Review and Prospects of Artificial Intelligence Technology in Virtual Power Plants

Abstract

With the rapid development of global renewable energy, the virtual power plant (VPP), as an emerging power management model, has attracted increasing attention. Traditional manual management is difficult to effectively deal with because of the complexity and uncertainty of the VPP. The application of artificial intelligence (AI) technology provides new solutions for the VPP to cope with these problems. This review presents the research of AI technology in the VPP. Firstly, the basic concepts and theoretical framework of the VPP are presented. Then, the application of AI technology in VPP functional modules is discussed. Finally, the challenges of the VPP in coping with uncertainty, improving algorithmic interpretability and ensuring data security are pointed out, which provides theoretical support for subsequent research in the field of VPPs.

1. Introduction

Driven by the global carbon neutrality target, the development of clean energy has become the key to the reform of the global power system, and a significant number of distributed energy resources (DERs) are being integrated into power grids, particularly from rapidly growing renewable sources such as wind and solar energy. The global distributed energy resources market was valued at USD 45,651.79 million in 2024 and is projected to reach USD 47,039.61 million in 2025 [1]. With the transformation of the global energy structure and the rapid development of renewable energy, the VPP, as an emerging power management model, has gradually received extensive attention from academia and industry. The VPP forms a flexible and adjustable power supply entity through the integration and optimization of DERs. This helps to improve the stability and economy of the power system and effectively cope with the uncertainty and volatility brought by renewable energy [2].

AI is an emerging technology that simulates and extends human intelligent behaviors through computers. Its core lies in enabling computers to learn, reason, make decisions, perceive, and act autonomously [3]. Based on this foundation, techniques such as machine learning (ML) [4] and natural language processing (NLP) [5] have gradually emerged and been widely used in the AI field. In areas like pattern recognition, data analysis, and predictive decision-making, AI’s performance has gradually surpassed that of traditional methods, demonstrating a more powerful ability to handle complex data and uncertainty problems. As an intelligent scheduling system aggregating and optimizing distributed energy resources, the operation of the VPP involves complex multi-source data processing, real-time dynamic optimization scheduling, and uncertainty management. Traditional VPP scheduling and management methods rely on mathematical optimization algorithms (e.g., linear programming, nonlinear programming), heuristic algorithms, and statistical prediction models based on historical data. Traditional methods have limitations in dealing with resource uncertainty, real-time optimization computation, and the collaborative control of large-scale resources, so AI technology has become a key technological tool to improve the operational efficiency and intelligence level of the VPP, among other aspects. In addition to the above functions, it not only improves the accuracy of load forecasting, but also optimizes resource allocation, enhances the flexibility of scheduling decisions, and improves the adaptability of the VPP to the grid and its economic benefits. For example, deep reinforcement learning (DRL) is widely used in the optimal scheduling of the VPP, enabling real-time strategy adjustment in a dynamic environment and improving the resource utilization rate and economic benefits [6]. In addition, the application of machine learning techniques in load forecasting enables VPPs to predict power demand more accurately, thus realizing more refined dispatch management [7]. Despite the increasing number of studies on the application of AI technologies in VPPs, the existing literature mostly focuses on the optimization of specific technologies or single functions. There is a lack of systematic sorting out and analysis of the application of AI technologies from the perspective of the overall architecture of VPPs. Meanwhile, in the face of the rapid development of AI technologies and the continuous expansion of VPP application scenarios, a comprehensive summary of the research status, technical characteristics, and development trends in this field is helpful to provide references for subsequent research and practical applications.

Therefore, the purpose of this review is to systematically sort out the current status of the application of AI technology in the VPP. The structure of the article is divided into three main parts. Firstly, it introduces the basic concepts and theoretical framework of the VPP to establish the foundation of the analysis. Secondly, it explores in detail the specific application of AI technology in various functional areas of the VPP and its implementation outcomes. Finally, it summarizes the main challenges faced in the current application of the technology and identifies the key directions for future research. This structured analysis provides theoretical references and practical guidance for the intelligent development of the VPP field.

2. Basic Concepts and System Architecture of VPP

The VPP is an intelligent system that integrates various flexible resources, including DERs, controllable loads, energy storage systems, etc., through information and communication technology (ICT) for unified scheduling and management [8]. Meanwhile, it also has the ability to interact with the power market. It can adjust its power generation and energy storage strategies in real-time according to market demand, price signals, and grid loads, participate in the power market bidding, and provide auxiliary services such as peak shifting and frequency regulation [9].

The VPP consists of three parts: DERs, ICT, and a control center [10]. DERs, as the main energy resources in the VPP, usually consist of geographically dispersed photovoltaic (PV), wind power (WP), and energy storage systems, as well as controllable loads such as electric vehicles and air conditioners. These resource outputs are unstable and intermittent [11], and their fluctuations can be smoothed and the energy output optimized through the integration and management of the VPP [12]. ICT is used to monitor, collect, and transmit the operational data of distributed energy in real-time to provide accurate information support for resource scheduling and load forecasting. The control center serves as the “nerve center” of the VPP, responsible for performing data analysis, resource scheduling, market bidding, and other operations based on market and grid demand signals. As an independent entity in the power market, the control center participates in power trading and coordinates the operation of distributed resources. The composition of the VPP is shown in Figure 1.

Based on the overall composition of the VPP and its functional positioning in energy management, the VPP can establish a clear hierarchical system architecture, which consists of the infrastructure layer, the sensing and analytics layer, the decision–execution layer, and the market participation layer [13,14,15,16].

(1)

The Infrastructure Layer: As the physical and information infrastructure of the VPP, this layer integrates the functions of the “physical layer” and “communication layer” in traditional energy systems. Through data collection systems, communication networks, and security assurance mechanisms, it ensures the real-time monitoring of distributed resources as well as the secure transmission and storage of data.

(2)

The Sensing and Analytics Layer: This layer is in the middle of data and decision-making, corresponding to the “information layer” of the traditional energy system framework, and is responsible for transforming massive raw data into knowledge with decision-making value. Through multi-source data fusion, feature extraction, dynamic modeling, and other technical means, the accurate perception and characterization of distributed resource state are achieved, providing data support and analysis basis for upper-level decision-making.

(3)

The Decision–Execution Layer: As the “brain” and “central nerve” of the VPP, this layer corresponds to the “functional layer” of the traditional framework, and its core function is to generate optimal resource scheduling strategies and issue control instructions based on the data analysis results of the sensing layer. In this process, it needs to balance economic efficiency and system safety, realize dynamic optimization decision-making under multi-objectives, and ensure the accurate execution of control instructions.

