COMPARATIVE ANALYSIS OF NEURAL NETWORK ARCHITECTURES FOR REACTIVE POWER FORECASTING IN AC TRACTION POWER SUPPLY SYSTEM

Abstract

The problem of reactive power compensation in AC traction power supply systems is critical for ensuring energy efficiency, voltage stability, and power quality. Traditional compensation methods often lack the adaptability required to handle the highly dynamic and non-linear loads characteristic of electric railway networks. This study aims to conduct a comprehensive comparative analysis of five artificial neural network architectures (specifically Multi-Layer Perceptron, Convolutional Neural Networks, Recurrent Neural Networks, Gated Recurrent Units, and Long Short-Term Memory) for short-term reactive power forecasting. The research relies on experimental data obtained throughout 2024 from an operating traction substation using a high-precision SATEC PM175 power quality analyzer with a resolution of 128 samples per cycle. The modeling process was implemented in the MATLAB environment, examining two scenarios: preserving the temporal sequence of data and using mini-batch shuffling to evaluate generalization capabilities without data leakage. A key finding is that the LSTM architecture demonstrates superior performance when instantaneous values of current, voltage, and phase angle are used as input parameters, achieving the highest precision with a Root Mean Square Error (RMSE) of 13.70 kvar and a Mean Absolute Percentage Error (MAPE) of 2.76 % in the sequential mode. The study confirms that incorporating physically justified electrical parameters significantly improves prediction accuracy compared to using historical reactive power data alone. These results validate the effectiveness of recurrent neural networks, particularly LSTM, for modeling complex processes in traction networks and provide a reliable algorithmic basis for implementing predictive real-time intelligent control of compensating devices, such as FACTS and UPQC systems. Keywords: reactive power, traction power supply, artificial neural networks, LSTM neural network, short-term forecasting, predictive control, power quality, railway infrastructure.