An Intelligent Framework for Multiscale Detection of Power System Events Using Hilbert–Huang Decomposition and Neural Classifiers

Abstract

This article proposes a multiscale classification framework for detecting voltage disturbances in electrical distribution systems using artificial neural networks (ANNs) combined with the Hilbert–Huang transform (HHT). The framework targets four core power quality (PQ) events defined in the IEEE 1159-2019 standard: normal operation and voltage sag, swell, and interruption. Unlike traditional methods that operate on a fixed disturbance duration, our approach incorporates multiple time scales (0.2 s, 0.4 s, and 0.8 s) to improve detection robustness across varied event lengths, a critical factor in real-world scenarios where disturbance durations are unpredictable. Features are extracted using empirical mode decomposition (EMD) and Hilbert spectral analysis, enabling accurate representation of the signals’ non-stationary and nonlinear characteristics. The ANN is trained using statistical descriptors derived from the first two intrinsic mode functions (IMFs), capturing both amplitude and frequency content. The method was validated in MATLAB on the IEEE 33-bus radial distribution test system using simulated disturbances. The proposed model achieved a classification accuracy of 94.09% and demonstrated consistent performance across all time windows, supporting its suitability for real-time monitoring in smart distribution networks. This study contributes a scalable and adaptable solution for automated PQ event classification under variable conditions.

1. Introduction

The stability of electrical distribution networks is one of the most important parts of maintenance engineering structures and requires the implementation of proper monitoring strategies. As emphasized in the existing literature, a sustainable and stable supply of electrical energy is of crucial importance for modern society [1].

At the same time, the existence of backup systems does not absolve distribution networks from being subjected to a wide range of possible threats that may compromise their performance. These problems, which can be subdivided into internal and external categories, include lightning, storms, operational failures, generator start or stop cycles, and even simple load switching. All these processes constitute a considerable challenge in regard to the operation and dependability of the electrical system [2,3].

Specifically, electrical disturbances comprise the class of phenomena that disrupt the energy equilibrium of a certain system, with perturbations of voltage being the most common manifestation of this category. More specifically, these are voltage sags and swells and complete loss of power [3].

Such processes may result from pulse transients, changes in frequency, equipment, or out-of-order devices and may affect system performance from increased wear of electric parts to irreversible and prolonged suspension of intended system operations [4].

There exists a problem in the effective monitoring structures in the context of interdependent electrical grids due to their intricate nature and the intelligent technologies associated with them. In other cases, more advanced devices such as artificial neural networks are employed for detecting and analyzing power system disturbances [4].

Neural networks have been proven to aid in problem solving due to their ability to complexly identify patterns and modify themselves according to continuously changing settings [5].

This paper proposes a methodological approach in the form of disturbance detection and classification at the level of the artificial neural network application paired with Hilbert–Huang transform-based power signal processing according to IEEE 1159 [5,6,7].

This methodology in the IEEE 1159 standard allows for monitoring the power supply and identifying phenomena that can impact the quality of electrical power supplied, hence categorizing them into electrical events. The proposed methodology allows for unrivaled depth analysis and classification in addition to disturbance detection, which further strengthens the existing electrical frameworks [5,6,7].

In this case, neural networks enhance detection capabilities and provide great value through effective data collection and analysis throughout large datasets. The Hilbert–Huang transform is capable of detailed spectral analysis of non-stationary signals decomposition enabling identifying characteristic patterns of disturbances [6,7].

This proposal stands out as it attempts to solve one of the almost perennial problems lingering in modern electrical systems: the early detection and accurate classification of disturbances. Precise detection allows for remedial action to be taken well ahead of the disturbances degrading grid performance, thereby minimizing the impact to users [6,7].

Accurate event classification enhances understanding of the phenomenon which can help in designing specific mitigation strategies. This proposal aims to provide groundwork for subsequent advanced research which integrates highly sophisticated techniques for dynamic and flexible monitoring of electrical systems. In this regard, the integration of neural networks with the Hilbert–Huang transform contributes significantly to the development of electrical networks which are more robust and effective [6,7,8].

1.1. Literature Review

The challenges of fault detection in active distribution networks (ADNs) have increased significantly due to the growing integration of inverter-based distributed generation (IBDG) [9], especially from solar and wind sources. Conventional protections designed for systems with unidirectional power flows and synchronous generators are insufficient in this new operating context. It is estimated that approximately 70% of faults in distribution networks are single-phase ground faults, which are often difficult to detect using traditional protection devices due to their low current amplitudes [10].

Three key gaps have been identified in the literature: the absence of clear guidelines for protection in ADNs, the omission of Fault Ride Through (FRT) requirements, and the limited consideration of the dynamic response of inverters. Faced with this problem, modern techniques such as machine learning and wavelet transform have been explored, highlighting the need for adaptive protection schemes that adjust to the dynamic conditions of the networks. Finally, future trends that point to the development of faster, more accurate, and intelligent protection systems for evolving electrical grids are presented.

On the other hand, protecting active distribution networks that integrate microgrids with distributed generation also faces significant limitations when applying traditional schemes [11]. Adaptive protection techniques have been proposed that improve both selectivity and response speed to faults. The results obtained through simulations on realistic grid models demonstrate that these methods allow for more effective protection in highly dynamic configurations. The study emphasizes the need for flexible protection algorithms and highlights the importance of coordination between microgrids and the central system. Validations were carried out in simulation environments using MATLAB/Simulink R2024a.

Another study investigated the penetration limits of solar photovoltaic generation in low-voltage networks, emphasizing the voltage variations induced by such integration [12]. A methodology was developed to calculate the technical penetration limit considering load profiles, solar irradiance, and grid topology. The results show that photovoltaic plants’ location and size significantly influence voltage stability. Based on these findings, safe operational limits were proposed to avoid swell events. The study, implemented through simulations in DIgSILENT PowerFactory, also suggests incorporating energy storage systems to mitigate adverse effects. This research represents a key contribution to planning smart grids with high penetration of renewable generation.

Protection schemes in distribution power grids with high penetration of distributed generation from renewable sources have received increasing attention in the technical literature. In [13], the main challenges associated with these types of networks are identified, and current solutions are classified into three categories: artificial intelligence-based approaches, conventional techniques, and hybrid methods. The study results confirm that traditional protection schemes are unsuitable for environments with high renewable penetration due to the variability and unpredictability of these sources. Consequently, a modern, flexible protection architecture is proposed, capable of adapting to the dynamic conditions of the system. In addition, several international case studies are discussed and regulatory and interoperability considerations are addressed, concluding with key recommendations for future research and technological developments.

