The study aimed to quantitatively assess the effectiveness of neural network-based control of dynamic reactive power compensation in the traction network of Ukrzaliznytsia. The methodology involved high-frequency monitoring of 118 substations (≈16 trillion measurements), aggregation of 52 million Supervisory Control and Data Acquisition records, Theil-Sen trend estimation, Seasonal-Trend decomposition using Loess, the Brown Forsythe test, and a deep neural network with 64-32-16 layers optimised through a Tree-structured Parzen Estimator, validated in 120 Monte Carlo scenarios in Simscape. The results demonstrated that the average daily Q increased from 309.8 to 367.2 Mvar (18.5%) with a trend of 0.95 Mvar/month (95% CI 0.82-1.08; p < 0.001). Winter peaks reached 520 Mvar, while variance after February 2022 increased 1.5-fold (F = 34.9; p < 0.001). The neural network achieved Mean Squared Error = 184 Mvar² and R² = 0.982; in testing, R² = 0.975, Mean Absolute Error = 11.2 Mvar, and the median error of winter peaks did not exceed 3%. In simulations, the algorithm maintained cos φ between 0.985 and 0.991 for loads ranging from 0.3 to 1.2 p.u., reduced daily losses by 6.1-9.4% (≈1.2 MWh/day), and decreased the number of Static Var Compensator switchings by 38%. Under failure of 20% of Insulated-Gate Bipolar Transistor modules, the duration of cos φ < 0.98 was 0.7%, compared with 19% for the Proportional Integral-Derivative controller. Quantitative comparison of compensation strategies confirmed that the simulated STATCOM-DL system achieved the lowest average cos φ deviations (0.015 vs. 0.030-0.048 for competitors) and minimal annual losses (≈ 430 MWh), outperforming all alternative approaches. The findings confirmed that the combination of deep learning and statistical validation enhances network stability and improves energy efficiency. The practical significance lies in the fact that the resulting algorithms may be employed by railway dispatch centres and energy integrators for real-time optimisation
Maksym Shpilievyi (Sun,) studied this question.