A deep learning model for fault detection in distribution networks with high penetration of electric vehicle chargers

• Detecting faults in distribution networks with high penetration of EV chargers. • Deep learning scheme based on black widow optimization bi-directional LSTM model. • No need to determine thresholds or utilize communication links. • Detecting all types of faults under different conditions of the network operation. • Significant increase in detection accuracy compared to the KNN and LSTM models. Integration of a significant number of domestic electrical vehicle (EV) charging stations into the power distribution infrastructure can give rise to several protection problems. Therefore, we propose a new method to detect short-circuit faults in distribution networks with high penetration of residential EV chargers. In this method, first, the features of voltage and current waveforms in various operational scenarios are extracted through a two-dimensional modeling. These features are then used to train a deep learning model based on black widow optimization bi-directional long short-term memory (BWO-BiLSTM) technique. In contrast with the conventional adaptive protection schemes, the proposed method can perform accurately in the presence of fast and unpredictable network topology, without requiring to determine a large number of threshold values to detect a fault, or relying on communication links. The effectiveness of the proposed method is investigated through a series of case studies on a modified IEEE 69-bus distribution network with a substantial penetration of residential EV chargers. The results show the proposed method's ability to detect all types of faults within 5 ms. Since it employs a machine learning algorithm for fault detection, the method's accuracy is 98.5 %, surpassing the accuracy of k-nearest neighbors (KNN) and conventional LSTM models. Additionally, the results confirm its optimal performance under noisy conditions. Even with noise in the sampled signals at a level of 10 dB, the method's accuracy remains higher than that of other methods, with a value of 96.9 %.