In the context of microgrids, one of the most significant challenges is ensuring adequate protection in interconnected and islanded operation modes. The study presented in [14] analyzes various protection strategies and evaluates their performance under both operating conditions. It is demonstrated that adaptive schemes with active communication offer more excellent reliability and responsiveness to internal faults. The article provides detailed models and experimental validations performed in real microgrids, highlighting the fundamental role of power control and the coordination between different protection levels. The analysis is based on experiences gathered in European systems, emphasizing the need to establish specific standards for designing and operating protections in microgrids.

For its part, the work by [15] offers a comprehensive review of fault detection techniques and protection schemes applicable to networks with distributed generation, addressing technical challenges such as the bi-directionality of power flow and the high variability of renewable sources. A classification of existing methods is proposed based on accuracy, computational complexity, and scalability criteria. The results reveal that artificial intelligence-based techniques, combined with phasor synchronization systems (PMU), offer superior performance in speed and reliability. Based on the analysis of 20 case studies, a methodological framework is proposed for the comparative evaluation of protection schemes, also integrating considerations for their interoperability with SCADA platforms. The study concludes with suggestions for improving the robustness and reliability of these solutions in real-world environments.

Finally, the classic study by [16] addresses protection problems in distribution networks with dispersed generation, highlighting how the presence of distributed sources significantly alters the coordination of protection devices. Fault currents are modified by the effect of DG units, which can reduce the effectiveness of conventional schemes. Proposed solutions include directional protection, dynamic parameter adjustment systems, and adaptive protection strategies. The analysis was based on simulations of real medium-voltage networks, where recurring conflicts in relay coordination were also identified. This work has served as the basis for subsequent investigations focused on protecting microgrids and hybrid systems.

Several recent studies have explored advanced artificial intelligence techniques for classifying power quality events. In [17], a review focuses on machine learning and deep learning algorithms applied to electrical signals. Approaches based on convolutional neural networks (CNNs), support vector machines (SVMs), and hybrid methods are highlighted, demonstrating accuracies above 95% in identifying events such as interruptions, transients, and harmonic distortions. The study used real databases and synthetic datasets generated for training and validation. The robustness of the models to noise and data variability was also analyzed, identifying the need for large volumes of labeled data as the main challenge. Strategies such as federated learning and knowledge transfer (transfer learning) are proposed as a future projection to mitigate these limitations.

In an innovative line of work, the work in [18] employs topological tools, specifically simplicial complexes and visibility graphs, for fault classification in distribution systems. These techniques were applied to multivariate time series under noisy conditions, achieving an accuracy of 98.7% in classifying symmetric and asymmetric faults. The model proved highly robust to nonlinear distortions and was validated with simulated data and real records. Due to its low computational requirements, its application in real-time automated protection systems is suggested.

On the other hand, ref. [19] proposes a method for early prediction of faults in electric motor bearings (CMOR) based on variational modal decomposition (VMD) combined with the autocorrelation function. This technique was optimized to improve extracting relevant signal features from critical systems. The methodology demonstrated high accuracy in predicting imminent faults, surpassing the performance of classical techniques such as the fast Fourier transform (FFT) and wavelet. Its non-invasive nature and compatibility with industrial monitoring platforms make it an effective solution for predictive maintenance schemes. Its applicability to the diagnosis of other rotating machines is also discussed.

In the study by [20], a high-impedance fault detection method is developed using a combination of the Hilbert and Fourier transforms. These types of faults, difficult to detect using conventional techniques due to their low fault currents, were addressed through time-frequency domain characteristic analysis. The proposed model achieved detection rates above 90% in simulations of typical distribution networks. Its high sensitivity to minor signal variations makes it a viable option for real-time applications, especially in urban environments. Furthermore, the study suggests its integration with intelligent sensors to improve electrical safety.

Finally, the work of [21] presents a fault diagnosis system in electric drives based on a Deep Echo State Network (DESN)-type neural network with scaled activation functions. The approach was validated with a database that includes multiple types of faults, such as insulation faults, phase imbalance, and rotor bar breakage. The model achieved an accuracy of 98.5% with an inference time of less than 15 ms, standing out for its speed and efficiency compared to traditional models such as RNN and LSTM. Furthermore, it demonstrated robustness against noise and uncertainty and is considered highly efficient for implementation in embedded monitoring and protection systems.

An innovative approach to fault detection in transmission lines is presented in [22], which uses 2D convolutional neural networks (2D-CNNs) applied to scalogram images generated using the continuous wavelet transform. The study considered five types of faults, including LG, LL, and LLG, and generated 3600 images for training and testing. The four-layer CNN model achieved a classification accuracy of 99.46%, identifying the location and type of fault in less than 0.1 s. The system demonstrated viability for real-time applications integrated with SCADA platforms. It maintained its performance under conditions with a signal-to-noise ratio of 20 dB, demonstrating robustness in noisy environments.

In [23], the detection of intermittent ground faults in compensated and isolated medium-voltage networks is addressed using data-driven models. Unlike conventional methods that rely on zero-sequence current, this approach combines voltage and current signals with machine learning algorithms, specifically Random Forest and XGBoost. Tests conducted with real event records yielded detection rates above 96% for faults lasting less than 200 ms. The model’s adaptability to smart grids positions it as a promising alternative for improving operational reliability.

An extensive review of fault diagnosis techniques in distribution networks is presented in [24], which analyzes more than 110 recent publications on artificial intelligence and signal processing. The study ranges from classical methods such as SVM, ANN, and CNN to advanced transforms such as Hilbert–Huang and wavelet. Average accuracy of 97% in controlled environments and the effective use of differential current signals and spectral analysis are highlighted. In addition, hybrid approaches are recommended to improve multiple fault detection, and significant gaps in the availability of labeled data are identified, especially for non-conventional faults.

In [11], an adaptive protection technique is proposed for active distribution networks with microgrids integrating multiple distributed generation (DER) technologies. Using PSCAD simulations, the system’s performance was evaluated against internal and external faults. The results reveal significant improvements in selectivity and detection speed, reducing the response time to less than 20 ms compared to the 60 ms observed in conventional schemes. The methodology is based on adaptive logic for microgrid status, complemented by continuous current and voltage monitoring, making it especially suitable for urban hybrid networks.

Regarding fault diagnosis in electrical machines, ref. [25] presents a technique for analyzing parasitic flux in three-phase induction motor stators combined with machine learning algorithms. Decision trees, Random Forests, and k-nearest neighbors (kNN) were used on signals captured by flow sensors. The Random Forest model achieved the highest accuracy, with 99.06%, demonstrating high robustness across different load levels and fault severities, making it an effective tool for industrial applications.

Likewise, in [26], a differential protection scheme for microgrids was developed using a hybrid SVM-CNN architecture. The method was tested on lumped systems with different load and generation configurations, achieving 99.4% accuracy in fault classification, with a mean response time of 25 ms. The approach demonstrated high adaptability to topological changes and transient conditions, making it especially useful for dynamic environments.

