Edge Convolution Predicts Power Grid Carbon Flow Under Renewable Energy Uncertainty

A new predictive framework using edge-conditioned convolution models dynamic carbon flow rates across power systems operating under renewable energy uncertainty. Developed by researchers at the State Grid Gansu Electric Power Research Institute of Economics and Technology, the computational method addresses the limitations of conventional carbon calculation models designed for traditional electrical grids. Traditional accounting methods typically rely on deterministic power flow calculations, which struggle to capture the operational volatility introduced by fluctuating generation sources and changing electrical loads under high renewable energy penetration.

To establish a foundation for dynamic accounting, the research team created a carbon flow rate calculation framework that incorporates the full lifecycle carbon emissions of renewable generators alongside a network loss allocation mechanism. Rather than assigning transmission losses through single-direction assumptions, this mechanism integrates upstream and downstream power flow tracing to divide network losses between the generation side and the consumption side via a designated allocation coefficient.

Building on this framework, the investigators developed a deep learning architecture based on edge-conditioned convolution to process transmission network structures directly. The edge-conditioned convolution weights enable the model to integrate specific transmission branch attributes during information aggregation across interconnected electrical nodes. To maintain consistency with physical network behavior, the training objective incorporates a physics-informed loss function formulated to encourage the conservation of the carbon flow rate throughout the simulated power grid.

To train and evaluate the architecture under diverse grid conditions, the authors generated multi-scenario datasets utilizing Sobol sequences and Inverse Transform Sampling algorithms. This sampling strategy extends the operational scenarios of the prediction model to power systems with substantial renewable energy integration. The generated multi-scenario data sets also serve to enhance the overall generalization performance of the edge-conditioned convolution network when subjected to variable source and load patterns.

In validation case studies, the edge-conditioned convolution model achieved a mean absolute error of approximately 0.144 tonnes of carbon dioxide per hour.

Further comparative analysis showed that the proposed approach characterizes carbon flow rate distributions across different renewable penetration levels more accurately than conventional carbon flow tracing methods. The model also adapted more effectively to source-load uncertainty across test scenarios, mitigating the errors that typically arise when deterministic flow equations encounter non-stationary supply and demand conditions.

The study authors noted that the model provides a potential computational basis for low-carbon power system dispatch, supporting China’s broader dual-carbon targets through fine-grained emission tracking.

The study was published in Frontiers in Energy Research within the Smart Grids section, identifying the work as an original research article. The authors, Chong Shao, Yongcheng Liu, Qian Yang, Boyang Chen, and Yalu Sun, are affiliated with the State Grid Gansu Electric Power Research Institute of Economics and Technology in Lanzhou, China. The manuscript was received on 11 September 2026 and accepted for publication on 26 October 2026. The article is distributed under a Creative Commons Attribution License, permitting use and reproduction provided the original authors and journal are credited.

The research specifically targets the modeling requirements of new-type power systems, which differ significantly from the traditional grids for which most existing carbon calculation tools were designed. The authors argue that the realization of China’s dual-carbon goals depends on the ability to perform accurate and fine-grained carbon emission modeling. Conventional methods, which rely on deterministic power flow calculations, are ill-suited for environments with high renewable penetration because they cannot adequately capture the impacts of source and load uncertainty. By shifting to a predictive model based on Edge Conditioned Convolution, the study aims to bridge this gap, offering a tool that can handle the variability inherent in modern electrical networks.

The carbon flow rate calculation framework explicitly includes the full lifecycle carbon emissions of renewable generators, ensuring that the environmental impact of the energy source is accounted for from the outset. Additionally, the model employs a network loss allocation mechanism that uses an allocation coefficient to distribute losses between the generation and consumption sides. This approach moves beyond simple directional assumptions, utilizing both upstream and downstream power flow tracing to provide a more nuanced accounting of where carbon emissions are effectively allocated within the grid structure.

The application of Sobol sequences and Inverse Transform Sampling algorithms is central to the model’s ability to generalize. These algorithms generate multi-scenario datasets that simulate various operational conditions, allowing the prediction model to be tested against a wide range of potential source and load configurations. This synthetic data generation process is designed to enhance the model’s robustness, ensuring that it can maintain accuracy even when faced with the non-stationary conditions typical of high renewable-energy penetration systems.

Subject of Research: Deep learning prediction of carbon flow rates in power grids with high renewable penetration

Article Title: An edge-conditioned convolution model for carbon flow rate prediction under source-load uncertainty

Article References: Shao, C., Liu, Y., Yang, Q., Chen, B., & Sun, Y. (2026). An Edge-Conditioned Convolution Model for Carbon Flow Rate Prediction under Source-Load Uncertainty. Frontiers in Energy Research, 14, 1995670. https://doi.org/10.3389/fenrg.2026.1995670

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1995670

Keywords: carbon flow rate, edge-conditioned convolution, source-load uncertainty, renewable energy, power systems, physics-informed loss function, network loss allocation, power flow tracing, Sobol sequences, Inverse Transform Sampling, smart grids, low-carbon dispatch, dual-carbon goals, lifecycle emissions, transmission networks, energy research