German country-wide renewable power generation from solar plus wind mined with an optimized data matching algorithm util
- PDF / 2,255,703 Bytes
- 43 Pages / 439.37 x 666.142 pts Page_size
- 48 Downloads / 200 Views
German country-wide renewable power generation from solar plus wind mined with an optimized data matching algorithm utilizing diverse variables David A. Wood1 Received: 27 March 2019 / Accepted: 7 July 2019 © Springer-Verlag GmbH Germany, part of Springer Nature 2019
Abstract Country-wide, hourly-averaged solar plus wind power generation (MW) data (8784 data records) published for Germany in 2016 is compiled to include ten influential variables related weather, ground-surface environmental and a specifically calculated day-head electricity price index. The transparent open box (TOB) learning network, a recently developed optimized nearest neighbour, data matching, prediction algorithm, accurately predicts MW and facilitates data mining for this historical dataset. The TOB analysis results in MW prediction outliers for about 1.5% of the data records. These outliers are revealed via TOB analysis to be related to uncommon conditions occurring on a few specific days typical over hourly sequences involving rapid change in weatherrelated conditions. Such outliers are readily identified and explained individually by the TOB algorithm’s data mining capabilities. A slightly filtered dataset (excluding 129 identified outliers) improves TOB’s prediction accuracy. The TOB algorithm facilitates accurate predictions and detailed evaluation over a range of historical temporal scales on a country-wide basis that could also be applied to regional spatial predictions. These attributes of the TOB method are conducive with it eventually being incorporated into forward-looking renewable forecasting frameworks. Keywords Country-level solar and wind power generation · Machine learning for transparent data mining · Combining weather/environmental/market variables · Prediction outlier analysis · Data filtering
Electronic supplementary material The online version of this article (https://doi.org/10.1007/s12667019-00347-x) contains supplementary material, which is available to authorized users.
B 1
David A. Wood [email protected] DWA Energy Limited, Lincoln, UK
123
D. A. Wood
1 Introduction Data-driven, machine learning forecasting of long-term and short-term spatial–temporal power generation from time-series data related solar [4, 22, 36, 58, 63, 69] and wind [11, 43, 60, 65] facilities involves a range of techniques and algorithms [21]. Their primary aim is to provide forward-looking forecasts of power generation on an hour-by-hour basis, either to assist in efficient balancing of the electricity grids at national, regional and local levels, or to predict potential plant outputs and revenue streams. Another important objective is to evaluate historical time-series datasets to verify the accuracy of the power generation predictions they are able to provide and data mine that information to better understand the causes of prediction errors. This work focuses on the latter prediction and data-mining objective rather than forwardlooking forecasts. The favoured techniques, typically applied to forecast and/or predict wind and solar da
Data Loading...
