A passive islanding detection method using deep neural bidirectional LSTM-CNN Authors: Hakan Ozturk, Secil Varbak Nese Published: 2026 ABSTRACT AbstractMicrogrids based on local generation principle has advantages as reducing power losses and improving voltage profile. One of the biggest operational concerns in presence of a microgrid connected to a power system is unintentional islanding. Unintentional islanding events have potential to cause equipment damage and endanger human safety. According to IEEE Std 1547–2018, unintentional islanding must be detected and controlled in 2s. Therefore, a novel One-Dimensional Convolutional Neural Network-Bidirectional Long Short-Term Memory (1D CNN-BiLSTM) method is introduced for islanding detection. In the proposed method, power system is firstly controlled and observed to select busbars that are sensitive to islanding, then selected critical busbars are used for measurement. The optimally selected current and voltage time series data are then processed by neural network. Simulation results on the IEC 61850-7-420 test system demonstrate that the introduced method achieves a detection time of 10ms and a Non-Detection Zone (NDZ) rate of only 0.02%. These results indicate that the method significantly outperforms existing techniques in terms of both detection speed and reliability.Keywords: Deep neural networks, Islanding detection, Microgrid, Distributed generationSubject terms: Energy science and technology, Engineering, Mathematics and computing Introduction Microgrids containing Distributed Generation (DG) combine the positive advantages of DG as environmentally friendly, energy security, reliability, stability with better management. So, such microgrids are becoming more increasingly integrated. The situation supplied portion of grid by local generation source is electrically isolated from utility is called as islanding. While islanding can made as intentional to ensure energy security and continuity, it can also occur as an unintentional that can cause risk consumer and utility. Microgrid Islanding Detection (ID) is more important in modern power systems, as a result of increased penetration. In this perspective, effective ID is a necessity for microgrid operation stability by preventing serious security risks and operational problems that may arise. Passive and active approaches are used for detection, each with its own advantages and disadvantages. While passive methods monitoring system characteristics, active approaches typically inject external signals into the system for ID. Numerous studies on ID methods have introduced to understand and improve of detection performance to the literature. Wang et al. proposed a frequency-reactive power feedback method that accelerates frequency deviation during islanding, enabling fast detection without dead zone and with little impact power quality1. Azim et al. introduced a passive method using a set of derived from voltage and current measurements and decision tree-based classifiers for ID2. Harikrishna et al. identified sequence-based phase angle and compared several approaches in hybrid microgrids for passive detection to improve accuracy during power imbalances3. Emphasizing the need of post-detection frequency management Eshraghi et al. concentrated on demand response strategies to control frequency in islanded microgrids4. Ghalavand et al. has also been used mathematical morphology as dilation and erosion filters to improve detection accuracy under varying conditions5. Kim et al. categorized ID methods, stressing the need of fast and reliable detection mechanisms in the context of increasing DG penetration6. Among the studies that attempt to effectively detect islanding, Roy et al. investigated a method to detect islanding for DC microgrid. The machine learning-based method using voltage parameters employs support vector machines obtained high accuracy even with small power imbalances7. Using a harmonic filter-based approach, Dhua et al. greatly improved detection accuracy reducing false detection during load switching8. Moreover, to reduce power quality issues with current injection methods Ganivada et al. proposed low frequency current injection during grid disturbance. The technique was successful effectively detects islanding conditions and minimize power quality disturbances9. Raju et al. assessed passive techniques for single line to ground fault detection in microgrids, proposed method has better detection time, lower and stable NDZ10. Xia et al. developed a novel algorithm using deep learning techniques with advanced computational methods to categorize islanding conditions with high accuracy11. Despite advances in ID methods, challenges remain. Minimal impact on power quality and fast detection remain the primary goals in this field. Furthermore, standards and regulations such as IEEE-1547-201812 influence the direction of research and applications. It is expected that integration of smart systems and machine learning algorithms will play an important role in improving ID in microgrids. Also, there are methods that use voltage at Point of Common Coupling (PCC) between grid and microgrid and/or voltage of DG for ID13,14. In another study, a deep learning method was applied for a microgrid that includes a diesel generator and a battery15. Neural networks with multiple Long Short-Term Memory (LSTM) units have been studied for a grid containing solar, wind and hydroelectric power plants16. Hussain et al. introduced a statistical feature-based Deep Neural Network (DNN) technique for ID hybrid systems with both synchronous and inverter-based DGs17. Furthermore, time based analyses and resonance-based fault detection techniques used for ID of microgrid18. However, decision trees are used for ID in microgrids. An ID method has been developed which is based on sensitivity analysis using transient kinetic energy of wind turbines and uses kernel density estimation to take into account cyberattack19. A decision tree based fuzzy logic method is proposed for fast ID for a system containing multiple microgrids20. A decision tree method for ID is developed combining noise-assisted ensemble empirical mode decomposition and a hybrid optimization based random forest classifier. In the method, binary grey wolf optimizer is also used to reduce computational burden21. Choudhury et al. showed that an AlexNet-based deep learning algorithm strengthens decision mechanism for microgrids22. Another method in the literature for ID is 1D CNN-LSTM based deep learning by using total harmonic distortion at the PCC23. Moreover, reinforcement learning and variational mode decomposition method are applied in microgrids for ID24. Consequently, ID is an area requiring improvement in modern power systems to ensure reliable and efficient operation of microgrids. In this study, a novel hybrid neural network architecture is proposed to address challenges of ID for Microgrids containing DGs. Differently previous studies in the literature, this study developed a 1D CNN-BiLSTM method for ID challenges. The proposed method utilizes both local feature extraction and temporal dependencies from sequential electrical signals. Furthermore, a comparative analysis of the method was conducted with a baseline LSTM network and a hybrid CNN-LSTM model. The method significantly outperformed existing techniques in terms of both detection speed and reliability. Despite the extensive use of deep learning in islanding detection, existing studies primarily focus on improving performance metrics using variations of established architectures. However, limited attention has been given to structured framework design, measurement selection, and physically grounded validation. To address these gaps, this study proposes an integrated islanding detection framework that combines model architecture, signal selection, and system-level validation. The main contributions of this work are summarized as follows: (i)First application of BiLSTM to islanding detection While prior deep learning-based ID studies employ unidirectional LSTM or CNN-LSTM architectures, which process temporal sequences in only one direction, this work is the first to incorporate a Bidirectional LSTM in the ID context. By processing each windowed segment in both forward and backward directions, the BiLSTM captures symmetric transient patterns around the islanding event that are inherently missed by causal-only models. First application of BiLSTM to islanding detection While prior deep learning-based ID studies employ unidirectional LSTM or CNN-LSTM architectures, which process temporal sequences in only one direction, this work is the first to incorporate a Bidirectional LSTM in the ID context. By processing each windowed segment in both forward and backward directions, the BiLSTM captures symmetric transient patterns around the islanding event that are inherently missed by causal-only models. (ii)Optimal measurement bus selection Rather than relying solely on PCC measurements as in most prior works, the proposed method introduces a sensitivity-based bus screening procedure that identifies the most islanding-responsive nodes in the network. The selected combination of Bus 1 (PCC) voltage and current and Bus 2 current provides richer discriminative information than single-bus approaches, without increasing sensor cost disproportionately. Optimal measurement bus selection Rather than relying solely on PCC measurements as in most prior works, the proposed method introduces a sensitivity-based bus screening procedure that identifies the most islanding-responsive nodes in the network. The selected combination of Bus 1 (PCC) voltage and current and Bus 2 current provides richer discriminative information than single-bus approaches, without