Global temperature anomaly prediction by using additive twin LSTM networks.

Due to the complexity of climate systems, data-driven modeling based on observed time series data is essential for predicting future climatic trends. This study aims to improve the long-term global temperature anomaly forecast performance of Long Short-Term Memory (LSTM) based neural network models. Although several LSTM variants and hybrid architectures have been suggested for time series data prediction problems, the long-term forecast performance of these models may not be satisfactory in practice. To address solution of these problems, firstly, authors focused on evaluating the forecast performance of models and suggested performance and test assessment procedures. Secondly, authors suggest an Additive Twin LSTM (AT-LSTM) model that can improve the forecast performance for the global temperature anomaly. Our test on the Berkeley Global Temperature Anomaly dataset demonstrates that the proposed AT-LSTM can improve performance relative to conventional LSTM variants in long-term forecasting. Authors observed that global temperature trend projections of the AT-LSTM models for 20 years in future are consistent with expectations of climate organizations and projections in other works. The AT-LSTM models forecasted an average of 1.415 °C with ± 0.073 °C error in the year 2042 and this indicates the strong potential of major climate changes in the near future of Earth.