Finally, the study by [27] proposes an electrical fault detection system based on a comparison of algorithms such as SVM, kNN, and artificial neural networks. Preprocessed from a simulated network, current and voltage signals enabled the SVM model to achieve 98.3% accuracy in classifying three-phase, two-phase, and single-phase faults. This system’s ability to effectively differentiate between various fault types makes it a competitive solution for advanced protection systems.

In [28], the Hilbert–Huang transform (HHT) is used for fault identification and classification in alternating current transmission lines. The study used simulated data and real records, applying an artificial neural network for processing. The system was able to identify faults with 98.5% accuracy in times under 0.1 s, demonstrating high sensitivity to high-impedance faults and events located near network nodes.

A more advanced strategy is presented in [29], which combines variational mode decomposition (VMD), the HHT, and convolutional neural networks (CNNs) for autonomous fault diagnosis in permanent magnet synchronous motors. Using current signals, the system identified faults in the rotor and stator with high accuracy. The CNN model achieved 99.7% accuracy on the test set, demonstrating real-time detection capabilities even under noisy and variable load conditions.

Meanwhile, [30] presents an approach to fault detection and classification in photovoltaic networks, combining an optimized HHT with ensemble learning algorithms such as Random Forest and Boosting. The system achieved 99.2% accuracy in detecting faults such as arcing, short circuits, and disconnections. The model was validated using MATLAB simulations and experimental data, including scenarios with varying solar irradiance, demonstrating its applicability in real-world environments.

The review in [31] performs analysis using the HHT in fault diagnosis in mechatronic devices, including motors, sensors, and nonlinear actuators. It is concluded that combining HHT with machine learning techniques improves diagnostic accuracy and early fault detection. According to the collected studies, the average accuracy exceeds 97%, and its future implementation in embedded systems with autonomous diagnostic capabilities is recommended.

The work by [32] proposes an early fault detection system for BLDC motor controllers using current signals processed with the HHT. The system can identify subtle variations associated with incipient faults with an accuracy of 98.9%, maintaining its performance even under variable load and dynamic speed conditions. Tests were conducted in a laboratory environment, demonstrating the approach’s effectiveness for experimental applications and its potential in industrial settings.

In [33], an HHT-based approach is developed to diagnose dynamic eccentricity faults in dual-rotor axial flux generators. Signals acquired under different load and speed conditions were used, and an accuracy of 98.1% was achieved using an artificial neural network. The methodology was validated by simulation and on a physical test bench, demonstrating its feasibility for real-world operating environments.

Finally, ref. [34] proposes a method for detecting inter-turn short circuits in DFIG generators using external leakage flux sensors and combined VMD-HHT analysis. Multiple load scenarios and fault levels were simulated, achieving efficient decomposition of non-stationary signals and identifying characteristic spectral patterns of faults. The detection rate exceeded 96%, even in signals with low signal-to-noise ratios. The study also evaluated different sensor locations, thus optimizing the system sensitivity. Experimental validation confirmed the applicability of the approach in real power systems.

1.2. Organization

Section 1 highlights the goals of this paper and presents a review of the literature relevant to the theme of this research.

Section 2 outlines the theoretical foundation of electrical disturbances in distribution systems and the Hilbert–Huang transform (HHT) application as a signal processing technique for analyzing nonlinear, non-stationary data. This section also describes the methodology implemented in this study, which involves simulating a 33-bus IEEE distribution system and controlling the injection of voltage sag, swell, and interruption events at specific nodes, with predefined durations of 0.2 s, 0.4 s, and 0.8 s.

Section 3 discusses the development of an artificial neural network model using MATLAB, where features from voltage signals were processed with the HHT, and the proposed algorithm was able to detect and classify electric disturbances in multiple test cases executed in the simulated distribution network.

Section 4 summarizes the key contributions of this work, highlighting the effectiveness of the proposed methodology for event detection and classification, its robustness under simulated conditions, and its potential applicability in stability analysis and monitoring of modern power distribution systems.

2. Methodology

Before developing and addressing the problem posed, it is essential to review certain key concepts related to power systems, the most common electrical disturbances, algorithms based on artificial neural networks, and the application of the Hilbert–Huang transform (HHT), an effective tool for the decomposition and analysis of nonlinear and non-stationary signals. This section briefly presents the essential theoretical principles underlying the approach used in this work.

The study is based on the IEEE 33-bus distribution system simulation, a standard network widely used for power system analysis in academic and technical settings. In this model, sag, swell, and interruption disturbances were introduced in a controlled manner at various system nodes to evaluate the performance of the proposed algorithm under different operating scenarios. To this end, three specific disturbance duration intervals were established, 0.2 s, 0.4 s, and 0.8 s, which allows for the analysis of the dynamic behavior of the signal as a function of the disturbance time.

The proposed methodology is based on artificial neural networks trained in MATLAB/Simulink R2024a, developed by MathWorks Inc. (Natick, MA, USA). These networks process voltage signals previously processed using the Hilbert–Huang transform. This processing facilitates extracting relevant features that allow for accurate event classification. The model was trained using the disturbances generated in the simulation, evaluating its ability to correctly identify and classify the type of event present in the distribution system.

The following sections detail the methodological procedure developed, which includes data preparation, the configuration of the neural network used, and the implementation of the disturbances, along with their respective evaluation metrics. Finally, the results are presented, and the performance of the proposed system under different operating scenarios is analyzed.

2.1. Electrical Power Systems

According to IEEE standards, the topology of electric systems is divided into three main processes, generation, transmission, and distribution, with the aim of providing a reliable electricity supply that is capable of meeting energy demand 24/7 and operates efficiently [1].

Power flow is one of the fundamental factors in the functioning of these systems and it enables the study of the movement of electricity through the system. This defines the performance of the grid and highlights critical issues such as the possibility of overloads on power transmission lines, voltage drop forecasting, and optimal generation distribution, and then all of this is needed in order to improve the functions of the electric system [2]. In addition, properly managed power flow ensures the seamless incorporation of renewables, such as solar and wind, and allows the system to curb the changing energy demands of electric grids that are modern and sustainable.

The electric system comprises a set of elements and subsystems that operate in collaboration for the purpose of generating, transmitting, and distributing electrical energy in a safe and efficient manner [35].

The initial step in the electrical system is Electric Power Generation. It is strategically important to optimally utilize the available resources at the generation units. In addition, this step takes care of frequency and voltage control, which, like other parameters, is also critical for electric grid operation [36,37,38].

From the generation center, electric power needs to be delivered, or more accurately, transmitted over long distances to regions of consumption. With regards to this step, the system should be able to provide enough energy at precise voltage levels. IEEE sets standards concerning the insulation values, load capacity, and maximum and minimum distances between lines to obviate interference and overloads. Along with the materials used in construction and their scheduled maintenance, these factors are important to ensure energy losses and failures are at an acceptable minimum [36,37,38].

