Description:DESCRIPTION:
Field of the invention:
[0001] The present disclosure generally relates to the technical field of power electronics and grid protection systems, and in specific relates to a machine learning-based, self-adaptive system employing a Hoeffding Tree classifier for passive islanding detection in inverter-based distributed generation systems operating under weak grid conditions.
Background of the invention:
[0002] With the increasing integration of distributed energy resources (DERs) such as photovoltaic inverters and wind turbines into modern power systems, ensuring grid stability and operational safety has become a critical challenge. One of the key threats to grid integrity is the phenomenon of unintentional islanding, where a portion of the distribution network continues to be energized by local inverters even after disconnection from the main grid.
[0003] Unintentional islanding can result in severe consequences, including the safety hazards for utility personnel, damage to equipment, power quality issues, and violation of grid codes. Consequently, accurate, fast, and reliable islanding detection methods are essential for compliance and system resilience. Various islanding detection techniques (IDTs) have been developed.
[0004] Passive methods that monitor local parameters such as voltage, frequency, and harmonic distortion at the point of common coupling (PCC). While non-intrusive, passive methods exhibit large Non-Detection Zones (NDZs), especially under near power-balanced conditions. Their fixed-threshold-based operation often results in false positives or missed detections, particularly in weak grid scenarios (low short-circuit ratio, SCR < 3).
[0005] Active methods that inject small disturbances (e.g., reactive power shifts or frequency perturbations) into the grid and observe the response. However, these reduce NDZs by injecting disturbances but degrade power quality and may negatively impact sensitive loads. Furthermore, they are unsuitable for weak grids where natural fluctuations mask artificial perturbations.
[0006] Communication-based methods that use control signals and data exchange between inverters or between inverters and utility control centers. While the communication-based methods are effective, these require expensive infrastructure like SCADA systems and are prone to data loss or synchronization failures. Their complexity limits scalability and reliability.
[0007] Intelligent methods that leverage machine learning (e.g., support vector machines, neural networks) to classify operating conditions based on extracted features. Each approach attempts to distinguish between islanded and grid-connected states under varying load and generation conditions. However, traditional ML-based islanding detectors typically rely on offline-trained models that cannot adapt to changing grid dynamics. Additionally, they often demand significant memory and computational resources, making them unsuitable for real-time embedded applications.
[0008] Therefore, there is a need for an islanding detection method that is adaptive, lightweight, fast, and accurate, capable of operating under weak grid conditions without introducing disturbances or requiring communication systems. There is also a need for a self- adaptive system for passive islanding detection in inverter-based microgrids operating under weak grid conditions
Objectives of the invention:
[0009] The primary objective of the invention is to provide a machine learning-based, self-adaptive system that employs a Hoeffding Tree classifier for passive islanding detection in inverter-based distributed generation systems operating under weak grid conditions.
[0010] The other objective of the invention is to provide a self-adaptive, machine learning-based islanding detection technique that is capable of continuously learning and adapting to changing grid and load conditions without the need for manual retraining.
[0011] The other objective of the invention is to provide a self- adaptive system that ensures accurate and rapid detection of unintentional islanding events in inverter-based distributed generation (IBDG) systems.
[0012] Another objective of the invention is to provide a self- adaptive system that eliminates non-detection zones (NDZs) commonly associated with conventional passive detection methods, especially under power-balanced conditions or low short-circuit ratio (SCR) scenarios.
[0013] The other objective of the invention is to provide a self- adaptive system that designs a non-intrusive, disturbance-free detection method that operates solely on real-time measurements without injecting external signals, thereby maintaining grid power quality.
[0014] Yet another objective of the invention is to provide a self- adaptive system that enables compatibility with both grid-forming (GFMI) and grid-following (GFLI) inverters, ensuring wide applicability across diverse microgrid configurations.
[0015] Another objective of the invention is to provide a self- adaptive system that develops a decentralized and communication-free solution, avoiding the need for supervisory control and data acquisition (SCADA) systems or inter-inverter communication infrastructure.
[0016] Another objective of the invention is to provide a self- adaptive system that supports efficient real-time implementation on embedded platforms, such as microcontrollers and FPGA systems, using a lightweight Hoeffding Tree classifier with minimal memory and computational overhead.
Summary of the invention:
[0017] The present disclosure proposes a self-adaptive system and method for passive islanding detection in microgrids operating under weak grid conditions. The following presents a simplified summary in order to provide a basic understanding of some aspects of the claimed subject matter. This summary is not an extensive overview. It is not intended to identify key/critical elements or to delineate the scope of the claimed subject matter. Its sole purpose is to present some concepts in a simplified form as a prelude to the more detailed description that is presented later.
[0018] In order to overcome the above deficiencies of the prior art, the present disclosure is to solve the technical problem to provide a machine learning-based, self-adaptive system employing a Hoeffding Tree classifier for passive islanding detection in inverter-based distributed generation systems operating under weak grid conditions.
[0019] According to an aspect, the invention provides a self-adaptive system for passive islanding detection in inverter-based microgrids operating under weak grid conditions. The self-adaptive system comprises a computing device having a processor and a memory, which stores one or more instructions executable by the processor.
[0020] In one embodiment, the computing device is in communication with a server via a network. The processor is configured to execute plurality of modules. The plurality of modules comprises a parameter acquisition module, a feature extraction module, a filter module, and a Hoeffding Tree classifier module.
[0021] In one embodiment, the parameter acquisition module is configured to receive one or more measured electrical parameters at a point of common coupling (PCC). The measured electrical parameters comprise, but are not limited to, voltage, current, frequency, active power, and reactive power at the point of common coupling (PCC).
