A Deep auto encoder based Framework for efficient weather forecasting

Weather forecasting has plethora of benefits in different domains. Traditional weather forecasting approaches applied science and technology towards predicting weather conditions in given place and time. With the emergence of Artificial Intelligence (AI) there are increased possibilities in the area of weather forecasting research. Instead of ground level observations, AI approaches learn from historical data and also current atmosphere data to come up with predictions. We suggested a framework for autonomous weather forecasting based on deep learning. Our framework is a variant of Convolutional Neural Network (CNN) model which exploits encoder and decoder to learn parameterizations from the given data and forecast weather. The proposed model is capable of interpreting spatial information associated with geopotential field and automatically infers forecasting knowhow with higher accuracy levels. A variable selection process is incorporated to determine geopotential height that has impact on the weather conditions. We proposed an algorithm known as Deep Weather Forecasting (DWF) to realize the proposed framework. Our empirical study has revealed that the proposed framework is used to evaluate different deep learning models and comparing their performance. Our deep learning models outperformed many existing regression models. U-Net showed highest performance with least MAE 0.2268 when compared with all other models.