Application of Machine Learning to Statistical Evaluation of Artificial Rainfall Enhancement

Abstract

As an important part of weather modification operation, the scientific effectiveness evaluation of artificial rainfall enhancement has gradually attracted attentions of government and public. In order to evaluate effects of artificial rainfall enhancement objectively and quantitatively, combing linear fitting, polynomial regression, spline regression and 3 other machine learning methods including decision tree, support vector machine and neural network, the relationship model between the rainfall in the target area and the contrast area is established based on rainfall data and operation information of recent 10 years in Fujian. Different historical regression statistical test schemes of rainfall enhancement effects are compared and analyzed, aiming to further optimize the best natural rainfall estimation algorithm based on alterable contrast area with statistical method, which can provide reference for the assessment of artificial rainfall enhancement effects. Results show that historical rainfall data samples are mainly concentrated in the weak rainfall grade. Using multiple regression methods (linear regression, polynomial regression and spline regression), the piecewise statistics of rainfall data does not significantly improve the linear regression model between two regions, and its root mean square error (RMSE) is generally higher than the statistical results. By comparing various machine learning and linear regression models, it is found that CNN and quomial regression perform relatively well when the regional average surface rainfall is taken as the statistical variable, with the determination coefficient of CNN being 0.516 and RMSE being 1.097 mm. Each statistical model is greatly improved after six root square transformations of rainfall data. The performance of the model established by CNN method is relatively optimal, with determination coefficient R2 up to 0.658 and RMSE only 0.2 mm1/2, followed by SVM statistical model based on sixth-order root square transform data, with R2 being 0.41 and RMSE being 0.203 mm1/2. In order to further overcome the time asynchronization and uneven spatial distribution of rainfall in the two regions, the convolutional neural network CNN optimizers (RMSP, ADAM and SGD) are used to establish the contrast-target region rainfall relationship model based on the grid data of natural rainfall plane. Comparison results show that the ADAM optimizer model is the best with the RMSE of 0.61 mm, and its ability to estimate natural rainfall in the affected area is enhanced, by which method the disturbance of the heavy rainfall center will be reduced.

Keywords

• artificial rainfall effect evaluation
• regional historical regression
• machine learning
• statistical test