There has been increasing demand in Australia for extended-range forecasts of extreme heat events. An assessment is made of the subseasonal experimental guidance provided by the Bureau of Meteorology’s seasonal prediction system, Predictive Ocean Atmosphere Model for Australia (POAMA, version 2), for the three most extreme heat events over Australia in 2013, which occurred in January, March, and September. The impacts of these events included devastating bushfires and damage to crops. The outlooks performed well for January and September, with forecasts indicating increased odds of top-decile maximum temperature over most affected areas at least one week in advance for the fortnightly averaged periods at the start of the heat waves and for forecasts of the months of January and September. The March event was more localized, affecting southern Australia. Although the anomalously high sea surface temperature around southern Australia in March (a potential source of predictability) was correctly forecast, the forecast of high temperatures over the mainland was restricted to the coastline. September was associated with strong forcing from some large-scale atmospheric climate drivers known to increase the chance of having more extreme temperatures over parts of Australia. POAMA-2 was able to forecast the sense of these drivers at least one week in advance, but their magnitude was weaker than observed. The reasonably good temperature forecasts for September are likely due to the model being able to forecast the important climate drivers and their teleconnection to Australian climate. This study adds to the growing evidence that there is significant potential to extend and augment traditional weather forecast guidance for extreme events to include longer-lead probabilistic information.
Heat waves are a regular feature of the Australian climate, with often severe impacts on a wide range of sectors, including health, agriculture, utilities, infrastructure, tourism, and commerce (Nairn and Fawcett 2013). With a warming climate, the frequency and intensity of these heat waves is increasing (Plummer et al. 1999; Collins et al. 2000; Griffiths et al. 2005; Alexander et al. 2007), and this trend is set to continue throughout the current century (IPCC 2013). The benefits of producing enhanced and actionable weather and climate forecasts of extreme heat events are numerous and of substantial value for mitigating impacts.
“At risk” sectors make considerable use of weather forecasts (days 1–7) and warnings (a day or two ahead) for extreme heat events. Example uses of forecasts include heat health warnings, provision of specialized care for vulnerable sectors of society (e.g., the elderly, young, and ill), fire bans, and proactive irrigation and shading for crops. There is currently a significant gap in the forecast information on time scales from a week to a season (i.e., intraseasonal/subseasonal/multiweek) around extreme heat events. To the extent that these forecasts are skillful, they can provide significant value in addition to the more traditional warnings, weather forecasts, and seasonal outlooks, which constitute the current operational service of the Australian Bureau of Meteorology (BoM) for extreme weather and climate events.
The BoM is currently engaged in developing the science and systems necessary to form a subseasonal forecast service for Australia. This has been driven by strong stakeholder interest, particularly from the agricultural (http://www.managingclimate.gov.au/publications/climag-18/ and http://www.grdc.com.au/Media-Centre/Ground-Cover-Supplements/Ground-Cover-Issue-87-Climate-Supplement) and emergency management sectors. This capability is based on the Predictive Ocean Atmosphere Model for Australia (POAMA, version 2), the BoM’s fully coupled ocean–atmosphere seasonal prediction system (Hudson et al. 2013). This dynamical model has shown particularly good skill in the prediction of maximum temperatures on subseasonal and seasonal time scales and has been found to produce forecasts that are statistically reliable (Hudson et al. 2013).
Recent work has explored the ability of POAMA-2 to forecast heat extremes on subseasonal time scales. The skill of POAMA-2 across a large number of events represented by a 30-yr hindcast set is well documented (Hudson et al. 2011a, 2013, 2015; Marshall et al. 2013b; White et al. 2013). However, there has been rather limited analysis of the forecast performance for individual extreme events. Cases studies are important because they provide the opportunity to examine not just the skill of the forecast but also to facilitate a more in-depth understanding of the drivers and reasons behind the performance of a particular forecast. Cases studies also help to promote the potential utility of the forecasts to the user and application communities, particularly when focusing on events that are of high societal impact. Here, we assess the experimental subseasonal guidance provided by POAMA-2 for the most extreme heat wave events over Australia in 2013, as identified in the BoM’s special climate statements (http://www.bom.gov.au/climate/current/special-statements.shtml), which provide a summary of significant weather events. Three heat events were reported on for 2013 and as such represent the most significant heat events of the year for Australia in terms of the extremity of the observations and their impacts. The events are
January 2013 (Australia’s hottest January and calendar month on record to date),
March 2013 (a major long-lived heat wave in southeastern Australia), and
September 2013 (Australia’s hottest September on record to date).
