【天津大学】2024人工智能驱动的分布式光伏数据虚拟采集技术报告

Summary of Virtual Collection Technology for Distributed PV Data Empowered by Artificial Intelligence

Core Content

Leijiao Ge, an Associate Professor at Tianjin University, presented a research on Virtual Collection Technology for Distributed PV Data Empowered by Artificial Intelligence. The technology aims to address the challenges of data monitoring in distributed photovoltaic (DPV) systems by using artificial intelligence to infer power data from a subset of reference power stations (RPSs) rather than deploying extensive sensing networks.

Main Points

1. Personal Information Introduction

• Name: Leijiao Ge
• Birth Year: 1984
• Education:
  • Bachelor's Degree: Beihua University (2002.9–2006.7)
  • Master's Degree: Hebei University of Technology (2006.9–2009.1)
  • PhD: Tianjin University (2012.9–2016.6)
• Work Experience:
  • Lecturer at Tianjin University (2016.11–2021.7)
  • Postdoctoral Fellow at Tianjin University (2017.3–2021.3)
  • Master's Supervisor and Doctoral Associate Supervisor at Tianjin University (2019.1–Now)
  • Associate Professor at Tianjin University (2021.7–Now)
• Academic Positions:
  • IEEE Senior Member
  • Chairman of the 2023 ICEPT Conference
  • Executive Director of IEEE PES Youth Expert Committee
• Editorial Roles:
  • Special Editor in Chief for 12 journals
• Invited Reports:
  • Over 30 invited reports at academic forums
• Publications:
  • Published 82 SCI papers (including 4 ESI papers and 3 hot topic papers)
  • Authored 6 monographs (chapters)
• Patents:
  • 48 authorized invention patents
• Standards:
  • Revised/developed 7 international and domestic industry standard specifications

2. Feasibility Study Performance

• Awards:
  • Received 24 provincial and ministerial level scientific and technological progress awards
  • 10 provincial government awards and 14 industry awards
• Projects:
  • Led 6 national level projects and 6 provincial/ministerial level projects
• Research Focus:
  • Intelligent distribution network situation awareness

3. Virtual Collection Technology

3.1 Hybrid Virtual Connection Method of IMOWOA and LightGBM

• Contribution 1:
  • Proposed an autoencoder similarity analysis method based on gridded area division to identify photovoltaic power station sets with similarity requirements.
  • Divided the area using geographical location and environmental conditions.
  • Screened power stations based on reconstruction errors to ensure accurate data inference.
• Contribution 2:
  • Introduced an improved multi-objective whale optimization algorithm (IMOWOA) to optimize the selection of RPSs and LightGBM hyperparameters.
  • Enhancements include:
    • Initialization strategy based on reverse learning
    • Leader selection mechanism based on crowded distance
    • Local binary conversion function
• Case Study:
  • Conducted a case study on 29 DPVs in Jiangsu Province.
  • Results showed that the method could reduce collection costs significantly by using a subset of PVs for data acquisition.

3.2 Virtual Data Connection Model Considering Spatiotemporal Coupling and Affine Optimization Reference

• Contribution 1:
  • Proposed a deep trained recursive denoising autoencoder (D-RDAE) to capture the spatiotemporal correlation of DPV data.
  • Added memory refresh and reproduction modules to improve the model's performance.
• Contribution 2:
  • Constructed affine artificial neural networks (AANN) to overcome the uncertainty of solar radiation intensity.
  • AANN outperformed traditional SO and RO methods in reducing MAPE.
• Case Study:
  • Conducted virtual collection testing on 33 DPVs in Nanjing, Jiangsu Province.
  • D-RDAE reduced the error of traditional DAE by nearly 70%.
  • AANN reduced MAPE of RO and SO by up to 1.8% and 0.9%, respectively.

4. Demonstration Application

• Application in Badong County, Nanjing City:
  • Deployed the DPV virtual collection software module on the National New Energy Cloud Platform.
  • Applied to 371 DPV power stations in Nanjing, Jiangsu, and other areas.
  • Improved operational efficiency of situational awareness by 14%.
• Performance Metrics:
  • Power injection error at power flow nodes: Not exceeding 5%
  • Node voltage error: Below 2.5%
  • Electricity theft detection accuracy: Over 95%
  • Photovoltaic prediction accuracy improved
  • Communication costs reduced

5. Application and Prospect

• Future Research Directions:
  • Propose a "real-time+virtual" state perception method for intelligent distribution networks.
  • Build a potential risk analysis system for intelligent distribution networks.
  • Develop lean operation and disturbance resistance technology for intelligent distribution networks.
  • Address the dual optimization challenge of economic and technological trends in situational guidance.
• Empowerment Scenarios:
  • Demand-driven applications
  • High reliability and high-quality power supply under multiple uncertain conditions of source, network, load, and storage

Key Information

• Technology Focus:

  • Virtual collection technology for DPV data using AI.
  • Combines gridded area division and autoencoder similarity analysis.
  • Utilizes IMOWOA and LightGBM for optimal RPS selection.
  • Employs AANN for handling solar radiation uncertainty.

• Applications:

  • Applied in Jiangsu Province and Nanjing City.
  • Integrated into the intelligent operation and maintenance system.
  • Used in State Grid Beijing Electric Power Company's projects.

• Outcomes:

  • Reduced data collection costs.
  • Improved accuracy in power injection, voltage, and theft detection.
  • Enhanced system perception and operational efficiency.

• Supporting Achievements:

  • 24 provincial and ministerial level awards.
  • 6 monographs and 48 invention patents.
  • 7 industry standard specifications.

This technology represents a significant advancement in the efficient and accurate management of distributed photovoltaic systems through the use of AI and optimization algorithms, with promising applications in real-world scenarios.