Reinforcement learning-based allocation of fog nodes for cloud-based smart grid

Real-time monitoring in smart grids requires efficient handling of massive amount of data. Fog cloud nodes can be strategically located within the smart grid to: pull readings from smart meters, implement local processing and control, and make all data available to the smart grid control center with minimum overall latency. Unlike existing studies in literature, we propose a novel Fog node allocation strategy that is tightly coupled with the power grid structure, and hence, accounts for the spatial distribution of data traffic sources (e.g., smart meters) within the power grid. Furthermore, the allocation strategy considers the diverse latency requirements of fixed scheduling and event driven data services within the power grid. The proposed allocation strategy first implements an unsupervised machine learning approach to determine initial number and locations of Fog nodes that can serve the data traffic with minimum overall latency. Then, a reinforcement-based mechanism is applied to minimize the required number of Fog nodes, and hence capital cost, through efficient mapping between Fog nodes and smart meters while still complying with the latency requirements. Our simulation studies demonstrate that a 50% reduction in required number of Fog nodes can be achieved while minimizing overall latency when the proposed allocation strategy is adopted.