Energy management in a campus of office buildings is the difference between millions of dollars in operating costs and environmental impact. Smart buildings deploy sensors and central monitoring to control lighting, HVAC, and electric car chargers to tilt energy usage towards non-peak hours.
In this use case, Emcien is giving property management at a smart campus the tools needed to predict when buildings in the campus will be using peak energy. When operators know ahead of the peak events they can reconfigure energy consumption away from the high load periods.
This article outlines the step-by-step directions to demonstrate the concept of using time series data. In the real world, operators would add to existing BI dashboards a peak usage screen that alerts operators before it is too late to act. Emcien provides the data for these dashboards to be valuable to property managers.
Getting into the data we see that we have ten months of hourly measurements. A quick look at the data reveals columns like:
|Visitors||How many people are actively inside the building consuming energy|
|Energy Use||Measurement of kilowatts used|
|Day of Week||Indicating if the measurement is performed on a Monday, Tuesday, etc.|
In total there are eleven columns in the data. Please download the sample data here.
Selecting an Outcome
Selecting your outcome is one of the most important steps. When starting you should ask:
In this use case operators want to know when the campus will hit peak energy usage. But how much before is the question. If the warning people is too far in advance, like twelve hours, then the effectiveness of the operators is diminished. In our example, operators wants to know one to three hours before peak energy usage occurs. With this window, operators can adjust the buildings while being close enough to the event to observe the results.
Adding a Time-to-live
Following the Emcien time series steps we add a column called “Hours Until Peak Energy Usage” to our source data. We define peak energy when the “Energy Use” goes above ten thousand. Once this is established, we add a derived column called “Hours Until Peak Energy Usage.” This derived column counts down to the peak energy hour.
With this new column we are now able to define the ideal ranges for the building operators. Using our count down column we define “Hours Until Peak Energy Usage” as:
Step 1: Learning from Historical Data
Emcien is a supervised machine learning application. This means historic data is required so that Emcien can automatically detect and extract signals leading up to the targeted outcome. To begin this learning process, we use Emcien Bandit to automatically detect and place numeric data into predictive ranges.
From the Home Page, click the “Bandit” link at the top right corner.
Drag the Smart-Buildings-LEARN.csv file onto Bandit:
With the input from the building operators, we have determined a set of ranges that is ideal of our use case. We will take these peak energy ranges and apply them to the learning data. This is called “User Bands” since we are defining the ranges rather than allowing Bandit to define the ranges.
To apply our user bands, we first select our outcome (highlighted in red) and then drag User-Breaks.csv file onto the input.
Then click “Transform”. Once Bandit is done click the “Analyze” button.
On the “New Analysis” page click the “Analyze” button because Bandit has filled out the form for you.
Once the analysis is done you will be able to view the rules on the Rules page. This is purely optional. Emcien creates thousands of strong and mixed signal rules to use together when making predictions. These rules do not require any human review, audit, or publishing. Emcien provides them for review when wanting to understand how predictions are made.
Optional: Viewing the Knowledge Graph
Looking deeper into how Emcien works, it is useful to see the knowledge graph. Emcien provides an easy way of viewing the relationships in data. To start simply click on any ‘blue pill’ on the screen.
Once on the “Item Detail” screen click the “Explore Graph” link at the bottom of the right column.
This will open the knowledge graph for you to explore. Here you can see the ‘friends’ of an item and the ‘friends-of-friends’. The weight and color the the lines between items indicates the strength of the relationship. The heavier the line, the stronger the relationship.
Step 2: Testing the Model
In the real world, the model we just built using Emcien would be used to test real time building data. Again, operators would be using BI dashboards and operations alerts to let them know when Emcien predicts a peak energy event will occur in one to three hours. In this document, we will use Emcien to test the model by predicting on a test dataset. This test data is one week of data which includes the known results. These know results allow us to determine if we got the answer correct (true positive) and if we caught all of the test events (capture).
To test data click the upload button on the Emcien home page.
Next upload the Smart-Buildings-TEST.csv file. Once the file is uploaded click the Home button.
Next click the “Predict” link located next to the Analysis you just created.
Next select the “Smart-Buildings-TEST.csv” file on the New Predictions page.
Once Emcien is done applying the rules from the model you built to the test data, you will be able to “View Predictions”.
On the Predictions Dashboard you will see that Emcien did very well predicting peak energy usage. We are most interested in the one to three hour time frame, therefore click on that link in the right column to select that outcome.
We see that our model was 100% accurate for the sixteen predictions made. In the test data there are twenty peak events, we captured 80% of them. This means that everytime Emcien alerted the building operators they successfully altered the energy consumption of the buildings and avoided brownouts, penalties, and excessive wear of systems.
One of the most common questions operators have when using Emcien is:
A strength of Emcien is that anyone can understand why a prediction is made simply by clicking on a prediction. Once on a predictions the matching rules, or reasons, are available for you to view. Outlined in red is information that is valuable when understanding a single prediction.
Emcien is a powerful tool to helping people understand when something will happen. With Emcien, property managers are able to get their promised return on investments when upgrading to “smart buildings’. Emcien had perfect accuracy meaning each event predicted was a real life issue, not a false alarm. The system was easy for operators to adopt because it plugged into existing systems via the Emcien APIs.
This walk-through highlighted the ease and value Emcien provides when solving time series use cases.
For any further questions please contact firstname.lastname@example.org