A Multi-objective Residual TrajGRU Model for Wind Field Forecasting
- Resource Type
- Conference
- Authors
- Zhang, Wei; Jiang, Yueyue; Song, Xiaojiang; Guoan, Boyu; Pang, Renbo
- Source
- 2022 IEEE International Conference on Big Data (Big Data) Big Data (Big Data), 2022 IEEE International Conference on. :4893-4900 Dec, 2022
- Subject
- Communication, Networking and Broadcast Technologies
Computing and Processing
Engineering Profession
Geoscience
Robotics and Control Systems
Signal Processing and Analysis
Training
Wind speed
Predictive models
Big Data
Feature extraction
Data models
Numerical models
wind forecast
wind components
multi-objective
deep learning
- Language
Numerical weather prediction models inevitably exhibit systematic bias when forecasting winds. In the present era of big data, model improvements are essential. We subjected wind data to deep learning and built an efficient data-driven model. We devised the "Multi-Objective Optimization ResTrajGRU" (MOO-ResTrajGRU) model to handle the three-dimensional (3D) spatiotemporal sequences of the U- and V- components of the wind field. The model has an encoder-forecaster architecture, and a residual connection mechanism that effectively extracts spatiotemporal features at different scales. The periodic characteristics of directional data are solved by modeling the U- and V-components; an additive loss function simultaneously optimizes wind speed and wind direction accuracy. We forecast wind fields during the four seasons of the western North Pacific (WNP) at scales of 12–120 h. The forecast data were more accurate than those of other models; the multi-objective mode of our model halves the training time and storage space requirements with little effect on performance.