At the distribution stage, electricity delivery to end users should be sustained, reliable, and continuous. Like other stages, the aim is to operate at low voltage levels specially tailored for domestic, commercial, or industrial use while preventing service interruptions, sudden voltage and frequency changes, or other phenomena that compromise power quality [36,37,38].

The system structure includes components that interact in a complex way to ensure the power supply, playing a fundamental role in the system design [36,37,38]. Transformers play a crucial role in transmission systems, increasing the voltage level to minimize energy losses during long-distance transmission. At the distribution stage, they reduce the voltage to safe and adequate levels for delivery to the end user. Meanwhile, protection devices, such as relays, fuses, and circuit breakers, are activated in the event of abnormal conditions, including faults, short circuits, or overloads, thereby ensuring the safety and reliability of the system [38,39].

The distribution system is the final link in the electricity supply chain. It includes facilities like distribution transformers, power lines, switches, and protection devices which are critical to the uninterrupted and efficient delivery of energy. Of these elements, the transformer is particularly important in stepping down the voltage level from transmission to a level suitable for residential, commercial, or industrial use [38,39].

2.2. Power Quality Events According to IEEE 1159

Voltage disturbances in distribution networks are characterized by deviations in magnitude or continuity that impact power quality and operational stability. According to the IEEE 1159-2019 standard [3], these disturbances can arise from various system conditions and are categorized to facilitate monitoring and automated detection.

This study considers four event types relevant to real-time classification tasks using artificial neural networks (ANNs): normal, voltage sag, voltage swell, and interruption. These categories were selected due to their practical significance in power system operations and are formally defined in

Table 1. Classification of power quality events based on IEEE 1159 and ANN labels.

Label	Type	IEEE 1159 Description
1	Interruption	Voltage = 0 for more than 0.5 cycles
2	Normal	Nominal conditions, no disturbance
3	Sag	RMS voltage drops below 90% for 0.5 cycles to 1 min
4	Swell	RMS voltage exceeds 110% for 0.5 cycles to 1 min

, along with their assigned classification labels used throughout this work.

These event types cover the most critical RMS-based deviations in grid performance. The corresponding labels were embedded into the disturbance generation process to ensure consistency across training, validation, and test datasets. More complex disturbances, such as harmonics, flicker, or transients, are excluded from the current scope and reserved for future studies focused on high-frequency and waveform distortion effects.

2.3. Artificial Neural Network-Based Classification

Artificial neural networks (ANNs) are effective for classification tasks involving nonlinear, non-stationary signals like voltage disturbances [40,41]. This work employs a supervised ANN to identify and classify power quality events using features obtained from the Hilbert–Huang transform (HHT).

The input to the network consists of six features derived from the first two intrinsic mode functions (IMFs) of each voltage signal window. Specifically, the feature vector includes the following:

• The mean and standard deviation of the instantaneous amplitude of IMF1;
• The mean and standard deviation of the instantaneous amplitude of IMF2;
• The mean of the instantaneous frequency of IMF1;
• The mean of the instantaneous frequency of IMF2.

These features capture the signal’s energy content and spectral dynamics, providing a compact and discriminative representation for event classification. The ANN architecture consists of an input layer with 6 neurons, two hidden layers with 20 and 15 neurons, and a Softmax output layer with 4 neurons corresponding to the classes normal, sag, swell, and interruption. This topology is shown in Figure 1.

The number of neurons in the hidden layers was determined empirically via multiple trials to maximize classification accuracy. However, no formal optimization method (e.g., grid search, Bayesian tuning, or genetic algorithms) was applied.

Each hidden layer employed the Rectified Linear Unit (ReLU) activation function to enhance nonlinearity and improve convergence speed, while the output layer used the Softmax function for multi-class classification.

The selection of this network structure was based on iterative experimentation. Several configurations were evaluated, including architectures with fewer neurons (e.g., 10–5) and those with additional hidden layers. Networks with fewer parameters exhibited underfitting and reduced generalization, while larger or deeper models increased training time without substantial improvements in accuracy. The 20–15 topology was selected as it offered a favorable trade-off between classification performance and computational efficiency.

No formal hyperparameter tuning methods (e.g., grid search or Bayesian optimization) were applied in this study as they were not part of this research objectives.

The network was implemented and trained in MATLAB using the Levenberg–Marquardt backpropagation algorithm and a cross-entropy loss function. Training data were generated through voltage simulations in the IEEE 33-bus system under various disturbance types and durations (0.2 s, 0.4 s, and 0.8 s). The dataset was partitioned into 70% for training, 15% for validation, and 15% for testing.

The proposed ANN configuration achieved a classification accuracy of 94.09% on the testing set, demonstrating strong generalization across different disturbance types and bus locations. Each input to the classifier corresponds to a 0.8 s window of voltage data, sampled at 10 kHz, resulting in 800 samples per event. Although the current implementation operates in offline mode, the compact feature set and lightweight network architecture are well suited for future integration into real-time disturbance monitoring systems.

2.4. Hilbert–Huang Transform

The Hilbert–Huang transform (HHT) is a time-frequency analysis method tailored for non-stationary and nonlinear signals. Unlike traditional techniques such as the Fourier or wavelet transforms that rely on predefined basis functions, the HHT is data-driven and adaptive [42,43]. It consists of two main stages:

• Empirical mode decomposition (EMD);
• Hilbert spectral analysis (HSA) via the Hilbert transform.

2.4.1. Empirical Mode Decomposition (EMD)

EMD decomposes a complex signal into a finite set of intrinsic mode functions (IMFs). Each IMF represents a simple oscillatory mode inherent to the original signal. For a function to qualify as an IMF, it must satisfy the following:

• The number of extrema and the number of zero crossings must either be equal or differ at most by one.
• The mean of the upper and lower envelopes—defined by local maxima and minima—must be zero at all points.

The step-by-step EMD process for a signal $x\left(t\right)$ is as follows:

• Identify local maxima and minima. Construct the upper envelope $e_{\max }\left(t\right)$ and lower envelope $e_{\min }\left(t\right)$ using interpolation.
• Compute the mean:

  $$m\left(t\right)={\displaystyle \frac{e_{\max }\left(t\right)+e_{\min }\left(t\right)}{2}}$$

  (1)
• Subtract the mean from the signal (sifting process):

  $$h_{k+1}\left(t\right)=h_k\left(t\right)-m_k\left(t\right)$$

  (2)
• When IMF conditions are met, set:

  $$c_k\left(t\right)=h_k\left(t\right)$$

  (3)
• Subtract $c_k\left(t\right)$ from $x\left(t\right)$ to compute the residue:

  $$r_k\left(t\right)=x\left(t\right)-\sum _{j=1}^k{c}_j\left(t\right)$$

  (4)
• Continue the decomposition until the final residue is obtained:

  $$x\left(t\right)=\sum _{k=1}^n{c}_k\left(t\right)+r_n\left(t\right)$$

  (5)

This adaptive decomposition preserves both time and frequency localization [44,45,46].

