Many important application data streams can be modeled as a complex graph of entities, where each graph node or entity is associated with a multi-variate time-series. The overall system’s behavior is modeled as a dynamical system, and it evolves through structural changes in the graph, and/or changes in the node-level measurements. Such networks include the power grid, smart cities and even social networks driven by countless handheld devices. Due to the enormity, complexity and dynamic nature of these systems real-time state prediction is infeasible to realize with centralized architectures. We propose to develop a framework where a distributed network of autonomous agents (subsequently referred to as “holons”) will coordinate to detect events in such dynamic networks and provide interpretable recommendations to users for alerting against critical events.
The framework will be co-designed with a power grid application as a starting use case. It is well known that the integration of distributed energy resources and the growth of advanced sensors have led to a tremendous increase in the volume and throughput of data from the power grid. Due to the existing centralized system architecture, use of physical model driven controls, and lack of advanced data-driven techniques, today’s operational models are not capable of utilizing all the data for better observability and control. A concrete artifact from this project will be an engine that recommends mitigation strategies that can be employed as a distributed application in the power grid to prevent failures that have the potential to cascade, which are the cause of most large-scale system outages.
We aim to make the following contributions for domain-aware machine-learning:
- Develop physics-informed neural network models of a networked dynamical system by combining recent developments in graph representation learning and deep learning for approximating ordinary differential questions. Using the power grid use case we will develop and test data-driven predictive models that capture the power system transient evolution for variable prediction lengths without the need to retrain the model when subjected to a disturbance. These will serve as benchmarks for performing prediction in the centralized setting.
- Develop methodologies for automated learning and maintenance of domain-specific knowledge bases (KB) that capture symbolic representations of events that manifest in the dynamic network.
- Develop new capabilities for advancing autonomous distributed decision making:
- Unify domain-aware ML and data analytics for prediction of transient states in networked dynamical systems using a distributed architecture.
- Compare performances of data parallel and distributed data parallel methodologies (where in data parallel the process is computed in multiple GPUs in a single server whereas in distributed data parallel the process is logically divided to be performed in multiple servers).
- Unify domain-aware ML and HPC to simulate models for power grid use case and develop distributed predictions via aggregation of localized models with graph attention network.
- Determine feasibility of employing edge computing paradigm to power grid applications (which are currently implemented in a centralized manner) for providing local observability to regional agents and reducing computation delays.