Towards Cost-effective and Interpretable Sea Level Prediction: MEMS Sensors and Temporal Fusion Transformer

Yiyuan Zhang

Abstract: Accurate sea level prediction is crucial for coastal communities facing rising sea levels and increasing storm-surge threats. However, the black-box nature of existing predictive models limits their acceptance in operational decision-making contexts, creating an urgent need for transparent, cost-effective, and reliable forecasting solutions. Here, we propose a novel framework that combines Micro-Electro-Mechanical System (MEMS)-based ocean sensors with an interpretable machine learning approach to address this need. Utilizing two months of data collected from a MEMS accelerometer array on submarine cables, our approach integrates Variational Modal Decomposition (VMD) to isolate key temporal patterns from raw sensor measurement data, and Principal Component Analysis (PCA) to optimize feature selection. These refined features with meteorological and calendar information are fed into an interpretable machine learning model, Temporal Fusion Transformer (TFT), for sea level prediction. The TFT provides interpretable outputs, including the importance ranking of input variables and attention distribution of different time steps. Our hybrid VMD-PCA-TFT model achieved a root mean square error of 1.57 cm and a coefficient of determination (R²) of 0.98 compared to tide gauge records, outperforming existing models. These results demonstrate that MEMS-based ocean measurement systems can match the accuracy of traditional methods at a fraction of the cost. Moreover, the TFT’s interpretable analysis reveals that the VMD-PCA features accounted for more than 80% predictive power, provide transparent insights into prediction mechanisms. This dual advantage of cost-effectiveness and interoperability of our framework could potentially promote widespread deployment of ocean monitoring systems and advance prediction capabilities for climate change adaptation.