1. Introduction
Methods for improving the prediction of rainfall is an ongoing area of research in the field of meteorology. Interest stems not only for academic reasons, but also due to the implication that accurate rainfall forecasts have the potential to save lives and avoid damage to infrastructure through better planning and awareness. For instance, [
1] estimated that floods kill 12,700 people annually and some hard-hit countries such as China have faced damages of 1–3% of Gross National Product due to flooding [
1]. In their latest report (AR6), the International Panel on Climate Change (IPCC) have confirmed that some areas of the globe have seen significant increases in precipitation since the beginning of the 20th century [
2] and that the world may see a 15% increase in the intensity of both 1-in-10-year and 1-in-50-year extreme precipitation events under a 2 °C warmer climate [
3]. Thus, the accurate prediction of such events requires the attention of meteorologists.
Methods for forecasting rainfall vary depending on the location and time horizon. For short time horizons (i.e., 0–6 h) and less chaotic locations, the methods of nowcasting have useful predictive power. Ref. [
4] reviewed the use of Convolutional Neural Networks (CNNs) for nowcasting from radar images. Ref. [
4] showed that for a heavy rain event in Braunsbach, Germany, the CNN is able to compare with more traditional techniques such as optical flow. Others have also tried blending radar and observations in a nowcasting method [
5], while Singapore (the focus area of this study) [
6] utilised GPS-estimates of precipitable water vapour to conduct very short (5 min ahead) prediction of rain events in 2014 and 2015.
However, for locations that experience rapidly evolving atmospheric processes, or for longer time horizons (i.e., beyond 4–6 h), physical models that can both spin-up and rain-out convective systems are necessary. Such models fall into the category of Numerical Weather Prediction (NWP) models. NWP models approximate the fluid and thermodynamic equations—often through discretisation of the governing equations and stepping forward in time through methods like Runge–Kutta. Other processes important for weather prediction, such as the transfer of radiation through the atmosphere and small-scale processes such as microphysics (phase changes within a cloud) require parameterisation in order for NWP models to be computationally efficient. How best to estimate the processes requiring parameterisation is an active area of research and models such as the Weather and Research Forecasting (WRF) model [
7] have a vast array of options for each process that is parameterised.
Ref. [
8] investigated over 100 possible choices of WRF model configuration, in addition to the performance of various spatial resolutions, for the forecast of precipitation over British Columbia, Canada. Over complex terrain, Ref. [
8] found the Kain–Fritsch to be generally better than the Grell–Frietas cumulus scheme and that there was clear seasonal dependence in the best performing scheme combinations. In addition, the authors found better representation of the distribution of precipitation values from the higher resolution domains (3 km versus 27 km) but that the coarser resolutions had higher equitable threat scores (overall performance). On the contrary, Ref. [
9] found greater skill in higher resolution, convection-permitting simulations over tropical east Africa. In Singapore, a collaboration between the United Kingdom’s Met Office and Meteorological Services Singapore, which gave rise to the SINGV model [
10], also found better skill at higher resolution. Analysis of the performance of SINGV in Ref. [
10] shows a clear improvement in rainfall forecast skill from the 1.5 km model when compared to the coarser 11 km grid.
The SINGV model represents a significant step forward for operational weather forecasting in Singapore, which is a challenging location to forecast. Located at roughly 1° N and 104° E, Singapore is an island state and maritime continent that experiences significant rainfall throughout the year. Following the shift in the meteorological equator, the prevailing winds deliver two monsoon seasons and two inter-monsoon/transition seasons. During the north-east monsoon (December–February), prevailing winds in addition to interactions with land–sea thermal contrasts can produce prolonged periods of rain. During the south-west monsoon (June–September), so-called Sumatran squalls are most common, which often bring heavy outbursts of rainfall in the early hours/pre-dawn. Inter-monsoon periods are characterised by generally light winds and diurnally forced afternoon storms. Ref. [
11] also investigated the use of WRF for rainfall modelling in Singapore. The authors investigated several physical scheme combinations on 14 rain events from 2011. Ref. [
11] found that even at 1 km resolution, the maximum intensity of rainfall from their output was much lower than the observations and that the cause could be due to low resolution geographical data leading to insufficient representation of land–sea processes.
In this paper, we explore the influence of spatial resolution on rainfall modelling over Singapore. As was mentioned previously, SINGV at 1.5 km showed improved forecast skill when compared to the control experiment at 11 km. Others have found similar results when it comes to increased skill in rainfall modelling with increased spatial resolution; the general idea being that higher spatial resolutions begin to resolve convection processes explicitly without the need for parameterisation. Refs. [
9,
12] both note the importance of running NWP models at the convection permitting scale. Meanwhile, Ref. [
13] compared 1, 4, 8 and 16 km resolutions for flood forecasting in north-eastern Italy and found generally larger errors from the 16 km domain compared to the 1 km resolution. Similar results were also found by Ref. [
14] in Peru showing that 0.75 km resolution WRF grid was able to better produce local processes and that the 18 km domain had the lowest skill.
We therefore wish to test the hypothesis that higher resolution WRF grids will also achieve greater skill when forecasting rainfall over Singapore. The rest of this paper is structured as follows:
Section 2 describes the methods and observations,
Section 3 presents the results for both the 15-day testing set and the heavy rain event, while
Section 4 concludes the study and suggests future work arising from the results.
4. Summary and Conclusions
In this study, we have compared the rainfall performance of various WRF model domain resolutions (1 km, 3 km, 9 km and 12 km) over Singapore. We selected a representative set of 15 days with a mixture of seasons, types of rain events and rain amounts. On the testing set, we showed that the highest resolution (1 km) had the best performance (highest CSI) for rain rates above 0.5 mm/h and when compared to ground observations. This result was then confirmed when comparison was made against radar-derived rain rates. We showed the coarser domains had better overall performance at the lower rain rates when ground observations were used as the truth, but when analysing certain days in more detail, it was clear that the coarser domains were unable to correctly capture any of the heavier rain rates (particularly the 9 km and 12 km domains above 2 mm/h). Additionally, we calculated the Fractions Skill Score (FSS) for three highlighted days over the testing set and showed that, in general, the 1 km domain exhibited highest skill for the heavier rain events. We then analysed a particularly heavy rain day not from the original 15-day set. The 10 January 2021 heavy rain day saw in excess of 100 mm of rain on average across the island of Singapore. On this day, the 3 km domain had highest CSI for the heavier rain rates at 4 mm/h and 16 mm/h—contrary to the 15-day set. However, the FSS showed highest skill from the 1 km domain, which was concordant with a visual inspection of the event. In general, we found the 1 km domain to be able to capture the most realistic-looking features and heaviest rain amounts. This is reflected in the FSS values, but the timing and location can be incorrect leading to lower hits, higher false alarms and thus lower CSI. We also found that all domains tended to experience the same level of mismatch in the onset of rainfall indicating the cause is likely downstream of the global model input (ECMWF). In order to rectify this issue, special attention is needed in either the assimilation of local data or the inclusion of other global model inputs in an ensemble.