Evolving Electricity Demand Modelling in Microgrids Using a Kolmogorov-Arnold Network

Abstract

Electricity demand estimation is vital for the optimal design and operation of microgrids, especially in isolated, unelectrified, or partially electrified areas where demand patterns evolve with electricity adoption. This study proposes a causal model for electricity demand estimation that explicitly considers the electrification process along with key factors such as hour, month, weekday/weekend distinction, temperature, and humidity, effectively capturing both temporal and environmental demand patterns. To capture the electrification process, a “Degree of Adoption” factor has been included, making it a distinctive feature of this approach. Through this variable, the model accounts for the evolving growth in electricity usage, an essential consideration for accurately estimating demand in newly electrifying areas as consumers gain access to electricity and integrate new electrical appliances. Another key contribution of this study is the successful application of the Kolmogorov–Arnold Network (KAN), an architecture explicitly designed to model complex nonlinear relationships more effectively than conventional neural networks that rely on standard activation functions, such as ReLU or sigmoid. To validate the effectiveness of the proposed electricity demand modelling approaches, comprehensive experiments were conducted using a dataset covering 578 days of electricity consumption from El Espino, Bolivia. This dataset enabled robust comparisons among KAN and conventional neural network architectures, such as Deep Feedforward Neural Network (DFNN) and Multi-Layer Perceptron (MLP), while also assessing the impact of incorporating the Degree of Adoption factor. The empirical results clearly demonstrate that KAN, combined with the Degree of Adoption, achieved superior performance, obtaining an error of 0.042, compared to DFNN (0.049) and MLP (0.09). Additionally, integrating the Degree of Adoption significantly enhanced the model by reducing DFNN estimation error by approximately 10%. These findings validate the effectiveness of explicitly modelling electricity adoption dynamics and confirm KAN’s relevance for electricity demand estimation, highlighting its potential to support microgrid design and operation.