Training Data Generation Strategies for Data-driven Security Assessment of Low Voltage Smart Grids

Juan J. Cuenca, Emanuel Aldea, Eloann Le Guern-Dall'o, Raphaël Féraud, Guy Camilleri, Anne BlavetteISGT EUROPE 2024 · DOI

TL;DR

This paper tackles the challenge of generating realistic training data for AI models that assess the security of low-voltage smart grids. By proposing novel data generation strategies beyond random sampling, the authors demonstrate that these new methods significantly improve model performance, especially when tested against real-world scenarios, highlighting the critical role of data quality in AI for grid security.

What Problem Does This Paper Address?

Modern electricity grids are becoming increasingly complex with the decentralization of energy resources. This makes it crucial for Distribution System Operators (DSOs) to have robust tools for "security assessment" – essentially, ensuring the grid operates within safe limits (voltage, current, etc.) to prevent outages. Traditionally, this is done using power flow (PF) simulations, which are computationally intensive. Data-driven methods, particularly AI, offer a faster alternative by learning to classify grid states as "safe" or "unsafe." However, a major hurdle is the lack of real-world data for training these AI models, especially for low-voltage grids where privacy concerns limit measurement collection. Current practice relies on synthetically generated data, often created through random sampling, but this approach has been shown to lead to models that perform poorly when faced with actual grid conditions. This paper addresses this gap by investigating new, more realistic data generation strategies for training AI models for smart grid security assessment.

Methods

The authors compare three popular data-driven methods:

  • Decision Trees (DT): A hierarchical rule-based system, good for interpretability.
  • Gradient Tree Boosting (GTB): An ensemble method known for its speed and scalability.
  • Deep Neural Networks (DNN): Capable of learning complex, non-linear relationships.

To generate training data, they propose six novel strategies that go beyond simple random sampling:

  • Guided Single Random Step: Iteratively adjusts load to move towards a balanced pool of safe/unsafe scenarios.
  • Guided Single Fixed Step: Similar to the above but uses a predefined step size.
  • Guided Global Random Step: Adjusts all nodes' consumption simultaneously.
  • Guided Single Random Step with Multiple Explorations: Combines guided steps with random restarts to explore diverse frontiers.
  • Guided Single Fixed Step with Multiple Explorations: Similar to the above but with fixed step sizes.

They also employ Synthetic Minority Oversampling Technique (SMOTE) for data augmentation to balance imbalanced datasets. The performance of models trained with these strategies is evaluated using precision and recall metrics, focusing on the "unsafe" classification.

Datasets

The study uses the IEEE European Low Voltage Test Network (ELVTN), a standard radial distribution system.

  • Training Data: Generated synthetically using the proposed strategies and SMOTE. One million operational points (OPs) were generated for each of the six strategies, resulting in 12 distinct training datasets (6 strategies + 6 augmented versions). Four subsets of these datasets (10³, 10⁴, 10⁵, 10⁶ OPs) were used to assess the impact of dataset size.
  • Testing Data: Two test datasets were created:
  • Test 1 (Random Generation): One million OPs generated randomly, mirroring the common practice in the literature.
  • Test 2 (Realistic Scenario): One million OPs generated by superposing permutations of daily load profiles from the ELVTN documentation and electric vehicle (EV) charging patterns. This dataset is considered more representative of real-world conditions.

Key Insights and Ideas

  • The "Random Generation" Trap: The paper strongly argues that the standard practice of randomly generating training data is insufficient and leads to models that fail to generalize to real-world scenarios. This is a critical finding for anyone working with data-driven models in similar domains.
  • Guided Strategies Shine: The novel "guided" data generation strategies, especially those involving multiple explorations or fixed steps, show a much greater ability to cover the operational space and produce models that perform well.
  • Data Quality Over Quantity (to an extent): While larger datasets generally improve performance, the quality and representativeness of the data generated by the novel strategies are shown to be far more impactful than simply having more randomly generated data.
  • DNNs Show Promise for Generalization: Deep Neural Networks, when trained with appropriate data generation strategies (like 'f' and 'd' with SMOTE), demonstrate the best potential for generalizing across both random and realistic test scenarios, even with smaller training sets.
  • Trade-off Between Training and Operation: The paper highlights a significant trade-off between the computational effort required for training models and the speed of operational evaluation. This is an important practical consideration for deploying AI in real-time grid management.

Key Results

  • Test 1 Performance: Models trained with data generation strategies that mimic the literature's random approach performed well on Test 1 (randomly generated test data). However, this success did not translate to Test 2.
  • Test 2 Performance: When tested on the realistic scenario (Test 2), most Decision Tree (DT) and Gradient Tree Boosting (GTB) models failed to achieve good performance (precision and recall above 0.98), regardless of the training data generation strategy.
  • Promising DNN Performance: DNN models trained using strategies 'f' and 'd' (with SMOTE) showed promising results on Test 2, even with a training dataset size of 10⁵ OPs, and also performed well on Test 1.
  • Computational Gains: Data-driven models offer substantial speedups over traditional PF simulations, ranging from 2x10³ to 6x10⁴ times faster for evaluation.
  • Training Time: Training times vary significantly, with DNNs exhibiting exponential increases with larger datasets, while DTs and GTBs show more linear scaling. Training effort can be orders of magnitude higher than operational effort.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that even the "realistic" test dataset is not a perfect representation of all real-world complexities.
  • No formal guarantees of generalization are provided for the proposed strategies.
  • Future Work:
    • Include a wider range of data-driven methods in the benchmark.
  • Investigate hyperparameter tuning more thoroughly to measure training computational efforts.
  • Utilize other open-source, synthetically generated datasets.
  • Explore "deep active learning" to enhance model resilience and reduce training times, especially given the observed trade-off between training and operational computational costs.

Summary updated

BibTeX
@inproceedings{cuenca2024training,
  author = {Juan J. Cuenca and Emanuel Aldea and Eloann Le Guern-Dall'o and Raphaël Féraud and Guy Camilleri and Anne Blavette},
  title = {{Training Data Generation Strategies for Data-driven Security Assessment of Low Voltage Smart Grids}},
  booktitle = {ISGT EUROPE},
  year = {2024},
  doi = {10.1109/isgteurope62998.2024.10863199},
}
Figure