increasing sensor cost disproportionately. (iii)Attention-augmented architecture An attention mechanism is integrated into the CNN-BiLSTM to allow the model to selectively focus on the most informative temporal segments within each input window. This is the first use of an attention layer in a deep learning-based islanding detection framework. Attention-augmented architecture An attention mechanism is integrated into the CNN-BiLSTM to allow the model to selectively focus on the most informative temporal segments within each input window. This is the first use of an attention layer in a deep learning-based islanding detection framework. (iv)Comprehensive multi-disturbance dataset on an IEC-compliant test system A dataset of 1761 scenarios covering islanding under varying power mismatch (ΔP ± 40%, ΔQ ± 20%) and five classes of non-islanding disturbances (LG, LLG, LLLG faults, load switching, and capacitor switching) is constructed on the IEC 61850-7-420 test system. This system uniquely combines both synchronous and inverter-based DG units, reflecting realistic heterogeneous microgrid conditions. Comprehensive multi-disturbance dataset on an IEC-compliant test system A dataset of 1761 scenarios covering islanding under varying power mismatch (ΔP ± 40%, ΔQ ± 20%) and five classes of non-islanding disturbances (LG, LLG, LLLG faults, load switching, and capacitor switching) is constructed on the IEC 61850-7-420 test system. This system uniquely combines both synchronous and inverter-based DG units, reflecting realistic heterogeneous microgrid conditions. (v)State-of-the-art detection performance The proposed method achieves a detection time of 10 ms and an NDZ of 0.02%, surpassing all methods reported in the manuscript (in literature from 2022 to 2025) on both metrics simultaneously. No prior study achieves sub-20 ms detection with an NDZ below 0.1% without relying on active signal injection. State-of-the-art detection performance The proposed method achieves a detection time of 10 ms and an NDZ of 0.02%, surpassing all methods reported in the manuscript (in literature from 2022 to 2025) on both metrics simultaneously. No prior study achieves sub-20 ms detection with an NDZ below 0.1% without relying on active signal injection. (vi)Validation contribution The effectiveness of the proposed method is validated using both the standard test system and an extended IEEE 33-bus distribution network, demonstrating strong generalization capability under both seen and unseen scenarios. Validation contribution The effectiveness of the proposed method is validated using both the standard test system and an extended IEEE 33-bus distribution network, demonstrating strong generalization capability under both seen and unseen scenarios. Materials and methods Long short-term memory (LSTM) LSTM networks are a specialized form of Recurrent Neural Network (RNN). LSTM introduced in the mid-1990s to overcome the difficulty of learning long-range temporal dependencies in sequence data25,26. Traditional RNNs suffer from vanishing or exploding gradients when trained by backpropagation through time, complicated to capture information over long intervals. LSTM cells were designed with an internal memory state and learnable gating mechanisms so that gradients can flow unchanged across many time steps26. The core of an LSTM network is its memory cell maintaining a hidden state vector Ct across time. A single LSTM cell contains three multiplicative gates (input, forget, and output) that regulate information flow into and out of the cell as shown in Fig. 1. The forget gate controls how much of previous cell state should be retained or discarded, input gate controls how much new information to add to cell, and output gate controls how much of cell state to output at this time. This gating structure allows cell to accumulate information additively in Ct, thereby preserving gradient flow (often referred to as constant error carousel). In other words, LSTM cell can carry information forward with minimal decay, while gates (via sigmoid activations) learn which parts of that information to keep. Fundamental architecture of an LSTM cell27. Formally, at each time step t, given input vector Xt and previous hidden state Ht−1, LSTM cell updates its gates and states as in (1)–(6). In Eqs. (1)–(6), each gating vector (Ft, It, Ot) and the candidate cell vector () is computed from an affine transformation of the concatenated inputs , followed by a nonlinearity. Here, is the logistic sigmoid function (mapping to [0,1]), and tanh() is the hyperbolic tangent (mapping to [− 1,1]). The symbol × denotes element-wise multiplication. Thus, each gate produces a vector of values in [0,1] that scale parts of cell state or candidate. This design—memory cell plus gated update—has been extensively analyzed and shown to be the standard (“vanilla”) LSTM architecture28. Bi-directional long short-term memory (Bi-LSTM) RNNs are designed to process sequential data maintaining a hidden state that captures information about past inputs. However, standard RNNs have limitations in capturing long-range dependencies and contextual information from both past and future inputs29. To address these limitations, Bi-LSTM networks were developed. Bi-LSTMs enhance capabilities of traditional LSTMs by processing input sequences in both forward and backward directions, allowing model to capture dependencies from both past and future contexts. This makes Bi-LSTMs particularly effective in tasks understanding the context surrounding an element in a sequence is crucial. Concretely, one LSTM layer reads sequence in normal (left-to-right) order, producing a sequence of forward hidden states , while another LSTM reads in reverse (right-to-left), producing backward states . These two sub-networks are independent, so forward and backward states do not feed into each other30. Each sub-network uses same gate Eqs. (1)–(6) but on its respective time indexing. Because backward LSTM uses future context relative to any position, a Bi-LSTM can “leverage . past and future” information simultaneously. At each time step t, the outputs of two LSTMs are combined. Typically, the forward and backward hidden vectors are concatenated to form Bi-LSTM output. In other words, if comes from forward pass and from backward pass (both computed via (6)) the Bi-LSTM output is a concatenated vector as (7)–(10). This doubling of dimension means each output carries information from both directions. In summary, a BiLSTM is “a specialized variant” of LSTM architecture consisting of two parallel LSTMs31,32. It thus captures bidirectional temporal dependencies by computing LSTM gate updates (forget, input, output, cell update) in both directions and combining the resulting hidden states at each time step. Test system IEC 61850-7-420 DG power system model in Matlab/Simulink is used to test the proposed method. The model consists of a multi-zone microgrid and utility grid, 7 Buses (B), 5 Distribution Lines (DL) and 5 Circuit Breakers (CB) as it can be seen in Fig. 233,34. The IEC 61850-7-420 test system for distributed energy management. The test system contains multiple DG units, as synchronous and inverter-based technologies. It reflects architecture of a multiple DG integrated microgrid test system coherent with standard IEC specifications. The presented model encompasses a Doubly-Fed Induction Generator (DFIG) based wind farm and a solar photovoltaic array, indicated as DG 1, 2 and DG 3, 4 respectively35,36. These are interconnected in a meshed configuration at PCC. The PCC is linked to a 120 kV utility grid through a 25 kV/120 kV step-up transformer. PCC is connected to 2 main zones and 1 Load (L). There are 6 loads and 5 transformers. Table 1 presents power values of DGs and loads in the test system. Specifically, DG 1 and DG 2 are synchronous generators as a combined apparent power 8 MVA, while DG 3 and DG 4 are inverter-based units as a total of 6 MVA, resulting in DG capacity of 14 MVA. On load side, six distinct consumer groups (L1–L6) are modeled. The total demand from all loads is 7 MW and 2.1 MVAr, as totally 7.31 MVA and an average power factor of approximately 0.96. These parameters ensure a realistic and balanced configuration for evaluating performance the proposed system. Table 1Power of DGs and loads.ComponentTypeP(MW)Q(MVAr)cosφS(MVA)DG 1 with 2Synchronous-DG---8DG 3 with 4Inverter based DG---6L1Load1,50,60,931,62L2Load1,50,30,981,53L3Load10,20,981,02L4Load1,50,60,931,62L5Load0,50,10,980,51L6Load10,30,961,04Total load-72,10,967,31Total DG OutputSynchronous+ Inverter---14 Power of DGs and loads. The proposed 1-D BILSTM-CNN Framework In the test system, measurements were taken from various buses to identify those particularly susceptible to islanding conditions. According to measurements results, Bus 1 (PCC) and Bus 2 were defined as critical buses. Voltage and current of PCC and only current of Bus 2 data were collected. The collected data was partitioned into training (40%), validation (30%), and test (30%) sets at the scenario level -prior to sliding-window expansion- to ensure complete separation between sets and eliminate any risk of data leakage between temporally correlated windows. The flow of the proposed CNN-BiLSTM-based ID framework is shown in Fig. 3. Flow of the proposed CNN-BiLSTM-based ID framework. Sensitivity-based bus selection method Consider a microgrid with buses, where three-phase voltage and current measurements are recorded at each bus. Let the total number of sampled time instants be N. The measured three-phase voltage data are stored in matrix , and the current data are stored in matrix . Each bus is represented by three consecutive columns corresponding to phase-a, phase-b, and phase-c quantities. To