The benefits of using raw CML attenuation data to predict urban runoff

Commercial microwave links (CMLs) have recently shown great potential in urban drainage modelling due to their ability to provide rainfall-runoff dynamics. Studies investigating potential of CMLs to improve rainfall-runoff modelling typically used mechanistic hydrodynamic models driven by quantitative precipitation estimates (QPEs) derived from CML attenuation data. Naturally, some errors are introduced, primarily related to CML rainfall retrieval model, including uncertainties in wet antenna attention correction, as well as errors originated from path-averaged character of CML QPEs. These processing steps not only generate some new uncertainties but also result in a loss of valuable information contained in raw data. Besides, mechanistic models require high-quality pre-processed input rainfall data, which adds complexity to the application. We address these issues by employing raw CML attenuation data without QPE derivation using a data-driven discharge model.

A Random Forest (RF) model is employed to estimate 2-minute urban runoff using the raw CML data. The study area is a small urban catchment (1.3 km²) with a lag time of approximately 20 minutes. Datasets consist of 1-minute rainfall data from 3 rain gauges, 10-second CML data from 14 CMLs and 2-minute flow data collected during the year 2014 to 2016. A calibrated SWMM hydrological model driven by the 3 local rain gauges is used as a benchmark.

 We find that: (1) Compared with rainfall data as inputs, CML attenuation data performs equally well or better in runoff simulation. The RF model with CMLs inputs achieves NSE of 0.90, PCC of 0.95, RMSE of 0.03 m³/s, and MAE of 0.02 m³/s; (2) The RF model produces comparable results to the SWMM model benchmark; (3) In addition, the RF using CML data can be used for runoff prediction exceeding horizon of the lag time. It accurately forecasts runoff up to 40-minute ahead, with NSE greater than 0.77 and PCC exceeding 0.88, whereas the RF using rain-gauge data struggles to forecast runoff more than 30-minute ahead. These results demonstrate that CML raw data can accurately yield runoff dynamics and volumes, and can be used for short-term runoff predictions.