The Study on Sea Surface Temperature Prediction Based on TCN-LSTM Model

This study presents a novel hybrid model-TCN-LSTM, based on temporal convolutional networks (TCN) and long short-term memory networks (LSTM), aimed at improving the prediction accuracy of sea surface temperature (SST). Through sensitivity experiments conducted at four ocean observation stations, we analyzed the impact of hyperparameters on model stability. By controlling key parameters such as the number of iterations, TCN kernel size, and LSTM neuron count, and employing ANOVA variance analysis, we found that all p-values were greater than 0.05. This indicates that these parameters have no significant effect on the Mean Absolute Percentage Error (MAPE), thus supporting the stability of the model. The TCN-LSTM model effectively combines the advantages of TCN in local feature extraction and LSTM in capturing long-term dependencies, enabling it to learn important features from SST data. In predicting sea surface temperature for the next seven days, the model demonstrates excellent adaptability and generalization capabilities, with outstanding performance in relevant evaluation metrics. Furthermore, by comparing the predictive performance across different combinations of hyperparameters, we validated the model's consistency and reliability. In summary, this research provides an innovative methodological framework for SST prediction. Although it mainly focuses on univariate time series models, future studies will explore multivariate models and spatiotemporal feature extraction to further enhance prediction accuracy.