Use data and machine learning to improve the operational performance of wind turbines

Effectively manage wind turbine icing with a modern data platform

The amount of money that utilities spend on operations and maintenance (O&M) is increasing with no slowdown in sight. Operational performance is an area organizations look to most often when trying to optimize staffing levels while reducing costs. In this use case, we’ll explore how a modern data platform can help manage wind turbine icing and improve operational performance.

Wind energy is quickly evolving into a main pillar of the world’s green energy initiative, and as such its operations are being looked at with greater scrutiny. Unlike other operations within the industry, the operation of wind turbines must be evaluated a little differently because each turbine is affected differently by weather conditions, altitude, and other factors based on its location.

Enhance wind turbine performance using machine learning and advanced analytics

When optimizing the performance of wind turbines, you need to consider several different factors, including the engineering specification of the turbine and its blades along with its location and the weather affecting its performance. To make sense of all this data, you need a data platform that will allow you to combine the data and apply machine learning (ML) as quickly as possible to provide insights to better optimize your operational performance. In the case of wind turbines, ice or even frost on the blade has been shown to affect the turbine's aerodynamic efficiency substantially, and it can reduce power generation by up to 80%. The ability to use ML and advanced analytics to swiftly understand, prepare for, and manage this loss is imperative to minimize the overall impact and maintain operational efficiency.

This image shows how Oracle Data Platform for energy and water can be used to support a use case around Operational Performance specifically regarding wind turbine icing. The platform includes the following five pillars:

1. 1. Data Sources, Discovery
2. 2. Ingest, Transform
3. 3. Persist, Curate, Create
4. 4. Analyze, Learn, Predict
5. 5. Measure, Act

The Data Sources, Discovery pillar includes three categories of data.

1. 1. First-Party data is comprised of asset metadata, GIS data, and video.
2. 2. Applications includes SCADA, device data hub, and outage and maintenance management systems.
3. 3. Third-party data includes data from GPS and weather sources.

The Ingest, Transform pillar comprises three capabilities.

1. 1. Batch ingestion uses OCI Data Integration, and Data Studio.
2. 2. Change data capture uses OCI GoldenGate.
3. 3. Streaming ingestion is comprised of OCI Streaming and OCI Connector Hub.

All three capabilities connect unidirectionally into the serving data store, and cloud storage within the Persist, Curate, Create pillar.

The Persist, Curate, Create pillar comprises four capabilities.

1. 1. The serving data store uses Oracle Autonomous Data Warehouse.
2. 2. Cloud storage uses OCI Object Storage.
3. 3. Batch processing uses OCI Data Integration, Functions, and Data Flow.
4. 4. Governance uses OCI Data Catalog.

These capabilities are connected within the pillar. Cloud storage is unidirectionally connected to the serving data store; it is also bidirectionally connected to batch processing.

The metadata lines unidirectionally connect from the serving data store and cloud storage to governance.

Two capabilities connect into the Analyze, Learn, Predict pillar: The serving data store unidirectionally connect to the analytics and visualization, predict , learn and AI services. Whereas the Cloud Storage connects bi-directionally to learn but unidirectionally to the analytics and visualization, learn and AI services.

The Analyze, Learn, Predict pillar comprises six capabilities.

1. 1. Analytics and visualization use Oracle Analytics Cloud and Spatial Studio.
2. 2. Data products, APIs uses OCI API Gateway, Oracle Integration Cloud, and OCI Functions.
3. 3. Predict uses OCI Data Science and Oracle Machine Learning services.
4. 4. Learn uses OCI Data Science and Oracle Machine Learning notebooks.
5. 5. AI services uses OCI Vision, OCI Language, and third-party services.
6. 6. The predict capability is unidirectionally connected to the data products, APIs capability.

The Measure, Act pillar captures how the data analysis may be applied to support a wind turbine icing delivery model and monitor performance. These applications are divided into two groups.

The first group Peoples and Partners includes operations and maintenance.

The second group Applications includes Oracle Field Service, Oracle Utilities Work and Asset Management, enterprise asset management, work management system, and field service management.

The three central pillars—Ingest, Transform; Persist, Curate, Create; and Analyze, Learn, Predict—are supported by infrastructure, network, security, and IAM.

Wind turbine operational performance logical architecture

There are three main ways to inject data into an architecture to enable utilities to effectively evaluate their operational performance strategy for wind turbines.

