A High-Precision Ionospheric TEC Forecasting Method Combining Large Language Models and Cross-Modal Temporal Framework

Objectives: Accurate and reliable forecasting of ionospheric total electron content (TEC) holds significant research value for navigation, communication, and related domains. Traditional mathematical methods face challenges in precisely predicting the nonlinear and complex dynamic systems of the ionosphere.Methods: We pioneers the application of cutting-edge large language models (LLMs) including DeepSeek, Qwen, GPT2, and Llama, integrated with a cross-modal temporal forecasting framework AutoTimes, to construct a high-precision intelligent TEC prediction model for ionospheric applications. We systematically investigated and analyzed the model's global ionospheric TEC forecasting capabilities during both solar minimum and maximum periods. Using post-processed ionospheric TEC products from the Center for Orbit Determination in Europe (CODE) as training data and reference values, we conducted the first comprehensive evaluation of different LLMs' performance in ionospheric TEC forecasting.Results: Experimental results demonstrate that the LLM-based cross-modal framework significantly outperforms CODE's C1PG (CODE's 1-Day Predicted Global Ionospheric Map) forecasting product. The maximum correlation coefficient between predicted values and observational ground truth reaches 0.976 5. The optimal root mean square error (RMSE) values for solar minimum and maximum periods are 1.843 TECU (total electron content unit) and 5.577 TECU respectively, representing reductions of 10.8% and 10.4% compared to C1PG products.Conclusions: The proposed method establishes a new benchmark for ionospheric forecasting, demonstrating superior adaptability across different solar activity phases, and it provides a robust theoretical foundation and methodological framework for subsequent studies in ionospheric TEC prediction.