⚡ electrical engineering

Power Grid Operational Risk Assessment Based on Source-Grid Multi- dimensional Feature Fusion and Multi-gate Mixture-of-Experts Model

This paper proposes a power grid operational risk assessment method utilizing a multi-gate mixture-of-experts (MMoE) model that fuses multi-dimensional source-grid features and empirical dispatch error simulations to achieve high-accuracy prediction and adaptive interpretability of spatiotemporal balancing risks in renewable energy-integrated systems.

Original authors: Shupeng Zhang, Chenguang Qiu, Chaojie Ding, Jie Li, Xing Luo

Published 2026-10-27

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Shupeng Zhang, Chenguang Qiu, Chaojie Ding, Jie Li, Xing Luo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The modern electrical grid is no longer just a one-way street where power flows from giant factories to homes. It has become a complex, two-way system where the weather itself acts as a generator. As nations shift toward cleaner energy, wind turbines and solar panels are replacing traditional power plants. However, the sun does not always shine, and the wind does not always blow with the same force. This unpredictability creates a delicate balancing act. If too much renewable energy is generated at once, the grid can become overloaded, forcing operators to waste clean power. If too little is generated, the grid can struggle to meet demand, risking blackouts. Managing this tension requires knowing exactly when and where these risks will appear, but the sheer number of variables makes this a daunting task for human planners.

Researchers at State Grid Jiangsu Electric Power Co., Ltd. and State Grid Nanjing Power Supply Company have developed a new way to see these risks before they happen. They created a computer system that acts like a highly experienced dispatcher, capable of learning from the past to predict the future. Instead of relying on simple rules or looking at wind and solar data in isolation, this system fuses information from the power sources, the physical network of wires, and the history of how the grid has behaved over time. By combining these different types of information, the system can identify two specific dangers: the risk of having to waste renewable energy because it cannot be used, and the risk of not having enough power to keep the lights on.

To teach their system what to look for, the researchers first had to create a realistic training ground. They built a detailed simulation that mimics how grid operators actually work. This simulation runs in two stages. First, it plans the day ahead based on what the weather is expected to be, deciding which power plants should turn on and how much reserve power to keep ready. Then, it runs a second stage that simulates what happens when the actual weather arrives, which is often different from the forecast. By running this simulation thousands of times with different possible errors in the weather forecast, the system generates a massive library of "what-if" scenarios. These scenarios serve as the answer key, teaching the computer what a dangerous situation looks like in real life.

The core of their new method is a sophisticated learning model called a multi-gate mixture-of-experts. Imagine a team of specialists, each with a different area of focus. One specialist is an expert at reading spreadsheets of numbers, such as how much wind or solar power is expected and how much electricity people are likely to use. Another specialist is an expert at maps, understanding how the physical layout of power lines and the connections between different cities affect the flow of electricity. A third specialist is an expert at time, looking at how wind and solar output change hour by hour to spot dangerous trends.

In older systems, a single model tries to do everything at once, often getting confused when different types of information conflict. In this new approach, the system uses a smart switchboard, or "gating" mechanism, to decide which specialist should lead the conversation for any given moment. If the risk comes from a sudden drop in wind power, the system listens to the specialist who tracks time. If the risk comes from a bottleneck in the power lines, it listens to the map expert. This allows the system to handle two opposing problems at the same time: the risk of wasting power and the risk of running out of power. These two risks often move in opposite directions, a situation that confuses simpler models, but this team-based approach navigates the conflict smoothly.

The results of their testing show that this method works with high precision. When the system was tested on real data from a provincial power grid, it correctly identified situations where renewable energy would be wasted 92.7% of the time. It also correctly identified situations where the grid would face a shortage of power 84.2% of the time. Perhaps more importantly, the system does not just give a warning; it explains why. Because the system knows which specialist was most active in making the decision, it can tell operators exactly where the problem lies. It can point to a specific group of power lines that are under stress, or a specific time of day when the wind is likely to die down.

This ability to explain its reasoning is a major breakthrough. Many advanced computer programs are "black boxes," meaning they give an answer without showing the work, leaving human operators unsure of whether to trust the warning. This new system is transparent. It can trace a risk back to its source, showing that a potential shortage is due to a specific combination of low wind forecasts and a heavily loaded power line in a particular region. In one test case, the system correctly identified that a specific set of buses, or connection points on the grid, were the primary sources of risk, matching the actual conditions of the day perfectly. It even distinguished between buses where power was actually being wasted and those where the risk was simply high due to the flow of electricity nearby.

The researchers found that the system is particularly good at avoiding false alarms when predicting wasted energy, getting it right 92.7% of the time with a very low error rate. When predicting power shortages, it is slightly more cautious, sounding the alarm a bit more often to ensure safety, which is the desired behavior for a system designed to prevent blackouts. By combining a realistic simulation of grid operations with a team-based learning model, the researchers have created a tool that not only predicts danger but also illuminates the path to solving it. This approach offers a clear, interpretable way to manage the complex, shifting balance of a modern power grid, ensuring that the transition to clean energy remains stable and reliable.

Technical Summary: Power Grid Operational Risk Assessment Based on Source-Grid Multi-dimensional Feature Fusion and Multi-gate Mixture-of-Experts Model

1. Problem Statement

The transition to a new-type power system characterized by high penetration of renewable energy (wind and photovoltaic) introduces significant stochasticity and volatility. This creates complex challenges for grid operational risk assessment, specifically regarding the simultaneous management of curtailment risk (renewable absorption limitations) and security of supply risk (upward regulation pressure).

