ENVIRONMENTAL MODEL ACCURACY IMPROVEMENT FRAMEWORK USING STATISTICAL TECHNIQUES AND A NOVEL TRAINING APPROACH

Abstract

It is challenging to predict environmental behaviors because of extreme events, such as heatwaves, typhoons, droughts, tsunamis, torrential downpour, wind ramps, or hurricanes. In this thesis, we proposed a novel framework to improve environmental model accuracy with a novel training approach. Extreme event detection algorithms are surveyed, selected, and applied in our proposed framework. The application of statistics in extreme events detection is quite diverse and leads to diverse formulations, which need to be designed for a specific problem. Each formula needs to be tailored specially to work with the available data in the given situation. This diversity is one of the driving forces of this research towards identifying the most common mixture of components utilized in the analysis of extreme events detection. Besides the extreme event detection algorithm, we also integrated the sliding window approach to see how well our models predict future events. To test the proposed framework, we collected coastal data from various sources and obtained the results; we improved the predictive accuracy of various machine learning models by 20% to 25% increase in R2 value using our approach. Apart from that, we organized the discussion along with different extreme event detection types, presented a few outlier definitions, and briefly introduced their techniques. We also summarized the statistical methods involved in the detection of environmental extremes, such as wind ramps and climatic events.