(4)

The Market Participation Layer: As the interface between the VPP and the external economic environment, this layer corresponds to the “business layer” in the traditional framework and is responsible for formulating market trading strategies and participating in the diversified markets of electricity spot, auxiliary services, and capacity. Through the analysis of market signals and the optimization of trading strategies, this layer transforms the VPP’s technical capabilities into economic value, maximizes operational returns, and provides flexible grid support services.

Based on these four hierarchical architectures, the six core functional modules of the VPP are identified, which are resource analysis, power prediction, resource aggregation, optimal scheduling, market trading strategy decision, and data security. These six modules are interrelated and together support the efficient operation of the VPP. The VPP four-layer architecture and functional modules is shown in Figure 2.

By exploring the basic concepts and system architecture of the VPP, the core components and functional modules of the VPP are clarified, which lays the foundation for understanding how AI technology plays a role in each module. As the complexity and variability of the power system increases, traditional dispatch methods have struggled to meet the demands of modern power grids. Therefore, the next chapter will further explore the application of AI technology in VPP, specifically analyzing how AI can enhance the capabilities of VPP in terms of intelligent dispatch, load forecasting, resource optimization, and demand response. Through the integration of AI technology, the VPP can better cope with the volatility and uncertainty of renewable energy sources and achieve more accurate and efficient resource management in complex electricity market environments.

3. Application of AI Technology in the VPP

Among the six functional modules of the VPP, resource analysis includes resource data analysis and resource modeling, and power prediction includes load prediction, power generation prediction, and regulation capacity prediction. The VPP functional modules are shown in Figure 3. These functional modules are not isolated but closely linked and mutually supportive as a whole. The application of AI technology in each module exhibits progression and crossover. The application of AI technology in the six functional modules of VPP will be described in detail below.

3.1. Resource Analysis

Resource analysis is a fundamental precursor to VPP construction, aiming at parameter extraction, the characterization and dynamic modeling of dispersed DERs to solve the problems of resource observability and provide data support for subsequent aggregation, scheduling, and trading. Its input is raw resource data, and its output is resource classification labels, dynamic model parameters, and dispatchable potential assessment results. Resource data analysis is mainly the collection, processing, and analysis of data, which can include historical data, meteorological data, etc., to model distributed energy output characteristics, analyze distributed energy generation patterns through statistical modeling and data mining, and support power prediction and scheduling optimization; resource modeling is to convert resource data into mathematical models that can be used for power prediction, optimal scheduling, and market trading decisions. The main modeling approaches include data-driven modeling for complex, nonlinear energy resources; physical modeling for systems with clear physical laws; and hybrid modeling that combines physical mechanisms with data analysis to improve modeling accuracy.

AI technology in the module of VPP resource analysis and modeling mainly contains the following technical applications. Federated learning (FL) is a distributed machine learning method that performs model training without centralizing the data [17], ensuring that the data do not leave the device or the local server. Through FL, the system is able to model and analyze the data of multiple users, which can solve the data silo problem of distributed resources in VPP and realize collaborative modeling across topics. The generative adversarial network (GAN) generates synthetic samples consistent with the distribution of real data through the adversarial game between generators and discriminators [18]. The GAN can be applied to supplement resource operation data in scarcity scenarios and construct an extreme scenario library to test the robustness of the VPP. Big data analytics technology mines massive data to discover potential laws and complex associations between data, providing scientific basis for decision-making on practical issues [19]. It enables pattern recognition of historical operation data of different types of distributed energy resources in the VPP, helping the VPP predict resource output and demand more accurately by analyzing the underlying patterns. It can also help the VPP process data from the electricity market and identify market price trends, providing accurate support for market trading decisions. The application of AI technology in the resource analysis module of the VPP is shown in Figure 4.

Big data analytics methods are now being put to use in the area of VPP resource analysis. Reference [20] constructs a VPP platform based on a cloud-management-edge fusion architecture. In this platform, big data analysis helps the VPP centrally manage various flexible resources and achieve global optimal scheduling by real-time collecting, processing, and mining urban distributed energy data. The study on the VPP platform mentioned above shows that big data analysis technology can effectively improve the prediction accuracy and coordinated management capabilities of VPPs for flexible resources such as renewable energy and EVs, providing strong support for the intelligent scheduling and optimized operation of distributed energy resources. In addition, reference [21] provides an important practical case support for the realization of more accurate and efficient VPP management through the introduction and the use of big data analysis in the VPP in the operation scheduling part of the development route of Shanghai VPP. However, the current application of big data analysis technology is still mainly limited to historical data mining and static analysis. Facing the urgent need for real-time response in VPPs, existing technologies still have significant deficiencies in processing massive real-time data streams and supporting rapid-response decision-making, necessitating further research and breakthroughs.

In addition to the extensive use of big data analysis methods mentioned above, the application of FL has also been gradually emerging. Reference [22] proposes a novel VPP modeling method based on multi-mode and multi-task federated learning, which allows individual generator agents to process their own data locally without transmitting all data to a central server. This effectively reduces the computational burden on the central energy management system (EMS). The FL-QLMS algorithm proposed in Reference [23] enables FL to integrate data from different EVs in the VPP, allowing the model to learn more comprehensive features and enhancing its ability to process complex data, thereby enabling more accurate prediction of power consumption.The application of FL in VPP resource analysis not only addresses the challenges of data silos and privacy protection faced by traditional centralized learning methods but also provides new ideas for the collaborative optimization of distributed energy resources. With the increasing diversification of trading entities in the power market, this paradigm shift of “models to data” will become a critical technological backbone for future data sharing and collaborative decision-making in VPPs.

At present, big data analysis technologies have been widely applied in VPP resource analysis and modeling to enhance the fine-grained management and optimized scheduling capabilities for distributed resources. However, research on VPP resource analysis and modeling still predominantly relies on traditional machine learning and big data processing methods, with a relatively limited involvement of AI technologies such as FL and GAN. These technologies possess enormous potential in areas such as data privacy protection, distributed intelligent learning, data generation, and completion. In the future, they are likely to play a more critical role in large-scale data fusion, heterogeneous resource optimization, and uncertainty modeling for VPPs.

3.2. Power Forecasting

Power forecasting provides data support for VPP dispatch optimization, market participation, and demand response through accurate forecasting that mainly includes load forecasts, power generation prediction, and regulation capacity forecasts.

(1)

Load forecasting serves as the foundation for VPP operation. Accurate load forecasting helps the dispatch system prepare in advance, avoiding grid overload or resource waste.

(2)

The power output of renewable energy sources such as PV and wind power has high volatility and uncertainty, which brings challenges to grid scheduling. Accurate power generation prediction can effectively improve renewable energy consumption capacity and reduce dependence on traditional power generation resources.

