Enhanced Electricity Fraud Detection in Smart Grids Using Clustering and Super Capacitor Control System

Since the world's modern civilization depends on electricity, the demand is growing ever more for power systems that are efficient, secure and resilient. Smart grids (SGs) are technological innovations that transform the way in which electrical networks are operated by leveraging advanced sensing, computing and communications to enhance the way in which the grid is operated. This level of evolution helps to address issues like energy efficiency, system reliability and reduction of technical and non-technical losses (NTLs) [1, 2]. The secret of SGs is an ‎advanced metering infrastructure (AMI) that allows energy ‎usage to be monitored and controlled in real-time by the intelligence of a smart meters (SMs) deployment. Monitoring devices can provide real-time feedback on power use all the way to the household level, and enable utilities to better predict energy loads and facilitate dynamic pricing and responsive demand management strategies that gently nudge consumers to adjust their own demand during times of peak load [3, 4]. Such capabilities are important for the supply and demand matching, and improving efficiency in operations and integrating renewable energy resources into the grid. Now, the problem is that ‎SGs are confronted with a growing number of security‎ challenges and threats to the integrity of networks. In particular, NTLs and energy theft are two of the biggest threats to the economic health of utilities and electric grids [5-7].

Energy theft, which by definition involves unauthorized use of electricity by tampering or bypassing the meter or by cyber manipulation, results in substantial-economic losses all over the world [8]. Energy theft has many causes ranging from high electricity tariffs to social and economic pressures. Equally important, new means of attack to AMI networks and smart meters have been opened up by digital technologies [9, 10]. Because detection and prevention of energy theft are already very important characteristics, it is necessary now for subtle methods in the combination of machine learning (ML) methods and artificial intelligence (AI). In this context, recent papers have found that ensemble learning models from the domain knowledge embedded in the premises or from heterogeneous data sources outside the house are effective tools in identifying anomalous trends that may suggest the occurrence of energy theft [11, 12]. Such models can be used to analyze large amounts of SM data, including consumption patterns, time-based power usage data, and behavioral analysis. Still, unsupervised learning algorithms such as principal component analysis (PCA) can be added to the mix to help detect terrain events without prior labeled cases of fraud or intrusion [13, 14].

Traditionally, electricity theft has only been detected through human intervention by detecting faults in the meter devices, periodic manual examination of the electricity meters and wires, or by receiving information from users of suspected electricity theft. However, because of the inherent complexity of power infrastructure, it has been difficult for human-centric methodologies to detect instances of electricity theft [15-17]. Therefore, strong and effective investigations with machine learning algorithms or energy theft detection systems or on-site inspections are required to act as deterrent and to give the authorization for dishonest actions, such as the study [18]. Many approaches to identify power use fraudulent schemes have been studied. Most of them are based on supervised machine-learning techniques, known as clustering-classification [19-21]. Addressing these issues is crucial for the sustainable and secure use of SG. By leveraging the distinctive characteristics of SGs, they become indispensable as we push forward in the quest for enhancements in the yield and reliability of electric grids. However, the same properties of SGs at the same time create unprecedented security problems. Finally, the above challenges can be overcome with a mix of cutting-edge ML methods, strong cybersecurity countermeasures, and transformative regulatory models.

Hence, considering the aforementioned issues, we propose a supercapacitor-based control strategy to mitigate the frequency instability issues among microgrid clusters for safe and reliable operation of SG. In this paper, we integrate Quantum Key Distribution (QKD) with Rolling Optimization Strategy (ROS), which can effectively reduce the computational burden of SG communication without affecting the fluctuation of MDs, and ensure the security and computational efficiency of the system. We attempt to realize capable performance of the electron fraud detection through the application of Extreme Gradient Boosting (EGB) and the Coati Optimization Algorithm (COA) method, which maximize the classification accuracy and reduce the false alarm rate. Besides, to achieve the trade-off between consumer trust and confidentiality, the proposed framework uses privacy-workability trade-off mechanism for SMs to achieve the requirement of optimal balance between effective tracking and consumer privacy protection in the SG system. In this paper, we examine some of the recent methods on electricity fraud detection in SGs based on ML algorithms and how secure AMI implementation can help reduce energy theft. While methods for detecting electricity theft are advancing, they often fail to address several critical, interconnected challenges in a real-world smart grid.

