Accurate and reliable prediction of geomagnetic data is essential for its management and application. To improve the accuracy of geomagnetic data prediction and overcome the limitations of traditional machine learning algorithms in network structure and prediction time, a hybrid prediction model combining wavelet packet decomposition (WPD) and the DLinear model is proposed. The model uses WPD to decompose the original geomagnetic data in multiple sub-layers, effectively eliminating high-frequency noise while retaining low-frequency signal input to the DLinear model to complete accurate time series prediction. To verify and compare the prediction efficiency of the proposed model with those of other models, four sets of geomagnetic data collected by the Beijing Geomagnetic Observatory in China were used to perform single- and multi-step prediction experiments. The experimental results showed the following: the proposed hybrid model had the best prediction performance in one-, three- and five-step predictions, and with the introduction of the WPD method, the prediction accuracy of the original model was effectively improved. The combination of WPD and deep learning technology in this study provides a new way to predict future geomagnetic data with high precision and reliability.
