Microgrid Fault Detection with CNN-GRU and DRL Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fault detection and classification methods in microgrids face challenges such as high computational complexity, inefficiency in handling high-impedance faults, and instability under varying network topologies and noise conditions, particularly in AC microgrids operating in islanded mode.
Innovation Solution
A hybrid model combining a convolutional neural network (CNN) and gated recurrent unit (GRU) with deep reinforcement learning (DRL) is employed to detect and classify faults, utilizing local measurements and adaptive learning to handle AC microgrid topology changes, bidirectional power flow, and fault current levels, with a focus on fast tripping and stability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional fault detection methods are used in microgrids, then the protection system can identify faults, but the computational complexity increases and detection accuracy decreases for high-impedance faults
Solution Approach 1:
The fault detection process is segmented into distinct phases: training phase where the CNN-GRU hybrid model learns from labeled fault data, and detection phase where the trained model processes real-time measurements. This segmentation allows complex computations to be performed only during training, while runtime detection becomes computationally efficient.
Solution Approach 2:
The CNN-GRU hybrid model is pre-trained offline using comprehensive fault data and simulations before deployment. This preliminary action prepares the model with learned features and patterns, enabling it to perform accurate fault detection during operation without requiring complex real-time computations.
2Adaptability or versatility
If adaptive protection schemes are implemented to handle topology changes, then the system can adapt to islanded mode operation, but the stability against noise and varying conditions deteriorates
Solution Approach 1:
The CNN-GRU hybrid model acts as an intermediary between raw measurements and protection decisions. It processes and filters the input signals, extracting meaningful features while suppressing noise and disturbances, thereby providing stable and reliable fault detection under varying topological conditions.
Solution Approach 2:
The model adapts to different operating conditions by learning from diverse training data that includes various topologies, noise levels, and fault scenarios. This parameter variation during training enables the model to maintain stability and accuracy across different microgrid configurations without requiring reconfiguration.
3Speed
If fast tripping is implemented to improve protection speed, then the operating time decreases, but the accuracy of fault classification and detection reliability worsens
Solution Approach 1:
The CNN-GRU hybrid model is pre-trained offline using comprehensive fault data and simulations before deployment. This preliminary action prepares the model with learned features and patterns, enabling it to perform accurate fault detection during operation without requiring complex real-time computations.
Solution Approach 2:
The protection approach transitions from traditional time-domain analysis to a hybrid spatial-temporal dimension by combining CNN's spatial feature extraction with GRU's temporal sequence processing. This dimensional transformation enables the model to capture both local fault characteristics and temporal evolution patterns, achieving fast and reliable classification simultaneously.
Data Source
AI summary
A method for the classification and detection of faults in a microgrid having a distance relay, includes measuring the first plurality of voltage signals and the first plurality of current signals of the microgrid, calculating a plurality of fault-loop impedance signals, comparing the plurality of fault-loop impedance signals and a plurality of reference impedance values, inputting the plurality of difference values to a convolutional neural network (CNN)-gated recurrent unit (GRU) hybrid model, generating a reference tripping signal, defining a deep reinforcement learning (DRL) agent for a deep reinforcement learning (DRL) model. The method further includes processing a second plurality of voltage signals and current signals of microgrid and classifying the one or more fault signals into one or more fault types.