Next Article in Journal
Influence of Joint Stiffness and Motion Time on the Trajectories of Underactuated Robots
Next Article in Special Issue
The Detection of False Data Injection Attack for Cyber–Physical Power Systems Considering a Multi-Attack Mode
Previous Article in Journal
Drive-by Methodologies Applied to Railway Infrastructure Subsystems: A Literature Review—Part I: Bridges and Viaducts
Previous Article in Special Issue
Distribution System State Estimation Using Hybrid Traditional and Advanced Measurements for Grid Modernization
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Applied Sciences
Volume 13
Issue 12
10.3390/app13126937
applsci-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

Artificial Intelligence Applications in Electric Distribution Systems: Post-Pandemic Progress and Prospect

by
Sungjoo Chung
* and
Ying Zhang
*
Department of Electrical and Computer Engineering, Montana State University, Bozeman, MT 59717, USA
*
Authors to whom correspondence should be addressed.
Appl. Sci. 2023, 13(12), 6937; https://doi.org/10.3390/app13126937
Submission received: 30 April 2023 / Revised: 20 May 2023 / Accepted: 25 May 2023 / Published: 8 June 2023
(This article belongs to the Special Issue Research Progress on Cyber-Physical Distribution System)
Browse Figures
Versions Notes

Abstract

:
Advances in machine learning and artificial intelligence (AI) techniques bring new opportunities to numerous intractable tasks for operation and control in modern electric distribution systems. Nevertheless, AI applications for such grids as cyber-physical systems encounter multifaceted challenges, e.g., high requirements for the quality and quantity of training data, data efficiency, physical inconsistency, interpretability, and privacy concerns. This paper provides a systematic overview of the state-of-the-art AI methodologies in the post-pandemic era, represented by transfer learning, deep attention mechanism, graph learning, and their combination with reinforcement learning and physics-guided neural networks. Dedicated research efforts on harnessing such recent advances, including power flow, state estimation, voltage control, topology identification, and line parameter calibration, are categorized and investigated in detail. Revolving around the characteristics of distribution system operation and integration of distributed energy resources, this paper also illuminates prospects and challenges typified by the privacy, explainability, and interpretability of such AI applications in smart grids. Finally, this paper attempts to shed light on the deeper and broader prospects in the realm of smart distribution grids by interoperating them with smart building and transportation electrification

1. Introduction

First defined in 2006 by the U.S. National Science Foundation (NSF), cyber-physical systems (CPS) are “engineered systems that are built from, and depend upon, the seamless integration of computation and physical components” [1]. The technological advancement of information and communications technologies (ICTs) vastly strengthens the coupling of the cyber and physical layers of many systems [2]. Among them are cyber-physical distribution systems (CPDSs), which integrate the primary components of the distribution systems and digital technologies. The CPDS is an integral part of power systems to feed the electricity from transmission systems to communities and industries, significantly affecting decisions regarding the planning, monitoring, and coordination of the power grid [3].
The deployment of advanced metering infrastructure (AMI) integrated with the Internet of Things (IoTs) has transformed the traditional CPDSs. Smart meters, phasor measurement units (PMUs), smart appliances, and smart plugs enable the collection of various data [4], while advanced IoT architecture allows for fast and accurate data transmission for system monitoring [5]. Although these advancements have opened up opportunities to improve the reliability, security, and efficiency of the CPDS operation, numerous challenges have arisen. Traditional techniques might be ineffective in handling the collected big data, which can be characterized by high volume and dimension, diversity in variety, velocity, veracity, and values [6]. Furthermore, deep penetration of distributed energy resources (DERs) introduces new challenges and complexities, such as variations in voltages, power quality issues, uncertainties of DER generation, and bi-directional power flow [7]. More specially, distribution systems are distinguished from traditional transmission systems due to high r/x ratio, multi-phase unbalanced operation, limited sensor installation, poor observability, and frequent topology changes. Nonetheless, the traditional methods, most of which are originally developed for transmission systems, struggle to address such challenges [8]. As such, it is crucial to deploy techniques that are capable of processing, monitoring, and utilizing historical and real-time data of large volume, dimension, and heterogeneity.
The COVID-19 pandemic has led to a surge in data availability as more people transit their routine work online, as well as research funding and investment, accelerating the development of artificial intelligence (AI). Recent successes in AI techniques across various fields such as object detection [9,10] and machine translation [11,12] have proved their effectiveness in extracting information from a massive amount of data. Hence, AI applications in CPDSs are gaining an explosive amount of research interest as the promising solution to the above-mentioned challenges. The complexity of coupling the physical and cyber layers in distribution systems often requires developing and even combining various AI techniques to provide dedicated and sophisticated solutions to certain domains; Figure 1 summarizes these application domains and algorithm layers. Yet, currently, there is a lack of timely review of AI techniques developed during this period specifically for the CPDSs. In order to fill this gap, this paper provides a comprehensive and detailed review of various state-of-the-art AI-based algorithms developed in the CPDS. Furthermore, current challenges and potential research directions are discussed, while interoperability with smart buildings and intelligent transportation systems, as the latest trend, is highlighted.
The rest of this paper is organized as follows. Section 2 briefly overviews recent advances in AI techniques, especially those that are witnessed spurs of research interests after the COVID-19 pandemic. Section 3 details the latest application of these techniques in CPDSs. Section 4 presents the opportunities and challenges faced by the applications of AI. Finally, Section 5 concludes this paper.

2. Latest Algorithmic Progress in Artificial Intelligence

Among the massive and rapid development of AI techniques that benefit from the availability of large data sets and increased computational power in recent years, we here introduce several representatives among the latest research progress. Moreover, these algorithms are widely explored for power system operation and control in CPDSs, including transfer learning (TL), graph learning (GL), deep attention mechanism, deep reinforcement learning (DRL), and physics-guided neural networks (PGNNs). These latest algorithmic efforts address some conventional limitations of plainly structured neural networks (NNs), e.g., overfitting, low data efficiency, low adaptivity, absence of utilization of graph information, and physical inconsistency.

2.1. Transfer Learning

TL aims to improve the learning of a task in the target domain by leveraging knowledge gained by its learning in the source domain. Differentiated from traditional machine learning (ML) algorithms, TL is capable of re-purposing different related tasks to improve generalization in another setting without the assumption of the same distributions of training and testing data [13]. The instance weighting strategy is a common data-based TL approach where different weights are assigned to the source domain instances of the loss function based on their usefulness to the target domain task. More relevant source domain instances will be utilized as additional resources in the training of the model, improving the efficiency and performance of TL [14].

2.2. Graph Learning

GL aims to explore the relationship among the nodes and the system topology modeled as a graph. The input to GL is a graph, a non-Euclidean data set that assumes dependence between each instance (node) represented by the edge connections. The motivation behind GL is the realization that graphs are natural representations of many real-world systems such molecular fingerprints [15], traffic networks [16], and social networks [17]. Although GL can be categorized in various ways, an important categorization comes from discerning static and dynamic graphs. Static graphs maintain constant states over time, and they can be further categorized, depending on the convolution method, into spectral-based and spatial-based graph convolution networks (GCNs). Spectral-based GCN, rooted in spectral graph theory, is based on the Laplacian matrix in the spectral domain, while spatial-based GCN is based on the propagation of node information through the edges. Unlike static graphs, dynamic graphs contain critical information that is sensitive to time, and nodes and edges can be added and removed from the graph.
In recent years, graph neural networks (GNNs), a subset of GL, have received a surge of interest due to their success in natural language processing [18] and image recognition [19]. In particular, GNNs are classified into four categories [20]: recurrent GNN, convolutional GNN, graph autoencoders (GAEs) and spatial-temporal GNN. Convolutional GNN is a generalization of convolution from 2-D input to graphs, while GAE reconstructs graph data from nodes graphs encoded into latent vector space.