(3)

Energy storage systems and adjustable loads in the VPP play an important role in regulating grid loads and providing auxiliary services. Accurate prediction of the charging and discharging capacity of energy storage systems and the responsiveness of adjustable loads enable the dispatch system to manage the grid more flexibly and improve overall system stability.

The general applications of AI technologies in VPP power forecasting mainly include long short-term memory (LSTM) networks, convolutional neural networks (CNN), and support vector machines (SVM). Among them, LSTM is a special type of recurrent neural network (RNN). In power forecasting, LSTM models can leverage historical load and meteorological data to predict future power demand, providing highly reliable forecasting results for load management [24]. The CNN, typically applied to image processing, is suitable for handling data containing multiple factors in forecasting to enhance the accuracy and adaptability of prediction models [25]. Especially in the context of VPP multi-dimensional power forecasting, where multiple time-series data dimensions are involved, the CNN can process these multi-dimensional data simultaneously, learning the relationships between different dimensions through convolutional layers to further optimize prediction accuracy. SVM, an MT algorithm for classification and regression, is well-suited for addressing nonlinear relationships [26]. It can handle multiple input dimensions in forecasting and make predictions based on such data. Notably, the SVM still demonstrates high prediction accuracy in small-sample data scenarios, enabling VPPs to predict grid load fluctuations and resource output in advance, thereby improving grid stability and operational efficiency. The application of AI technology in the power forecasting module of the VPP is shown in Figure 5.

Regarding the application of LSTM in VPP power forecasting, Reference [27] uses LSTM to obtain load, wind power, and PV output forecasts within a 24 h period for the VPP’s jurisdiction. Reference [28] focuses on power demand forecasting in VPP electricity trading platforms, proposing a novel method of applying one-hot encoding to input/output variables in deep learning-based power demand forecasting models and selecting the LSTM model for power demand prediction. Reference [29] employs a multimodal LSTM network that integrates sky images and historical irradiance data to achieve high-precision short-term solar irradiance forecasting, providing accurate irradiance predictions for PV power generation in VPPs. These studies fully demonstrate LSTM’s advantages in processing multi-dimensional time-series data in power systems, enabling high-precision forecasting for VPPs to support optimal energy resource allocation and enhance system resilient scheduling capabilities. However, LSTM models still have limitations such as long training times, high computational resource requirements for large datasets, as well as difficulty in capturing power fluctuations under extreme weather conditions, necessitating optimization through integration with other technical approaches in practical applications.

In light of the performance bottlenecks exhibited by LSTM models in applications, researchers have continuously explored innovative model architectures and hybrid algorithm integrations to enhance VPP forecasting and scheduling performance. Reference [30] developed an integrated gated recurrent unit and bidirectional long short-term memory network (GRU-BiLSTM) for the day-ahead time-series forecasting of multi-energy demands and renewable resources in multi-energy VPPs (MEVPPs). Reference [31] combined LSTM with a multi-objective particle swarm optimization (MOPSO) algorithm for the bi-objective scheduling of multi-energy VPPs, where LSTM was applied to improve wind power forecasting accuracy. Reference [32] adopted an improved K-means clustering algorithm integrated with a bidirectional long short-term memory network (BiLSTM) for VPP load forecasting. Reference [33] employed a fuzzy decision support system (FDSS) combined with BiLSTM for VPP power forecasting. Reference [34] utilized a CNN-LSTM hybrid model for short-term electrical and thermal load forecasting in VPPs, where the CNN precisely extracted data features and LSTM captured temporal dependencies, significantly improving forecasting accuracy. Reference [35] proposed a hybrid model combining LSTM and a transformer, which first processes input data through LSTM to extract temporal features and then refines them using the transformer, thereby remarkably enhancing forecasting accuracy to support economic dispatch and optimal operation of VPPs. Reference [36] designed a GCN-BiLSTM combined model to predict the power of aggregation nodes (ANs) and proposed a power shortage measurement method considering forecasting uncertainties, enabling VPPs to respond to AGC commands from the main grid and effectively improve grid frequency stability. Reference [37] introduced a novel decision support system (DSS) by constructing a DSS-BiLSTM model, demonstrating higher accuracy and stability in forecasting VPP power generation. Reference [38] combined LSTM, BiLSTM, and DSS-BiLSTM models to predict VPP power generation. The above studies fully demonstrate that the use of LSTM prediction algorithms provides strong data support for real-time scheduling and resource optimization in VPPs, adapting to the fine-grained requirements of practical VPP applications. However, while improving prediction accuracy, the modified model structures and multi-algorithm integrations also increase the difficulty of model parameter tuning and training, which makes them prone to causing overfitting issues. This is unfavorable for real-time applications and may affect the overall scheduling flexibility and economic benefits of VPPs.

In parallel with the extensive application of LSTM and its derivative models in VPP forecasting, researchers have also actively explored other algorithmic models to address the specific requirements of different forecasting scenarios. Reference [39] proposed a short-term photovoltaic power forecasting method based on improved grey relational analysis (IGRA), the efficient channel attention network (ECANet), and the temporal convolutional network (TCN) in an edge-computing-based VPP architecture, where TCN overcomes the limitations of traditional recurrent neural networks in time-series forecasting. Reference [40] used least squares support vector machines (LSSVM) to predict wind speed and solar radiation based on meteorological data, thereby deriving wind and solar power generation to address risk-based scheduling challenges in VPPs. The application of these two types of algorithms in VPP power forecasting fully utilizes historical data to predict load curves, wind speed, solar radiation, etc., effectively capturing the inherent patterns and temporal characteristics in the data, thereby providing decision support for risk-based scheduling and optimized operation of VPPs. However, as both algorithms rely on historical data patterns, they may fail to quickly adapt to drastic changes in data distribution during extreme weather conditions or unexpected events, thus affecting prediction accuracy.

In summary, significant progress has been made in VPP power forecasting research in recent years, particularly in load forecasting, renewable energy prediction, and demand response optimization. Deep learning models based on LSTM and its variants, hybrid architectures combining advanced algorithms such as the CNN and transformer, and traditional machine learning methods like support vector machines have all demonstrated good prediction performance in their respective applicable scenarios, providing a solid data foundation for the precise scheduling and resource optimization of VPPs. However, these forecasting algorithms still have limitations in historical data dependence, adaptability to complex scenarios, parameter tuning difficulties, and responsiveness to extreme events. Future research should focus on algorithmic fusion innovation, the optimization of multi-source data collaborative processing mechanisms, and the improvement of prediction uncertainty quantification methods to further enhance the operational stability and scheduling efficiency of VPP systems. As one of the core functional modules of VPPs, the output of the power forecasting module serves as an important input to the resource aggregation module. Through accurate forecast data, it can provide a data foundation for subsequent resource aggregation and optimization, thereby achieving the efficient collaboration of resources.

