Attention-Guided KAN-Transformer hybrid model for power prediction

Accurate power prediction such as energy and transient energy is essential for the efficient operation of modern electrical systems. Traditional methods often fail to capture complex dynamics of power load data due to their linear assumptions. Recent deep learning methods, including Recurrent Neural Networks (RNNs) and Transformer-based models, have shown promise but still face challenges in handling long-term dependencies and high-dimensional data. To address these limitations, we propose the Attention-Guided KAN-Transformer Hybrid Model (AKTHM), which integrates the Kolmogorov-Arnold Network (KAN) with the Transformer architecture to capture both short-term fluctuations and long-term dependencies. Our model introduces several innovations: an attention mechanism to dynamically learn the fusion weights of different-term KAN features, a deep residual structure to enhance feature extraction and training stability, and a specialized loss function that incorporates periodicity and trend components. Through extensive experiments on four benchmark datasets (ETTh1, ETTh2, ETTm1, and ETTm2), AKTHM consistently outperformed the state-of-the-art models. On the ETTm1 dataset, AKTHM achieved an average MSE of 0.362 and MAE of 0.384, significantly lower than Autoformer (MSE: 0.560, MAE: 0.498) and Informer (MSE: 0.580, MAE: 0.527). These results highlight the superior performance of our AKTHM in capturing the complex dynamics of power data, making it a valuable tool for accurate power prediction and even for the intractable energy and transient energy prediction in electrical systems. • Propose a KAN–Transformer hybrid model for adaptive nonlinear feature fusion. • Develop a structure-aware loss to jointly learn periodic and trend components. • Experiments on four datasets verify the proposed method improves forecasting accuracy and robustness.