Short-Term Load Forecasting Using Conventional and AI Techniques

Abstract

Short-term load forecasting (STLF) ensures efficient energy management, economic dispatch, and grid stability in modern power systems. However, their performance is often constrained by assumptions of linearity and stationarity in the data. In contrast, Artificial Neural Networks (ANN) offer a data-driven, non-linear approach that captures complex relationships in time series data, thereby demonstrating significant potential in enhancing forecasting accuracy. This paper comprehensively compares conventional statistical methods (AR, ARMA, ARIMA) with Artificial Neural Networks for STLF. The analysis explores their performance under various forecasting horizons, data resolutions, and real-world scenarios, emphasizing their adaptability to changing load patterns. The research highlights that while ARIMA performs well for linear and stationary data, its accuracy diminishes when dealing with highly non-linear, dynamic loads. On the other hand, ANNs, with their inherent capacity to model non-linearities, exhibit superior accuracy, particularly when integrated with data pre-processing techniques and optimized training algorithms. The study assesses the effectiveness of different approaches using datasets of actual electrical loads. Findings show the accuracy, resilience, and scalability benefits of AI-based methods, opening the door for their incorporation into contemporary energy systems. The limits of conventional forecasting techniques for trustworthy energy management are addressed by artificial intelligence, as this comparative study highlights.