[0022] In one embodiment, the feature extraction module is configured to extract one or more dynamic electrical features from the electrical parameters. The dynamic electrical features comprise at least one of dynamic frequency change rate (DFTR), dynamic voltage change rate (DVTR), dynamic active and reactive power change rates (DPTR, DQTR), dynamic voltage-to-frequency change rate ratio (DVTFTR), dynamic power-to-frequency change rate ratios (DPTFTR, DQTFTR), total harmonic distortion of voltage and current (THDv, THDi), and voltage unbalance factor (VUF).
[0023] In one embodiment, the filter module is configured to normalize and pre-process the dynamic electrical features to produce a pre-processed feature stream. The pre-processed feature stream is generated by applying a moving average filter with a 10-sample window and min-max normalization in the range of 0 to 1.
[0024] In one embodiment, the Hoeffding Tree classifier module is configured to classify an operating state as grid-connected or islanded based on the pre-processed feature stream to obtain one or more classified samples, thereby generating an islanding detection signal if a temporal consistency condition of multiple consecutive classified samples is achieved. The temporal consistency condition comprises confirming islanding status based on at least three to five consecutive samples classified as islanded with a probability threshold of at least 90%. An islanding detection time is less than 15 milliseconds for both single and multi-inverter-based systems.
[0025] In one embodiment, the Hoeffding Tree classifier module comprises Hoeffding Tree-based Passive Islanding Detection Technique (HT-PIDT) model, wherein the HT-PIDT model is initially trained offline using a training dataset simulated under different grid-forming inverter (GFMI) and grid-following inverter (GFLI) operational scenarios. The Hoeffding Tree classifier module is adapted to operate effectively under short-circuit ratio (SCR) conditions ranging from 1.5 to 3. The Hoeffding Tree classifier module is configured to incrementally learn from the pre-processed feature stream using a Hoeffding bound-based statistical criterion for node splitting, and updates a decision structure in real time. The Hoeffding Tree classifier module is configured to perform binary classification based on continuously updated statistics at each decision node without requiring storage of historical data. The HT-PIDT model is implemented in real time on an embedded controller selected from an STM32F7 microcontroller or an FPGA-based NI sbRIO controller.
[0026] In one embodiment, class imbalance in the training dataset is addressed using a synthetic minority oversampling technique (SMOTE) to achieve a target distribution of 70% grid-connected and 30% islanded samples. The self-adaptive system does not require any communication infrastructure or disturbance injection, thereby being fully decentralized and non-intrusive.
[0027] According to another aspect, the invention provides a self-adaptive passive islanding detection method for inverter-based microgrids operating under weak grid conditions. First, one or more measured electrical parameters is acquired by the parameter acquisition module at a point of common coupling (PCC). Next, one or more dynamic electrical features are extracted by the feature extraction module from the one or more measured electrical parameters.
[0028] Next, the one or more dynamic electrical features are normalized and filtered by the filter module to produce a pre-processed feature stream. Later, an operating state as grid-connected or islanded is classified by the Hoeffding Tree classifier module based on the pre-processed feature stream to obtain one or more classified samples, thereby generating an islanding detection signal if a temporal consistency condition of multiple consecutive classified samples is achieved.
[0029] Further, objects and advantages of the present invention will be apparent from a study of the following portion of the specification, the claims, and the attached drawings.
Detailed description of drawings:
[0030] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate an embodiment of the invention, and, together with the description, explain the principles of the invention.
[0031] FIG. 1 illustrates a block diagram of a self-adaptive system for passive islanding detection in inverter-based microgrids operating under weak grid conditions, in accordance to an exemplary embodiment of the invention.
[0032] FIG. 2 illustrates a test system to evaluate the proposed self-adaptive system, in accordance to an exemplary embodiment of the invention.
[0033] FIG. 3 illustrates a single line diagram of the inverter with the block diagram of the invented HT-PIDT model, in accordance to an exemplary embodiment of the invention.
[0034] FIG. 4 illustrates a graph depicting performance accuracy comparison of the self-adaptive system with SVM and ELM based methods, in accordance to an exemplary embodiment of the invention, in accordance to an exemplary embodiment of the invention.
[0035] FIG. 5 illustrates a graph depicting performance accuracy comparison of the self-adaptive system with SVM and ELM under weak grid condition (SCR=1.5), in accordance to an exemplary embodiment of the invention.
[0036] FIG. 6 illustrates a flowchart of a self-adaptive passive islanding detection method for inverter-based microgrids operating under weak grid conditions, in accordance to an exemplary embodiment of the invention.
Detailed invention disclosure:
[0037] Various embodiments of the present invention will be described in reference to the accompanying drawings. Wherever possible, same or similar reference numerals are used in the drawings and the description to refer to the same or like parts or steps.
[0038] The present disclosure has been made with a view towards solving the problem with the prior art described above, and it is an object of the present invention to provide a machine learning-based, self-adaptive system employing a Hoeffding Tree classifier for passive islanding detection in inverter-based distributed generation systems operating under weak grid conditions.
[0039] According to an exemplary embodiment of the invention, FIG. 1 refers to a block diagram of a self-adaptive system 100 for passive islanding detection in inverter-based microgrids operating under weak grid conditions. The self-adaptive system 100 comprises a computing device 102 having a processor 104 and a memory 106, which stores one or more instructions executable by the processor 104. These instructions may be executed to cause the self-adaptive system 100 to perform the various functionalities. The processor 104 acts as the central processing unit (CPU) of the self-adaptive system 100, responsible for coordinating different tasks and carrying out complex operations, data processing, and decision-making by fetching instructions from the memory 106, thereby decoding the instructions and executing the necessary actions.
