Using Keras as networks API. Keras version is 2.1.6 tensorflow version is 1.8.0

This version of code just use dense and/or conv2d to contruct networks.
There is a very serious problem of overfitting.

2018/9/26
The overfitting may be solved. This problem may be related to different historical field.

2018/11/2
Run ZC model freely. The results also look good.

2018/11/17 Processing historical data. SST from HadISST1: https://climexp.knmi.nl/select.cgi?id=someone@somewhere&field=hadisst1
In the equatorial Pacific, the depth of 20C isotherm is widely used to represent the thermocline depth.
Z20C form GODAS(Can not get data from GODAS directly):http://iridl.ldeo.columbia.edu/SOURCES/.NOAA/.NCEP/.EMC/.CMB/ .GODAS/.monthly/.BelowSeaLevel/.POT/Y/%28-20%29%2820%29RANGE/X/122.25/288.75/RANGE/%28Celsius_scale%29unitconvert/ Z/20/invertontogrid/datafiles.html
The monthly mean is calculated using the historical data from 198001 to 201809.
The wind stress anomaly data for ZC prediction: https://iridl.ldeo.columbia.edu/SOURCES/.FSU/index.html 196401 to 200202

2019/01/25 Add experiment of SPB, including plotting.

2019/02/25 Add experiment of seasonal circle as input. And for 'plot_nino34_prediction_for_month', add input way.

2019/03/22 Determined the final model (no BN, with SC input). Add some script for analysis.