3.3. Resource Aggregation

The resource aggregation module involves the integration and collaborative management of multiple distributed energy sources and flexible loads to achieve the optimal scheduling, flexible regulation, and efficient utilization of power resources. Resource aggregation is a way for the VPP to optimize its overall resource utilization by integrating different types of distributed energy sources and adjustable loads, treating them as a VPP; different types of resources are synergistically optimized through an intelligent dispatch system, which adjusts the output of resources and the load response according to the market demand, price signals, and so on during the fluctuation of the grid demand. Through collaborative optimization, the VPP can improve the response speed of power supply and demand balance, optimize resource allocation, and to a certain extent, reduce dispatching costs and the impact of electricity price fluctuations.

The general applications of AI techniques in VPP resource aggregation and cooperative optimization are as follows. Clustering algorithms (e.g., K-means, hierarchical clustering) are used in resource aggregation to group resources with similar characteristics to achieve more targeted management [41]. In the VPP, K-means can classify resources such as WP, PV, energy storage, and EVs, to ensure that the optimal scheduling strategy can be matched when resources are aggregated, and hierarchical clustering can divide resources into multiple management levels, which facilitates the hierarchical optimization of scheduling; multi-agent systems (MASs) treat different resources as independent intelligences and realize automatic aggregation based on the synergistic strategy, which allows different types of resources to autonomously coordinate their output and demand when resources are aggregated, making the aggregation of resources more flexible and self-sustaining. The aggregation of resources is more flexible and adaptive [42]. In the use of the VPP function, the MAS allows different resource intelligences to make decisions independently and optimize the overall operation of VPP through collaboration; genetic algorithms and particle swarm optimization can help the system to find the optimal combination of resources in the VPP according to the current demand and resource characteristics. The application of AI technology in the resource aggregation module of the VPP is shown in Figure 6.

Reference [43] indicates that the technical methods for resource aggregation encompass physical modeling and data modeling, with the data modeling component mainly employing statistical methods and artificial intelligence algorithms to characterize the typical features of distributed resources. The K-means clustering algorithm, a widely used artificial intelligence technique, is exemplified in multiple studies: Reference [44] presents a resource aggregation method for PV output forecasting in VPP peak regulation, whose core is to use K-means clustering to categorize historical PV data by weather type, extract data features of similar days, and aggregate distributed PV resources; Reference [45] generates random scenarios via the Monte Carlo method, processes them through dimensionality-reduced K-means clustering, and proposes a new rolling horizon mixed-integer linear programming model for the real-time operation management of VPPs; and Reference [46] applies K-means clustering analysis with Chebyshev distance in the comprehensive correlation and clustering analysis of long-term power quality data in VPPs to identify and eliminate outliers, thereby preventing their influence on correlation analysis results. These studies provide important support for the efficient aggregation and optimal scheduling of distributed resources through their data-driven classification and feature extraction capabilities, while also laying a foundation for integrating more advanced algorithms to address complex energy system problems.

In terms of innovations in the K-means clustering algorithm, Reference [47] employs a hybrid algorithm combining agglomerative hierarchical clustering and K-means clustering to cluster the load output curves of all aggregated loads, identify load curve clusters of the same category and their cluster centers, analyze clustering results, establish a matching evaluation system, and select appropriate load combinations for VPP aggregation through comprehensive evaluation. For long-term energy consumption forecasting in commercial VPPs, Reference [48] uses unsupervised K-means learning to cluster the seasonal patterns of daily energy consumption, combines K-means clustering with linear regression, divides forecast results into peak and non-peak periods, and considers the additional costs of power distribution companies during high-demand periods, demonstrating high accuracy in aggregated load environments. Reference [49] integrates agglomerative hierarchical clustering to determine a reasonable number of clusters, then applies the K-means clustering algorithm to group data, and proposes a TCN-BiGRU-attention prediction model. By combining the improved clustering analysis method with a deep learning model, this approach enhances prediction accuracy and robustness.

Notably, multiple studies have explored the combined application of Latin hypercube sampling (LHS) and K-means clustering: Reference [50] generated a set of representative scenarios using this combination to develop an economic model for VPPs, Reference [51] effectively addressed the uncertainties in wind and solar power generation through this technique to reduce the number of scenarios while retaining critical information and provide a basis for efficient solving of VPP optimal scheduling models, and Reference [52] applied this method to dynamic VPP formation discrimination and transaction matching schemes to evaluate market risks and assist in game-theoretic decision-making. Overall, the K-means clustering algorithm is widely used in the field of VPP resource aggregation due to its advantages of simple calculation and high efficiency, but it also has limitations such as sensitivity to initial cluster centers and difficulty in handling non-convex shape data, and future research may need to explore hybrid clustering methods integrating more advanced technologies such as deep learning and graph neural networks to enhance the generalization ability and dynamic adaptability of aggregation models.

In addition to the K-means algorithm, researchers have actively explored the application and innovation of other diverse clustering methods in VPP resource aggregation. For example, Reference [53] employs a two-stage clustering analysis algorithm combining hierarchical clustering and fuzzy C-means to analyze adjusted daily load curves, providing a basis for leveraging price mechanisms to deeply exploit urban public building complexes as VPP resources. Reference [54] proposes a distributed dynamic clustering algorithm to achieve the power demand-based aggregation of heterogeneous VPP resources. These diversified clustering methods offer more targeted technical pathways for addressing the complex and variable load characteristics and heterogeneous resources in VPPs.

With the increasing complexity of VPP resource aggregation, research has gradually evolved from static clustering methods to dynamic and adaptive intelligent systems, among which MASs have garnered extensive attention due to their distributed collaboration and self-organization capabilities. Reference [55] proposes a VPP decision-making method that uses fuzzy logic and a novel “unsafety” index based on human psychology, modeling the VPP as a MAS to minimize carbon emissions and energy costs through an aggregation structure similar to that of energy or carbon markets. Such dynamic and adaptive multi-agent-based methods break through the limitations of traditional clustering algorithms, offering more flexible and robust technical pathways for the collaborative optimization and real-time decision-making of large-scale heterogeneous distributed energy resources.

In summary, as an efficient data processing method, clustering algorithms have been widely applied in the VPP resource aggregation module. From basic K-means to composite technologies integrating multiple algorithms and emerging MASs, they all provide powerful decision support for the efficient operation of VPPs. It is worth noting that although research on multi-agent systems in VPP resource aggregation and collaborative optimization is relatively limited, their adaptability and self-organization capabilities in handling heterogeneous resources, dynamic market environments, and demand fluctuations make them an important direction for future research. Through reasonable resource classification and integration, resource aggregation technologies lay a solid foundation for optimal scheduling, significantly reducing scheduling complexity and improving overall system efficiency. The next section will further explore how artificial intelligence technologies play a key role in the VPP optimal scheduling module.