The impacts of these heat waves were wide ranging and included devastating bushfires (in January and September), damage to crops and accelerated plant growth [e.g., in September; Collis (2013)], and impacts on animals and natural ecosystems (e.g., more than 10 000 flying fox deaths in January; information online at http://theconversation.com/killer-climate-tens-of-thousands-of-flying-foxes-dead-in-a-day-23227). There were likely other impacts, such as on human health and on infrastructure and utilities (e.g., increased electricity demand). In some respects, 2013 represents a challenging year for subseasonal and seasonal forecasts, since the two key drivers of Australian climate variability, El Niño–Southern Oscillation (ENSO) and the Indian Ocean dipole (IOD), were both in neutral states. In fact, ENSO was on the La Niña side of neutral during most of 2013, and La Niña’s are usually associated with cooler than normal temperatures over Australia (White et al. 2013). Although both ENSO and the IOD are associated with low-frequency seasonal variations, they have been shown to be a source of subseasonal climate predictability (Hudson et al. 2011a; White et al. 2013).
The paper is structured as follows. Section 2 provides a brief description of the POAMA-2 forecast system, along with the forecast and observational datasets. Sections 3–5 examine the three heat events, respectively, in terms of the ability of POAMA-2 to provide advance warning and the potential role of the large-scale circulation. A summary and conclusions are presented in section 6. We emphasize that looking at only a limited number of forecast cases is not a complete assessment of the skill of a forecast system. It is important to demonstrate that POAMA-2 forecasts are skillful and reliable over a large number and wide range of cases. Although we do not show overall skill in this paper, the skill for the upper-tercile, top-quintile, and top-decile maximum temperature subseasonal forecasts from the full hindcast set (covering the 30-yr period 1981–2010) have been published elsewhere (Hudson et al. 2013; Marshall et al. 2013b; White et al. 2013; Hudson et al. 2015).
2. Forecast system and methodology
a. POAMA-2 forecast system
We assess forecasts from the current operational version of POAMA (version 2; hereafter, POAMA), which was producing real-time forecasts during 2013. This system produces the BoM’s operational seasonal forecasts (available online at http://www.bom.gov.au/climate/ahead/), but remains experimental for forecasts on the subseasonal time scale. Full details of the model, data assimilation, and ensemble generation are provided in Hudson et al. (2013; the system is referred to in their paper as P2–M).
In brief, POAMA is a fully coupled ocean–atmosphere model and data assimilation system that produces an ensemble of future climate states based on perturbed initial conditions. The atmospheric model is the Bureau of Meteorology Research Centre (BMRC) Atmospheric Model, version 3 (BAM3; Colman et al. 2005; Wang et al. 2005; Zhong et al. 2006), which is a spectral transform model with a horizontal resolution of T47 (approximately 250 km) and 17 vertical levels. This horizontal resolution, together with the grid configuration, means that Tasmania is not resolved as land. The land surface component of BAM3 is a simple bucket model for soil moisture (Manabe and Holloway 1975) and three active soil layers for temperature. The ocean model is the Australian Community Ocean Model 2 (ACOM2; Schiller et al. 1997, 2002), which is based on the Geophysical Fluid Dynamics Laboratory Modular Ocean Model (MOM), version 2. The ocean model grid has a zonal resolution of 2° and a meridional resolution of 0.5° equatorward of 8° latitude, gradually increasing to 1.5° near the poles. ACOM2 has 25 vertical levels, with 12 in the top 185 m and a maximum depth of 5 km. The atmosphere and ocean models are coupled using the Ocean Atmosphere Sea Ice Soil (OASIS 2.4) coupling software (Valcke et al. 2000).
Forecasts are initialized from observed oceanic and atmospheric states. Ocean initial conditions are provided by the POAMA Ensemble Ocean Data Assimilation System (PEODAS; Yin et al. 2011) and the atmosphere and land initial conditions by the Atmosphere–Land Initialization (ALI) scheme (Hudson et al. 2011b). Perturbations to the atmosphere and ocean initial conditions are produced by a coupled-model breeding technique (Hudson et al. 2013). For each forecast, a 33-member ensemble is generated in “burst” mode (i.e., all initial conditions are valid for the same date and time).
b. Forecasts and verification
The real-time forecasts were initialized once per week (0000 UTC on Thursdays) up to the end of January 2013 and twice per week thereafter (0000 UTC on Mondays and Thursdays). The exact forecast periods and lead times that are examined in each case study are restricted by this forecast frequency and the overlap with the heat wave period of interest. We assess forecasts of maximum temperature (Tmax) anomalies over Australia both in terms of the ensemble mean anomaly forecast and the ensemble probability that a given fortnight or month will be in the top quintile or top decile. The probabilistic forecasts are obtained by calculating the percentage of ensemble members forecasting above the respective quintile or decile threshold. We acknowledge that the definition of a heat wave and its resulting impacts are characterized by more than just maximum temperature, in particular the roles of minimum temperature, humidity, and wind may be important (e.g., Steadman 1984; Nairn and Fawcett 2013; Perkins and Alexander 2013). The forecast Tmax anomalies are formed relative to the ensemble mean climatology from the hindcasts. The hindcasts comprise forecasts starting on the 1st, 11th, and 21st of each month for the period 1981–2010. This climatology (as well as the quintile and decile thresholds, which are defined by ranking the data) is a function of both start date, lead time, and grid box, and thus a correction for model mean bias is made to the real-time forecasts (e.g., Stockdale 1997). The hindcast start date closest to the real-time forecast date is used to bias correct the forecasts (note that the hindcast set has since been expanded to having six initializations per month, i.e., a hindcast every 5 days, but in 2013 it was still based on three initializations per month).
