Executive Summary : | Forecasting the rapid intensification (RI) of Tropical cyclones (TCs) over North Indian Ocean (NIO) is a challenging and essential task. Several studies depict that warm ocean, moist middle atmosphere and low vertical wind shear are favourable for having RI in TC, but recent event shows there must a wind surge at surface. Hence physical processes and dynamics associated with RI is not well explored. However, there is a separate Statistical cyclone intensity prediction (SCIP) scheme over NIO but this scheme unable to capture RI. Present endeavour is toward developing a hybrid system that considers the effect of physical processes through numerical model and their statistical inter-relation with RI and set up an algorithm which can learn the data pattern based on this understanding. To evaluate and explore pre-condition of RI, a composite and statistical analysis will be performed with RI systems by appraising mean for different lead hours considering onset of RI as reference time. For non-RI cases, onset of maximum intensity will be considered as reference time. To extract significant physical processes for RI, rotate principal component analysis will be performed to extract key indicator. In all cases, difference in magnitude can be effectively diagnosed using proper hypothesis testing. Through observational analysis, key predictor can be identified, and this predictor can be used in artificial neural network that will be designed through deep learning multi-layer perceptron with best suitable training algorithm. For preparation of input layers, data will be extracted and standardized at each grid points within and on radius starting from 25 km to 500 km with interval of 50 km. With this data, one best ANN-MLP model will be obtained for each radius for each lead hour. Using this model for each radius, an ensemble probabilistic prediction will be prepared for each lead hour. Afterwards each ensemble probabilistic prediction will be evaluated based on error metrics (bias, Root mean square error, mean absolute error, prediction error) and skill metrics (POD, FAR, CSI, Hit rate) along with Willmott’s index, d-index, and prediction error. Evaluation at each lead hour will help us to identify the best prediction hour for RI. However, in real-time prediction, key predictors at different radius of TCs will be taken from numerical modelling which will be assessed through simulations of RI events using different sets of physical schemes using two different ocean-atmosphere couple experiments. Through evaluation, one single model with best configuration will be chosen from each couple experiments and thereafter key parameters will be used to test the best ANN model. For observational analysis and model evaluation, ERA-5 and IMDAA reanalysis will be used. Improvement of the scheme depends on the skill of model which can be improved through assimilation. However, improvement of scheme will be not part of this study. |