Keunju Song
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Keunju Song has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Hongseok Kim
Sogang University
7 shared publications - Minsoo Kim
Department of Energy Engineering, Korea Institute of Energy Technology, 21, Kentech-gil, Naju-si, Jeollanam-do, Republic of Korea
4 shared publications Hongseok Kim
3 shared publicationsYoungmin Choi
3 shared publications- Haejoong Lee
Sogang University
2 shared publications Hien Thanh Doan
2 shared publications
Latest Publications
- 7 September 2026
- Published by Elsevier in Electric Power Systems Research
- Vol. 265
Topology transferable physics-informed graph learning for national-scale AC optimal power flow
- 7 September 2026
- Published by Elsevier in Electric Power Systems Research
- Vol. 265
Topology transferable physics-informed graph learning for national-scale AC optimal power flow
- Open Access
- 5 March 2026
- Published by Elsevier in SSRN Electronic Journal
Alternative Learning with Semi-Supervised Relaxation for AC Optimal Power Flow
Recent advancements in next-generation power systems highlight the need for fast and scalable AC optimal power flow (AC-OPF) solutions, as traditional solvers struggle with computational demands on large networks. This paper introduces AltOPF, a machine-learning (ML) framework that employs an alternating learning strategy between the original nonconvex AC-OPF problem and a relaxed auxiliary form. By embedding augmented-Lagrangian multipliers into a compact deep neural network, AltOPF enforces operational constraints while improving solution quality and inference speed. We further establish an ergodic convergence guarantee to Clarke stationarity for the resulting alternating stochastic subgradient method. Empirically, we evaluate AltOPF on IEEE 162-, 300-, and 1354-bus systems and a real 4492-bus Korea Power Exchange (KPX) network. Across these cases, AltOPF achieves a 99% constraint satisfaction compared with six established baselines, and delivers microsecond-level inference for AC-OPF instances. It also reduces checkpoint size by 75% and peak GPU memory by about 40% under identical batch settings, demonstrating the potential for real-time, resource-efficient AC-OPF on large-scale systems.
- 5 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAlternative Learning with Semi-Supervised Relaxation for AC Optimal Power Flow
Recent advancements in next-generation power systems highlight the need for fast and scalable AC optimal power flow (AC-OPF) solutions, as traditional solvers struggle with computational demands on large networks. This paper introduces AltOPF, a machine-learning (ML) framework that employs an alternating learning strategy between the original nonconvex AC-OPF problem and a relaxed auxiliary form. By embedding augmented-Lagrangian multipliers into a compact deep neural network, AltOPF enforces operational constraints while improving solution quality and inference speed. We further establish an ergodic convergence guarantee to Clarke stationarity for the resulting alternating stochastic subgradient method. Empirically, we evaluate AltOPF on IEEE 162-, 300-, and 1354-bus systems and a real 4492-bus Korea Power Exchange (KPX) network. Across these cases, AltOPF achieves a 99% constraint satisfaction compared with six established baselines, and delivers microsecond-level inference for AC-OPF instances. It also reduces checkpoint size by 75% and peak GPU memory by about 40% under identical batch settings, demonstrating the potential for real-time, resource-efficient AC-OPF on large-scale systems.
- Open Access
- 4 March 2026
- Published by Elsevier in SSRN Electronic Journal
Alternative Learning with Semi-Supervised Relaxation for AC Optimal Power Flow
Recent advancements in next-generation power systems highlight the need for fast and scalable AC optimal power flow (AC-OPF) solutions, as traditional solvers struggle with computational demands on large networks. This paper introduces AltOPF, a machine-learning (ML) framework that employs an alternating learning strategy between the original nonconvex AC-OPF problem and a relaxed auxiliary form. By embedding augmented-Lagrangian multipliers into a compact deep neural network, AltOPF enforces operational constraints while improving solution quality and inference speed. We further establish an ergodic convergence guarantee to Clarke stationarity for the resulting alternating stochastic subgradient method. Empirically, we evaluate AltOPF on IEEE 162-, 300-, and 1354-bus systems and a real 4492-bus Korea Power Exchange (KPX) network. Across these cases, AltOPF achieves a 99% constraint satisfaction compared with six established baselines, and delivers microsecond-level inference for AC-OPF instances. It also reduces checkpoint size by 75% and peak GPU memory by about 40% under identical batch settings, demonstrating the potential for real-time, resource-efficient AC-OPF on large-scale systems.