[0040] In one embodiment herein, the memory 106 serves as the storage component of the self-adaptive system 100, holding the executable instructions, as well as any data or information required by the processor 104 to perform its tasks. The data includes user inputs, system configurations, and any other relevant data needed for the system's operations. Through the communication between the processor 104 and the memory 106, the self-adaptive system 100 is able to process the user inputs, access stored information, perform computations, and make decisions accordingly.
[0041] In one embodiment, the computing device 102 is in communication with a server 110 via a network 112. The processor 104 is configured to execute plurality of modules 108. The plurality of modules 108 comprises a parameter acquisition module 114, a feature extraction module 116, a filter module 118, and a Hoeffding Tree classifier module 120.
[0042] In one embodiment herein, the computing device 102 represents any electronic device that the user can utilize to interact with the self-adaptive system 100. The computing device 102 can be, but not limited to, a smartphone, a laptop, a tablet, a personal computer, or any other suitable electronic device. The computing device 102 serves as the user's gateway to accessing and interacting with the self-adaptive system 100. The computing device 102 is configured to enable the user to engage with the system's functionalities and capabilities through a user interface.
[0043] In one embodiment herein, the user interface is a crucial component of the computing device 102, which allows the users to input commands, receive information, and control the self-adaptive system 100. The user interface can be, but not limited to, a touch screen, a keyboard, a mouse, voice recognition modules, gesture recognition sensors, and virtual reality interfaces. The versatility of the user interface ensures that the users can engage with the self-adaptive system 100 in a manner that is most intuitive and comfortable for the users, thereby catering to a wide range of user preferences and accessibility needs. The computing device 102 empowers the users to interact with the self-adaptive system 100 seamlessly and efficiently by providing multiple user interface options, thereby leveraging the most appropriate input and output modalities for their specific needs and preferences.
[0044] In one embodiment herein, the computing device 102 is in communication with the server 110 via the network 112. The network 112 acts as a communication that allows the computing device 102 to interact with the other components of the self-adaptive system 100, thereby facilitating the exchange of data, commands, and information. In one embodiment herein, the network 112 can be a wireless communication infrastructure, which offers the users flexibility and convenience when interacting with the self-adaptive system 100. This wireless connectivity enables the users to access the self-adaptive system 100 from various locations, without being tethered to a fixed physical connection.
[0045] In one embodiment herein, the network 112 can be, but not limited to, Local Area Network (LAN), Cellular Network, Wide Area Network (WAN), Intranet, Virtual Private Network (VPN), and wireless networks that use radio frequency (RF) or infrared (IR) technology to transmit data without the need for physical cables, thereby providing mobility and flexibility. The versatility of the network 112 ensures that the computing device 102 can seamlessly connect to the server 110, thereby enabling the users to access the system’s 100 functionalities and resources from a variety of locations and devices. This wireless connectivity enhances the overall accessibility and convenience of the self-adaptive system 100 for the users.
[0046] In one embodiment, the parameter acquisition module 114 is configured to receive one or more measured electrical parameters at a point of common coupling (PCC). The measured electrical parameters comprise, but are not limited to, voltage, current, frequency, active power, and reactive power at the point of common coupling (PCC).
[0047] In one embodiment, the feature extraction module 116 is configured to extract one or more dynamic electrical features from the electrical parameters. The dynamic electrical features comprise at least one of dynamic frequency change rate (DFTR), dynamic voltage change rate (DVTR), dynamic active and reactive power change rates (DPTR, DQTR), dynamic voltage-to-frequency change rate ratio (DVTFTR), dynamic power-to-frequency change rate ratios (DPTFTR, DQTFTR), total harmonic distortion of voltage and current (THDv, THDi), and voltage unbalance factor (VUF).
[0048] In one embodiment, the filter module 118 is configured to normalize and pre-process the dynamic electrical features to produce a pre-processed feature stream. The pre-processed feature stream is generated by applying a moving average filter with a 10-sample window and min-max normalization in the range of 0 to 1.
[0049] In one embodiment, the Hoeffding Tree classifier module 120 is configured to classify an operating state as grid-connected or islanded based on the pre-processed feature stream to obtain one or more classified samples, thereby generating an islanding detection signal if a temporal consistency condition of multiple consecutive classified samples is achieved. The temporal consistency condition comprises confirming islanding status based on at least three to five consecutive samples classified as islanded with a probability threshold of at least 90%. An islanding detection time is less than 15 milliseconds for both single and multi-inverter-based systems.
[0050] In one embodiment, the Hoeffding Tree classifier module 120 comprises Hoeffding Tree-based Passive Islanding Detection Technique (HT-PIDT) model, wherein the HT-PIDT model is initially trained offline using a training dataset simulated under different grid-forming inverter (GFMI) and grid-following inverter (GFLI) operational scenarios. The Hoeffding Tree classifier module 120 is adapted to operate effectively under short-circuit ratio (SCR) conditions ranging from 1.5 to 3. The Hoeffding Tree classifier module 120 is configured to incrementally learn from the pre-processed feature stream using a Hoeffding bound-based statistical criterion for node splitting, and updates a decision structure in real time. The Hoeffding Tree classifier module 120 is configured to perform binary classification based on continuously updated statistics at each decision node without requiring storage of historical data. The HT-PIDT model is implemented in real time on an embedded controller selected from an STM32F7 microcontroller or an FPGA-based NI sbRIO controller.
[0051] In one embodiment, class imbalance in the training dataset is addressed using a synthetic minority oversampling technique (SMOTE) to achieve a target distribution of 70% grid-connected and 30% islanded samples. The self-adaptive system 100 does not require any communication infrastructure or disturbance injection, thereby being fully decentralized and non-intrusive.
