PA-ResTCN: A Phase-Aware Residual Calibration Model for Radar Tide Gauges

Abstract

Non-contact radar water level gauges are widely used in marine and coastal observation because of their convenient deployment and low maintenance cost. However, in practical operation, radar observations are often affected by installation zero drift, weak temporal lag, wave disturbance, and local nonlinear fluctuations, which leads to systematic deviation from reference gauges. To address this problem, this paper proposes a phase-aware residual temporal convolutional network (PA-ResTCN) for tide level calibration. The proposed framework first performs deterministic lag correction and constant-bias correction according to the physical consistency between radar and buoy-gauge observations, and then learns only the dynamic residual component by a multi-branch temporal convolutional architecture. Specifically, the model contains a trend branch for long-range tidal evolution, a derivative branch for short-term local fluctuation modeling, and a phase feature branch for representing tidal-state priors. A weighted residual-learning objective is further introduced to emphasize turning-point intervals. Experiments were conducted on minute-level paired radar-buoy tide observations collected at a tide station established by our team. The results show that the proposed method reduces the RMSE from 8.526 of raw radar observations to 1.585 on the test set, and further outperforms lag-bias correction, linear residual regression, and vanilla TCN-based calibration models. Ablation experiments verify that residual learning, derivative-aware modeling, and phase-aware auxiliary features are the main sources of performance gain. The proposed method provides an effective and interpretable calibration solution for practical radar water-level monitoring.