Forecasting failures of 35-kV overhead power lines using a neural network to enhance power supply reliability in the agro-industrial complex

Abstract

The development, training, and verification of an intelligent model for forecasting the quantity and types of technological failures in 35 kV overhead power lines are discussed. The research goal was to enhance the reliability of power supply to agro-industrial complex facilities under the specific and dynamically changing climatic conditions of the Shali District of the Chechen Republic. The relevance of the research is determined by the critical dependence of modern agro-industrial complex on stable energy supply and the limitations of traditional overhead power line diagnostic approaches which are unable to account for complex, non-linear interdependencies among numerous factors. A fully connected neural network was employed as the methodological basis, trained on an extensive time series dataset from 2018 through 2023. Data aggregation involved information from operational logs, SCADA systems, and local meteorological stations facilitating the incorporation of destructive factors such as temperature fluctuations, wind loads, and ice-frost depositions. Neural network architecture optimization utilized the Bayesian approach, and its predictive capability was evaluated using the Mean Absolute Error (MAE) metric. Empirical experiments revealed that the optimal neural network architecture was a three-layer configuration with 16 × 64 × 64 neurons achieving a minimum MAE of approximately 0.17 across all classes of predicted failures. Further increases in network depth or width did not yield significant improvements, thereby confirming an optimal balance between model complexity and training data volume. The developed model demonstrates significant potential for the timely planning of preventive maintenance which may substantially enhance the operational reliability of power supply to critical agro-industrial complex facilities.