ML-Based Load Shedding Framework

Updated 28 September 2025

Machine learning-based load shedding frameworks are advanced strategies that leverage deep and reinforcement learning to automate optimal emergency load reduction in power systems.
They utilize decentralized, data-driven policies and surrogate models to achieve scalable, robust, and real-time control even in complex grid environments.
Practical implementations demonstrate significant speedups and safety improvements compared to traditional rule-based methods, highlighting their potential for modern grid management.

Machine learning–based load shedding frameworks constitute a rapidly expanding research direction in power and event processing systems, leveraging data-driven or learning-based strategies to automate, decentralize, and accelerate corrective actions such as emergency load reduction. These frameworks address operational challenges posed by growing system complexity, increased uncertainty (e.g., through renewable integration or cyber-physical contingencies), and the need for safe, scalable, real-time control. A unifying feature is the use of ML, including deep and reinforcement learning, to construct optimal or near-optimal shedding strategies—often surpassing the limitations of traditional rule-based and optimization-based approaches.

1. Foundations and Key Principles

Load shedding is a critical emergency operation for ensuring voltage stability, frequency control, and constraint satisfaction within power grids or resource-limited event-processing systems. Traditional schemes (e.g., undervoltage, underfrequency, rule-based, or threshold-triggered) are known for their conservativeness, rigidity, and poor adaptiveness to evolving conditions. Recent research has shifted towards ML-based load shedding frameworks to address core limitations:

Scalability: Ability to handle high-dimensional control/action spaces as grid size or event complexity grows.
Adaptivity: Utilization of real-time or large-scale PMU/SCADA/event data for learning policies that generalize to previously unseen faults or operating conditions.
Decentralization: Design of local control policies leveraging only locally available measurements, drastically reducing communication bottlenecks.
Safety and Robustness: Integration of explicit safety constraints into learning or prediction, often through constrained RL formulations or post-hoc margin tuning.

New ML-based methods typically operate by either direct policy learning (e.g., actor-critic or random search in DRL), supervised mapping from local states to actions (e.g., neural network regression), or surrogate modeling to accelerate classic optimization (e.g., binding constraint identification via neural nets).

2. Deep Reinforcement Learning and Accelerated Policy Search

Deep reinforcement learning (DRL) has been intensively explored for adaptive grid stability and emergency voltage control. The Accelerated Random Search (ARS)-inspired PARS algorithm ("Accelerated Deep Reinforcement Learning Based Load Shedding for Emergency Voltage Control" (Huang et al., 2020)) is exemplary:

Key mechanism: PARS perturbs network policy weights $\theta_t$ in $N$ sampled directions $\{\delta_i\}$ , evaluates plus/minus rollouts, and aggregates top- $b$ directions with the highest reward differentials to update $\theta_{t+1}$ :

$\theta_{t+1} = \theta_t + \frac{\alpha}{b \sigma_b} \sum_{i=1}^{b} (r_{t,i}^+ - r_{t,i}^-) \delta_i$

where $\alpha$ is the step size and $\sigma_b$ is the standard deviation of top-b rewards.

Computational properties: Derivative-free, forward-pass–only structure; supports nested massive parallelization (across perturbations and tasks) using frameworks like Ray. Demonstrated $136\times$ speedups on IEEE 300-bus versus serial baselines.
Practical advantages: FNN and LSTM policies are supported; LSTMs capture temporal correlations directly, leading to 50% training time reduction versus FNN for similar performance in 39-bus trials. Consistently outperformed model-predictive control (MPC), PPO, and UVLS baselines in final rewards, convergence speed, and robustness.

3. Safety-Constrained Learning and RL Extensions

Safety guarantees in RL-based load shedding frameworks are addressed by augmenting the reward with explicit time-dependent safety functions and using constrained optimization. For example, "Safe Reinforcement Learning for Emergency LoadShedding of Power Systems" (Vu et al., 2020) applies a Lagrangian approach:

Formulation:

$\max_\pi \mathbb{E}[r(s_t, a_t)]\ \text{subject to}\ \mathbb{E}[f(s_t, a_t, t)] \geq 0$

with $f$ encoding transient voltage safety envelopes; the dual variable $\lambda$ is updated to penalize/favour safety compliance during training.

Implementation: Retains random search policy optimization, but polices combined reward and safety metrics; parallel rollout evaluations accelerate convergence. Empirically, policies learned this way met voltage recovery standards in all simulated fault cases and were robust to unseen contingencies.

