The Bureau of Meteorology has deployed machine learning systems that improve short-term rainfall prediction accuracy by 35% compared to conventional numerical weather models, providing better warnings for flash flooding and severe weather.

The system, called NowcastNet, uses neural networks trained on decades of radar imagery, satellite data, and surface observations to predict rainfall patterns over the next 1-3 hours. This “nowcasting” timescale is critical for emergency response but has been challenging for traditional physics-based weather models.

The improved forecasts help emergency services position resources before flash floods develop, allow water utilities to manage stormwater systems more effectively, and provide construction sites and outdoor events with better short-term weather information.

The Nowcasting Challenge

Weather forecasting uses different approaches depending on the forecast period. Multi-day forecasts rely on global numerical models that simulate atmospheric physics. These work well beyond about six hours but struggle with rapid weather changes at shorter timescales.

For the next one to three hours, extrapolating recent weather observations often outperforms physics-based models. Radar shows a storm moving northeast at 40 kilometres per hour, so simple extrapolation predicts where it will be in an hour. But this breaks down when storms intensify, change direction, or new storms develop.

Machine learning offers a middle path. Neural networks can learn complex patterns in how weather evolves by training on vast historical data. They capture behaviours that are difficult to model with physics equations but are evident in observations.

Dr Susan Chen, who leads the Bureau’s AI forecasting program, said the goal wasn’t to replace numerical models but to complement them. Physics-based models excel at some tasks while machine learning excels at others. Using both provides better forecasts than either alone.

Technical Approach

NowcastNet takes radar imagery from Australia’s weather radar network, satellite observations, surface station data, and lightning detection as inputs. The network architecture is based on convolutional neural networks that excel at processing spatial data like images.

The model was trained on ten years of Australian weather radar data, learning patterns in how rainfall develops and moves. Training required substantial computing resources, using specialised AI processors running for several weeks to optimise millions of neural network parameters.

The system generates probabilistic forecasts, predicting not just whether rain will occur but the likelihood of different rainfall amounts. This uncertainty information helps users make risk-based decisions. Emergency managers treat a 30% chance of heavy rain differently than 80%.

Integration with existing Bureau systems was critical. The machine learning forecasts feed into the same dissemination systems as traditional forecasts, appearing on Bureau websites and apps, and flowing to emergency services and other users through established data feeds.

The system updates every ten minutes using the latest radar observations. This frequent updating allows forecasts to adapt quickly as weather conditions change. Traditional numerical models typically update hourly at best.

Validation and Performance

The Bureau validated NowcastNet using a year of held-out data not used in training. This tests whether the model generalises to new weather situations or just memorised training examples.

Results showed 35% improvement in skill scores for one-to-three-hour rainfall forecasts compared to extrapolation methods previously used for nowcasting. Improvement was largest for heavy rainfall, exactly where better forecasts provide most value.

The model performed well across different Australian climate zones, from tropical northern regions to temperate south. This was encouraging because weather patterns differ substantially between regions and the model needed to handle that diversity.

Some weather situations challenged the model more than others. Rapidly developing thunderstorms were harder to predict than steadier rainfall from frontal systems. Coastal rainfall affected by sea breezes showed mixed results. These limitations are being addressed in ongoing development.

Independent evaluation by emergency services users found that the improved nowcasts helped with decision-making. Flood warnings could be issued with more confidence and better timing. Transport operators could better plan for weather disruptions.

Emergency Services Applications

State emergency services in New South Wales and Queensland were early adopters, using the enhanced nowcasts for flood response positioning. When heavy rainfall is forecast, they pre-position rescue teams and equipment to enable faster response to flood calls.

The probabilistic forecasts help with this positioning. If there’s a 60% chance of heavy rain in a particular region, that’s enough to justify moving resources but not enough to be certain they’ll be needed. Understanding forecast uncertainty enables better risk management.

