Cost-Effective Strategies for Wind Farm O&M: Topics in Structural Reliability, Load Analysis, Predictive Maintenance and Decision Making

Over the decade preceding this writing, important technological advancements in wind turbine design have enabled a massive deployment of large scale offshore wind farm projects.The main question is now shifting from enhancing turbine and farm design characteristics to optimize the energy capture, to operating and maintaining these facilities in an economically convenient manner. This research primarily concentrates on the development of aspects related to wind farm Operation and Maintenance (O&M). This field is by now an established discipline embracing the latest developments in reliability and Big-data, which are transforming the way to operate in the context of the new industrial revolution. There is a considerable drive in decreasing wind energy costs, where innovations in O&M are targeted among the most relevant to achieve the goal. This thesis addresses relevant issues related to wind farm O&M, which specific targets are 1) evaluation of the suitability of environmental models for load assessment of offshore substructure 2) assessment the economic viability of monitoring systems based on current machine learning techniques and 3) correlating loads with experienced failures on a wind farm scale. With regards to load analyses, design load cases (DLC) corresponding to operational states and standstill conditions are considered throughout this work along with normal turbulence models and normal sea states. For development of predictive maintenance techniques, non-operational states are taken out from the dataset. The original contributions of this thesis are readily listed below. The list is uncategorized with respect to the different topics treated.
•Investigation of the suitability of the fatigue equivalent turbulence percentile for monopiles as defined by IEC61400-1 [2] through Monte Carlo (MC) simulations of the joint probability distribution of the environmental variables, such as mean windspeed, turbulence, significant wave height and wave peak period. Contribution from wind farm wakes is also considered based on the model developed by Frandsen [3].•Derivation of the analytical error of Wheeler stretching at finite water depths, used as corrective factor in the original formulation to satisfy the Laplace equation describing the two dimensional wave motion.•Quantification of the wave kinematic model uncertainty with respect to fatigue damage in standstill and operational conditions. Nonlinear waves and Wheeler stretching correctionare considered. This uncertainty is defined as the ratio between DELs in the two cases divided by the DEL resulting from using the linear wave model and Wheeler stretching.•A reliability analysis based on fatigue damage limit state of the most fatigue-loaded section of the monopile is performed, with respect to the variability of environmental models for load assessment, namely turbulence and waves. The limit state functionincludes standstill and operational damage, where their respective uncertainties are adopted in both the load and resistance term of the limit state function. The analysis is performed over the turbine lifetime.•A qualitative analysis about possible correlation between blade root DELs and experienced shut-down events due to malfunctions in the pitch mechanism on an offshore windfarm. This is achieved by recreating the fatigue load map of the blade root bending moments of a wind farm, and compare it with recorded shut-downs due to malfunctions.•A methodology based on least absolute shrinkage and selection operator (LASSO)regularization to select the minimum necessary number of signals to predict damage-sensitive features for regression-based prediction models. The model is based on tracking the normal behaviour fluctuations of main bearing temperatures.•Classification of SCADA alarms for early fault detection. The method is based on avoiding the occurrence of critical failures by predicting low-severity SCADA alarms,which could potentially lead to improved wind farm operations. Neural network and Naïve Bayes classifiers are implemented and their probabilistic output is used to build Receiver Operator Characteristic (ROC) curves. A multi-testing approach is used to generate several ROC curves, in order to quantify the prediction uncertainty.•A methodology to assess the economic viability of classification-based prediction models by coupling machine learning and event tree analysis, considering risk of failures and false alarms. An efficiency parameter is introduced to model the reduced probability of failure of mechanical components given maintenance interventions performed consequently to an alarm. Thus, the quantification of its effect on the total utility is addressed. The efficiency acts on the probability of failure given an intervention from the operator and increases linearly from 0 to 1. The same approach is used for regression-based prediction models.
Thus, in connection to the previous list of developments, the main findings and implications of this work can be listed as follows:•The fatigue equivalent percentile defined by IEC standards is suitable for mechanical components like blades but too conservative for monopiles.•The probabilistic analysis on the hydrodynamic coefficients shows that the DEL distributions are not sensitive to the variability of inertia and drag coefficients, which is an relevant information in reliability analyses.•The wave model uncertainty analysis of the monopile revealed that a substantial difference between standstill and operational loads. These increases dramatically in standstill, due to absence of aerodynamic damping.•The importance factors computed through the first order reliability method revealed that despite the wave model uncertainties are not the major drivers for fatigue reliability,the analysis shows significant differences in terms of lifetime and annual probability of failure when different models are combined. This has impact on design load analyses as well as lifetime reassessment through aeroelastic simulations.•The qualitative assessment between blade root flapwise DEL normalized with respect to its maximum value and probability map of pitch system malfunctions, reveal that the in this case it is not possible to predict these malfunctions through a load map. However,since the assessment is not quantitative, these study cannot confirm a direct correlation between loads and failures, but rather a potential for using the load maps for achieving an improved farm configuration to mitigate the occurrence of critical failures.•The analysis of damage-sensitive features in case of main bearing failures shows that the tower-top acceleration in the fore-aft direction and main bearing vertical acceleration show variation when a failure is present. This information can be used for building multivariate output models based on regression.•The LASSO regularization enables a significant reduction of the dimensionality of the dataset as well as providing a physical interpretation of the failure process. The decision analysis carried out on this problem, revealed that a repair policy is cost-effective compared to replacement, and the first varies as function of the efficiency of intervention.The method thus helps decide on the economic viability of intelligent systems before their implementation.•Coupling machine learning based predictive models with event-trees to quantify the reliability of data-driven monitoring systems provides a criteria to select a risk-based threshold for online classifiers. The approach could lead to improved wind farm operations and smooth running conditions.•Training the classifiers multiple times with a random testing batch for each hold/out enables a quantification of the prediction uncertainty on the ROC curve. This information is useful to ascertain the reliability of the algorithm adopted. Thus, modeling of the reduced probability of failure given intervention by the user, allows to evaluate the efficiency at different lead times, which will help making decisions about the type of system to implement. At last, the analysis of the SCADA data prior a shut-down event revealed that wind speed statistics higher than those normal conditions and abnormal operation may increase the risk of failures. 
Future research should complement key-aspects of this dissertation, from investigating the effect of improved statistical analyses to provide a more general conclusion about the reliability of wind turbine foundations under different environmental models for load assessment; inclusion of detailed analysis of vibration data into prediction models and quantification of their benefit;standardization of monitoring system by merging SCADA and vibration data; scale-up the monitoring systems to a wind farm level and allow flexible learning; extension of decision models to comprise longer lead times, more robust predictions and more accurate cost models;the load maps should be extended with real component failures coupled with measured vibration features and SCADA data.