A Genetic Neural Network ensemble prediction model based on Locally Linear Embedding for typhoon intensity

Based on the nonlinearity of typhoon intensity, the Locally Linear Embedding (LLE) method is employed to reduce the dimensions of the factors obtained from the climatology and persistence (CLIPER) prediction method for predicting typhoon intensity in the Western Pacific Ocean (WPO). Emulating the theory of the numerical weather prediction in ensemble forecast, a new nonlinear artificial intelligence ensemble prediction model has been developed using an evolutionary genetic algorithm (GA) and based on multiple neural networks with the same expected output. Using identical sample cases from typhoon intensity, predictions of the Genetic Neural Network based on LLE (GNN-LLE) are compared with the traditional multiple regression prediction model. Results show that the mean absolute errors of the former in the 12-72h, June-October forecasts are reduced by 8.87%, 16.63%, 22.42%, 20.24%, 28.34% and 23.0%, respectively, in comparison to the corresponding multiple regression models; 93.33 percent of the predictions of the GNN-LLE model are more accurate than those of the multiple regression prediction model. Therefore, for the nonlinear problems, such as typhoon intensity, the GNN-LLE model is better than the traditional multiple regression prediction model and the prediction performance is stable.