2.3. Deep Attention Mechanism

The deep attention mechanism is able to improve the performance of machine learning models by being selectively attentive to relevant parts of the input. The concept of attention was first introduced for machine translation tasks [21], where the proposed model specifically searched for parts of the sentences that are more relevant in the prediction of a target word, thereby achieving desirable prediction performances.
The main motivation behind the deep attention mechanism is to address the limitations of traditional neural networks that apply the same weights to all input elements. By assigning appropriate weights to each input element based on its relevance to the target task, it enables the model to be more efficient in allocating its computational resources. Moreover, by focusing on critical features of the target tasks, it allows models to capture the characteristics of complex systems such as long-range dependencies and integrability [22].

2.4. Deep Reinforcement Learning

Inspired by the success of AlphaGo, DRL has been widely used to solve complex power system decision and control problems in a time-varying and stochastic environment [23]. The decision-making process for distributing networks is formulated as a Markov decision process (MDP). Diversified reinforcement learning algorithms, roughly divided into Q-learning and policy gradient, are implemented to solve the MDP. In DRL, NNs are constructed as decision-making agents to offer control actions in an adaptive manner. Moreover, the efficiency of the actions is consistently optimized by the interaction of the agents and environment. During the offline training process, operational knowledge is extracted and embedded into the RL agents. In contrast, the model-free nature of RLs captures the characteristics of DERs and is capable of navigating through dynamic, real-time environments [24].

2.5. Physics-Guided Neural Networks

Existing simply structured NN-based approaches for distribution systems often produce results that do not conform with physics laws and are hard to generalize to out-of-sample boundaries [25]. To overcome overfitting and the dependency on the training data quality, PGNNs, as one of the latest advancements, are proposed to integrate the knowledge of physics laws regarding power system operation into the NN design [26]. Different types of physics knowledge are incorporated into NN architectures for dedicated applications in power systems, such as equation correspondence [27] and network topology [28] for power flow. Furthermore, PGNN shows promise in integrating different deep-learning techniques, as introduced before, to overcome the challenges of limited or noisy labeled data and varying operating environments in distribution systems by leveraging physics knowledge to supplement or guide the learning process.

3. Latest Applications in Distribution Systems

The proliferation of affordable and smart measuring instruments enables the collection of high-resolution accurate data. Moreover, the abundance of internet bandwidth and improvements in wireless communications allows for those devices to communicate reliably and effectively with the users and the power system. Furthermore, the widespread deployment of DERs produces bi-directional distribution system power flows. Thus, to fully utilize these benefits and harmoniously integrate cyber and physical components of the distribution system, key grid techniques ought to undergo appropriate improvements. Conventional computational techniques have proven to struggle to deal with challenges in four aspects [6,7,8]: (1) considerable computational burdens from the increasing volume and dimension of data; (2) the stochastic nature and intermittency of DER output; (3) dynamic complexity of inverter-connected resources, and (4) scalability in large-scale power systems. Thus, we investigate state-of-the-art AI-enabled algorithms developed in recent years for crucial functionalities in distribution systems, including power flow, state estimation, voltage control, topology identification, and line parameter calibration. Table 1 provides a concise technical summary that classifies these algorithms into various techniques introduced in Section 2.
Table
Table 1. The technical summary classifies the AI application in distribution systems into various learning-enabled techniques.
Table 1. The technical summary classifies the AI application in distribution systems into various learning-enabled techniques.
Research TopicAlgorithmComments and Keynotes for Selected Papers
Power FlowGL [29,30,31,32,33][29,30]: GNN deals with varying network topology and outperforms FCNN and CNN models
[33]: GNN is equipped with ARMA layers to prevent oversmoothing
Attention [31,32][31]: Inductive learning is used to enhance computation efficiency
[32]: Solar and wind power generations are accounted for probabilistic power flow.
PGNN [27,28]
DRL [34,35]		[28]: Network topology as physics knowledge
[34]: Adapt to topology changes and DER uncertainty
State Estimation
/Load Monitoring		GL [36,37,38,39][36]: Two-layer framework of GCN and GRU
[37]: Simultaneously captures spatiotemporal correlation of data
[38]: Pseudo-measurements are generated via GNN
[39]: Low pass property of voltage phasors are utilized to reconstruct network
TL [40,41,42][40]: A temporal convolutional network is developed to learn the dynamic features of individual appliance load
[42]: Privacy-preserving TL for non-intrusive load monitoring
PGNN [43,44,45][43]: PMU allocation are optimized and embedded into NN design
[44]: Jacobian matrix as physics knowledge to construct the loss function
[45]: Encoder–decoder architecture for evaluating physical consistency with measurement equations
Volt-Var ControlDRL [46,47,48,49,50,51,52,53,54][51]: Achieve near constraint satisfaction in time-varying conditions via a safe DRL algorithm
[53]: An adversarial SAC algorithm enhances robustness against modeling errors of distribution networks
[46,48,49,52,54]: Multi-agent DRL; physics-based shielding mechanism in PV/battery energy storage systems [48]
GL [49][49]: Privacy is preserved by training agents locally via GCNs
Attention [54][54]: Attention is used to automatically assign the learning weights for cooperative learning
Topology
Identification		GL [55,56,57,58][55]: Graph learning algorithm with backtracking
[56]: Power flow equation is used to estimate missing node measurements
[57]: Topology is estimated from a constructed topology graph bank
Attention [59,60][59]: Separate convolution layers for voltage and power measurements
[60]: Meta-paths are constructed via graph transformer network
Parameter
Calibration		GL [61][61]: Effects of temperature and humidity on line
PGNN [62,63][62]: The backward/forward sweep distribution flow impedances are accounted for formulation is used to construct the graph learning
[63]: Kron reduction is implemented to reduce computational complexity

3.1. Power Flow Analysis

Power flow analysis is fundamental to power system planning, monitoring, and control. Approaches that utilize numerical optimizers, such as the Newton–Raphson (NR) method, are conventionally implemented as the solution. Power flow requires meeting load demands while not violating the physical constraints of the system. The problem of achieving the optimal steady-state operating point of the system that minimizes the generation cost while being reliable and safe is referred to as the optimal power flow (OPF) problem. The OPF problem is NP-hard [64], and some numerical methods are computationally expensive and do not guarantee a global optimum [65]. Furthermore, they lack sufficient adaptability to changes happening in the system, e.g., DER output, and remodeling might be required, exacerbating computational costs [29].
The effectiveness of GNNs in dealing with varying operating conditions, by exploiting underlying correlations between nodes learned implicitly from modeling the power system network as a graph, is verified for the OPF problem in [29]. Meanwhile, such effectiveness of GNNs is extended in [30] by utilizing local marginal prices (LMPs) and voltage magnitudes, which are topology-dependent, to significantly reduce the model complexity. In parallel with centralized GNN-based methods, by framing the power flow balancing as a supervised vertex regression task, [33] develops a decentralized GNN to enable communication between vertices on the graph with no need for centralized operators, in which the deep GNN is equipped with convolutional auto-regressive moving average (ARMA) layers [66]. Results indicate that the decentralized GNN outperforms conventional multi-layer perceptron (MLP) approaches for fixed and varying topologies.
On the other hand, the deep attention mechanism is integrated into GL for deterministic and probabilistic power flow [31,32]. A physics-informed graph attention network is adopted for power flow analysis in [31], and the inductive learning ability of the attention-based model enhances computational efficiency. By aggregating neighboring information selectively through the graph attention-enabled convolution layer, [32] captures the complex correlation and stochasticity of wind and solar power generation, realizing improved performance in probability power flow calculation.
While GNNs are implemented to adapt to varying topologies, several efforts in DRL, for example, proximal policy optimization (PPO) in [34] and soft actor–critic (SAC) in [35] are made to improve the adaptability of the OPF problem to the varying operating conditions. The well-trained agent is able to achieve near-optimal status efficiently while dealing with uncertainties, such as topology changes and uncertainties of DERs in [34].