determine the buses most sensitive to islanding, the measured signals are divided into two operating regions: pre-islanding and post-islanding. Let denote the last sample before islanding. Therefore, samples correspond to the grid-connected mode, while samples correspond to the islanded mode. For the i-th bus (), the corresponding three-phase voltage and current signals are extracted. The average pre-islanding voltage vector of bus i is defined as in (11) and the average post-islanding voltage vector is in (12). Similarly, the average pre-islanding current vector is in (13) and the average post-islanding current vector is (14) The voltage sensitivity index of bus iis defined as is (15) and the current sensitivity index is in (16). The total islanding sensitivity score is then obtained as in (17). A higher value of indicates that the corresponding bus undergoes a stronger electrical change after islanding and is therefore more sensitive to the islanding event. All buses are ranked in descending order according to , and the buses with the highest values are selected as the most sensitive buses for monitoring. This systematic selection strategy ensures that the most informative measurement location is used, thereby improving the robustness and reliability of the proposed islanding detection method. As illustrated in the sensitivity ranking Fig. 4, the total islanding sensitivity of the six-bus microgrid was evaluated based on the aggregated response variations between pre- and post-islanding operating conditions, and the results reveal a clear differentiation among buses in terms of their dynamic response to the disturbance event. The analysis indicates that Bus 1 exhibits the highest sensitivity level, followed by Bus 2, while the remaining buses demonstrate comparatively lower sensitivity values, and since the figure represents only the total sensitivity index, these results reflect the overall system-level impact of each bus on islanding detection performance. Notably, Bus 1 corresponds to PCC, which inherently experiences the most pronounced dynamic variations during grid disconnection, and the elevated sensitivity observed at Bus 2 further confirms its strong involvement in post-disturbance system reconfiguration. Based on these findings, Bus 1 and Bus 2 are identified as the most influential locations for islanding detection and are therefore selected as priority monitoring points within the proposed framework to enhance the observability and reliability of islanding event detection. Sensitivity scores of buses with respect to islanding in the IEC 61850-7-420 test system. The architecture integrates 1-D CNN with BiLSTM layers, thereby exploiting both local feature extraction and temporal dependencies from sequential electrical signals. This framework is specifically designed to classify dynamic patterns that characterize islanding and non-islanding conditions with high temporal precision as shown in Fig. 5. The basic graphical structure of islanding classification by CNN-BiLSTM model. The model input is a three-dimensional sequence tensor in size 9 × 1 × 20, where 9 is number of extracted features (channels), 1 is batch dimension, and 20 is time steps in each windowed segment. The network initiates with a 1-D convolutional layer (conv1) that applies 64 filters in size 3 × 9, capturing local temporal patterns across all features. Batch normalization (bn1) and a ReLU activation (relu1) follow to stabilize and introduce nonlinearity, respectively. Subsequent max pooling (pool1) reduces temporal resolution, and dropout (drop1) prevents overfitting. Second convolutional block (conv2), similarly to the conv1, has one more additional convolutional layer for feature expansion with 128 channels filters (5 × 64 pad), normalization (bn2), ReLU (relu2) and then pooling (pool2), dropout (drop2). Convolutional stage 3 (conv3) applies 256 filters in size 7 × 128 for more sophisticated features-in detail normalization (bn3), ReLU (relu3), pooling (pool3) and dropout (drop3) provide output tensor size of 256 × 1 × 2. The extracted feature sequences are passed through two stacked BiLSTM layers. The first BiLSTM layer (bilstm1) processes the time series bidirectionally, producing a 512-dimensional context-aware sequence output. A dropout layer (drop_bilstm1) is applied to reduce overfitting, followed by a second BiLSTM layer (bilstm2) that reduces the representation to a 256-dimensional output vector. These layers enable the model to retain information from both past and future contexts in the sequence, essential for accurately modeling transient grid behaviors such as those occurring during islanding. Result flattening and attention mechanism [a fully connected layer (fc_attn) with a tanh activation function and dropout (drop_attn)] is applied to extract final feature sequence (this is how it is handled temporal data). Used here for attention mechanism that test model to learn focus on temporally most informative features before classification. Output layer (FC), finally, followed by softmax activation, classifying an islanding/non islanding state in binary form. The proposed 1-D BILSTM-CNN framework thus combines hierarchical feature extraction via CNN with sequence modeling via BiLSTM and incorporates an attention mechanism to enhance discriminative learning. The modular and deep structure of this network facilitates the detection of slight anomalies under various power mismatch conditions, including challenging cases with zero power imbalance. Each input segment consists of multivariate measurements. Each input vector at time step t is defined in (18). Here, V1a(t), V1b(t), and V1c(t) is PCC voltages, while I1a(t), I1b(t), and I1c(t) are PCC currents, and I2a(t), I2b(t), I2c(t) are Bus 2 currents. These are segmented using a sliding window of T = 20 time steps, resulting in an input tensor . A sequence of one-dimensional convolutional layers is applied to extract local temporal patterns. The output of a convolutional layer l is computed by (19)37. Where W(l) and b(l) are the filter weights and biases, ∗ is convolution operation, BN(·) denotes batch normalization, and ReLU(·) is the rectified linear unit activation function. Commonly used in deep learning architecture because of its easy computation and means of gradient propagation, Rectified Linear Unit or ReLU. Suppresses negative input values but leave unchanged positive ones, therefore providing nonlinearity. The ReLU function may be defined as in (20) mathematical form38. In here, x is the input of activation function. The output is zero for all negative inputs and linearly increases for positive values. This property not only mitigates the vanishing gradient problem associated with sigmoid or tanh functions but also accelerates convergence during training. However, ReLU units can sometimes suffer from the so-called “dying ReLU” problem, in which a neuron may become inactive if it only outputs zeros, typically caused by large negative weights or biases. Batch normalization is applied to stabilize and accelerate training by normalizing input distribution of each layer. For a mini-batch input {x1, x2,., xm}, normalized output is given by (21)39. Here and are batch mean and variance, ε is a small constant for numerical stability, and γ, β are learnable parameters. The pooling operation reduces spatial dimensions of input feature maps while retaining the most significant information. In the case of max pooling with a window size of k×k, the output at location (i, j) is computed as in (22)40. In here, Ωi, j denotes k×k pooling region centered at (i, j) in input X. To reduce overfitting, dropout is applied after convolution and LSTM blocks. During training, each unit is retained with a probability of 1 − p. The dropout operation is defined by (23)41. Where r is a random binary mask and ⊙ is element-wise product. Temporal dependencies are learned using BiLSTM layers. At each time step t, BiLSTM outputs concatenation of forward and backward hidden states as in (24)30. The internal operation of each LSTM unit is described previously. The flatten operation transforms input tensor from a multi-dimensional array into a one-dimensional vector, which is required before passing data to fully connected layers. Given an input tensor , the flatten operation reshapes it into a vector by (25). To enhance focus of model on critical temporal features, final LSTM output is passed through a fully connected layer with a non-linear transformation by (26)42. The final class probabilities are obtained using a softmax activation applied to the last fully connected layer as in (27)43. In here C is total number of output classes and denotes logit score for class i. To quantify prediction error in binary classification tasks such as ID, the Binary Cross-Entropy (BCE) loss is employed. Given the ground truth label and the predicted probability for each sample i, the loss function is defined by (28)44. This loss function penalizes predictions based on their deviation from actual labels, encouraging model to output probabilities that reflect true class distributions accurately. And finally, layer-wise architecture overview of the proposed CNN-BiLSTM model also provides an important reference for layer-wise architectural design as detailed in Table 2. In the table Activations column represents output tensor dimensions of each layer, where C, B, and T denote the number of channels, batch size, and temporal length, respectively. The Learnable Properties column describes trainable parameters such as convolutional weights and biases, as well as batch normalization parameters including scale (γ), offset (β), and the learned running mean and variance. The States column refers to internal states maintained during