Hybrid pre-classification technique – artificial neural network for lightning severity classification

This thesis is presents the classification of lightning severity from meteorology characteristic using the computational intelligence; the Artificial Neural Network (ANN). The meteorology parameters used are very basic and economical as it is designed for public. The targeted user group is for those who have a higher risk to be strike by lightning and also for those users without any meteorology background. Examples of these targeted user groups are recognized as those who enjoys outdoor activities, the event organizer, building maintenance workers, and skyscraper crane operator. This group of user is prone to lightning strikes since their working environments are constantly exposed to the lightning strikes possibility. The weather forecast broadcasted on mass media does not fully describe the condition of the daily weather qualitatively. Hence, the qualitative interpretation given to the public usually too general and does not provide sufficient information needed, in this case the lightning severity information. Therefore, by analyzing the meteorology parameters quantitatively, the severity of lightning can be determined, thus revealing the risk of lightning strikes on that particular day. This piece of information may benefits user in order to avoid the risk of casualties and property losses due to lightning. During the study, three objectives are listed. First objective is to establish a practical scale; the Daily Lightning Severity Scale (DLSS). Second, the application of ANN in classifies the severity of the lightning. And third, to propose and test a new technique of separating data for ANN Training, Validation and Testing (TVT) datasets, known as PreClass Test (PrCT) technique. The study outcome revealed that the proposed scale of DLSS is practical to be used for the study area. The DLSS listed out four levels of lightning severity denoted as Safe, Normal, Frequent, and Very Frequent. While developing ANN, two networks were prepared for this study based on two datasets, known as RandSet and PrCTSet. The RandSet utilize common method of separating the TVT dataset using random separation ratio whilst the PrCTSet applied the new proposed technique for TVT separation. The result indicates the PrCT techniques have faster training result at approximately 50% reduction of number of epochs required and shortening almost 50% of training time compared to random separation method. It was observed that networks developed from both datasets yields good performance. PrCTSet score 92.9% of accuracy, while the RandSet network scores similar accuracy at 92.9%. It is suggested that the PrCT method is suitable for ANN application which requires faster training time and at minimal computational effort.

Preview Text FK 2014 133IR.pdf Download (1MB) | Preview

Actions (login required)

View Item