Search

This letter presents an impressive optimization method for determining the optimal model hyperparameters of a deep neural network (DNN) targeted to model the characteristics of antennas. In this paper we propose an innovative approach of efficient yield analysis for modeling and sizing antennas. It is based on the long short-term memory (LSTM) DNN aiming to forecast the extended frequency responses, where various stochastic methods are applied for determining the optimal hyperparameters while training a DNN. Among the various methods, the one which models the antenna accurately in terms of input scattering parameter, gain, and radiation patterns is the winner. The proposed method is compact and addresses the problem of heavy reliance to the designer experience in determining the hyperparameters. Additionally, forecasting the future frequency responses of the antenna reduces the designers effort substantially in measuring large frequency band; hence, measuring whole frequency band would not be needed. For validating the effectiveness of the proposed method, the fabricated two element antenna array is used for modeling where the results demonstrate that the Thompson sampling (TS) algorithm can determine optimal hyperparameters with minimum error in comparison with other reported stochastic methods leads to predict the future frequency band accurately. IEEE

This study aimed to evaluate the estimation accuracy of rainfall erosivity (R-factor) in Italy and Switzerland through five Machine learning (ML) models (Decision Tree (DT), K-Nearest Neighbors (KNN), Random Forest (RF), Gradient Boosting (GB), and eXtreme Gradient Boost (XGB)) tuned with optimal hyperparameters. To build the ML model, high temporal resolution (HTR) rainfall data were collected from 297 rain-gauge stations located in the study area. To estimate the RUSLE-based R-factor through the models, the rainfall amount for each rainfall event, the rainfall duration, and the maximum 60-min intensity were used as input data. The datasets for training/validation and testing consisted of rainfall data from 287 and 10 stations, respectively. In a second phase, each ML model was trained through 10-fold cross validation based on training and validation data. For hyperparameter adjustment, the models were optimized using the Bayesian optimization algorithm (BOA). The R-factor estimation performance of each ML model through cross validation improved from 6.1% to 62.8% as hyperparameters were optimized through BOA. In particular, ensemble models such as RF, GB, and XGB were superior to other models with an accuracy performance of 0.9 or even more. And the RF showed an excellent estimation performance (R 2 = 0.965, NSE = 0.958, RMSE = 44.993 MJ mm ha−1h−1, and MAE = 13.901 MJ mm ha−1h−1) for test stations, followed by GB and XGB with similar performance. However, the R-factor for the extremely intense rainfall event estimated by the ML models showed a significant difference from the RUSLE-based R-factor. This result implies that although the ML model built in this study can reasonably estimate the R-factor in the general rainfall event, additional training and validation through securing various rainfall event data is required to improve estimation accuracy on an extreme rainfall event.