Power Grid Behavioral Patterns and Risks of Generalization in Applied Machine Learning

Shimiao Li,Jan Drgona,Shrirang Abhyankar,Larry Pileggi

2023-06-02

Abstract:Recent years have seen a rich literature of data-driven approaches designed for power grid applications. However, insufficient consideration of domain knowledge can impose a high risk to the practicality of the methods. Specifically, ignoring the grid-specific spatiotemporal patterns (in load, generation, and topology, etc.) can lead to outputting infeasible, unrealizable, or completely meaningless predictions on new inputs. To address this concern, this paper investigates real-world operational data to provide insights into power grid behavioral patterns, including the time-varying topology, load, and generation, as well as the spatial differences (in peak hours, diverse styles) between individual loads and generations. Then based on these observations, we evaluate the generalization risks in some existing ML works causedby ignoring these grid-specific patterns in model design and training.

Systems and Control,Machine Learning

What problem does this paper attempt to address?

The paper primarily aims to address a series of generalization risk issues that arise when applying machine learning methods in power systems due to the neglect of grid-specific behavior patterns (such as topology, temporal and spatial variations in load and generation).

Specifically:

1. **Research Background**: In recent years, data-driven methods for power systems have been increasing, including methods for tasks such as state estimation and power flow analysis. However, these methods often overlook the spatiotemporal behavior patterns unique to power grids, leading to infeasible or meaningless prediction results in practical applications.

2. **Core Issue**: Ignoring the physical knowledge specific to power systems, such as topology, load demand, and spatiotemporal characteristics, may result in machine learning models producing inaccurate or even invalid results when faced with new inputs. For example, models that do not consider changes in grid topology cannot effectively generalize on dynamic graphs; models that do not fully consider the variability in load distribution may perform poorly on real system configurations.

3. **Research Content**:
- **Spatiotemporal Behavior Pattern Analysis**: By studying real grid data, the research reveals the behavior patterns of grid topology, total load and generation, as well as individual load and generation units, and the physical reasons behind them.
- **Risk Assessment of Existing Methods**: Based on the above observations, the research evaluates the generalization risks faced by some existing machine learning works due to the neglect of these grid-specific patterns during design and training.

4. **Experimental Validation**: Through anomaly detection experiments and data-driven optimal power flow (AC-OPF) tasks, the research demonstrates the impact of different data generation strategies on model performance, emphasizing the importance of considering grid-specific patterns.