The new Local Forecast on Ferny Grove Weather is produced by a forecasting system called LACO — Local Atmospheric Corrected Output.
In the previous blog post, we introduced the Local Forecast and explained what you can expect to find when using it.
This article explains in more detail what happens behind the forecast.
The current operational implementation of LACO focuses on temperature and produces a single best-estimate forecast, commonly described as a deterministic forecast. Temperature is the first published variable, with the broader system designed so that additional weather variables can be added over time.
LACO takes weather model guidance from major weather forecasting centres around the world and learns from how they perform at Ferny Grove based on weather observations reported on this website. It applies locally learned corrections, combines the model guidance according to recent performance, and can respond to new observations after the forecast has been produced.
It therefore takes the enormous amount of weather model guidance available globally and turns it into a forecast specifically tailored to Ferny Grove.
Why use multiple weather models?
Major forecasting centres around the world operate their own numerical weather prediction models.
LACO uses guidance from a range of these models, including ECMWF, GFS, ICON, the UK Met Office, GEM, JMA and others.
These models use different modelling systems to predict how the atmosphere may evolve, so they do not always agree.
One might forecast a warmer afternoon than another. One might predict a considerably cooler overnight minimum. Their forecasts can also change from one model run to the next.
There is another important complication: no single model is always best at a given location.
Choosing a favourite model therefore discards potentially useful information from all the others.
A simple average makes use of more of that information, but introduces another assumption: that every model deserves exactly the same influence.
Instead, following an approach commonly used in operational forecasting, LACO asks: Based on what has actually happened at Ferny Grove, how much influence should each model have on the forecast?
Learning from observations
The observations recorded by the Ferny Grove weather station provide the evidence needed to answer that question.
LACO can take an earlier forecast from each model and compare it with the temperature that was subsequently observed.
Repeated over many forecasts, this builds a picture of how the models have actually been performing at this location.
But LACO doesn’t simply assign each model a permanent score.
Performance can vary according to how far into the future the forecast extends. A model that performs particularly well during the first part of a forecast isn’t necessarily the one that performs best several days later.
Performance can also differ between model runs.
LACO therefore learns at a finer level than simply deciding that one model is ‘good’ and another is ‘bad’.
Recent performance matters
Model performance also isn’t necessarily constant over time.
For that reason, LACO doesn’t treat an error from months ago as being equally important as an error from a recent forecast.
Recent forecast performance receives greater weight, while the influence of older evidence gradually declines.
This gives LACO a degree of adaptability.
If a model has recently been forecasting Ferny Grove particularly well, it can gain greater influence in the consensus. If its relative performance deteriorates, that influence can subsequently decline.
The model weights used by LACO are therefore learned from actual forecast performance rather than permanently predetermined.
Correcting systematic forecast behaviour
Weighting is only one part of the analysis.
Suppose a particular model produces useful forecasts but has recently tended to forecast Ferny Grove slightly too warm.
Simply reducing that model’s weight throws away some potentially valuable information.
If the behaviour is sufficiently systematic, another possibility is to learn from it.
LACO analyses each day previous forecasts and observations for patterns in forecast error and can apply several forms of correction to the incoming model forecasts.
These corrections can account for things such as overall offsets, recurring behaviour at particular times of day, and differences in the shape or progression of the forecast temperature curve.
LACO can also consider behaviour around particularly warm or cool conditions.
The objective isn’t to make arbitrary changes to the models. Corrections are based on evidence from previous forecasts and observations.
So before producing the consensus, LACO is effectively asking two different questions: How much influence should this model have? And what have we learned about the way this model forecasts Ferny Grove?
Producing the consensus forecast
Once the individual model forecasts have been analysed and corrected, LACO combines them using the learned weights.
The result is the underlying consensus forecast. This is different from simply averaging all of the model temperatures: each model can contribute differently according to the evidence LACO has accumulated about its performance. As new forecasts are verified, that evidence changes and the composition of the consensus can change with it.
This forms the basis of the Local Forecast.
But the process doesn’t end when that forecast is generated.
Responding to what is happening now
New global weather model guidance is not available every hour, but the Ferny Grove weather station continues observing what is actually happening between model updates. LACO may therefore have evidence that the temperature is developing differently from the existing forecast before replacement model guidance becomes available.
This is where the nowcasting component of the system ran each hour, becomes involved.
Nowcasting compares recent observations with the forecast.
It deliberately doesn’t react to every error. A single observation being slightly warmer or cooler than forecast isn’t necessarily evidence that the forecast has gone wrong.
Instead, the system looks for a sufficiently meaningful and persistent difference.
When its triggering conditions are met, nowcasting can apply a controlled adjustment to the short-range forecast.
The adjustment is constrained so that a temporary discrepancy doesn’t produce an unreasonable forecast change. It can also taper away with time, smoothly returning the forecast towards its underlying model-based trajectory.
This gives LACO two distinct feedback mechanisms. Historical observations help it learn how to construct future forecasts, while current observations help it respond when a forecast already in progress begins to depart from what is being observed.
A continuous forecast between the hours
The resulting temperature forecast is based on hourly values.
But the atmosphere doesn’t operate in hourly steps.
If the forecast is 24 °C at 2 pm and 25 °C at 3 pm, the temperature doesn’t remain at 24 °C before suddenly jumping to 25 °C.
