Short-term forecasting of cloud movement based on sky images

V diplomski nalogi je opisan postopek razvijanja algoritmov, ki napovedujejo prekritost sonca z oblaki. V času ogromnega naraščanja števila elektrarn na omrežju, ki jih poganjajo obnovljivi viri, postaja izjemno pomembno tudi kratkoročno napovedovanje proizvodnje električne energije. Prav tako pa je kratkoročno napovedovanje temelj za trgovanje z energenti, predvsem na krajši rok. V nalogi iščemo rešitev specifično za sončne elektrarne, ki za omrežje, v večjem številu, predstavljajo veliko obremenitev. Program pa najde svoje mesto tudi pri trgovcih z energenti. Naloga detajlno opisuje vsak del algoritma. Prvo opisujemo postopek kalibracije kamere oz. pridobivanje parametrov za namen preslikave iz polkrogle, ki jo predstavlja nebo na ravno površino. Naslednji korak pri izdelavi programa je bilo izkustveno določanje barvnih kriterijev, pod katere spadajo oblaki in Sonce za namen programskega zaznavanja njihovih centrov. Preko zaznanih centrov oblakov in Sonca je nato nastal algoritem za sledenje objektom na nebu, preko njihovih koordinat centrov. Iz dveh zaporednih slik lahko program namreč razbere, kateri centri pripadajo istemu zaznanemu objektu, na podlagi Evklidove razdalje. Ko program zajame šest zaporednih posnetkov zaslona iz videoposnetka, program prične z napovedjo pozicij oblakov in Sonca. Medtem, ko program izvaja branje neba se koordinate centrov vseh zaznanih oblakov in Sonca shranjujejo (vsak šesti shranjeni nabor koordinat se z novim posnetkom zaslona osveži), z namenom, da te podatke naprej poda algoritmu, ki je najbolj primeren za iskanje krivulje, ki opisuje potovanje centrov oblaka ali Sonca. Opisana je tudi primerjava metod za napovedovanje, med drugimi: ARIMA, eksponentno glajenje, polinomska regresija in metoda, ki temelji na povprečju razdalj med vsemi zgodovinskimi nabori koordinat. Za najboljšo, se je po prilagoditvi njenih parametrov izkazala metoda z eksponentnim glajenjem, ki je v večini primerov lahko napovedala koordinato na piksel natančno za korak naprej (ena minuta). Program po izračunih napovedi, na posnetek zaslona nariše Sonce na napovedano koordinato in šele nato nariše še oblake na napovedane pozicije. Sonce preslika z neko barvo, ki pa se nujno razlikuje od barv, s katerimi program preslika oblake. Program na koncu preveri koliko pikslov v barvah Sonca je prisotnih in na podlagi tega izračuna prekritost. Na koncu pa program preveri še barvna razmerja pikslov oblakov, ki prekrivajo Sonce. Glede na barvna razmerja program razvrsti piksle v razred, ki procentualno opisuje ali gre za temen, delno temen ali svetel del oblaka. Prozornost oblaka pa je povprečje vseh teh vrednosti. Opisan je tudi postopek izdelave uporabniškega vmesnika.This thesis describes the development of algorithms, which can forecast the coverage of the Sun by the clouds. In times of increased renewable source installements on the grid, the importance of short term forecasting is becoming more relevant by the day, for production of electricity. It also represents the basis on which short term energy traders work on. In this diploma tehsis we are trying to find a solution spcifically for solar plants, which are capable of stressing the power grid in case of weather changes. Likewise, the program could also be useful to energy traders. This literature describes every step of the development. Firstly, camera calibration or gathering of extrinsic parameters of the camera is described, which is done to help the program remap the objects on the sky from a semicircle onto a flat surface. Next step in the development of the program was to set thresholds of HSV values for cloud and sun centre detecting purposes. With detected centres the development could then be focused on a tracking algortihms. With two consecutive screenshots of the sky, the program is able to guess which centres on both screenshots belong to the same detected object, based on Euclidian formula for distance. When the program detects and saves six consecutive collections of centres it can then start with the position forecasting of the sun and the clouds. While program is reading the video of the sky, all the centres of clouds and the sun are being recorded (every sixth list of coordinates is being updated upon arrival of new frames of the sky), for forecasting method data feeding purposes. Forecasting method was chosen based on the best curve fitting capability. Comparison of a few forecasting methods is also described, among those are: ARIMA, exponential smoothing, polynomial regression and a method which is based on the average distance between the collection of corresponnding coordinates. Best among those methods proved to be Exponential Smoothing (DES), which was able to predict the coordinates most accurately based on a one-step ahead (1 minute) test. After all the forecasting, the program draws a sun and after that it draws all the clouds onto a screenshot of the sky. Sun is drawn with a color that must be different from the colors with which the clouds are being drawn with. Program then checks how many pixels are left with the same color as the sun, to then calculate the coverage by the clouds. At the end program also iterates through every pixel that represents the clouds over the sun and calculates their color ratios. Based on those ratios it puts every pixel into one of three classes that descirbe the transparency of the cloud. One of the classes describe dark pixel, another one describes semi dark and the last one represents the light ones.Transparency is the average of all those pixels. Front-end development of the program is also described towards the end of the paper