TCN-BiLSTM-Attention: A Hybrid Neural Network Model for Atmospheric Propagation Delay Prediction Accounting for Nonlinear Time-Varying Characteristics

Zenith tropospheric delay (ZTD) and vertical total electron content (VTEC) are key parameters for describing atmospheric propagation delay. However, high-accuracy reference products are typically released with latency, which limits their use in near-real-time remote sensing applications.

To address the periodicity, short-term fluctuations, nonlinear time-varying behavior, and multi-factor drivers of ZTD and VTEC series, this paper proposes a TCN-BiLSTM-Attention model for atmospheric propagation delay parameter prediction. Historical delay sequences, periodic temporal factors, spatial location information, and external driving parameters are integrated as model inputs. Local temporal variation features are extracted, historical sequence dependencies are captured, and informative time steps are emphasized through the TCN, BiLSTM, and attention modules, respectively.