the overall cost while persistently fulfilling the demand and fast reaction and convergence of our distributed algorithm, is demonstrated on a benchmark case study.