The emergence of a new-type power system characterized by high renewable integration and complex cyber-physical interactions poses significant challenges in forecasting, analysis, and decision making. Artificial intelligence for Science (AI4S) offers a promising paradigm combining data-driven learning with physical knowledge. This paper investigated the application of AI4S to key power system functions, including renewable and load forecasting, analysis and computing, and optimization and decision making. Insights from scientific domains such as protein folding and atmospheric simulation highlighted the scalability and potential of this paradigm. Critical technical challenges are analyzed, including multi-scale prediction, and source-grid-load-storage coordination. The foundational principles and core technologies underpinning AI4S in renewable and load forecasting, analysis and computing, and optimization and decision making are presented, addressing fundamental issues such as temporal dependency learning, nonlinear system modeling, and high-dimensional optimization. Finally, future application trends are discussed, focusing on the evolution of AI4S itself and the development of comprehensive model evaluation frameworks.

Multi energy flow coupling and clean energy substitution are significant characteristics of Multi-Energy Systems (MES). For the coordinated operation of interconnected MESs, it is crucial to construct an operation mechanism that considers both regional autonomous operation and multilateral interaction. However, the intertwined coupling between intermittent renewable energy and speculative behaviors of participants poses considerable challenges to the game decision making of multiple stakeholders. To this end, this paper proposes a distributed robust hybrid game mechanism to support the collaborative operation of interconnected MESs. This enables us to construct a bilevel robust game model within the MES to coordinate the multilevel interactions between the operator and photovoltaic (PV) prosumers. In the upper level, the Multi Energy System Operator (MESO) adopts distributionally robust optimization based on multiple discrete scenarios to formulate operation plans satisfying system constraints. In the lower level, the energy interactions among PV prosumers are modeled as a Generalized Nash Equilibrium model with distributionally robust chance constraints, and variational inequalities are utilized to precisely characterize its optimality conditions. At the inter-MES interaction level, an asymmetric Nash bargaining model is introduced to pursue the maximization of social welfare. Subsequently, a quantitative analysis model of MES speculative behavior is proposed to reveal the specific impact of such behavior on the distributed operation mechanism. To efficiently solve the aforementioned hybrid game model, this paper proposes a distributed optimization algorithm incorporating robust operation constraints to obtain the optimal scheduling strategy. Numerical case studies demonstrate that the proposed method effectively quantifies speculative behavior and enhances system economy and robustness under uncertain scenarios.

Accurate uncertainty quantification (UQ) is pivotal for mitigating the operational risks associated with the inherent intermittency and variability of renewable energy sources. However, existing methods, such as Bayesian inference, often struggle to adequately capture spatial heterogeneity among distributed sites and evolving temporal patterns. To this end, we propose a Spatio-Temporal Adaptive Conformal Forecasting (STACF) framework that integrates spatio-temporal graph neural networks with adaptive online conformal inference. STACF captures the inherent spatio-temporal uncertainty of renewable generation by employing a dynamic graph diffusion network for feature extraction. We further apply a control conformal inference mechanism to adaptively calibrate forecast uncertainty. This technique allows for valid uncertainty quantification without relying on restrictive distributional assumptions. Furthermore, we establish a theoretical guarantee for the asymptotic coverage validity of the proposed framework under spatio-temporal distribution shifts. Finally, we validate our proposed method on multi-regional wind datasets from the United States, covering 1395 wind farms. These results verify the superior performance of our proposed method compared to state-of-the-art baselines.

