Naslov
MACHINE LEARNING APPROACH TO FORECASTING DAY-AHEAD AND INTRADAY ELECTRICITY PRICES

Autori
Jurčević, Jura ; Vlah Jerić, Silvija ; Zoričić, Davor

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, prošireni sažetak, znanstveni

Izvornik
Book of Abstracts / - , 2022, 14-15

Skup
International Scientific Conference Technology, Innovation and Stability: New Directions in Finance

Mjesto i datum
Zagreb, Hrvatska, 5-6.05.2022

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Recenziran

Ključne riječi
random forest, external predictors, predictor relevance, benchmark out- performance

Sažetak
The growing challenge of integrating Renewable Energy Sources (RES) has given rise to aggregators as new market participants in the power markets and the power system. By aggregating consumers on one side and power generation units on the other aggregators can provide both supply- and demand-side flexibility in order to help balance the power system. The aggregated resources which the aggregators operate are referred to as Distributed Energy Resources (DERs). In fulfilling the role of flexibility providers one particular type of DER is exceptionally in the focus of various research papers – the battery storage. Storages play a vital role in the de- scribed setting due to the fact that they can be used to provide both aspects of flexi- bility. Battery storages have come to the fore with the advent of Li-ion batteries which are nowadays most widely used technology. Their power and energy characteristics coupled with high efficiency make them highly suitable for various operations which could promise multiple revenue sources for aggregators if operated properly. Howev- er, the high price remains their main disadvantage which makes accurate forecasting of electricity prices in various markets the biggest challenge regarding economic viability. This research is limited only to forecasting of the day-ahead and the intra- day electricity prices which is related to the possible arbitrage revenue of the battery storage. The aim of the research is twofold - to analyse accuracy of forecasts and the relevance of variables used as predictors. To this end a machine learning model based on random forest algorithm and a full set of variables used as predictors is compared to two benchmark models: one based on the same machine learning algorithm with no external predictors and a naïve model. The described methods are applied to the German power market for the two-year period (2017- 2018). We use the 2017 data as the initial training data set and estimate day- ahead and intraday prices up to one month ahead. We then repeat the forecasting process by rolling the training set forward one month while keeping its length fixed. Thus, we obtain a moving window of training data serving as a base for each of the subsequent forecasts. Our findings suggest that the model which includes full set of variables outperforms both analysed benchmarks in both analysed markets. The analysis shows that the outperformance can be traced to the relevance of the variables used as external predictors, such as load and wind generation forecasts. For the intraday electricity prices the imbalance price on the balancing power market is also highlighted. We also find evidence of intraday price as a useful predictor of day-ahead prices. As far as the accuracy of the forecasts is concerned our findings show an increasing RMSE with the forecast horizons. However, it can be noticed that RMSE of forecasts for the last three months of the 2018 is significantly higher across forecast horizons for all models and both sampling methods tested. This seems to be related to the missing data for certain periods in 2018 and possibly to the changes in the market bidding zones. Overall, our findings corroborate earlier research suggesting the relevance of external predictors as variables in the model and point out the benefits of using shorter forecast horizons in the case of the day-ahead and intraday electricity prices.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Ekonomija

POVEZANOST RADA