2.4.2. Hilbert Spectral Analysis (Hilbert Transform)

After extracting the IMFs via EMD, the Hilbert transform is applied to each IMF to extract instantaneous frequency and amplitude information [47,48,49].

Given a real-valued signal $x\left(t\right)$, its analytic representation is defined as:

$$\widehat{x}\left(t\right)=x\left(t\right)+j\tilde{x}\left(t\right)$$

(6)

where $\tilde{x}\left(t\right)$ is the Hilbert transform of $x\left(t\right)$:

$$\tilde{x}\left(t\right)={\displaystyle \frac{1}{\pi }}\mathcal{P}{\int }_{-\infty }^{\infty }{\displaystyle \frac{x\left(\tau \right)}{t-\tau }}d\tau$$

(7)

and $\mathcal{P}$ denotes the Cauchy principal value.

From the analytic signal $\widehat{x}\left(t\right)$, the following quantities are derived:

• Instantaneous amplitude:

  $$A\left(t\right)=|\widehat{x}\left(t\right)|=\sqrt{x^2\left(t\right)+{\tilde{x}}^2\left(t\right)}$$

  (8)
• Instantaneous phase:

  $$\varphi \left(t\right)={tan}^{-1}\left({\displaystyle \frac{\tilde{x}\left(t\right)}{x\left(t\right)}}\right)$$

  (9)
• Instantaneous frequency:

  $$f\left(t\right)={\displaystyle \frac{1}{2\pi }}{\displaystyle \frac{d\varphi \left(t\right)}{dt}}$$

  (10)

These quantities provide a high-resolution time–frequency representation of the original signal, enabling the identification of time-localized spectral features.

2.4.3. Feature Selection and Implementation Details

For each voltage signal, EMD typically yields between six and eight intrinsic mode functions (IMFs), depending on the duration and shape of the disturbance. However, only the first two IMFs are retained for feature extraction. These contain the most relevant high-frequency content associated with transient anomalies. Higher-order IMFs, which reflect slower oscillations and baseline drift, were excluded to reduce redundancy and dimensionality.

The sifting process follows the standard algorithm with a stopping criterion based on the standard deviation between successive iterations, with a threshold of 0.2. This ensures convergence while avoiding excessive iterations. No additional denoising methods were applied, as EMD inherently acts as a signal-adaptive filter by isolating oscillatory modes.

After obtaining the relevant IMFs, the Hilbert transform is applied to each to compute the instantaneous amplitude $A\left(t\right)$ and frequency $f\left(t\right)$ as defined in Equations (8) and (10). From these, the following statistical features are extracted:

• Mean of the instantaneous amplitude;
• Standard deviation of the instantaneous amplitude;
• Mean of the instantaneous frequency.

These features were selected due to their ability to characterize the magnitude and frequency behavior of transient disturbances in a compact form. Preliminary testing with additional features (e.g., entropy, energy, and higher-order moments) showed marginal gains in classification accuracy while increasing the risk of overfitting and computational complexity. Therefore, the chosen features offer a robust and efficient representation for training the neural network classifiers.

2.5. Study Cases

Electrical disturbances can result from environmental factors, external events, or internal operational failures. These disruptions manifest mainly through voltage variations, such as sags, surges, or complete interruptions, which compromise system stability, accelerate equipment wear, and reduce power quality.

2.5.1. System Overview

The electrical network analyzed corresponds to the IEEE 33-bus radial distribution system. Due to its realistic configuration and widespread use, this testbed is a reference model for evaluating advanced monitoring, control, and diagnostic methodologies.

Its key features include the following:

• 33 buses and 32 distribution lines.
• Nominal voltage of 12.66 kV.
• Inclusion of various types of loads and line impedances.
• A radial structure prone to disturbances such as sags and interruptions.

Table 2. Voltage and load summary in the IEEE 33-bus system.

Metric	Min	Max	Average
Bus Voltage (kV)	12.12	12.66	12.41
Real Power Demand (kW)	0	420	105.5
Reactive Power Demand (kW)	0	600	73.2

summarizes the voltage range and power values across the buses.

2.5.2. Line Impedance Profile

The impedance values across lines introduce variation in voltage drops, directly affecting disturbance propagation. Rather than listing all 33 line parameters,

Table 3. Summary of line impedance values in the IEEE 33-bus system.

Parameter	Resistance (R)	Reactance (X)
Min Value ($\mathsf{\Omega }$)	0.0922	0.0477
Max Value ($\mathsf{\Omega }$)	1.7114	1.7210
Average Value ($\mathsf{\Omega }$)	0.692	0.638

presents statistical metrics to capture impedance characteristics.

2.5.3. Simulation and Disturbance Modeling

This system was simulated using MATLAB/Simulink to test its response to voltage disturbances under controlled conditions.

The IEEE 33-bus test system is widely adopted for its radial structure, realistic voltage drops, and load diversity, making it a suitable benchmark for evaluating power quality event detection methods. It includes buses with varied electrical distances from the substation, enabling robust analysis of spatially distributed voltage anomalies.

To evaluate the classifier’s generalization across network conditions, voltage disturbances were injected at four representative buses: 4, 12, 18, and 30. These were selected to reflect a range of topological characteristics:

• Bus 4 is located near the source, experiencing minimal voltage drop.
• Bus 12 lies in the middle of the feeder.
• Bus 18 and Bus 30 are at the feeder end, where voltage sag and transient sensitivity are more severe.

Each bus underwent three disturbance durations (0.2 s, 0.4 s, 0.8 s) across all event types. This setup produced a diverse labeled dataset to train and validate the ANN classifier. Labels used for classification were previously listed in Table 1.

The IEEE 33-bus system provides a reliable environment for studying the behavior of distribution networks under anomalous operating conditions. Its structure supports the development and validation of detection algorithms and facilitates dynamic analysis of power quality events. The system’s topology is detailed in Figure 2, in which Bus 0 is referred to as the connection point to the grid.

2.6. Methodology Algorithm

The following Algorithm 1 outlines the proposed methodology for classifying electrical disturbances in distribution systems using a hybrid approach that combines the Hilbert–Huang transform (HHT) with artificial neural networks (ANNs). The procedure begins with the controlled injection of sag, swell, and interruption events into an IEEE 33-bus test system. This is followed by decomposing the resulting voltage signals into intrinsic mode functions (IMFs) through empirical mode decomposition (EMD). Subsequently, the Hilbert transform is applied to each IMF to extract time–frequency features such as instantaneous amplitude and frequency.

