Interpretable real-time TN soft sensor via multi-periodicity and dimension-aware temporal representation learning:...

Continuous monitoring of stream total nitrogen (TN) is essential for tracking nutrient dynamics and supporting early warning and rapid response. However, long-term in-situ TN monitoring remains unreliable due to biofouling, sensor drift, and frequent data gaps. This study proposes IRIS-TN, an interpretable real-time in-situ soft sensing framework designed for robust TN estimation under dynamic river conditions. First, a data-quality enhancement pipeline was developed to identify abnormal observations and reconstruct corrupted TN records using statistical screening and CatBoost-based regression, leveraging sensor relationships and temporal context to generate physically plausible inputs. Then, the reconstructed data were used to train an edge-deployable soft sensor built upon a frequency-aware temporal modeling architecture capable of capturing dominant periodicities and long-range dependencies in non-stationary river dynamics. Finally, gradient-based attribution analysis was conducted to quantify the contribution of input drivers of hourly TN estimates and to assess whether learned dependencies remain physically consistent across varying hydrological conditions. The results showed that IRIS-TN achieves strong agreement with analyzer-based TN measurements with MAE of 0.191, RMSE of 0.259, and R² of 0.809, consistently outperforming representative recurrent and Transformer-based baselines. Robust performance is maintained across flow-regime subsets, and edge deployment achieves 7.69 ms inference latency. Attribution results indicate dominant reliance on nitrate-related signals with flow-dependent contributions from hydraulic and physicochemical context. Overall, IRIS-TN provides an interpretable and real-time soft-sensing solution with strong potential for reliable deployment in urban river systems.