The proliferation of distributed energy resources introduces multi-source uncertainties, including implicit uncertainties arising from third-party operators’ partial observability of security constraints, challenging traditional distribution network planning methods dependent on model simplification and predefined scenarios. We address this gap via an adaptive hierarchical learning architecture that co-optimizes distributed energy resources location, capacity, and operational strategies data-drivenly, enabling autonomous learning of implicit constraints without full model knowledge. Our framework embeds a bi-level Stackelberg structure where Monte Carlo Tree Search autonomously generates planning schemes at the upper level, while multi-agent reinforcement learning directly learns operational policies from real-time data at the lower level under partial observability. Validation on both benchmark and large-scale practical distribution systems shows lower investment costs and faster solutions while maintaining voltage stability, demonstrating superior scalability and adaptiveness to implicit uncertainties versus scenario-based methods. Yue Xiang and colleagues propose an adaptive hierarchical learning framework to address implicit uncertainties in distributed energy resource planning when third-party operators lack full grid visibility. Their method co-optimizes investment and operational decisions data-drivenly, reducing costs while maintaining voltage stability without relying on predefined scenarios or model simplifications.

With the continuous expansion of data centers, their potential to serve as new flexible loads in power system demand response services has become increasingly significant. In response to the problem of low internal resource utilization efficiency in single-site data center demand response under a multi-loop power supply structure, this paper first proposes a coordinated response strategy. The strategy improves overall response capability by jointly scheduling the UPS backup battery and flexibly adjusting the topological connections between UPS systems, photovoltaic and energy storage systems. Furthermore, a block order mechanism based on integrated learning is developed to address the multi-period non-convex optimization problems and user privacy challenges that arise during the bidding process in the electricity market. A neural network is used to fit user bidding behavior, which is then embedded into the Independent System Operator's market-clearing optimization model as a Mixed-Integer Linear Programming formulation. This integration enables scheduling decisions to be made under privacy-preserving conditions. Finally, case studies are conducted on the modified PJM 5-bus system and the IEEE 118-bus system. The results show that the proposed strategy increases the regulation potential of data centers by 26%, while improving their revenue compared with the traditional block orders approach.

With the development of smart grid, the significant role of artificial intelligence technologies in microgrid state analysis and energy scheduling is becoming increasingly evident. The microgrids constantly generate a massive amount of computational requests. However, the distributed architecture of microgrids poses challenges for processing computing requests, such as limited resources at single nodes and load imbalance. In this paper, we propose FLARE, an end-edge-cloud collaboration computing architecture in microgrids, to optimize request scheduling. The optimization objective of the scheduling algorithm is to maximize long-term comprehensive utility considering factors such as delay and energy consumption. We utilize Deep Reinforcement Learning (DRL) to tackle this NP-hard problem. To address the deployment and update issues of distributed schedulers on end devices, we introduce model pruning and Federated Learning (FL). Experimental results demonstrate that FLARE outperforms state-of-the-art heuristic and centralized algorithms in terms of performance and scalability.

Modem power systems serve as a critical component of new energy architectures and are essential carriers for achieving the dual carbon goals. With increasing renewable energy penetration and widespread use of various uncertain and flexible distributed resources, the variety of generation and load types in power systems will continue to expand, resulting in a dramatic increase in operational complexity and control challenges. Collective intelligence, identified as one of the key research directions in China's new generation artificial intelligence development plan, can leverage multi-agent cooperative optimization capabilities to support coordinated optimization and control analysis in modern power systems. This paper begins with a brief overview of collective intelligence technologies. The current state of key research scenarios in modern power systems is then analyzed, with a focus on the emerging features and challenges encountered in these environments. Subsequently, the application potential of collective intelligence in modern power systems is summarized, along with a review of relevant research efforts in this domain. Finally, future research directions are discussed, highlighting advanced technologies that are expected to further enhance the integration of collective intelligence into modern power system control and optimization.

