Forecasting of atmospheric variables based on ECMWF analysis data using Machine Learning approaches

Recent advancements within the arena of Artificial Intelligence have widened the potential applications of Machine Learning (ML) frameworks in climate prediction and weather forecasting. For any modern forecasting system, a core objective is linked with handling uncertainty and scientists are interested in the accuracy of the forecasts.  The time series forecast of air temperature using ML approaches is available in the literature. But for this study, we have selected major atmospheric variables- air temperature, dew point temperature, wind components and mean sea-level pressure (MSL-P) retrieved from the ECMWF analysis system and which are to be used in perturbation of the ocean forecasting system. In our previous approach, we analysed the probability distributions of the selected atmospheric variables. In this study, we intend to forecast those atmospheric variables using machine learning algorithms to compare with the analysis dataset produced by ECMWF. Under the initial approach, a Convolution Neural Network (CNN) approach is built to predict the time series for the atmospheric variables. The predicted results from the forecasts have shown minimal differences in comparison to the observations.  Based on the results produced from the CNN, we would like to apply other ML approaches to compare the accuracy and in the process of selecting a better ML model.