Algorithm 1 Disturbance classification method using HHT and ANN

Step 1: Initialize Test System Load IEEE 33-bus topology, line impedances, and load conditions: $System\leftarrow LoadIEEE33Bus\left(\right)$ Step 2: Inject Disturbances Apply $d\in \{\mathrm{Sag},\mathrm{Swell},\mathrm{Interruption}\}$ at selected nodes $i\in \{1,\dots ,33\}$ with durations $T_d=\{0.2,0.4,0.8\}$ seconds. Record voltage signal $x_{i,d}\left(t\right)$ for each scenario. Step 3: Perform Empirical Mode Decomposition (EMD) Decompose signal into Intrinsic Mode Functions (IMFs): $$x\left(t\right)=\sum _{k=1}^n{c}_k\left(t\right)+r_n\left(t\right)[\mathrm{Equation}\left(5\right)]$$ Step 4: Apply Hilbert Transform to Each IMF Obtain analytic signal ${\widehat{c}}_k\left(t\right)=c_k\left(t\right)+j{\tilde{c}}_k\left(t\right)$ [Equation (6)] Compute instantaneous amplitude $A_k\left(t\right)$, phase ${\varphi }_k\left(t\right)$, and frequency $f_k\left(t\right)$: $$f_k\left(t\right)={\displaystyle \frac{1}{2\pi }}{\displaystyle \frac{d{\varphi }_k\left(t\right)}{dt}}[\mathrm{Equation}\left(10\right)]$$ Step 5: Construct Feature Vectors For each signal: ${\mathbf{F}}_i=[A_k\left(t\right),f_k\left(t\right),{\varphi }_k\left(t\right)]$ across selected IMFs. Step 6: Assign Disturbance Labels Use classification scheme from Table 1: $$Y_i\in \{1=\mathrm{Interruption},2=\mathrm{Normal},3=\mathrm{Sag},4=\mathrm{Swell}\}$$ Step 7: Train Artificial Neural Network Design ANN with input $\mathbf{F}$ and output Y. Train with supervised learning to minimize loss $\mathcal{L}(Y,\widehat{Y})$. Step 8: Validate Performance Use testing set ${\mathbf{F}}_{test}$ to compute: Accuracy, Precision, Recall, and Confusion Matrix. Step 9: Return Outputs Return: trained weights ${\theta }^{*}$, predictions $\widehat{Y}$, and evaluation metrics.

These features are used to construct descriptive input vectors for training a supervised ANN model, which accurately identifies the type of disturbance present in each scenario. The algorithm includes all stages from simulation and preprocessing to model training and validation, providing a robust framework for detecting and classifying power quality events under dynamic operating conditions.

3. Results

3.1. Case Study: Voltage Sag

The IEEE 33-bus radial distribution system is used as the simulation platform to evaluate the proposed methodology’s performance. Disturbances are introduced in the MATLAB/Simulink environment by injecting voltage sags into selected nodes using block-level modifications to the transmission lines. Figure 3 shows a visual representation of voltage sags. Three distinct sag durations are defined, 0.2 s, 0.4 s, and 0.8 s, allowing for a multi-scale analysis of the neural network’s ability to detect, classify, and localize transient events.

Results Validation

As defined in Table 1, each disturbance is assigned a numerical label to facilitate classification. Voltage sags are represented by label 3. The detection system was evaluated across multiple scenarios by introducing voltage drops at different time points. Figure 4 illustrates the network’s response to a disturbance with a $t=0.2$ s duration. The model successfully identifies the event onset and assigns the correct label with high temporal precision.

Figure 5 shows the response to a sag with a 0.4 s duration. The classification remains consistent, with the model correctly identifying the transition between normal (label 2) and disturbed (label 3) states, and returning to label 2 once the sag ends.

The final test, presented in Figure 6, evaluates the detection of a longer-duration sag with 0.8 s duration. The classifier maintains its performance, demonstrating robustness in identifying the anomaly across varying time windows.

These results confirm the classifier’s ability to accurately detect the onset, duration, and clearance of voltage sag events. The ANN demonstrates strong generalization to different temporal scales, a critical real-time power quality monitoring capability. Unlike traditional detection methods that often rely on fixed thresholds or frequency-domain analysis, the data-driven approach employed here adapts to dynamic signal behavior, offering improved flexibility and precision.

Moreover, the model’s accurate identification of the transition boundaries between normal and abnormal states supports its application in intelligent fault detection systems, where rapid response and reliable classification are essential for maintaining grid reliability and service continuity.

3.2. Case Study: Voltage Level Interruption

The IEEE 33-bus distribution system assesses the proposed methodology’s ability to detect voltage interruption events. These are modeled in MATLAB/Simulink by inserting block-level modifications that simulate a complete voltage loss at selected nodes. Three distinct durations are defined for the analysis: 0.2 s, 0.4 s, and 0.8 s. Figure 7 presents the signal behavior during one such simulated interruption. The primary objective is to validate the neural network’s capacity to identify and classify these events under varying temporal conditions.

Results Validation

As Table 1 indicates, voltage interruptions are represented by label 1. The neural network is evaluated across three scenarios in which the interruption begins at $t=0.2$ s, $0.4$ s, and $0.8$ s, respectively. These events simulate the sudden loss of voltage due to faults, protection device operation, or equipment malfunction, all of which may lead to severe system consequences if undetected.

Figure 8 illustrates the detection output for the first case. The ANN correctly identifies the interruption onset, maintains label 1 during the disturbance window, and returns to label 2 upon signal recovery.

Figure 9 displays the detection behavior for a 0.4 s interruption. The network’s consistent response, despite variation in timing, demonstrates its generalization capability.

In the final case, shown in Figure 10, the ANN again exhibits accurate temporal labeling of the event with no misclassifications, even for the longest duration scenario.

The model’s consistent performance across these scenarios validates its robustness in detecting interruptions under varying durations and onset conditions. Such capabilities are essential for real-world deployment, where faults may occur unpredictably in time and magnitude.

Accurate classification of interruption events not only enables timely fault isolation but also supports the implementation of automated mitigation strategies. Therefore, the proposed supervised learning architecture strengthens early detection systems and enhances the intelligence and adaptability of modern distribution networks.

3.3. Case Study: Voltage Swell

The IEEE 33-bus distribution system is employed as the simulation framework to evaluate the model’s ability to detect swell events. In this case study, swell disturbances are introduced at selected nodes in the MATLAB/Simulink environment. These surges represent temporary increases in RMS voltage levels. Three durations are considered: 0.2 s, 0.4 s, and 0.8 s. Figure 11 illustrates the signal profile during a simulated swell event.

Results Validation

According to the classification scheme in Table 1, swell events are labeled with identifier 4. The network was evaluated across three time-based scenarios, each representing a different moment of surge onset: 0.2 s, 0.4 s, and 0.8 s. These events are critical in power quality analysis, as they can severely damage sensitive electronic equipment or lead to misoperation of protection devices if not adequately addressed.

Figure 12 shows the neural network’s output for the 0.2 s surge event. The system accurately detects the start and end of the disturbance and correctly assigns label 4 throughout the duration.

