Optimising With Big Data
Can optimisation save the world? Well it depends on the problem.
Optimisation may not be able to solve all of humanity’s problems but it has recently made important strides to help avert the impending global warming disaster. As I’m sure you’re aware, the global warming crisis is rated by scientists worldwide as being one of the most important issues of the twenty-first century. For decades now scientists have been warning us that we need to wean ourselves off our dependency on fossil fuels. Although replacements to fossil fuel-based energy have existed for over half a century, none of them seem to have made a significant impact on the global warming issue yet. For example wind power can currently produce electricity at a price that is close to fossil fuels based electricity. However, in spite of the impending crisis, wind power is still to enter our energy mix in a big way. One of the reasons for this is not simply because of lack of political will but because of a lack of good optimisation tools. Although, in the last five years it seems that the winds of change are blowing because some of wind power’s biggest opponents have gone silent. So what has happened?
To be able to appreciate what has happened and why optimisation is so important, we need to understand what the problem is and to be honest, the problem is wind. Unfortunately, like visits from your in-laws, wind does not come when we want it to nor does it behave like we want it to (imagine generating wind power during a cyclone!). So, due to its unpredictability, naively using wind power requires companies to have back-up power plant on standby just in case the wind suddenly stops. Back-up power plants are much less efficient than a permanently on power plant and counter-intuitively a situation like this can lead to more CO2 pollution than if we had just used a coal fired power plant!
What has happened in the recent past, which has turned wind energy from a much debated dream into a reality, is an optimal use of wind turbines to generate energy. Due to such optimal use of wind turbines, in Colorado, USA, the energy company Xcel Energy (which mailed its customers to oppose a meagre 10% renewable energy contribution), is now planning on profitably sourcing up to 30% of its energy from wind by 2030 (Reference). This is more than the famous 20% by 2020 proposed by Germany (which is now coming under fire due to high energy costs) and owes its success to better weather prediction and an optimal use of the wind turbines.
The case of optimising wind turbine electricity production is obviously a complicated affair given the fact that wind can change its direction and strength continuously throughout the whole day. So, in its simplest form, the challenge for those operating wind turbines is, given predictions about wind strength and direction in the next moment, how should the turbine’s parameters be adjusted to maximise electricity output. These parameters typically translate into physical properties of the wind turbine such as the direction the turbine is facing, the angle of the turbine blades, the speed and torque of the turbine drivetrain, and so on. Given this significant challenge, even if we could perfectly predict the wind conditions in the next moment, we probably would have no way of knowing the perfect parameter values for each given set of wind conditions. To find a solution to this problem, scientists have turned to a set of tools used in big data called machine learning tools to predict optimal parameters (and also to help predict the weather too!).
Machine learning tools are typically used when there is a large amount of data available to teach a system to behave in a certain optimal way. In our case, it is to teach the wind turbines to adjust their parameters optimally to maximize their electricity output given the predicted wind conditions. Typical examples of machine learning techniques include Bayesian Networks (used everywhere from biology to economics), Support Vector Machines (often for recognizing handwriting) and, probably most well-known, Artificial Neural Networks (which were used famously in 2012 by Google to pursue the noble cause of identifying pictures of cats in videos).
In the current case of optimising parameter selection for optimal electricity output of wind turbines, one promising tool to find the best choice of parameters is Artificial Neural Networks. For those working in the industry, this choice may seem natural, however, implementing a good neural network is wrought with difficulties. One needs to choose a correct model for the function to be optimised, given the available data one needs to choose a good learning algorithm to train the neural networks correctly and finally the computations need to be robust to new and unseen inputs. It is precisely in these fields that we have seen significant advances (and also stellar buyouts such as DeepMind for $450 million in 2014) which have led to more efficient algorithms for parameter selection. In turn this has led to the recent adoption of more and more wind turbines to generate renewable energy due to the more electricity produced by each turbine. In 2013 in Australia alone, plans were underway to install another eleven wind farms adding 1627 MW to the grid (Reference).
However, when talking about advances in wind turbine technology what constitutes a significant advance? Let’s take an example of a country that is not even a major player on the world scene but exemplary when it comes to wind energy generation, Spain. According to the Guardian, wind energy was Spain’s top source of electricity in 2013. It generated approximately 54,000 GWh of electricity in 2013 from wind turbines. So assuming that we are able to improve our optimisation methods so that electricity production increases by a mere 0.1%, then, in Spain alone, we would generate an extra 54 GWh of electricity. Given that 1 MWh supplies up to roughly 200 households, this would mean that simply by increasing output by 0.1% in Spain alone, we can supply two cities with Sydney’s population with the extra electricity!
Facts like this are what cause insignificant sounding headlines by Siemens and GE to be such game changers. Recently Siemens claims to have had a breakthrough in their neural network algorithms that enables their turbines to produce one percent more electricity annually under moderate wind conditions. Given our previous calculation we can see how important this is. Even more impressive than this, given what a one percent difference can do, was GE’s headline that they are able to increase their wind turbine output by five percent (reference)! They haven’t stated specifically how they did it but their description of “big data” and “predictive algorithms” would lead one to guess at similar machine learning algorithms as for the Siemens success story.
As we can see all these small gains can add up and as the reader can guess, without the optimisation tools available today these feats would not have been possible. We can only hope that these improvements point to an industry that is poised to revolutionise our energy industry and, with the help of optimisation, maybe save the world.