3.4. Optimal Scheduling

Optimal scheduling in the VPP function module is mainly responsible for coordinating distributed resources on different time scales to ensure stable operation of the power system, maximize economic returns, and improve the capacity of renewable energy consumption. The main roles of the optimization scheduling module include supply and demand balance, resource coordination, market transaction optimization, and grid security and stability assurance.

The applications of AI technology in VPP optimization scheduling include reinforcement learning (RL), MASs, and DRL. Among them, RL is an adaptive learning technology that can optimize resource allocation through a trial-and-error mechanism in uncertain environments. The intelligent agent within RL obtains feedback through continuous interaction, continuously refines scheduling strategies, and, ultimately, achieves optimal resource scheduling. In VPPs, RL can be used for real-time optimization scheduling to ensure power supply–demand balance and handle multi-objective optimization problems, dynamically adjusting scheduling schemes to ensure optimal decision-making. MAS technology allows multiple independent agents to collaborate in the scheduling process [42]. In VPPs, each agent can represent different distributed resources or scheduling units, achieving global resource optimization through collaborative interaction and policy sharing. The MAS is particularly suitable for scheduling problems in distributed systems, enabling the flexible allocation of different resources across time and space. DRL is a machine learning method that combines deep learning (DL) and RL. It maps states and actions in RL through deep neural networks (DNN) to learn optimal decision-making strategies in complex environments [6]. In VPPs, DRL can analyze electricity market price dynamics and optimize market bidding strategies to derive optimal decision-making strategies. The application of AI technology in the optimal scheduling module of the VPP is shown in Figure 7.

In the complex and dynamic environment of VPPs, traditional scheduling methods struggle to address the highly uncertain distributed resources and multi-market environments. As a method capable of learning optimal decisions through interaction with the environment, RL offers new solutions for VPP scheduling problems. By continuously testing different scheduling strategies and learning from feedback, RL can adapt to the uncertainties and complexities of VPP operating environments. Reference [56] employs a two-layer game model based on an improved reinforcement learning algorithm (IRLA) to optimize VPP scheduling and internal electricity price decisions, which positively improves scheduling efficiency, achieves multi-objective balance, and enhances system stability. Reference [57] designs an energy storage system controller framework for VPPs, using the deep Q-network (DQN) method in RL to train a neural network and maximize profits. Overall, these RL-based methods provide flexibility and adaptability to VPP optimal scheduling systems but require further algorithmic improvements and optimizations when facing high-dimensional state spaces and more complex decision-making environments.

Although traditional RL algorithms have demonstrated potential in VPP scheduling, their efficiency and convergence are often limited when facing high-dimensional state spaces and complex decision-making environments. By integrating the powerful representation capabilities of DL with RL’s decision optimization mechanism, DRL can more effectively handle large-scale, high-dimensional scheduling problems in VPPs and achieve end-to-end optimization decisions. For example, Reference [58] constructs a DRL-based Stackelberg game model for VPPs with EV charging stations, setting separate agents for EVs and VPPs, alternately training the agents’ network parameters via DRL, and calculating strategies and solutions at game equilibrium to optimize DER scheduling and bidding strategies in power markets. Reference [59] proposes an improved two-stage DRL method to provide frequency regulation services for VPPs and achieve higher economy and accuracy in practical operations. Reference [60] develops a day-ahead and real-time coordinated scheduling optimization model for VPPs, using the deep Q-network (DQN) in DRL to address the complexity and nonlinearity of VPP models and achieve maximum benefits and minimum carbon emissions. Reference [61] presents a DRL-based economic scheduling optimization method for VPPs to realize real-time online optimization decisions for VPP economic scheduling within the Internet of Energy (IoE) framework. Reference [62] addresses the challenge of centralized scheduling and optimization of distributed power sources by proposing a DRL-based VPP optimization scheduling method and constructing an optimal scheduling model for each generating unit in the VPP. Reference [63] divides VPP energy management into three stages—internal price setting, microgrid (MG) scheduling, and energy storage system (ESS) management—with each stage using DRL for optimization decisions, where DRL agents learn optimal scheduling strategies through continuous interaction with the environment. Reference [64] proposes a voltage control model for VPP collaborative optimization and transactive energy (TE), solved by an enhanced multi-agent deep reinforcement learning (MADRL) method to optimize distribution network (DN) voltage control and VPP operating profits.

In the field of VPP scheduling, DRL methods documented in the literature have demonstrated significant advantages. Through continuous environmental interaction and policy optimization, they effectively enhance economic benefits, resource utilization, and the collaborative scheduling capability of multi-energy systems. However, these methods still face numerous challenges, including a high dependence on computational resources during training, the time cost of sample collection, and potential shortcomings in generalization ability and robustness in complex and variable VPP environments. Additionally, the black-box nature of DRL models leads to insufficient transparency in decision-making processes, which may affect the trust of power grid operators in their scheduling strategies. Future research should focus on improving model interpretability, reducing training costs, enhancing algorithm robustness, and developing more efficient transfer learning methods to promote the large-scale deployment and implementation of DRL technologies in practical VPP applications.

In the optimal scheduling module of VPPs, the application of MASs has been gradually increasing. For example, Reference [65] designs a hierarchical DRL algorithm, Hierarchical-TD3, where the upper layer solves the economic scheduling model of VPPs based on the single-agent TD3 algorithm. The lower layer, following the upper-layer scheduling instructions, uses the multi-agent MATD3 algorithm to achieve the real-time response of distributed energy within VPPs, ensuring privacy protection and coordinated scheduling between VPPs and DE to reduce operating costs. Reference [66] proposes a multi-agent optimization framework to capture the distributed characteristics of VPPs and facilitate their optimal economic scheduling. Reference [67] develops a model-aided multi-agent reinforcement learning (MARL) method, formulating the VPP scheduling problem as a decentralized, partially observable Markov decision process. Reference [68] devises a distributed multi-agent algorithm based on graph theory and the Lagrangian multiplier method to promote communication between units and minimize the total system operating cost. Reference [69] addresses privacy issues through DTDE training strategies, introducing adjacent agent communication and low-dimensional fingerprint information processing for non-stationary environments to improve algorithm convergence and stability, thereby achieving the optimal coordination of multiple VPPs. Reference [70] treats each multi-energy building in the VPP as an independent agent, where the multi-agent learning algorithm leverages the attention mechanism of the transformer to continuously interact with the environment and learn optimal energy scheduling strategies, maximizing the overall benefits of the VPP. These studies demonstrate that MAS methods can effectively decompose VPP scheduling problems, reflect the distributed characteristics of the system, and achieve collaborative decision-making and resource optimization while protecting user privacy. However, current research still lacks in-depth analysis of system stability and algorithm convergence, suffers from imperfect agent coordination mechanisms, and lacks reasonable incentive mechanisms in complex market environments, limiting the application effectiveness of models in practical scenarios. Future research needs to strengthen the theoretical foundation, improve communication protocols and benefit distribution mechanisms, and enhance the reliability and adaptability of MAS in actual VPP operations.