[0052] According to another exemplary embodiment of the invention, FIG. 2 refers to a test system 200 to evaluate the proposed self-adaptive system 100. The self-adaptive passive islanding detection method initiates with the generation of the training dataset, which serves as the foundation for the initial offline training of the Hoeffding Tree classifier module 120. The training dataset is constructed by extracting a plurality of dynamic electrical features from the point of common coupling (PCC), facilitating the differentiation between grid-connected and islanded operating states. For the purpose of dataset generation, a test system 200 is configured and simulated within a MATLAB-based environment, encompassing a range of operational scenarios incorporating both grid-forming inverter (GFMI) and grid-following inverter (GFLI) control schemes. A representative set of such scenarios is detailed in Table 1. During each scenario, key electrical parameters at the PCC, including voltage, current, system frequency, and active and reactive power, are continuously monitored and recorded. These measurements are subsequently processed to compute the required feature vectors for use in the training of the Hoeffding Tree classifier module 120.
[0053] Table 1:
Case No. Scenario Grid breaker statues
1. 𝑃𝑟𝑒𝑓 = 0.4 𝑡𝑜 0.8 𝑝𝑢 0
2. 𝑃𝑟𝑒𝑓 = 0.8 𝑡𝑜 0.5 𝑝𝑢 0
3. 𝑃𝑟𝑒𝑓 = 0.5 𝑡𝑜 1 𝑝𝑢 0
4. 𝑃𝑟𝑒𝑓 = 1 𝑝𝑢 0 𝑡𝑜 1
5. 𝑃𝑟𝑒𝑓 = 0.7 𝑝𝑢 0 𝑡𝑜 1
6. 𝑄𝑟𝑒𝑓 = 0 𝑝𝑢 0 𝑡𝑜 1
7. 𝑄𝑟𝑒𝑓 = 0 𝑡𝑜 0.5 𝑝𝑢 0
8. 𝑄𝑟𝑒𝑓 = 0.5 𝑝𝑢 0 𝑡𝑜 1
9. 𝑄𝑟𝑒𝑓 = 0.5 𝑡𝑜 0.2 𝑝𝑢 0
10. 𝑄𝑟𝑒𝑓 = 0.2 𝑝𝑢 0 𝑡𝑜 1
11. 𝑃𝐿𝑜𝑎𝑑 = 0.6 𝑡𝑜 1 𝑝𝑢 0
12. 𝑃𝐿𝑜𝑎𝑑 = 1 𝑡𝑜 0.4 𝑝𝑢 0
13. 𝑃𝐿𝑜𝑎𝑑 = 0.4 𝑝𝑢 0 𝑡𝑜 1
14. 𝑃𝑙𝑜𝑎𝑑 = 0.8 𝑝𝑢 0 𝑡𝑜 1
15. 𝑄𝐿𝑜𝑎𝑑 = 0.2 𝑡𝑜 0.5 𝑝𝑢 0
16. 𝑄𝐿𝑜𝑎𝑑 = 0.5 𝑝𝑢 0 𝑡𝑜 1
17. 𝑄𝑙𝑜𝑎𝑑 = 0.5 𝑡𝑜 0.3 𝑝𝑢 0
18. 𝑄𝑙𝑜𝑎𝑑 = 0.3 𝑝𝑢 0 𝑡𝑜 1
19. 𝐿𝑔 = 0.4 𝑡𝑜 0.6 𝑚𝐻 0
20. 𝑉𝑔 = 1 𝑡𝑜 1.05 𝑝𝑢 0

[0054] Where, ensures high-resolution monitoring of frequency instability.



[0055] To improve the sensitivity of the classification process of the Hoeffding Tree classifier module 120, a derived feature, namely the Voltage-to-Frequency Change Rate Ratio (DVTFTR), is introduced, which represents the ratio between the dynamic rate of change of voltage and the dynamic rate of change of frequency. This parameter captures the coupling behavior between voltage and frequency transients, which is particularly informative under conditions of islanding where the loss of grid reference often results in disproportionate or anomalous variations in these electrical quantities. The DVTFTR feature enables the Hoeffding Tree classifier module 120 to identify subtle system disturbances that are indicative of islanding, especially in microgrid environments utilizing heterogeneous inverter control strategies, such as mixed grid-forming and grid-following inverters, or operating under weak grid conditions. Additionally, the normalization inherent in the ratio formulation reduces the impact of absolute measurement noise, thereby improving the robustness and reliability of the feature for real-time classification.

[0056] The feature extraction module 116 further utilizes the Active Power-to-Frequency Change Rate Ratio (DPTFTR) and the Reactive Power-to-Frequency Change Rate Ratio (DQTFTR) as additional discriminative features. These ratios quantify the relative response of active and reactive power, respectively, to variations in system frequency, thereby characterizing the transient behavior of the power system during potential islanding events. The Hoeffding Tree classifier module 120 employs these normalized indicators to enhance the detection of islanding conditions by identifying characteristic deviations in power-frequency coupling. These features are particularly advantageous in weak grid environments, where conventional detection metrics are prone to inaccuracies due to high impedance, reduced fault current availability, and inverter-dominated dynamic responses. The normalized nature of DPTFTR and DQTFTR mitigates the effects of absolute signal magnitudes, contributing to improved robustness and generalization of the classification model under diverse operating conditions.


[0057] Total Harmonic Distortion (THD) for voltage and current:


[0058] Voltage Unbalance Factor (VUF):

[0059] A Pearson correlation analysis identified rate-of-change-of-frequency (ROCOF), 𝑇𝐻𝐷(𝑉), and ROTOP as the most informative features (ρ > 0.85). Less discriminative parameters such as reference power setpoints (𝑃𝑟𝑒𝑓, 𝑄𝑟𝑒𝑓) were excluded to reduce computational overhead. Table 2 summarize those selected features with their benefits.
