Tools


Siemens Worldwide

Pictures of the Future

Contact

Contact

sts.components.contact.mr.placeholder Sebastian Webel
Mr. Sebastian Webel

Editor-in-Chief

Tel: +49 (89) 636-32221

Fax: +49 89 636-35292

Werner-von-Siemens-Straße 1
80333 Munich


sts.components.contact.mr.placeholder Arthur F. Pease
Mr. Arthur F. Pease

Executive Editor English Edition

Tel: +49 (89) 636-48824

Fax: +49 89 636-35292

Otto-Hahn-Ring 6
81739 Munich
Germany

Pictures of the Future
The Magazine for Research and Innovation
 

Artificial Intelligence

The Science of Prediction

Dr. Hans-Georg Zimmermann of Siemens Corporate Technology has developed a learning system capable of forecasting everything from the price of copper to the best place to locate a factory.

What’s the best time for a company to purchase electricity or key raw materials? Is it possible to predict the hourly output of a wind park with enough accuracy to plan the use of back-up gas generators? Siemens is developing methods that can identify, track, and learn the key parameters that underlie such systems and trends, resulting in the ability to predict many processes with amazing accuracy.

Take a break and look out the window for a second. What do you see? Partially identifiable shapes — or buildings and trees? Chances are that if you had never seen a building or a tree, and had never even heard of such things, the view might indeed appear to be a confusing jumble. The reason that it isn’t is that you have models in your mind that organize the enormous amount of data in your field of vision into things that are instantly identifiable.

Now, the same sort of challenge is being confronted by complex man-made systems; but the models in question make sense of patterns that are so multifaceted and so invisible to our senses that no human being could ever see them. They are the patterns from which, with steadily growing success, predictions can be made.

Precise Demand Predictions

And they work! The predictive technologies now evolving at Siemens offer surprisingly sharp snapshots of the future output, behavior, and maintenance needs of systems ranging from turbines to wind parks and of the development of economic trends such as the prices of raw materials and the direction of the stock market. Indeed, Siemens already gets decision support for its electricity purchases in Germany and its huge worldwide copper purchases from predictions provided by its Software Environment for Neural Networks (SENN) learning system, which is, according to Senior Principal Research Scientist Dr. Hans-Georg Zimmermann, “the most advanced high-dimensional, non-linear modeling system of its kind.” Thanks to more then 20 years of experience in integrating mathematical research, software development and real world applications, SENN has been able to focus on the science of prediction more consistently and continuously than any other program.

Zimmermann, who has laid the mathematical groundwork for over 60 predictive industrial applications, registered 22 patents to protect associated software system architecture models, and holds university lectures in quantitative finance, explains that neural networks offer significant advantages over conventional predictive systems based on linear logic. “Neural networks can cope with real world applications, no matter how non-linear or multi-dimensional the underlying problem is. In addition, neural networks are an elegant framework for the modeling of temporal structures,” he says. For instance, in a recent study designed to predict demand for 16 types of electrical control cabinets, Zimmermann’s team pitted SENN against a linear model. The two systems predicted sales volume on a month-by-month basis for each cabinet type over a full year. But SENN took factors such as foreign exchange rates and fluctuations in automation systems markets into account. The result: SENN achieved an average error of only 23.3 percent (compared to actual demand) — much better than the linear model’s error, which was 52.6 percent. “This kind of highly accurate demand forecast can help to optimize a supply chain and reduce costs,” says Zimmermann.

SENN is also playing an important role in predicting the supply of wind energy. A SENN model is now being tested on data from a large offshore wind farm in Denmark. The model uses forecasts for wind speed, temperature, and humidity to predict the farm's electricity output for the next three days to within 7.2 percent. For example, if the system forecasts an output of 100, the actual value would be between 92.8 and 107.2. “The accuracy of the forecast depends mainly on the quality of the data,” says Zimmermann’s colleague Dr. Ralph Grothmann. “All in all, we can predict the weather fairly accurately three days in advance.”

“This method provides a range of different future scenarios to be played out and evaluated.”

"With the rising stake of renewable energy sources such as wind in the total energy mix,” says Zimmermann, “utilities not only need to predict demand, but also supply. Prediction is important because it allows them to estimate when to activate back-up gas-fired generation.” With this in mind, Zimmermann’s team developed a neural network based on the major parameters that can affect wind power generation. “In such cases, the goal is to create a software model that is a mathematical representation of the real world,” says Zimmermann. But initially, he explains, the model does not know how important each parameter is — and that’s where learning from data comes into play (for more, see article “From Biological Systems to Machines, Learning is the key” http://www.siemens.com/innovation/apps/pof_microsite/_pof-fall-2011/_html_de/wie-maschinelles-lernen-funktioniert.html).