- 4 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAlternative Learning with Semi-Supervised Relaxation for AC Optimal Power Flow
Recent advancements in next-generation power systems highlight the need for fast and scalable AC optimal power flow (AC-OPF) solutions, as traditional solvers struggle with computational demands on large networks. This paper introduces AltOPF, a machine-learning (ML) framework that employs an alternating learning strategy between the original nonconvex AC-OPF problem and a relaxed auxiliary form. By embedding augmented-Lagrangian multipliers into a compact deep neural network, AltOPF enforces operational constraints while improving solution quality and inference speed. We further establish an ergodic convergence guarantee to Clarke stationarity for the resulting alternating stochastic subgradient method. Empirically, we evaluate AltOPF on IEEE 162-, 300-, and 1354-bus systems and a real 4492-bus Korea Power Exchange (KPX) network. Across these cases, AltOPF achieves a 99% constraint satisfaction compared with six established baselines, and delivers microsecond-level inference for AC-OPF instances. It also reduces checkpoint size by 75% and peak GPU memory by about 40% under identical batch settings, demonstrating the potential for real-time, resource-efficient AC-OPF on large-scale systems.
- Open Access
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Foreseeable Implicit Training for ESS Operation based on Deep Reinforcement Learning
The recent increase in behind-the-meter renewable energy (RE) has accelerated the deployment of residential energy storage systems (ESS). However, uncertainty in RE generation presents significant challenges in ESS operation, such as maintaining power balance and minimizing costs. To address these issues, we propose a deep reinforcement learning (DRL) framework. Unlike conventional DRL approaches, the proposed method reformulates the control problem as a reference tracking task, where the agent learns to follow reference trajectories generated offline by an optimization solver. This design enables the agent to implicitly consider future uncertainties while adapting to Time-of-Use (TOU) price signals without relying on any forecasting model. Additionally, we introduce a knowledge peering structure, in which peer actors guide each other during training to improve generalization and prevent overfitting. Our method achieves up to 82% improvement in optimality gap compared to existing baselines and demonstrates near-optimal performance without requiring explicit future information. Extensive experiments confirm that the proposed algorithm, named FITNESS, is a robust and scalable solution for real-time ESS control under uncertainty.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Open AccessForeseeable Implicit Training for ESS Operation based on Deep Reinforcement Learning
The recent increase in behind-the-meter renewable energy (RE) has accelerated the deployment of residential energy storage systems (ESS). However, uncertainty in RE generation presents significant challenges in ESS operation, such as maintaining power balance and minimizing costs. To address these issues, we propose a deep reinforcement learning (DRL) framework. Unlike conventional DRL approaches, the proposed method reformulates the control problem as a reference tracking task, where the agent learns to follow reference trajectories generated offline by an optimization solver. This design enables the agent to implicitly consider future uncertainties while adapting to Time-of-Use (TOU) price signals without relying on any forecasting model. Additionally, we introduce a knowledge peering structure, in which peer actors guide each other during training to improve generalization and prevent overfitting. Our method achieves up to 82% improvement in optimality gap compared to existing baselines and demonstrates near-optimal performance without requiring explicit future information. Extensive experiments confirm that the proposed algorithm, named FITNESS, is a robust and scalable solution for real-time ESS control under uncertainty.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 170661-170679
Meta-PV: Wide-Area PV Forecasting With Robust Meta-Learning Under Insufficient Observations
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 170661-170679
Open AccessMeta-PV: Wide-Area PV Forecasting With Robust Meta-Learning Under Insufficient Observations
- Open Access
Power System Decision Making in the Age of Deep Learning: A Comprehensive Review
Modern power systems are facing growing complexity and uncertainty due to electrification, large-scale renewable integration, and evolving consumption behaviors. These changes have pushed traditional numerical optimization methods to their practical limits in terms of scalability and real-time applicability. In response, deep learning approaches that offer fast inference and robustness to uncertainty are gaining significant attention. This paper presents, to the best of our knowledge, the first systematic review from a functional perspective of deep learning research supporting power system decision making. Taking a functional perspective, we classify neural networks into three core roles: learning to predict, learning to surrogate, and learning to optimize. In the first role, neural networks forecast exogenous uncertainties serving as the instance input to operational optimization problems. In the second role, they approximate complex physical constraints, enabling the efficient formulation of problems that would otherwise be analytically intractable. In the third role, neural networks act as learning-based optimizers that either replace or augment conventional solvers. The core purpose of this paper is to emphasize that neural networks should not simply be regarded as generic data-driven tools, but rather as models serving distinct functional roles—each with its own objectives and considerations. In this regard, we introduce diverse approaches aligned with these roles, offering conceptual foundations for principled application in practice. Such functional insights will ultimately guide the design of modular and hybrid architectures that integrate these roles, which in turn may provide the basis for developing domain-specific foundation models for power system operations.