We compare the model’s forecast of Tmax to observations using the Australian Water Availability Project (AWAP) observed gridded dataset (Jones et al. 2009), regridded to POAMA’s T47 grid (approximately 250 km). The climatology period for calculating the anomalies and thresholds for the observations is the same as that for the model (1981–2010).
Quantification of the performance of the cases is provided in Table 1 in terms of spatial correlation r, hit rates (HRs), and false alarm ratios (FARs) calculated for mainland Australia (i.e., for all grid boxes defined as land on the POAMA grid). Spatial correlation measures the correlation between the respective observed Tmax anomaly (shown in column 1 in Figs. 1–3) and the ensemble mean forecast (column 3 in Figs. 1–3). The probabilistic forecasts (columns 4 and 5 in Figs. 1–3) are assessed against the observed outcome (column 2 in Figs. 1–3, i.e., the observed occurrence of either a top-quintile or top-decile event) using hit rates and false alarm ratios. This is done by converting the probabilistic forecast to a dichotomous (yes/no) forecast (i.e., “yes, the event will happen” or “no, the event will not happen”) using a probability threshold, above which the forecast will be yes and below which it will be no. We have chosen to assess two thresholds for each product. First, as is often done (Wilks 2006), we have chosen the climatological frequency of the event (i.e., it is a yes forecast if the probabilistic forecast is greater than 20% for the quintile product and greater than 10% for the decile product). Second, we have used slightly higher thresholds that directly match the depiction of the start of the color shading of “increased chance” in the figures, that is, greater than 30% (15%) for the quintile (decile) product is a forecast yes event. Note that to some extent this conversion of the probabilistic forecast to a nonprobabilistic forecast is arbitrary since in reality the threshold that is chosen depends on the user of the forecast and the decision being made (Wilks 2006). In addition, a lot of potentially useful forecast information is lost when doing the conversion. Nonetheless, the hit rates and false alarm ratios provide a simple quantification of the maps and aid comparisons. The HR (also referred to as the probability of detection) is defined as the percentage of observed yes events that were correctly forecast (Wilks 2006). The FAR is the percentage of forecast yes events that did not occur in reality (i.e., were false alarms) (Wilks 2006). The FARs and HRs shown in Table 1 are calculated by summing spatially over all land grid boxes in mainland Australia for each event (i.e., the total number of events contributing to the contingency table is equal to the total number of land grid boxes).
3. January 2013
a. Temperature forecasts
January 2013 was Australia’s hottest January and month on record and was also associated with major bushfires in southeastern Australia and Tasmania (Bureau of Meteorology 2013c). The heat wave in late December and the first weeks of January was notable in terms of the spatial extent of the high temperatures, and a new record for the hottest day for Australia as a whole was set on 7 January (Bureau of Meteorology 2013c). The area-averaged maximum temperature for Australia exceeded 39°C on seven consecutive days from 2 to 8 January (the previous record was four days in December 1972). The heat wave started over far southwestern Western Australia from 25 to 30 December and culminated in a final phase of prolonged extreme heat affecting South Australia, Victoria, and southern New South Wales on 17–18 January (Bureau of Meteorology 2013c).
The observed Tmax anomalies for the fortnight of 3–16 January were in the top decile across most of Australia (i.e., above the 90th percentile threshold for 3–16 January anomalies calculated using the 1981–2010 base period), except for far Western Australia (Figs. 1a,b). One week in advance of this fortnight and prior to the heat wave over central and eastern Australia (forecasts initialized on 27 December), POAMA was clearly able to warn of the upcoming extreme heat over mainland Australia (i.e., weeks 2 and 3 of the forecast; Figs. 1c–e). A noticeable poor forecast was made for Tasmania, where POAMA indicated a reduced chance of extreme conditions. The peak of the January heat wave over Tasmania fell within this fortnight, on 4 January (Bureau of Meteorology 2013c). As mentioned in section 2, Tasmania is not resolved as land in the model and the forecast is therefore reflecting ocean surface temperatures from the model. The observed sea surface temperature (SST) anomalies for that fortnight [calculated from the PEODAS ocean reanalysis; Yin et al. (2011)] were mostly between −0.5° and 0.5°C around Tasmania (not shown). POAMA did, however, show strongly increased odds of top-quintile and top-decile Tmax anomalies over virtually all of Australia, as eventuated (Figs. 1d,e). The probabilistic forecasts indicated greater than 50% and 30% chances of exceeding the top-quintile and top-decile thresholds, respectively, over large portions of Australia (Figs. 1d,e), that is, indicating from a doubling to tripling of the chance of hot conditions compared to climatology (20% and 10%, respectively). Increased odds (compared to climatology) of top-quintile (decile) conditions were forecast over 96.0% (92.2%) of the area where top-quintile (decile) heat eventuated and the false alarm ratios were low (FAR = 11% for the quintile product and 19.4% for the decile product) (Table 1).
