By: Samantha Ferrett
Forecasting weather in Southeast Asia
Southeast (SE) Asia is prone to high‐impact weather and is often subject to flooding and landslides as a result of heavy rainfall. Just last month Indonesia was hit by heavy rainfall that resulted in floods and landslides because of a rainy season that lasted longer than was initially forecast. Global computer models used for Numerical Weather Prediction (NWP) have been known to fail to accurately capture Maritime Continent rainfall, limiting predictions of high‐impact weather in the region. I am a Research Scientist on a Weather and Climate Science for Service Partnership (WCSSP) Southeast Asia project, “FORecasting for SouthEast Asia” (FORSEA), that aims to improve forecasts in SE Asia to reduce social and economic losses from high impact weather events. In this blog, I will provide an overview of some of my recent work that examines how well newly developed ensemble forecasts reproduce extreme precipitation in SE Asia.
What is an ensemble forecast?
A deterministic forecast is a single forecast from a computer model using one initial condition and producing one final estimate of future weather. The initial condition is an estimation of the observed weather at the start of the forecast. There are multiple reasons for an incorrect forecast. For example, one cause is that the model may not be able to fully replicate processes that drive weather in the real world. This is why forecasts are just an estimate of future weather. Unfortunately, there is also some uncertainty even to the observed weather that can result in large errors in the final forecast, even with a ‘perfect’ forecast model. An ensemble forecast consists of multiple forecasts from the same model, each with slightly different initial conditions, representing the uncertainty in observations. This results in an ensemble of estimates of future weather that can then be used to gain an understanding of the uncertainty of the forecast.
Are convection-permitting ensemble forecasts worth it?
Figure 1: Schematic of rainfall in a coarser resolution forecast model (left) and rainfall in a high-resolution convection permitting model (right). Darker blues indicate more rainfall.
A forecast model divides the region to be forecast into a grid. Convection-Permitting (CP) forecasts are those that use NWP models with such small grid sizes that they can better represent processes associated with rainfall. A schematic showing the difference between a coarser grid and a high-resolution grid used in CP models is shown in Fig. 1. The downside is that CP models, and ensembles, are more computationally expensive. Modellers face a difficult task in striking a balance between cost and benefit; this is where those of us who analyse such models hope to be useful! It’s important for the modelling community to know if the resources invested in these forecasts are worth it.
In my work I examine how “skilful” forecasts of extreme rainfall are for CP ensembles of forecasts in Malaysia, Indonesia and the Philippines. These are ensembles of 17 forecasts at a resolution of 4.5km (like the schematic in Fig. 1 shows) and were run by the Met Office between October 2018 to March 2019. SE Asia has a strong daily cycle of precipitation where precipitation is over land during the day and moves over ocean during the night. A question to answer is if these normal daily variations of rainfall remove the need for CP forecasts – is rainfall so dominated by the daily cycle that there is no need for these high resolution forecasts?
Figure 2: Fractions Skill Score (FSS) of 3 hourly accumulated precipitation at 8pm-11pm local time (Malaysia) exceeding 95th percentile aggregated over all forecasts in Oct 2018-Mar 2019 as function of spatial scale (x-axis) a) Malaysia, b) Indonesia and c) Philippines. The horizontal line shows the FSS=0.5 “skilful” threshold. Lines show results from the ensemble forecast for 1, 3 and 5 days into the forecast (black, mid grey and light grey solid lines) and results from a forecast based on observed weather from 1, 3 and 5 days before the day to be forecast (black, mid grey and light grey dashed lines).
I compare the skill (using a metric called the Fractions Skill Score) of the ensemble forecasts, shown by the solid lines in Fig. 2, to a “persistence” forecast, shown by the dashed lines in Fig. 2. The persistence forecast does not use a model but instead uses observed weather from the days prior to the day being forecast to estimate the weather. The forecast is considered skilful at the spatial scale shown on the x-axis if the metric exceeds the threshold shown by the horizontal black line. The ensemble forecast is much more skilful. The larger skill at lower spatial scales means that smaller scale features can be more accurately forecast by the ensemble. Even skill five days into the ensemble forecast (shown by light grey solid line) is higher than that of the first day of the forecast based on observations (black dashed line). This means there is value in using such a forecast in all three regions.
It’s not over…
This is promising news for the use of CP models in the tropics, but questions still remain to be addressed in FORSEA:
- How do common large scale features known to modulate SE Asia rainfall, such as the Madden Julian Oscillation or equatorial waves, influence forecast skill?
- Shall we go smaller? This suite of forecasts also includes sub-kilometre scale forecasts. Is there benefit to using these?