training and inference, particularly for recurrent layers such as BiLSTM, including hidden states and cell states. Non-parametric layers (e.g., ReLU, pooling, dropout) do not contain learnable parameters or internal states, and are therefore marked with “–”. The model consists of three convolutional layers, followed by two BiLSTM layers, a flattening stage, an attention mechanism, and a final classification layer. The convolutional layers are responsible for extracting local temporal features, whereas the BiLSTM layers capture bidirectional temporal dependencies within the sequential data. The attention mechanism enables the model to focus on the most informative temporal features prior to classification. Table 2Layer-wise architecture overview of CNN-BiLSTM Model.NameTypeActivationsLearnable PropertiesStates1inputSequence Input9(C) × 1(B) × 20(T)--2conv11-D Convolution64(C) × 1(B) × 20(T)Weights 3 × 9 × 64 Bias 1 × 64-3bn1Batch Normalization64(C) × 1(B) × 20(T)Offset 64 × 1 Scale 64 × 1Trained Mean 64 × 1 Trained Variance 64 × 14relu1ReLU64(C) × 1(B) × 20(T)--5pool11-D Max Pooling64(C) × 1(B) × 10(T)--6drop1Dropout64(C) × 1(B) × 10(T)--7conv21-D Convolution128(C) × 1(B) × 10(T)Weights 5 × 64 × 128 Bias 1 × 128-8bn2Batch Normalization128(C) × 1(B) × 10(T)Offset 128 × 1Scale 128 × 1Trained Mean 128 × 1 Trained Variance 128 × 19relu2ReLU128(C) × 1(B) × 10(T)--10pool21-D Max Pooling128(C) × 1(B) × 5(T)--11drop2Dropout128(C) × 1(B) × 5(T)--12conv31-D Convolution256(C) × 1(B) × 5(T)Weights 7 × 128 × 256 Bias 1 × 256-13bn3Batch Normalization256(C) × 1(B) × 5(T)Offset 256 × 1 Scale 256 × 1Trained Mean 256 × 1 Trained Variance 256 × 114relu3ReLU256(C) × 1(B) × 5(T)--15pool31-D Max Pooling256(C) × 1(B) × 2(T)--16drop3Dropout256(C) × 1(B) × 2(T)--17bilstm1BiLSTM512(C) × 1(B) × 2(T)Input Weights 2048 × 256 Recurrent Weights 2048 × 256 Bias 2048 × 1HiddenState 512 × 1 CellState 512 × 118drop_bilstm1Dropout512(C) × 1(B) × 2(T)--19bilstm2BiLSTM256(C) × 1(B)Input Weights 1024 × 512 Recurrent Weights 1024 × 128 Bias 1024 × 1HiddenState 256 × 1 CellState 256 × 120flattenFlatten256(C) × 1(B)--21fc_attnFully Connected128(C) × 1(B)Weights 128 × 256 Bias 128 × 1-22tanh_attnTanh128(C) × 1(B)--23drop_attnDropout128(C) × 1(B)--24fcFully Connected2(C) × 1(B)Weights 2 × 128 Bias 2 × 1-25softmaxSoftmax2(C) × 1(B)--26outputClassification Out2(C) × 1(B)-- Layer-wise architecture overview of CNN-BiLSTM Model. The detection latency of the proposed method is directly governed by the sliding window length used as the model input. In the Matlab/Simulink environment, voltage and current signals are sampled at a fixed step size of ms, corresponding to a sampling frequency of 2 kHz. Each input window spans T=20 consecutive time steps, yielding a window duration of as in (29). The model processes each incoming window and produces a softmax probability vector for two classes (islanding / non-islanding). A binary detection decision is triggered when predicted islanding probability exceeds a fixed classification threshold of , which corresponds to standard argmax decision rule of softmax output layer. No additional delay filter or confirmation timer is applied. Consequently, minimum achievable detection time equals window duration, ms, provided the islanding event generates a sufficiently discriminative pattern within that window. A potential concern when employing BiLSTM architectures in protection systems is their reliance on bidirectional processing, which is often interpreted as requiring access to future data. In the proposed framework, this issue is addressed through a segment-level (block-based) processing strategy. Instead of processing the signal in a fully streaming manner, the model operates on fixed-length windows of T = 20 time steps. Each window is fully collected before inference is performed. At the moment of classification, all samples within the window are already available; therefore, the backward LSTM pass operates only on previously acquired data within the closed window. From the perspective of the deployed system, no future information beyond the current observation window is required. Thus, the model behaves causally at the system level, while utilizing bidirectional processing internally within each window to enhance temporal feature extraction. The only latency introduced by this approach corresponds to the window duration which remains well within the 2-second detection requirement defined in IEEE Std 1547–2018. This mechanism is conceptually similar to frame-based processing used in conventional relay protection algorithms, where decisions are made after acquiring a finite data window. Training configuration and optimization strategy The training process of the proposed deep learning model was configured using the Adam optimization algorithm to ensure efficient and stable convergence. A bidirectional LSTM structure with 256 hidden units was employed to effectively capture temporal dependencies in the sequential data, while a dropout rate of 0.3 was introduced to reduce the risk of overfitting and improve generalization performance. The network was trained for a maximum of 250 epochs using a mini-batch size of 8 samples. The initial learning rate was set to 1 × 10− 2, and a piecewise learning rate scheduling strategy was adopted to progressively refine the optimization process. Specifically, the learning rate was multiplied by a factor of 0.3 every 5 epochs, allowing rapid learning during early stages and more precise weight updates in later epochs. In addition, an L2 regularization coefficient of 1 × 10− 4 was applied to further enhance model robustness. To improve statistical reliability, the training data were randomly shuffled at each epoch. Model validation was performed periodically using the validation dataset at intervals of every 10 iterations. Furthermore, an early stopping criterion was implemented through a custom output function, terminating the training procedure once the target classification accuracy of 99.8% was achieved. All simulations and training operations were executed in a CPU-based computational environment. These training strategies collectively ensure stable convergence while reducing the risk of overfitting and improving generalization capability. Dataset generation and preprocessing The datasets used in this study were generated using MATLAB/Simulink-based dynamic simulations performed on the IEC benchmark system and the IEEE 33-bus distribution system. Various operating conditions, including islanding and non-islanding scenarios, were simulated under different load levels and distributed generation operating conditions to improve dataset diversity and model generalization capability. The IEC benchmark system implementation and certain CNN-based modeling approaches were developed with reference to previously published open-source studies, which are appropriately cited in this manuscript. In addition, the IEEE 33-bus distribution system model was adapted from a publicly available implementation distributed under the BSD 3-Clause License. Voltage and current measurements were collected from buses identified as sensitive through preliminary sensitivity analysis. The generated signals were automatically labeled according to the operational state of the system during simulation. Since the datasets were entirely simulation-based, no human annotation process was involved. For the IEEE 33-bus system dataset, the processed data were stored in the MATLAB variable named “uu”, where the first six columns represent three-phase voltage and current measurements, and the seventh column represents the islanding status label (0 for non-islanding and 1 for islanding). For the IEC benchmark system dataset, the processed data were stored in the MATLAB variable named “uu”, where the first nine columns represent three-phase voltage and current measurements, and the tenth column represents the islanding status label. Duplicate data checks and preprocessing procedures were performed before training and testing to eliminate repeated simulation records and improve dataset consistency. Data availability The datasets generated and analyzed during the current study are publicly available in the Zenodo repository under the MIT License at: https://zenodo.org/records/20286297. The repository includes: (i)MATLAB/Simulink models of the IEC benchmark system and IEEE 33-bus distribution system, MATLAB/Simulink models of the IEC benchmark system and IEEE 33-bus distribution system, (ii)raw simulation files used for data generation, raw simulation files used for data generation, (iii)processed datasets used for deep learning training and testing, processed datasets used for deep learning training and testing, (iv)sensitive bus analysis scripts, and. sensitive bus analysis scripts, and. (v)custom MATLAB codes used in the study. custom MATLAB codes used in the study. The datasets are stored in MATLAB format. The processed datasets contain three-phase voltage and current measurements together with binary islanding labels. Code availability The custom MATLAB scripts, Simulink models, and deep learning implementation used in this study are publicly available in the Zenodo repository under the MIT License at: https://zenodo.org/records/20286297. The repository includes the CNN-BiLSTM implementation, dataset generation scripts, sensitive bus identification algorithms, and power system simulation models used to reproduce the main findings of this work. Performance evaluation Compared neural architectures Two different neural network architectures, a baseline LSTM network and a hybrid CNN-LSTM model, are developed and compared to evaluate the effectiveness of temporal and spatiotemporal modeling for ID. Each model is trained using same input data and