Open AccessPower System Decision Making in the Age of Deep Learning: A Comprehensive Review
Modern power systems are facing growing complexity and uncertainty due to electrification, large-scale renewable integration, and evolving consumption behaviors. These changes have pushed traditional numerical optimization methods to their practical limits in terms of scalability and real-time applicability. In response, deep learning approaches that offer fast inference and robustness to uncertainty are gaining significant attention. This paper presents, to the best of our knowledge, the first systematic review from a functional perspective of deep learning research supporting power system decision making. Taking a functional perspective, we classify neural networks into three core roles: learning to predict, learning to surrogate, and learning to optimize. In the first role, neural networks forecast exogenous uncertainties serving as the instance input to operational optimization problems. In the second role, they approximate complex physical constraints, enabling the efficient formulation of problems that would otherwise be analytically intractable. In the third role, neural networks act as learning-based optimizers that either replace or augment conventional solvers. The core purpose of this paper is to emphasize that neural networks should not simply be regarded as generic data-driven tools, but rather as models serving distinct functional roles—each with its own objectives and considerations. In this regard, we introduce diverse approaches aligned with these roles, offering conceptual foundations for principled application in practice. Such functional insights will ultimately guide the design of modular and hybrid architectures that integrate these roles, which in turn may provide the basis for developing domain-specific foundation models for power system operations.
- 28 August 2025
- Published by Elsevier in Applied Energy
- Vol. 401
Locational scenario-based pricing in a bilateral distribution energy market under uncertainty
- 28 August 2025
- Published by Elsevier in Applied Energy
- Vol. 401
Locational scenario-based pricing in a bilateral distribution energy market under uncertainty
- 21 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Sustainable Energy
- Vol. 16 (1) , 160-173
Graph-Based Large Scale Probabilistic PV Power Forecasting Insensitive to Space-Time Missing Data
In recent years, power systems integrated with distributed energy resources (DERs) have been considered to mitigate climate change. However, this makes power systems even more uncertain and complex, so uncertainty-aware accurate forecasting needs to be considered for the massive penetration of renewable energy. To this end, we propose a scalable and missing-insensitive framework for probabilistic multi-site photovoltaic (PV) power forecasting, specifically focused on large-scale PV sites and space-time missing data. By leveraging the graph neural network (GNN), the proposed scalable graph learning mechanism with random coarse graph attention and probabilistic spatio-temporal learning performs efficiently for large-scale PV sites in terms of forecasting accuracy and model training complexity. At the same time, our framework adaptively imputes the missing PV data in the space and time domain, respectively. Ablation study results demonstrate that our framework is effective for extracting complex spatial-temporal features across large-scale PV sites. Under extensive experiments, our framework shows 7 $-$ 10% and 6 $-$ 25% improvement on average for over 1600 PV sites and three types of space-time missing data, which ensures accurate and stable forecasting.
- 21 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Sustainable Energy
- Vol. 16 (1) , 160-173
Graph-Based Large Scale Probabilistic PV Power Forecasting Insensitive to Space-Time Missing Data
In recent years, power systems integrated with distributed energy resources (DERs) have been considered to mitigate climate change. However, this makes power systems even more uncertain and complex, so uncertainty-aware accurate forecasting needs to be considered for the massive penetration of renewable energy. To this end, we propose a scalable and missing-insensitive framework for probabilistic multi-site photovoltaic (PV) power forecasting, specifically focused on large-scale PV sites and space-time missing data. By leveraging the graph neural network (GNN), the proposed scalable graph learning mechanism with random coarse graph attention and probabilistic spatio-temporal learning performs efficiently for large-scale PV sites in terms of forecasting accuracy and model training complexity. At the same time, our framework adaptively imputes the missing PV data in the space and time domain, respectively. Ablation study results demonstrate that our framework is effective for extracting complex spatial-temporal features across large-scale PV sites. Under extensive experiments, our framework shows 7 $-$ 10% and 6 $-$ 25% improvement on average for over 1600 PV sites and three types of space-time missing data, which ensures accurate and stable forecasting.
