Modelo matemático para predicción del clima usando Compressive sensing y Deep learning.

Abstract

The weather forecast is an important item in all industry, from extract the raw matery, its processing to get the final product. Thus, it's necessary control the weather knowing its trend. This document describe development a mathematic model using Compressive Sensing to predict the weather with the daily mean of variables like Temperature, Relative Humidity Wind Speed, Precipitation and Insolation from 2015 to 2020 employing POWER NASA data base, this technique was able to reduce time series length with 130 points until 26, 39 and 52 compressing 80%, 70% and 60% respectively utilizing a Identity matrix and its Discrete Cosine Transform to compact the information, the learning was development with Keras and Tensorflow in Python environment through a Multilayer Neural Network with Adam optimizer to faster extract features. Generally, the best accuracy was to Temperature, Relative Humidity, Wind Speed and Insolation but the precipitation got the worst statistics. It affirmed this model comply with expectative because 4 historic years is very low for the traditional methods, but Compressive Sensing strengthened to good forecast.