[0060] Table 2:
Feature Benefits
DFTR Primary indicator of frequency instability during islanding; sensitive to grid disconnection
DVTR Captures fast voltage dips/swells in response to disturbances; effective in weak grids
DPTR Reflects real power imbalance; useful in detecting load-generation mismatch
DQTR Indicates reactive power imbalances; linked to voltage stability in weak grids
DVTFTR Highlights disproportionate voltage shifts for small frequency drifts; detects weak grid effects
DPTFTR Identifies cases where power mismatches cause frequency variations;
effective in inverter-driven grids
DQTFTR Diagnoses poor voltage regulation contributing to frequency deviations in islanded states
THDv Detects distortion caused by inverter switching and loss of grid reference
THDi Indicates current waveform distortion under non-linear load and islanded operation
VUF Identifies voltage asymmetry due to unbalanced loads or inverter mismatch post-islanding

[0061] All features are normalized to the range [0, 1] using min-max scaling to ensure uniform weighting:

[0062] In order to suppress high-frequency noise components in the measured voltage and frequency signals while preserving signal phase integrity, a non-causal moving average filter with a fixed window size of ten samples is applied. This preprocessing step enhances signal quality for subsequent feature extraction without introducing temporal distortion. Furthermore, to mitigate the impact of class imbalance in the training dataset, the Synthetic Minority Oversampling Technique (SMOTE) is employed. This algorithm generates synthetic samples for the underrepresented islanded class to achieve a balanced dataset distribution, maintaining approximately 70% of samples corresponding to grid-connected conditions and 30% to islanded conditions. This class rebalancing improves the learning stability and generalization capability of the Hoeffding Tree classifier module 120.
[0063] The processed and balanced dataset is subsequently utilized to perform offline initialization of the HT-PIDT model. This offline training phase establishes the initial decision structure of a Hoeffding Tree classifier by configuring the root and internal nodes based on statistically significant feature splits derived from the training data. A pre-trained model thereby achieves a satisfactory baseline classification performance prior to deployment. Upon deployment, the model transitions to real-time operational mode, wherein it continues to adapt incrementally to streaming data inputs, maintaining responsiveness to evolving grid conditions and system dynamics.
[0064] During real-time microgrid operation, electrical parameters at the point of common coupling (PCC) are continuously acquired and monitored. The same set of diagnostic features described during the offline training phase—comprising time-domain and frequency-domain indicators—are extracted from the real-time measurement signals. These features are subsequently subjected to signal conditioning, including filtering and normalization procedures in accordance with pre-processing steps B and C. The resulting feature vectors are then streamed sequentially into the HT-PIDT classification module. This data stream emulates live operational conditions and enables the classifier to perform both instantaneous state classification and incremental model updates, thereby facilitating continuous, online islanding detection under dynamic grid scenarios.
[0065] In one embodiment, the Hoeffding Tree classifier module 120 features self-adaptive learning capability. The HT-PIDT model is configured to perform continuous learning from newly acquired labeled or semi-labeled data samples during operational runtime. The HT-PIDT model is further configured to dynamically adapt to variations in self-adaptive system’s topology and configuration, including but not limited to changes in the number of inverter-based resources, integration of new inverter models, and fluctuations in grid impedance characteristics. The HT-PIDT model is further configured to incrementally update its decision structure using the Hoeffding bound-based statistical criterion, which evaluates the sufficiency of observed data to justify a node split and ensures statistically significant structural modifications to the decision tree.
[0066] In the streaming inference mode, each incoming data instance or batch is processed through the model to generate a classification output representing the probability of an islanding event. A predefined confidence threshold (e.g., 𝑃(islanding) ≥ 0.9) is applied to filter high-confidence detections. To enhance reliability and mitigate transient misclassifications, a temporal consistency window comprising a predefined number of consecutive samples (e.g., 3 to 5) meeting the threshold condition is required to confirm a valid islanding event. Upon confirmation, the self-adaptive system 100 may initiate appropriate protective actions, including but not limited to triggering alarms, disconnecting the inverter(s), or isolating the microgrid segment from the main grid.
[0067] According to another exemplary embodiment of the invention, FIG. 3 refers to a single line diagram 300 of the inverter with the block diagram of the invented HT-PIDT model. The core classification engine of the self-adaptive system 100 comprises a Hoeffding Tree (HT) classifier, which is an incremental decision tree algorithm optimized for data stream processing and adaptive learning applications. Unlike conventional batch-trained decision trees, the HT classifier operates on sequentially observed data and does not require access to the entire historical training dataset. Instead, it employs a statistical decision mechanism based on the Hoeffding inequality to evaluate and determine the optimal feature for node splitting using a finite, bounded number of observations.
[0068] The HT classifier incrementally ingests feature vectors derived from real-time electrical measurements at the point of common coupling (PCC), dynamically updating sufficient statistics at each node, including class distributions and attribute histograms. These statistics are utilized to assess candidate features for node bifurcation without the need to retain or reprocess the full dataset. The decision to perform a split is governed by the Hoeffding bound, which estimates the minimum number of samples required to assert, with high confidence, that the observed difference in information gain between the top two candidate attributes is statistically significant. The Hoeffding bound is mathematically defined as:

[0069] where (𝐺) denotes the range of the information gain metric (e.g., Gini index or entropy), (δ) represents the desired confidence level, and (𝑛) is the number of samples observed at the decision node. A split is initiated when the observed difference in information gain between the best and second-best attributes exceeds the computed value of (ϵ), thereby ensuring statistically robust tree growth under streaming conditions.
