LSTM-UKF framework for an effective global land-ocean index temperature prediction

Abstract

In the recent era of evolution in deep learning, several ANN techniques have been developed to forecast weather conditions. Temperature forecasting is a subset of weather forecasting which employ the use of these ANN techniques to predict the future trend in temperature variation. The modern ANN technologies have facilitated the prediction process due to its virtue of handling lengthy sequences which have been encorporated by many reserachers to train their models. These ANN models require an optimisation technique to mitigate the error between the forecasted and the true value, thus a lot of exploration has been done in the improvement of these optimisation strategies. In this research work, a unique derivative free optimiser known as Unscented Kalman Filter (UKF) is implemented for optimisation along with an upgraded form of Recurrent Neural Network called Long Short Term Memory (LSTM) to forecast the Global Land-Ocean index temperature. In the proposed model, LSTM is used as a base neural network due to its probity to handle very long sequences of data and resolving the issue of vanishing and exploding gradients suffered by RNN. The prediction results obtained from the propsed LSTM-UKF model was compared with Gradient Descent (GD) optimizers such as, Adam optimizer, RMSprop optimizer, and an evolutionary optimisation algorithm called Genetic Algorithm (GA) to demostrate its optimising strength. Well established predictive models such as Gated Recurrent Unit (GRU) and Convolutional Neural Network (CNN) were also trained using UKF and the results were compared with the base model along with the stated optimisation techniques.