LACO can therefore construct a continuous curve through the hourly forecast.
This interpolation is designed to preserve the actual hourly forecast points while representing the expected progression between them.
Finding the actual daily extremes
If we examined only the hourly forecast values, we would effectively assume that every daily minimum and maximum occurred exactly on the hour.
There is no meteorological reason that should be true.
LACO can mathematically examine its continuous temperature curve for turning points.
If the curve reaches its highest value at 2:42 pm, for example, LACO can identify that as the forecast maximum rather than forcing it to either 2 pm or 3 pm.
Similarly, if the temperature continues falling through the final hour of a calendar day, the minimum can continue through to 11:59 pm. Midnight itself belongs to the following day.
The minimum and maximum temperatures and their times can therefore be derived from the behaviour of the continuous forecast rather than merely selecting the highest and lowest hourly rows.
Verification closes the loop
Eventually, every forecast becomes history. Once observations become available, LACO can compare what was forecast with what was actually observed. This is the verification side of the system.
Errors can be examined by individual model, forecast lead and time of day, among other characteristics.
The final Local Forecast can also be verified.
That’s important because sophisticated processing is not evidence by itself that a forecast is better. The forecast has to be compared with observations over many forecasts and different weather situations. The Forecast Verification page provide a way to examine that performance directly, while the About the Local Forecast page explains the methodology and its limitations in more detail.
And verification feeds back into the next analysis.
The complete process therefore forms a cycle: Forecast → Observe → Verify → Learn → Forecast again.
Rather than being separate from forecasting, verification is part of the forecasting system.
The four main parts of LACO
LACO-A — Analysis and learning
LACO-A analyses previous model forecasts against observations. It produces the learned information used for model weighting and the various forecast corrections.
LACO-F — Forecasting
LACO-F processes the available model guidance, applies the learned corrections and weights, and generates the consensus forecast. It also produces the continuous interpolation and daily forecast extremes used by the operational forecast.
LACO-N — Nowcasting
LACO-N monitors recent observations against the current forecast and determines whether a short range adjustment is warranted.
LACO-V — Verification
LACO-V provides the tools for examining how forecasts actually performed, including the individual models and LACO’s resulting forecasts.
Together they form the broader LACO system.
Making it operational
There is another part of LACO that has little to do with forecasting but is essential in practice: keeping the system running reliably.
Forecast runs are tracked, individual processing steps are logged, failures can be detected, system health can be monitored, and diagnostics provide visibility into what the forecasting processes are doing.
The system also maintains a distinction between its internal forecasting database and the information intended for public consumption.
Only the forecast and verification information required by the website is published to the public environment.
The entire process from incoming forecast information through processing and ultimately publication is designed to operate automatically.
How does LACO compare with professional systems?
The scientific ideas behind LACO aren’t unique.
Multi-model blending, statistical post-processing, bias correction, performance-based weighting and the use of observations to improve forecasts all have counterparts in professional meteorology.
The US National Weather Service’s National Blend of Models, for example, combines and statistically processes guidance from numerous forecasting systems.
Commercial forecasting provider meteoblue operates its Learning MultiModel, which combines numerous models and can use observations and recent model performance to improve forecasts for particular locations.
Systems operated by national meteorological agencies and commercial providers are vastly larger than LACO, often producing forecasts across large areas.
LACO applies that broad philosophy in a more narrowly focused way. Instead of trying to forecast everywhere, it concentrates deeply on one location: Ferny Grove.
It is this local specialisation that allows observations and historical model performance from Ferny Grove itself to directly inform the forecast.
Avoiding the black box
Transparency has been another important objective during development.
The final temperature displayed on the Local Forecast page is only the end result of a much larger process.
Internally, the original model forecasts, corrections, weights, individual model contributions, original LACO forecast, nowcast adjustments, processing diagnostics and verification information can all be examined.
That has been particularly useful when something unexpected occurs.
Instead of simply seeing an unusual forecast number, it is possible to trace how that number was produced.
That doesn’t make the forecast infallible, but it does make the system being able to be inspected.
And ultimately, the strongest evidence comes afterwards: compare the forecast with what actually happened.
From an experiment to a forecasting system
LACO started with a relatively straightforward question: Could forecasts from multiple global weather models be combined intelligently using their actual performance at Ferny Grove?
Answering that question eventually required much more than combining a few numbers.
The resulting system now includes automated forecast processing, local performance analysis, dynamic model weighting, learned forecast corrections, consensus generation, observation-based nowcasting, continuous interpolation, precise daily extreme calculation, verification, diagnostics, health monitoring and automated publication.
Many of those components are invisible when looking at the Local Forecast.
Most of that complexity is intentionally kept behind the forecast page, where visitors can simply use the resulting Local Forecast.
The objective at the end of all that processing remains quite simple: Use the available models, what they have taught us about forecasting Ferny Grove, and what is being observed here now to produce the best local forecast LACO can at the time.
The Local Forecast can be viewed on Local Forecast page. A more detailed and permanent description of the methodology, including the system’s limitations, is available on the About the Local Forecast page, while measured performance is published on the Forecast Verification page.
And after a long period of building, testing and refining the system behind it, LACO is now operational.
As always, please reply below or contact us if you have encounter any issues or have any thoughts.