How is AI transforming industry in its pursuit of new energy savings?

Since the invention of the first statistical algorithms, the progress of AI has been considerable. This "learning" intelligence of the machine offers great opportunities in many sectors, including energy and industry. Indeed, statistical AI models, built on the basis of real industrial data, allow to visualize in real time energy drifts and to identify possible causes, but also to calculate a potential energy gain and to test various approaches to reach it (or even to exceed it!). 

After a brief historical review on the origins of artificial intelligence, discover in this article the potential of AI in the industry through the approach of our Data Scientists and a concrete application case on site.

How is artificial intelligence born?

Short history of AI

Historically, the notion of Artificial Intelligence dates back to the second half of the 20th century: in the 1950s, the first software capable of learning to play chess was created. The term AI was invented in 1956. The term machine learning, in 1959. 

Between the 60's and the 90's, computers are not very efficient, technical means are limited... AI is still young. We call this period "the winter of AI".

Between 1996 and 1997, the AI "Deep Blue" - a program launched by IBM - beat the world chess champion!

Again in 2016, the champion of the Go game - a more complex game than chess - is beaten by Deep Mind, an AI programmed by Google, about 20 years earlier than anyone expected!

What is the definition of artificial intelligence?

A simple definition of AI could be: "computer programs that are able to adjust their behavior according to the treatments and actions they have to execute". We speak of "learning" artificial intelligence.

Data Science, an inclusive discipline for developing algorithms for AI, is at the heart of 3 themes:

• Computer science,
• Mathematics,
• Data analysis.

At their convergence, we find different major competences in the field of energy: data manipulation, exploratory analysis, and modeling.

How does artificial intelligence work?

What is an artificial intelligence model?

In the industry, we can define a model as a mathematical representation of a phenomenon, for example energy consumption.

The objective of the model is to analyze this phenomenon by simplifying it, i.e. by "reducing" it to the sum of a few influencing variables in order to clarify its understanding. We also talk about explanatory variables, or Features (e.g. a data from the production).

Examples of artificial intelligence models for industry:

Machine learning and deep learning, two families of models in artificial intelligence

At Energiency, Data Scientists use 2 families of models:

• Machine learning. Ex: Gradient Boosting Tree. This is a family of models based on decision trees (random forests, with reinforcement techniques to improve the accuracy of the models by iterations). 

• Deep learning. Ex: neural networks. This is a family of models organized in layers of several neurons, which receive input data and will reproduce an output signal according to different parameters (output laws). For example, an activation threshold. These models allow to obtain a good accuracy in modeling the output data. 

Example of how artificial intelligence is used in industry: the method developed by Energiency

Diagram illustrating the steps involved in modeling an industrial phenomenon with artificial intelligence 

1. Formulate a problem and identify useful data

The first step is to determine what we want to look for and optimize!By analyzing the industrial context, Data Scientists will then ask the following questions:

• What is the variable to be explained?
  What are the possible explanatory variables?

A common mistake would be to collect a maximum of data at random. However, the best practice is to start by formulating a problem and then identify the useful data that will help answer the problem.

In the energy domain, these data are typically :

• Data related to energy and fluids,
• Data related to production and activity (activity rate, shift...),
• Data related to the weather (temperature, humidity...).

2. Clean up the data

After selecting the right data, the Data Scientists will then perform a cleaning and filtering step to exclude data that are not relevant.

3. Build the dataset

In building artificial intelligence models, we will of course use the initial data cleaned beforehand, but also new data built by the Data Scientists. This is called Feature Engineering.

These data are calculated from the raw variables, but have greater explanatory power. Generally, these types of constructed variables are what improve the accuracy of the models.

Example of calculated variables:

• time of day,
• summer / winter,
• production with a lag of x time.

4. Train the model

Then comes a period of training of the model, in order to adjust its accuracy by iterations. 

There are feedback loops to correct errors at each stage, depending on the performance obtained: this is the principle of learning the AI model. To learn well, a model must be fed with a "good" variety of data (in terms of history and number of variables available).

5. Test and measure performance via indicators

After learning, it's time for a test run! To evaluate the relevance of the model, it is subjected to a period of tests, then its predictions are measured against reality. For example, we can make sure that the model adapts well to the stopping and restarting of industrial machines (at the right speed, and with the right orders of magnitude).