Forecasts for this fortnight two weeks in advance (i.e., initialized on 20 December, prior to the heat wave) also showed strongly increased odds of top-quintile and top-decile conditions over most of Australia outside of southwestern Western Australia (i.e., weeks 3 and 4 of the forecast; Figs. 1g,h), with top-quintile (decile) conditions forecast over 93.1% (92.2%) of the area where top-quintile (decile) heat eventuated (with false alarm ratios below 13% for both the quintile and decile products) (Table 1). However, the spatial pattern of the forecast ensemble mean Tmax anomalies was poor for both the forecast initialized on 27 December (r = 0.09; Table 1, Fig. 1c) and the forecast initialized on 20 December (r = 0.33; Table 1, Fig. 1f), such that the magnitudes of the anomalies were underestimated over central Australia and parts of southern Australia, and were overestimated over northwestern Australia.
The model also forecast in advance that the month of January (Figs. 1i–m) was very likely to be extremely hot, particularly over the eastern half of the country, with high forecast probabilities for Tmax anomalies falling in the top quintile (decile) over large regions (Figs. 1l,m). This indication of an increased chance of extreme heat compared to climatology verifies well with the observed outcome, although the false alarm ratio for the top-decile forecast is relatively high (Fig. 1 and Table 1: HR = 98.8% and FAR = 17.5% for the quintile product; HR = 100% and FAR = 36.4% for the decile product). Probabilistic forecasts for the month of January produced a month in advance (i.e., forecasts initialized on 29 November) had a very similar pattern of increased chances of extreme heat as the shorter lead time forecasts in Figs. 1l and 1m, but were less emphatic (maps not shown), with top-quintile (decile) conditions forecast over 94.2% (85.7%) of the area where top-quintile (decile) heat eventuated (with a false alarm ratio of 14.7% for the quintile product and 33.3% for the decile product).
b. Large-scale circulation and drivers
The synoptic conditions associated with the periods of heat for different regions of Australia during January 2013 are described by Bureau of Meteorology (2013c). The end of December 2012 was characterized by a high pressure in the Great Australian Bight (i.e., the large open oceanic bay off the southern coastline of mainland Australia), which brought heat to southwestern Australia and then began to move eastward at the end of the month, bringing hot conditions progressively farther east over Australia (Bureau of Meteorology 2013c). By 4 January, the high pressure had moved into the Tasman Sea and was directing hot air over southeastern Australia (while Western Australia cooled somewhat). A second high pressure moved into the bight around 8 January, bringing more heat to Western Australia and then to the interior of Australia as it moved eastward on 11 January (Bureau of Meteorology 2013c). These transient pressure variations are visible in the time–longitude plot of 850-hPa geopotential height anomalies at 45°S for the period from 28 December to 19 January from ERA-Interim (Dee et al. 2011) (Fig. 4). The POAMA ensemble mean forecast, initialized on 27 December, was able to correctly persist the high pressure anomaly near the bight, which was present in the initial conditions, and POAMA’s sequence of geopotential height anomalies is similar to what was observed for about the first 10 days of the forecast (Fig. 4). Beyond that, memory of the initial conditions diminishes and the forecast ensemble diverges, resulting in small ensemble mean anomalies. Interestingly, there is an indication that the POAMA ensemble does capture the influence of a second high pressure system moving eastward from about 11 January, suggesting that some ensemble members correctly forecast the timing and position of the transitions (Fig. 4). After the first week of January, the POAMA ensemble mean forecast shows a persistent high pressure anomaly east of Australia over the Tasman Sea and New Zealand (~170°E), not seen in the observations (Fig. 4). This could account for the differences in the spatial pattern of the Tmax anomalies shown in Figs. 1a and 1c for 3–16 January, since a high over the Tasman Sea would focus the heat more on coastal eastern Australia, as seen in POAMA.
In the mean, the heat in January was associated with above average 500-hPa heights over southern Australia, associated with a poleward shift of the upper-level westerly winds [Fig. 5, and reported and shown in more detail in Climate Prediction Center (2013)]. The POAMA ensemble mean forecast for January, initialized on 27 December, was able to capture the higher than normal heights over Australia and the southward shift of the upper-level winds, particularly over eastern Australia and the Tasman Sea (Fig. 5).
It is not clear what process or driver is providing the predictability for these forecasts. In an attribution study using a suite of coupled-model experiments with natural forcings compared to natural and anthropogenic forcings, Lewis and Karoly (2013) concluded that anthropogenic climate change greatly contributed to the extreme heat of the summer of 2013 (from December 2012 to February 2013) and that natural climate variability alone, including ENSO (which was in a neutral-cool state at that time), was unlikely to explain the record temperatures over Australia. POAMA is able to include the contribution of the long-term warming trend through its use of observed initial conditions in the forecasts, from both the oceans and the land surface, and this may be providing some predictability in this case.
