The Blog of operational excellence and energy management for industry


Browse our articles, expert advices and clients testimonials: our experience save your time!
Industry machine learning energy efficiency

Applying Machine Learning to energy

Technology – in the form of Machine Learning, Deep Learning, Artificial Intelligence, Big Data, neural networks, genetic algorithms and more – crops up regularly in marketing discourse, but often the concepts are used rather loosely. Let’s take a closer look at Machine Learning and see what industrial energy efficiency stands to gain from it.

 

Definition: Machine Learning

Machine Learning, or Statistical Learning, is one of the disciplines of Artificial Intelligence. It uses a set of methods and algorithms to develop stand-alone software programs that are capable of learning to recognise a complex system’s states from its data history.

Several types of algorithms can be combined in machine learning (decision trees, linear discriminant analysis, clustering, etc.), including neural networks, which became well known for their initial ambition to reproduce the workings of the human brain. Deep Learning, for its part, is “simply” (sic!) one application of neural networks.

machine learning industrie énergie

 

A broad spectrum of applications

From a functional viewpoint, Machine Learning can be broken down into two phases:

1- Initial learning: created from a set of input data and, of course, the corresponding output data. Let’s take an example from the agrifood industry. The input data includes the temperature of a freshly-cooked biscuit, the hygrometry and ambient temperature in the laboratory, the biscuit’s dimensions and a photo. The output data includes its conformity or otherwise and, if applicable, the reasons it was discarded (degree of cooking, size, appearance, etc.). The biscuit’s conformity is assessed “manually”.

2- Operation: the algorithm is then able to return an output value, calculated from the input data provided. For our example: based on the data collected by the sensors, the application tells the automatic control unit which biscuits to discard and sends only acceptable biscuits for packaging. Note that the operating phase may also contain learning functionalities to refine the algorithm’s performances.

  |  Many applications concern the perception of an environment and/or of a body of complex and varied data:

  • Recognition of objects in an image, image indexing, voice recognition, etc.
  • Self-driving car
  • Fraud detection
  • Medical diagnosis
  • Financial analysis
  • Preventive and predictive industrial maintenance

 

Statistics vs physics

Machine Learning uses statistical approaches. This makes it different to physical models, which rely on an understanding of the physical and chemical features (thermodynamics or fluid mechanics, for example) or mathematical features (economics, finance, etc.) of the phenomena observed.

Three outcomes have to be taken into account:

1- Machine Learning does not indicate the reasons for its “decision”: its statistical approach establishes correlations between several measurements, but does not consider their causality.

2- Its capacities for recognition remain valid for a scope and conditions that are “constant” or at least known to the algorithm: whenever the nature or type of input data changes, a fresh learning phase becomes necessary. In practice, any change in the industrial process observed (e.g. a change of machine or a change in the temperature regime), in the means of observation (e.g. sensors) or in the outside conditions (e.g. climate, building alteration) should prompt an evaluation of the consequences on the Machine Learning algorithm.

3- The initial learning process and any subsequent changes require full sets of historic data: all of the possibilities must be covered, which generally means observing at least one full cycle (seasonality). Sometime “virtual” data can be created from a partial physical model (this is known as a hybrid algorithm).

 

What are the best tools for the energy sector?

Process engineers and energy managers will find Machine Learning a very useful tool for:

  • Predicting consumption levels: weather forecasts, load schedules, intrant quality, etc. can be used to accurately estimate the energy intensity necessary for the hours ahead and adjust the operating conditions accordingly beforehand (energy storage, machine start-up/shutdown, etc.).
  • Organising procurement: whether the supplies are fuels or intrants whose quality has an impact on energy requirements, Machine Learning makes it possible to select the optimal source of supplies, depending on the weather conditions, the load schedule, the quality of the intrants and the facilities’ performance at a given point in time.

 

How Machine Learning can be applied to a biomass power plant

A biomass boiler’s performance is heavily dependent on the hygrometry of the plant waste it burns and which comes from several localities in the region (within an 80 km radius). There is limited capacity for buffer storage in the plant’s immediate vicinity, so fuel deliveries have to be timed as close as possible to the boiler’s actual requirements in order to meet the demand for heating. The efficiency of the heat production depends on the quality of the fuel. The Machine Learning algorithm works on the following principle:

marchine learning energy industry

As a general rule, it is a good idea to check the operating conditions of a Machine Learning algorithm at least every quarter.

 

In conclusion, Machine Learning is a useful tool for industrial energy efficiency. It helps with both understanding the phenomena and managing the operational side. Relax! Machine Learning is no longer a (total) mystery for you!