Power System Dispatch Based on Improved Scenario Division with Physical and Data-Driven Features

Wenqi Huang, Shang Cao, Lingyu Liang, Huanming Zhang, Xiangyu Zhao, Hanju Li, Jie Ren, Liang CheEnergies 2023 · DOI

TL;DR

This paper tackles the challenge of efficiently dispatching power systems with high renewable energy penetration by proposing a novel multi-scenario approach. By combining data-driven features from Variational Autoencoders (VAEs) with crucial physical system indicators, they create a more robust scenario division method. This improved division then enables a more accurate and faster reinforcement learning-based dispatch model, outperforming traditional methods in decision-making efficiency and renewable energy integration.

What Problem Does This Paper Address?

The increasing integration of renewable energy sources (like solar and wind) into power grids introduces significant uncertainty and variability. Traditional power system dispatch methods, often based on complex physical models, struggle to adapt to these rapid changes and can be computationally intensive, hindering real-time decision-making. Data-driven methods, particularly those using reinforcement learning (RL), offer faster decision-making and better adaptability. However, a major hurdle for these RL methods is the "distributional shift" problem: their performance degrades when the real-world operating scenarios differ from the limited scenarios they were trained on. This paper aims to bridge this gap by developing a more comprehensive scenario division technique that better captures the complexities of power system operations, thereby improving the robustness and efficiency of data-driven dispatch.

Methods

The core of this paper lies in its improved scenario division method. It combines two types of features:

  • Data-Driven Features: A Variational Autoencoder (VAE) is employed to extract latent features from high-dimensional historical power system operational data. The VAE's encoder maps the data to a lower-dimensional representation, and its decoder reconstructs the data, minimizing reconstruction loss to capture essential patterns.
  • Physical Features: Recognizing that VAEs might lose some critical information, the authors incorporate key physical indicators related to generation and load. These include peak load, average load, renewable peak and average output, and the renewable peak-to-valley difference. These features are chosen for their interpretability and ability to reflect the system's operational state.

These combined features are then fed into a Gaussian Mixture Model (GMM) for scenario division. The GMM models the data as a mixture of Gaussian distributions, where each component represents a distinct operational scenario. The optimal number of scenarios is determined using the Akaike Information Criterion (AIC), and the effectiveness of the division is validated using the Silhouette Coefficient (SC).

Finally, a multi-scenario data-driven dispatch model is established. For each identified scenario, a specific reinforcement learning agent is trained using the Proximal Policy Optimization (PPO) algorithm. This allows the system to select the appropriate agent for the current operating scenario, leading to faster and more accurate dispatch decisions.

Datasets

The simulations were conducted using 300 days of historical grid data from a real power system in a province of China. This dataset included information on 748 nodes, 55 renewable units, and 187 thermal units, with a dispatch interval of 15 minutes. The data was pre-processed and normalized before training. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Synergy of Data and Physics: The paper strongly advocates for the combined use of data-driven (VAE) and physical features for scenario division. This hybrid approach offers a more comprehensive and interpretable representation of the power system's state than either method alone.
  • Addressing Distributional Shift: By creating more distinct and representative scenarios, the proposed method directly tackles the distributional shift problem that plagues single-scenario RL approaches, leading to more reliable dispatch decisions.
  • GMM for Scenario Granularity: The Gaussian Mixture Model proves effective in clustering the complex operational data into meaningful scenarios, allowing for tailored dispatch strategies.
  • RL for Real-time Dispatch: The use of PPO-based RL agents for each scenario enables rapid decision-making, crucial for the dynamic nature of modern power grids.

Key Results

  • Improved Scenario Division: The proposed method, integrating physical and VAE features, achieved a Silhouette Coefficient (SC) of 0.4522, significantly outperforming the VAE-only features (SC of 0.0093), indicating a much better clustering effect.
  • Enhanced RL Training: The multi-scenario approach led to higher mean reward curves during agent training compared to single-scenario methods, demonstrating improved learning efficiency and effectiveness. For instance, in scenario A, the mean reward increased by 43.51% compared to VAE-based scenarios.
  • Faster Decision-Making: The proposed method achieved an average single-step decision time of 0.003 seconds, a dramatic improvement over traditional optimization methods which took 19.82 seconds, representing a 99.98% increase in efficiency.
  • Better Renewable Energy Integration: The renewable energy consumption rate was significantly improved with the proposed scenario division method.
  • Reduced Operating Costs: The overall power system operating cost was reduced by 5.196% with the proposed method compared to without scenario division.

Limitations and Future Work

  • Scenario B Performance: The paper notes that scenario B, with fewer samples, showed a lower mean reward during training compared to the original scenario. This suggests that even with scenario division, very sparse scenarios might still pose challenges for RL agent training.
  • OPT vs. RL Trade-off: While the proposed RL method offers superior decision-making speed, traditional optimization methods (OPT) can sometimes achieve higher mean rewards due to their ability to find global optima in complex problems. The paper acknowledges this trade-off.
  • Future Work: The authors suggest that further research could explore more advanced scenario division techniques and potentially hybrid approaches that combine the strengths of both optimization and reinforcement learning for even more robust and efficient power system dispatch.

Summary updated

Abstract

In power systems with high penetration of renewable energy, traditional physical model-based optimal dispatch methods suffer from modeling difficulties and poor adaptability, while data-driven dispatch methods, represented by reinforcement learning, have the advantage of fast decision making and reflecting long-term benefits. However, the performances of data-driven methods are much limited by the problem of distribution shift under insufficient power system scenario samples in the training. To address this issue, this paper proposes an improved scenario division method by integrating the power system’s key physical features and the data-driven variational autoencoder (VAE)-generated features. Next, based on the scenario division results, a multi-scenario data-driven dispatch model is established. The effectiveness of the proposed method is verified by a simulation conducted on a real power system model in a province of China.
BibTeX
@inproceedings{huang2023power,
  author = {Wenqi Huang and Shang Cao and Lingyu Liang and Huanming Zhang and Xiangyu Zhao and Hanju Li and Jie Ren and Liang Che},
  title = {{Power System Dispatch Based on Improved Scenario Division with Physical and Data-Driven Features}},
  booktitle = {Energies},
  year = {2023},
  doi = {10.3390/en16227520},
}
Figure