Introduction
The integration of renewable energy sources (RESs) and electric vehicles (EVs) is a cornerstone of the modern sustainable smart grid. However, these technologies introduce significant stochasticity that challenges grid stability. Traditional control methods often struggle to balance operational costs with reliability when faced with "messy" or incomplete data. This paper presents a novel Wasserstein-based Distributionally Robust Model Predictive Control (WDR-MPC) framework designed to manage both static and dynamic uncertainties while maintaining computational efficiency.
The Challenge of Uncertainty in Smart Grids
Grid operators must contend with two distinct categories of uncertainty:
- Static Uncertainties: These are instantaneous errors in forecasting, such as the immediate power output of a solar array or wind turbine.
- Dynamic Uncertainties: These errors propagate through time. A primary example is the charging pattern of an EV; an unpredictable charging event doesn't just impact the current load but alters the battery's State of Charge (SoC), thereby influencing all future control decisions.
Existing stochastic methods often assume a known probability distribution (e.g., a Gaussian distribution) for these errors. In practice, the true distribution is rarely known, and ignoring the "memory" inherent in battery systems can lead to conservative operations or, worse, grid failure.
Methodology: The WDR-MPC Framework
The proposed WDR-MPC approach addresses these challenges through a two-stage optimization process that leverages Wasserstein ambiguity sets.
Wasserstein Ambiguity Sets
Rather than assuming a single fixed probability distribution, the model constructs a "ball" of all possible distributions that are close to historical data, as measured by the Wasserstein metric. This ensures the system remains robust against "out-of-sample" scenarios where real-world data deviates from historical patterns.
Two-Stage Optimization and Ambiguity Tubes
- Stage 1 (Robustness): The model establishes "ambiguity tubes." These are mathematical bounds that define the worst-case evolution of dynamic errors over a given time horizon. By decoupling these uncertainties from the primary control logic, the system ensures that constraints—such as voltage limits—are never violated.
- Stage 2 (Optimization): The stochastic problem is reformulated into a "nominal" convex form. This simplified version is pre-hardened against uncertainty, allowing for rapid solving without the computational overhead typically associated with robust optimization.
Computational Acceleration
A significant technical contribution of this work is a reformulation method that prevents the number of constraints from growing linearly with the number of data samples. This convex acceleration allows the WDR-MPC to remain scalable for large-scale power systems, a common bottleneck in distributionally robust frameworks.
Experimental Validation
The researchers validated the WDR-MPC framework using the IEEE 38-bus and IEEE 94-bus systems. To ensure realistic testing, the simulations incorporated:
- Actual solar power generation data.
- New York electricity demand records from June 2019.
- Market pricing data from Finland (June 2022).
Key Results
- Reliability: In out-of-sample testing, the WDR-MPC achieved nearly 100% reliability, significantly outperforming standard MPC and Sample Average Approximation (SAA) methods, which often failed to maintain constraints under high uncertainty.
- Speed: The acceleration technique proved vital. On the 38-bus system with 100 samples, the accelerated WDR-MPC solved in approximately 2,660 seconds. In contrast, the standard WDR-MPC formulation failed to converge even after 50,000 seconds.
- Cost Efficiency: By being distributionally robust rather than purely robust (which assumes the absolute worst-case), the system avoided overly conservative settings, leading to lower operational costs as more data samples were integrated.
Conclusion and Future Work
The WDR-MPC framework provides a scalable, reliable solution for managing the complex uncertainties of modern smart grids. While the current model assumes a centralized controller for radial distribution networks, future research will explore distributed scenarios for local coordination and the integration of advanced machine learning predictors to further refine forecasting accuracy within the WDR framework.
Summary updated