Utilizing the new Kolmogorov-Arnold network for short-term electric load forecasting
Muhammad Abbas, Yanbo Che, Zafar Ahsan
Abstract
Power utilities rely on accurate load forecasting for making effective real-time decisions and managing system operations, which, in turn, significantly impacts their financial performance. Accurate load forecasting enables these utilities to reduce operational costs, optimize generation, and maintain system stability. However, despite extensive research efforts, the nonlinear characteristics of power system and complex nature of the load datasets make accurate load forecasting challenging. This paper presents a novel application of the new Kolmogorov-Arnold Networks (KANs) for short-term load forecasting (STLF). KANs offer a promising predictive ability compared to Multi-Layer Perceptrons by replacing their linear weights with unique learnable activation functions. These learnable activation functions enable KAN to dynamically learn and effectively capture the non-linearities and complexities in power system load data. The proposed network is applied for STLF based on the latest ISO-NE hourly sampled load data in the period 2019–2023. The results demonstrate that KANs deliver higher accuracy in forecasting load over various prediction horizons while utilizing significantly reduced number of learnable parameters. This efficiency shows not only the KAN's predictive power but also a reduced computational complexity. By leveraging the dynamic learning capabilities of KAN, power utilities can achieve more precise load forecasts, leading to optimized operations and improved financial performance.