Hybrid multi-agent deep reinforcement learning for multi-type mobile resources dispatching under transportation and power network recovery

Rainstorm waterlogging or typhoon can not only cause seriously failure of power network (PN), but also damage the normal traffic of transportation network (TN). Equipment fault of PN affects normal power supply of critical loads, and the interruption of TN severely limits the flexible transfer of mobile resources for recovery of transportation and power network (TPN). Previous work only addresses dispatching of multi-type mobile resources (MMRs) for power network recovery on the assumption of healthy TN, which makes dispatching strategy impractical. To fill this gap, this paper proposes a dispatching model of MMRs for collaborative recovery of TPN, embedding road repair crews (RRCs) dispatching behaviors into road repair constraints. To solve the above model, firstly road island and various topology update strategies are introduced to simplify shortest path searching for MMRs routing. Then, the dispatching model of MMRs is described as a parameterized action Markov decision process, in which MMRs are modeled as different types of intelligent agents considering various discrete-continuous dispatching characteristics. And, a hybrid multi-agent deep reinforcement learning (HMADRL) method characterizing master-slave architecture is developed to improve the solving efficiency and convergence speed of model, where the master module describes the recovery process of TN with dispatching of RRCs, and the slave module is constructed to recovery PN based on the path update strategies. The case analysis based on 15-node PN (18-node TN), 33-node PN (45-node TN) and practical example demonstrates that this approach can elevate the practicality of dispatching strategies and the recovery efficiency of TPN.