Forecasting DK1 Electricity Prices: Comparison and Performance Evaluation of ARIMAX and XGBoost.
Author
Prosad, Manik Chandra
Term
4. semester
Publication year
2024
Submitted on
2024-12-01
Pages
70
Abstract
At forudsige elpriser er blevet stadig vigtigere i et marked præget af hurtige udsving. I Danmarks DK1-prisområde påvirkes priserne blandt andet af mere vedvarende energi og eksterne geopolitiske forhold. Dette speciale sammenligner to metoder til at forudsige elspotpriser: ARIMAX, en statistisk tidsrække-model der bruger tidligere priser og eksterne forklarende variable, og XGBoost, en maskinlæringsmetode der kombinerer mange små beslutningstræer. Analysen bygger på timeserier af DK1-elspotpriser og energiproduktionsdata. Før ARIMAX-modellen anvendes, gennemføres standardtjek for stationaritet (om mønstre er stabile over tid), autokorrelation (sammenhæng mellem nuværende og tidligere værdier) og normalitet. Begge modeller finjusteres med udbredte metoder som AIC og BIC (modelkriterier), grid search og krydsvalidering for at vælge passende parametre. Derpå vurderes deres evne til både kortsigtede og langsigtede forudsigelser for at belyse behovet for robuste modeller, der kan håndtere elprisers volatilitet. Resultaterne viser, at ARIMAX forudsiger meget præcist på de data, modellen er tilpasset (in-sample), men har svagere evne til at generalisere, mens XGBoost giver mere præcise forudsigelser på nye, usete data.
Forecasting electricity prices has become increasingly important as markets experience rapid swings. In Denmark’s DK1 price area, prices are shaped by the growth of renewable energy and external geopolitical factors. This thesis compares two approaches to forecasting spot prices: ARIMAX, a statistical time-series model that uses past prices and external explanatory variables, and XGBoost, a machine learning method that combines many small decision trees. The study uses hourly DK1 spot prices together with energy production variables. Before applying ARIMAX, the data are checked for stationarity (stable patterns over time), autocorrelation (links between current and past values), and normality. Both models are tuned with common techniques such as AIC and BIC (model selection criteria), grid search, and cross-validation to choose suitable parameters. Their performance is then evaluated for both short-term and long-term forecasts, highlighting the need for robust models that can capture electricity price volatility. The findings show that ARIMAX fits the training data very well (in-sample) but generalizes less well, whereas XGBoost delivers more accurate predictions on unseen data.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents