Short-Term Regional Temperature and Salinity Prediction Based on Deep Learning Long Short-Term Memory

Abstract

The El Nino-Southern Oscillation (ENSO) is the most significant year-to-year climate variation, affecting weather and climate systems worldwide. However, current prediction models, both dynamic and statistical, struggle with accuracy due to the complex mechanism of ENSO. This study introduces a regional temperature and salinity prediction method using a Long Short-Term Memory (LSTM) deep learning model, which is well suited for identifying long-term patterns in sequential data. The model is applied to three specific regions using in-situ data from Argo floats: the centraleastern Pacific, the central tropical Pacific Nino 3.4 region, and the Western Pacific Warm Pool (WPWP). These regions are chosen because they play key roles in ENSO dynamics. Results show that the LSTM model performs best in the WPWP, where the average mean squared error (MSE) is low (0.03), indicating high accuracy and stability. This is likely due to lower noise in the original data. In contrast, the model performs poorly in the central-eastern Pacific, where the average MSE is much higher (7.03), suggesting instability due to high noise in original data. These findings highlight the potential of deep learning for regional climate predictions and suggest that LSTM models could improve local weather forecasting and fisheries management.