Figure 13 presents the network’s classification for the 0.4 s case. The system demonstrates consistency by accurately labeling the anomaly and transitioning back to label 2 after the event concludes.

The last test, shown in Figure 14, evaluates the most extended event duration (0.8 s). Again, the model performs reliably, marking a clear boundary between normal and disturbed states with no false positives or missed detections.

The experimental results confirm the robustness of the proposed ANN-based classifier in recognizing Swell events of varying durations. The system’s ability to maintain accuracy under multiple disturbance windows supports its deployment in predictive monitoring and adaptive protection schemes. Moreover, its strong generalization across different time scales reinforces its value as a dynamic fault detection tool within innovative grid environments.

3.4. Results Summary

The Hilbert–Huang transform (HHT) was applied to extract relevant features from voltage disturbance signals in the IEEE 33-bus system. This adaptive time–frequency analysis combines empirical mode decomposition (EMD) with the Hilbert transform, enabling non-stationary signals to be decomposed into a set of intrinsic mode functions (IMFs), each capturing oscillatory behavior at different time scales. This decomposition facilitates the separation of fundamental signal components from transient irregularities, allowing for high-resolution and localized characterization of electrical disturbances.

Applying the Hilbert transform to each IMF yields instantaneous amplitude, phase, and frequency parameters. These descriptors form a detailed representation of the signal’s dynamic evolution, particularly beneficial for identifying the timing, duration, and severity of non-stationary disturbances. Figure 15 illustrates this process for the first IMF, where high-resolution time–frequency data enables precise anomaly detection.

As shown in Figure 16, the visualization of each IMF’s amplitude and frequency spectrum allows for identifying both abrupt and subtle alterations in the signal. This capability makes the HHT framework a powerful tool for advanced power quality assessment and transient event localization in modern distribution networks.

Once the HHT-transformed signals were used as training inputs, the neural network was optimized using a performance-based validation approach. As shown in Figure 17, the model achieved a classification error of 0.125 across 250,000 data samples. The rapid convergence observed demonstrates successful generalization and training stability.

To evaluate the generalization of the model under complex operating conditions, additional case studies were developed using a realistic simulation of the IEEE 33-bus system. In each scenario, simultaneous disturbances (sags, interruptions, and swells) were injected across all three phases at multiple time points ($t=0.2$ s, $0.4$ s, and $0.8$ s). In Case 1 (Figure 18), the model successfully identified the type, timing, and duration of each event, demonstrating its ability to differentiate among label classes with no misclassification.

Figure 19 shows the system’s response to the second event set. The classifier maintained high precision in distinguishing transitions between normal (label 2), interruptions (label 1), sags (label 3), and swells (label 4). This highlights the model’s robustness against temporal variability in fault occurrence.

The third test, presented in Figure 20, validated the network’s ability to detect extended and overlapping anomalies under multiphase symmetrical conditions. The classifier’s consistent behavior across all tests confirms its high generalization capacity and resilience to complex dynamic operating states.

Overall, the combined use of the HHT and neural networks provides a comprehensive and adaptive strategy for real-time monitoring. The model not only detects and classifies transient disturbances with high precision but also retains performance under multiphase, time-variable conditions. This positions the approach as a scalable and intelligent solution for power quality assurance and advanced fault diagnostics in modern distribution grids.

The results obtained demonstrate the effectiveness of the proposed HHT-based ANN classification framework in fulfilling the primary objectives of this study. The model achieved an overall classification accuracy of 94.09%, with strong generalization across varying disturbance durations (0.2 s, 0.4 s, 0.8 s) and spatial locations within the IEEE 33-bus system. The confusion matrix shows that the classifier reliably distinguishes between voltage sag, swell, and interruption events, each associated with distinct voltage signatures defined by IEEE 1159.

The relatively shallow ANN architecture enabled fast convergence (as seen in the training curve) while maintaining low computational complexity, making the approach suitable for real-time deployment. These results confirm that the integration of HHT features (instantaneous amplitude and frequency) provides a rich, non-stationary feature space that enhances the classifier’s ability to detect subtle transient patterns.

Furthermore, the method’s robustness across buses and time windows suggests that it can be effectively deployed in smart distribution networks where operating conditions are dynamic and disturbance propagation is topology-dependent. Thus, the presented results validate the proposed approach not only in terms of raw performance metrics but also in terms of its alignment with real-world monitoring requirements.

3.5. Expanded Experimental Validation

The training dataset was constructed using simulated voltage signals in the IEEE 33-bus test system under four event classes: normal operation and voltage sag, swell, and interruption. Each class contains 300 labeled signals per duration (0.2 s, 0.4 s, and 0.8 s), yielding a total of 3600 signals. Disturbance events were injected with randomized onset times, amplitudes, and durations within the IEEE 1159-2019 guidelines. Voltage sags varied between 60 and 90% of nominal voltage, swells between 110 and 130%, and interruptions dropped below 10% for specified intervals. No synthetic noise was added, but signal diversity was ensured through perturbation variability and randomized bus locations.

To assess generalization capability, we adopted five-fold cross-validation over a stratified 80/20 training/test split, ensuring that all event types and durations were proportionally represented across folds. This method provides robust performance estimation and reduces the risk of overfitting, comparable to using an independent test set, without requiring additional simulations. The average classification accuracy across folds was 94.09%, with standard deviation below 1%, confirming stable model behavior.

A confusion matrix summarizing classification outcomes is shown in Figure 21, with strong per-class separability.

Table 4. Per-class performance metrics.

Class	Precision	Recall	F1-Score
Normal	0.95	0.93	0.94
Sag	0.94	0.96	0.95
Swell	0.92	0.89	0.90
Interruption	0.97	0.98	0.98

provides precision, recall, and F1-score metrics for each class, highlighting the model’s balanced accuracy across all types of disturbances.

To formally define these metrics:

• Precision:

  $$\mathrm{Precision}={\displaystyle \frac{TP}{TP+FP}}$$

  (11)
• Recall:

  $$\mathrm{Recall}={\displaystyle \frac{TP}{TP+FN}}$$

  (12)
• F1-score:

  $$\mathrm{F}1-\mathrm{score}=2·{\displaystyle \frac{\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

  (13)

Although a comparison with other classifiers (e.g., SVM, decision tree) could offer additional insight, the focus of this work was to validate the neural network approach under multiscale event durations using a consistent, diverse dataset. Given the satisfactory results and robust generalization observed through cross-validation, extending the study to alternative classifiers is considered future work.

Real-Time Feasibility and Computational Cost

Although real-time deployment was not directly implemented, the methodology was designed with computational efficiency in mind. The average processing time per event, including both Hilbert–Huang decomposition and ANN classification, was approximately 0.13 s using MATLAB on a standard Intel i7-11800H CPU with 16 GB RAM. This runtime supports near real-time monitoring scenarios (e.g., 5 Hz sampling) with minimal optimization.