First, existing models fail to consider the physical stability of the grid. They do not have mechanisms for frequency maintenance of microgrid clusters and are very sensitive to data outlier. More, because of data sharing in these models, the computational overhead may add complexity that reduces the effectiveness of frequency regulation attempts. Second, detection models themselves have poor classification accuracy, particularly with the rare and skewed nature of the data with energy theft. This leads to high false alarm rates, which are operationally expensive. Finally, privacy is another major, and often unaddressed, challenge. Many of the existing solutions require access to granular data, which could violate consumer privacy and raise ethical and regulatory concerns. This paper presents a unified solution that simultaneously addresses these gaps in stability, accuracy and privacy. In this work, we propose an advanced and effective framework for electricity theft and fraud detection in smart grids (SGs) based on cutting edge machine learning techniques. To meet this challenge, we propose an efficient algorithm to reduce the effect of theft detection interruptions on the frequency stabilization of microgrid clusters to provide continuous supply to the load at fault. Robust techniques are incorporated into the system to improve detection accuracy by lessening the effects of data variations.

Moreover, the optimized procedures for information sharing are also implemented to reduce redundancies while transmitting and processing the data, thus improving the overall performance of the system. Furthermore, it incorporates the state-of-the-art techniques for handling the rare occurrence of electricity theft, in order to reduce the number of false alarms and enhance the classification performance. Moreover, to ensure consumer privacy and security, the SG infrastructure achieves a trade-off between fine-grained electricity consumption measurement and privacy-preserving techniques on system functionality availability, and establishes consumer trust for consumer-side services in SG infrastructure. The key contributions of this research are as follows:

• Microgrid Clustering Based Topology for Higher Stability: This work proposes a clustering based on distance, load and energy capacity metrics, which provides autonomous integration and self-tuning primary control in a MESH topology from microgrid and other sub-systems. Thus, improves communication, stability, reliability and scalability of SG operation.
• Frequency Control of Microgrid Clusters Based on Supercapacitors: Supercapacitor-based energy storage mechanism is proposed to solve the problem of frequency instability of microgrid cluster, which can ensure the reliability and stability of the operation of SG.
• QKD-ROS: Secure Communication with QKD and ROS: The integration of QKD and ROS greatly improves the security of communication, reduces the computational complexity, and reduces the volatility of SG networks.
• Next Level Defection Detection using EGB-COA: The proposed framework uses the EGB combined with the COA for electricity theft classification with higher accuracy. Moreover, we use GANs to reduce the number of false positives and increase the detection accuracy.
• Privacy-Preserving Smart Metering: Trade-off strategies are available to preserve privacy while enabling and protecting against excessive monitoring by SMs. This not only ensures privacy but also maintains the transparency and confidentiality of the SG.

The rest of this paper is organized as follows: Section 2 summarizes the survey of existing works, with a research gap. In Section 3, we present the research methodology for the proposed approach with related graphs, mathematical formulation, and pseudocode. Section 4 shows the experimental results and comparison of proposed and existing methods. Section 5 is the proposed work conclusion.

The present section aims to establish the proposed framework for SGs electricity fraud detection by selecting an algorithm/technique that can sustain an optimality between accuracy, efficiency, scalability, and robustness. This section explains the reasons for the clustering approach, machine learning specifications, control systems, and secure communication protocols adopted here. An overview of the proposed approach is shown in Figure 1, which consists of the following major components:

• Microgrid Clustering — Grouping microgrid based on key parameters for better operating stability.
• Secure Smart Grid Fluctuation — Security mechanisms to respond to the vibration and security of the system parameters.
• Electricity Theft and Fraud Detection — Most advanced ML techniques to determine fraudulent activities.
• Monitoring — Balancing fraud detection with consumer privacy through secure data monitoring strategies.

Figure 1. The structure of the suggested approach

3.1 Microgrid clustering

The resistance to electricity theft within SGs, initially, involves the strategic clustering of microgrids. This clustering process is led by metrics such as distance, load, energy availability, and capacity to generate cohesive units capable of independent operation. In alignment with the scalability goals of this framework, our approach utilizes an adaptive combination of k-means and GMM. This decision deliberately prioritizes the high computational efficiency required for real-time operation over the more complex, computationally expensive methods like DBSCAN, which, as noted in our literature review, do not scale as effectively for this application. Through this approach, the microgrid integrates autonomously, developing adaptive primary control among them. Once the microgrid is clustered a connection framework is established through MESH topology, this improves the whole communication within clusters. The decentralized controllers and agents within the framework strengthen the stability and fault tolerance while the interconnect. The MESH topology improves reliable communication and easy scalability for the new microgrid. MESH structure for communication from the inside and outside. Each microgrid creates links with all surrounding microgrids through an integrated power transmission and communication network according to the extensive interconnections between individual microgrids created by this MESH frame-work. It provides several channels of redundant communication. If one connection fails, there are many paths that can be taken. Microgrids and microgrid clusters provide for flexibility in data exchange and control, and they can establish connections with another microgrid or cluster inside the net-work. MESH topologies can help with load balancing by distributing data and communication traffic over many channels.

