Time Series Data Prediction Using CNN-BiLSTM based Attention-KAN Model
Zhiquan He, Zhen Guo · 2024
Time series refers to a sequence of data points arranged in chronological order, typically used to represent the trend of a variable over time. These data points can be continuous, such as hourly temperature records, or discrete, such as annual demographic statistics. Precise time series prediction plays a pivotal role in many areas, for instance, the domain of natural language processing. However, achieving precise predictions for time series data remains challenging. The methods encompass traditional machine learning algorithms, such as Random Forests (RF), as well as recent advancements in deep learning, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), along with various improvements of these neural networks. While these approaches have achieved favorable results in practical time series forecasting applications, they still exhibit some limitations. This paper proposes a hybrid network to improve the accuracy of time series training and prediction. The network is based on one-dimensional Convolutional Neural Networks (1DCNN), Bidirectional Long Short-Term Memory (BiLSTM), dual-head attention mechanisms, and the novel Kolmogorov-Arnold Network (KAN). The model is subsequently applied to the preprocessing, training, and testing of stock price trends as a time series. Comparative analysis with existing CNN-LSTM time series prediction models demonstrates that the proposed CNN-LSTM-KAN hybrid approach achieves improvements in fitting degree, RMSE, R2, and accuracy.