KANLINEAR: A Lightweight Model for Multi-Step Feedforward Prediction

This paper introduces KANLINEAR, a lightweight multi-step feedforward predictive model designed for monitoring the operational status of electrical equipment. KANLINEAR is based on a linear model framework and incorporates several innovative elements to enhance its predictive performance. Firstly, we introduced a linear layer based on the Kolmogorov-Arnold Network (KAN), termed KANlayer, which significantly enhances the model's ability to express nonlinear relationships. Secondly, by designing a hybrid strategy for trend decomposition, we have overcome the limitations of traditional moving average methods. Moreover, we have developed a Temporal Volatility Capturing Loss function (TVCLoss), which further enhances the model's ability to capture fluctuations in time series data. Experimental results on an electrical data set demonstrate that KANLINEAR outperforms the baseline model, DLinear, across various performance metrics, particularly in capturing fluctuations in equipment operational status. Ablation studies further validate the effectiveness of these improvements. This research not only showcases the potential application of KANLINEAR in the field of electrical equipment status monitoring but also provides new directions and insights for related research areas.