In summary, research on optimal scheduling for VPPs has achieved three breakthroughs through AI technologies such as RL, DRL, and MAS: first, the shift from single static optimization to dynamic adaptive decision-making, which enhances the system’s ability to cope with uncertainties; second, the transition from centralized control to distributed collaboration, effectively balancing global benefits and local autonomy; and third, the expansion from single economic objectives to multi-objective optimization, taking into account economic benefits, environmental benefits, and system security. Although these methods have significantly improved the flexibility and benefits of VPPs, they still face technical challenges such as insufficient model interpretability, unstable algorithm convergence, and limited generalization ability. Future research should focus on lightweight algorithm design, enhanced robustness, and theoretical foundation improvement to promote the large-scale application of these advanced methods in actual power systems. It is worth noting that the ultimate goal of optimal scheduling is to improve the economic benefits and competitiveness of VPPs in the power market, which naturally leads to the core content of the next section: the application of AI technologies in the functional module of VPP market trading strategy decisions.

3.5. Market Trading Strategy Decisions

Market trading strategy decision-making is a core functional module for VPPs to develop optimal power trading strategies, bidding schemes, and resource allocation through intelligent decision-making algorithms to maximize revenues, minimize costs, and ensure grid stability under various trading environments such as the power market, auxiliary service market, and carbon trading market. Its main objectives under VPP participation in market trading are maximizing market revenue, reducing operating costs, improving market competitiveness, and supporting grid stability. This module is an important part of the VPP’s value realization and is closely linked to the aforementioned optimized scheduling module, which transforms the VPP’s technical advantages into economic benefits through market-based means.

The general applications of AI technology in VPP market trading strategy decision-making are as follows:

(1)

lDL [71] is a branch of machine learning, which is mainly based on the multi-layer architecture of neural networks, and deals with and learns complex data patterns through multi-level feature abstraction. It can be used for market tariff prediction, through learning historical data, weather factors, supply and demand changes, etc., to predict future market prices, to provide accurate data support for the VPP trading strategy, mainly using LSTM, the CNN, the transformer, and other models to capture the time series trend.

(2)

RL can be used in market trading strategy decision-making for the optimization of the dynamic tariff response strategy, and the cross-market bidding strategy RL can be used for dynamic tariff response strategy optimization and cross-market bidding strategy formulation in market trading strategy decision-making, mainly using Markov decision process modeling (MDP) and the deep deterministic strategy gradient (DDPG).

In the realm of electricity price forecasting, Reference [72] employs LSTM neural networks to forecast uncertain factors such as electricity prices, wind speeds, and solar irradiance, and then uses scenario generation and reduction methods to handle prediction uncertainties, achieving maximum revenue for VPPs in day-ahead markets, futures markets, and bilateral contracts. Reference [73] determines the optimal market clearing price through the Fischer Black model and uses a linear regression model to predict hourly electricity prices for the next day, effectively maximizing VPP profits in energy markets. Reference [74] adopts a price forecasting method based on $\epsilon$-support vector regression ($\epsilon$-SVR) to address inaccuracies in electricity price forecasting for VPP market trading strategy decision-making. Overall, these electricity price forecasting strategies provide VPPs with accurate predictions of market price fluctuations, effectively supporting bid decision optimization and cross-market resource allocation, enabling VPPs to maximize revenues and minimize risks in complex electricity market environments.

In market trading strategy decision-making, electricity price forecasting is just one link, with the core being the formulation of electricity price strategies. For example, Reference [75] uses a DDPG algorithm with prioritized experience replay (PER) to develop real-time electricity price plans, continuously optimizing VPP decision-making in market participation and reducing user electricity costs. Reference [76] addresses the bi-level mathematical programming problem (MPEC) between VPPs and the main grid based on the DDPG algorithm, modeling market trading strategies using Markov decision processes (MDP) to provide efficient, robust, and economical optimization solutions for VPP market trading strategy decisions. Reference [69] employs an independent twin delayed deep deterministic policy gradient (ITD3) algorithm to achieve optimized coordination of multiple microgrid VPPs through multi-agent collaboration and optimize internal market prices. Reference [77] uses a combined experience pool replay twin delayed deep deterministic policy gradient (CEPR-TD3) algorithm to optimize incentive coefficients, effectively improving the participation rate of DER participants. Reference [78] applies a multi-agent twin delayed deep deterministic policy gradient (MATD3) algorithm to solve the problem of price-setting VPPs participating in day-ahead (DA) markets. Reference [79] proposes an LSTM-encoded DDPG (LDDPG) algorithm to optimize VPP trading strategies and resource allocation in power markets. Reference [80] constructs a DRL-based Stackelberg game model for VPPs with EV charging stations, using the twin delayed deep deterministic policy gradient (TD3) algorithm to calculate strategies and solutions at game equilibrium through alternating training of agent network parameters. Reference [81] explores the use of the DDPG algorithm to optimize decision-making for VPPs in complex urban environments.

Studies show that such algorithms can continuously optimize the decision-making process of VPPs in market participation, achieving multiple objectives such as reduced electricity costs, improved participation rates of DER, optimized resource allocation, and coordinated internal market prices. However, these methods also have limitations, including high computational complexity, poor interpretability, reliance on large volumes of data, and difficulties in parameter tuning. Nevertheless, they can effectively address market uncertainties, provide efficient, robust, and economical optimization solutions, offer strong support for intelligent decision-making of VPPs in complex power market environments, and represent an important development direction in the research of power market trading strategies.

In addition to DDPG and its variants, other RL methods have also demonstrated promising results in optimizing VPP market trading strategies. Reference [82] uses DL to construct a VPP market competition model and proposes a trading strategy optimization method based on an Actor–Critic DRL framework. This approach approximates the Nash equilibrium to solve for the optimal bidding strategies of multiple VPPs in market games and addresses uncertainties in power markets through nonlinear approximation. Reference [83] constructs a community VPP (cVPP) decision model, transforming the decision problem into a partially observable Markov game (POMG) model, and uses a transformer-based multi-agent DRL method to enable cVPPs to autonomously generate optimal bidding and scheduling strategies. Reference [84] proposes a model-free incentive-based demand response (DR) scheme combining RL and deep neural networks, constructing a VPP model that maximizes profits while considering factors such as renewable energy generation deviations, load demand, and DR incentive costs. These diverse RL methods provide broader technical pathways for VPP market trading strategy research and collectively advance the development of intelligent decision-making in power markets.