classification targets to ensure an accurate comparison. The LSTM-based architecture is designed to capture long-range temporal dependencies in sequential input data. It consists of a single LSTM layer with a reduced number of hidden units to produce a compact representation using the ’last’ output mode, followed by a fully connected layer, a softmax classifier, and a classification output layer. This model is particularly suitable for modeling pure temporal patterns without spatial convolution. In contrast, CNN-LSTM architecture integrates convolutional layers before LSTM block to extract higher-level spatial features from raw input signals. The network includes three successive 1-D convolutional layers, each followed by batch normalization, ReLU activation, and max pooling. These layers progressively downsample and transform input sequence, enabling network to capture localized temporal patterns. The final convolutional output is then passed to an LSTM layer, which summarizes temporal dynamics of learned features. As in LSTM-only model, output is processed through a fully connected layer, softmax layer, and classification output. An extensive dataset was constructed using simulation-based data mining techniques to ensure effectiveness of proposed ID algorithm. The dataset comprises a total of 1761 scenarios, divided into islanding and non-islanding conditions. As shown in Table 3, each scenario designed to represent realistic operational disturbances in a distribution network. Table 3Feature extraction and dataset structuring for ID.ScenarioDisturbance typeLocationVariation / RangeCasesPhysical BasisIslandingPower Mismatch by changing ∆P and ∆Q-∆P (-40/+40%) & ∆Q (-20/+20%) Variation (2% Increment)861Non-IslandingSingle Line-Ground FaultBus 1 and 2Changing fault resistance from (0.1–100 ohm)200Double Line-Ground FaultBus 1 and 2Changing fault resistance from (0.1–100 ohm)200Triple Line-Ground FaultBus 1 and 2Changing fault resistance from (0.1–100 ohm)200Load SwitchingLoad 3Range (100 kVA − 35 MVA)50Grid supported; PCC adjacent zone stress testLoad 4Range (20 kVA − 600 kVA)50Within local DG capacity boundsLoad 5Range (500 kVA − 100 MVA)50Utility infeed supported worst-case transientCapacitor SwitchingBus 3Range (100 kVA − 35 MVA)50Substation level bank switchingBus 4Range (5 kVA − 200 kVA)50Load bus PF correctionBus 5Range (1 MVA − 10 MVA)50Feeder Level compensationTotal1761 Feature extraction and dataset structuring for ID. The disturbance ranges in Table 3 were selected based on two complementary principles: physical plausibility within the grid-connected test system and worst-case stress coverage for classifier robustness. For load switching scenarios, it is important to note that the test system operates in grid-connected mode, with the PCC linked to a 120 kV utility grid through a 25 kV/120 kV step-up transformer. In this configuration, the short-circuit power of the upstream utility network is on the order of several hundred MVA, making it physically capable of supplying large transient load steps without causing system collapse. Load switching events of up to several tens of MVA therefore represent credible, grid-supported disturbances from the perspective of the PCC measurement point — even though they exceed the local DG capacity of 14 MVA. These large-step events are particularly valuable for training purposes because they generate the most aggressive voltage and current transients among non-islanding disturbances, constituting the hardest cases for the classifier to distinguish from islanding. Specifically, the three load switching locations were assigned different ranges to reflect their distinct positions in the network topology and their associated short-circuit levels. Load 3 (nominal 1 MVA) covers 100 kVA – 35 MVA to stress the PCC-adjacent zone; Load 4 (nominal 1.62 MVA) covers 20 kVA – 600 kVA within a physically bounded range; and Load 5 (nominal 0.51 MVA) covers 500 kVA – 100 MVA to represent a feeder-level large switching event supported by the utility infeed. While the 100 MVA upper bound for Load 5 is indeed far beyond the local generation capacity, it is not intended to represent a stand-alone islanded switching event but rather a grid-connected worst-case transient that, in a real distribution system, would be fed primarily by the upstream network. For capacitor switching, the ranges span 5 kVA to 35 MVA across three buses. These reflect the wide range of compensation equipment that may be present in a multi-zone distribution system, from small power factor correction capacitors at load buses to larger substation-level bank switching events visible at the PCC. This broad coverage is a deliberate design choice: a classifier trained only on mild disturbances would be poorly equipped to distinguish islanding from aggressive non-islanding events. By including extreme cases, the model learns to focus on the distinctive temporal patterns of islanding rather than relying on simple amplitude thresholds. For islanding conditions, 861 cases were generated by systematically varying power mismatch between DGs and local loads. Specifically, active power (∆P) was altered in range of ± 40%, and also reactive power (∆Q) in range of ± 20%; with 2% increments. The other 900 cases represent non-islanding events, covering a wide variety of disturbances: Fault scenarios (600 cases) were simulated by applying LG, LLG, and LLLG faults at Buses 1 and 2, with fault resistances in range 0.1–100Ω. Fault scenarios (600 cases) were simulated by applying LG, LLG, and LLLG faults at Buses 1 and 2, with fault resistances in range 0.1–100Ω. Load switching events (150 cases) were performed at Load points 3, 4, and 5, with loading capacities ranging from 100 kVA to 100 MVA. Load switching events (150 cases) were performed at Load points 3, 4, and 5, with loading capacities ranging from 100 kVA to 100 MVA. Capacitor switching events (150 cases) were generated switching capacitive elements ranging 5kVA-10MVA at Buses 3, 4, and 5. Capacitor switching events (150 cases) were generated switching capacitive elements ranging 5kVA-10MVA at Buses 3, 4, and 5. Approximately 50 h of simulations were carried out in the Matlab/Simulink environment to generate a comprehensive dataset for analysis. This comprehensive scenario modeling supports training and evaluation of machine learning models by ensuring dataset. Dataset expansion and leakage-free split protocol Each of the 1761 simulated scenarios produces a continuous multivariate time series of approximately 0.75 s duration with a sampling interval of 0.5 ms. To construct the dataset, the signals are segmented into fixed-length windows of T = 20 time steps. The segmentation is performed such that each raw sample is assigned to only one window, resulting in approximately 75 windowed segments per scenario and a total dataset of approximately 132,000 windows across all scenarios. To prevent data leakage -a critical concern in time-series analysis- the train/validation/test split is performed exclusively at the scenario level, prior to any windowing operation. Specifically, the 1761 scenarios are partitioned into training (40%, n = 705), validation (30%, n = 528), and test test (30%, n = 528) subsets using stratified random sampling to preserve the islanding/non-islanding class distribution. Windowed segments are then generated independently within each subset. This protocol guarantees that no segment from a test scenario appears in the training or validation sets, even partially, ensuring a leakage-free evaluation. The resulting dataset sizes are summarized in Table 4. Table 4Dataset expansion summary after scenario-level split.SubsetScenariosNumber of windows (approximately)Islanding windowNon-islanding windowTraining705 (40%)~ 52,833~ 16,011~ 36,822Validation528 (30%)~ 39,625~ 12,284~ 27,341Test528 (30%)~ 39,624~ 12,093~ 27,531Total1761~ 132,082~ 40,388~ 91,694 Dataset expansion summary after scenario-level split. The total test set size of approximately 39,600 windows is consistent with the scenario-level partitioning, as 528 test scenarios yield approximately 39,624 windowed segments. This confirms the internal consistency and integrity of the evaluation protocol. Evaluation metrics Multiple statistical metrics that capture different aspects of prediction quality are considered to comprehensively evaluate a classification model performance. Among the most commonly used are sensitivity, specificity, and the F1-score45. Sensitivity, also known as recall or true positive rate, measures proportion of actual positive cases that are correctly identified by the model. It reflects ability of the model to detect positive instances and is particularly important in classification situation in which islanding takes place. Mathematically, it is defined by (30)40. Here, TP is number of true positives, and FN is number of false negatives. Specificity, also referred to as true negative rate or dependability, is proportion of actual negative cases that are correctly predicted. It is useful in assessing the model capacity to avoid false alarms. The formula is (31)40. Here, TN represents true negatives and FP false positives. F1-score is a harmonic mean of precision and recall, providing a balanced measure that accounts for both false positives and false negatives. It is especially useful when data is imbalanced. The F1-score is given by (32)45. Results Case studies Zero power mismatch In microgrid islanding, the zero power mismatch condition is a scenario immediately after grid disconnection when local generation exactly equals local demand, so that there is virtually no net power flow with the utility46. In this worst-case, voltage and frequency at the PCC change negligibly, so passive islanding detectors (which