A Hybrid Approach of Weather Forecasting using Data Mining

Abstract

In the paper, the work focuses on weather prediction by using real time data from day to day. Weather Prediction has proven to be a very important application of Machine Learning since the beginning. Different models were studied and found out ways how prediction could be made more accurate by aban-doning the classical models and adopted a hybrid method of including more than hundred decision trees bagged to form an aggregate total. The aggregate results achieved from each tree was considered to be a random split of data, saving a lot of computation time. Gradient Boosting was used to increase accuracy significantly making it a very efficient model to work with. The boosting helped the weak learner Decision Tree to select a random sample of data, fit it with a model and train it sequentially to compensate for the weakness of its predecessor. To improve the accuracy of a model in boosting, a combination of a convex loss function, which measures the gap between the expected and goal outputs, and a penalty term for the complexity of the model were used to reduce a regularized objective function that included both L1 and L2 regression tree functions. The resulting model achieved a significantly high level of accuracy when tested with new data.