My thanks to Per Strandberg for this update on his ENSO modelling effort. This is looking good, and is based on a neural network which uses lunar and solar data for its input.
There are two main drivers of ENSO. ENSO stands for El Niño Southern oscillation and is Earth’s most influential weather phenomena after seasonal changes. When ENSO changes it causes changes in currents and of temperatures in the tropical Pacific Ocean.
The most important ENSO driver is linked to variations in gravitational tidal forcing associated with Moon’s Perigee. Moon is in what is called Perigee, when the Moon is at it closest point during its elliptical orbit around Earth. This is also when the tidal force caused by the Moon is at its strongest.
The second most important forcing is linked to variation in solar activity.
I have identified these two basic underlining forces for ENSO by creating and utilizing an ANN I’ve built ANN stands for Artificial Neural Network and is a pattern recognition technique and is a form of Artificial Intelligence which is used in many different types of applications. This can be in everything from different types of forecasts, in robotics, data mining and for different kinds of identification.
This is a result I recently got from the ANN that I’m using.
The ENSO value up to and including October 2015 is from MEI ENSO index.
MEI is an acronym for the Multivariate ENSO Index which is an ENSO index complied by NOAA.
In the Calc 12 and 15 graph lines I use training time from 1979 up to the end of 2004.
CALC12 uses as testing period the time between 2005 and up to 2012 and make predictions from 2012 and up to 2020 which includes simulated solar and magnetic data for the this predictive period.
CALC15 use as testing period the time between 2005 and up to 2015 and make predictions from 2015 and up to 2020 which includes simulated solar and magnetic data for the this predictive period.
Here’s a close up picture over ENSO predictions up to 2020.The ENSO value is for the real MEI value up to and including October 2015.
Note: The weight values inside the ANN are based on a training period which is saved at the point where the testing part reaches its minimal variance error value.
The input values that are used in the neurons use no ENSO data. Only input values from the Perigee gravitational pulse values in the form of a vector size and its angel against the equator, plus Ap, Kp and solar wind data are used.
I think myself that that is very impressive.
I’m not finished with my work to get optimal ENSO predictions. There are still a number of improvements I’m going to make.
One thing I’m going to make is to reduce the overfitting problem with my ANN. The remedy should be simple. Currently I’m saving the weight values when the variance error reaches its minimum at the testing part. Instead I should save the weight values just a bit before the error value reach its minimum. By doing this I should be able to reduce extra statistical noise which is increasingly introduced near the minimal value before the error value start to diverge. In my case however this is not a big problem. The data I use in my ANN response and convergence quite quick and good.
One other thing I’m going to do is use several runs and creating many different predictions. By doing this I’m going to get an ensemble of predictions and by using their mean values I expect better and more robust prediction.