Control of Rainwater Harvesting Systems: Optical-Flow based Radar Nowcasting

Sammanfattning

Urban water systems in southern Sweden face increasing pressure from both short- duration intense rainfall, which can cause pluvial flooding and pollutant discharges, and more frequent dry periods, which motivate water reuse (IPCC 2023; Schimanke et al. 2022; Canedo Rosso et al. 2025). Rainwater harvesting (RWH) tanks can contribute to both goals (Campisano et al. 2017), but their detention benefit is reduced when storage is unavailable ahead of storms. Forecast-based control (FBC) can create capacity through pre-storm releases (Kerkez et al. 2016; Xu et al. 2022), but its viability depends on precipitation forecast skill at short lead times. This thesis investigates whether high-resolution radar-based nowcasting can provide forecast information of sufficient quality to support FBC of RWH systems in the Malmö area. Using the PySTEPS framework (Pulkkinen et al. 2019), four extrapolation-based nowcasting configurations were generated from Swedish radar composites and evaluated for 19 intense precipitation events at 0.5 km / 5 min resolution, with additional verifi- cation at the 2 km / 15 min configuration used by SMHI’s operational KNEP product (SMHI 2025a). Forecasts were compared against Eulerian and Lagrangian persistence baselines using continuous, categorical, and spatial verification metrics over lead times of 5–90 min. Nowcast skill depends strongly on season and event duration. Winter events retain useful correlation beyond 60 min for 3 h durations, whereas convective summer events decorrelate within about 15–30 min, consistent with the conventional decorrelation framing based on Pearson correlation (Germann et al. 2002). At the lead times rele- vant to control, S-PROG at native resolution and ANVIL at KNEP-equivalent reso- lution maintain probability of detection above 0.7 and frequency bias close to unity; the latter supports the operational choice of ANVIL for KNEP (Falahat 2026). The direct effect of increasing radar resolution within a fixed method is modest, but the best-performing method changes with resolution: S-PROG performs best at native resolution, while ANVIL performs best at the coarser operational configuration. Interpreted against required tank drawdown times from parallel emptying-time sim- ulations (Sundstedt 2026), the results indicate that 30–90 min nowcast horizons can support operational pre-storm releases under permissive outflow strategies and urban- scale outflow constraints, while stricter constraints and strategies requiring a fully empty tank at event onset require longer horizons and motivate blending with numer- ical weather prediction.