3.2. State Estimation/Load Monitoring

Distribution system state estimation (DSSE) is the backbone of modern distribution management systems (DMSs) to capture voltage phase angles and magnitudes at all buses, from available measurements, such as power injection/flow and voltage, for distribution system monitoring and control. Motivated by the increasing penetration level of DERs, it becomes vital to develop an accurate, efficient, and robust state estimation technique that addresses the complications that they generate [67]. Conventional DSSE methods that use measured data of the supervisory control and data acquisition (SCADA) system, such as the weighted least square (WLS) method, might be incapable of dealing with the system-wide unobservability, increased computational complexity, and topological variations emerging in modern distribution systems. Furthermore, these methods heavily rely on real-time measurements and fall short of clarifying important correlations hidden in historical DER data. Therefore, appropriate AI techniques, such as GNN, have been applied to address these shortcomings [68]. Moreover, this has also motivated the development of the wide area measurement system (WAMS) based on PMUs, and its integration with state-of-the-art AI techniques, which this section will explore in detail.
To capture the spatial and temporal dependencies of power system measurements, a two-layer framework is adopted in [36], where a graph convolution layer that captures the spatial dependencies is followed by a gated recurrent unit (GRU) layer that captures the temporal dependencies. Compared with support vector regression and Bayesian multivariate regression, ref. [36] shows improved accuracy, reduced latency, and computation time. Spatiotemporal correlations of historical measurement data are simultaneously captured in [37] by the proposed unrolled spatiotemporal GCN. This is accomplished by splicing spatial graphs of neighboring time steps which allows the capturing of the effects of adjacent nodes within current and adjacent time steps directly and simultaneously. Ref. [37] is also capable of capturing long-range dependencies by leveraging multi-module layers.
DSSE often suffers from poor network observability as economic constraints limit the number of sensors installed in distribution systems. To address this issue, ref. [38] leverages the inter-dependencies of nodes learned by the GNN architecture to generate pseudo-measurements to achieve accurate estimations with lower computational times. Similarly, ref. [39] deals with the problem of poor network observability with sparse AMI measurements by utilizing the low pass property of voltage phasors of the buses. Then, graph signal processing (GSP) and Shannon’s sampling theorem are applied to reconstruct AMI measurements to enhance the robustness and scalability.
Reference [43] develops an optimization strategy of PMU placement for NN design and then a physics-aware NN is proposed to utilize the PMU placement to determine the NN structure, which reduces the complexity of the NN architecture as well as enhances robustness against over-fitting. In the context of PGNN, ref. [44] incorporates line/shunt admittance and grid topology information into the NN architecture. Estimation accuracy is significantly improved when compared to the widely used WLS method, without specific PMU placement requirements. Rather than embedding physical information into the architecture, ref. [45] proposes a framework where the estimated states produced by the encoder-decoder architecture are evaluated against power flow equations governed by physical laws. Then, the loss function is shaped as the measurement weighted residual, which is back-propagated to adjust weights and biases in the NN for improved estimations. The accuracy and robustness of DSSE are greatly improved relative to conventionally used techniques such as WLS and long short-term memory (LSTM) networks.
On the grid edge, non-intrusive load monitoring (NILM), which aims to disaggregate whole-house power readings into appliance-wise power readings, has been a popular area of TL applications recently [40]. In [41], the effectiveness of TL in the form of seq2point learning is demonstrated across not only different appliances but also power readings. This is made possible due to the CNN architecture extracting domain-invariant features from the power readings. As domain-invariant features can be extracted, computational costs of NILM models are significantly reduced as they do not need to be re-trained for different appliances or power readings from different regions. Privacy-preserving TL framework is introduced in [42], and a transferable tree-based ensemble model is developed and only requires training samples from the target domain, thus reliably protecting the privacy of the users in the source domain.

3.3. Volt/VAR Control

Volt/VAR control (VVC), as another core module in DMSs, aims to minimize voltage violation and reduce power losses by optimizing voltage profiles in distribution systems [69]. It determines the best available combination of settings from controllable devices, including on-load tap changers (OLTCs), capacitor banks (CBs), inverter-connected photovoltaics (PVs), voltage regulators, and dispatchable electric vehicles (EVs). Conventionally, model-based optimization methods such as conic relaxation and mixed integer linear programming are adopted. However, not only does DER penetration cause variations in the network topology, but it also introduces intermittency and uncertainty in DER generation [46]. These changes might cause sharp fluctuations in voltage, which lead to serious overvoltage/undervoltage issues.
Since 2019, the development of VVC has centered around data-driven approaches, with a particular emphasis on DRL-based adaptive decision-making. Instead of depending on accurate power flow models to formulate an optimization model for VVC, decision-making agent(s) are represented as NNs and trained to determine the control variables in a data-driven adaptive manner. This framework allows for sample-efficient training and effective online exploration of the state-action space, which enhances the adaptability of the strategy to capture the stochastic behaviors of DER generation. The DRL algorithms for VVC methods evolve from deep Q-network (DQN) [46], deep deterministic policy gradient (DDPG) [47], twin delayed deep deterministic (MATD3) policy gradient [48], PPO [49], and SAC [50]. This topic gradually migrates to two branches: safe DRL [51,52] for satisfying inequality constraints and robust DRL against modeling errors of distribution networks [53]. Safe DRL seeks to feasible control actions within required constraints towards actions or states, since certain control actions violating these constraints might cause equipment damage and undermine the operating reliability. In [53], an adversarial MDP (AMDP), which utilizes an adversarial agent representing modeling errors, is formulated. Furthermore, a joint adversarial SAC algorithm is proposed, which enhances the robustness through offline training and guarantees an efficient convergence of the training process. Another recent development in DRL is the use of meta-learning, which trains a model to learn how to learn. This approach allows the model to quickly adapt to new operating environments; see a paradigm in [70] for power system emergency control.
Beyond the centralized VVC, multi-agent DRL, featuring “centralized/decentralized training, decentralized implementation”, is developed for cooperative control by multiple agents to learn optimal policies that maximize their individual and collective rewards, meanwhile preserving privacy. Among them, the attention mechanism is embedded to weigh the impact of other neighboring agents during offline training [54]. Reference [49] adopts a GCN which captures the known topology-based relationships between neighboring nodes in a PPO-based DRL framework. Furthermore, the safety of battery energy storage systems (BESSs) is often overlooked in MARL-based VVC methods, and an agent’s actions that cause excessive charging or discharging of the BESS shorten the lifespan. Thus, ref. [48] proposes a physics-based shielding mechanism in PV/battery energy storage systems.

