Introduction and Background

Typically retail energy companies buy electric power and natural gas from generators at the wholesale level on behalf of its customers, each of which fall under one of many fixed-cost, variable-cost, or hybrid energy plans offered by the company.   The plan dictates how much each customer pays for their energy needs month to month.  In this sense the plan can be thought of as a set of rules which convert a batch of demand to the consumer’s cost of energy every billing cycle.  Companies spend significant resources designing plans (“products”) that will appeal to consumers in the areas they serve.

While there are many factors that drive the success of providers, three important aspects of company operations are paramount:

• Customers should be fitted to plans that create the most value for them, while providing the company the flexibility to buy their power in the most advantageous way
• The company must retain as many good customers as it can
• The company must purchase power for the least possible cost to ensure attractive operating margins

For these reasons, we have devised a “program” of analytical models that support the strategies directly.  The models serve to aid human decision makers every day/every hour to move toward optimal performance in the three areas.

Part 1. Plan Optimization for a Customer Segment of 1

It is very common for companies to divide customers into segments for the purpose of devising marketing plans and pricing for that group.  Each segment is clustered along some set of measured attributes – income, payment history, demand size, etc. In the era of data analysis tools we do not have to accept an imperfect fit of segment to customer. We can, in essence, make each segment one customer by fitting a given plan precisely to a given customer’s unique needs.

For every customer there is a load profile—the seasonal demand pattern.  When that pattern is placed against a given plan, the result is a demand and payment schedule for that customer across the year.  The amount of demand is also driven by weather.  In some cases there is feedback from the spend back to the behavior – if the customer sees very high electric bills in the summer he/she may scale back on air conditioning load.

In any case, it is possible to “model” the behavior under a range of plans for each and every individual consumer.  Once done, we can see which plan would be optimal for the company as well as the consumer, and perhaps suggest/encourage a different plan for that consumer if the current plan is not optimal.

Part 2. Decreasing Customer Churn

Customers come and go as with any consumer service.  However the cost of acquiring a new customer to replace the one that just left is very high, so reducing the number of defections as a percentage of the whole customer base, known as churn, is a high value activity.

Organizations attempt to reduce churn by persuading customers that have signaled a desire to defect to reverse their decision.

By this point it is most often too late.  The better alternative is to seek to predict which customers or customer groups are likely to defect given the pattern of defections from historical data.  Consumer studies show that defections are not random but rather have complex patterns that are difficult to detect.  This is where tools such as machine learning can be brought to bear on the problem of predicting churn.

Machine learning involves “training” a software model on which customers leave per their individual attributes. The machine learning training is accomplished by presenting a list of known defectors and their attributes (payment history, plan, location, etc.) to the model. The model will then seek to “predict” the attrition probability of all customers using the training set. The model will be a living model in that it will be continuously trained and re-trained to account for changes in the population and also to narrow its results for better predictive capability.

The Benefits of Building and Using the Model

• To build the model we would by necessity create a detailed process blueprint/value map of the “calculus” of churn. The blueprint in and of itself is a valuable document.
• Just Energy personnel spend less time on the mechanics of constructing analysis and more time thinking about the strategy of delivering service efficiently to customers.
• The act of building this first model lays down important infrastructure, which will make the next round of analysis faster and cheaper.

Part 3. Supply Side Optimization

Retail energy companies purchase electric power and natural gas to fulfill demand obligations on behalf of its customers.  This is done on the open marketplace in each of the regions in which it operates.  In Texas, for example, wholesale power purchases are coordinated through ERCOT

At present this is a stand-alone activity—the power obligation is communicated to the department responsible for supplying power and the power is acquired.  In other words, the demand is given, and the company tries as best it can to supply the power to meet that demand.
There is very little feedback from the demand end of the business to the supply side in the sense that smarter, more advantageous purchases could be made with knowledge about the nature of demand, or predictions of demand—akin to the shopkeeper that knows a big snowstorm is approaching and buys snow shovels in bulk in advance of the storm.We propose to create models that use demand side signals and forward predictions to enhance the power buyer’s position to the wholesale marketplace.

After Steps 1 to 3: End to End Optimization

The ultimate goal is to make companies work as one coordinated whole, sensing and responding to real time information as it arrives—an agile, omnipotent organization.  Once steps 1 to 3 above are completed, it is a relatively straightforward matter to link all of the optimization models together, in the same way that purchasing a shirt at Wal-mart starts a ripple effect of actions that span the organization, resulting in the replacement of that store’s inventory with a new shirt.

Summary

These are just a few ideas that relate directly to typical retail energy company strategies and a means to act on the strategies with operational systems that hold the organization true to its strategic intent.  Modern progressive firms use data and analytics to make operations work better, faster, and cheaper (as in the popular book Moneyball), and retail energy companies have the opportunity to lay a foundation of analytical competence for many years to come by engaging in projects such as those described here.