All the system knows at first is that, given the input it receives during its training phase, it will have to produce an output that is as close to the actual power output of the wind park as possible over time.

At first, the discrepancy between model output and actual data is huge. But over time, the learning algorithm begins to modify the individual parameters within its model so that the predicted and actual results become closer and closer.

By measuring its level of error over thousands of iterations, the system gradually moves from producing random outputs to identifying which combinations of weights on which input parameters result in which effects. “It’s like learning how to score a goal in a soccer game,” says Zimmermann. “All you know is that your output should be to get the ball into the net. Through a process of trial and error, and given the thousands of possible circumstances that can influence the result, you may learn to get it just right.”

Quantifiying the Unkown

And SENN did get its prediction of the wind park’s output right. Its average error in terms of predicting the total energy supply of the park per day (calculated in terms of root mean square deviation) is now down to 7.2 percent — a full three percent better than the closest competing physics-based model. Similar models are currently being developed for photovoltaic plants.

Similarly, Zimmermann’s team has developed a neural network to model the nitrous oxide (NOX) emissions of gas turbines. Such a model can be used to analyze the relationships between numerous input variables and the output of a turbine over time. As with the case of the wind park, SENN began with only raw data and a mandate to describe actual output over time. Nevertheless, as it learned the relationships between variables the model grew closer and closer to duplicating the turbine’s behavior, and was eventually able to predict its behavior in real time with almost perfect accuracy.

But of course there’s a lot more going on in a turbine — or any other complex system for that matter — than just its known variables. As Zimmermann points out, “There are variables that you cannot measure; and then there are those you do not even know about.” Such invisible variables can add up to a mountain of uncertainty. “In view of this,” says Zimmermann, “we have discovered a new way of explaining uncertainty — one that frames it as the interaction between observable and hidden variables.”

By comparison, the standard approach to measuring uncertainty in mechanical and economic dynamic systems is to translate the deviation between what the model predicts and what actually happens in the real world into an estimate of risk. The underlying assumption is that the model of uncertainty measured in the past is a good estimator of future risk.

 

“But this does not generally apply to predictions in the world of finance, which can include copper and electricity prices,” cautions Zimmermann. “Here, the idea is that uncertainty spreads from the present into the future as a diffusion process — scaled by measured historical model error — becoming larger and larger as we move forward through time.” In contrast, according to Zimmermann’s solution, since it is not possible to reconstruct hidden system variables unambiguously, you can quantify the amount of uncertainty in a prediction by analyzing the distribution of different scenarios that take shape. Here, the range of fluctuation between scenarios is interpreted as the level of risk, and a scenario based on the mean values from the different scenarios — all of which have the same probability — can be assumed to be the most probable future trend. “The resulting market risk is thus characterized by the variation between the scenarios,” says Zimmermann, who explains that, given a finite number of observations, there will always be multiple ways to reconstruct hidden variables, thus resulting in different scenarios for the future.

Siemens already uses these methods to augment procurement decisions for energy and copper. “Instead of just a single model of the future,” adds Zimmermann, “this method provides a range of different future scenarios to be played out and evaluated.”

How might the science of prediction evolve over the next few years? Clearly, if the past is any guide, we will see a steady progression toward increased accuracy. As Zimmermann points out, not only are SENN models learning more each day, but its creators are learning from the models it generates as they morph into closer and closer representations of reality.

Beyond forecasting energy and raw materials prices, beyond predicting the outputs of wind parks and turbines, SENN offers the potential for virtually limitless numbers of applications. It could help with some of the most challenging, complex and costly decisions of our time, namely those associated with urban and regional investment decisions in areas such as road, air traffic, water, and electrical infrastructures. Indeed, SENN’s potential as a decision support system is already being tested at Siemens to help determine, for instance, the relative long-term advantages of different sites before building a factory.

And beyond that? A different model for our relationship with the future is taking shape in the form of a demonstration SENN Forecast Server now running on Siemens’ intranet. The system is being used to introduce internal customers to SENN’s potential.

Fast forward ten years and we may be downloading SENN apps to monitor, learn from, diagnose, and optimize the functions of our homes, vehicles, businesses, and supply chains. SENN’s future versions may even be able to offer scenarios that support optimized, personalized nutritional, healthcare, educational, and financial paths. Every question, after all, has an answer that lies somewhere in the future.

“The science of prediction,” says Zimmermann, “is a race between the increasing complexity of the real world and our accelerating ability to mathematically represent it by means of information-technology-related capabilities, such as SENN models.”

 

Arthur F. Pease
Picture credits: from top: 2.picture bluemagenta, 3. Prisma