In summary, the application of AI technologies in the decision-making module of VPP market trading strategies has demonstrated promising prospects. Although these methods perform well in complex power market environments, they still face challenges such as high computational complexity, poor model interpretability, and data dependence. Future research should focus on addressing issues such as algorithm training efficiency, model stability, and market environment adaptability, while strengthening algorithm interpretability and lightweight design. In the overall operation of VPPs, data security and privacy protection have become particularly important throughout the processes of resource analysis, power forecasting, resource aggregation, optimal scheduling, and market trading strategy decision-making. This naturally leads to the next core functional module—data security.

3.6. Data Security

Data security is a key assurance component for VPP operation. The VPP integrates a large amount of distributed energy sources, which requires data collection, real-time monitoring, and intelligent scheduling through ICT. However, due to the distributed storage of data, complex remote communication, and transactions involving multiple parties, VPPs face many security risks, such as data leakage, cyber-attacks, transaction fraud, and privacy protection. As the basic support of the first five functional modules, data security permeates the entire VPP operation process and plays a decisive role in ensuring the safe and stable operation of the system.

The main applications of AI technology for data security in VPPs include FL, blockchain technology, and DL and pattern recognition technologies. FL is a decentralized ML method [85], which allows multiple VPP nodes to share the model parameters without the need to directly exchange raw data, which helps in data analysis and model training while ensuring data privacy. Blockchain technology [86] can effectively ensure data integrity and traceability through distributed ledger and tamper-proof characteristics, and in the VPP and smart grid, blockchain can help manage the flow and use of distributed energy data, ensure the security of the data when it is transmitted among multiple parties, and prevent data leakage and illegal access. DL and pattern recognition technologies in cybersecurity can detect potential cyber threats and anomalous activities in real-time, with AI-based systems identifying unusual patterns in data transmission [87]. The application of AI technology in the data security module of the VPP is shown in Figure 8.

In relevant review articles on VPPs, the importance of blockchain and federated learning technologies in VPP data security is emphasized [88,89]. AI technologies ensure the security and privacy of VPPs during data collection, transmission, and storage through encryption algorithms, blockchain, and FL. For example, Reference [85] employs a horizontal federated learning framework, where each VPP uses a deep transformer Q-network for local training to protect sensitive data privacy. Reference [90] proposes a privacy-preserving intelligent energy management scheme based on consumer electronic chips and FL, aiming to reduce reliance on sensitive information through model sharing between the control center and chips. The privacy-preserving horizontal federated reinforcement learning (PHFRL) method proposed in Reference [91] enhances data security by introducing local differential privacy and CKKS homomorphic encryption technologies. In summary, FL addresses the issues of data silos and privacy protection in VPP systems through a distributed training model of “data remaining in local domains while models collaborate.” Especially when combined with technologies such as differential privacy and homomorphic encryption, it enables efficient collaborative analysis and optimal decision-making while safeguarding the security of sensitive data from distributed energy resources, providing critical technical support for the secure operation of VPPs.

Data security measures not only prevent information leakage and cyberattacks but also ensure the isolation and independence of user data in multi-agent collaborative scheduling environments, enhancing the overall security of VPPs and user trust, while the application of blockchain technology can effectively improve the data security of VPPs. Reference [86] designs a virtual power plant blockchain network (VPPBN) and its collaborative mechanism (BNCM) to ensure the privacy and security of power transaction data in multi-node distributed systems, significantly reducing transaction and collaboration delays and improving the overall operational efficiency and reliability of VPPs. Reference [92] uses the blockchain-based VPPBN to address DER coordination issues in VPPs and the security and efficiency of information transmission. Reference [93] proposes a peer-to-peer (P2P) energy trading scheme for VPPs using smart contracts on the Ethereum blockchain platform, which solves cost and security problems through an auction mechanism operated by smart contracts based on a public blockchain network. Reference [94] introduces consortium blockchain technology to achieve transaction matching among multiple VPPs, with the transaction process including VPPs determining market behaviors and electricity prices, uploading transaction information, conducting direct transactions and signing smart contracts, while a credit scoring mechanism is introduced to constrain node behaviors. Reference [95] details the algorithms and processes for VPP initialization, prosumer joining, energy settlement and delivery tracking, and financial settlement, fundamentally protecting VPP data security through blockchain technology. Reference [96] indicates that the consensus mechanism of blockchain ensures the transparency of P2P transaction information among prosumers within VPPs, enabling flexible transactions and automatic settlement between VPPs and prosumers through smart contracts. Reference [97] proposes a blockchain-based decentralized P2P multi-layer energy trading framework that uses blockchain technology to record energy transactions and ensure data transparency and immutability. Reference [98] suggests using private blockchains to store production information required for VPP scheduling calculations and consortium blockchains to store private blockchain summaries and transaction information, ensuring data security and traceability. Reference [99] combines blockchain with the alternating direction method of multipliers (ADMM) to ensure that VPP optimization data are processed locally and only necessary variable updates are transmitted, protecting VPP commercial data privacy during distributed optimization.

These studies fully leverage the advantages of blockchain technology in information security, transparency, and decentralization to ensure the tamper-proofing and credibility of VPP data, providing a solid foundation for the secure operation of VPPs. However, they insufficiently address the deep integration of blockchain technology with traditional VPP scheduling mechanisms and cross-domain technical integration, leaving the overall solutions facing challenges in technical integration and efficiency balance during practical deployment and large-scale application.

In summary, data security is of utmost importance in the operation of VPPs, and AI technologies provide strong security guarantees for VPPs through various means. By integrating these technologies, VPPs can effectively manage distributed energy data, enhance system security and reliability, and strengthen user trust. In the future, with the further development of AI and blockchain technologies, the data security of VPPs will continue to be optimized, providing a more solid foundation for the intelligent operation of power markets.

3.7. Comprehensive Comparison of Functional Modules

In Section 3.1, Section 3.2, Section 3.3, Section 3.4, Section 3.5 and Section 3.6 above, the current status and main approaches of AI technology application in each of the six functional modules of VPPs are explored. In order to summarize them as a whole and further sort out the core role of each module, the AI technologies used, and the existing deficiencies, the following table systematically summarizes the application of each module, as summarized in

Table 1. The application of AI technology in the six functional modules of VPP.