4. March 2013
a. Temperature forecasts
The heat wave during the first two weeks of March 2013 was a fairly localized event, affecting mainly southern Victoria and Tasmania, but broke numerous records (Bureau of Meteorology 2013a). Parts of central northern Australia also recorded higher than average maximum temperature anomalies (in the top decile over parts), while at the same time southwestern and eastern Australia experienced colder than normal temperatures (Figs. 2a,b). Impressively, POAMA’s forecasts for the fortnight of 4–17 March and initialized on 4 March (i.e., weeks 1 and 2 of the forecast; Fig. 2) generally captured this overall pattern of anomalies over Australia (r = 0.52; Table 1). However, the ensemble mean forecast overestimated the hot conditions over the north and the cooler conditions over the west and east and did not capture the magnitude of the warm temperature anomaly over Victoria and Tasmania (Fig. 2c). The probabilistic forecasts do indicate very strongly increased odds for having a top-quintile and top-decile event over northern Australia and Tasmania (Figs. 2d,e), as eventuated (Fig. 2b), but over Victoria the increased odds are very much restricted to the coastline and adjacent ocean. The forecast for this fortnight initialized a week earlier (25 February) did not capture the pattern of temperature anomalies over Australia, instead predicting cooler than normal conditions over most of Australia and warmer than normal conditions over southwestern Australia and the Cape York Peninsula (to the far northeast) (i.e., weeks 2 and 3 of the forecast; Figs. 2f–h, Table 1). Over the southeast, warmer than normal conditions were largely restricted to the surrounding ocean (Figs. 2f–h).
b. Large-scale circulation and drivers
The heat wave was associated with a high pressure ridge in the midlatitudes, generally weak winds over far southern and southeastern Australia, and exceptionally high SST anomalies off the southern coast of Australia (Bureau of Meteorology 2013a). POAMA forecasts for 4–17 March (initialized on 4 March) were able to capture the high pressure anomaly over the Tasman Sea and south of Australia as well as the abnormally high SST anomalies around southern Australia (Fig. 6). The latter explains the good forecast over Tasmania. The forecast for this fortnight initialized a week earlier (25 February) did not capture the high pressure anomalies mentioned above, but did still represent the positive SST anomalies (not shown). The warm SSTs offshore are thought to have contributed to the prolonged warm conditions over southern Australia through reducing the effect of cooling sea breezes (Bureau of Meteorology 2013a). It is unclear whether POAMA can adequately capture the nature and extent of the sea-breeze circulations, particularly localized effects, and this may have played a role in the underestimation of the magnitude and spatial extent of the heat over southern mainland Australia by the model.
5. September 2013
a. Temperature forecasts
September 2013 was Australia’s hottest September on record (Bureau of Meteorology 2013b). The most extreme heat occurred between the last week of August and the first two weeks of September (Figs. 3a,b; Bureau of Meteorology 2013b). The largest Tmax anomalies occurred over the central and eastern interior (Fig. 3). Numerous maximum temperature records were broken (including stations in all states and territories) and warmer than normal temperatures were also unusually prolonged in many regions (Bureau of Meteorology 2013b). The occurrence of these high temperatures during spring, while not extreme by summer standards, had various impacts on agriculture, such as accelerated plant growth and increased water stress (e.g., Collis 2013), and were also associated with bushfires in eastern New South Wales (Bureau of Meteorology 2013b). In the southeast, they resulted in a premature end to the snow-related winter sports season.
The POAMA forecasts shown in Figs. 3c–e are for the fortnight from 26 August to 8 September (initialized at the start of the week prior, i.e., weeks 2 and 3 of the forecast). POAMA provided good warning (except perhaps over the far north) a week in advance of the warmer than normal temperatures experienced in this fortnight. Although the ensemble mean anomaly underestimated the observed anomaly, the location of the above normal temperatures is reasonable (r = 0.84; Table 1). The probabilistic forecasts show strongly increased chances of top-quintile and top-decile Tmax over most of central, southern, and eastern Australia (Figs. 3d,e), with good hit rates (>67%; Table 1) and low false alarm ratios (<18%; Table 1). The forecasts for the same fortnight two weeks in advance (i.e., initialized on 12 August) were not as good (i.e., weeks 3 and 4 of the forecast; Figs. 3f–h, Table 1), but did provide indications of the upcoming extreme heat over parts of eastern Australia.
POAMA’s forecast for the month of September, initialized in August (i.e., prior to the main event), showed very strong indications of upcoming heat over the whole of Australia (Figs. 3l,m), although this forecast verifies poorly for far Western Australia (Figs. 3i,j). Hit rates are high (>86%; Table 1) and false alarm ratios are relatively low (<22%). However, forecasts for the month of September issued a month or more in advance did not provide good warning of the upcoming heat in September (not shown).
b. Large-scale circulation and drivers
The 850-hPa circulation for the fortnight from 26 August to 8 September (corresponding to the temperature forecasts in Figs. 3a–e) featured a high pressure anomaly over the Tasman Sea and a low pressure anomaly southwest of Australia, both acting to bring anomalous southerly flow across the southern half of Australia (Fig. 7). POAMA’s ensemble mean forecast initialized a week prior (on 19 August, i.e., weeks 2 and 3 of the forecast) was able to capture these anomalies, although weaker than observed (Fig. 7).
Unlike for January and March, the heat in September was associated with some strong atmospheric climate drivers. To try and understand the drivers behind the September 2013 heat event, Arblaster et al. (2014) used both a multiple linear regression model built from observed predictors, and POAMA sensitivity forecast experiments. The predictors they examined in the regression model included monthly anomalies of ENSO, the IOD, the southern annular mode (SAM), global mean temperatures (as an indicator of large-scale warming), and August soil moisture. They concluded that the record Tmax over Australia resulted from a combination of background warming, abnormally dry soils, and a strongly anomalous atmospheric circulation, which included a strong negative SAM and a low pressure anomaly southwest of Australia. Their reconstruction of the event did not capture the full magnitude of the observed Tmax anomalies, and they suggested that this may be partly due to missing key predictors, specifically the Madden–Julian oscillation (MJO; Madden and Julian 1971).
Building on the work of Arblaster et al. (2014), we have examined the large-scale drivers using daily data instead of monthly data to show the evolution of the drivers over the heat event. In addition, in our analysis we have also included the MJO, as well as blocking and persistent subtropical ridge highs over the Tasman Sea (STRH events), all of which have been shown to influence the occurrence of heat events over Australia (Marshall et al. 2013b). We do not include ENSO and the IOD in the analysis since they did not play a role in the event [being both neutral; Arblaster et al. (2014)].
Daily mean indices are calculated for the SAM, STRH, and blocking over the period of the event and are normalized by their respective standard deviations (using the period 1981–2010). This has been done for the forecasts and observations and the results are plotted in Fig. 8 (and described later). Full details of the definitions of these indices are provided in Marshall et al. (2013b). The SAM index is constructed using the eigenvalues of the leading empirical orthogonal function (EOF) of daily mean zonal-mean mean sea level pressure (MSLP) data between 25° and 75°S. During a positive (negative) SAM index the pressure anomaly is low (high) near the pole and the midlatitude westerly jet is shifted poleward (equatorward). Persistent high pressures over the Tasman Sea can be described in terms of 1) split-flow atmospheric blocking (referred to here as blocking events) and 2) high pressure systems located in the vicinity of the subtropical ridge about 10° north of the split-flow blocking region (referred to here as STRH events). These two types of events have different impacts on the occurrence of extreme heat over Australia (Marshall et al. 2013b). The blocking index used is the Bureau of Meteorology’s definition, which describes the degree of splitting of the 500-hPa westerly airstream (Pook and Gibson 1999). The positive polarity of the blocking index represents a physical situation of reduced midlatitude westerly flow experienced during episodes of strong blocking (conversely, a negative blocking index indicates enhanced zonal westerly flow in the Australian region). The STRH index is based on daily mean MSLP data averaged over a region of the Tasman Sea and is calculated as an anomaly relative to its annual climatology. A positive STRH index is used for identifying persistent anticyclones in the vicinity of the subtropical ridge. The MJO is depicted using the Real-Time Multivariate MJO (RMM) index (Wheeler and Hendon 2004).
We also examine the general influence that the climate drivers have on the chance of top-decile extreme heat occurring during August and September, calculated from a 30-yr period. These results are shown in Fig. 9 (and described later). This is done by computing composites of the probability of occurrence of top-decile Tmax events associated with each driver in August and September, calculated from the AWAP observations (Jones et al. 2009) over the period 1981–2010. Composites are calculated for each phase of each driver and are based on weekly mean data. Refer to Marshall et al. (2013b) for details of the formulation of these composites. Essentially, each composite is constructed by counting the number of weeks at each grid location for which the weekly averaged Tmax anomaly is greater than the 90th percentile, and then dividing by the total number of weeks in that composite to form a probability. The composites are displayed as a ratio to the mean probability (i.e., 10%) such that values greater (less) than one indicate an increased (reduced) chance of extreme heat. We do not show the corresponding composites from POAMA, although they are shown and discussed for the traditional four seasons in White et al. (2013) and Marshall et al. (2013b). In most cases POAMA can capture the sense of the teleconnection between the climate driver and extreme heat, but tends to underestimate the magnitude of the teleconnection.
Figure 8a shows the observed evolution of the drivers during and preceding the heat event. The indices are calculated from the NCEP–NCAR reanalyses data (Kalnay et al. 1996), and a 3-day running mean is used to smooth the index data. The shaded areas of the plots indicate the primary period of the heat wave and shading is indicated for values of the normalized indices that exceed plus or minus one standard deviation. The daily Tmax anomaly [from AWAP; Jones et al. (2009)] area averaged over the heat wave region (demarcated by the box in Fig. 3i; land only) is also plotted (smoothed with a 3-day running mean). The amplitude of the RMM index is shown as a number, which represents the phase of the MJO, and only MJOs that are strong (amplitude exceeds one standard deviation) are plotted. It is clear that the heat wave was associated with a strong negative SAM, with excursions near or exceeding two standard deviations around the times of the peaks in Tmax (Fig. 8a). The SAM was in a negative phase during most of August and September. At this time of year a negative SAM is associated with increased odds of extreme heat over most of Australia (Fig. 9d).
Apart from the SAM, the first peak in Tmax, around the start of September, was also associated with a strong STRH index, negative blocking, and an MJO moving through phases 8, 1, and 2 (Fig. 8a). These phases of the drivers are typically associated with extreme heat occurring over various regions of Australia (Fig. 9). A strong STRH and strong zonal westerly flow (essentially no blocking) tend to increase the chances of having extreme heat over the southern half of the country in August–September (up to about 2.5 times more likely in places; Figs. 9c,e), whereas the impact of the MJO in phase 1 is much weaker and increased odds of extreme heat are more localized over eastern Australia (generally only 1.1–1.5 times more likely) (Fig. 9a). In the middle of September, just prior to a second peak in Tmax, the MJO was in phase 5 (Fig. 8a), which is associated with areas over northern and central Australia of small increased odds of having extreme heat (Fig. 9b), and then it moved into phase 6, which generally reduces the odds of having extreme heat over the region (composite not shown) and may have contributed to the demise of the heat event. Although the average increased odds (calculated over 30 yr of data) of having extreme heat are fairly small for the MJO during phases 1 and 5 (i.e., the phases coinciding with the peaks in Tmax in this event; Fig. 9), this does not negate the effect the MJO could have had in a specific case, such as this. The exact role of the MJO in this particular heat wave requires more detailed investigation, beyond the scope of this study.
The observed indices are compared to the ensemble mean indices predicted by POAMA for 1-week lead-time forecasts. For example, in Fig. 8b the daily means plotted against 19 September are from the forecast initialized on 12 September. The points are spaced every 3 or 4 days apart, corresponding to the twice-weekly real-time forecasts. For the RMM index, the forecast is interpolated onto a daily time grid so that a missing MJO index value for a particular day specifies a weak MJO forecast (and not the absence of forecast data). The magnitude of POAMA’s predicted ensemble mean indices for the SAM, STRH, MJO, and blocking are much smaller than observed (Fig. 8b). This is partially indicative of the spread of the ensemble of forecasts; the ensemble mean forecast hides the often large range of forecasts from individual members, especially for the atmospheric variables. In contrast, the observations are essentially one realization. POAMA’s ensemble mean forecast was indeed for negative SAM conditions over the whole period shown in Fig. 8b. In the forecasts, the first peak in Tmax near the start of September is also associated with an MJO in phase 1 (or 2), a strong STRH index, and no blocking. As in the observations, in the POAMA forecasts the slight cooling of the Tmax anomalies in the middle of September is associated with a weakening of the STRH index and enhanced blocking conditions in the Australian region (Fig. 8), reducing the odds of extreme heat occurring over the heat wave region (not shown). However, this would be offset by the SAM index still being in a negative phase (Fig. 8). In the POAMA forecasts the second peak in Tmax at the end of September is associated with an MJO in phase 5 and a negative SAM exceeding one standard deviation, as was observed, but the exact timing of these events with respect to the daily Tmax anomaly differs in the observations and POAMA.
Another key conclusion in Arblaster et al. (2014) was that anthropogenic climate change played an important role in the record-breaking heat of September. The background warming trend was one of the dominant contributors to the heat experienced over Australia. This trend is expressed in the seasonal forecasts through the trend in the ocean and land initial conditions and their POAMA seasonal sensitivity experiments indicated that it was the land temperature initial conditions that were the dominant contributor to the temperature anomaly over Australia (Arblaster et al. 2014).
6. Summary and conclusions
This paper examines how well POAMA would have provided advanced warning of the three major heat events experienced in 2013 in Australia. We could consider 2013 to be a challenging year for climate predictions, since both ENSO and the IOD were in neutral states, and there is some evidence that prediction skill for Australian climate is greater during the extremes of these drivers (Hudson et al. 2011a; White et al. 2013). However, in general, the POAMA guidance for two of the three heat waves was reasonably good, that is, the larger spatial scale and longer-lived events in January and September. In contrast, the March event was fairly localized, focused mainly over Victoria and Tasmania. POAMA’s performance for the upcoming fortnight of 4–17 March 2013 (at zero lead) was notable for correctly capturing the pattern of temperature anomalies over Australia, although the forecast of the heat event over Victoria was more spatially restricted to the coastline and surrounding oceans (did not penetrate inland). A finer-resolution model may be needed to translate the predictability provided by the SSTs into predictability over the adjacent land, particularly since the warm SSTs offshore were thought to have contributed to the prolonged warm conditions by reducing the effect of cooling sea breezes (Bureau of Meteorology 2013a).
