Performance Comparison Analysis on Weather Prediction using LSTM and TKAN

Abstract

The development of machine learning methods in the last few decades has shown great potential in various predictive applications, including in domains such as financial prediction, medical diagnosis, and big data analysis. One of the most widely used methods in prediction tasks is Long Short-Term Memory (LSTM). LSTM has become popular because of its ability to handle time series data by retaining relevant information in the long term and the ability to forget irrelevant information through the forget-gate mechanism. However, along with the development of technology and the need to improve accuracy and efficiency, new methods such as the Kolmogorov Arnold Network (KAN)  have emerged. KAN was then developed into the Temporal Kolmogorov Arnold Network (TKAN), which was designed to match or even surpass the performance of LSTM. The TKAN architecture has produced significant improvements in the management of both new and historical information. Because of this capability, TKAN can excel in multi-step predictions, demonstrating a clear advantage over conventional models such as LSTM and GRU, particularly in the context of long-term forecasting. This research goal is to give insight into the comparison of both the TKAN and LSTM models for weather prediction using model loss and mean absolute error evaluation (MAE). The model for both LSTM and TKAN achieved 0.09 and 0.11 for model loss and 0.08 and 0.96 for MAE.