Modeling Global Ice Volume Changes: A Nonlinear Autoregressive Neural Network Approach

A single time series prediction problem is solved with a neural network.The nonlinear autoregressive (NAR) type of network is used.The network is trained in an open loop and then transformed to closed loop for multistep prediction.The prediction is made 20 time steps into the future.The delay is removed from the network to get the prediction one time step earlier.The shallow neural network is trained on the global ice volume dataset, which contains 219 measurements of global ice volume over 440,000 years.The network is able to predict future ice volume based on past values with a high degree of accuracy.Three different backpropagation training algorithms were used to train the network: Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient.The Levenberg-Marquardt algorithm achieved the lowest MSE (0.02257 at epoch 13) and the highest R (0.99254).The Bayesian Regularization algorithm achieved an MSE of 0.027209 at epoch 4 and an R of 0.99192.The Scaled Conjugate Gradient algorithm achieved an MSE of 0.01878 at epoch 3 and an R of 0.99018.This work contributes to the field of climate studies by providing a tool for predicting future ice volume.This information can be used to better understand Earth's glacial cycles and to develop strategies for mitigating the effects of climate change.