Neural Network forecast of geopotential height
Frame Controls: This product is experimental and should be applied as such. The algorithm will be modified from time to time, so anticipate changes.
Predictors are selected from the first 40 time extended PCs of 30 to 87.5N geopotential height anomalies.

Treat results like you would a 7-day centered moving average. Synoptic events are largely removed.

The constructed analog method is appled to reconstruct the data fields from the PCs, but the neural network predicts the PCs.
Red and blue contours represent positive and negative 300 hPa height anomalies, with a 20m interval.
Magenta and cyan contours represent the constructed analog of the absolute value of the geopotential height
minus the absolute value of the constructed analog of the geopotential height. PCs are retained as predictors only when the skill
of the network at a given lead time exceeds 0.3 relative to climatology. Skill is assessed by -(MSEforecast-MSEclim)/MSEclim.
Since the number of excluded PCs increases with lead time, the fraction of the variance explained by the retained PCs must decline.
The PCs are calculated using data centered on the period 45 days before to 45 days after the present day of year, and the skill chart
is updated nightly on the new PCs.

Loop Mode:


Animate Frames:


Dwell First:

dec start dwell
inc start dwell

Dwell Last:

decrease end dwell increase end dwell

Adjust Speed:

Slower Faster

Advance One:

Backward Forward
Frame No:
Processed OLR data