Probabilistic forecasting of wind farm output
Files
(Published version)
Date
2013
Authors
Agrawal, M.
Huang, J.
Boland, J.W.
Editors
Piantadosi, J.
Anderssen, R.S.
Boland, J.
Anderssen, R.S.
Boland, J.
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Conference paper
Citation
MODSIM2013, 20th International Congress on Modelling and Simulation, 2013 / Piantadosi, J., Anderssen, R.S., Boland, J. (ed./s), pp.1475-1481
Statement of Responsibility
Conference Name
20th International Congress on Modelling and Simulation (1 Dec 2013 - 6 Dec 2013 : Adelaide)
Abstract
We have previously developed a short time scale forecasting tool for solar radiation [Huang et al., 2013], and also a mechanism for estimating the conditional variance of wind farm output at particular time scales using data at a higher frequency, see [Agrawal et al., 2010, 2013]. The term conditional variance reflects the idea that the variance is changing with time (heteroscedastic) rather than being homogeneous in time (homoscedastic). In this paper, we will describe the application of the solar radiation forecasting tool (which is referred as CARDS model) to wind farm output to obtain forecasts of the level of output on two specific time scales, five minute and half hour. These are the time scales at which the Australian Electricity Market operates. Hence for efficient operation of the electricity grid, it is crucial to have knowledge of forecast of wind energy 5 minute ahead as well as half an hour ahead together with appropriate error bounds. This is exactly the aim of this paper which we achieve using the techniques developed in [Huang et al., 2013] and [Agrawal et al., 2013].
An interesting outcome is that 93:5% of the data coverage is contained in the interval Ft+1± r σ t+1 for the 30 minute ahead forecast, while for 5 minute ahead forecast 94.2% of the data coverage is contained in the constructed interval Ft+1± r σ t+1 with r = 0.65. In other words, a lower rate of conditional standard deviation suffices to contain most of the observations at the 5 minute time scale.
In more explicit terms, knowledge of {Fτ}tτ =t0 , the history of the wind energy output series up to time t allows us to forecast the level of output at time t + 1, this we achieve using the forecasting tool developed in [Huang et al., 2013]. We then estimate the conditional variance at time t using the techniques developed in [Agrawal et al., 2013]. To facilitate this, we did have high frequency data available at the 10 second time scale. Once we obtained a time series of conditional standard deviation, {στ}tτ = t0 , up to the current time step t, we reinvoke CARDS model to obtain a forecast of the conditional standard deviation at time t + 1, that is, to get σ t+1. Upper and lower bounds of the forecasted wind farm output are thus constructed as Ft+1± r σ t+1 where r is a positive real number.
This allows us to not only have a forecast of the output but to also put error bounds on that forecast. This type of information is crucial for efficient operation of the electricity grid. This is particularly true in South Australia where wind farms provided 26 % of the electricity generation in the financial year ending June 2012.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2013 The Modelling and Simulation Society of Australia and New Zealand