Zhengqing Liu

Mr. Zhengqing Liu


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
27 October 2025 in Research Square
Verified
Preprint
Open access
PDF file
ZLZhengqing Liu
Zhengqing Liu
JYJunfeng Yang
Junfeng Yang
DLDan Liu
Dan Liu
JWJianmei Wang
Jianmei Wang
YZYiming Zhang
Yiming Zhang
HXHu Xiong
Hu Xiong
Artificial intelligence (AI) weather forecasting has advanced rapidly owing to its high prediction accuracy and exceptional computational efficiency. However, research on stratospheric forecasting and multi-step prediction strategies remains relatively underdeveloped compared to studies focused on the troposphere or model architecture improvements. Stratospheric wind plays a crucial role in the flight performance of balloon experiments. Although Numerical Weather Prediction can generate forecasts by solving atmospheric partial differential equations, its expensive computational cost fundamentally limits the capability of short-term prediction. In recent years, artificial intelligence technology has been increasingly applied to atmospheric predictions. In this study, we build convolutional neural network (CNN) and long and short-term memory (LSTM) network models for single point (95.5° E, 37.5° N) and area (90–100° E, 30–40° N) predictions of the short-term zonal wind in the stratosphere. The convolution architecture is found to outperform the LSTM models. Four prediction strategies including multi-input multi-output (MIMO), DirRec, and hierarchical time aggregation (HTA) and DirRec are implemented and compared to achieve multi-step forecasting. MIMO, DirRec, and HTA strategies are found to have higher prediction accuracy than the recursive strategy, and the DirRec strategy requires significantly higher computational costs. Hence HTA and MIMO are considered more suitable in stratospheric wind pre-diction. The HTA strategy is better for forecasts in 1-6th steps, while the MIMO strategy is optimal for forecasts in 7-12th steps. The HTA performs faster in training time than multi-input multi-output, but slower inference time. Our comparison is helpful for selecting appropriate strategies for different neural network–based forecast scenarios.