The delta then represents the error of the model. If the model is good, the error is centered around 0. 

Once the model is sufficiently trained, we can then calculate a potential energy gain. To do this, we look at the most important drifts and, via a hypothesis matrix, we estimate the gains that can be made by using the model.

Pitfalls to avoid in model building: 

• Overfitting: a model that perfectly reproduces the basic values. It is very accurate in the training period, but less efficient in the period when it discovers new data. To understand, imagine a student who has learned by memory, but without understanding!
• Underfitting: a model that is not at all adjusted, too imprecise. Here, it would be a student who has just not learned, or has learned poorly, his lesson.

It is therefore necessary to ensure that the performances are as similar as possible between learning and training (we can say that we have consolidated a generalization of the model to all real situations), while being as precise as possible.

Once again, just like a student learning to perform operations!

Use of AI models at Energiency 

1. Gain potential study using historical data

Data Scientists start with a potential gain study on historical data (12 to 18 months of records). Our teams will then:

1. Identify the relevant influencing factors and quantify the impact of these factors on consumption,
2. Develop the models,
3. Calculate the potential energy gain.

We can identify factors with a significant impact (strong influence), but also other factors that are lighter and can be adjusted manually or automatically, such as the settings of a machine (e.g. temperature of an oven, frequency of maintenance, flow of fluid used, etc.).

This step allows to define a target of potential gain (e.g. percentage of energy savings achievable on a workshop). This energy target will be the "course" to be maintained in all the following steps.

2. Implementation of statistical AI models on the platform 

Once the AI models have been tested and sufficiently trained, the data scientists will implement the models in the energy performance software. 

The platform allows, among other things, to visualize expenses and to track energy performance in real time. This is also known as an Energy Management System (EMS).

This configuration will allow, among other things, to visualize:

1. Over-consumption alerts, in real time, based on the reference consumption model (reference to be "respected" with regard to the desired savings objectives for the period).
2. The factors that cause energy consumption variations.

3. Identification of concrete energy saving solutions by the Energy Managers

Once the potential savings have been detected and the platform is operational, the challenge is to be able to make the right assumptions and to take the relevant decisions based on the energy drifts and consumption patterns observed. 

The Energy Managers, who are responsible for the operational monitoring of energy consumption and the search for concrete ways to make savings, then start working.  Their missions include:

1. Analyze deviations from the reference energy consumption, 
2. Testing corrective approaches and validating the energy gains made (or continuing to adjust),
3. Carrying out regular assessments of the savings achieved.

4. Update of the artificial intelligence models

Data Scientists update the models every year - if performance has been good - based on the data collected over the new period.

The overall approach is therefore rooted in a logic of continuous improvement, in order to identify new sources of energy savings, thanks to a very detailed analysis of consumption and influencing factors.

In conclusion: Artificial Intelligence does not go without ... Human Intelligence!

To be effective, AI must be :

• relevant: with accurate models, relevant alerts...
• easy to use: the platform must be ergonomic, reactive...
• easy to interpret: the factors explaining the differences should not be too difficult to identify.

But be careful, AI is only a tool! We talk a lot about digital, but digital alone doesn't go very far. Manufacturers need to arm themselves with skills, experience and good human intuition in parallel. This is what we call Human Intelligence.

This is why our teams work closely with industrial sites, industrial directors, maintenance managers, energy managers and operators to adjust models and identify the right corrective measures based on field experience. It is these human exchanges that ensure the quality of our energy saving researches.

How industrials can benefit from the potential of artificial intelligence: the SKF success story

To discover a concrete application of the statistical models developed by Energiency, download the success story of SKF, a major global bearing manufacturer:

You will discover:

• SKF's initial challenge in its search for additional energy savings on its most energy-intensive workshop.
• The meeting with Energiency.
• The implementation of 2 AI models based on 18 months of rich production and consumption data (electricity and compressed air).
• Deployment of the Energiency application and operational follow-up with the Energy Managers (visualization on weekly reports, monitoring on a 10-minute time step, over-consumption alerts, visual management by color, search for continuous improvement, etc.).

You are wondering if artificial intelligence and a data science study would allow you to identify new energy saving opportunities?Each industrial situation being unique, do not hesitate to contact us so that we can study your context and your current energy management.