PREDICTION OF NON-SINUSOIDAL VOLTAGE CHANGES BASED ON MACHINE LEARNING ALGORITHMS

Abstract

The analysis of the harmonic composition of the voltage at the substation of the Irkutsk region shows that for all three phases, the average values of the coefficients of the harmonic component of the voltage for the harmonic spectrum from 2 to 9 exceed the corresponding GOST normally permissible values. The results of the measured values of the currents demonstrated that the harmonic level depends on the magnitude of the load. (Research purpose) The research purpose is studying the models for predicting non-sinusoidal parameters at medium-voltage power substations and nodal voltages at 0.4 kilovolt substations connected by a single grid and determining the smoothing functional dependencies necessary for calculating the grid mode. (Materials and methods) Presented data on measurements of electricity quality indicators at the Novozhilkino substation and its connections, conducted an analysis of measurement emissions. The simulation of operating modes was carried out using the Google Colaboratory package. (Results and discussion) We tested the hypothesis that the non-sinusoidal voltage coefficient at one of the connected substations can be predicted with acceptable accuracy at known voltage indicators and non-sinusoidal voltage coefficients at the head and other connected substations. Emissions were analyzed and visualizations were constructed based on measurement data at the supply substation and outgoing connections, and a correlation matrix was constructed. The predicted data of the non-sinusoidal coefficient with a scattering characteristic, which is about 0.7 for all substations of the system, were obtained experimentally. It was determined that the results of the forecast based on machine learning models correlate with the data of measurements in electrical networks. (Conclusions) Proposed a methodology for the effectiveness of forecasting non-sinusoidality based on existing machine learning models, which allow to implement automatic calculation to solve the problem, to correct models, to increase the accuracy of forecasting under uncertainty. The effectiveness of the proposed approach was shown by the example of real experimental data at substations of the central networks of the Irkutsk region.