Short-term natural gas load forecasting using a new informer framework based on temporal kolmogorov-arnold networks

Natural gas load (NGL) forecasting is an important task for enterprises to develop scheduling plans. At present, existing prediction models have disadvantages such as low prediction accuracy and poor stability. In order to solve this problem, an innovative combination model is established in this work. Firstly, to eliminate the influence of noise on the time series of NGL, wavelet transform (WT) is used to decompose it; Secondly, combining the advantages of Informer and TKAN, TKAN is incorporated into the Informer framework; Again, the established prediction model is used to predict the components decomposed by WT and reconstruct them to obtain the final prediction result. To validate the performance of the prediction model developed in this work, data collected from two sites in Greece are studied. The results suggest that: (1) compared with several benchmark prediction models (WT+LSTM, WT+GRU, and WT+CNN), the proposed prediction model has certain advantages in both forecasting stability and accuracy in view of different indicators for different data; (2) the developed forecasting model has improved forecasting accuracy compared to various modules; (3) the proposed prediction model has excellent scalability.