This result aligns with reported empirical runtimes for Hilbert–Huang transform (HHT)-based feature extraction in non-parallelized settings, where typical values range between 0.1 and 0.2 s per signal for similar data lengths [50,51]. Furthermore, the proposed ANN architecture remains lightweight, with fewer than 100 trainable parameters, enabling fast inference and compatibility with embedded or edge-device platforms in smart grid environments.

Comparative Analysis with Existing Methods

To evaluate the results from the implementation of the HHT-derived ANN model, in

Table 5. Comparison of recent PQD classification methodologies (2020–2025).

Author	Year	Methodology	Test System	Accuracy (%)	Remarks
Li et al. [50]	2021	S-Transform + Convolutional Neural Network (CNN)	Simulated PQD dataset	High	Robust classification using S-Transform; tolerant to noise and distortion
Wang et al. [51]	2020	DWT + Hierarchical Extreme Learning Machine (H-ELM)	Synthetic signals	Not specified	Multi-level structure improves generalization and handles imbalanced data
Bhargava et al. [52]	2023	VMD + CNN	Simulated signals with noise injection	Not specified	Emphasized complex transient behavior using adaptive decomposition
Güvengir et al. [53]	2025	Multiple ML models (SVM, XGBoost, KNN)	Real-world transmission data	Not specified	Compared algorithms on real SCADA/PQD data for grid-scale analysis
Dubey [54]	2022	DWT + Online Sequential ELM (OS-ELM)	Simulated PQD cases	Not specified	Designed for real-time learning; good for streaming classification
Present Work	2025	Hilbert–Huang Transform (HHT) + Artificial Neural Network (ANN)	IEEE 33-bus system	94.09	Strong performance with low computational cost and multiscale sensitivity

, this paper reviews other approaches to power quality disturbance (PQD) classification, focusing on more recent published works, which naturally enables understanding where the gaps in the research lie. The table features the most impactful works published after 2020 that use other signal processing and machine learning techniques like Discrete Wavelet Transform (DWT), Variational Mode Decomposition (VMD), and S-Transform, alongside Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Extreme Learning Machine (ELM) classifiers.

While some methods like DWT combined with FPGA-accelerated ELMs reach classification accuracies over 99%, these methods often rely on ideal signal conditions or need specialized equipment. In comparison, the approach presented in this paper attained 94.09% classification accuracy while processing each event in 0.13 s on a standard CPU. This trade-off between accuracy and efficiency highlights the method’s potential for real-time application in smart distribution systems. In addition, the application of Hilbert–Huang transform allows for multiscale analysis of changing voltage signals and provides a powerful complement to standard time–frequency decomposition methods.

4. Conclusions

The neural network model trained using features obtained with the Hilbert–Huang transform reached a classification accuracy of 94.09% for electrical system disturbance classification. The level of accuracy attained confirms the model is well-suited for the automatic detection of transient events, underscoring its value for proactive power quality monitoring.

The architectural model described herein displayed a high level of generalization, performing voltage dip detection and interruption detection, as well as recording overrides of the voltage at 0.2, 0.4, and 0.8 s. In all scenarios studied, the neural network was successful in detecting the beginning, active period, and termination of the event without any false detections.

As the ANN relies on features derived from the Hilbert–Huang transform—namely, the instantaneous amplitude and frequency extracted from the first two IMFs—the HHT remains an integral part of the inference pipeline. However, due to the localized and adaptive nature of the EMD–Hilbert process, feature extraction can be performed efficiently, supporting the feasibility of deploying the proposed method in real-time disturbance monitoring systems.

While simulating simultaneous active disturbances on all three phases of the 33 node IEEE system, the model consistently demonstrated accurate classification devoid of errors. The ability to sustain this level of performance illustrates the capacity of the system to function in highly dynamic scenarios, which is necessary in modern distribution systems that are subject to high-load fluctuations and transients.

The current framework was evaluated using synthetic, noise-free voltage signals under controlled scenarios. Although the results demonstrate strong classification accuracy, practical deployment may face challenges such as measurement noise, data incompleteness, or signal distortions. Future research will aim to validate the model on field-acquired datasets and explore alternative neural architectures, such as convolutional or recurrent networks, to enhance generalization under real-world conditions.

5. Future Work and Challenges

Based on the results obtained, one of the central future projections is to apply the proposed model to the analysis of short-circuit conditions in distribution networks. This type of event represents one of the most critical scenarios in the operation of electrical systems, and its early detection is key to avoiding severe damage to the infrastructure. Incorporating short circuits as a new disturbance class within the classification system would enable the evaluation of the model’s ability to identify high-severity events, accurately differentiating them from other, less severe, transient phenomena.

Another relevant aspect of future work will be the validation of the model with real signals obtained directly from operating electrical networks. This approach will allow for the analysis of its behavior under practical conditions characterized by noise, uncertainty, and multiple sources of interference. Additionally, the exploration of more sophisticated neural architectures, such as convolutional networks (CNNs), recurrent networks (RNNs), or hybrid networks, is proposed, which could improve accuracy, reduce response times, and increase the system’s robustness to non-ideal or incomplete data.

Additionally, while this study focused on the voltage response at the point of disturbance injection, it is important to note that power quality events often propagate to adjacent buses, especially in radial networks. These neighboring locations may experience more severe or earlier voltage deviations due to impedance paths and electrical proximity. As such, future work will explore spatial feature integration by capturing the voltage behavior of adjacent buses to enhance detection accuracy and resilience. This approach will support distributed sensing schemes and better reflect real-world deployment scenarios in active distribution networks.

The proposed methodology was designed with implementation feasibility in mind. By leveraging a lightweight neural architecture and using HHT-derived statistical features, the framework avoids the high computational overhead typical of more complex models such as deep CNNs or ensemble classifiers. This makes it suitable for deployment on embedded platforms or edge devices for real-time monitoring applications.

While the current validation is based on simulation, the approach is modular and can be adapted to process real-time voltage signals from existing smart meters or phasor measurement units (PMUs). While the classifier achieved high accuracy (94.09%) and consistent recall across all event types, future studies could explore real-time adaptation and alternative network architectures to enhance robustness and reduce latency in practical deployment scenarios. Integration efforts will also focus on hardware-in-the-loop validation and compatibility with modern grid monitoring infrastructure, paving the way for scalable, accurate, and responsive power quality monitoring in active distribution networks.

Finally, we propose integrating the model into real-time monitoring platforms, such as SCADA systems or embedded environments (such as Raspberry Pi or FPGA), to enable continuous monitoring of electrical service quality. This implementation would not only enable efficient disturbance detection but also facilitate the development of automatic response schemes that respond promptly to anomalous events. Together, these future directions position the proposed approach as a solid foundation for designing intelligent protection and diagnostic systems in modern distribution networks.