基于机器学习方法的内蒙古地区格点风场模式预报产品研究

Research on Grid Wind Field Model Prediction Products in Inner Mongolia Based on Machine Learning Method

为了提高格点风场预报产品的准确性,提出了一套基于深度学习和集成学习的风速、风向预报方法。使用时间序列上的站点风实况数据和空间范围上的数值模式预报产品建立时空信息匹配模型,使用了长短期记忆法(LSTM)和极端梯度提升(XGBoost)建立时空匹配的联合预报产品订正模型,形成72小时隔3小时、5千米时空分辨率的SCMOC风速、风向预报产品。评估结果显示,相较于SCMOC预测,联合模型实现风速平均绝对误差(MAE)降低了14.17%,风向平均绝对误差(MAE)降低了23.61%。模型对SCMOC风场产品的风速、风向的准确率有较好的提升,产品释用效果良好。

In order to improve the accuracy of grid wind field prediction products,a set of wind speed and direction prediction methods based on deep learning and ensemble learning is proposed.The real-time data of station wind in the time series and the numerical model prediction products in the spatial range are used to establish the spatio-temporal information matching model,and the Long short-term memory method(LSTM)and extreme gradient lifting(XGBoost)are used to establish the joint prediction product correction model with spatio-temporal matching,forming the 72 hour SCMOC wind speed and direction prediction products with spatio-temporal resolution of 3 hours and 5 kilometers.The assessment results show that the mean absolute error(MAE)of wind speed realized by the joint model is 14.17%lower than that of SCMOC prediction,and the mean absolute error of wind direction is 23.61%lower than that of SCMOC prediction.The model has significantly improved the accuracy of wind speed and direction for SCMOC wind farm products,and the product has a good interpretation effect.