VPP Functional Modules	Use of AI Technology	Corresponds	Existing Deficiencies
resource analysis	FL; GAN; Big Data analytics technology	Accurately identify and classify the characteristics of various types of energy resources in a virtual power plant to provide data support for subsequent aggregation and optimization.	Imperfect real-time analysis and model fusion for cross-regional heterogeneous data; challenges in privacy protection and distributed data management.
power forecasting	LSTM; CNN; SVM	Improve the accuracy of load fluctuation and resource output forecasting for the grid and reduce dispatch risk.	Insufficient algorithm efficiency and robustness under extreme weather conditions; lack of systematic validation in large-scale power grids.
data security	FL; blockchain technology; DL and pattern recognition technologies	Improve data tampering and traceability and reduce the risk of centralized data breaches.	Insufficient characterization of heterogeneous resource differences; scalability in large-scale distributed scenarios requires further research.
resource aggregation	clustering algorithms (e.g., K-means, hierarchical clustering); MAS; genetic algorithms and particle swarm optimization	Improve resource utilization, reduce scheduling complexity, improve the response speed of power supply and demand balance, and optimize resource allocation.	Algorithm convergence speed and robustness remain challenges; lack of safety and stability assessment under extreme conditions.
optimal scheduling	RL; MAS; DRL	Achieve coordinated scheduling of multiple resources to ensure stable operation of the power system and maximize economic benefits.	Facing complex environments with changing market rules and information asymmetry; need for improved mechanism design and practical validation.
market trading strategy decisions	DL; RL	Improve the profitability and competitiveness of virtual power plants in many types of markets.	Insufficient integration between blockchain and traditional scheduling mechanisms; technical integration and efficiency balance still face challenges.

.

4. Future Challenges and Outlook

As the role of VPPs in modern energy management continues to expand, how to more effectively utilize AI technologies to achieve optimal resource allocation, privacy protection, and intelligent scheduling has become a key topic. The following section explores the future challenges and research outlook of the VPP.

4.1. Challenges Faced

(1)

Current VPP research mainly focuses on the optimization of a single functional module and lacks systematic research on the synergistic operation between modules. Resource aggregation models and optimal scheduling algorithms are often designed independently of each other, resulting in aggregated resource characteristics that cannot be fully utilized by scheduling algorithms. The adaptability and migration ability of AI technology under different power market environments need to be improved. With the expansion of VPP scale, flexible access to massive heterogeneous resources and collaborative interaction among VPPs require more scalable AI solutions. Various AI models have inherent tradeoffs in terms of accuracy, computational efficiency, interpretability, and data requirements, and a more systematic research approach is needed to select the right combination of AI technologies based on the characteristics of different functional modules of the VPP.

(2)

VPPs face multiple sources of uncertainty, including fluctuations in renewable energy output, changes in user behavior, and changes in market conditions. Existing AI algorithms are not robust enough in highly uncertain environments. Especially in multi-timescale operation, short-term optimization may conflict with long-term planning, and the decision-making mechanism for balancing short-term gains and long-term sustainability is in urgent need of a breakthrough. As a support for the safe operation of the power system, the AI system of the VPP must meet the requirements for high reliability of the power system, including the ability to provide security under fault conditions.

(3)

The “black box” nature of AI technology raises issues of transparency and interpretability, and the VPP involves multiple stakeholders with high demands for understandability and fairness in AI decision-making. It is critical to develop hybrid models that can both leverage AI optimization capabilities and provide explanations for decisions. AI-driven VPPs will change the competitive landscape and value distribution mechanisms of traditional electricity markets, and these socio-economic impacts have not been fully studied. Inadequate regulatory policies and legal frameworks also constrain the actual deployment and commercialization of VPPs.

4.2. Future Prospects

(1)

In the future, the VPP can build a multi-layered hybrid intelligence architecture covering sensing, analyzing, decision-making, and execution. It will fuse physical models with data-driven methods to improve modeling accuracy and develop a hierarchical decision-making framework with both real-time response and global optimization. The computing architecture of edge intelligence and cloud collaboration will deploy lightweight algorithms at the edge to handle time-sensitive tasks and perform complex global optimization in the cloud to effectively respond to the computational challenges of resource scale expansion, and ensure the stable performance of the system at various scales through an elastic computing resource allocation mechanism.

(2)

Future VPP decision-making will shift from purely economic optimization to integrated consideration of economic, environmental, and social sustainability. The combination of deep reinforcement learning and system dynamics can achieve a balance between short-term gains and long-term sustainability. The future VPP will require introducing the theory of cooperative games and mechanism design, and constructing an operation model that promotes the voluntary participation of all parties and the equitable distribution of benefits.

(3)

VPPs should build a unified data exchange and model-sharing platform to promote knowledge migration and experience sharing among VPPs. The industry should formulate AI model evaluation and certification standards to establish an industry trust foundation. Organizations should form a data-sharing mechanism based on value contribution and clarify data property rights and value assessment methods. All stakeholders should strengthen interdisciplinary research and industry–university–research cooperation, integrate knowledge in multiple fields such as power systems, computer science, economics, etc., cultivate composite talents, and jointly promote the intelligent development of VPPs.

5. Conclusions

This paper centers on the application of the VPP in modern power systems, focusing on the role of AI technology in enhancing the intelligence level of the VPP.

First, the analysis introduces the basic concept of the VPP, which improves the flexibility and economy of the power system by integrating and managing distributed energy sources to cope with the uncertainty brought by the introduction of renewable energy sources; based on the theoretical framework of the VPP, it is derived that it has six main functional modules.

Secondly, the application of AI technologies in the six functional modules of the VPP is discussed in detail. In optimal scheduling and power prediction, the potential of technologies such as DL and RL is demonstrated; in resource analysis and market trading strategy decision-making, efficient resource scheduling is achieved through machine learning and NLP; and in resource aggregation and data security, innovation is also provided to achieve the flexible integration of DER and privacy protection solutions.

Finally, this article discusses the main challenges that the VPP may face in the future, including distributed resource heterogeneity exacerbating the modeling difficulty, a high misjudgement rate of prediction models under extreme events, and the leakage risk of user data, privacy protection, etc., and it puts forward the future development outlook based on these challenges.

The research results indicate that this study constructed a four-layer theoretical framework for the VPP, systematically reviewed the current application status of AI technologies in various functional modules of the VPP, and found that current AI technologies demonstrate certain advantages in the VPP but still have limitations such as insufficient coordination between modules and poor algorithm interpretability. Overall, with the continuous development of AI technology, the application of the VPP in multi-energy systems is promising, but multiple technical and market challenges need to be overcome in order to make further progress in terms of intelligence, data security, and efficient collaboration.