monitor such parameters) unable to sense transition. The NDZ is defined as the set of load-generation imbalance conditions (active/reactive power mismatch) under which an islanding event may not be detected47,48. Therefore, the zero-mismatch case is at the center of NDZ for many passive methods. This is a critical safety issue. For secure and reliable operation, anti-islanding schemes are designed to minimize NDZ so that even zero mismatches detection. Figure 6 shows performance of the proposed ID model under zero power mismatch condition. In this challenging case, the islanding event is intentionally triggered at 0.5 s, as indicated by vertical dashed red line on the first subplot. The model responds with a clear change in detection signal (blue line) in 10 ms, as denoted by horizontal arrow. This fast response shows high sensitivity and responsiveness of the model, even when conventional electrical patterns such as voltage and frequency exhibit minimal disturbances. Islanding classification of the model and corresponding PCC Voltage, Frequency, and ROCOF time series for zero power mismatch case. The voltage subplot displays the system voltage profile in per-unit values. As expected under zero power mismatch conditions, the voltage remains relatively stable and does not show any instantaneous deviations post-islanding. This lack of voltage disturbance confirms difficulty of using purely threshold-based methods for detection in such cases. Also, the frequency subplot shows similar response, that remains in a narrow band around 50 Hz, with only a slight decline following the islanding point. Such small changes are typically insufficient to trigger conventional detection schemes. The Rate of Change of Frequency (ROCOF) subplot is a more dynamic indicator showing a noticeable transition shortly after the islanding event. As seen the model successfully uses ROCOF temporal variations to detect change in system dynamics, despite their low amplitude. The sudden drop in ROCOF following 0.5 s, although not impressive, alines closely with detection signal change in the first subplot. Double line to ground fault Misclassifying a double Line-to-Ground (LLG) fault as islanding can delay proper fault-clearing and allow fault currents to continue. Therefore, risks such as damaging equipment, igniting fires, and endangering maintenance personnel arise. In this case study, fault resistance as 110 Ω was chosen, a value outside LLG fault resistances used to train neural network. A LLG fault was initiated at 0.3 s, assuming a critical clearing time as 0.1 s. As shown in Fig. 7, the proposed model did not issue a false islanding alarm, even for a fault resistance value it was not trained. This indicates the generalization capability of the model. Furthermore, the model accurately predicts the islanding, which occurs at 0.45 s, with a delay of only 10 ms. Islanding classification of the model and corresponding PCC Voltage, Frequency, and ROCOF time series for LLG fault case. Time-varying noise interval Figure 8 shows response of the model, bus voltages, and line currents at Bus 1 and Bus 2 under a time-segmented noise injection scenario during 0.75 s. The simulation time is divided to three equal segments of 0.25 s, during which different additional noise as 40 dB, 25 dB, and 15 dB SNR were sequentially injected to the system. As the signal-to-noise ratio decreases over time, noticeable waveform distortions are observed, particularly in current profiles of both buses and voltage wave-forms. These distortions become clearer in the final segment (15 dB), highlighting effect of increasing noise on signal integrity and system observability. Despite rising noise intensity, the model demonstrates a degree of robustness, maintaining a reasonably stable output in the early segments, with observable deviations appearing under severe noise conditions. This behavior highlighting the relevance of evaluating AI-based detection models under varying levels of signal degradation to ensure their resilience in real-world noisy environments. Islanding classification of the model, voltages and currents under time-varying noise levels. Evaluation of the proposed method The test errors histogram of the model is presented in Fig. 9. The errors distribution is concentrated around zero with minimal deviation from ideal zero error (orange line), indicating high accuracy and reliability of model. Also, it shows precision learning capability and robust generalization performance of the model when applied to unknown data. Histogram of classification errors in test phase. The confusion matrices of training were created to evaluate effectiveness of the proposed deep learning architectures, but are not presented here for reasons of visual simplicity. For training, LSTM model has 99.6% and 97.9% correct classification rate for classes 0 and 1, respectively. However, it has a relatively higher misclassification rate for class 1 as 2.1%. The CNN-LSTM model has higher accuracy rates, at 99.3% and 98.1%, respectively. However, the CNN-BiLSTM model clearly outperforms the other two, with rates 99.9% for class 1 and 99.6% for class 0 for training. In this regard Fig. 10 qualitatively compares three architectures using confusion matrices for test phase under microgrid environments. The confusion matrix of LSTM model (Fig. 10a) has a great success in accuracy, as it gets 27,529 non-islanding and 11,732 islanding with just a few wrongly classified. Nevertheless, there is a bit more false negatives in ID that hint at some kind of sensitivity limitation for sequential-only modeling. The matrix of CNN-LSTM (Fig. 10b) shows better performance with fewer false classification. However, CNN-BiLSTM model performance (Fig. 10c) has near perfect classification of 27,808 non-islanding and 11,712 islanding events with minor misclassification. The addition of bi-directional LSTM layers provide a good temporal patterns recognition thus CNN-BiLSTM model is suitable for real-time ID as precise decisions are a must. Performance comparison of LSTM (a), CNN-LSTM (b), and CNN-BiLSTM (c) models for ID in Microgrids. For convergence and generalization of three models, training and validation loss and accuracy curves of each model trained for over 600 epochs was recorded as in Fig. 11. The LSTM model (Fig. 11a) exhibits relatively stable training in loss and accuracy, achieving convergence after approximately 400 epochs. However, the CNN-LSTM model (Fig. 11b) has more consistent performance with smoother validation curves and a rapid decline in loss at early stages of training, indicating faster learning and better generalization. Notably, the CNN-BiLSTM architecture (Fig. 11c) achieves the most rapid convergence and the lowest final loss values, while maintaining consistently high training and validation accuracy. This suggests superior learning capacity and reduced overfitting due to ability of bidirectional LSTM layer to capture temporal dependencies in both forward and backward directions. Training and validation loss and accuracy curves for LSTM (a), CNN-LSTM (b), and CNN-BiLSTM (c) models over epochs. The performance of three deep learning-based architectures is investigated using key evaluation metrics: F1-score, sensitivity, and specificity. As shown in Fig. 12, the CNN-BiLSTM model outperformed the other models, achieving the highest F1-score (0.9978), sensitivity (0.9989), and specificity (0.9968). The CNN-LSTM model followed closely with an F1-score of 0.9875, sensitivity of 0.9815, and specificity of 0.9935. Meanwhile, the LSTM model achieved comparable performance, with an F1-score of 0.9880, sensitivity of 0.9811, and specificity of 0.9950. These findings underscore advantage of the CNN-BiLSTM model, for robust and accurate ID in complex microgrid environments. Comparative evaluation of LSTM, CNN-LSTM, and CNN-BiLSTM models based on sensitivity, specificity, and F1-score metrics. Mechanistic interpretation of model performance The observed performance differences between the evaluated architectures can be explained by their ability to capture the temporal structure of islanding transients. Islanding events are characterized by short-duration, non-stationary transients whose exact position within the observation window depends on the switching instant. As a result, an effective detection model must not only capture local transient features but also interpret their temporal context relative to both pre- and post-disturbance signal behavior. The LSTM-only model processes the input sequence in a unidirectional manner, accumulating temporal information progressively. This structure limits its responsiveness to transient events occurring at arbitrary positions within the input window, as reliable representation can only be formed after the transient has fully propagated through the sequence. This behavior explains its relatively higher misclassification rate. The CNN-LSTM architecture improves upon this limitation by introducing convolutional layers that extract localized temporal features prior to sequence modeling. These layers act as feature compressors, transforming raw signals into higher-level representations that emphasize transient patterns. As a result, the LSTM receives a more informative input, leading to faster convergence and improved classification accuracy. However, the sequence modeling stage in CNN-LSTM remains unidirectional, which restricts its ability to fully exploit the temporal symmetry inherent in islanding events. The CNN-BiLSTM model overcomes this limitation by employing bidirectional temporal processing. Within each input window, the forward pass captures the evolution of the signal leading up to the transient, while the backward pass encodes the post-transient settling