4. Decentralized and Distributed Learning Approaches

Scalable learning-for-OLS frameworks treat each load center as an intelligent agent trained to make autonomous shedding decisions:

Algorithmic design ("Scalable Learning for Optimal Load Shedding Under Power Grid Emergency Operations" (Zhou et al., 2021); "Machine Learning for Scalable and Optimal Load Shedding Under Power System Contingency" (Zhou et al., 9 May 2024)):
- Each load center receives a local feedforward NN $f_i(x_i; \phi_i)$ mapping local voltage, flow, and demand features $x_i$ to the optimal shedding amount or dual variable $\alpha_i$ .
- Training is fully offline using AC-OLS optimization to generate pairs $(x_i, y_i)$ or $(x_i, \alpha_i)$ under a rich set of disturbances.
- Inference is decentralized, enabling sub-second online action even in high-dimensional (e.g., 2000-bus) grids, as there is no need for synchronous global optimization or state exchange.
- Demonstrated mean absolute errors on the order of $<1\%$ in simulated rollouts.
Distributed dual gradient tracking ("Distributed Dual Gradient Tracking for Priority-Considered Load Shedding" (Fitri et al., 2021)):
- Priority-based load shedding is formulated as a convex program with equality constraints representing priority groups.
- Distributed Lagrangian and dual scaling update each agent's slack and shedding variable using local minima; consensus-based dual updates ensure coordination without central control.
- Converges to centralized optima and preserves priority structure in simulations.

5. Data-driven Predictive Control and Koopman Operator Approaches

Data-driven model predictive control (MPC) for underfrequency load shedding (UFLS) leverages machine-learned models to approximate nonlinear grid dynamics:

KLS ("A Data-driven Under Frequency Load Shedding Scheme in Power Systems" (Cao et al., 2023)):
- Neural network–based latent extractor $\phi$ lifts a window of frequency and voltage measurements into a latent space where the system is linearly propagating under the Koopman operator $(g_{t+1} = Ag_t + B u_t)$ .
- Safety margin $\zeta$ is analytically selected to safeguard against error and discretization in load shedding actions.
- Achieves $<0.1$ Hz prediction error in 95% of cases and robust frequency stability across a broad range of perturbations, including non-envisioned faults.

6. Optimization, Event Systems, and Emerging ML Directions

Alternative ML-based load shedding frameworks target fairness, event stream management, or integration into classic optimization:

Fairness-aware load shedding ("Machine Learning for Fairness-Aware Load Shedding: A Real-Time Solution via Identifying Binding Constraints" (Zhou et al., 25 Jul 2024)): Neural nets are trained to instantaneously predict the set of binding constraints given the current load vector, reducing the original QP to a KKT-based linear system. Achieves millisecond-level solution times for both economic and equity-aware load shedding objectives, verified on RTS-GMLC and toy systems.
Complex event stream processing ("gSPICE: Model-Based Event Shedding in Complex Event Processing" (Slo et al., 2023)): Probabilistic models and ML classifiers (decision trees, random forests) are used to estimate event utility for selective event dropping, maintaining low latency and high quality under overload. Feature-rich event descriptions and efficient storage techniques (e.g., Zobrist hashing) further increase applicability to real-world event streams.

7. Challenges, Limitations, and Future Directions

ML-based load shedding frameworks face several theoretical and practical challenges:

Training Data Representativeness and Coverage: The generalization ability of trained policies critically depends on the diversity and completeness of offline simulation scenarios. Extreme or rare contingencies may require targeted data generation.
Safety and Robustness Guarantees: Integrating explicit, physics-informed safety margins, risk-aware loss functions (CVaR), or robust constraint satisfaction remains a priority for deployment in safety-critical applications.
Real-time Computation and Scalability: Leveraging decentralized architectures enables massive scaling, but highly distributed implementations depend on reliable low-latency measurement and inference infrastructure.
Integration with Physics-based Models: AC-aware models ("AC-aware Optimization Framework for Under-Frequency Load Shedding" (Elsaadany et al., 6 Jan 2025)) show that retaining or embedding non-linear AC power flow and voltage effects significantly increases fidelity for modern grids, especially under high DER penetration.
Interfacing with Classical Optimization: ML surrogates (e.g., regression trees, Tobit models) can be embedded in MILP for UFLS estimation and operational scheduling (Rajabdorri et al., 2023).
Event/Resource-Type Extension: ML-driven load shedding methods are actively extending to distributed, priority-adaptive, and event-driven contexts, matching the requirements of both classic power systems and modern event streams.

These directions suggest an ongoing convergence between data-driven policy learning, robust optimization, and domain-aware modeling in the field of intelligent emergency control and load shedding.