Deep Unsupervised Learning for Condition Monitoring and Prediction of High Dimensional Data with Application on Windfarm SCADA Data

Abstract

In this work we are addressing the problem of statistical modeling of the joint distribution of data collected from wind turbines interacting due to collective effect of their placement in a wind-farm, the wind characteristics (speed/orientation) and the turbine control. Operating wind turbines extract energy from the wind and at the same time produce wakes on the down-wind turbines in a park, causing reduced power production and increased vibrations, potentially contributing in a detrimental manner to fatigue life. This work presents a Variational Auto-Encoder (VAE) Neural Network architecture capable of mapping the high dimensional correlated stochastic variables over the wind-farm, such as power production and wind speed, to a parametric probability distribution of much lower dimensionality. We demonstrate how a trained VAE can be used in order to quantify levels of statistical deviation on condition monitoring data. Moreover, we demonstrate how the VAE can be used for pre-training an inference model, capable of predicting the power production of the farm together with bounds on the uncertainty of the predictions.

Examples employing simulated wind-farm Supervisory Control And Data Acquisition (SCADA) data are presented. The simulated farm data are acquired from a Dynamic Wake Meandering (DWM) simulation of a small wind farm comprised of nine 5 MW turbines in close spacing using OpenFAST.

The contribution of this work lies in the introduction of state-of-the-art machine learning techniques in the general context of condition monitoring and uncertainty quantification. We show how the high dimensional joint probability distribution of condition monitoring parameters can be analyzed by exploiting the underlying lower dimensional structure of the data imposed by the physics of the problem. The process of making use of the trained joint distribution for the purposes of inference under uncertainty and condition monitoring is clearly exposed.