This paper proposes a multi-agent cooperative operation optimization strategy for regional power grids considering the uncertainty of renewable energy output and flexibility of electric vehicle (EV) scheduling, which not only improves the economy of networked microgrid (NMG) scheduling but also reduces the impact on active distribution network (ADN). EV condition matrix and model of the adjustable charge-and-discharge capacity of the EV may be built up by simulating the trip rule of an EV using the driving behavior of the vehicle model. In the day-ahead stage, by taking into account NMG operating cost, distribution network loss, and EV owners' payment cost, a multi-objective optimal scheduling model is developed, and the day-ahead scheduling contract for EV is obtained. Generative Adversarial Network (GAN) generates a significant number of intraday scenarios of photovoltaic (PV), load, and EV based on historical scheduling data as training data for the intra-day scheduling model multi-agent PPO (MAPPO). In the intra-day scheduling stage, intra-day ultra-short-term forecast data is input into the intra-day scheduling model, and the trained multi-agent model realizes NMG distributed real-time optimal scheduling. Finally, the economy and effectiveness of the proposed strategy are verified by Day-after optimal scheduling results.

Unit Commitment (UC), a core power market clearing problem, is non-deterministic polynomial hard (NP-hard), leading to the "dimensionality curse" in large-scale power grids; traditional numerical methods fail to adapt to the growing number of market participants, while existing artificial intelligence (AI)-based approaches lack interpretability. To address these issues, this paper proposes a distributed optimization method with Transformer-based warm-start for large-scale local market clearing: it uses the Alternating Direction Method of Multipliers (ADMM) to decompose the problem into subproblems for handling multi-participant constraints, and designs an encoder-decoder Transformer to extract UC temporal features via self-attention, generating feasible initial solutions from historical data to reduce optimizer iterations. Validated on an IEEE 33-bus system (with 3 dispatchable distributed generations (DGs), 1 PV farm, and 1 wind farm), results show the traditional linear programming solver takes 120 seconds, while the proposed method reduces calculation time to 48 seconds, with optimization error rising slightly from 0.1% to 0.5% (practically acceptable), effectively balancing efficiency and accuracy to meet Distributed System Operators (DSOs)’ real-time needs.

The operation of power grid has intensively complicated due to grid expansion and renewable resources, posing unprecedented challenges for power grid dispatching. Traditional data-driven dispatching methods are highly dependent on the quantity and quality of data samples. To address the optimization problem of regional power grid dispatching, this paper proposes a data-knowledge hybrid-driven multi-agent optimization scheduling model. The model combines the multi-agent deep reinforcement learning framework with scheduling knowledge by incorporating scheduling experience knowledge into the agent's loss function as regularization terms, guiding the agent to take reasonable and effective actions during the optimization process and improving learning efficiency. The model aims to minimize system operation costs, reduce wind and solar curtailment and load shedding, and considers constraints such as nodal power balance, line flow constraints, and unit output limits. Experimental results show that compared with traditional data-driven reinforcement learning algorithms, this method has significant advantages in terms of training time, convergence performance, cost-effectiveness, and security.

As modern power systems increasingly integrate renewable energy sources and face stricter reliability requirements, conventional dispatch strategies struggle to meet the demands of adaptability, safety, and autonomy. This paper presents a novel intelligent operator for optimal power system dispatch, designed to autonomously perceive system states, analyze constraints, and make decisions within a safety-aware learning framework. The operator is built upon a Safe Reinforcement Learning (Safe RL) paradigm, modeled as a Constrained Markov Decision Process (CMDP), to ensure both economic efficiency and strict adherence to operational safety constraints such as generator limits, ramping capabilities, and power balance. The proposed approach leverages Constrained Policy Optimization (CPO), integrating real-time constraint critics and a safety correction layer to maintain policy feasibility throughout the learning and deployment phases. Extensive simulations on an enhanced IEEE 39-bus system demonstrate that the intelligent operator achieves near-optimal dispatch performance while markedly reducing constraint violations, outperforming baseline methods including DDPG, MPC, and rule-based dispatch. These results highlight the potential of intelligent operators, guided by Safe RL, as a scalable and dependable solution for the future of autonomous power system operation under uncertainty.

