Predictive analytics for wind power forecasting

  • Rosemary Tawn

Student thesis: Doctoral Thesis

Abstract

At the most basic level, forecasts are needed when decisions must be made in the present but are based on future conditions. There are many uses for forecasts within the energy industry, and in particular in relation to renewable generation types where future generation is uncertain and depends on weather conditions, be that wind, solar irradiation, cloud cover or sea conditions. Thus, forecasts are an essential part of the running of the electricity system, from day-to-day scheduling and trading decisions to long-term system planning. This thesis selects and aims to address three current problems in the practical implementation of wind power forecasting methods: firstly, the affect of several different occurrences of missing data in the forecasting process and how this can be mitigated; secondly, the difficulty in producing a forecast that accurately predicts ramps; and thirdly, the need for skilful site- and task-specific forecasts weeks ahead to inform maintenance decisions. Missing data is the result of incomplete datasets, data latencies and new sites lacking historic power data. Missing historic values affect the ability of statistical models to learn site characteristics and _t an accurate model while data latencies mean forecast inputs are not always available when new forecasts are issued. Several different mitigation methods for each occurrence of missing data are explored via case studies for a Vector Autoregressive model implementation. Complex relationship between wind speed and power, sudden ramps in power and imperfect models can reduce the skill of individual forecast models. Improvements through combination of several different forecasting models is explored and a forecast combination method that explicitly incorporates forecasts of ramp rate proposed to improve power forecasts around times of ramps. A lack of suitably tailored forecasts for a given decision also reduces the likelihood of uptake of a new model or data source for certain applications. A novel job-specific index of `number of useful hours' is forecast for subseasonal-to-seasonal timescales and its calibration and usefulness for the case of crane hire for maintenance decisions is assessed. This work has also produced guidance and recommendations for the implementation of a very short-term statistical forecasting system for Natural Power Consultants.
Date of Award16 May 2022
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorJethro Browell (Supervisor) & David McMillan (Supervisor)

Cite this

'