The neural network training process produces black-box models with low explainability. In addition, the process itself is numerical, with parameters (such as learning rate, momentum, and early stopping trigger) being chosen ad hoc. During the training with chosen parameters, after each calculated update of the weights, the observed total change of weights indicates which training stage the network is currently in. At the same time, neural networks are limited in the data they can model due to various reasons, such as architecture, activation functions, data itself, and the training approach. This limitation is expressed in the phenomenon of the efficient computational frontier, which, it seems, cannot be crossed, no matter the hyperparameters of the network. This paper tackles the efficient usage of information regarding the total change of weights and the efficient computational frontier to determine when the training should be stopped. The results demonstrate the efficiency of training of simpler models compared to more complex models and prove that the general weight structure of models is formed very quickly in the training, while the forming of finer details takes up much more time.

The architecture of a Deep Neural Network (DNN) plays a major role in determining its performance, yet the traditional methods for optimizing these architectures often depend on iterative trial-and-error processes requiring substantial expertise and manual effort. Neural Architecture Search (NAS) has emerged as a rapidly advancing field focused on automating the optimization of hyperparameters and network architectures. This study presents a comparative analysis of three heuristic approaches for NAS: Evolutionary Genetic Algorithms, Reinforcement Learning, and Random Forest Optimization. The efficacy of these methods is evaluated on two widely recognized benchmark classification datasets—MNIST and CREDIT CARD FRAUD—as well as a synthetically generated dataset. A comprehensive evaluation of performance metrics provides insight into the strengths, limitations, and relative effectiveness of each NAS methodology in optimizing neural network architectures for diverse data distributions.

With the growing requirements to keep the security of supply higher than ever the room for failures is getting smaller in today's power systems, while the increased integration of distributed renewable energy sources is additionally complicating fault detection. By using big data that is collected in modern power systems, artificial intelligence algorithms can significantly improve the capabilities of traditional protection schemes. However, the choice of the artificial intelligence algorithm can significantly impact the scheme accuracy. This paper analyses a novel approach for power system fault detection and classification by using automated machine learning procedure that iterates over different data transformations, machine learning algorithms, and hyperparameters to select the best model. By simulating and testing tens of thousands of fault scenarios on a realistic test system, the suggested approach resulted with robustness and high accuracy.

Despite the rapid improvements in the field of microgrid protection, it continues to be one of the most important challenges faced by the distribution system operators. With the introduction of this new operation concept, the existing protection devices are not able to successfully identify, classify and localize different types of faults that occur in the microgrids due to their dynamic behaviour, especially in the islanded mode of operation. This paper presents a methodology that provides the station protection functionalities that include detection and classification of faults, isolation of the faulty feeder and fault location estimation. The proposed method is based on discrete wavelet transform and artificial neural networks. The test system based on the real data, completely developed in MATLAB Simulink, is used to demonstrate the accuracy of all functionalities of the station protection algorithm that can be easily applied in microgrids. The presented results demonstrated the method accuracy and showed that it can be used as an upgrade of the existing protection equipment for the future implementation of the advanced microgrid station protection system.

This paper presents a method for distributed generation (DG) allocation in low voltage distribution network based on the total annual energy loss reduction and Artificial Neural Network (ANN). The proposed method is applied to the PV solar based DG allocation problem in the low voltage distribution network using realistic network data and measurements. This research is motivated by numerous realistic issues faced by the Distribution System Operator in the area of DG planning. The main objective of this work is to develop, test and validate a robust method for DG allocation which can be used in practical problems without the need for extensive system modelling and load flow analysis. The results confirm the importance of appropriate DG planning and show that the proposed method can be used as a promising tool for efficient and effective DG allocation in low voltage distribution network.