Existing risk assessment methods face three primary deficiencies:

Neglect of Spatiotemporal Coupling: Traditional methods often fail to capture the dynamic coupling between physical boundaries, source-side uncertainties, and grid topology.
Deviation from Actual Operations: Many probabilistic models rely on decoupled scenario generation or excessive reliance on wind/PV data-driven approaches, leading to risk labels that do not align with actual dispatch outcomes.
Poor Interpretability: Current neural network approaches often function as "black boxes." While post-hoc explanation methods (e.g., SHAP, GNNExplainer) exist, they are computationally expensive and unstable, failing to provide direct, actionable risk tracing for dispatch decisions.

2. Methodology

The paper proposes a comprehensive framework integrating a two-stage dispatch simulation for label generation and an Improved Multi-gate Mixture-of-Experts (MMoE) model for risk perception.

A. Risk Label Generation (Two-Stage Dispatch Simulation)
To ensure risk labels reflect actual operational constraints, the authors establish a simulation framework driven by empirical forecast error distributions:

Day-Ahead Pre-dispatch: Determines unit commitment and reserve plans based on forecast values, minimizing generation costs, startup/shutdown costs, and penalties for curtailment and line flow violations.
Intraday Redispatch: Fixes the day-ahead unit commitment but substitutes actual wind, PV, and load data. This stage allows for load shedding, deep peak regulation, and emergency slack, directly propagating forecast errors into physical risks (curtailment energy and reserve call amounts).
Label Aggregation: By sampling $K$ error scenarios from historical empirical distributions and aggregating the dispatch results, the system generates probabilistic risk labels (event occurrence probability) and expected power labels (mean curtailed energy or reserve call).

B. Risk Perception Model (Improved MMoE)
The core assessment model is an MMoE architecture designed to handle the negative correlation between curtailment and security of supply risks while fusing heterogeneous data sources:

Expert Modules:
- MLP Expert: Processes source-side and schedule-side tabular features (e.g., forecasts, reserve capacity, unit status).
- Hybrid Graph Neural Network (GCN + GAT) Expert: Captures grid-side topology features and spatial dependencies using Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT).
- GRU Expert: Extracts temporal correlation features from the previous 4 hours of renewable output and load data.
Multi-gate Control: Unlike standard MMoE models that share gates, this approach employs task-specific independent gating networks for curtailment and security of supply risks. This design mitigates gradient conflicts arising from the negative correlation between the two risks. The model selects the top-2 experts for each task to achieve sparse, task-specific fusion.
Output & Loss: The model outputs expected power values and event probabilities. The loss function combines Mean Squared Error (MSE) for numerical predictions, Binary Cross-Entropy (BCE) for probabilities, and a load-balancing term to prevent expert collapse.
Probability Calibration: Isotonic regression is applied post-training to calibrate predicted probabilities against actual event frequencies, ensuring reliability.

C. Native Interpretability and Risk Tracing
A key feature of the proposed method is its ability to provide adaptive interpretability without post-hoc computation:

Information Type Tracing: Gating weights indicate which expert (Tabular, Topology, or Temporal) dominates the risk prediction.
Spatial Tracing: Graph attention coefficients identify critical transmission lines and buses contributing to risk.
Temporal Tracing: GRU hidden state norms highlight critical time periods driving the risk.

3. Key Contributions

Physical-Consistent Label Generation: The establishment of a two-stage dispatch framework driven by empirical error distributions ensures that risk labels are physically grounded and consistent with actual dispatch constraints, addressing the deviation issues in traditional scenario-based methods.
Task-Specific MMoE Architecture: The introduction of independent gating networks for negatively correlated risks effectively alleviates gradient conflicts, enabling joint assessment of curtailment and security of supply risks.
Intrinsic Interpretability: The model utilizes gating weights and graph attention coefficients to provide structured risk tracing (information type, spatial location, temporal location) directly during forward propagation, eliminating the need for unstable post-hoc explanation tools.
Multi-dimensional Feature Fusion: The successful integration of tabular, topological, and temporal features allows for a holistic assessment of spatiotemporal coupling characteristics in the power grid.

4. Experimental Results

The method was evaluated on a full-year dataset from a provincial power grid (35,040 dispatch intervals), comparing against Gradient Boosting Decision Trees (GBDT) and Multi-head Graph Attention Networks (MGAT).

Prediction Accuracy: The proposed method achieved a determination accuracy of 92.7% for curtailment events and 84.2% for security of supply events. These results outperformed GBDT and MGAT by 1.0–4.9 percentage points.
Risk Identification Performance:
- Curtailment: 85.5% hit rate with a low false-alarm rate of 4.8%.
- Security of Supply: 88.5% hit rate with a false-alarm rate of 21.0%. The authors note this asymmetry is intentional, prioritizing the avoidance of missed alarms for security of supply.
Calibration: Reliability diagrams confirmed that isotonic regression effectively corrected systematic biases, aligning predicted probabilities with observed frequencies.
Tracing Validation: Case studies demonstrated that the model's tracing results (e.g., identifying buses 61, 43, and 8 as high-risk locations) were consistent with actual operational data, validating the model's ability to identify both actual curtailment sources and topologically critical nodes.

5. Significance

The paper claims that this method addresses the critical gap in current power system risk assessment by providing a tool that is not only highly accurate but also interpretable and physically consistent. By moving away from black-box models and decoupled scenario generation, the proposed approach enables dispatch operators to:

Accurately identify the source of risks (whether driven by forecast errors, grid topology, or temporal trends).
Trace risks to specific spatial locations (buses/lines) and temporal windows.
Make informed decisions based on a model that inherently understands the complex coupling between renewable stochasticity and grid constraints.

The authors conclude that this framework offers a stable and reproducible pathway for intelligent risk assessment in new-type power systems, with future work planned to incorporate diverse energy storage types and demand-side response mechanisms.