[0070] In one embodiment, the Hoeffding Tree classifier module 120 employs continuous-valued features—such as Dynamic Frequency Change Rate (DFTR), Dynamic Voltage Change Rate (DVTR), Dynamic Active and Reactive Power Change Rates (DPTR, DQTR), Voltage-to-Frequency Change Rate Ratio (DVTFTR), Active and Reactive Power-to-Frequency Change Rate Ratios (DPTFTR, DQTFTR), Total Harmonic Distortion of voltage and current (THDv, THDi), and Voltage Unbalance Factor (VUF)—without applying discretization. The Hoeffding Tree algorithm implements continuous attribute split testing, wherein optimal threshold values for splitting each continuous feature are identified dynamically during online learning. These thresholds segment the feature space at each decision node into binary branches (e.g., DFTR ≤ 0.2 Hz/s), and are iteratively refined as additional data samples are observed, enabling continuous adaptation to evolving system dynamics.
[0071] Each decision node in the Hoeffding Tree maintains a set of sufficient statistics required for evaluating split criteria, including class label distributions and per-feature histograms. As the classification tree grows incrementally, it adapts efficiently to changing operating conditions, such as variations in load demand, the introduction or removal of inverter-based resources, and fluctuations in grid strength (e.g., short-circuit ratio). The lightweight, non-blocking architecture of the HT classifier is optimized for resource-constrained environments and is therefore well-suited for real-time deployment in embedded platforms such as smart inverters, edge computing devices, or local grid controllers.
[0072] The described architecture enables the HT-PIDT model to perform reliable and scalable islanding detection with low computational overhead. The system supports continuous self-training by integrating newly labeled operational data into the model during runtime, ensuring long-term adaptability and robust classification performance in dynamic microgrid environments.
[0073] In one embodiment, the proposed self-adaptive system 100 is evaluated using MATLAB Simulink by first simulating the test system 200 to generate and prepare the offline training dataset. The design and implementation of the Hoeffding Tree-based Passive Islanding Detection Technique (HT-PIDT) are performed outside the MATLAB environment due to the lack of native support for Hoeffding Trees within MATLAB. Consequently, the HT-PIDT model is developed and trained using the Python programming language.
[0074] Following the offline training phase, the trained HT-PIDT model is integrated into the MATLAB Simulink environment to enable real-time islanding detection. This integration is achieved via the MATLAB Engine API for Python, facilitating communication between the Python-based model and MATLAB Simulink.
[0075] In one embodiment, the Python source code for the HT-PIDT model design and offline training process is provided, assuming that the offline training dataset has been previously generated and exported from the MATLAB Simulink simulation. After successful installation and configuration of the MATLAB Engine, the HT-PIDT is trained using the offline dataset with the provided Python code.
[0076] Subsequently, the trained HT-PIDT model is employed for real-time islanding detection and adaptive self-training. The corresponding Python code is integrated with MATLAB Simulink through the MATLAB Engine interface, enabling continuous real-time operation and model updating within the simulation environment.
[0077] In one embodiment, to validate the practical feasibility of the proposed self-adaptive system 100, the proposed self-adaptive system 100 is deployed on a real-time embedded platform utilizing an STM32F7 microcontroller. The STM32F7 series is selected due to its high-performance ARM Cortex-M7 core, which provides the computational capability necessary for high-speed signal acquisition, digital signal processing, and low-latency inference. These features are critical to achieving real-time islanding detection within inverter-based microgrids. The implementation on the embedded platform comprises three primary functional components: real-time extraction of relevant features from sensor signals, execution of embedded inference using the pre-trained Hoeffding Tree decision model, and generation of system-level responses based on the inference outcomes to facilitate timely and accurate islanding detection.
[0078] The STM32F7 microcontroller acquires voltage and current signals at the point of common coupling (PCC) via its high-resolution analog-to-digital converter (ADC) channels. These signals are sampled at a frequency of 1 kHz and undergo preprocessing through digital filtering techniques to remove noise and improve signal quality. Subsequently, both time-domain and frequency-domain features are computed directly on the microcontroller in real time. The Derivative of Frequency at the Terminal (DFTR) is calculated by detecting zero-crossings of the voltage waveform and computing the derivative of the frequency over time. The Derivative of Voltage at the Terminal (DVTR), Derivative of Active Power at the Terminal (DPTR), and Derivative of Reactive Power at the Terminal (DQTR) are computed using finite difference methods applied to successive sampled measurements. Total Harmonic Distortion of voltage (THDv) and current (THDi) are estimated by performing real-time Fast Fourier Transform (FFT) analysis on the respective waveforms using the CMSIS-DSP software library optimized for ARM Cortex processors. The Voltage Unbalance Factor (VUF) is derived through symmetrical component analysis of the three-phase voltage inputs, quantifying the extent of voltage unbalance present at the PCC. Furthermore, feature ratios including Derivative of Voltage to Frequency Terminal Ratio (DVTFTR), Derivative of Active Power to Frequency Terminal Ratio (DPTFTR), and Derivative of Reactive Power to Frequency Terminal Ratio (DQTFTR) are computed to enhance detection sensitivity under weak grid conditions.
[0079] In one embodiment, the HT-PIDT model is initially trained offline using datasets generated from MATLAB/Simulink simulations and implemented in Python utilizing the River machine learning library. For deployment on the STM32F7 embedded platform, the trained HT-PIDT model is converted into a fixed-depth, rule-based decision tree expressed in embedded C code. Each decision node within the tree is implemented as a conditional if-else statement that evaluates one or more extracted features against predetermined thresholds derived during training. This streamlined model architecture ensures a memory-efficient inference engine capable of executing real-time classification with sub-millisecond latency, without reliance on floating-point hardware acceleration or external software libraries.
[0080] Upon classification of the system state as either islanded or grid-connected, the STM32F7 microcontroller initiates corresponding control actions. In response to detection of an islanding condition, the microcontroller can trigger a relay to disconnect the inverter from the grid, activate a fault-indicating LED via a general-purpose input/output (GPIO) pin, and/or transmit event status information over communication interfaces such as UART or CAN bus for higher-level system coordination and monitoring.
