科研之友
微信
新浪微博
Facebook
分享链接
摘要
Artificial neural network (ANN) models were developed to predict thunderstorm occurrence within three separate 400 km(2) regions, 9, 12 and 15 h (+/- 2 h) in advance. The predictors include output from deterministic Numerical Weather Prediction models and from sub-grid scale soil moisture magnitude and heterogeneity estimates. The feed-forward multi-layer perceptron ANN topology, with one hidden layer and one neuron in the output layer, was chosen. Two sets of nine ANN models each were developed; one set was developed after a filtering-based feature selection technique was used to determine the predictor subset from 43 potential predictors. The other models were developed based on all 43 predictors. For each of the 18 models, a wrapper technique was used to determine the optimal number of neurons in the hidden layer. Thunderstorm artificial neural network (TANN) model performance was compared to that of multi-linear regression (MLR) models, and to human forecasters (NDFD), based on a novel data set. Results reveal that for several of the nine box/prediction hour combinations with respect to at least one skill-based performance metric, the TANN model's performance exceeded that of the MLR models and NDFD. Yet, the performance of both the MLR models and NDFD were superior to that of the corresponding TANN models in several other cases. Results indicate that the TANN models can provide automated predictions with skills similar to that of operational forecasters. Comparisons of the two sets of TANN models reveal utility in the use of feature selection.
- 出版日期2015-7
