Loading...

Improving short-range wind intensity prediction based on multimodel meteorological ensemble forecasts and Genetic Programming

As part of the “HPC4E – High Performance Computing for Energy” project, the National Laboratory for Scientific Computing and the University of São Paulo in Brazil are working on the improvement of wind downscaling techniques based on genetic programming (GP). The ultimate aim is to provide forecasting tools to support the transition to a reliable, sustainable and competitive energy system based on wind power, a common goal between the EU and the Brazilian Government. 

GP is a robust optimization algorithm based on Darwin's theory of evolution by natural selection that evolves a population of computer programs, usually expressed as syntax trees. GP exhibits inherent parallelism and can potentially capture non-linear phenomena. One of its major advantages over other techniques is its capability of evolving human-interpretable solutions of potentially unbounded complexity. The main drawback of GP is the high computational cost, which would preclude its operational implementation for wind forecasting. However, as part of the HPC4e project, a fast version of the GP application to wind forecast has been developed for use with CPU accelerators (GPUs). The evolutionary tasks were divided between the CPU and an accelerator. The CPU manages the evolutionary process and performs the selection, reproduction and replacement steps in a parallel way, while an accelerator (GPU) is responsible for evaluating the solutions and finding the best solution at each iteration. The final result was a code about 40 to 60 times faster than the purely CPU version.

The parallel implementation of GP has been applied to post-process short-range ensemble wind intensity forecasting (10m above the surface) in NB. The input attributes are: date and time information, wind intensity predicted by 30 operational numerical weather predictions and climatic indices that influence the wind field in NE Brazil (indices related to sea surface temperature, in general) and horizontal temperature gradients. Wind observations came from the National Institute of Meteorology in Brazil. The training set consists of the period of January 2012 to December 2014 while the test set refers to the period of January to December 2015 (4 times per day). 

Figure 1 shows the time series of wind intensity predicted at Fortaleza for the period of January to December 2015 as an example of GP’s ability to predict short range wind intensities at different times of the day. The blue and red marks indicate, respectively, the correctly and incorrectly instances classified by GP. The confusion matrix (error) is given by Table 1 for different wind speed categories. The accuracy of the test case is of the order of 75%. 
 

Figure 1 Wind intensity forecasts (m/s) for the 00, 06,12 and 18 UTC at Fortaleza (NEB) from January 01, 2015 to December 31, 2015. Blue (red) dots represent correct (wrong) predictions.

Figure 1 Wind intensity forecasts (m/s) for the 00, 06,12 and 18 UTC at Fortaleza (NEB) from January 01, 2015 to December 31, 2015. Blue (red) dots represent correct (wrong) predictions.

 


Table 1:  Error (Confusion) matrix for the evaluation of the Genetic Programming wind intensity prediction by categories of wind speed (in m/s).

Table 1:  Error (Confusion) matrix for the evaluation of the Genetic Programming wind intensity prediction by categories of wind speed (in m/s).

 

In summary, GP obtained a higher performance relative to traditional statistical methods and it is computationally very efficient for operational application.  Other possible experimental designs are being explored to improve the accuracy of wind intensity forecasting and the results will be reported soon.

 

Pedro L. Silva Dias  - Institute of Astronomy, Geophysics and Atmospheric Sciences – São Paulo- Brazil
Amanda Sabatini - National Laboratory for Scientific Computing – Petrópolis - Brazil

 

 

This article also appeared in the LinkedIn GroupJoin us!

Other LinkedIn articles:

Link Wind energy in HPC4E
Link Investigation of flame structure of biomass-derived gaseous fuels
Link Finding the "not so easy" oil
Link How supercomputing can help improving the energy sector
Link Effects of fuel composition on biogas combustion in premixed laminar flames
Link New generation subsurface imaging gets a boost from HPC
Link Novel Hybridizable Discontinous Galerkin method paves the way to tackle realistic 3D problems in seismic imaging
Link Biogas Utilisation for Sustainable Power Generation
Link Preparing the oil and gas industry for the Exaflop era
Link Dynamical and statistical high resolution downscaling approaches for the surface wind
Link Innovative multiscale numerical algorithms for highly heterogeneous media extended to seismic problems
Link Do we really need exascale computers? Geophysicists say yes, we do
Link Providing resilient executions to the energy simulations