Wind Power Resource Estimation with Deep Neural Networks

The measure-correlate-predict technique is state-of-the-art for assessing the quality of a wind power resource based on long term numerical weather prediction systems. On-site wind speed measurements are correlated to meteorological reanalysis data, which represent the best historical estimate available for the atmospheric state. The different variants of MCP more or less correct the statistical main attributes by making the meteorological reanalyses bias and scaling free using the on-site measurements. However, by neglecting the higher order correlations none of the variants utilize the full potential of the measurements. We show that deep neural networks make use of these higher order correlations. Our implementation is tailored to the requirements of MCP in the context of wind resource assessment. We show the application of this method to a set of different locations and compare the results to a simple linear fit to the wind speed frequency distribution as well as to a standard linear regression MCP, that represents the state-of-the-art in industrial aerodynamics. The neural network based MCP outperforms both other methods with respect to correlation, root-mean-square error and the distance in the wind speed frequency distribution. Site assessment can be considered one of the most important steps developing a wind energy project. To this end, the approach described can be regarded as a novel, high-quality tool for reducing uncertainties in the long-term reference problem of on-site measurements.

[1]  Nitish Srivastava,et al.  Improving neural networks by preventing co-adaptation of feature detectors , 2012, ArXiv.

[2]  James Martens,et al.  Deep learning via Hessian-free optimization , 2010, ICML.

[3]  Tom Howard,et al.  Correction and downscaling of NWP wind speed forecasts , 2007 .

[4]  James F. Manwell,et al.  Comparison of the performance of four measure–correlate–predict algorithms , 2005 .

[5]  Volodymyr Mnih,et al.  CUDAMat: a CUDA-based matrix class for Python , 2009 .

[6]  Geoffrey E. Hinton,et al.  Training Recurrent Neural Networks , 2013 .

[7]  S. Schubert,et al.  MERRA: NASA’s Modern-Era Retrospective Analysis for Research and Applications , 2011 .

[8]  Frank Sehnke,et al.  Multimodal Parameter-exploring Policy Gradients , 2010, 2010 Ninth International Conference on Machine Learning and Applications.

[9]  Martin A. Riedmiller,et al.  A direct adaptive method for faster backpropagation learning: the RPROP algorithm , 1993, IEEE International Conference on Neural Networks.

[10]  A. Ng.  Feature selection, L1 vs. L2 regularization, and rotational invariance , 2004, Twenty-first international conference on Machine learning - ICML '04.

[11]  Yoshua. Bengio,et al.  Learning Deep Architectures for AI , 2007, Found. Trends Mach. Learn..

[12]  Frank Sehnke,et al.  Parameter-exploring policy gradients , 2010, Neural Networks.

[13]  Geoffrey E. Hinton,et al.  Restricted Boltzmann machines for collaborative filtering , 2007, ICML '07.