Accurate multi-step wind and solar power forecasting based on multi-scale convolutional Kolmogorov-Arnold network and improved Lemming-optimized attention fusion

Abstract

With the deepening of power market reform, the increasing share of wind and solar energy introduces significant challenges for power system stability due to the high volatility and uncertainty of weather-dependent generation. Accurate multi-step ultra-short-term forecasting is therefore essential for ensuring power balance and effective dispatch coordination in smart grids. To address this issue, we propose a novel hybrid deep learning framework that integrates a multi-scale convolutional Kolmogorov-Arnold network (MCKAN) to improve forecasting performance. This network is specifically designed to capture high-dimensional spatial and temporal features across multiple levels of abstraction. To improve feature selection and scale-specific weight allocation, we integrate an Efficient Additive Attention (EAA) mechanism, which is applied for the first time in the context of renewable energy forecasting. In addition, a Chaotic Quasi-Reverse Artificial Lemming Algorithm (CQALA) is proposed to automatically optimize the complex multivariate hyperparameters, enabling optimal hyperparameter selection and improving the model's overall predictive performance. Extensive experiments on a two-year wind and photovoltaic power dataset from the State Grid of China demonstrate that the proposed method outperforms several state-of-the-art models. For multi-step forecasting, the mean absolute error is reduced by up to 27.6 percent for photovoltaic power and 33.4 percent for wind power, highlighting the practical value of the proposed approach in real-world renewable energy management.