Napovedovanje cene električne energije na trgu za dan vnaprej s časovnimi serijami

MALI, BORUT (author), Pantoš, Miloš (mentor)

Trgovanje z električno energijo na trgu za dan vnaprej deluje na principu dražb, kjer se križata agregirani krivulji ponudb za nakup in prodajo. Te so dan prej podane s strani trgovcev in dobaviteljev, ki za ceno na dan trgovanja ne vedo, saj se ta izoblikuje šele po prejemu vseh ponudb. V času oddaje le-teh tako nastane negotovost, ki jo lahko vsak izmed sodelujočih izkoristi, če le ima orodja, s katerimi si pomaga pri boljšem predvidevanju cene. S tem bi si ustvaril ključno prednost pred ostalimi in lažje oddal dobičkonosno ponudbo. Namen naloge je bil opisati metode kratkoročnega (za dan vnaprej z urnim korakom) napovedovanja cene električne energije in jih preizkusiti na primeru trga za dan vnaprej v Sloveniji. Najprej smo si podrobneje ogledali znane informacije o trgu in prišli do podatkov o cenah električne energije v določenem časovnem obdobju. Po njihovi preučitvi smo izbrali najbolj primerne metode napovedovanja. Lastnosti in delovanja metod povprečne vrednosti, naključnega hoda, eksponentnega glajenja in modela ARIMA so bila vsa temeljito predstavljena in tako smo si ustvarili podlago za nadaljnje delo, napovedovanje cene. Tu smo vzeli dva dvotedenska intervala iz različnih obdobij, ki sta služila kot nabor podatkov za napoved cen dni za intervaloma, kjer je bil en izmed njiju stabilen, drugi pa nestabilen. S primerjavo cenilk napak napovedi RMSE in MAPE smo prišli do zaključkov, da je sicer najbolj kompleksen med modeli, sezonski model ARIMA, opravil najboljšo napoved, vendar so si bili nekateri modeli pri rezultatih precej blizu, prav tako smo opazili, da so bili posamezni dnevi, predvsem vikendi, problematični zaradi popačenja stalnega trenda delavnikov. Na podlagi teh ugotovitev smo se odločili primerjavo med modeli razširiti na intervale skozi celo leto in poizkusiti z razdelitvijo le-teh na delavnike in vikende ter jih posebej vključiti pri samih napovedih. Končni rezultat je še podkrepil zaključke iz prvega dela. Sezonski model ARIMA se je tu izkazal za daleč najboljši model, še najbližje mu je prišel en najbolj intuitivnih in enostavnih modelov, model naključnega hoda, kjer so bile napovedi cen enake vrednostim iz prejšnjega dneva. Prav tako se izkaže, da se razdelitev intervalov, kjer posebej obravnavamo delavnike in vikende, običajno izplača in prinese boljše rezultate.

• Language: Slovenian
• Keywords: trg električne energije za dan vnaprej, cena električne energije, kratkoročno napovedovanje, eksponentno glajenje, model ARIMA
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2020
• PID: 20.500.12556/RUL-117059
• Publication date in RUL: 22.06.2020

Secondary language

• Language: English
• Title: Electric energy price forecasting in day-ahead market with time series

Abstract

Electric energy trading on the day-ahead market operates on the principle of auctions, where intersect the curves of bids and offers to sell. Bids and offers to sell are given the day before by traders and suppliers who do not know the price since the prices are formed only after receiving all offers. At the time of submitting them, there is uncertainty, of which each of the participants can take advantage of if they have the tools to help them better predict the price. These tools can help them create a key advantage over others and make it easier to submit a profitable offer. The purpose of work was to describe the methods of short-term forecasting (for the day-ahead hourly loads) of the price of electric energy and to test these methods on the example of the day-ahead market in Slovenia. First, we inspected the known market information and came up with data on electric energy prices over a period. After examining them, we selected the most appropriate forecasting methods. We created a basis for price forecasting and further work and by thoroughly presenting the properties and functions of the methods of average value, random walk, exponential smoothing and the ARIMA model. We took two two-week intervals from different periods that served as a data set for forecasting the price on the days after the intervals, where one of them was stable and the other unstable. By comparing forecasts with error estimators RMSE and MAPE, we concluded that the most complex among the models, the seasonal model ARIMA, gave the best prediction, but some other models also had close results. We also noticed that some days, especially weekends, were problematic because of the distortion of the constant trend of working days. Based on these findings, we expanded the comparison between the models to a comparison of intervals throughout the year and to split them to weekdays and weekends and include them separately in the forecasts. The results further supported the conclusions from first part. ARIMA seasonal model proved to be by far the best model, the closest to him came one of the most straightforward and most intuitive models, the random walk model, where price forecasts were equal to the values on the day-ahead. It also turned out that the distribution of intervals, when we treat the weekdays and weekends separately, pays off and brings better results.

• Keywords: day-ahead electric energy market, electric energy price, short-term forecasting, exponential smoothing, ARIMA model