[0081] In one exemplary embodiment, the self-adaptive system 100 is evaluated using simulated feature data generated from MATLAB, alongside real-time signal emulation, to validate the classification performance of the HT-PIDT model. All processing and inference computations are executed within a fixed control loop interval of 10 milliseconds, thereby satisfying the sub-cycle islanding detection timing requirements as specified by the IEEE 1547 standard. Consistency between simulation and hardware implementations is confirmed by comparing the classification outputs of the embedded STM32F7 system with those of the Hoeffding Tree classifier module 120. The embedded implementation confirms that the proposed self-adaptive system 100 can be effectively deployed on cost-efficient microcontroller hardware, achieving high detection accuracy and rapid response times. Moreover, the system maintains compatibility with practical inverter-based distributed energy resource configurations, demonstrating its suitability for real-world applications.
[0082] In one embodiment, the Hoeffding Tree classifier module 120 of the self-adaptive system 100 is evaluated under two distinct scenarios: a multi-inverter-based resource (IBR) system as illustrated in FIG. 2, and a single-IBR test bed system. The real-time embedded test system 200 setup for the latter. The test system 200 utilizes the NI sbRIO GPIC (Single-Board RIO Grid Processor-in-the-Loop Controller) Evaluation Kit to manage control of the three-phase inverter. The HT-PIDT algorithm is implemented on the sbRIO-9606 FPGA controller board.
[0083] In both scenarios, the islanding detection performance of the Hoeffding Tree classifier module 120 is assessed over 200 test cases, incorporating variations in short-circuit ratio (SCR) from 1.5 to 3, as well as diverse load and power generation conditions under both balanced and unbalanced states. The average detection accuracy is computed and benchmarked against established methods, including support vector machine (SVM)-based and extreme learning machine (ELM)-based islanding detection techniques. Furthermore, the non-detection zone (NDZ) and detection time metrics are analyzed to comprehensively evaluate the effectiveness and reliability of the Hoeffding Tree classifier module 120.
[0084] According to another exemplary embodiment of the invention, FIG. 4 refers to a graph 400 depicting performance accuracy comparison of the self-adaptive system 100 with SVM and ELM based methods. The Hoeffding Tree classifier module 120 of the self-adaptive system 100 achieved an average accuracy of 99.2% across 200 test scenarios, with a detection time of 13.3 𝑚𝑠 for single IBR system, and average of 98.9% accuracy with 14.1 𝑚𝑠 of the detection time in the case of multi-IBR system. Precision and recall reached 98.7% and 99.1%, respectively, indicating minimal false positives and missed detections. The F1-score of 98.9% underscores balanced performance in distinguishing islanding events from normal grid fluctuations. Key to this success is the model’s ability to leverage transient features like DFTR and 𝑇𝐻𝐷𝑉, which exhibit pronounced deviations during islanding in weak grids (SCR < 3). FIG. 4 shows the average accuracy of the proposed islanded detection method compared to SVM and ELM based methods for both experimental and simulation cases.
[0085] Conventional passive methods, such as rate-of-change-of-frequency (ROCOF) and voltage unbalance, exhibited a large NDZ under power-balanced conditions. In contrast, the proposed method eliminates the NDZ, even when 𝑃𝐿𝑜𝑎𝑑 matched inverter output within 1%. This improvement stems from the ML model’s sensitivity to multi-feature interactions (e.g., simultaneous spikes in (DVTFTR, DPTFTR, DQTFTR and 𝑇𝐻𝐷𝑉) that single-parameter thresholds fail to capture.
[0086] According to another exemplary embodiment of the invention, FIG. 5 refers to a graph 500 depicting performance accuracy comparison of the self-adaptive system 100 with SVM and ELM under weak grid condition (SCR=1.5). The Hoeffding Tree classifier module 120 of the self-adaptive system 100 is tested across SCR values ranging from 1.5 (very weak) to 3 (moderately weak). Detection accuracy remained above 98.5% for SCR ≥ 2 but dropped slightly to 97.1% at 𝑆𝐶𝑅 = 1.5 due to increased harmonic distortions masking islanding signatures. However, even at SCR = 1.5, the model outperformed ROCOF-based methods, which achieved only 85.3% accuracy under the same conditions. FIG. 5 shows the performance accuracy comparison of the proposed islanded detection method with SVM and ELM based methods for both single and multi IBR systems under very weak grid scenario (SCR=1.5).
[0087] According to another exemplary embodiment of the invention, FIG. 6 refers to a flowchart 600 of a self-adaptive passive islanding detection method for inverter-based microgrids operating under weak grid conditions. At step 602, one or more measured electrical parameters is acquired by the parameter acquisition module 114 at a point of common coupling (PCC). At step 604, one or more dynamic electrical features are extracted by the feature extraction module 116 from the one or more measured electrical parameters.
[0088] At step 606, the one or more dynamic electrical features are normalized and filtered by the filter module 118 to produce a pre-processed feature stream. At step 608, an operating state as grid-connected or islanded is classified by the Hoeffding Tree classifier module 120 based on the pre-processed feature stream to obtain one or more classified samples, thereby generating an islanding detection signal if a temporal consistency condition of multiple consecutive classified samples is achieved.
[0089] Numerous advantages of the present disclosure may be apparent from the discussion above. In accordance with the present disclosure, a machine learning-based, self-adaptive system employing a Hoeffding Tree classifier for passive islanding detection in inverter-based distributed generation systems operating under weak grid conditions.