behavior. This dual perspective enables the model to capture subtle asymmetries between pre- and post-islanding signal patterns, which are particularly critical under low power mismatch conditions. The attention mechanism further enhances this capability by dynamically assigning higher importance to time steps that contain the most discriminative information. Instead of uniformly processing the entire window, the model selectively focuses on transient segments where islanding signatures are most pronounced. Overall, the superior performance of the CNN-BiLSTM architecture can be attributed to its ability to jointly model local transient features, bidirectional temporal dependencies, and selective temporal emphasis, all of which align with the physical characteristics of islanding phenomena. This interpretation indicates that the performance gain is structurally grounded rather than purely data-driven. Spatiotemporal feature attribution and model interpretability To provide insight into the decision-making behavior of the proposed model, a gradient-based feature attribution analysis is performed on the trained CNN-BiLSTM network. The analysis evaluates the relative importance of each input channel over time using the selected nine-channel measurement set . The resulting spatiotemporal heatmap (Fig. 13) reveals that the model decisions are predominantly influenced by current measurements at Bus-1, particularly the phase- c component , which exhibits the highest saliency values within the intermediate time interval (approximately time steps 7–10). This indicates that transient current responses at the electrically sensitive bus carry the most discriminative information for islanding detection. Spatiotemporal feature attribution heatmap. In addition, the phase- b current component also demonstrates a noticeable contribution, albeit at a lower magnitude compared to . The involvement of multiple phase currents suggests that the model captures subtle phase-dependent asymmetries in the transient response, rather than relying on a single dominant signal component. In contrast, voltage measurements and remote bus currents contribute less significantly to the classification process. This suggests that the model primarily exploits localized dynamic current signatures rather than steady-state voltage variations, which are typically less sensitive to islanding conditions. From a temporal perspective, the concentration of high attribution values within a narrow time interval demonstrates that the model focuses on post-disturbance transient behavior instead of uniformly processing the entire observation window. This selective emphasis is consistent with the attention mechanism embedded in the architecture, which enhances sensitivity to critical temporal segments. Importantly, these findings align with the physical characteristics of islanding phenomena, where abrupt changes in current profiles provide more reliable indicators than voltage signals under near power balance conditions. Therefore, the feature attribution analysis not only improves interpretability but also validates the physical relevance of the selected measurement configuration and the learned decision patterns. NDZ analysis and comparison In microgrids, islanding occurs when the utility grid is disconnected while local distributed generation continues to supply the connected load. The detectability of islanding is fundamentally governed by the mismatch between generation and load at the instant of disconnection. The active and reactive power imbalances are defined as (33) and (34). In here and denote the total generated active and reactive power, and , represent the corresponding load demand. The NDZ is defined as the set of operating points for which the islanding detection method fails to detect islanding within the required 2-second limit specified in IEEE Std 1547–2018. In this study, the NDZ is evaluated over a two-dimensional mismatch space defined by and , both with increments of 2%. This results in a total of 861 operating points. For each operating condition, an islanding event is simulated by opening the grid-side circuit breaker at t=0.32s under the specified mismatch. The resulting time-series data are processed and segmented into non-overlapping windows of 20 samples. To ensure robustness, only windows corresponding to steady post-islanding conditions with consistent labels are considered, while transition periods and non-informative segments are excluded. These validated windows are then evaluated using the trained CNN-BiLSTM model, which outputs binary classification results indicating detection success or failure. Finally, the classification results are mapped onto the plane. Operating points where islanding is not detected constitute the NDZ region. This representation enables direct visualization of detection sensitivity and demonstrates the robustness of the proposed method under varying power mismatch conditions. The NDZ map of the CNN-BiLSTM architecture is presented in Fig. 14. According to evaluation results, the proposed algorithm achieves a very low NDZ rate, approximately 0.02%, demonstrating its high sensitivity and reliability. NDZ Identification for the CNN-BiLSTM model. Detection time is defined as the time difference between the occurrence of the islanding event and the first stable classification of the islanding condition by the proposed model. A sliding window of 20 time steps is used, and classification is performed sequentially as new data becomes available. Therefore, the detection delay depends on the sampling frequency and the window size. In this study, the proposed method detects islanding within approximately 10 ms after the event, corresponding to the earliest consistent classification of the islanding state. The obtained detection time satisfies the IEEE 1547–2018 requirement of detecting unintentional islanding within 2 s. Table 5 presents a comprehensive comparison between the proposed ID method and several recent approaches in the literature from 2022 to 2025. The comparison based on key performance indicators including detection time, non-detection zone (NDZ), methodological approach, and robustness against disturbances such as faults, noise, and switching events. It is evident that many existing methods involve trade-offs between detection speed, accuracy, and implementation complexity. Passive approaches are generally sensitive to system conditions, while several machine learning and signal processing techniques suffer from larger NDZ or slower response times. In contrast, the proposed CNN-BiLSTM-based approach achieves a significantly faster detection time with a minimal NDZ, while maintaining strong robustness under diverse disturbance scenarios, making it highly suitable for real-time applications. Table 5Comparison with studies in the literature.YearRef.Method(s)Detection TimeNDZMethod TypeDisturbance RobustnessRemarks202317Deep Neural Network43.4 ms0.18%Deep LearningMediumModerate NDZ and detection speed202313Threshold-based20 msNearly zeroPassiveLowSensitive to threshold selection202418Voltage trajectory + Resonance0.17 s-Signal ProcessingMediumRelatively slow detection202319Statistical + KDE30 ms1%StatisticalMediumLimited generalization202421Decision Tree34.27 ms< 1%Machine LearningMediumModerate classification performance202349RF + Optimization22 ms0.25%Hybrid MLMediumIncreased computational complexity202422AlexNet – based DL-1%Deep LearningMediumHigh model complexity202250Threshold - basedVariableVariablePassiveLowHighly dependent on system conditions202223CNN-LSTM50 ms2%Deep LearningMediumLarger NDZ compared to proposed202351Artificial Neural Network-2.11%Machine LearningMediumLimited detection performance202252Communication- Based2.5 cycles-PassiveHighRequires communication infrastructure202216Multi-LSTM50 ms2%Deep LearningMediumSlower detection performance202253OSVM + PSO25 ms-Hybrid MLMediumComplex implementation202354Lissajous + Adaptive Threshold-1%Signal ProcessingMediumSensitive to signal variations202355µPMU-Based Detection< 250 msNearly zeroMeasurement - basedHighRequires high-cost measurement units202315VMD + DBNVariable2.4%Hybrid Deep LearningMediumModerate NDZ and complexity202324Reinforcement Learning + VMD60 ms-Hybrid MLMediumHigh training complexity 2026 Proposed CNN-BiLSTM 10 ms 0.02% Deep Learning (Hybrid) High Fast, accurate, low NDZ, robust under multiple disturbances Comparison with studies in the literature. The introduced method, based on a hybrid CNN-BiLSTM architecture, achieves a remarkably fast detection time as 10 ms and a minimal NDZ as 0.02%, significantly outperforming existing techniques in both speed and accuracy. This performance highlights the capability of model to rapidly and reliably detect islanding conditions under various system disturbances and scenarios. It is also clear in Table 5 that most prior studies either lack complete data on NDZ or report values significantly higher than the proposed model. Thus, the findings confirm that the introduced approach not only enhances detection speed but also minimizes the NDZ, making it highly suitable for real-time implementation in modern smart grid environments. Validation on the IEEE 33-bus distribution system To further evaluate the scalability and robustness of the proposed method, an additional case study was conducted using the IEEE 33-bus distribution system. The IEEE 33 Bus system is based on the Matlab implementation by Wong56. To obtain the conventional IEEE 33-bus radial distribution network and to represent an active microgrid-integrated system the model was modified. In order to increase the practical relevance of the test system, distributed generation (DG) units were integrated at selected buses with different generation