Optimizing energy storage plant discrete system dynamics analysis with graph convolutional networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Energy Systems Engineering

Background:

• Suboptimal performance and high computational costs challenge current Graph Convolutional Network (GCN) models in energy storage power plant optimization.
• Existing GCN models struggle with efficient information flow and feature amalgamation for diverse network topologies.

Purpose of the Study:

• To introduce a novel packet-switched graph convolutional network (PS-GCN) to address limitations in GCN-based energy storage power plant optimization.
• To enhance model performance by improving information exchange and feature fusion across different network groupings and time sequences.

Main Methods:

• Developed a GCN extreme learning machine as a foundational model.
• Introduced a group exchange graph convolution module utilizing group graph convolution techniques for feature amalgamation.
• Designed a timing depth separation convolution module, including a packet-switching mechanism with 1x1 convolutional layers for inter-packet information exchange.

Main Results:

• The proposed PS-GCN model achieved RMSE of 46.08 and MAE of 26.22 for single-step prediction at 60 minutes.
• Demonstrated significant error reduction in multi-step predictions: 14.71% RMSE decrease at 15 minutes and 9.29% at 60 minutes compared to benchmark models.

Conclusions:

• The novel packet-switched graph convolutional network effectively optimizes energy storage power plants by improving prediction accuracy and reducing computational complexity.
• Enhanced model performance leads to increased operational efficiency and reliability, especially crucial for dynamic time-series data in energy storage systems.