3.4. Topology Identification

Key functionalities, such as state estimation, power flow analysis, and voltage control, are built on known network topology. With the expanding integration of DERs, the network topology changes more frequently, making it difficult for real-time topology identification in distribution systems. On the other hand, the recent surge in smart grid technologies has led to the development of AMIs and PMUs, increasing the availability of high-resolution real-time measurements. Thus, it is of interest in distribution networks to utilize this information to develop topology identification methods [7].
Conventional model-based approaches for topology identification assume complete observability from available nodal voltage and line measurements [71,72]. Furthermore, works that consider systems with missing nodal voltage measurements, such as [73,74], leverage relevant historical measurements for data imputation. Realistic distribution systems are featured by partial observability due to limited available sensors that should be considered for applicable topology identification [56]. The metrics of voltage sensitivity are exploited [55] to reconstruct the network topology with partial observability. Based on node-to-node distances, candidate topologies are generated, and the optimally matched topology is determined by GL to minimize a multi-criteria objective regarding voltage sensitivity, magnitudes, and correlation coefficients. A Bayesian network (BN) is adopted in [57] for its robustness against missing data, as BN can capture the probability of partial observability of the network. Then, the switching relationship between different topologies is exploited to develop a topology graph bank. Specifically, by exploiting a limited number of possible posterior topologies produced from switch positions, an estimated topology is obtained within the according bank based on real-time measurements, and any unobservable nodes can be estimated by power flow analysis or state estimation techniques from the estimated topology. A similar approach, in regard to utilizing the set of possible topologies that can be achieved from a specific topology, is adopted in [58].
References [59,60] combine GL with a deep attention mechanism to overcome the traditional scalability and generalizability limitations of GL. Considering different features of voltages and power, [59] implements a dual-graph attention-based framework that applies separate convolution layers for voltage and power measurements to learn their features at a deeper level. Then, an attention pooling layer is adopted to select the most relevant nodes for topology identification. A graph transformer network (GTN), which introduces the concept of transformers [12] to GNN, is adopted in [60] mainly to establish meta-paths and extract useful correlations within these paths that traditional GNNs fail to capture.

3.5. Line Parameter Calibration

Besides topology, the operation, control, and analysis of distribution systems are also built on accurate model parameters such as line impedance. However, such information is often unavailable, incomplete, or outdated due to the system upgrade, especially in secondary distribution systems. Some utilities only have single-line diagrams of their systems without detailed three-phase line parameters; other utilities possess system models, but they are often incomplete or outdated due to frequent system expansion and reconfiguration [75]. Conventional parameter calibration methods either require the widespread installment of costly sensors such as PMUs or assume a simplified single-phase distribution network mode, which is not suitable for industrial practice. To overcome these challenges, several AI techniques are applied to solve the parameter calibration problem for three-phase distribution lines.
In [61], a dynamic Bayesian network (DBN) with the consideration of external environment factors such as temperature and humidity is constructed for line parameter calibration. Such consideration is based on the observation of seasonal changes of line impedance parameters, due to the line absorbing moisture from the surrounding environment resulting in a decrease in insulation resistance. Thus, the DBN models time-varying correlations between specified environmental factors, network measurement data, and line parameters. By representing these correlations as probabilistic distributions, robustness to data noise and errors is enhanced. Physics knowledge, in the form of power flow equations, is embedded in the graphical learning models of [62,63]. Specifically, a three-phase power flow-based transition function is constructed in [62] to replace the transition functions of GNNs. Then, the gradient of the voltage magnitude loss function with respect to line parameters is derived and utilized to update line parameter estimates through the iterative application of the stochastic gradient descent method. Reference [63] reduces the computation complexity of the proposed PGNN model and addresses the partial observability issue by eliminating unobservable nodes of the power flow equation via Kron reduction [76].

4. Opportunities and Challenges

4.1. Privacy Concerns

With the widespread deployment of AMIs, IoTs, and ICT systems, deep integration of the cyber and physical layers of distribution systems raises serious privacy and cybersecurity concerns. Traditional centralized model training frameworks require large amounts of raw data to be transmitted to a central server for effective model training. However, the transmission of raw data in the network is especially vulnerable to cyberattacks.
One solution to the above-mentioned privacy concerns is federated learning (FL). First introduced in 2016, FL is developed to address these privacy concerns of traditional learning models [77]. FL allows local devices to collaboratively train a global model without sharing their local data [78]. As models are trained locally, raw data are kept locally, which alleviates the computational communication burdens of the centralized model. Several attempts to integrate FL into power system operation and control are explored in [79,80]. Reference [79] proposes a decentralized federated DRL-based VVC in distribution networks. Specifically, private measurements and policies of virtual power plants (VPPs) are all kept locally to agents, and these local agents only exchange a few scalars with the coordinated agent, significantly enhancing communication efficiency and data privacy. An FL framework is adopted in [80] to detect false data injection attacks (FDIAs). Not only is privacy preserved by training the data locally for the detection of FDIA, but generated models are also deployed locally, enabling efficient detection as communication delays caused by transmitting data between nodes and a central server are alleviated. To further protect data privacy from hackers, the Paillier cryptosystem is integrated with the FL framework by encrypting model parameters exchanged during training.
Generally, privacy protection in machine learning addresses three objectives: privacy of the data used for learning a model as input, privacy of the model, and privacy of the model output. The related techniques include privacy-preserving probabilistic inference [81] and computing on encrypted data [82], and however, their application to CPDSs remains open. Future development of AI techniques must take into consideration the privacy implications, not only to combat security breaches that can cause data theft but also to motivate the users to share their data for more efficient and reciprocal demand response. Securing privacy in big data analytics for smart grids remains an ongoing challenge that requires consistent attention and combined endeavors for adequate solutions throughout data collection, storage, analysis, and sharing.

4.2. Explainability and Interpretability

Deep learning algorithms typically built on NNs are difficult to be trusted due to their hard-to-explain “black-box” nature. Despite their attractive merits, as mentioned, their practical use in power systems is sometimes limited, considering any unreliable decision-making and control demands might result in devastating damages and even blackouts [83]. Therefore, it is crucial to sufficiently understand how the learning model reaches its prediction. Thus, the explainability of AI techniques has been receiving the attention of many researchers, and this concept is introduced by explainable AI (XAI) in the literature.
One XAI technique, SHapley Additive exPlanation (SHAP) value, already introduced in power systems, explains the results of the ML model based on game theory by providing measures of feature importance [84]. Deep-SHAP-based methods are exploited for power system emergency control [85] and fault identification in grid-connected PV systems [86]. Even though various works have explored the application of XAI to power systems, most of them focus on transmission systems [87], while branching out to one grid-edge application, i.e., NILM [88,89,90]. It can be speculated that the complexity of the distribution system hinders the extension of such application, which demands deliberation.
The performance of XAI techniques is vulnerable to the quality of the training population. Thus, the training conducted on limited data could introduce “biases” to the XAI model, leading to an inaccurate explanation of the result. A potential approach to mitigating this effect could be inspired by the combination of GL and PGNN with XAI. It is demonstrated, as discussed before, that these techniques have been utilized to address partial observability in distribution systems. Therefore, XAI techniques should be developed with realistic assumptions about the data quality, enhancing its robustness. Yet another issue arises due to the problem of partial observability of distribution systems: the distinction between model interpretation and certification [91]. However, there is no consensus regarding the effectiveness of explanation and evaluation on such distinction.

4.3. Interoperability with Smart Buildings and Intelligent Transportation Systems

Technological developments made at the grid edge, in the form of smart buildings, EVs, and intelligent transportation systems, evolve these components from passive to active participants in the system. Figure 2 summarizes the interaction of smart buildings and intelligent transportation systems with a power system. This presents numerous opportunities to enhance the efficiency and robustness of the grid, as components work harmoniously in the context of smart cities.
Grid-interactive buildings are capable of adapting to time-of-use electricity pricing and responding to external weather conditions using prediction control strategies on resources, such as heating, ventilation, and air conditioning (HVAC) and storage systems [92]. Accompanied by the development of AI techniques, such prediction control strategies become more efficient, sustainable, and robust. Specifically, DRL has shown considerable effectiveness in overcoming several challenges posed in this domain, such as dealing with varying building environments [93]. Furthermore, the deep penetration of DERs converts users to prosumers and introduces unforeseen complexities through a bi-directional interactive system between the prosumers and the grid in a building. Building energy management systems (BEMSs) seek to improve energy efficiency and conservation. The field of AI brings novel and efficient approaches to the development of smart BEMSs in areas such as pattern prediction of energy usage, load profiling, detection of electricity anomaly/theft, and load profiling [94,95].
EVs bridge between power networks and transportation systems and augment the grid with the capability to store and deliver electricity in a variety of locations, therefore, optimizing energy efficiency. According to [96], California currently has the greatest number of EVs nationwide, i.e., more than half a million by 2022. This represents a significant potential to make the power system more efficient. Specifically, through a collaborative communication infrastructure, EV owners are able to optimize their charging behavior based on multi-dimensional data, such as traffic data, charging prices, commuting time, locations of EV charging stations, etc. In return, power grids enable the dispatch of servable EVs as DERs for voltage regulation [97]. Furthermore, advances in technology have enabled not only vehicle-to-grid (V2G) interactions but also vehicle-to-vehicle and vehicle-to-home interactions. This enables a more flexible allocation of resources that can be utilized to meet real-time demand. Therefore, it is essential that the development of technologies efficiently model the behaviors of these technologies and their interaction with the grid.

