Predictive/Prescriptive Modeling on Residential Energy Consumption

Predictive analysis is about using different techniques to predict the future even though, most likely, the prediction is imperfect. Prescriptive analysis, on the other hand, is about turning the sketchy prediction to actionable orders. Like Biff Tannen in ”back to the future”, having access to the sports almanac is all it takes to win Warren Buffett’s $1 billion bracket challenge to fill out a perfect NCAA bracket. But we know precise information about future is not possible except through a time machine that can move between past, present, and future. Predicting the future might happen on wide range of timescales from milliseconds in wall street ’s trading business, to monthly prediction of number of citations in transportation, to billions of years of predicting the titanic collision of our Milky Way galaxy with the neighboring Andromeda galaxy. Similar to mentioned examples, wherever we look, we can see clear competitive advantage of knowing more about the future.

One of the most inspiring emergences of predictive and prescriptive analysis and modeling is in energy consumption and specifically residential sector. With wide spread use of smart meters and other related sensors, a tremendous amount of energy consumption data are available. My smart meter data somehow captures who I am regarding my lifestyle, what I do in terms of activities, and where and how I live. The smart meter data reveals roughly what time I wake up, what time I leave the house, and what time I come back home. Even the data can potentially identify my eating habits, e.g., weekday’s peak load in the evening can be a sign of activities/cooking in the kitchen. The data carries time of use consumption, trend and seasonal patterns of my behavior. More broadly speaking, the raw data even can potentially encapsulate sentimental information, e.g., figuring out whether a person is happy based on the energy consumption patterns. Figure (1) demonstrates the heatmap of a subject in a single-family home in Bay Area over a length of one year.

Figure 1. Heatmap of a subject in a single-family home in Bay Area over a length of one year. Darker spectrum indicates higher level of energy consumption.

Focusing on predictive and prescriptive modeling in energy, the overall goal of these techniques is to identify behavioral patterns from diverse set of historical and temporal data, e.g., energy consumption, building information, weekly schedules, even maybe shopping habits, and so on. In this regard, the most intuitive starting point is to perform time-series analysis on just historical energy consumption. Note that the diverse nature of most people’s lifestyles imposes very volatile energy consumption. State of the art techniques can be used to implement algorithms to deal with the highly volatile data, and output probability distributions that fit the historical data. Figure (2) displays an example of 4 days forecast using historical energy consumption.

Figure 2. Hourly prediction of energy consumption over 4 days

Predictive and prescriptive analysis and pattern recognition are beneficial to many modern industries. In energy domain, they can also potentially have wide range of advantages. Here we provide a list of potential applications:

• The first potential is a direct application of predicting temporal energy consumption using either just the historical energy consumption or combination of historical energy consumption and other potentially available information such as calendar data. Statistical information on the aggregated sum of the predicted energy consumption over the neighborhood or the city can be very valuable to the utilities in order to more effectively and efficiently plan ahead, e.g., to more efficiently perform power purchase agreements.
• Utilities come up with different incentive programs, rebates, demand response, different tariffs, and so on. One of the big issues is to anticipate people’s reaction to these programs, 
 but a better understanding of behavioral patterns can be extremely helpful to better predict the success chances of different programs.
• Adopting new behaviors and habits is the key for many energy programs to succeed. On the other hand, lessons learned from fields such as experimental economics and quantitative behavioral finance tell us that human’s behavior and their decisions are more heuristic. We believe there is an analogy between the way people consumes energy and the way they make financial decisions. Long history of research in experimental economics along with predictive modeling and extracting past behavioral patterns will create the right environment to facilitate behavioral changes.
• These types of behavioral and predictive analysis on energy consumption can have broader applications. More insights on people’s lifestyles are the key in effective targeting in marketing industries. Marketing goods and services based on insights obtained from analyzing energy consumption can potentially provide new business opportunities. Predicting how likely it is for a person to purchase a product, e.g., an Electrical Vehicle (EV), based on the person’s approach to energy consumption can provide competitive advantages. 
We are living in a new world with new rules especially regarding the human-machine interactions. More sensors, more data, more complex scenarios, and more complicated needs and behavioral patterns provide tremendous amount of opportunities to be harvested using analytical techniques. Analytics considers an essential part of many projects at PARC.

The application of model predictive control (MPC) in the context of energy needs to be addressed as well. MPC is an advanced and well-studied branch in control theory. MPC optimizes over a finite time horizon, but only implements the current time slot. The main advantage of MPC is its ability to anticipate future events and take control actions accordingly. The intersection of MPC and predictive/prescriptive analysis and its ramification in energy and other domains can bring fruitful insights to the table.

By Dr. Marzieh Nabi, Research Scientist and Technical Project Lead at PARC; other contributors to the project include Dr. Yu An Sun.