TL;DR
This paper introduces a novel data-driven approach to multi-stage optimal power flow (OPF) for electrical grids, bypassing the need for explicit network topology and parameter identification. By leveraging Willems' fundamental lemma and descriptor system theory, the method directly learns system dynamics from historical operational data, enabling efficient and robust power scheduling, particularly in the presence of renewable energy sources and battery storage.
What Problem Does This Paper Address?
The efficient operation of large-scale power systems relies heavily on numerical optimization, specifically multi-stage optimal power flow (OPF). Traditionally, this requires a detailed understanding of the grid's topology, line impedances, and bus specifications. Obtaining this information involves a complex, multi-step process of identifying the network graph and estimating its parameters. This process is not only computationally intensive but also prone to errors and requires significant tuning. The authors aim to overcome these challenges by developing a data-driven method that can construct and solve the OPF problem directly from measured operational data, eliminating the need for explicit model identification. This is particularly relevant for modern grids with increasing complexity due to renewable energy integration and the deployment of energy storage systems like Battery Energy Storage Systems (BESS).
Methods
The core of the proposed method lies in applying concepts from behavioral systems theory and data-driven predictive control to the power system's DC power flow model.
- Descriptor System Representation: The power system's dynamics, including BESS charge/discharge cycles and power balance equations, are first formulated as a linear time-invariant descriptor system. This representation captures both the dynamic (BESS state) and algebraic (power flow) aspects of the system.
- Willems' Fundamental Lemma: The authors leverage Willems' fundamental lemma, which provides a non-parametric representation of linear systems using input-output trajectories. This allows them to construct Hankel matrices from historical data, which encapsulate the system's behavior without explicit model identification.
- Data-Driven Predictive Control: Using these Hankel matrices, a data-driven optimal control problem (OCP) is formulated. This OCP directly optimizes future control actions (generator and BESS dispatch) based on past and predicted future operational data, subject to system constraints.
- Segmentation and Regularization: To improve computational performance and handle potential noise in real-world data, the authors employ trajectory segmentation and introduce a quadratic regularization term in the objective function. This regularization helps prevent overfitting, especially when dealing with noisy measurements.
The key novelty is the direct application of these data-driven control techniques to the multi-stage OPF problem, bypassing the traditional model identification pipeline.
Datasets
The paper uses a modified IEEE 118-bus system for evaluation. This system is augmented with Battery Energy Storage Systems (BESS). The authors generated operational data for this system to train and test their data-driven controller. They do not introduce a new benchmark dataset but rather use a well-established power system model for validation.
Key Insights and Ideas
- Eliminating Model Identification: The most significant insight is that explicit knowledge of grid topology and line parameters is not strictly necessary for solving the multi-stage OPF problem. Data-driven methods can learn the system's behavior directly from operational history.
- Bridging Control Theory and Power Systems: The paper successfully bridges advanced concepts from control theory (descriptor systems, Willems' lemma) with the practical challenges of power system operation, demonstrating their applicability beyond traditional control domains.
- Data-Driven OPF as a Viable Alternative: The results suggest that data-driven OPF can be a competitive alternative to traditional model-based approaches, especially in scenarios where obtaining accurate models is difficult or costly.
- Implicit Topology Learning: The data-driven approach implicitly learns aspects of the network topology and parameters through the Hankel matrices, without requiring explicit estimation algorithms.
Key Results
- The proposed data-driven multi-stage OPF (DD OPF) controller achieved performance comparable to a classical approach using exact topology data. The realized closed-loop costs were only 0.5% higher than the exact model-based approach.
- The DD OPF controller does not require knowledge of generator, demand, or BESS locations, which is a significant advantage over traditional methods.
- While the OCP solution time for DD OPF was higher (5.09s median) compared to the exact DC model (0.82s) and sequential ID (0.91s), it is considered satisfactory for the chosen time scale (15-minute intervals).
- The resulting power flows, even with regularization, were kept within the specified upper bounds, demonstrating the controller's ability to respect system constraints.
Limitations and Future Work
- Measurement Noise: The paper acknowledges that the theoretical foundation (Willems' lemma) assumes noise-free data. While regularization is used to mitigate the impact of noise, it introduces a bias. The authors suggest exploring more advanced noise handling techniques.
- Extension to AC Power Flow: The current work focuses on the DC power flow model for simplicity. Future work will involve extending the data-driven approach to the more realistic AC power flow model.
- Inclusion of Voltage Data: The current approach primarily uses power flow and BESS state data. Incorporating voltage data could further enhance the controller's performance and robustness.
- Data Requirements: While the method bypasses explicit model identification, it still requires a sufficiently long and persistently exciting historical trajectory for training. The paper discusses lower bounds for this data length.
Summary updated