Machine Learning Applications in the Energy Sector

Sepehr Ramyar

by Sepehr Ramyar
Department of Technology & Information Management, University of California Santa Cruz

The energy industry, and particularly the power sector, is undeniably an essential component of any modern society. In fact, electrification has been determined as the “greatest engineering achievement of the 20th century” by the National Academy of Engineering (NAE, 2018). However, this mighty engineering achievement of the previous century faces just as great of challenges in the 21st century that can no longer be addressed using analytical tools of the past. Increasing scale and complexity of the power systems, inclusion of new active players and stake-holders in the power sector, and rapid evolution of institutions and regulations that govern and operate the energy industry are the main reasons that necessitate the application of robust analytical tools that can efficiently and effectively address these challenges. Machine learning has proved to be a powerful tool in extracting and processing information from large sets of data and has been used extensively for different applications in the energy sector. In this article, we will discuss how machine learning has been applied to enable a more efficient power system.

The power system traditionally was a monopoly in which a single utility company produced, transmitted, distributed, and ultimately sold electricity to end-users. This paradigm, however, started to change in the 1980’s and 90’s as competition was introduced to the power systems to induce economic efficiency. In the new framework, power is traded in an electricity market, usually called the wholesale market, in which generators and consumers (the load serving entities1) bid their supply and demand quantities and the market consequently clears under a set of conditions2, optimally allocating power by merit order: the cheapest producers (generators) supply electricity to the consumers with the highest willingness-to-pay. In technical terms, this is a type of double auction. For example, imagine there are three generators respectively offering 5, 10, and 15 megawatts (MW) for 3, 4, and 5 dollars per MW ($/MW) for a specific hour of the day. There are also three consumers bidding 1, 3, and 6 $/MW each demanding 5 MW for the same time/hour of the day. Then, the least-cost generator, i.e. the 3 $/MW generator, is allocated its full 5 MW to meet the demand of the consumer with the highest willingness-to-pay, i.e. the 6 $/MW consumer. Because the remaining two consumers’ bids are lower than the remaining two generators’ asks per MW, the generators are not able to further supply the remaining customers. Thus, the market clears at 3 $/MW. Each day, a similar process is repeated for each of the 24 hours of the following day in what is called a day-ahead market to determine the price and quantity of the power to be generated and delivered.

It is evident that if one is able to accurately predict the market clearing price, then they can place bids that have a much higher chance of being cleared (and compensated). Moreover, the Independent System Operator (ISO) that operates the power market wishes to identify and intercept fraudulent bids to limit strategic behavior by participants. As mentioned before, this auction is carried out every day for each of the 24 hours of the following day. This means that there is plenty of data available to exploit and this is where machine learning tools come in handy. One particularly useful method is Support Vector Machine (SVM). This machine learning technique maps inputs to a feature space and then the predicted outcome is calculated as a linear function in the new feature space. The strength of this method lies in its ability to linearly separate a dataset even though it may not be separable in the original space of the training data. In the case of power price forecasting, the feature space could include elements such as time (day and/or hour), temperature (max, min, average), humidity, etc. in the location of demand. The linearity saves a lot of computation time and makes SVM a particularly useful analytical tool in the context of an electricity market. Specifically, in the real-time market3(as opposed to day-ahead markets) the entire bidding and market clearing process is carried out in a very short time window (less than five minutes) and consequently it is essential for market participants to be able to carry out massive computations for predicting the price is the shortest time possible, and this is where SVM’s computational efficiency comes in.

Hourly prices for power through a market mechanism introduce a host of opportunities for new businesses to grow in the energy sector. Once particular example is energy demand management. This stems from the fact that there are different energy-consuming activities (e.g. lighting, heating, electric vehicle charging, etc.) and each consumer has his own preferences/ranking over them or can be incentivized to rank them. Using machine learning algorithms, it is possible to recognize the energy consumption pattern of each consumer and identify energy savings opportunities based on the variable market prices. For example, one can automatically, using smart energy management systems that run on machine learning algorithms, shift their electric vehicle charging from peak hours to off-peak hours with lower prices. One particularly useful machine learning tool here is reinforcement learning. This method dynamically learns and adopts the behavior that yields maximum reward which in this context could be translated into monetary savings. The applicability of the reinforcement learning in demand management is that no historical data is required, and the algorithm would be able to navigate and detect optimal action in real-time.

Scaling up this solution, a new participant emerges in the power market: aggregators. An aggregator operates vast fleet of demand management systems throughout a community. Each of these units is capable of identifying consumption patterns and energy saving opportunities based on the machine learning algorithms. This energy saving could be in form of forgoing consumption (e.g. reduced lighting) or shifting consumption (delaying electric vehicle charging to off-peak hours). The aggregator then bundles or aggregates these energy savings and offers it to the wholesale market in form of demand response i.e. power that otherwise had to be generated and consumed and receives compensation in return. In addition, when the power grid becomes under stress, the ISO can issue demand response (ask for reduced demand) and aggregators would then be able to secure the required demand response from their pool of customers and receive (and share) compensation for it. This guarantees reliability and efficiency of power systems.

So, we can see how different elements of the current institutions that operate power systems can be improved using machine learning algorithms. By enabling the computation and extracting information from massive data sets, machine learning algorithms have enabled various agents in the power system such as consumers, generators, market operator, and aggregators to scale up solutions and effectively exploit new opportunities introduced in the new energy sector paradigm while at the same time improving economic efficiency and reliability of the greatest engineering achievement of the 20th century.

1  Load serving entities are electric service providers, i.e. retail electric providers or municipally owned utilities
2  This is usually referred to as market clearing conditions that in addition to trying to set demand equal to supply, also involve physical characteristics of the power grid in terms of capacity, congestion, etc.
3  After the day-ahead market clears and quantities are determined, there might still be deviations from the allocations due to unexpected failures or demand surge. These deviations are settled in the real-time market on the same day that units are dispatched.


 NAE [National Academy of Engineering]. (2018). Greatest Engineering Achievements of the 20th Century. Retrieved from