5. Conclusions

This paper presents technical development and AI-enabled solutions in distribution systems since the COVID-19 pandemic and then reviews the recent works, their technical merits, and ongoing trends in the AI-energy interdisciplinary research community. Combined with operating practice and model-based algorithmic developments in the last decades, learning-enabled methodologies for smart grid operation and control are heading toward the future with next-generation distribution systems. To facilitate seamless transitions towards a clean and intelligent energy future, it is crucial to help researchers and stakeholders understand the application scenarios of the AI technologies in practical power systems. Moreover, rationalizing the theoretical fundamental behind the state-of-the-art AI and meanwhile combing with existing research efforts and industrial practice will efficiently push forward its success from computer science to dedicated implementation of distribution system operation and control. Therefore, regarding this, this paper provides a review of several emerging developments in the past three years and specifically clarifies the limitations of the conventional approaches for power flow, state estimation, voltage control, topology identification, and line parameter calibration and the AI-based algorithms that are capable of addressing these limitations.
This paper also illuminates opportunities and challenges typified by the privacy, explainability, and interpretability of such AI applications in smart grids. These emerging challenges emphasize the importance of privacy-preserved AI and XAI, which will lead to better insights and more informed decision-making and meanwhile securing utility and individual privacy. However, these still remain open for distribution systems with increasing grid-edge energy resources. On the other hand, the smart grid is rapidly expanding to couple with smart buildings and intelligent transportation systems by jointly dispatching and managing these energy resources. Such extended cyber-physical systems pose higher requirements for their interoperation.

