Forecasting electrical losses in the power system of the Republic of Kazakhstan based on artificial intelligence methods

Abstract

Abstract. At the current stage of development, the traditional power industry requires qualitative renewal and deep modernization. Amidst the rapid growth of digital technologies, innovative digital solutions are increasingly replacing classical approaches. In the current era of digitalization, there is an unprecedented surge of interest in machine learning and artificial intelligence (AI) technologies, which are becoming key tools for solving complex engineering problems.

This article addresses one of the most pressing issues in modern power engineering: improving the accuracy of forecasting technical electrical energy losses in the national power grid of the Republic of Kazakhstan, managed by JSC "KEGOC". Traditional statistical calculation methods used in current practice often exhibit significant margins of error. This is due to the inability of classical models to fully account for non-linear dependencies, dynamic changes in meteorological conditions, and the complex nature of load processes in high-voltage networks.

As an innovative solution, the article proposes a methodology based on the use of recurrent neural networks with Long Short-Term Memory (LSTM) architecture. Unlike standard algorithms, LSTM is capable of effectively processing time series by considering long-term dependencies and cyclic trends. The work includes a detailed description of the intelligent data preprocessing stage and a comparative analysis (benchmarking) of LSTM, ARIMA, and gradient boosting models. The results demonstrate that the implementation of AI reduces the forecasting error to less than 5%, which is critical for optimizing the operating modes of the power system.