The value of streamflow data in improving TSS predictions - Bayesian multi-objective calibration

- To better model TSS we explore the benefits of Bayesian multi-objective calibration.
- We describe model structure error and input error as an autocorrelated error process.
- Including streamflow data into model calibration gives more reliable TSS predictions.

SummaryThe concentration of total suspended solids (TSS) in surface waters is a commonly used indicator of water quality impairments. Its accurate prediction remains, however, problematic because: (i) TSS build-up, erosion, and wash-off are not easily identifiable; (ii) calibrating a TSS model requires observations of sediment loads, which are rare, and streamflow observations to calculate concentrations; and (iii) predicted TSS usually deviate systematically from observations, an effect which is commonly neglected. Ignoring systematic errors during calibration can lead to overconfident (i.e. unreliable) uncertainty estimates during predictions. In this paper, we therefore investigate whether a statistical description of systematic model errors makes it possible to generate reliable predictions for TSS. In addition, we explore how the reliability of TSS predictions increases when streamflow data are additionally used in model calibration. A key aspect of our study is that we use a Bayesian multi-output calibration and a novel autoregressive error model, which describes the model predictive error as a sum of independent random noise and autocorrelated bias. Our results show that using a statistical description of model bias provides more reliable uncertainty estimates of TSS than before and including streamflow data into calibration makes TSS predictions more precise. For a case study of a small ungauged catchment, this improvement was as much as 15%. Our approach can be easily implemented for other water quality variables which are dependent on streamflow.

116