A Novel LSTM-KAN Networks for Demand Forecasting in Manufacturing

This paper proposes a novel LSTM-KAN networks that combines Long Short-Term Memory Networks (LSTM) and Kolmogorov-Arnold Networks (KAN), and tests the performance of this network in handling time series forecasting tasks. By comparing the performance of this integrated neural networks with traditional LSTM networks in handling the product demand forecasting task, we found that it has significant advantages in training efficiency. In addition, the prediction accuracy is also improved by an average of $\mathbf{8 . 2 6 \%}$ compared to traditional LSTM networks. This study verifies the feasibility of KAN in processing time series forecasting tasks, and it is valuable for exploring the performance potential of KAN.