Precipitation and Temperature Data for the Sydney Catchment Area, Australia

This dataset contains time series for monthly precipitation over six sites (Blackheath, Braidwood, Darkes Forest, Goulburn, Lithgow and Moss Vale) in the Sydney Catchment Area (SCA) and monthly mean maximum and mean minimum temperature for three sites (Goulburn, Lithgow, and Moss Vale) in the SCA. This data was used in the study Attribution and Prediction of Precipitation and Temperature Trends within the Sydney Catchment Using Machine Learning. The data was originally from the Australian Bureau of Meteorology Climate Data Online (http://www.bom.gov.au/climate/data/index.shtml), but has been updated to have missing values (8% of data) filled using a moving average centred on the year for which the data is missing. 

Below is the abstract for the paper:

Droughts in southeastern Australia can profoundly affect the water supply to Sydney, Australia's largest city. Increasing population, a warming climate, land surface changes, and expanded agricultural use increase water demand and reduce catchment runoff. Studying Sydney's water supply is necessary to manage water resources and lower the risk of severe water shortages. This study aims at understanding Sydney water supply by analysing precipitation and temperature trends across the catchment. A decreasing trend in annual precipitation was found across the Sydney catchment area. Annual precipitation also is significantly less variable, due to fewer years above the 80th percentile. These trends result from significant reductions in precipitation during spring and autumn, especially over the last 20 years. Wavelet analysis is applied to assess how the influence of climate drivers has changed over time. Attribute selection was carried out using linear regression and machine learning techniques including random forests and support vector regression. Drivers of annual precipitation included Niño3.4, SAM, DMI and measures of global warming such as the Tasman Sea Sea Surface temperature anomalies. The support vector regression model with a polynomial kernel achieved correlations of 0.921 and a skill score compared to climatology of 0.721. The linear regression model also performed well with a correlation of 0.815 and skill score of 0.567, highlighting the importance of considering both linear and non-linear methods when developing statistical models. Models were also developed on autumn and winter precipitation but performed worse than annual precipitation on prediction. For example, the best performing model on autumn precipitation, which accounts for approximately one quarter of annual precipitation, achieved an RMSE of 418.036 mm² on the testing data while annual precipitation achieved an RMSE of 613.704 mm². However, the seasonal models provided valuable insight into whether the season would be wet or dry compared to the climatology.