capacities. Specifically, a 0.35 MW DG unit was connected at Bus 18, a 0.5 MW unit at Bus 6, and a 1 MW unit at Bus 30, as illustrated in Fig. 15. This modification transforms the passive radial feeder into an active distribution network, enabling realistic islanding and non-islanding scenarios. The modified IEEE 33 Bus system. A set of representative operation scenarios, including both islanding and non-islanding events, was defined as summarized in Table 6. These scenarios include breaker-based islanding events, fault conditions, load variations, and capacitor switching operations to ensure comprehensive evaluation. Table 6Islanding and Non-Islanding Operation Scenarios Applied to the Modified IEEE 33-Bus Microgrid Test System.Scenario IDConditionLocationEvent TypeParameterJustificationReferenceIS-1IslandingIslanding Breaker (Microgrid Tie Switch)Breaker openingt = 0.1 sStandard microgrid separation case where the local network is disconnected through the dedicated islanding switchIEEE Std 1547IS-2IslandingIslanding Breaker (Microgrid Tie Switch)Breaker openingLocal load connectedEvaluates detection under normal operation load after intentional separationMicrogrid protection studiesIS-3IslandingIslanding Breaker (Microgrid Tie Switch)Breaker openingOne DG offline before eventTests islanding detection under reduced local generation reserveDistributed generation islanding literatureNI-1Non-IslandingBus-2SLG Fault1 ΩClose-in fault near source side may create severe transient signatures without actual islandingIEEE 33-bus protection studiesNI-2Non-IslandingBus-10LL Fault2 ΩMid-feeder line fault used to verify immunity against false islanding tripsDistribution protection literatureNI-3Non-IslandingBus-18Load Increase+ 25% step loadSudden load variation produces voltage/current disturbance while remaining grid-connectedMicrogrid dynamic studiesNI-4Non-IslandingBus-30Capacitor Switching300 kvarCapacitor energization introduces oscillatory transients similar to islanding signaturesPower quality literature Islanding and Non-Islanding Operation Scenarios Applied to the Modified IEEE 33-Bus Microgrid Test System. To determine the most informative measurement location, a sensitivity analysis was performed across system buses. As shown in Fig. 16, Buses 18, 19, and 32 exhibited high sensitivity to islanding events. Among these, Bus 18 demonstrated the highest sensitivity and was selected as the critical measurement point. The three-phase voltage and current signals obtained from this bus were used as input to the proposed CNN-BiLSTM model. Sensitivity scores of buses with respect to islanding in the Modified IEEE 33-Bus System. The model was trained using the defined scenarios, and convergence was achieved within approximately 60 epochs, with both training and validation accuracies reaching near-perfect levels, as shown in Fig. 17. The corresponding confusion matrices are presented in Fig. 18. Training and validation and accuracy curves for the proposed CNN-BiLSTM model over epochs. Training (a) and Test Confusion Matrices (b) of the proposed model for the modified IEEE 33-Bus System. To evaluate generalization capability, the trained model was tested on unseen scenarios. In a non-islanding test case, a phase-to-phase fault with a resistance of 1.7 Ω was applied at Bus 9, along with a 400 kVAr capacitor switching event at Bus 30. Despite not being included in the training dataset, the model correctly classified the condition as non-islanding (Fig. 19). Non-islanding classification performance of the proposed model for the Modified IEEE 33-Bus System. Furthermore, a challenging zero power mismatch scenario was created by matching the generation and load levels. Additionally, a load reconnection event was introduced. Even under these conditions, the model successfully identified the islanding event without misclassification (Fig. 20), demonstrating strong robustness and generalization capability. Islanding classification performance of the proposed model for the Modified IEEE 33-Bus System. These results confirm that the proposed method maintains high performance when applied to a larger and more complex distribution system. Conclusions In this study, a novel CNN-BiLSTM-based deep learning architecture was developed for accurate and fast detection of islanding events in microgrids. The proposed method achieved a detection time of only 10 ms, demonstrating its suitability for real-time protection applications. The effectiveness of the approach was further validated by an extremely narrow NDZ of just 0.02%, indicating high sensitivity in islanding conditions from normal operation modes. Performance metrics including accuracy and loss were evaluated comprehensively. The results were visualized through training and testing confusion matrices, highlighting the ability of classifier to generalize well across unseen data. Comparative analysis with previous works confirmed that the proposed method outperforms existing algorithms in terms of detection reliability and speed. The proposed method is developed and evaluated using the IEC 61850-7-420 simulation test system. While the results demonstrate strong performance, several factors influencing generalization to real-world deployments should be considered. The generalization capability of the model is primarily supported by the diversity of the training dataset and the use of physically meaningful input features. The dataset includes a wide range of operating conditions, such as varying active and reactive power mismatches, multiple non-islanding disturbance types, and different fault scenarios. This diversity enables the model to learn invariant temporal patterns associated with islanding events rather than system-specific signatures. In addition, the model predominantly relies on transient current measurements, which are fundamental indicators of islanding behavior across different microgrid configurations. This physics-consistent representation improves the transferability of the learned features to other systems. The generalization capability is further supported by validation on an independent IEEE 33-bus distribution network, where the model maintains high detection performance under unseen system conditions. Nevertheless, several limitations should be acknowledged. The current dataset is generated under balanced three-phase operating conditions, and therefore does not capture unbalanced loading scenarios or single-phase distributed generation units. Furthermore, the noise robustness analysis considers moderate measurement noise levels, while more severe sensor degradation or failure conditions are not included. Finally, the study is based on simulation data, and validation using hardware-in-the-loop (HIL) platforms or real measurement systems remains an important step for practical deployment. In future studies, the introduced model can be validated by testing it with real measurement data (PMU/SCADA) or laboratory-based microgrid hardware (RTDS, CHIL/HIL). In the future, accuracy of the proposed model can be increased by adding multiple features (power, harmonics, dq components, THD, flicker measurements). As part of future work, the proposed method will be further validated on larger and more complex distribution systems, such as the IEEE 69-bus and IEEE 118-bus networks, in order to investigate its scalability and performance under more realistic grid conditions. Beyond performance improvements, this study provides a comprehensive evaluation framework for islanding detection. The proposed approach integrates model architecture design with physical interpretability and system-level validation. The NDZ-based analysis offers a physics-driven assessment of detection sensitivity, while the feature attribution results reveal that the model decisions are aligned with meaningful transient current signatures. Furthermore, the validation on an independent test system and the discussion of practical deployment aspects demonstrate the robustness and applicability of the proposed method under diverse operating conditions. Acknowledgements The present study constitutes part of the Ph.D. dissertation of the first author, conducted in Marmara University. Author contributions All authors contributed to writing the paper, critically reviewed the research design, data analysis, and final revisions. H.O. contributed to conceptual development, formal analysis, research protocol design, result interpretation, and initial manuscript drafting. S.V.N. contributed to research design and supervision and directed targeted research tasks. The authors read and approved the final manuscript. Data availability Data sets generated during the current study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets generated and analyzed during the current study are publicly available in the Zenodo repository under the MIT License at: https://zenodo.org/records/20286297. The repository includes: (i)MATLAB/Simulink models of the IEC benchmark system and IEEE 33-bus distribution system, MATLAB/Simulink models of the IEC benchmark system and IEEE 33-bus distribution system, (ii)raw simulation files used for data generation, raw simulation files used for data generation, (iii)processed datasets used for deep learning training and testing, processed datasets used for deep learning training and testing, (iv)sensitive bus analysis scripts, and. sensitive bus analysis scripts, and. (v)custom MATLAB codes used in the study. custom MATLAB codes used in the study. The datasets are stored in MATLAB format. The processed datasets contain three-phase voltage and current measurements together with binary islanding labels. Data sets generated during the current study are available from the corresponding author on reasonable request.