Background to why I built this ANN program to analyze ENSO

With a background as an engineer and physicist, I became interested in the climate debate, because I like so many others, reacted to the increasingly extreme claims of a coming climate catastrophe caused by human activity. This led me to unconditionally dive deeply into data and information about what causes changes in climate, whether it is natural or man-made. I found that changes in the sun’s electromagnetic activity and changes in ocean currents play a critical role as drivers of climate change.

I had already experimented earlier with Artificial Neural Network (ANN) and was very familiar how ANN worked. So I thought ANN programs and climate data analysis was something that in this case could fit well together. This started for me more as a hobby and I experimented with building an ANN program. I locate various inputs data that could be connected the climate. What I first started analyzing was the global temperature variation measured from the satellite which could be paired with some of the input I have loaded as inputs. I knew that ENSO affects the temperature, so I put ENSO’s impact as a norm, to 100%.

Then I tried to estimate whether there were correlations between the input and the derivative of the temperature.

In doing so, I began to uncover interesting results. I measured each input parameters impact on the derivative of temperature with respect to the impact that ENSO derivative has on the temperature.
These are the results I then got: SST 110%, ENSO 100%, LOD 68.5%, Solar wind speed 49.5%, Kp index 27.4%, Solar wind temperature 26.3%, AMO 22.4, Ap index 13.3%, Solar wind density 9.7%.

It was shown that SST which are variations of the global ocean temperature has the greatest impact on global temperature which is something that seems logical. After ENSO comes LOD that are small variations in the Earth’s rotation, which may not seem logical. In fact, it has a natural explanation, because LOD is correlated to ENSO. LOD signal is thus a ENSO signal. Then this is followed the solar wind and Earth Magnetism values. The reason I included LOD in the calculations was that I looked at this video from a lecture by Dr. Vincent Courtillot where he said that there are correlations between LOD and the global temperature.

These data and this information caused me to take a closer look at ENSO which is the same as the El Niño index and I was able to discover the underlining drivers of ENSO variability. When ENSO and LOD have strong correlation, so should this correlation be due to something or some type of force. Closest at hand should be the tidal effect, I thought. To calculate the tidal energy’s gravitational influence, I needed exact data on the Moon and Sun position and distance. I found a computer program which is used by amateur astronomers, where I could print the data to a file. I got a real breakthrough when I discovered that I got good correlation between the derivative value of ENSO and tidal gravity changes at the time Moon’s Perigee. Perigee is when the Moon is closest to Earth which happens once every 27.5 days.

What is the reason I use an ANN program for studying ENSO?

To answer this without going into detail about how ANN program works I will here try to put this in context. Besides, there are plenty of details of ANN both on the Internet and in books.
But first, what are the methods to study the connection between input data and output data?

A common method is by linear regression, i.e. comparing the input with the output. Another way is by some form of frequency analysis. Typically, with some variation of FFT analysis.
A third way is by creating a flow chart-related control model. The problem with this method is that a working ENSO model is too complex to build and also the connection between in and output signals has significant time delays.

Climatologists use to make ENSO forecasts through some variation of GCM models which use differential equations that follow thermodynamic and flowing dynamic equations by doing massive calculations in small cubes and small time increments. The current ENSO models’ forecasting ability is poor. Only a few months. The main reason for this is that they ignore the underlying driving forces linked to strong tidal pulses and variations in solar activity.

But how do you make the best analysis of ENSO that can provide reliable ENSO forecasts when ENSO is a nonlinear system, where the output follows the input after considerably time delays, the input signals do not follow distinct frequencies, where one does not know initially which input signals have no connection with output, but where relationships between in and output signals follow stable physical laws?
The answer to this is to use an ANN program. ANN methodology works perfectly for this.
Simplified, one can say that the ANN based on a network of asymptotic transfer functions where the relationships between the signals and transfer functions are controlled by weight values that are modified recursively.

I started my ANN analysis by first studying individual input parameters to determine if the input parameters had statistical relationship with ENSO. The input parameters which had no relation to ENSO then were excluded as inputs in my ANN calculations in order to reduce statistical noise. I had as my hypothesis that ENSO variability is largely caused by forces outside internal weather fluctuations. Therefore, I excluded parameters PDO, SOI and SST from being included in my ANN calculations for ENSO.

I use monthly based data in my calculations. From the network, I get for every month an estimated output value. By comparing this value with the actual output value I get an error value. The program then calculates the variance based on these monthly error values. I use an algorithm and change the weights in the network in an iterative way which reduces the variance value of the error over time. The program also uses a test period using the same weight values used in the training part, but it calculates its own variance value which is different from training period’s variance value. If then both training period and the test period variance values decrease, I then know that the input parameter used in calculation has a relationship with ENSO.

By working with input parameters that have an impact on ENSO and optimize weight values, it then becomes possible to create through asymptotic transfer functions a model using the ANN program and to do so by solve unknown physical relationships. This is something which is done with ANN program, so to speak, recursively.

But to get this to work requires an understanding of how ANN program works and a lot of creative experimentations.

I think that the ANN method is underestimated, especially in control theory technology.

Comments are closed.