Improving Climate Change Predictions Using Time Series Analysis and Deep Learning

limate change forecasting can be considered one of the important branches of climatology as accurate estimates of future climate situations are needed to plan timely and such interventions which will minimize the impacts of change in the climate. General circulation models, or GCMs in short, although they are commonly used, have a problem with their ability to perform long term forecasts as they tend to be hugely resource demanding. In the present study, we aim to further examine climate time series data and its improvement through deep learning. More specifically, we implemented and tested two specific models: a Basic Long Short-Term Memory (BLSTM) model and an Autoregressive Long Short-Term Memory (AR-LSTM) model. The Basic LSTM model is developed to capture long-term dependencies in time series data, but it does not specifically account for linear models of the time series. The assumption about linear relations between variables in the AR-LSTM is removed. It includes autoregressive elements to model such relations, while LSTM units are reserved for representation of non-linear dependencies.