MEDNet: A Mixture of Experts and Dual Sampling Framework for Multivariate Climate Time Series Prediction
Year: 2025
Type: conference-paper
Abstract: In addressing the challenges of global climate change, multivariate climate time series prediction is a critical topic. Climate systems comprise multiple variables such as temperature, humidity, and precipitation, which dynamically interact in complex ways and significantly impact various aspects of human life. Accurately predicting multivariate climate time series is essential for understanding these interactions. However, existing time series prediction models based on Recurrent Neural Network (RNN) and Transformers often exhibit high computational complexity and resource requirements, limiting their applicability in real-world scenarios. To address this, we propose MEDNet (Mixture of Experts and Dual Sampling Network), an innovative and highly lightweight model for long-term multivariate climate time series prediction. MEDNet leverages an expert model allocation strategy and employs two distinct sampling methods—periodic sampling and sequential sampling—to extract correlations from different time points within the same phase and from the same time point across different phases, respectively. An attention mechanism is then used to assign weights to the results of these sampling methods, effectively capturing multi-scale features and global trends in the time series data. To validate the effectiveness of the proposed model, we conducted comprehensive experiments on the benchmark Jena Climate dataset. Compared with several state-of-the-art time series prediction models, MEDNet demonstrated superior performance across all evaluation metrics, achieving a minimum improvement of over 4.6% in Mean Squared Error (MSE) and over 4.3% in Mean Absolute Error (MAE), while significantly reducing computational overhead. These findings firmly establish MEDNet as a superior solution with broad application potential for climate prediction tasks.
Authors: Zeyu Zheng, Shujuan Zheng, Xiaoshuai Zhang
Institution: Huzhou Normal University
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: gold
OpenAlex ID: W4411697315