A multi-physics ensemble approach for short-term precipitation forecasts at convective permitting scales based on sensitivity experiments over southern parts of peninsular India

Abstract

The southern peninsular India is characterized by unique climatology with rainfall processes throughout the year from land–ocean contrasts. In addition, the complex terrain induces localized effects causing huge spatial and temporal variability in the observed precipitation. This study aims at evaluating the sensitivity of the high-resolution Weather Research and Forecasting (WRF) model (4 km) to multi-physics parameterizations, 3D variational data assimilation, and domain configuration, in the study domain covering southern peninsular India. Furthermore, the study focusses on the formulation of an ensemble method to improve the simulation of precipitation across seasons. A total of 120 experiments were set up across four crucial rainfall events, of varying spatial extent and duration, dominated by different rainfall generation mechanisms. The assessment of the experiments shows that the model’s cumulus and microphysics schemes have the highest impact on the location, intensity, and spread of the simulated 4-day long Quantitative Precipitation Forecasts (QPFs). Applying cumulus schemes at all domains represented the variability in the QPFs, across space and time, for the precipitation events dominated by convective activity. The cases without cumulus schemes at the convective scale domain (4 km), captured the higher intensity rains during organized cyclonic circulations in the north-east monsoon period. Hence, a 10-member multi-physics ensemble approach including members with and without cumulus parameterization at the fine resolution domain was adopted. The preliminary results demonstrate that the mean from the suggested ensemble approach (n-MPP) performed well in capturing the dynamics of QPFs across the rainfall events, as opposed to a single-member deterministic simulation and mean from larger member conventional multi-physics ensemble approach (c-MPP) without cumulus parameterization at the convective scale. The rank histogram, delta semi-variance plots, and outlier statistics at various lead times clearly showed that the suggested n-MPP was able to capture the high-intensity rainfall, increasing the spread of precipitation forecasts and consequently reducing the occurrence of outliers.

Highlights

• Evaluated the sensitivity of a high-resolution Numerical Weather Prediction (NWP) model (4 km) with 120 experiments in generating short-term (4-day long) quantitative precipitation forecasts (QPFs).

• The cumulus and microphysics schemes have the highest impact on the location, intensity, and spread of the simulated rainfall across events dominated by different rainfall generation mechanisms.

• A 10-member multi-physics ensemble approach including members with and without cumulus parameterization at the fine resolution domain was able to capture the high-intensity rainfall, increasing the spread of precipitation forecasts and consequently reducing the occurrence of outliers.

Appendices

Appendix

A.1 Formulation of FSS

Fractional skill score is helpful in determining the usefulness of a forecast for a given rainfall threshold and neighbourhood size. The higher the neighbourhood size, the FSS for same rainfall threshold will be high, since it averages spatially a larger areal extent, thus removing the errors due to spatial shift. The measure is subjective in the sense the score will depend on the selection of neighbourhood size and the rainfall threshold. In general, the FSS is derived for different rainfall thresholds and neighbourhood sizes. In this study, as mentioned in section 3.2, 5 × 5 pixels were selected as neighbourhood size (neighbourhood scale of 50 km) after testing neighbourhood scales of 20, 30, 50, and 70 km. Four different threshold levels, viz., 1 mm/3 hr, 2.5 mm/3hr, 5 mm/3hr, and 10 mm/3hr rainfall rates were selected based on the intensities of the selected events. The binary map (if the value of a pixel in the raster is above the selected threshold, the value 1 will be assigned to the pixel and vice versa) for four rainfall thresholds was derived for both observed and simulated average rainfall rate raster for every event. Later a non-overlapping 5×5 size window (50 km neighbourhood scale) was moved across the raster and the fractions within the window were calculated. The FSS for every sensitivity experiment was computed using equation (A1).

$$ {\text{FSS}} = 1 - \frac{{\frac{1}{N}\mathop \sum \nolimits_{1}^{N} \left( {P_{f} - P_{o} } \right)^{2} }}{{\frac{1}{N}\left[ {\mathop \sum \nolimits_{1}^{N} P_{f}^{2} + \mathop \sum \nolimits_{1}^{N} P_{o}^{2} } \right]}}, $$

(A1)

where $P_{f}$ is forecast fraction, $P_{o}$ is observed fraction, and N is the number of spatial windows in the evaluation domain.

A.2 Formulation of S–A–L

The quality measure S–A–L (as suggested by Wernli et al. 2008) has been designed to effectively assess the NWP performance of simulating single-type precipitation within a defined domain like a river basin. Furthermore, the metric does not require subjective decisions on the threshold or the size of the neighbourhood like FSS. The components of S–A–L, i.e., structure, amplitude, and location are derived based on the order of complexity which goes from A to L and then finally to S. The implementation of S–A–L involved the following steps:

• (1) Define domain: In the study, the evaluation domain 3 was selected as the domain for the derivation of S–A–L.

• (2) Identifying rain objects: As S–A–L is an object-based metric, the precipitation objects need to be defined. The threshold value (R*) to select the object was defined by $R{*} = \frac{1}{15}R^{95}$, where $R^{95}$ is the 95th percentile of all the values within the domain, calculated separately for each precipitation field considered. Binary map with 1 representing the values beyond $R{*}$ was derived. An algorithm was formulated to identify spatially connected cells with value 1 in the binary map and was declared as rain object.

• (3) Amplitude: A was calculated as the normalized difference of domain averaged precipitation between simulated and observed precipitation field. The values range from –2 to 2 are analogous to bias metrics (grid-based metrics). Positive amplitude values denote over-estimation, and negative values represent under-estimation of the simulated rainfall. Value 0 represents perfect prediction.

• (4) Location: The L component comprises two additional parts. The first part (L1) accounted for the distance between the center of mass of the simulated and observed precipitation fields. L1 ranges from 0 to 1 with 0 representing no deviation of the center of mass in the simulated precipitation field from the observed. Second part (L2) was calculated as averaged distance between the center of mass of the total rainfall and the individual rain objects as computed in step 2. L2 will become 0, if both the simulated and observed precipitation fields have single rain object. In other scenarios, the range of L2 will be between 0 and 1. Thus, the value of L ranges between 0 and 2 with 0 denoting no shift from the observed field, while higher values indicating a spatial shift from actual rainfall occurrences.

• (5) Structure: The structure component (S) compared the volume of rain objects, quantifying the size and shape of the objects. The values range between –2 and 2, with negative values signifying that the event is simulated as a small spread and peaked event. Positive values on the other hand represent a much broader or flat event simulated by the model when compared to the actual event. Values close to 0 represent closely simulated rain structures as compared to the observed.