Show simple item record

dc.contributor.advisorAndersson, Lars Jonas
dc.contributor.authorChong, Jae Meng
dc.contributor.authorKristiansen, Chris
dc.date.accessioned2021-04-23T07:36:09Z
dc.date.available2021-04-23T07:36:09Z
dc.date.issued2020
dc.identifier.urihttps://hdl.handle.net/11250/2739234
dc.description.abstractThis paper examines various short-term forecasting methods to forecast hourly wind energy production in Norway. Performance of forecasting methods were compared across different months, through different evaluation metrics, to analyze the uniformity and dependability of methods. More than a decade of hourly wind speed data spanning across 69 locations along the Norwegian coast were utilized in the study. Given the upcoming integration of EU and Nordic intraday electricity markets into the Cross- Border Intraday market (XBID), the study focuses on one-hour-ahead forecasting to be in alignment with the operations of intraday electricity market. With the final objective of predicting power production, a customized loss function, Power Curve Conversion Error with penalty, which takes into account both wind speed to power conversion and economic cost associated with over and under forecast, is included as part of the evaluation metrics to capture the true value of each model’s predictions. Forecasting methods undertaken consist of a mix of Statistical and Machine-Learning methods, with Naive forecasting used as the overall benchmark model. Other Statistical methods are ARIMA, and ARIMAX which includes the use of seasonal and time of day dummies. In terms of ML methods, Gradient Boosted Trees, Extremely Randomized Forest, and Neural Network are selected. Finally, a hybrid model of ARIMAX and Extremely Randomized Forest is also formulated. These methods are then evaluated on multiple evaluation metrics, namely: RMSE, MAE, Classification Accuracy, and the Power Curve Conversion Error with penalty. The general implication of the study reveals that accuracy of models are consistent with their required computational intensity, with ML models outperforming statistical methods in most situations. The findings also suggest the Hybrid model to be the most suitable forecasting method for one-hour-ahead forecast under almost all evaluation metrics employed. This conclusion holds true for wind power forecasting under different seasons of the year as well.en_US
dc.language.isoengen_US
dc.subjectbusiness analyticsen_US
dc.subjectEnergy, Natural Resources and the Environmenten_US
dc.titleA comparison of wind forecasting methods for Norwegian on-shore wind : a perspective into the nuances in wind speed to power conversion and the economic costs associated with wind forecast accuracyen_US
dc.typeMaster thesisen_US
dc.description.localcodenhhmasen_US


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record