• We’ll use batch ingestion to import data from systems that can’t support streaming (for example, older supervisory control and data acquisition (SCADA) systems or maintenance management systems). In this use case, asset data, weather data, and data from GPS, maintenance, and outage management systems will be ingested at varying intervals. We’ll use OCI Data integration to load these data sets into Oracle Cloud Infrastructure (OCI) Object Storage or directly into Oracle Autonomous Data Warehouse (ADW).
• In addition, we’ll use Oracle Cloud Infrastructure GoldenGate to ingest data from operational systems, such as outage systems, maintenance management systems, and asset data hubs, via change data capture.
• For any streaming data, such as weather data, we’ll use the seamless combination of OCI Streaming and OCI Connector Hub to capture, aggerate, and load the data directly into OCI Object Storage.

Data persistence and processing is built on three components.

• Ingested raw data from all sources is stored in cloud storage. Depending on the action needed, we can use the automated features within OCI, such as OCI Events Service, to initiate batch processing events. In this use case, we’re going to ingest the weather data and convert it to a readable form for later use. We’ll then use OCI Data Integration, OCI Functions, or OCI Data Flow for batch processing to consolidate, curate, or enhance the collected data as needed. The data pipelines are built and maintained using OCI Data Integration. Though OCI Data Integration comes with a wide array of connectors for varied data assets (databases, applications, object storage, REST APIs, and so on) it may not meet all your needs. If this is the case, you can build an OCI Data Flow application to take advantage of all the connectors that are available via Spark. In this example, asset data hub, GPS, weather, historical outage, and maintenance data are combined to build a model to identify the physical asset locations requiring attention; this information can then be used to improve the operational maintenance program.
• We have now created processed data sets that are ready to be persisted in optimized relational form for curation and query performance in the serving data store provided by ADW. This enables us to visualize the results of the model predictions. We can even use the built-in spatial capabilities to visualize turbines that may require immediate attention.

The ability to analyze, learn, and predict is facilitated by three technologies.

• Analytics and visualization services, such as Oracle Analytics Cloud and Spatial Studio, can deliver interactive dashboards we can use to visualize image information and predict the future impact of weather on specific turbines or turbine farms. These services provide
  • Descriptive analytics, which we can use to illustrate current and historical icing trends with histograms and charts to help identify areas requiring immediate maintenance
  • Predictive analytics, which we can use to plan and determine longer-term operational and maintenance needs by predicting future weather events, identifying trends, and determining the probability of uncertain outcomes
  • Prescriptive analytics, which can propose suitable actions to help optimize strategic operational performance management decision-making
• Alongside the use of advanced analytics, machine learning models are developed, trained, and deployed using OCI Data Science. These models use machine learning to analyze large amounts of asset, weather, maintenance, geographic information system (GIS), and other data to allow you to better understand and improve the operational performance of each wind turbine. This fine-grained understanding can help you continually prioritize tasks such as shutting down turbines and determine the work to be done and the teams required to do it in the most efficient and cost-effective manner. Some of the more common types of models used are XGBoost algorithms and those that can be found in deep learning, such as recurrent neural networks, deep neural networks, and transfer learning. Once these models are trained, they can be deployed in several ways depending on the user’s preference. The models can be called via REST endpoints using the OCI Data Science platform or the in-database Oracle Machine Learning Services REST API. Additionally, the user can package these models up in an Open Neural Network Exchange (ONNX) format and deploy them as part of an application.
• Our curated, tested, and high-quality data and models can have governance rules and policies applied using OCI Data Catalog in conjunction with other services and can be exposed as a “data product” (API) within a data mesh architecture for distribution across the organization.

Better your utility’s bottom line with improved operational performance

Inefficient maintenance strategies can degrade operational performance and profitability, and lead to unsatisfied customers. This wind turbine icing use case is just one example of how you can use ML and other advanced analytic techniques, including predictive and prescriptive analytics, to fine-tune your operational performance strategy. By using these techniques, you can now anticipate freezing events and asset breakdowns and generate actionable insights in real time. These insights trigger prescriptive workflows so you can take preemptive action and optimize your maintenance. The following examples are some of the possible outcomes you can realize when you use the right data platform to improve your operational performance:

• Improved reliability
• Reduced proactive/preventive maintenance response times
• Lower costs
• Reduced restoration time
• Improved convenience of supply