Neuromorphic Systems for Power Grid Cyber-Resilience

Legacy devices, non-standard physical infrastructure, decentralized distribution and communication networks, and the growing incorporation of renewable energy sources compound the complexity and vulnerability of the modern power grid. Detection and subsequent response to faults and failures caused by natural events, component degradation, and cyber-attacks is critical. The sheer magnitude of this sprawling system is of particular concern, with millions of potential cyber-intrusion access points in hardware, software, and network connections. Effective monitoring systems must do more than rely on simple logical protocols. The overall goal of this research is to investigate efficient, adaptable neuromorphic computing solutions to increase resilience of the power grid under natural and adversarial failure conditions. To achieve this goal, the project addresses several objectives: (1) Deploying grid models onto real-time simulators with failure and attack scenarios; (2) Implementing spike representation encoding of system parameters; (3) Designing and optimizing spiking neuromorphic architectures; (4) Benchmarking performance of the neuromorphic approach using multiple criteria.Power grid operation will be simulated in real-time, for which system parameters and state variables will be streamed through a neuromorphic spiking neural network (SNN) platform. The SNN will interface with the system via standard protocols, encode the data in spike representations, and process these event-based spatiotemporal patterns (STPs) to detect correlations and anomalies. Real-time response will be provided through a separate response SNN. Modeling and simulation will also guide neuromorphic SNN architecture development, paired with hardware validation and testing.SNNs have been selected for the neuromorphic platform due to their ability to handle streaming information from a large number of input sources. SNNs are able to employ a wide variety of training algorithms on the global and local levels, and this effort will focus on evolutionary optimization (a genetic algorithm) paired with Hebbian spike timing-dependent plasticity (STDP) learning localized at the synapses. The STPs on which SNNs operate provide very rich, high-dimensional representations of information. Combining these methods holds particular promise for extremely accurate and rapid detection of abnormal operating conditions in a broad range of distributed systems including the power grid.Outcomes of the research are also applicable to many industrial settings utilizing a variety of supervisory control and data acquisition (SCADA) systems and to edge computing and other complex distributed networks for which reliability and resilience are of concern. Achieving these outcomes will improve energy transmission and availability and bolster development of novel intelligent system monitoring techniques and response capabilities.