This paper discusses the application of deep reinforcement learning (DRL) to the economic operation of power distribution networks, a complex system involving numerous flexible resources. Despite the improved control flexibility, traditional prediction-plus-optimization models struggle to adapt to rapidly shifting demands. Modern artificial intelligence (AI) methods, particularly DRL methods, promise faster decision-making but face challenges, including inefficient training and real-world application. This study introduces a reward evaluation system to assess the effectiveness of various strategies and proposes an enhanced algorithm based on the Model-based DRL approach. Incorporating a state transition model, the proposed algorithm augments data and enhances dynamic deduction, improving training efficiency. The effectiveness is demonstrated in various operational scenarios, showing notable enhancements in rationality and transfer generalization.

The development of edge computing technologies has brought about challenges in resource management. Traditional resource scheduling policies often prove insufficient due to the dynamic nature of cloud-edge collaboration. Therefore, adopting an edge cloud-native approach becomes necessary. This paper proposed a unified resource scheduling approach for joint optimization across multiple scenarios in the edge cloud-native environment. Our approach can schedule dynamically mixed-service groups across multiple scenarios by utilizing the adversarial learning of the environment and agents. Our approach can address the latency issues arising from imbalances among multiple scenarios. We conduct experiments by considering some factors such as device-number, communication-distance, CPU-cycle, and task-generation-speed. The results show that our approach can achieve a higher offloading rate and better average performance.

This paper addresses the challenges in large-scale joint market clearing for electricity systems, focusing on computational complexity, coordination between market layers, and convergence bottlenecks. The core market clearing problem, such as security-constrained unit commitment (SCUC), is NPhard and becomes increasingly intractable as system size grows. The inclusion of network constraints and detailed generator models further complicates the optimization. Coordination between regional and inter-regional markets, along with managing day-ahead and real-time market clears, adds another layer of complexity. To tackle these challenges, the paper explores the application of machine learning (ML) to accelerate market clearing. ML techniques, including supervised learning, neural networks, and reinforcement learning, are used to predict solutions or guide solvers, significantly improving solving times without compromising optimality. Hybrid approaches, combining ML with traditional solvers, have demonstrated substantial reductions in computation time. Additionally, advanced solver techniques such as decomposition, parallelization, and heuristic methods have been developed to enhance the efficiency of market clearing. The paper concludes that while no single method can solve all challenges, a combination of these strategies, including ML-enhanced optimization and hybrid solvers, offers a promising path forward for efficient market clearing in increasingly complex electricity markets.

This paper investigates the distributed optimization problem for a class of power grids. A distributed optimization method based on multi-agent deep reinforcement learning is developed by leveraging the multi-agent trust region policy optimization (MA-TRPO) algorithm. In the proposed algorithm, the power grid is divided into several sub-regions, which is managed by an agent. To achieve coordinated optimization among agents, deep neural networks are used to approximate the optimal policy of each agent. Unlike existing works, the proposed algorithm outperforms not only deep reinforcement learning methods for centralized control but also distributed online optimization methods. A case study is presented to verify the effectiveness and feasibility of the proposed approach.

With the increasing integration of distributed energy resources (DERs) in power systems, driven by the rising demand for renewable energy, challenges such as intermittent energy outputs, complex bidirectional power flows, and dynamic grid topology have emerged. These challenges require efficient parameter identification methods to ensure grid stability, voltage control and fault detection. Traditional approaches are often computationally intensive and struggle to adapt to the complexities of grids. This study proposes a parameter identification method by integrating Graph Convolutional Networks and Long Short-Term Memory networks. The method uses GCNs to capture the spatial topology of the grid and LSTMs to model the temporal dynamics of grid parameters. experiments were conducted using real-world data, where the model demonstrated performance in parameter estimation tasks under varying grid conditions. Our Key findings include: the proposed GCN-LSTM model outperformed traditional methods in terms of accuracy and robustness, achieving RMSE of 0.2196; the model captured both spatial and temporal characteristics, ensuring reliability under grid conditions; and its adaptability to different operating scenarios, including high renewable penetration, underscores its practical applicability. This research offers contribution by combining spatial and temporal modeling to address the pressing challenges of parameter identification in power systems.

