About the Local Forecast

1. Forecast guidance from several models

The process begins with numerical weather-prediction guidance from major international forecasting centres. These models simulate the atmosphere independently, using different model designs, resolutions, initial conditions and update schedules. Their forecasts can therefore differ, even for the same place and time.

Using several models provides a broader evidence base than relying on a single forecast. Before analysis, the system also checks whether each model has enough recent data to support a meaningful local assessment. A model with insufficient coverage is not treated as though its performance were known.

2. When the system runs

LACO separates learning from forecast production. A daily analysis run refreshes what the system has learned from the rolling history of model forecasts and Ferny Grove observations. That analysis produces the local corrections, performance measures and weights used by later forecast runs. Keeping this as a separate daily process allows the evidence to be updated regularly without relearning the full history every time a new forecast is issued.

Forecast production runs four times each day, corresponding to the main six-hourly global forecast cycles. A Python controller requests the model data through Open-Meteo and records the availability of each model run. Because the models are not all published at the same time and some models use a different source cycle for a given forecast cycle the required model and initialisation time are resolved from configuration rather than assumed to be identical.

Before a forecast starts, the controller checks that every model marked as required for that cycle has been reported available and has actually been downloaded into the local model data store. If the required inputs are still arriving, it waits and checks again. It can stop waiting when either a certain amount of time has occured that would excessively delay publication or recent availability history shows that a required feed is stale.

Once readiness is confirmed, the forecast run uses one consistent, locked set of model initialisation times and the latest eligible daily analysis. The pipeline the forecast is ran under validates its configured processing steps, records the parameters used, runs each enabled forecast variable and logs failures. These controls make a forecast run reproducible and prevent different parts of the same issued forecast from accidentally using different model cycles.

3. Learning from local observations

Model forecasts are matched with observations recorded at Ferny Grove. The learning process uses a rolling recent history window, with newer evidence given more influence than older evidence. This allows it to adapt as model behaviour and seasonal conditions change without overreacting to only one or two recent events.

The comparison is performed at detailed levels rather than assigning one permanent score to each model. Performance can vary with how far ahead the forecast is and with the model's initialisation cycle, so those contexts are analysed and retained when enough evidence is available.

4. Correcting local behaviour

The analysis looks for systematic local differences between each model and the observations. Corrections are configured and learned separately for each forecast variable because different variables have different behaviour, units and useful adjustment methods. The current temperature forecast uses the following adjustments:

• Baseline offset: removes a model's recurring overall tendency to forecast too high or too low before finer patterns are assessed.
• Hour-specific residual bias: corrects error that remains after the baseline adjustment and varies by time of day.
• Daily slope: adjusts a rise or fall that is systematically too strong or too weak while retaining the model's broader pattern.
• Daily shape: adjusts the curvature and timing of the within-day temperature cycle after its general level and trend have been separated.
• Maximum anchoring and daily recentering: prevent the reshaped curve from drifting away from its learned daily maximum or mean after the other corrections are applied.
• Daytime and night-time range: treats the warm side and cool side of the daily cycle separately because daytime heating and overnight cooling do not necessarily have the same error behaviour.
• Cold and hot conditions: learns separate relationships for unusually cool minima and warm maxima, using locally and seasonally appropriate thresholds rather than one fixed definition of an extreme.

These temperature adjustments can improve the forecast by removing repeated local errors that weather models cannot resolve well at a specific location. Future variables can use their own appropriate correction sequence rather than inheriting temperature specific assumptions. Within each variable, enabled adjustments are applied in a defined sequence so that a later adjustment works on the result of the earlier one rather than independently moving the same value several times.

The system does not assume every learned pattern is trustworthy. It requires sufficient evidence, gives newer observations more influence, constrains correction strength, and can use longer history or monthly climatological fallbacks when recent extreme-weather samples are sparse or unstable. Individual correction stages can also be disabled. These controls are intended to retain the useful weather signal in the source guidance and reduce overcorrection;not to force every past forecast to match the observations.

5. Giving more influence to better performing guidance

Recent mean absolute error is used to estimate how much influence each model should have. Lower error guidance receives more weight; higher error guidance receives less. Weighting can reflect overall performance, forecast lead time and initialisation cycle.

Only models available for a particular forecast hour participate in that hour's calculation, and their effective weights are rescaled to form a complete total. This means a missing model does not leave a gap or make the remaining weights invalid.

The weights describe recent measured performance, not certainty. A highly weighted model can still have a substantial forecast error, and agreement between models does not guarantee the outcome.

6. Producing one locally corrected forecast

For each forecast time, the corrected model values are combined using their effective performance weights. This creates the raw consensus: one value representing the available guidance rather than a simple unweighted average.

The processing path is retained in stages: the Raw Consensus shows the weighted model guidance before local correction, the Preliminary forecast applies the earlier level and hourly-bias adjustments, and the Final forecast applies the complete enabled correction sequence. Keeping these stages allows internal verification to test whether the processing actually improved the source consensus rather than assuming that every adjustment helped.

Each published forecast is stored as an issued product. This is important for honest verification: past accuracy is calculated from the forecast that was actually available at the time, not from a later reconstruction using information that was not then known.

7. Updating the near-term forecast with recent observations

Enhanced forecast feature: this process is currently included for all users. It refines the near term after the core hourly forecast has been produced.

Weather models cannot represent every local change at exactly the right time. After the base forecast is issued, recent Ferny Grove observations are compared with its near-term values. When a difference is large enough and sufficiently persistent, a limited observation based adjustment can be applied to the early forecast period.

This process is known as nowcasting. The adjustment is capped, can be refreshed as new observations arrive, and is not carried unchanged into the long range. It is reduced later in the forecast, can cease when the forecast trend changes direction, and can also be removed when observations and the original forecast come back into agreement.

Nowcasting is intended to improve short term continuity with conditions already unfolding. It does not replace the broader model guidance and does not assume that a current local error will continue indefinitely.

8. Creating the 15-minute forecast and daily extremes

Enhanced forecast feature: the 15-minute series and its more precise estimates of daily extreme values and times are currently included for all users.

The core forecast is hourly. A shape preserving cubic interpolation is used to create values at 15 minute intervals between those hours. The method passes through every original hourly forecast point and uses the neighbouring hourly trend to create a smooth curve.

This is more informative than drawing straight lines between hours, especially near a daily minimum or maximum. The curve may reach its turning point between the hourly timestamps, so the reported daily minimum and maximum—and their times—are calculated from the interpolated curve rather than simply selecting the lowest and highest hourly rows.

Interpolation adds time detail to the same forecast; it does not add new atmospheric observations or make the underlying forecast more certain. When nowcasting changes the hourly forecast, the 15 minute series and daily extremes are recalculated from those updated values.

9. Publishing and verifying the forecast

The public page displays the newest eligible operational forecast. Times are shown in Brisbane local time (AEST). Where an enhanced product is available, it may also include recent observation based updates or finer time detail derived from the hourly forecast.

The current public verification is specifically temperature verification. Once temperature observations become available, issued hourly temperature forecasts are matched with the observation for the same hour. Only complete calendar days are included. Accuracy is reported in degrees Celsius using mean absolute error, average bias, the proportion of errors within stated temperature thresholds, performance by lead day, and daily minimum and maximum temperature accuracy.

This ongoing temperature verification makes recent performance visible and gives the daily learning process fresh evidence for later temperature forecasts. The published results describe the temperature forecast as published, including applicable near-term updates where available to be shown. LACO's verification framework is variable based: as other variables are published, each will be matched with its corresponding observations and assessed using measures, units and thresholds appropriate to that variable.