Artificial Neural Network-Based Fault Detection, Classification, and Localization in Hybrid Microgrids
- lms editor
- 7 hours ago
- 5 min read
Abstract
The transition toward decentralized energy systems has accelerated the deployment of hybrid microgrids, yet their protection remains a significant technical hurdle. Conventional protection coordination is often compromised by the complex fault dynamics of inverter-based resources (IBRs) and synchronous generators. This study proposes an intelligent diagnostic framework utilizing Artificial Neural Networks (ANN) for autonomous fault detection, classification, and localization. The system is modeled in MATLAB/Simulink, integrating a Solar Photovoltaic (PV) array, a Battery Energy Storage System (BESS), and a Diesel power plant. Three distinct neural network architectures are employed, utilizing Levenberg-Marquardt backpropagation to process electrical transients at the Point of Common Coupling (PCC). The methodology addresses 11 distinct fault types over a 10 km distribution line. Results indicate that the multi-network approach achieves high-precision localization and robust classification, with regression coefficients (R) exceeding 0.95. This research validates the efficacy of ANNs in providing real-time, adaptive protection for hybrid microgrids, ensuring stability across diverse operational modes and transient states.
Keywords:
Microgrid, Fault Detection, Neural Networks, MATLAB/Simulink, Power System Protection, Battery Energy Storage Systems (BESS).
I. Introduction
The resilience of modern power systems increasingly depends on the strategic integration of microgrids capable of operating in both grid-connected and islanded modes. However, the protection of these systems is inherently more complex than that of traditional radial distribution networks. Traditional overcurrent protection relies on high short-circuit levels provided by synchronous machines. In contrast, hybrid microgrids incorporate IBRs, such as Solar PV and BESS, which utilize power electronic interfaces with strict current-limiting characteristics. These characteristics suppress the fault current contribution, often leading to the failure of conventional relays to detect high-impedance or distant faults.
Furthermore, the simultaneous operation of synchronous Diesel generators and stochastic IBRs creates non-linear electrical signatures that vary based on the generation mix and load profile. To ensure system reliability, intelligent diagnostic tools are necessary to process multi-variable data from the Point of Common Coupling (PCC). This study develops a specialized ANN-based framework to analyze three-phase voltages (V{abc}), currents (I{abc}), and zero-sequence components (V0, I0). By leveraging the pattern recognition capabilities of neural networks, the system identifies fault types and estimates their location with minimal error, providing a scalable solution for modern protection coordination.
II. System Configuration and Modeling
Maintaining stability in a hybrid microgrid requires a coordinated balance between intermittent renewable sources and dispatchable generation. The configuration modeled in this study is designed to simulate a high-reliability distribution environment.
A. Component Synthesis
The hybrid system comprises the following elements:
· Generation Side: A Solar PV system and a BESS provide renewable energy and energy arbitrage capabilities. A Diesel power plant acts as the primary frequency-response unit, providing the necessary inertia during islanding transients.
· Distribution Line: A 10 km distribution line connects the hybrid microgrid to the utility. Faults are injected at specific intervals (1 km, 5 km, and 9 km from the PCC) to assess the localization accuracy of the proposed model.
· Loading: Parallel AC and DC load centers are integrated to represent a diversified modern demand profile.
B. Mathematical Representation
The electrical behavior at the PCC is defined by the summation of currents from the hybrid sources and the demand from the loads. The total current at the PCC, is formulated as:
For fault detection, zero-sequence components are extracted as critical features. These components are essential for distinguishing between grounded and ungrounded faults, as they are non-zero only during unbalanced conditions involving a ground path. They are derived from the instantaneous three-phase measurements:
These eight variables (Vabc}, Iabc, V0, I0) constitute the input vector for the diagnostic layer.
III. Neural Network-Based Control Strategy
A multi-network architecture is adopted to overcome the dimensionality challenges associated with a single global network. This modular approach enhances diagnostic resolution by isolating the classification and localization tasks.
A. Architecture Breakdown
NN 1 (Detection & Classification): Employs 8 inputs
1. to generate a 4-bit logic output. This network is responsible for identifying the fault topology.
2. NN 2 (LG/LLG Localization): Specifically trained for grounded faults (Line-to-Ground and Double-Line-to-Ground), this network utilizes V0 and I0 as primary inputs to exploit zero-sequence feature extraction.