Author Contributions

Conceptualization, Y.Z.; methodology, Y.Z. and S.C.; validation, Y.Z. and S.C.; formal analysis, Y.Z. and S.C.; investigation, Y.Z and S.C.; resources, Y.Z. and S.C.; data curation, Y.Z. and S.C.; writing—original draft preparation, Y.Z. and S.C.; writing—review and editing, Y.Z. and S.C.; visualization, Y.Z.; supervision, Y.Z. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Cyber-Physical Systems (CPS). Available online: https://beta.nsf.gov/funding/opportunities/cyber-physical-systems-cps (accessed on 9 April 2023).
  2. Rajkumar, R.; Lee, I.; Sha, L.; Stankovic, J. Cyber-physical systems: The next computing revolution. In Proceedings of the 47th Design Automation Conference, Anaheim, CA, USA, 13–18 June 2010; pp. 731–736. [Google Scholar] [CrossRef]
  3. Zhao, H.; Liu, H.; Gu, Y.; Han, L.; Zhao, W.; Du, Y. Overview of Architecture and Planning of Cyber Physical Distribution System. In Proceedings of the 2021 IEEE Sustainable Power and Energy Conference (iSPEC), Nanjing, China, 23–25 December 2021; pp. 4174–4180. [Google Scholar] [CrossRef]
  4. Bose, A. Smart Transmission Grid Applications and Their Supporting Infrastructure. IEEE Trans. Smart Grid 2010, 1, 11–19. [Google Scholar] [CrossRef]
  5. Nirmal, D. Artificial Intelligence Based Distribution System Management and Control. J. Electron. 2020, 2, 137–147. [Google Scholar] [CrossRef]
  6. Bhattarai, B.; Paudyal, S.; Luo, Y.; Mohanpurkar, M.; Cheung, K.; Tonkoski, R.; Hovsapian, R.; Myers, K.; Zhang, R.; Zhao, P.; et al. Big Data Analytics in Smart Grids: State-of-the-Art, Challenges, Opportunities, and Future Directions. IET Smart Grid 2019, 2, 141–154. [Google Scholar] [CrossRef]
  7. Zhang, Y.; Wang, J.; Li, Z. Uncertainty modeling of distributed energy resources: Techniques and challenges. Curr. Sustain. Energy Rep. 2019, 6, 42–51. [Google Scholar] [CrossRef]
  8. Javed, A.H.; Nguyen, P.H.; Morren, J.; Slootweg, J.H. Review of Operational Challenges and Solutions for DER Integration with Distribution Networks. In Proceedings of the 2021 56th International Universities Power Engineering Conference (UPEC), Middlesbrough, UK, 31 August–3 September 2021; pp. 1–6. [Google Scholar] [CrossRef]
  9. Howard, A.G.; Zhu, M.; Chen, B.; Kalenichenko, D.; Wang, W.; Weyand, T.; Andreetto, M.; Adam, H. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017, arXiv:1704.04861. [Google Scholar]
  10. Beery, S.; Wu, G.; Rathod, V.; Votel, R.; Huang, J. Context R-CNN: Long Term Temporal Context for Per-Camera Object Detection. In Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Seattle, WA, USA, 13–19 June 2020. [Google Scholar]
  11. Johnson, M.; Schuster, M.; Le, Q.V.; Krikun, M.; Wu, Y.; Chen, Z.; Thorat, N.; Viégas, F.; Wattenberg, M.; Corrado, G.; et al. Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation. arXiv 2017, arXiv:1611.04558. [Google Scholar] [CrossRef] [Green Version]
  12. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, L.; Polosukhin, I. Attention Is All You Need. arXiv 2017, arXiv:1706.03762. [Google Scholar]
  13. Pan, S.J.; Yang, Q. A Survey on Transfer Learning. IEEE Trans. Knowl. Data Eng. 2010, 22, 1345–1359. [Google Scholar] [CrossRef]
  14. Lee, K.; Han, S.; Pham, V.H.; Cho, S.; Choi, H.J.; Lee, J.; Noh, I.; Lee, S.W. Multi-Objective Instance Weighting-Based Deep Transfer Learning Network for Intelligent Fault Diagnosis. Appl. Sci. 2021, 11, 2370. [Google Scholar] [CrossRef]
  15. Duvenaud, D.; Maclaurin, D.; Aguilera-Iparraguirre, J.; Gómez-Bombarelli, R.; Hirzel, T.; Aspuru-Guzik, A.; Adams, R.P. Convolutional Networks on Graphs for Learning Molecular Fingerprints. arXiv 2015, arXiv:1509.09292. [Google Scholar]
  16. Yu, B.; Yin, H.; Zhu, Z. Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intelligence Organization, Stockholm, Sweden, 13–19 July 2018. [Google Scholar] [CrossRef] [Green Version]
  17. Zhang, M.; Chen, Y. Link Prediction Based on Graph Neural Networks. arXiv 2018, arXiv:1802.09691. [Google Scholar]
  18. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. arXiv 2014, arXiv:1412.3555. [Google Scholar]
  19. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep Residual Learning for Image Recognition. arXiv 2015, arXiv:1512.03385. [Google Scholar]
  20. Wu, Z.; Pan, S.; Chen, F.; Long, G.; Zhang, C.; Yu, P.S. A Comprehensive Survey on Graph Neural Networks. IEEE Trans. Neural Netw. Learn. Syst. 2021, 32, 4–24. [Google Scholar] [CrossRef] [Green Version]
  21. Bahdanau, D.; Cho, K.; Bengio, Y. Neural Machine Translation by Jointly Learning to Align and Translate. arXiv 2016, arXiv:1409.0473. [Google Scholar]
  22. Hernández, A.; Amigó, J.M. Attention Mechanisms and Their Applications to Complex Systems. Entropy 2021, 23, 283. [Google Scholar] [CrossRef]
  23. Li, F.; Du, Y. From AlphaGo to power system AI: What engineers can learn from solving the most complex board game. IEEE Power Energy Mag. 2018, 16, 76–84. [Google Scholar] [CrossRef]
  24. Chen, X.; Qu, G.; Tang, Y.; Low, S.; Li, N. Reinforcement Learning for Selective Key Applications in Power Systems: Recent Advances and Future Challenges. IEEE Trans. Smart Grid 2022, 13, 2935–2958. [Google Scholar] [CrossRef]
  25. Willard, J.; Jia, X.; Xu, S.; Steinbach, M.; Kumar, V. Integrating Scientific Knowledge with Machine Learning for Engineering and Environmental Systems. arXiv 2022, arXiv:2003.04919. [Google Scholar] [CrossRef]
  26. Raissi, M.; Perdikaris, P.; Karniadakis, G.E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 2019, 378, 686–707. [Google Scholar] [CrossRef]
  27. Hu, X.; Hu, H.; Verma, S.; Zhang, Z.L. Physics-Guided Deep Neural Networks for Power Flow Analysis. IEEE Trans. Power Syst. 2021, 36, 2082–2092. [Google Scholar] [CrossRef]
  28. Donon, B.; Clément, R.; Donnot, B.; Marot, A.; Guyon, I.; Schoenauer, M. Neural networks for power flow: Graph neural solver. Electr. Power Syst. Res. 2020, 189, 106547. [Google Scholar] [CrossRef]
  29. Falconer, T.; Mones, L. Leveraging power grid topology in machine learning assisted optimal power flow. IEEE Trans. Power Syst. 2022, 38, 2234–2246. [Google Scholar] [CrossRef]
  30. Liu, S.; Wu, C.; Zhu, H. Topology-aware Graph Neural Networks for Learning Feasible and Adaptive AC-OPF Solutions. IEEE Trans. Power Syst. 2022, 1–11. [Google Scholar] [CrossRef]
  31. Jeddi, A.B.; Shafieezadeh, A. A Physics-Informed Graph Attention-based Approach for Power Flow Analysis. In Proceedings of the 2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA), Pasadena, CA, USA, 13–16 December 2021; pp. 1634–1640. [Google Scholar] [CrossRef]
  32. Wu, H.; Wang, M.; Xu, Z.; Jia, Y. Graph Attention Enabled Convolutional Network for Distribution System Probabilistic Power Flow. IEEE Trans. Ind. Appl. 2022, 58, 7068–7078. [Google Scholar] [CrossRef]
  33. Hansen, J.B.; Anfinsen, S.N.; Bianchi, F.M. Power Flow Balancing With Decentralized Graph Neural Networks. IEEE Trans. Power Syst. 2022, 38, 2423–2433. [Google Scholar] [CrossRef]
  34. Zhou, Y.; Zhang, B.; Xu, C.; Lan, T.; Diao, R.; Shi, D.; Wang, Z.; Lee, W.J. A Data-driven Method for Fast AC Optimal Power Flow Solutions via Deep Reinforcement Learning. J. Mod. Power Syst. Clean Energy 2020, 8, 1128–1139. [Google Scholar] [CrossRef]
  35. Cao, D.; Hu, W.; Xu, X.; Wu, Q.; Huang, Q.; Chen, Z.; Blaabjerg, F. Deep Reinforcement Learning Based Approach for Optimal Power Flow of Distribution Networks Embedded with Renewable Energy and Storage Devices. J. Mod. Power Syst. Clean Energy 2021, 9, 1101–1110. [Google Scholar] [CrossRef]
  36. Hossain, M.J.; Rahnamay–Naeini, M. State Estimation in Smart Grids Using Temporal Graph Convolution Networks. In Proceedings of the 2021 North American Power Symposium (NAPS), College Station, TX, USA, 14–16 November 2021; pp. 1–5. [Google Scholar] [CrossRef]