This might enhance the utilization of resources inside microgrid clusters. A grid can be conceptualized as a particular example of a micro-grid. The mesh’s dimensions, node count, permitted power ratings, and geographic reach.

After establishing the microgrid cluster, the next step is to overcome the frequency instability within these clusters. This is achieved through super capacitor control. By using the super capacitors, the system efficiently manages the frequency instability and improves the consistent and reliable operation of the microgrid cluster. This step is crucial for maintaining the overall stability of SG. The energy storage system for microgrid frequency stabilization using supercapacitors is considered in this study. To obtain the required current and voltage, supercapacitor modules can be combined in a series-parallel configuration. An impedance can be coupled in series with a voltage source that depends on the state of charge () in order to imitate a supercapacitor. The following equation could have been obtained from the supercapacitor’s generalized model:

$\mathbb{V}_o=\mathbb{V}_{i n}(\mathfrak{H})-\mathbb{I}_{s c} \mathbb{Z}$     (1)

where, $\mathbb{V}_o$ is output voltage, $\mathbb{V}_{i n}$ is internal voltage, $\mathbb{I}_{s c}$ is supercapacitor current, and $\mathbb{Z}$ is impedance.

The following formula offers a simple approach for calculating the $\mathfrak{H}$ of supercapacitors:

$\mathfrak{H}=\mathfrak{H}_{\text {int }}+\frac{1}{N m_C} \int \mathbb{I} d t$     (2)

Whenever the manufacturer's defined nominal capacity $N m_C$ and the starting stage of charge $\mathfrak{H}_{\text {int}}$ are the supercapacitor's respective values. As the quantity of loss to be accounted for in the model rises, so does the supercapacitor model's complexity. To exclude higher-order non-linearities in small-signal modeling, the transfer function that we utilize is as follows:

$\mathfrak{G}_{S C}(S)=\frac{\mathcal{K}_{S C}}{\mathcal{T}_{\mathbb{C} \mathfrak{H}}+1}$     (3)

$\mathcal{K}_{S C}$ and $\mathcal{T}_{\mathbb{C} \mathfrak{H}}$ represent the supercapacitor’s gain and time constant, respectively. With fractional-order supercapacitors, microgrid clusters’ consistency in frequency will be improved. Fractional-order computation is one branch of statistics that works with integrals and equations of non-integer order. Fractional order controllers are employed in many different techno- logical applications due to their higher efficiency compared to conventional integer-order controllers.

$F(\mathcal{T})=\mathfrak{G}_P e(\mathcal{T})+\int_{\mathcal{T}}^{\Lambda} \mathfrak{G}_F e(\mathcal{T})$     (4)

The error signal in this case is represented by $e(\mathcal{T})$, the proportional gain by $\mathfrak{G}_P$, the integral gain by $\mathfrak{G}_F$, and any real number greater than zero by $\Lambda$. The fractional-order controller might be represented below in a frequency plane by using the Laplace transformation:

$\mathbb{I}(s)=\mathfrak{G}_P+\frac{\mathfrak{G}_F}{s^{\Lambda}}$     (5)

3.2 Secure smart grid fluctuation

With the stability of the microgrid clustered certain attention shifts to securing SG communication to overcome the fluctuations and prevent unauthorized access. The QKD keys are employed to improve the authentication security while minimizing computational overhead. The security measures are accompanied by ROS to design to minimize the sensitive fluctuation within SG further improving the stability and reliability. The QKD-ROS communication model data flow shown in Figure 2.

Figure 2. QKD-ROS communication model data flow

As shown in Algorithm 1, the QKD-ROS model does not generate a new quantum key for every single data packet, which would be computationally prohibitive. Instead, the ROS manages the key's lifecycle, re-keying the channel based on a pre-defined rolling time window (T_w). This balances the high security of QKD with the low-latency, high-efficiency requirements of a real-time smart grid, effectively reducing the computational burden.