- Open Access
Convolutional Autoencoder-Based Anomaly Detection for Photovoltaic Power Forecasting of Virtual Power Plants
Machine learning-based time-series forecasting has recently been intensively studied. Deep learning (DL), specifically deep neural networks (DNN) and long short-term memory (LSTM), are the popular approaches for this purpose. However, these methods have several problems. First, DNN needs a lot of data to avoid over-fitting. Without sufficient data, the model cannot be generalized so it may not be good for unseen data. Second, impaired data affect forecasting accuracy. In general, one trains a model assuming that normal data enters the input. However, when anomalous data enters the input, the forecasting accuracy of the model may decrease substantially, which emphasizes the importance of data integrity. This paper focuses on these two problems. In time-series forecasting, especially for photovoltaic (PV) forecasting, data from solar power plants are not sufficient. As solar panels are newly installed, a sufficiently long period of data cannot be obtained. We also find that many solar power plants may contain a substantial amount of anomalous data, e.g., 30%. In this regard, we propose a data preprocessing technique leveraging convolutional autoencoder and principal component analysis (PCA) to use insufficient data with a high rate of anomaly. We compare the performance of the PV forecasting model after applying the proposed anomaly detection in constructing a virtual power plant (VPP). Extensive experiments with 2517 PV sites in the Republic of Korea, which are used for VPP construction, confirm that the proposed technique can filter out anomaly PV sites with very high accuracy, e.g., 99%, which in turn contributes to reducing the forecasting error by 23%.
Open AccessConvolutional Autoencoder-Based Anomaly Detection for Photovoltaic Power Forecasting of Virtual Power Plants
Machine learning-based time-series forecasting has recently been intensively studied. Deep learning (DL), specifically deep neural networks (DNN) and long short-term memory (LSTM), are the popular approaches for this purpose. However, these methods have several problems. First, DNN needs a lot of data to avoid over-fitting. Without sufficient data, the model cannot be generalized so it may not be good for unseen data. Second, impaired data affect forecasting accuracy. In general, one trains a model assuming that normal data enters the input. However, when anomalous data enters the input, the forecasting accuracy of the model may decrease substantially, which emphasizes the importance of data integrity. This paper focuses on these two problems. In time-series forecasting, especially for photovoltaic (PV) forecasting, data from solar power plants are not sufficient. As solar panels are newly installed, a sufficiently long period of data cannot be obtained. We also find that many solar power plants may contain a substantial amount of anomalous data, e.g., 30%. In this regard, we propose a data preprocessing technique leveraging convolutional autoencoder and principal component analysis (PCA) to use insufficient data with a high rate of anomaly. We compare the performance of the PV forecasting model after applying the proposed anomaly detection in constructing a virtual power plant (VPP). Extensive experiments with 2517 PV sites in the Republic of Korea, which are used for VPP construction, confirm that the proposed technique can filter out anomaly PV sites with very high accuracy, e.g., 99%, which in turn contributes to reducing the forecasting error by 23%.
- Open Access
DTTrans: PV Power Forecasting Using Delaunay Triangulation and TransGRU
In an era of high penetration of renewable energy, accurate photovoltaic (PV) power forecasting is crucial for balancing and scheduling power systems. However, PV power output has uncertainty since it depends on stochastic weather conditions. In this paper, we propose a novel short-term PV forecasting technique using Delaunay triangulation, of which the vertices are three weather stations that enclose a target PV site. By leveraging a Transformer encoder and gated recurrent unit (GRU), the proposed TransGRU model is robust against weather forecast error as it learns feature representation from weather data. We construct a framework based on Delaunay triangulation and TransGRU and verify that the proposed framework shows a 7–15% improvement compared to other state-of-the-art methods in terms of the normalized mean absolute error. Moreover, we investigate the effect of PV aggregation for virtual power plants where errors can be compensated across PV sites. Our framework demonstrates 41–60% improvement when PV sites are aggregated and achieves as low as 3–4% of forecasting error on average.
Open AccessDTTrans: PV Power Forecasting Using Delaunay Triangulation and TransGRU
In an era of high penetration of renewable energy, accurate photovoltaic (PV) power forecasting is crucial for balancing and scheduling power systems. However, PV power output has uncertainty since it depends on stochastic weather conditions. In this paper, we propose a novel short-term PV forecasting technique using Delaunay triangulation, of which the vertices are three weather stations that enclose a target PV site. By leveraging a Transformer encoder and gated recurrent unit (GRU), the proposed TransGRU model is robust against weather forecast error as it learns feature representation from weather data. We construct a framework based on Delaunay triangulation and TransGRU and verify that the proposed framework shows a 7–15% improvement compared to other state-of-the-art methods in terms of the normalized mean absolute error. Moreover, we investigate the effect of PV aggregation for virtual power plants where errors can be compensated across PV sites. Our framework demonstrates 41–60% improvement when PV sites are aggregated and achieves as low as 3–4% of forecasting error on average.