The rate of data collection has grown exponentially across the globe. In our own industry, the advancement of remote sensing devices such as LiDARs, ubiquity of satellite-based measurements and the growth of operating assets rich in sensors has resulted in the storage of ever-increasing terabytes of data, which if analysed intelligently, can be used to help make highly impactful business decisions. In fact, this information boom has led to the growth of a whole new multi-disciplinary field: data science.
Data science utilises scientific methods to extract knowledge, gather insights and build predictive models from data which in turn can be used to design more efficient renewable projects or increase the performance of operational assets to maximise return on investment. What’s more, whereas in the past it would have taken months to build and run complex predictive models, high performance computing and powerful open-source software libraries now enables us at RES to do this in minutes.
One of the fields that modern-day data scientists excel in is machine learning, a subset of artificial intelligence. In many respects machine learning is simply a synonym for predictive modelling (i.e. using statistics to predict outcomes) and is defined as the use of algorithms and statistical models to learn patterns and trends within complex datasets.
While analytical models (developed from a fundamental understanding of a problem) are highly accurate and are based on mathematical models derived from centuries of academic literature, there is no mathematical formula to describe how, for example, groups of consumers behave. This is where machine learning is useful – we use masses of historical data to determine how consumers tend to behave, on average, then use that as our best guess.
Machine learning lets the data do the talking by quantifying the relationships between variables empirically. The technique does not require explicit instruction on how to quantify these relations. The relations – if they exist – are automatically “learned” via algorithms that search for structure within data. If relationships between variables change over time or space, the model can be programmed to adapt and learn.
Care must be taken, however, as machine learning models are only as good as the data on which they have been trained. For example, some search engines have had to refine their input datasets to avoid gender biased results for search terms such as ‘nurse’. Such unintentional consequences point to the need to carefully curate the data fed to machine learning algorithms. A solution to the example in question is to equalise the proportion of male and female examples in the training dataset to ensure the results are not gendered biased. Furthermore, machine learning, by nature of its fundamentally statistical approach, has a finite probability of returning the wrong answers. An algorithm that analyses images may be 99% accurate at identifying people. However, that leaves 1 out of 100 times when it will provide the wrong answer. Such relatively small inaccuracies may not matter for certain applications (counting the approximate number of people on a train platform for crowd management) whilst it will be critical for others (autonomous vehicles identifying pedestrians who have strayed onto the road). Careful use of the outputs of machine learning models with consideration for the confidence of the model prediction in the context of a specific application is, therefore, essential. Corroboration of predictions from multiple datasets also helps minimise errors e.g. autonomous cars analyse data from multiple data sources (cameras, RaDAR, LiDAR etc.) to provide a highly reliable identification of obstacles, hazards and pedestrians.
Machine learning and renewables
At RES, we have built up our data science capability across the globe with experts in the fields of computing, computational science, statistics and machine learning. Combining our skillsets and working together has led to advances across all sectors of our business and stages of our asset lifecycle.
In construction, scheduling is carried out at times that minimise delay by predicting the most likely periods that extreme weather events will not occur. In pre-construction wind energy resource assessment, machine learning models can now be trained with multi-height LiDAR data to accurately predict long term operational energy production surpassing the accuracy achievable with traditional met mast datasets. In energy trading, machine learning models can forecast day-ahead energy prices – these forecasts can then be used when trading on energy markets to maximise revenue. In asset management and O&M, machine learning is used to proactively schedule component maintenance ahead of when they are predicted to fail, not reacting once they have failed, reducing downtime and OPEX costs for owners. We can also use machine learning and onsite sensors to automate noise and shadow flicker management by automatically detecting potential issues and proactively changing turbine modes or temporarily shutting down turbines.
The applications are endless, but the key takeaway is that machine learning can be used to multiply the human capability of our subject matter experts, ensuring the scalability and customer responsiveness of our business.
The term “machine learning” is searched today ten times more frequently than it was in 2010 and has become a modern-day buzzword, grasping the attention of businesses and government.
The hype is fuelled by media attention alongside promises of magical algorithms that can solve any problem in the world. Just try opening Google images and search for machine learning – pictures of seemingly sentient, autonomous robots appear. Machine learning is not general intelligence, but a rather a technique for generating highly specialised synthetic intelligence for specific problems, based on the data available for training. A machine learning model trained on one problem will not be of any use to an unrelated problem and if based on poor data, will generate poor results. Conversely a well-defined problem with quality training data can be automated using machine learning and executed many times more efficiently by a computer than a human.
Machine learning solutions require data, computing power and above all, skilled data scientists who can diagnose the problem and use the appropriate methodologies suited for the job. Each solution is tailored to the specific problem at hand and can only solve that one specific task – machine learning is not a magical black box.
This doesn’t mean machine learning isn’t powerful – it is. Tech savvy data scientists are coming up with new ways to harness its power in almost every industry. Google use it to predict traffic flow on its map service. Tesla uses it in its driverless cars. We use it to help achieve our vision: a future where everyone has access to affordable zero carbon energy.