INTRODUCTION

As wind energy grows in market share, the more it needs to increase its output (performance) and reduce its cost (maintenance).

Wind energy systems stand out from other complex technical systems because of the combination of large levels of wind uncertainty and high levels of interaction of wind farm physics. Big data analytics techniques can significantly improve wind farm performance and reduce costs.

Data are estimated to be created at 2.5 quintillion bytes/day from sensors, social media, images, and myriad other sources. The growing use of big data in wind power operations and maintenance generates an estimated 25 trillion bytes/day. The ubiquitous availability of data has created a paradigm shift from information-poor to information-rich management and impacts virtually every area of modern life.

TRENDS IN BIG DATA ADVANCED ANALYTICS

The use of big data advanced analytics for knowledge discovery, especially machine learning (ML), has emerged as a means to enable smart decisions. It has been successfully used to address problems in many industrial domains, resulting in disruptive innovation that can be leveraged to solve challenges such as those in performance and maintenance costs of wind energy.

The design and development of high-quality large-scale analytics are complex, involving big, “noisy,” structured and/or unstructured datasets as well as a large pool of diverse models. Evaluating just a single model involves a search across all combinations of structures and parameter values, and finding the right scalable ML approach can require many attempts.

The availability of new infrastructures at scale, such as cloud platforms, has provided a new direction for efforts to solve these challenges. The emerging paradigm needs to involve automation of a significant portion of the current manual process involved in problem formulation (to select the appropriate ML algorithms) as well as data preparation, model selection, model tuning, and so forth. In addition, it is important to leverage parallel computing environments—through cloud computing (such as Hadoop), high-performance computing, and large-scale optimizations—to create, maintain, and deploy large-scale machine learning on big data.

Good-quality data are essential to the development of an effective predictive model. There are two main challenges when dealing with data quality: (1) The data are unlabeled even when there is a large pool, and (2) the features do not have predictive power.

In the first scenario, the data could be annotated by subject matter experts. Since their time is very expensive, the challenge is to determine what data points are the most informative to focus their time and effort. ML techniques, such as active learning (Settles 2010), can interactively query the expert to obtain the desired outputs at new data points and solve an optimization problem in order to get the highest performance from the predictive model with the smallest training set.

The second scenario is common when the predictive problem is very complicated and predictive features are missing. In this case, feature engineering (Heaton 2016) can be used for the design of the best (or at least a better) representation of the sample data to yield the necessary information for the predictive algorithm.

BIG DATA ANALYTICS AND THE IOT

The IoT, connecting machines both to machines and to people through networks, data, and analytics, is an important technology for dealing with challenges of big data analytics. As it shapes modern businesses from manufacturing to marketing, the IoT promises to remake global industry, boost productivity, and launch a new age of prosperity and growth.

Machines and other objects have long been used to issue early warnings, but in an inconsistent and unactionable way. The advent of networked machines with embedded sensors and advanced analytics tools has changed that reality. Most machines now either have or are in the process of getting multiple sensors and being connected. The sensors constitute a plethora of data sources that are often neither connected nor integrated, yielding a deluge of data from wind turbines.

To harness the power of data integration and systems-level analytics and optimization in applications such as wind energy O&M, it is critical to ensure interoperability among data sources. But concerns about privacy and cybersecurity are raised by both industry and government. The risk of connecting unsecure devices to the Internet should be properly mitigated through a combination of cyber- and physical security solutions. To accelerate secure data-driven innovation and discovery, new technologies, infrastructure (for networking, storage architecture, cloud computing), new platforms, and cybersecurity technologies are needed to enable industry to effectively tackle the flow of data from machines and objects.

CHALLENGES AND OPPORTUNITIES

The Industrial Internet—the combination of big data analytics with the IoT—is producing huge opportunities for companies in all industries, and renewable energy is no exception. But as one analysis put it, “Not all Big Data is created equal” (Kelly and Floyer 2013). According to the authors, “data created by industrial equipment such as wind turbines, jet engines and MRI machines . . . holds more potential business value on a size-adjusted basis than other types of Big Data associated with the social Web, consumer Internet and other sources.”

To support and accelerate data-driven innovation and discovery, new technologies and infrastructure are needed to empower industry. To that end, GE has invested significantly in a new Industrial Internet platform, Predix Asset Performance Management (Evans and Annunziata 2012; Floyer 2013),¹ to enable big data analytics to address complex systems such as wind farm O&M. Through the diagnostic and prognostic capabilities of GE's new platform, it is possible to reduce operating expenses and move away from traditional reactive O&M to sophisticated predictive and proactive O&M solutions.

For wind energy O&M, this approach extensively leverages physics-based modeling of the system and fuses it with data-driven models and statistical and ML techniques to increase performance and reduce maintenance costs in wind energy O&M. It does so by

• continuously collecting data from assets combined with other operational data to monitor, analyze, and improve performance and maintenance;
• delivering insights from asset-specific advanced analytics models; and
• providing the asset issues to enable smart decisions and the best course of action.

Wind farm operators can thus better understand what is happening in the field, plan ahead, and properly predict extended operating life, resulting in reduced maintenance costs and improved performance.

REFERENCES

• Etherington D. Google says it will hit 100% renewable energy by 2017. TechCrunch; Dec 6, 2016. https://techcrunch​.com​/2016/12/06/google-says-it-will-hit-100-renewable-energy-by-2017/
• Evans PC, Annunziata M. General Electric: Industrial Internet, pushing the boundaries of minds and machines. Nov 26, 2012. https://www​.ge.com/docs​/chapters/Industrial_Internet.pdf.
• Floyer D. Defining and sizing the Industrial Internet. Wikibon; Jun 27, 2013.
• Heaton J. Paper presented at IEEE SoutheastCon. Norfolk VA: 2016. [March 30-April 3]. An empirical analysis of feature engineering for predictive modeling.
• Kelly J, Floyer D. The Industrial Internet and big data analytics: Opportunities and challenges. Wikibon; Sep 16, 2013. http://wikibon​.org/wiki​/v/The_Industrial​_Internet_and_Big_Data​_Analytics:_Opportunities​_and_Challenges.
• Moodie A. Google, Apple, Facebook race towards 100% renewable energy target. The Guardian; Dec 6, 2016. https://www​.theguardian​.com/sustainable-business​/2016/dec/06​/google-renewable-energy-target-solar-wind-power.
• Saintvilus R. Microsoft makes largest wind energy investment. Investopedia; Nov 15, 2016. https://www​.investopedia​.com/news/microsoft-makes-largest-wind-energy-investment-msft​/?lgl=rirabaseline-vertical.
• Settles B. Computer Science Technical Report 1648. University of Wisconsin-Madison; 2010. Active learning literature survey. http://burrsettles​.com/pub/settles​.activelearning.pdf.