With the advancement of new power systems, the integration of a high proportion of renewable energy sources has introduced significant uncertainty and complex dynamic coupling challenges to power grid dispatching. Traditional dispatching methods struggle to balance physical constraints with real-time decision-making requirements, while existing artificial intelligence approaches still fall short in model coordination and safe integration. To address these issues, this paper proposes a collaborative integration method of large and small models tailored for power grid dispatching tasks. A multimodal collaborative representation framework incorporating “spatiotemporal-load-topology” dimensions is constructed to achieve dynamic decoupling and fusion of dispatching features. A Pareto frontier-based model selection mechanism is designed to balance computational efficiency and decision accuracy. Furthermore, a reinforcement learning framework integrating differential-algebraic equation (DAE) constraints is developed to ensure the safety and feasibility of decision-making. Experiments on the IEEE 39-bus system demonstrates the performances of the proposed intelligent power grid dispatch solution that combines efficiency, safety, and interpretability.

As microgrids evolve towards integrating diverse energy sources and accommodating interactive competition among various stakeholders, conventional centralized optimization methods encounter difficulties in addressing the game among multiple entities. Therefore, this study proposes a strategy to optimize the operation of multi-energy microgrids (MEMG) with shared energy storage based on a Stackelberg game. First, the system architecture is introduced, and operation optimization models are established for MEMG operator, user aggregator, and shared energy storage service provider, respectively. Second, the game relationship between MEMG operator and user aggregators is revealed, and a Stackelberg game framework is established considering shared energy storage between MEMG operator and user aggregator. Finally, the strategies of each entity were optimized based on a combination of heuristic algorithms and quadratic programming. Experimental results show the effectiveness of the Stackelberg game model, with the model proposed in this paper increases the revenue of user aggregator by 20.23

As the randomness and intermittency brought by the increasing renewable energy sources and emerging loads intensify, traditional scheduling plans and intraday dispatch methods struggle to fully address the source and load fluctuations. Achieving coordinated optimization between economic efficiency and system robustness has become a key research direction. This paper employs the Value-at-Risk (VaR) metric to quantify the potential risks posed by system uncertainties. In the day-ahead scheduling stage, the dispatch objectives include minimizing the day-ahead cost of generating units, interconnection line costs, reserve costs and system risk costs, and reserve availability was verified to ensure the economic efficiency and robustness of day-ahead scheduling plan. In the intraday dispatch stage, a multi-agent real-time intraday dispatch model is proposed, adopting the MAPPO framework. The primary objective is to minimize the redispatch cost of generating units and the penalty costs for curtailing wind, solar power, and load shedding. Finally, the feasibility and effectiveness of the proposed model and methodology are evaluated through a devised IEEE 39-bus case.

Regional integrated energy system (RIES) cluster, i.e., multi-source integration and multi-region coordination, is an effective approach for increasing energy utilization efficiency. The hierarchical architecture and limited information sharing of RIES cluster make it difficult for traditional game theory to accurately describe their game behavior. Thus, a hierarchical game approach considering bounded rationality is proposed in this paper to balance the interests of optimizing RIES cluster under privacy protection. A Stackelberg game with the cluster operator (CO) as the leader and multiple RIES as followers is developed to simultaneously optimize leader benefit and RIES utilization efficiency. Concurrently, a slight altruistic function is introduced to simulate the game behavior of each RIES agent on whether to cooperate or not. By introducing an evolutionary game based on bounded rationality in the lower layer, the flaw of the assumption that participants are completely rational can be avoided. Specially, for autonomous optimal dispatching, each RIES is treated as a prosumer, flexibly switching its market participation role to achieve cluster coordination optimization. Case studies on a RIES cluster verify effectiveness of the proposed approach.