Back to the programme printer.gif Print




Delegates are invited to meet and discuss with the poster presenters in this topic directly after the session 'Forecasting: Maximizing grid deliverability and leading your business processes to profitability' taking place on Thursday, 13 March 2014 at 11:15-12:45. The meet-the-authors will take place in the poster area.

Maria Bullido Garcia Meteodyn, France
Co-authors:

(1) Meteodyn, NANTES, France

Printer friendly version: printer.gif Print

Abstract

How to mix CFD down-scaling and online measurements for short-term wind power forecasting: an artificial neural network application

Introduction

One-day ahead Power Forecasting is more and more required on the energy markets, and its accuracy is more and more crucial since it affects the net income of operators. This paper addresses a new operational forecast system, and its application on some wind parks in different countries. In the whole paper, time step is 15 minutes and “power” is the active instant power.

Approach

The forecasting system is based on three approaches: 1/ a Weather Numerical Prediction (WNP), including CFD micro down scaling. 2/ Fresh measurements on site are another source of data for the forecasting system. The main idea is to complete the WNP with persistence model. 3/ Then, an Artificial Neural Network (ANN) is trained on old data in order to mix these two approaches (for intermediate horizons) and to correct systematic errors and biases.

Main body of abstract

The main work realized is the ANN implementation and its training. An ANN is a black box with some inputs, some outputs, and an “engine”. The output is here the instant output power of a windfarm. There are three kinds of inputs: some meteorological inputs, given by the WNP, such as wind speed, wind direction, air density, atmosphere stability, horizon, hour; a second kind of inputs are non linear transforms of these first inputs, mainly a forecasted power obtain by physical approach through a CFD micro-scaling. This transform deals with local topography and land use, turbines power curves (including cut-out), and wake effects. A last group of inputs are fresh measurements on site, and more specifically measured output power and its own horizon. Lot of questions arises about the “engine”. This black box has to be trained with past data, saying some examples where all the input parameters values as well as the output value (historical data) are available. Special attention is paid on time definitions, which become a matrix: for a single forecasted time, it exist several meteorological horizons and several fresh measurement horizons. This matrix vision of a single time step, with a single actual output value, makes the number of data points exploded, but with a high level of correlation. Thus, the problem of specifying training/test/validation sets –which is a common issue in machine learning– is solved by selecting data on random, but by keeping each day intact.

Conclusion

This approach is successfully used on several parks in different countries. It appears that with about one year of historical data, the use of “direct” NWP data as inputs does not bring any improvement: the “minimum” set of inputs, using only 4 data, is the CFD-based forecast power, the fresh measurements, its horizon, and the hour (to deal with day/night). Normalized Root Mean Squared Error highly depends on the wind farm and data quality. Generally speaking, this forecast system is 2,5 RMSE points better than simple persistence (from ~13% to ~10.5%) for the horizon between 0 and 4 hours.


Learning objectives
The audience will understand the global architecture of the Artificial Neural Network implemented. Special focus will emphasis the problems of data volume and correlation, and set selection. Last, some considerations about how to evaluate a Power Forecasting System will be addressed.