Prediction and Forecasting of Severe Convective Weather based on Hybrid Fully Convolutional Network with Long Short-Term Memory

In modern era, it is essential to accurately forecast the convective weather to reduce risks and improve the operational efficiency in fields such as aviation, agriculture and energy. However, the existing Fully Convolutional Network (FCN) based model faced drawbacks such as, lack of ability to capture temporal dependencies in sequential weather data and class imbalance, which led to poor prediction performance in critical weather events. To address these limitations, a hybrid framework integrating FCN with Long Short-Term Memory (LSTM), namely FCN-LSTM is proposed for enhanced forecasting and prediction of convective weather. Initially, the data of atmospheric from European Center for Medium range Weather Forecasts (ECMWF) is collected. Next, gathered data is preprocessed using data cleaning technique and normalized using min max scaling. Furthermore, the preprocessed data is converted into grid like images by using bilinear interpolation method. Lastly, these grid-like images are fed to the hybrid FCN-LSTM model, where FCN extracts the spatial features and LSTM captures the temporal dependencies to forecast and predict the future weather conditions. The proposed hybrid FCN-LSTM outperform the existing FCN by obtaining better results in context of Critical Success Index (CSI) with 82.7%, Probability of Detection (POD) with 90.5% and False Alarm Ratio (FAR) with 20.7% respectively.