Transport authorities use the forecasts for managing road and rail operations. Heavy rain affects visibility and road conditions. Better short-term forecasts allow variable speed limits and travel warnings to be activated at appropriate times.

Water utilities use the forecasts to manage stormwater systems. Urban stormwater infrastructure can be overwhelmed by intense rainfall, causing localised flooding. Advance warning enables operators to adjust pumping and drainage to maximise capacity.

Agricultural users have also found value. Knowing when heavy rain will arrive helps with decisions about field operations like harvesting or fertiliser application. A three-hour warning of rain helps avoid situations where expensive field operations are interrupted.

Integration with Traditional Forecasting

The Bureau’s forecasters use NowcastNet alongside traditional numerical models and their own experience-based judgment. The machine learning system provides one input to forecasting decisions but doesn’t replace human forecasters.

This human-machine teaming approach lets forecasters catch situations where the model performs poorly and apply context that algorithms miss. Forecasters know local geography, recent weather patterns, and user needs in ways that machine learning systems don’t capture.

Over time, forecasters learn when to rely heavily on the AI system and when to downweight its predictions. This calibration improves forecast quality beyond what either humans or algorithms achieve alone.

There’s ongoing research on how to best present AI forecasts to forecasters. Too much information creates cognitive overload. Too little leaves forecasters uncertain about the AI system’s reasoning. Finding the right balance involves user experience research and iterative design.

Limitations and Ongoing Work

Machine learning weather forecasting has limitations that physics-based models don’t face. The models can only learn from weather patterns that occurred in training data. Truly novel weather events might not be predicted accurately.

The models also struggle with rare extreme events that appear infrequently in training data. Neural networks optimise overall performance, which means rare events get less weight than common situations. This is problematic because rare extremes are often most important for warnings.

The Bureau is addressing these limitations through hybrid approaches that combine machine learning and physics-based models. Using physics models to constrain machine learning predictions could prevent physically impossible forecasts while retaining the pattern-recognition advantages of AI.

Additional training data could improve performance. The Bureau is exploring whether data from overseas weather services could augment Australian training data, though differences in climate and radar systems create challenges.

Explainability is another research direction. Understanding why the model makes specific predictions would help forecasters trust and effectively use the system. Techniques from explainable AI are being adapted to weather forecasting applications.

Global Context

Weather services worldwide are exploring machine learning for forecasting. Google’s MetNet, the European Centre for Medium-Range Weather Forecasts’ GraphCast, and several other systems show that AI can contribute to weather prediction.

Different approaches focus on different timescales and weather elements. Some target multi-day forecasts while others focus on nowcasting like the Bureau’s system. Some predict all weather variables while others specialise in specific phenomena like precipitation.

Australia’s relative geographic isolation means that weather systems develop locally more than in mid-latitude regions where global weather patterns dominate. This makes Australian forecasting somewhat distinct from European or North American situations.

The Bureau participates in international research collaborations on AI weather forecasting, sharing methods and insights. Weather knows no borders, and advances anywhere can benefit forecasting globally through technology transfer.

Future Developments

The Bureau plans to extend machine learning approaches to other forecasting challenges beyond rainfall nowcasting. Temperature forecasting, wind prediction, and severe weather warning all have potential for AI enhancement.

There’s also scope to improve the current nowcasting system. Higher resolution forecasts with more spatial detail would benefit some users. Longer forecast horizons extending to six or twelve hours would bridge the gap between nowcasting and numerical model forecasts.

Integration with impact forecasting represents another direction. Rather than just predicting rainfall, forecast likely flooding, traffic impacts, or infrastructure stress. This requires combining weather forecasts with information about infrastructure and vulnerabilities.

For organisations that depend on accurate weather information, the Bureau’s machine learning nowcasting represents meaningful progress in a critical capability. Better short-term rainfall forecasts won’t solve all weather-related challenges, but they enable more informed decision-making during the critical window when weather hazards are imminent.