3. NN 3 (LL/LLL/LLLG Localization): Optimized for ungrounded or balanced faults, this architecture relies exclusively on phase voltage (Vabc) inputs to determine the distance to the fault.
B. Logic Mapping for Classification
The classification network (NN 1) maps complex transients to a unique binary sequence for 11 distinct fault types. This logic allows for seamless integration with digital protection relays.
Table 1: 4-Bit Fault Classification Logic Mapping
Fault Condition | B1 | B2 | B3 | B4 |
Normal Operation | 0 | 0 | 0 | 0 |
Phase A to Ground (AG) | 0 | 0 | 0 | 1 |
Phase B to Ground (BG) | 0 | 0 | 1 | 0 |
Phase C to Ground (CG) | 0 | 0 | 1 | 1 |
Phase A–B to Ground (ABG) | 0 | 1 | 0 | 0 |
Phase A–C to Ground (ACG) | 0 | 1 | 0 | 1 |
Phase B–C to Ground (BCG) | 0 | 1 | 1 | 0 |
Three-Phase to Ground (ABCG) | 0 | 1 | 1 | 1 |
Phase A–B Line-to-Line (AB) | 1 | 0 | 0 | 0 |
Phase A–C Line-to-Line (AC) | 1 | 0 | 0 | 1 |
Phase B–C Line-to-Line (BC) | 1 | 0 | 1 | 0 |
Three-Phase Balanced (ABC) | 1 | 0 | 1 | 1 |
IV. Simulation Setup and Neural Network Training
High-fidelity data acquisition is paramount for training robust neural models. Utilizing the MATLAB nnstart environment, the fitting process maps the non-linear relationship between electrical signatures and fault characteristics.
A. Data Acquisition and Training Parameters
The input (X) and target (T) datasets were compiled from exhaustive simulations of 11 fault types at varying distances. The training parameters were set as follows:
· Data Split: 70% Training, 15% Validation, 15% Testing.
· Algorithm: Levenberg-Marquardt backpropagation.
· Performance Metric: Mean Squared Error (MSE) and Regression Coefficient (R).
· R-Value Achievement: Training was conducted until R ≈ 0.95 to 1.0, ensuring the model captures the transient complexities of the IBR-heavy system.
B. Model Integration
The trained networks were exported as Simulink blocks and integrated into the microgrid model. This allows the system to perform real-time inference, where PCC data is continuously analyzed to provide instantaneous fault identification and distance estimation.
V. Results and Discussion
The protection system’s dynamic response was evaluated by injecting faults at t = 0.1 s. This timing is critical to verify the ANN's ability to distinguish between steady-state operation and high-speed transients.
Case Study A: Double Line to Ground (ABG) Fault
A fault was injected at the 9 km mark as measured from the microgrid PCC.
· Observation: The classification network (NN 1) immediately identified the "ABG" logic (0100). The localization network (NN 2) estimated the fault distance at 9.4 km.
· Discussion: The error margin of 0.4 km is primarily attributed to the current-limiting controllers within the PV and BESS inverters. These controllers clip the fault current peak, slightly distorting the traditional distance-impedance relationship that the ANN must navigate.
Case Study B: Mid-Line Fault
A fault was simulated at the 5 km mark from the PCC.
· Observation: The localization output converged to 4.6 km.
· Discussion: The high degree of accuracy (within 8% error) demonstrates the model's effectiveness in mid-line scenarios. The results confirm that even when BESS and PV power outputs fluctuate during the transient phase, the ANN maintains diagnostic integrity without requiring the manual threshold adjustments typical of overcurrent relays.
VI. Conclusion and Future Scope
This research highlights the strategic advantage of Artificial Neural Networks in hybrid microgrid protection. By utilizing a multi-network architecture, the system successfully classifies 11 different fault types and provides precise localization across a 10 km distribution line. Unlike conventional protection schemes, this ANN-based approach effectively manages the non-linear fault signatures introduced by IBRs (PV/BESS) and synchronous machines.
Future research will focus on the integration of Total Harmonic Distortion (THD) analysis into the training features to enhance detection under high non-linear load conditions. Additionally, the development of adaptive protection logic for purely islanded scenarios—where fault currents are significantly lower—remains a critical area for improving the resilience of autonomous microgrids.
VII. YouTube Video
VIII. Purchase link of the Model
SKU: 0091
Comments