Use this URL to cite or link to this record in EThOS:
Title: Forecasting wind power for the day-ahead market using numerical weather prediction models and computational intelligence techniques
Author: Martínez-Arellano, G.
ISNI:       0000 0004 5365 3873
Awarding Body: Nottingham Trent University
Current Institution: Nottingham Trent University
Date of Award: 2015
Availability of Full Text:
Access from EThOS:
Access from Institution:
Wind power forecasting is essential for the integration of large amounts of wind power into the electric grid, especially during large rapid changes of wind generation. These changes, known as ramp events, may cause instability in the power grid. Therefore, detailed information of future ramp events could potentially improve the backup allocation process during the Day Ahead (DA) market (12 to 36 hours before the actual operation), allowing the reduction of resources needed, costs and environmental impact. It is well established in the literature that meteorological models are necessary when forecasting more than six hours into the future. Most state-of-the-art forecasting tools use a combination of Numerical Weather Prediction (NWP) forecasts and observations to estimate the power output of a single wind turbine or a whole wind farm. Although NWP systems can model meteorological processes that are related to large changes in wind power, these might be misplaced i.e. in the wrong physical position. A standard way to quantify such errors is by the use of NWP ensembles. However, these are computationally expensive. Here, an alternative is to use spatial fields, which are used to explore different numerical grid points to quantify variability. This strategy can achieve comparable results to typical numerical ensembles, which makes it a potential candidate for ramp characterisation. A major disadvantage of most ramp events studies is that they are based on a binary classification, which specifies a percentage of change in power within a defined time window. This may produce artifacts, as ramp detection tools might miss potential changes due to errors in the forecasts. Moreover, a change just below the threshold could be equally damaging as a change that meets the definition. The novel contribution of this project is the application of computational intelligence techniques for wind power forecasting and ramp event characterisation. To achieve this, two stages are required. In the first stage, Genetic Programming (GP) is used to generate an ensemble of wind power forecasts based on the idea of spatial fields. This in its own is an important contribution as the approach will allow the development of computationally cheap wind speed-to-power conversion models, without making any assumptions of their shape or properties. In the second stage, wind power forecasts are converted into a set of filtered signals in order to study ramp events at different time scales. These signals, when applied to a set of Fuzzy Logic rules, indicate the probabilities of a ramp event happening, avoiding the binary classification, which is another important contribution of this work. The observation data used for this investigation was obtained from a real wind park in Galicia, Spain and some observation points in Illinois, USA. The numerical data was obtained by running locally a Mesoscale model. Experiments showed that the accuracy of wind power forecasts obtained using GP as a downscaling/conversion method are comparable to traditional forecasting tools as it is able to achieve an 87% of accuracy. At the same time the computational effort was significantly reduced. The novel ramp detection approach that is introduced here, is able to outperform a basic binary-based detection algorithm. In addition, the fuzzy rules can provide a probability of other events happening; events that might not meet the crisp definition. Using colour maps, which are easier to interpret by human non-experts, it is possible to show how an event is developing in different time windows. Finally, it is shown how neighbouring points can help modelling events that might not be detected using only the closest point of the grid. Having a detailed characterisation of future ramp events can help grid operators to make more informed decisions on the scheduling of back-up units needed and hence to potentially reduce costs and the environmental impact.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available