  37. Wu, H.; Xu, Z.; Wang, M. Unrolled Spatiotemporal Graph Convolutional Network for Distribution System State Estimation and Forecasting. IEEE Trans. Sustain. Energy 2023, 14, 297–308. [Google Scholar] [CrossRef]
  38. Madbhavi, R.; Natarajan, B.; Srinivasan, B. Graph Neural Network-Based Distribution System State Estimators. IEEE Trans. Ind. Inform. 2023, 1–10. [Google Scholar] [CrossRef]
  39. Saha, S.S.; Scaglione, A.; Ramakrishna, R.; Johnson, N.G. Distribution Systems AC State Estimation via Sparse AMI Data Using Graph Signal Processing. IEEE Trans. Smart Grid 2022, 13, 3636–3649. [Google Scholar] [CrossRef]
  40. Lin, J.; Ma, J.; Zhu, J.; Liang, H. Deep Domain Adaptation for Non-Intrusive Load Monitoring Based on a Knowledge Transfer Learning Network. IEEE Trans. Smart Grid 2022, 13, 280–292. [Google Scholar] [CrossRef]
  41. DIncecco, M.; Squartini, S.; Zhong, M. Transfer Learning for Non-Intrusive Load Monitoring. arXiv 2019, arXiv:1902.08835. [Google Scholar] [CrossRef] [Green Version]
  42. Chang, X.; Li, W.; Xia, C.; Yang, Q.; Ma, J.; Yang, T.; Zomaya, A.Y. Transferable Tree-Based Ensemble Model for Non-Intrusive Load Monitoring. IEEE Trans. Sustain. Comput. 2022, 7, 970–981. [Google Scholar] [CrossRef]
  43. Zamzam, A.S.; Sidiropoulos, N.D. Physics-Aware Neural Networks for Distribution System State Estimation. arXiv 2019, arXiv:1903.09669. [Google Scholar] [CrossRef] [Green Version]
  44. Ostrometzky, J.; Berestizshevsky, K.; Bernstein, A.; Zussman, G. Physics-Informed Deep Neural Network Method for Limited Observability State Estimation. arXiv 2020, arXiv:1910.06401. [Google Scholar]
  45. Wang, L.; Zhou, Q.; Jin, S. Physics-guided Deep Learning for Power System State Estimation. J. Mod. Power Syst. Clean Energy 2020, 8, 607–615. [Google Scholar] [CrossRef]
  46. Zhang, Y.; Wang, X.; Wang, J.; Zhang, Y. Deep Reinforcement Learning Based Volt-VAR Optimization in Smart Distribution Systems. IEEE Trans. Smart Grid 2021, 12, 361–371. [Google Scholar] [CrossRef]
  47. Sun, X.; Qiu, J. Two-Stage Volt/Var Control in Active Distribution Networks with Multi-Agent Deep Reinforcement Learning Method. IEEE Trans. Smart Grid 2021, 12, 2903–2912. [Google Scholar] [CrossRef]
  48. Chen, P.; Liu, S.; Wang, X.; Kamwa, I. Physics-Shielded Multi-Agent Deep Reinforcement Learning for Safe Active Voltage Control with Photovoltaic/Battery Energy Storage Systems. IEEE Trans. Smart Grid 2022, 1. [Google Scholar] [CrossRef]
  49. Wang, Y.; Qiu, D.; Wang, Y.; Sun, M.; Strbac, G. Graph Learning-Based Voltage Regulation in Distribution Networks with Multi-Microgrids. IEEE Trans. Power Syst. 2023, 1–15. [Google Scholar] [CrossRef]
  50. Hu, D.; Peng, Y.; Yang, J.; Deng, Q.; Cai, T. Deep Reinforcement Learning Based Coordinated Voltage Control in Smart Distribution Network. In Proceedings of the 2021 International Conference on Power System Technology (POWERCON), Haikou, China, 8–9 December 2021; pp. 1030–1034. [Google Scholar] [CrossRef]
  51. Wang, W.; Yu, N.; Gao, Y.; Shi, J. Safe off-policy deep reinforcement learning algorithm for volt-var control in power distribution systems. IEEE Trans. Smart Grid 2019, 11, 3008–3018. [Google Scholar] [CrossRef]
  52. Nguyen, H.T.; Choi, D.H. Three-Stage Inverter-Based Peak Shaving and Volt-VAR Control in Active Distribution Networks Using Online Safe Deep Reinforcement Learning. IEEE Trans. Smart Grid 2022, 13, 3266–3277. [Google Scholar] [CrossRef]
  53. Liu, H.; Wu, W. Two-Stage Deep Reinforcement Learning for Inverter-Based Volt-VAR Control in Active Distribution Networks. IEEE Trans. Smart Grid 2021, 12, 2037–2047. [Google Scholar] [CrossRef]
  54. Cao, D.; Zhao, J.; Hu, W.; Ding, F.; Huang, Q.; Chen, Z. Attention enabled multi-agent DRL for decentralized volt-VAR control of active distribution system using PV inverters and SVCs. IEEE Trans. Sustain. Energy 2021, 12, 1582–1592. [Google Scholar] [CrossRef]
  55. Pengwah, A.B.; Fang, L.; Razzaghi, R.; Andrew, L.L.H. Topology Identification of Radial Distribution Networks Using Smart Meter Data. IEEE Syst. J. 2022, 16, 5708–5719. [Google Scholar] [CrossRef]
  56. Deka, D.; Chertkov, M.; Backhaus, S. Joint Estimation of Topology and Injection Statistics in Distribution Grids With Missing Nodes. IEEE Trans. Control. Netw. Syst. 2020, 7, 1391–1403. [Google Scholar] [CrossRef] [Green Version]
  57. Liu, Y.; Ren, P.; Zhao, J.; Liu, T.; Wang, Z.; Tang, Z.; Liu, J. Real-Time Topology Estimation for Active Distribution System Using Graph-Bank Tracking Bayesian Networks. IEEE Trans. Ind. Inform. 2023, 19, 6127–6137. [Google Scholar] [CrossRef]
  58. Anderson, O.; Yu, N. Detect and Identify Topology Change in Power Distribution Systems Using Graph Signal Processing. In Proceedings of the 2021 IEEE PES Innovative Smart Grid Technologies Europe (ISGT Europe), Espoo, Finland, 18–21 October 2021; pp. 1–6. [Google Scholar] [CrossRef]
  59. Lin, G.; Liu, S.; Shi, D.; Wang, X.; Liu, S. A Dual-Graph Attention-Based Approach for Identifying Distribution Network Topology. In Proceedings of the 2022 IEEE 10th International Conference on Computer Science and Network Technology (ICCSNT), Dalian, China, 22–23 October 2022; pp. 29–33. [Google Scholar] [CrossRef]
  60. Zhao, Z.; Qiao, J.; Li, J.; Shi, M.; Wang, X. Distribution Network Topology Identification with Graph Transformer Neural Network. In Proceedings of the 2022 4th International Conference on Smart Power & Internet Energy Systems (SPIES), Beijing, China, 9–12 December 2022; pp. 1580–1585. [Google Scholar] [CrossRef]
  61. Liu, J.; Tang, H.; Wu, Z.; Huang, Z.; Wu, Y.; Sun, J.; Chen, S. Dynamic Parameter Identification Technology of Distribution Network Line Based on Dynamic Graph Model. In Proceedings of the 2021 IEEE 11th Annual International Conference on CYBER Technology in Automation, Control, and Intelligent Systems (CYBER), Jiaxing, China, 27–31 July 2021; pp. 253–258. [Google Scholar] [CrossRef]
  62. Wang, W.; Yu, N. Estimate Three-Phase Distribution Line Parameters With Physics-Informed Graphical Learning Method. IEEE Trans. Power Syst. 2022, 37, 3577–3591. [Google Scholar] [CrossRef]
  63. Pagnier, L.; Chertkov, M. Physics-Informed Graphical Neural Network for Parameter & State Estimations in Power Systems. arXiv 2021, arXiv:2102.06349. [Google Scholar]
  64. Lehmann, K.; Grastien, A.; Van Hentenryck, P. AC-Feasibility on Tree Networks is NP-Hard. IEEE Trans. Power Syst. 2016, 31, 798–801. [Google Scholar] [CrossRef] [Green Version]
  65. Yan, M.; Shahidehpour, M.; Paaso, A.; Zhang, L.; Alabdulwahab, A.; Abusorrah, A. A Convex Three-Stage SCOPF Approach to Power System Flexibility With Unified Power Flow Controllers. IEEE Trans. Power Syst. 2021, 36, 1947–1960. [Google Scholar] [CrossRef]
  66. Bianchi, F.M.; Grattarola, D.; Livi, L.; Alippi, C. Graph Neural Networks With Convolutional ARMA Filters. IEEE Trans. Pattern Anal. Mach. Intell. 2022, 44, 3496–3507. [Google Scholar] [CrossRef]
  67. Zhang, Y.; Wang, J.; Li, Z. Interval State Estimation With Uncertainty of Distributed Generation and Line Parameters in Unbalanced Distribution Systems. IEEE Trans. Power Syst. 2020, 35, 762–772. [Google Scholar] [CrossRef] [Green Version]
  68. Wu, Z.; Wang, Q.; Liu, X. State Estimation for Power System Based on Graph Neural Network. In Proceedings of the 2022 IEEE 5th International Electrical and Energy Conference (CIEEC), Nangjing, China, 27–29 May 2022; pp. 1431–1436. [Google Scholar] [CrossRef]
  69. Zhao, T.; Zhang, Y.; Yue, M. Scalable Deep Reinforcement Learning-based Volt-VAR Optimization in Distribution Systems: A Mean-field Approach. In Proceedings of the 2022 IEEE Power & Energy Society General Meeting (PESGM), Denver, CO, USA, 17–21 July 2022; pp. 1–5. [Google Scholar] [CrossRef]
  70. Huang, R.; Chen, Y.; Yin, T.; Huang, Q.; Tan, J.; Yu, W.; Li, X.; Li, A.; Du, Y. Learning and fast adaptation for grid emergency control via deep meta reinforcement learning. IEEE Trans. Power Syst. 2022, 37, 4168–4178. [Google Scholar] [CrossRef]
  71. Bolognani, S.; Bof, N.; Michelotti, D.; Muraro, R.; Schenato, L. Identification of power distribution network topology via voltage correlation analysis. In Proceedings of the 52nd IEEE Conference on Decision and Control, Firenze, Italy, 10–13 December 2013; pp. 1659–1664. [Google Scholar] [CrossRef]
  72. Park, S.; Deka, D.; Chcrtkov, M. Exact Topology and Parameter Estimation in Distribution Grids with Minimal Observability. In Proceedings of the 2018 Power Systems Computation Conference (PSCC), Dublin, Ireland, 11–15 June 2018; pp. 1–6. [Google Scholar] [CrossRef] [Green Version]
  73. Deka, D.; Backhaus, S.; Chertkov, M. Structure Learning in Power Distribution Networks. IEEE Trans. Control. Netw. Syst. 2018, 5, 1061–1074. [Google Scholar] [CrossRef]
  74. Deka, D.; Backhaus, S.; Chertkov, M. Learning topology of the power distribution grid with and without missing data. In Proceedings of the 2016 European Control Conference (ECC), Aalborg, Denmark, 29 June–1 July 2016; pp. 313–320. [Google Scholar] [CrossRef] [Green Version]
  75. Guo, Y.; Yuan, Y.; Wang, Z. Distribution Grid Modeling Using Smart Meter Data. IEEE Trans. Power Syst. 2022, 37, 1995–2004. [Google Scholar] [CrossRef]
  76. Dorfler, F.; Bullo, F. Kron Reduction of Graphs with Applications to Electrical Networks. arXiv 2011, arXiv:1102.2950. [Google Scholar] [CrossRef] [Green Version]
  77. Li, T.; Sahu, A.K.; Talwalkar, A.; Smith, V. Federated learning: Challenges, methods, and future directions. IEEE Signal Process. Mag. 2020, 37, 50–60. [Google Scholar] [CrossRef]
  78. McMahan, H.B.; Moore, E.; Ramage, D.; Hampson, S.; Agüera y Arcas, B. Communication-Efficient Learning of Deep Networks from Decentralized Data. arXiv 2023, arXiv:1602.05629. [Google Scholar]
  79. Liu, H.; Wu, W. Federated Reinforcement Learning for Decentralized Voltage Control in Distribution Networks. IEEE Trans. Smart Grid 2022, 13, 3840–3843. [Google Scholar] [CrossRef]
  80. Li, Y.; Wei, X.; Li, Y.; Dong, Z.; Shahidehpour, M. Detection of False Data Injection Attacks in Smart Grid: A Secure Federated Deep Learning Approach. IEEE Trans. Smart Grid 2022, 13, 4862–4872. [Google Scholar] [CrossRef]
  81. Pathak, M.; Rane, S.; Sun, W.; Raj, B. Privacy preserving probabilistic inference with Hidden Markov Models. In Proceedings of the 2011 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Prague, Czech Republic, 22–27 May 2011; pp. 5868–5871. [Google Scholar] [CrossRef] [Green Version]
  82. Xie, P.; Bilenko, M.; Finley, T.; Gilad-Bachrach, R.; Lauter, K.; Naehrig, M. Crypto-nets: Neural networks over encrypted data. arXiv 2014, arXiv:1412.6181. [Google Scholar]
  83. Machlev, R.; Heistrene, L.; Perl, M.; Levy, K.Y.; Belikov, J.; Mannor, S.; Levron, Y. Explainable Artificial Intelligence (XAI) techniques for energy and power systems: Review, challenges and opportunities. Energy AI 2022, 9, 100169. [Google Scholar] [CrossRef]
  84. Lundberg, S.; Lee, S.I. A Unified Approach to Interpreting Model Predictions. arXiv 2017, arXiv:1705.07874. [Google Scholar]
  85. Zhang, K.; Zhang, J.; Xu, P.D.; Gao, T.; Gao, D.W. Explainable AI in Deep Reinforcement Learning Models for Power System Emergency Control. IEEE Trans. Comput. Soc. Syst. 2022, 9, 419–427. [Google Scholar] [CrossRef]
  86. Wali, S.; Khan, I. Explainable Signature-based Machine Learning Approach for Identification of Faults in Grid-Connected Photovoltaic Systems. arXiv 2021, arXiv:2112.14842. [Google Scholar]
  87. Xu, C.; Liao, Z.; Li, C.; Zhou, X.; Xie, R. Review on Interpretable Machine Learning in Smart Grid. Energies 2022, 15, 4427. [Google Scholar] [CrossRef]
  88. Machlev, R.; Malka, A.; Perl, M.; Levron, Y.; Belikov, J. Explaining the Decisions of Deep Learning Models for Load Disaggregation (NILM) Based on XAI. In Proceedings of the 2022 IEEE Power & Energy Society General Meeting (PESGM), Denver, CO, USA, 17–21 July 2022; pp. 1–5. [Google Scholar] [CrossRef]
  89. Murray, D.; Stankovic, L.; Stankovic, V. Explainable NILM Networks. In Proceedings of the 5th International Workshop on Non Intrusive Load Monitoring (NILM), Virtual Event, 18 November 2020. [Google Scholar] [CrossRef]
  90. Houidi, S.; Fourer, D.; Auger, F. On the Use of Concentrated Time–Frequency Representations as Input to a Deep Convolutional Neural Network: Application to Non Intrusive Load Monitoring. Entropy 2020, 22, 911. [Google Scholar] [CrossRef]
  91. Samek, W.; Montavon, G.; Lapuschkin, S.; Anders, C.J.; Müller, K.R. Explaining Deep Neural Networks and Beyond: A Review of Methods and Applications. Proc. IEEE 2021, 109, 247–278. [Google Scholar] [CrossRef]
  92. Al Dakheel, J.; Del Pero, C.; Aste, N.; Leonforte, F. Smart buildings features and key performance indicators: A review. Sustain. Cities Soc. 2020, 61, 102328. [Google Scholar] [CrossRef]
  93. Yu, L.; Qin, S.; Zhang, M.; Shen, C.; Jiang, T.; Guan, X. A Review of Deep Reinforcement Learning for Smart Building Energy Management. IEEE Internet Things J. 2021, 8, 12046–12063. [Google Scholar] [CrossRef]
  94. Aguilar, J.; Garces-Jimenez, A.; R-Moreno, M.D.; García, R. A systematic literature review on the use of artificial intelligence in energy self-management in smart buildings. Renew. Sustain. Energy Rev. 2021, 151, 111530. [Google Scholar] [CrossRef]
  95. Selvaraj, R.; Kuthadi, V.M.; Baskar, S. Smart building energy management and monitoring system based on artificial intelligence in smart city. Sustain. Energy Technol. Assess. 2023, 56, 103090. [Google Scholar] [CrossRef]
  96. U.S. DOE, Office of Energy Efficiency and Renewable Energy. Maps and Data—Electric Vehicle Registrations by State. Available online: https://afdc.energy.gov/data (accessed on 22 April 2023).
  97. Sun, X.; Qiu, J. A Customized Voltage Control Strategy for Electric Vehicles in Distribution Networks With Reinforcement Learning Method. IEEE Trans. Ind. Inform. 2021, 17, 6852–6863. [Google Scholar] [CrossRef]
Applsci 13 06937 g001 550
Figure 1. CPDSs consist of cyber and physical layers and interconnect architectures in a physical distribution system environment by utilizing communication, control, and computing resources through two-way information and data stream, as well as algorithms.
Figure 1. CPDSs consist of cyber and physical layers and interconnect architectures in a physical distribution system environment by utilizing communication, control, and computing resources through two-way information and data stream, as well as algorithms.
Applsci 13 06937 g001
Applsci 13 06937 g002 550
Figure 2. Smart grid of today and tomorrow, especially distribution systems, tightly interacts with residential and commercial buildings and transportation electrification by combining energy efficiency and demand flexibility in clean and flexible energy resources such as HVAC systems and EVs.
Figure 2. Smart grid of today and tomorrow, especially distribution systems, tightly interacts with residential and commercial buildings and transportation electrification by combining energy efficiency and demand flexibility in clean and flexible energy resources such as HVAC systems and EVs.
Applsci 13 06937 g002
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Chung, S.; Zhang, Y. Artificial Intelligence Applications in Electric Distribution Systems: Post-Pandemic Progress and Prospect. Appl. Sci. 2023, 13, 6937. https://doi.org/10.3390/app13126937

AMA Style

Chung S, Zhang Y. Artificial Intelligence Applications in Electric Distribution Systems: Post-Pandemic Progress and Prospect. Applied Sciences. 2023; 13(12):6937. https://doi.org/10.3390/app13126937

Chicago/Turabian Style

Chung, Sungjoo, and Ying Zhang. 2023. "Artificial Intelligence Applications in Electric Distribution Systems: Post-Pandemic Progress and Prospect" Applied Sciences 13, no. 12: 6937. https://doi.org/10.3390/app13126937

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop