2. Development History


FlowCast has been developed over a period of ten years with funding from a range of sources. It was originally developed to generate probabilistic forecasts of streamflow and irrigation allocations in the project ‘A decision support system for improving water use efficiency in the northern Murray-Darling Basin’ (Abawi et al. 2001). It has since been used to develop seasonal outlooks for South East Queensland’s water supplies, and a range of in-house analyses. Most recently, it has been reengineered as a part of the ACIAR funded project, ‘Seasonal climate forecasting for better irrigation system management in Lombok’ (SMCN/2002/033), which aims to improve agricultural decision-making in Lombok, Indonesia. It has now evolved into a powerful generic piece of software suitable for any empirically based climate-forecasting scenario, extending the basic functionality of other similar products.


2.1 Evolution of Seasonal Climate Forecasting Software


FlowCast  is one of at least five stand-alone Windows-based software applications for seasonal climate forecasting to have been developed in the last 20 years. It was originally developed to overcome many of the limitations at that time of using Rainman (Clewett et al. 1999) as a systems’ modeling tool for forecasting streamflow. Rainman was the pioneer in the field of seasonal climate forecasting software and was available for sale to the general public, winning many awards. Although developed within the same organization (the former Queensland Centre for Climate Applications, DPI), FlowCast  and Rainman evolved with significantly different roles despite both (originally) using similar forecasting and skill testing methodologies. Rainman was always the more refined and user-friendly software of the two, packaged with high quality documentation, tutorials, and support, including the book ‘Will it Rain? – The effects of the Southern Oscillation and El Niño on Australia’ (Partridge et al. 1991). FlowCast  was designed primarily as an in-house researchers’ tool in a systems’ modeling and post-processing role, with powerful interactive analyses and flexible data inputs, but limited documentation and no formal support or training. Rainman was distinguished by the inclusion of a preinstalled climate database (rainfall, temperature, and later on, streamflow) and a limited number of predictive systems that greatly simplified its operation, but had limited support for data outside of that installed. In contrast, FlowCast  didn't come with any predictand data preinstalled, but was designed to easily input any number and type of predictor and predictand data. Different software engineering practices have impacted heavily on both designs, with Rainman’s development being limited by its use of mixed Visual Basic and Fortran technologies. On the other hand, FlowCast  was developed using a more modern and unified C++ based technology leading to simpler development and maintenance while providing greater flexibility and power.

The development of Rainman spanned over thirteen years from 1991 to 2004 with its evolutional changes playing a significant role in the design of FlowCast . The first version of Rainman (Clarkson & Owens, 1991) provided the operational functionality for future versions despite being DOS-based with a textural interface and written in FORTRAN and BASIC. Forecasts were in the form of tabulated probability distributions based on stratification using a three category SOI–based predictor. It contained information on the climate of over 400 stations in Queensland with analyses allowing the user to examine the wettest and driest months, and to study historical droughts. Planting opportunities could also be assessed based on the daily rainfall accounting of wetting and drying cycles. The second version released in 1994 (Clewett et al. 1994) added Australian wide coverage of 3700 climate stations, and a DOS-based graphical user interface with tabular, chart and map-based outputs. Additional predictive systems were incorporated including the SOI-based five-phase system (Stone, 1992), and Indian Ocean SSTa system (Drosdowsky, 1993).

A Windows-based version of Rainman (Version 3) was introduced in 1999 (Clewett et al. 1999) with internet-based updating of data and integrated electronic users guide, tutorials, and the second edition of ‘Will it Rain?’ (Figure 1a). It was developed in a mixture of Visual Basic, FORTRAN and C-based code and dynamic link libraries, which continued through to the current version. A supplement was released in 2001 called STREAMFLOW, containing monthly (and some daily) data from over 500 gauging stations on major rivers in Australia, and with the addition of a ‘persistence’ analysis. The two products were later merged in 2003 to form Version 4, “Rainman + Streamflow”. An international version was released in 2002 featuring rainfall data from over 12,000 locations, and powerful spatial analyses for a range of outputs (Figure 1b).

Rainman’s development ceased in 2004 due to a change in organizational management and direction, although it is currently still available for sale in ‘Standard’, ‘Educational’ and ‘Professional’ forms with limited support. Its evolution through four versions was both engineering and data motivated, catering for a wide range of stakeholder needs. The fact that it is still in demand and being used today is testament to the quality of its market research and design.

Figure 1: Screenshots of Rainman Version 2 and 3

Figure 1 : Rainman version 3 and 4.

On the other hand, FlowCast has also evolved through four versions (Figure 2), but its revisions have been forced through the need to enhance its technical capabilities. The first version of FlowCast was released in early 2000 containing much of the functional capacity of Rainman at that time (sans database), including stratification-based forecasting algorithms, non-parametric hypothesis testing (Kruskal-Wallace, Kolmogorov Smirnov, and Wilcoxon-Mann-Wittney), and several analyses to review raw input data, summary information and forecast results. It was principally designed to import, manipulate and combine modeled time-series data from the IQQM water allocation model (NSW Department of Land and Water Conservation, 1998a,b,c,d,e), and originally only supported the SOI-based five-phase predictor system.

A second version was released in the same year with a revised interface, and advanced predictor functionality for composing stratifications from a limited number of predictor data. It featured a custom designed ‘period-setter’ user interface tool for adjusting the predictor and predictand periods that was synchronized with the analysis engines to provide real-time updating of results (unlike Rainman’s equivalent tool which was modal in behavior). The period-setter contained separate controls for each stratification, and custom editor windows to create and manipulate rules, and define stratifications from multiple time-series. These features were further enhanced with Version 3 in 2003, which also focused on improving the flexibility of the data inputs and refining the graphical user interface. This version was specifically designed to support any type of predictand and predictor data, moving on from its streamflow-based heritage.


Figure 2: Evolution of FlowCast’s Graphical user interface.

Figure 2: Evolution of FlowCast's Graphical user interface.

In 2004, the developers of FlowCast won a contract with the Australian Bureau of Meteorology (BOM), in an AusAid-funded project, to develop the seasonal prediction software SCOPIC (Seasonal Climate Outlooks in Pacific Island Countries) for Pacific Island Country (PIC) meteorological services. The software was required to simplify the process of generating seasonal climate forecasts using the discriminant analysis methodology of BOM’s operational forecast system, while being customizable for individual PIC needs. That neither Rainman nor FlowCast  were deemed suitable for this purpose reflected the need for an even simpler interface than either provided, and that each employed a stratification methodology instead of a discriminant analysis system. BOM also wanted to retain ownership of the product and maintain tighter control over the software development given that they had earlier commissioned another product called ‘ENSO’ (BOM, 2003) though a private company, which had failed to live up to expectations. While ENSO was initially released to the PIC members as a training tool, it never saw operational use as it was poorly designed with a primitive user interface, limited graphical outputs, and inflexible inputs (Figure 3a).

Figure 3: Software developed for Pacific Island Countries: (a) ENSO; and (b) SCOPIC.Figure 3: Software developed for Pacific Island Countries: (a) ENSO; and (b) SCOPIC.

At the onset of this project, the current version of FlowCast (Version 4) was being completely reengineered to develop an open stable platform to accommodate future functionality requirements (as part of the ACIAR project in Indonesia) focusing on both point-based (‘station’) and spatial analyses. This framework was immediately used to rapidly prototype the first version of SCOPIC with the newly developed station-analysis functionalities. From thereon, the development of SCOPIC and FlowCast evolved in unison, with FlowCast being used as a test-bed for SCOPIC design. New analyses and algorithms were initially prototyped within the FlowCast platform before being refined and customized inside of SCOPIC. It was during this period that FlowCast gained the discriminant-analysis forecasting algorithms, while its spatial analysis tools were being developed to their present form. In their latest versions, both programs share a significant amount of common code, which is mostly analysis-related. However, the internal data algorithms, storage and computation sequencing are significantly different since FlowCast has a predominantly ‘spatial analysis’ role.

In 2005 the first version of SCOPIC was released to PIC members, accompanied by an extensive training program though workshops in each participating country. SCOPIC graphical user interface design and operational functionality was based around five steps for generating a forecast: (1) Organise Project; (2) Explore Data; (3) Analyses Relationships; (4) Assess Skill; and (5) Generate Report. Each step was presented in separate tab-sheets with multiple analyses and functionalities it in each. Dynamic report using logic-based XSLT technology allows written reports to be automatically generated (in rich-text format) in the language of each country. Version 2 (Figure 3b) was released in 2006 adding a range of drought analyses employing both ‘decile-method’ and standardized precipitation index (SPI) methodologies. Interactive ‘eLearning’ tutorials have also now been developed as well as extensive online help and support. SCOPIC is currently being used operationally in ten Pacific Island countries and is continually being refined and revised. The software was rewritten in 2012 to modernise its software architecture while providing essentially the same functionality.

In 2005, another program for seasonal climate forecasting called NSFM (‘Non-parametric Seasonal Forecasting Model’ - Chiew and Siriwardena, 1995) was also being developed in Australia as part of the catchment modeling toolkit (http://www.toolkit.net.au) (Figure 4b). This software employs discriminant analysis methodology to produce exceedence probabilities for streamflow. NSFM allows the use of two predictor variables including an ENSO-related variable (SOI or SST), and antecedent streamflow (persistence). However, this capability could potentially artificially enhance skill since predictor components in discriminant analysis are assumed to be orthogonal, and there is often correlation between ENSO and streamflow (in Australia). Nevertheless, using persistence in streamflow forecasts is usually highly beneficial, which has also been captured in all versions of FlowCast and Rainman Streamflow. NSFM is a robust and userful program although its interface is technical and inflexible, lacking the in-depth analyses and refinement of Rainman, SCOPIC and FlowCast .

Another software application deserving mention is the US-developed Climate Predictability Tool (CPT) (http://iri.columbia.edu/outreach/software), which although it hasn’t played a role in influencing FlowCast's design, is currently being used operationally in many countries around the world including South America and South-East Asia (Figure 4b). CPT also has a long development history from 2001 to the present day going through nine revisions. It is designed to use downscaled GCM projections or statistical analysis of SST fields to generate forecasts. It offers a range of analysis methods include principal components regression, canonical correlation analysis, and multiple linear regression. Extensive verification tests are provided including most of those recommended by the World Meteorological organization. It produces a range of quality textural, chart and spatial outputs, but its usability is hindered by a technical and bland menu-driven user interface.

Figure 4: Other softwareFigure 4: Other software

Version 4 of FlowCast was completed and released to the Indonesian Meteorological Service (BMG) in April 2008, and is currently being used in research and systems modeling roles within the Queensland Climate Change Centre of Excellence. Version 4 has nearly all of the technical capacity of the previously described software, except for the CPT program. Notable omissions include the drought analysis and dynamic report generation functionalities of SCOPIC, and the preinstalled climate database of Rainman. Also most of the advanced stratification tools that were incorporated in earlier versions of FlowCast have since been omitted since it is easier to capture this detail though simple XML-based definition files. Non-parametric hypothesis testing has also been removed in favor of hindcast-based skill assessments.