The January event was not associated with strong activity of the large-scale drivers that are known to influence Australia’s natural climate variability. Extreme heat in the month of January was associated with above average 500-hPa heights over Australia and a poleward shift of the upper-level westerly winds, which POAMA was able to capture. The specific episodes of heat over various regions of Australia during the month were associated with transient high pressures south of Australia, in the bight, and over the Tasman Sea. The POAMA forecast from 27 December showed that the model was able to correctly persist the high pressure anomaly near the bight, which was present in the initial conditions, and that the forecast sequence of pressure anomalies was similar to observed for about the first 10 days of the forecast, but beyond that memory of the initial conditions diminishes and the forecast ensemble diverges. Differences in the positioning of the forecast high pressure anomalies beyond day 10 compared to the observed, led to differences in the spatial pattern of the forecast temperature response over Australia. However, the probabilistic forecasts did provide good advance warning of the upcoming heat, showing increased odds of top-quintile and top-decile conditions over the affected areas (high hit rates) for the key fortnight of the heat wave both one and two weeks in advance. The forecasts also effectively warned that the month of January would be associated with extreme heat.
In contrast to January, the extreme heat in September was related to strong forcing from climate drivers that are associated with an increased chance of having extreme heat over Australia. The anomalous atmospheric circulation during the periods of heat included a strong negative SAM, MJO activity (phases 1/2 and 5/6), a persistent region of subtropical high pressure over the Tasman Sea, and an absence of blocking. POAMA ensemble mean forecasts of these climate driver indices one week in advance captured the sense of these drivers, but the magnitude was weaker than observed. Previous research, evaluating a 30-yr hindcast set, has demonstrated that POAMA can in general predict the MJO up to three weeks in advance (Marshall et al. 2012a) and subseasonal variations of the SAM and blocking up to two weeks in advance (Marshall et al. 2012b, 2013a). POAMA is also largely able to reproduce the impact of these drivers on extreme heat over Australia, although it tends to underestimate the magnitude of the teleconnections (Marshall et al. 2013b). It appears that POAMA’s reasonably good probabilistic forecasts of extreme Tmax over Australia, initialized in advance of the warming, for the fortnight at the start of the heat wave (with a 1-week lead time) as well as for the month of September, is a result of the model being able to forecast the important climate drivers and their teleconnection to Australian climate reasonably well. Having some knowledge of the climate drivers and their expected teleconnections to Australian climate can be useful for understanding and diagnosing a particular real-time forecast of extreme heat from POAMA, including identifying those events that may be more predictable in advance.
Other studies have suggested that climate change may have played a role in the heat experienced during January and September (Lewis and Karoly 2013; Arblaster et al. 2014). To place 2013 in context, the record September temperatures followed Australia’s second-warmest winter on record for maximum temperatures and Australia’s hottest summer (from December 2012 to February 2013) on record, and in fact Australia’s warmest January–September period on record (Bureau of Meteorology 2013b). Although not considered in this study, POAMA’s ability to capture the long-term warming trend through its use of observed initial conditions may be providing some predictability for the forecasts. In a study of the September event, Arblaster et al. (2014) did seasonal forecast sensitivity experiments with POAMA and concluded that the land temperature initial conditions were a key contributor to the predicted temperature anomalies over Australia for September.
Our assessment of the POAMA forecasts is performed at the native resolution of the model and therefore looks at fairly broad spatial-scale performance. It is clear from the forecast products that, owing to the limited model resolution, local detail is absent in the forecasts. Interpretation of the POAMA forecasts at the local scale may result in errors in as much as the local conditions contribute to or modify the nature of the heat event, since local conditions are often not adequately captured by the model (e.g., land–sea-breeze circulations, local-scale topography and aspect, influence of coastlines, and finescale land surface conditions). The next version of POAMA will improve on the spatial resolution (Zhao et al. 2013). In addition, appropriate statistical downscaling could be applied to add local detail to the forecasts. The view of the performance and usefulness of these POAMA forecasts by an end user will also ultimately depend on the nature of the application.
This paper has focused on POAMA’s performance for just three heat wave events and as such does not enable one to draw conclusions about the overall skill of the system. The latter has been documented in previous studies using a large set of hindcasts spanning a 30-yr period (Hudson et al. 2011a, 2013, 2015; Marshall et al. 2013b; White et al. 2013). The case studies provide a valuable complement to these retrospective skill assessments. We conclude that POAMA provided relatively good warning and guidance on subseasonal time scales for two of the three major heat wave events occurring in 2013. This work adds to the growing body of evidence of the potential to extend traditional weather forecasts and warnings for extreme events to include longer-lead probabilistic guidance.
This work was supported by the Managing Climate Variability Program of the Grains Research and Development Corporation. The authors thank Eunpa Lim, Pandora Hope, and three anonymous reviewers for their useful comments during the preparation of this manuscript.