[0090] The self- adaptive system 100 is capable of continuously learning and adapting to changing grid and load conditions without the need for manual retraining. The self- adaptive system 100 ensures accurate and rapid detection of unintentional islanding events in inverter-based distributed generation (IBDG) systems. The self- adaptive system 100 eliminates non-detection zones (NDZs) commonly associated with conventional passive detection methods, especially under power-balanced conditions or low short-circuit ratio (SCR) scenarios.
[0091] The self- adaptive system 100 designs a non-intrusive, disturbance-free detection method that operates solely on real-time measurements without injecting external signals, thereby maintaining grid power quality. The self- adaptive system 100 enables compatibility with both grid-forming (GFMI) and grid-following (GFLI) inverters, ensuring wide applicability across diverse microgrid configurations.
[0092] The self- adaptive system 100 develops a decentralized and communication-free solution, avoiding the need for supervisory control and data acquisition (SCADA) systems or inter-inverter communication infrastructure. The self- adaptive system 100 supports efficient real-time implementation on embedded platforms, such as microcontrollers and FPGA systems, using a lightweight Hoeffding Tree classifier with minimal memory and computational overhead.
[0093] The self- adaptive system 100 ensures robust operation under weak grid conditions, with consistent performance in systems having SCR values as low as 1.5. The self- adaptive system 100 provides system scalability, allowing the islanding detection technique to be effectively deployed in microgrids with multiple inverter-based resources (multi-IBR systems) without degradation in performance. The self- adaptive system 100 provides an adaptive, feature-rich classification framework, using multi-dimensional features such as frequency and voltage change rates, harmonic distortion, and voltage unbalance, for improved sensitivity and reliability.
[0094] It will readily be apparent that numerous modifications and alterations can be made to the processes described in the foregoing examples without departing from the principles underlying the invention, and all such modifications and alterations are intended to be embraced by this application.
, Claims:CLAIMS:
I / We Claim:
1. A self-adaptive system (100) for passive islanding detection in inverter-based microgrids operating under weak grid conditions, comprising:
a computing device (102) having a processor (104) and a memory (106) for storing one or more instructions executable by the processor (104), wherein the computing device (102) is in communication with a server (110) via a network (112),
wherein the processor (104) is configured to execute plurality of modules (108), wherein the plurality of modules (108) comprises:
a parameter acquisition module (114) configured to receive one or more measured electrical parameters at a point of common coupling (PCC);
a feature extraction module (116) configured to extract one or more dynamic parameters from the one or more measured electrical parameters received through the parameter acquisition module (114);
a filter module (118) configured to normalize and pre-process the dynamic electrical features using statistical filtering processes to produce a pre-processed feature stream; and
a Hoeffding Tree classifier module (120) configured to classify an operating state of a microgrid as either grid-connected or islanded based on the pre-processed feature stream to obtain one or more classified samples, and to generate an islanding detection signal if a temporal consistency condition of multiple consecutive classified samples is achieved, thereby enabling real-time islanding detection without requiring human intervention.
2. The self-adaptive system (100) as claimed in claim 1, wherein the one or more measured electrical parameters comprise at least one of voltage, current, frequency, active power, and reactive power at the point of common coupling (PCC).
3. The self-adaptive system (100) as claimed in claim 1, wherein the one or more dynamic electrical features comprise at least one of dynamic frequency change rate (DFTR), dynamic voltage change rate (DVTR), dynamic active and reactive power change rates (DPTR, DQTR), dynamic voltage-to-frequency change rate ratio (DVTFTR), dynamic power-to-frequency change rate ratios (DPTFTR, DQTFTR), total harmonic distortion of voltage and current (THDv, THDi), and voltage unbalance factor (VUF).
4. The self-adaptive system (100) as claimed in claim 1, wherein the Hoeffding Tree classifier module (120) comprises a Hoeffding Tree (HT) classifier, wherein the HT classifier is initially trained offline using a training dataset simulated under multiple grid-forming inverter (GFMI) and grid-following inverter (GFLI) operational scenarios.
5. The self-adaptive system (100) as claimed in claim 1, wherein the Hoeffding Tree classifier module (120) is adapted to operate effectively under short-circuit ratio (SCR) conditions varies between 1.5 and 3.
6. The self-adaptive system (100) as claimed in claim 1, wherein the temporal consistency condition comprises confirming islanding status based on at least three to five consecutive samples classified as islanded with a probability threshold of at least 90%.
7. The self-adaptive system (100) as claimed in claim 1, wherein the pre-processed feature stream is generated by applying a moving average filter with a 10-sample window and min-max normalization in the range of at least 0 to 1.
8. The self-adaptive system (100) as claimed in claim 1, wherein the Hoeffding Tree classifier module (120) is configured to incrementally learn from the pre-processed feature stream using a Hoeffding bound-based statistical criterion for node splitting, and updates a decision structure in real time.
9. The self-adaptive system (100) as claimed in claim 1, wherein the Hoeffding Tree classifier module (120) is configured to perform a binary classification based on continuously updated statistics at each decision node without requiring storage of historical data.
10. A self-adaptive passive islanding detection method for inverter-based microgrids operating under weak grid conditions, comprising:
acquiring, by a parameter acquisition module (114), one or more measured electrical parameters at a point of common coupling (PCC);
extracting, by a feature extraction module (116), one or more dynamic electrical features from the one or more measured electrical parameters;
normalizing and filtering, by a filter module (118), the one or more dynamic electrical features to produce a pre-processed feature stream; and
classifying, by a Hoeffding Tree classifier module (120), an operating state of a microgrid as either grid-connected or islanded based on the pre-processed feature stream to obtain one or more classified samples, and to generate an islanding detection signal if a temporal consistency condition of multiple consecutive classified samples is achieved, thereby enabling real-time islanding detection without requiring human intervention.