3.3 Electricity theft and fraud detection

The EGB algorithm was chosen for its high performance on structured data and its native capability of dealing with complex and non-linear relationships in electricity consumption patterns. To improve its prediction ability even more, we use the COA for hyperparameter optimization. COA is a recently developed metaheuristics algorithm, which has a good ratio of exploration and exploitation, and is able to prevent the EGB model from falling into local optima and achieve a more robust classification with higher accuracy. While a direct comparison to other optimizers like PSO or GA is outside the scope of this framework-focused study, COA was specifically chosen based on strong evidence from recent literature demonstrating its superior convergence and ability to escape local optima in complex, non-linear problems, making it an ideal candidate for tuning the EGB model [38]. After the grid is stable and secure the focus is on identifying electricity theft and fraud using EGB-COA for classification. EGB can identify the electricity theft instance with high accuracy, and COA can improve the classification performance while improving the overall accuracy.

Table 2. Proposed GAN architecture

* N-dim Feature Vector corresponds to the number of features used for classification

Table 3. GAN training hyperparameters

3.3.1 GAN-based data augmentation

Class imbalance is one of the most important challenges in fraud detection and is well known to be frustrating traditional classifiers. In our synthetic dataset fraudulent profiles only account for 5% of the overall consumers. In order to correct this imbalance, we employ Generative Adversarial Network (GAN) to oversample the minority (fraudulent) class. The GAN learns the underlying distribution of fake data and generates high fidelity synthetic data. By doing so it provides the EGB-COA classifier with a more balanced and robust data set which helps decrease the false alarm rate and increase the recall, results that we have validated. The GAN consists of two feed-forward neural networks, a Generator (G) and a Discriminator (D). The Generator is a generator that creates the synthetic data vectors from the random noise, while the Discriminator attempts to distinguish the genuine fraudulent samples from the synthetic samples. The detailed architecture is presented in Table 2 and the training parameters are shown in Table 3.

3.3.2 Privacy threat model

To formalize the privacy and security guarantees of our framework, we define the following threat model, which identifies potential adversaries, their capabilities, and the corresponding mitigations provided by our system. Table 4 illustrates the privacy and security threat model.

Table 4. Privacy and security threat model

3.4 Monitoring

To enhance trust and confidentiality in the system, the monitoring activities within the SG must strike a balance between functionality and privacy for the consumer. This is accomplished by using SMs that come with privacy functionality trade-off mechanisms. As shown in Figure 3, the operation of the strategy in the proposed model is as follows. The distribution-level substations (Sub) that serve families with electricity are included in the model, as are customers, "energy suppliers", "network operators", and "third parties". The proposed system's SM updates its measurements via a private platform which can be PC or smartphone. Rather than having a direct dialogue with the energy provider. Private platform is equipped with basic computing and storage capabilities for saving power usage and bill payment. Via TOU channel, the dynamic TOU tariff is enabled. In conventional smart metering, in order to obtain the TOU bills, the SM reports the energy consumption of each charging station. The more charging points a utility installs, the more personal information the utility can collect about an individual, which increases the risk of privacy violations. The data is sent back the other way via our TOU billing channel. Algorithm 2 shows the formula used to calculate the TOU billing. Every half hour the TOU pricing is sent from the ES to the SM. The SM combines current TOU pricing and the last 30 minutes of energy use. Bills computation. On the last day of every month, the total TOU bill amount in rupees is calculated and sent to the ES who then sends the bill to the customers.

These procedures provided effective monitoring while protecting the privacy of the monitored individuals, which in turn enhanced the consumer trust and confidence in the SG system.

The privacy assuring power of this model, as explained in Algorithm 2, is given by the fact that it uses local data aggregation. Conventional smart metering devices send high-granularity (e.g., 15 minutes or 30 minutes) consumption signals directly to the utility, posing a serious privacy threat. Such fine-grained data can be used to do non-intrusive load monitoring (NILM) and behavioral inference, which can show when occupants are at home, what appliances they are using, and their everyday lifestyle trends. This threat is clearly addressed in our framework. Not only by processing all high-granularity E_h and TOU_h data locally on the user “private platform” and only sending final aggregated monthly totals ($\mathbb{M}_E$ and $\mathbb{M}_{\mathfrak{B}}$) will our system ensure utility provider is never in a position to infer such inference attacks.

Figure 3. Smart metering system