Operational data mining for offshore wind farm maintenance

Abstract

Utilising operational data from offshore wind farms is one lever which can be used to reduce the cost of energy in the industry. However, the value of operational data in reducing operations and maintenance costs is not yet fully leveraged. This is especially true for maintenance (O\&M) records. Due to corporate sensitivity of the data; poor data collection and processing practices and; a lack of scrutiny into "data fusion" of maintenance data with other data streams, research into the effectiveness of maintenance interventions has significant room for improvement.

This thesis therefore addresses the research question: "How can operational maintenance data be better leveraged to support decision making and therefore reduce O\&M costs in the industry?". This thesis presents a series of analyses which address the research question. First, there is a review of the offshore wind-data ecosystem to identify opportunities for improvement from data processes. Then, the available dataset is used to calculate and scrutinise relevant key performance indicators describing maintenance intervention. From the calculated key performance indicators, the thesis goes on to address two questions posed by the operator of the wind farm: "How effective are night shifts in increasing power production and availability?" and "What is the effect that annual services have on proceeding corrective works and in turn the reliability of wind turbines?".

Frequentist statistics are initially used to analyse a database of work procedures. The data is broken down into different maintenance types and presented in terms of number of interventions per year and mean downtime. Two case studies are then presented which utilise Bayesian methodologies. Bayesian methodologies were selected as they present several advantages which map well to the problem of analysing maintenance data. A Bayesian hierarchical model was used to address the question “How effective are night shifts in increasing power production and availability?". The methodology allowed for conditional probability distributions to be derived for weekly lost production and technical availability from a small sample size, which could in turn be used for decision making. A Bayesian reliability analysis was used to address the question "What is the effect that annual services have on proceeding corrective works and in turn the reliability of wind turbines?". The methodology extends traditional reliability models by incorporating time-dependent variables, which could be used to quantify the effect of annual services on wind turbine time-to-failure. Employing a Bayesian regime in the reliability model provided in-built uncertainty quantification.

A high-quality database from a currently operational offshore wind farm in the UK is used for the above analyses. Results from the work procedure analysis show that tidally-restricted turbines reduce median availability by 0.89% and that failure rates range from below 1 to over 10 failures per turbine per year using different failure definitions. Results from the night shift analysis show a potential 0.64% increase in availability from a night shift strategy involving one crew transfer vessel employed at night. Results from the annual services analysis show a higher failure intensity for the first 6 days after an annual service takes place, after which the failure intensity is decreased up until 137 days after servicing. The case studies show where better processing of operational maintenance data can lead to insight and therefore aid decision making. The Bayesian hierarchical modelling approach facilitates the updating of prior beliefs as new, although limited, data becomes available. The Bayesian reliability analysis increased the complexity of failure modelling, which is useful in quantifying the impact of a maintenance procedure with a time-dependent effect.