Work

Multiscale adaptive Kolmogorov-Arnold networks with cross-variable learning for short-term electricity price forecasting

Year: 2026

Type: article

Abstract: With the transition of electricity markets toward low-carbon and market-oriented structures, the electricity price time series exhibits strong non-stationarity, multi-scale fluctuations, and complex nonlinear dynamics, which pose significant challenges for Electricity Price Forecasting (EPF). However, existing models often fail to completely capture these complex patterns. To address the limitations of current methods, we propose a Multiscale Adaptive Kolmogorov-Arnold Network with Cross-Variable Learning (MKAN-CL), which is combined with the Adaptive Lemming Algorithm (ALA) for automatic hyperparameter optimization. MKAN-CL integrates spectral decomposition with adaptive modeling to capture complex temporal dynamics in short-term electricity price series. Specifically, the method consists of four key components. First, a multi-lag cross-variable modeling mechanism is designed to capture cross-scale dependencies between electricity prices and exogenous variables. Then, hierarchical spectral decomposition is applied to separate non-stationary sequences into distinct frequency components. Subsequently, Chebyshev-based Kolmogorov-Arnold Networks (KAN) are employed to perform adaptive nonlinear modeling across frequency bands. Moreover, the Adaptive Lemming Algorithm (ALA) is employed to automatically select key hyperparameters for the proposed model, reducing reliance on manual tuning. Finally, experiments on the Australian and Singapore electricity markets demonstrate the effectiveness of MKAN-CL. On the AEMO dataset, MKAN-CL reduces RMSE, MAE, and MAPE by 4.97%, 13.75%, and 12.95%, respectively, relative to the strongest baseline, CrossLinear. Overall, MKAN-CL provides an effective solution for short-term electricity price forecasting in complex environments and offers data-driven support for price volatility assessment and operational decision-making in renewable-integrated power systems.

Source: Electric Power Systems Research

DOI: 10.1016/j.epsr.2026.114352

Authors: Xin Wen, Jiqiang Tang, Xiaofei Gao

Institution: Chongqing University of Technology

Language: English

Topic: Energy Load and Power Forecasting

Subfield: Electrical and Electronic Engineering

Field: Engineering

Domain: Physical Sciences

SDGs: Affordable and clean energy, Climate action, Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,810