Statistical Deep Learning-based Spatio-temporal Long-term Forecasting Method

Sea surface temperature (SST) is an important parameter of ocean hydrography, and accurate forecasting of SST is of great significance for ocean economic development, extreme weather prevention and marine vessel navigation. Therefore, the rapid development of SST forecasting is necessary. To address the characteristics of SST time series data with nonlinearity and multiple noise, this paper use EOF decomposition method to reduce the influence of data dimensionality and noise on prediction results. By constructing the Conv-BiLSTM prediction model, the convolutional neural network (CNN), bidirectional long-and short-term memory network (BILSTM) and deep neural network (DNN) are effectively combined. The SST spatio-temporal sequence evolution features are also extracted, thus effectively improving the SST long-term prediction accuracy. The model is statistically compared with the baseline model and the conventional machine learning model in a forecast experiment at four independent stations in the Pacific Northwest. The results show that the model not only has a significant improvement in SST forecast accuracy, but also has better robustness.