Exemplary Energy’s work on Precipitation Disaggregation, converting daily data as was recorded by the Bureau of Meteorology (BoM) in the 1990s and much of the 2000s, has been reported on in Exemplary Advances during its development over several years. It was presented in extended abstract format at the Asia Pacific Solar Research Conference (APSRC) in December 2024. Our full paper on that ground-breaking work has now been published in the learned journal Stochastic Environmental Research & Risk Assessment.
Our team has successfully developed an innovative machine learning approach that generates high-resolution precipitation data (mostly rainfall) critical for modern building performance and hygrothermal simulations such as in WUFI. Our work introduces a long short-term memory (LSTM) neural network that can disaggregate daily precipitation readings into half-hourly intervals—a significant advancement over existing machine learning models limited to hourly resolution.
The table below summarizes model performance across the five climate zones, showing Root Mean Square Error (RMSE), wet half-hour detection, and skill score accuracy on test data set.
| Location | NCC Climate Zone | RMSE (mm) | Wet Half-Hour Detection (%) | Skill score (Half-Hourly) |
| Cairns | Zone 1 | 0.6899 | 65.88 | 0.68 |
| Brisbane | Zone 2 | 0.5670 | 60.68 | 0.65 |
| Sydney | Zone 3 | 0.5136 | 70.85 | 0.73 |
| Melbourne | Zone 6 | 0.2345 | 55.86 | 0.62 |
| Canberra | Zone 7 | 0.2857 | 68.14 | 0.70 |
Note:
- RMSE measures average prediction error (lower is better).
- Wet Half-Hour detection shows the percentage of rainy half-hours correctly identified in the exact half-hour, reflecting timing accuracy.
- Skill score evaluates how well the model captures the frequency and distribution of precipitation events (higher is better, 1.0 is perfect).
In comparison to the previous model (Ferrari et al. (2022)), the RMSE improved by over 30% for Canberra (0.4513 vs 0.65 mm).
In the Australian context, the extreme weather conditions like draughts and heavy rains make precise data essential. By incorporating meteorological variables and a specialised normalisation layer, our model maintains statistical consistency while preserving daily precipitation totals.
This breakthrough enables more precise modeling for building designs and water management systems, especially valuable in regions like late 20th century Australia where historical sub-hourly precipitation data is limited.
You can read the full paper here: https://doi.org/10.1007/s00477-025-02996-0
Weather and climate files incorporating precipitation data in ACDB and EPW formats are available through our sales portal at no extra charge.