Random-forest based adjusting method for wind forecast of WRF model

Nowadays, machine learning (ML) methods have gained much attention and have been applied in some important related applications in earth science field, including observation data mining, geoscience image recognition, remote sensing image classification and so on. These ML-based applications play important roles in our daily life. However, in meteorological and oceanographic forecast, numerical is still the most popular method. Although researchers have proposed some ML-based prediction methods to overcome the shortcomings of numerical weather forecast methods, the explainability for the forecast result of artificial intelligence (AI) technology is still not as good as numerical weather forecast methods. Therefore, in this paper, we propose a random forest based adjusting method, which introduces AI technology to correct wind prediction results of numerical model. The proposed adjusting method greatly improves the accuracy of forecast results. Furthermore, the physical meanings of parameters in the numerical model are retained in adjusting results. From experimental evaluations, it is obvious that the root mean square error (RMSE) of each feature is reduced greatly. In detail, the average RMSE of 10m wind decreased by more than 45%, and the average RMSE of sea level pressure decreased by more than 50%. It is worth noting that the improvement here is the average of all forecasts for whole region within 7 days. •A RF based adjusting method is designed to correct the output of WRF model.•A DNN model is applied as the feature selector to replace the original data.•The RMSE of wind gets an average decrease of 40% compared with WRF model.•The DNN selected features can improve the adjusting accuracy by 5%.