AAU Student Projects is unavailable between June 15th 1.30pm and 17th 1.30pm due to planned system maintenance. The projects cannot be downloaded during this period.
AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


Forecasting Potential Wind Power Production in the Faroe Islands: Predictive Analytics for Húsahagi, Flatnahagi and Gellingarklettur

Authors

;

Term

4. semester

Publication year

2026

Submitted on

Abstract

The shift to renewable electricity makes wind power increasingly central, but its variability poses operational challenges, especially in small, isolated grids without interconnection. This thesis examines how accurately short-term models can forecast potential wind power production for the Tórshavn wind-farm cluster. Using 12 months of operational data from Húsahagi, Flatnahagi and Gellingarklettur together with meteorological observations, the study evaluates ten model variants spanning baselines and four trained model classes (including LSTM, XGBoost, ARIMAX and Chronos). Models are assessed with expanding-window cross-validation and a two-month holdout, covering point accuracy, probabilistic calibration, statistical significance and operational impact. Because historical weather observations are used as a proxy for numerical weather prediction and aligned with the forecast target, the reported scores represent an upper bound on real-world performance. The LSTM delivers the strongest holdout point forecasts—with a cluster-mean Skill Score of +0.562 and 12.14% nMAE—and is the only model with substantially positive Skill Score (above +0.47) at all three sites; however, on cross-validation it is not statistically distinguishable from XGBoost or an empirical power-curve baseline, indicating that much of the measured skill comes from the wind-to-power relationship. ARIMAX achieves near-nominal 80% prediction-interval coverage (78.6% on holdout), making it the most credible probabilistic component, while full-parameter fine-tuning of Chronos yields no significant gain and worsens holdout calibration. Translated into operations under explicit reserve-sizing assumptions and perfect weather foresight, the LSTM forecast reduces estimated diesel-reserve cost by roughly 34 million DKK per year relative to a persistence baseline. Practically, the thesis proposes a deployment architecture that combines LSTM, XGBoost, ARIMAX and Chronos in complementary roles, and methodologically it advances time-series forecast evaluation in small, isolated, renewable-dominated power systems—providing a proof-of-concept foundation for forecast-based decision support at SEV and comparable utilities within clearly delimited conditions.

Overgangen til vedvarende el gør vindkraft central, men dens variabilitet skaber driftsmæssige udfordringer, især i små, isolerede net uden sammenkobling. Dette speciale undersøger, hvor præcist korttidsmodeller kan forudsige potentiel vindkraftproduktion for Tórshavn-vindpark-klyngen. Med udgangspunkt i 12 måneders driftsdata fra Húsahagi, Flatnahagi og Gellingarklettur samt meteorologiske observationer vurderes ti modelvarianter, fra simple baselines til fire trænede modelklasser (herunder LSTM, XGBoost, ARIMAX og Chronos). Modellerne evalueres med udvidende-vindue cross-validation og et to-måneders holdout, med fokus på punktnøjagtighed, probabilistisk kalibrering, statistisk signifikans og driftsmæssig effekt. Da historiske vejrobservationer anvendes som proxy for numeriske vejrudsigter og tidsmæssigt er justeret til prognosemålet, udgør de rapporterede resultater en øvre grænse for operativ ydeevne. LSTM giver den stærkeste holdout-punktforudsigelse med en klyngegennemsnitlig skill score på +0,562 og 12,14% nMAE og er den eneste model med markant positiv skill score (over +0,47) på alle tre parker; i cross-validation er den dog ikke statistisk adskillelig fra XGBoost eller en empirisk effektkurve-baseline, hvilket indikerer, at en stor del af den målte evne stammer fra vind-til-effekt-relationen. ARIMAX opnår næsten nominel 80% dækningsgrad for prognoseintervaller (78,6% på holdout) og er dermed den mest troværdige probabilistiske komponent, mens fuld-parameter finjustering af Chronos ikke giver statistisk signifikant gevinst og forværrer kalibreringen. Oversat til drift, og under eksplicitte antagelser om reservesizing og perfekt viden om vejret, reducerer LSTM-prognosen den estimerede dieselreserve-omkostning med cirka 34 mio. DKK om året i forhold til en persistens-baseline. Specialet bidrager praktisk med et forslag til implementeringsarkitektur, der kombinerer LSTM, XGBoost, ARIMAX og Chronos i komplementære roller, og metodisk med at fremme evaluering af tidsrækkeprognoser i små, isolerede, vedvarende dominerede elsystemer—som et proof-of-concept for beslutningsstøtte hos SEV og lignende forsyninger under klart afgrænsede betingelser.

[This apstract has been generated with the help of AI directly from the project full text]