A SIMPLE PREDICTION METHOD FOR TEMPORAL CHANGES IN SEA ICE THICKNESS USING RNN AND LSTM

Abstract

 This study aims to provide more accurate forecasts of future sea ice thickness in response to climate change, and to track past variations in sea ice thickness for statistical evaluations of these trends. Therefore, we proposed a method using neural networks (RNN, LSTM) for predicting the temporal changes in sea ice thickness and the heat balance on ice, which serve as indicators of coldness, from meteorological data including solar radiation and temperature, and investigated its accuracy. We found that even a simple RNN could make good predictions for thermal balance with daily periodic variations, and that LSTM, with its high capability for long-term memory in sequential data, could also predict relatively well for sea ice thickness without periodic variations. Furthermore, we confirmed a further improvement in accuracy by adding the cumulative value of atmospheric temperature data (Freezing degree days) as input data, by normalizing the input data, and by evaluating with ensemble averaging using multiple initial weights.