Combining Seasonal and Trend Decomposition Using LOESS With a Gated Recurrent Unit for Climate Time Series Forecasting
Xiao Liu, Qianqian Zhang
Abstract
Temperature, as a key indicator of climate change, is a constant object of research focus due to its importance in accurate forecasting. Traditional meteorological models have limitations in handling complex temperature data, and deep recurrent network techniques, known for their excellent performance in capturing long-term dependencies in time series, offer new possibilities for climate prediction. This paper adopts a decomposition strategy to address the temporal features of meteorological data and proposes a decomposition-prediction model combining seasonal and trend decomposition using LOESS and gated recurrent unit (STL-GRU) neural networks. By applying this model to datasets from three different regions in Gansu Province, China, the effectiveness of the model is demonstrated. The results show that the combination of decomposition methods and deep learning techniques improves the accuracy of seasonal variations and long-term prediction trends of temperature data, and the root-mean-square errors of the proposed model in the prediction of three real surface temperature datasets are 1.7707, 1.2681, and 1.4166, respectively.