Powering Predictions in Congestion Trading: A new era for analytics
In the world of power trading, the ability to predict congestion demand, and price accurately has become a vital competitive differentiator. Across the sector, decreased volatility has led to more organizations pursuing a greater number of smaller opportunities—and it’s clear that traditional approaches are no longer enough.
Trading instruments such as Financial Transmission Rights (FTR) and Congestion Revenue Rights (CRR), for example, rely on a series of calculations to forecast congestion and determine the best strategy. In the past, traders could succeed with their intuition, relationships and an array dashboards and analytics, including supply-demand power flow models. However, as pressure mounts to extract greater profit, organizations must analyze multiple pricing inputs, congestion reports and other external data to build models that are more accurate. As model accuracy increases, so too does the potential for greater gross margin.
Traditionally, increasing model accuracy meant first deploying advanced enterprise-wide platforms. However, technology is moving at a fast pace and the platform itself is just the beginning. Organizations that marry the right infrastructure with more innovative approaches, including machine learning and artificial intelligence, are achieving the real competitive edge.
This paper examines the journey that organizations should take to reach the desired maturity in their analytics. This requires rethinking and deploying an intelligent data strategy, including the sourcing and curation of data, and delivering that through an integrated platform. These two foundations will enable them to create new models and harness more accurate insights to increase their gross margin.
How mature are you?
A company can define its overall maturity for insight generation based on two factors: its analytical maturity and its data maturity.
- Analytical maturity: Analytics starts with the fundamentals of reporting, which equip firms with initial visibility into the market and the supply chain. Companies can then add further layers to their capabilities, such as business intelligence and visual analytics, to become more analytically mature. The most mature companies are those already leveraging machine learning and artificial intelligence, which is where they will be able to create the biggest impact.
- Data maturity: There are different levels of data maturity based on how effectively a company manages each element, such as data sourcing and curation. Firms with an undefined data strategy will likely struggle with processes that are performed differently and in separate ways by individuals across the organization. In contrast, those with an optimized strategy will use continuous feedback and responses to metrics and other inputs to improve processes.
Analytical maturity must evolve in line with a company’s data maturity. If a company is not mature in terms of its data, efforts to deliver advanced analytics become futile. Similarly, if it is “data mature” but underperforming on analytics, it is not fully utilizing the available information.
Companies need to develop both analytical and data maturity, but how? With so many factors to consider, they can look to those that have already achieved success to help construct their own roadmap for change.
Machine learning in action: predicting load, congestion and LMP
As part of Sapient’s research and development efforts, we wanted to apply the power of machine learning to improve congestion forecasting, but doing so required having a cohesive data and platform strategy. This meant deploying an iterative approach to test and refine its models and generate the optimum results. The first step was to build a model to more accurately predict congestion forecasting compared to a standardized model (power flow model based on supply and demand) through a proof of concept. Subsequent steps included looking at historical data and identifying the top 10 node pairs in an ERCOT footprint, which offered the best opportunity to action congestion prediction based on historical price sensitivity. We deployed an iterative approach to test and refine our models to generate optimum results. However, the plan began with selecting the right platform.
A framework for accelerating advanced analytics at scale
The platform powering the analytics consists of multiple elements, from the underlying infrastructure and data components through to refined reporting tools.
The biggest challenge for many organizations is that much of this exists in a fragmented structure, with many elements having evolved in isolation. For companies to access the insights they need, they must first unite their technology within an integrated framework.
The most effective way to achieve this is to start with the key elements. This includes a single place to house all of the data, along with the machine learning models and other relevant tools for advanced analytics.
Then, access to the essentials needs to extend to all the traders and analysts across all desks, as well as the supporting personnel, including IT. It is only through having this shared view of the same information that companies can truly collaborate and get the most from their data.
With this enterprise-wide setup in place, the platform can then support the ongoing evolution of the data strategy.
Developing the data strategy
With decisions made about the platform, organizations can implement their data strategy using an iterative, ongoing approach. In general terms, this process includes multiple iterations of data and proof of concepts:
- Proof of concepts: For each use case, companies develop a series of proof of concepts to test the data sets and the models
- Refining the data: Each proof of concept output enables them to change variables and refine the tools, such as discarding invalid data that does not generate meaningful results.
Once the proofs of concept have proven successful in addressing a specific use case, companies can deploy that approach across the wider business, such as in other ISO regions or other desks. They can then turn to their next set of use cases and repeat the same iterative process using the same platform and tools.
This continual cycle of development has proved successful in helping companies deliver more advanced analytics that generate new levels of insights. Over time, companies become better equipped to capitalize on opportunities such as these and, ultimately, increase profits.
Collecting the data
We began our implementation with one proof of concept for each desk (e.g., ERCOT), using the data already available within our organization. The most crucial was the time-series data: generation data, load data, weather data, price data and congestion data. We then built the right data strategy around these sources, including how to handle the time-series elements.
Handling time-series data
To mimic real-life scenarios related to the handling of time-series data, we crafted a challenge that involved addressing gaps across multiple inputs due to missing data or dissimilar granularity of data across inputs. In some cases, we could disregard the missing data point but in other cases, to avoid introducing bias, it was better to substitute it using a unit or item imputation.
Training the machine—feature engineering
The next step was to decide how the data would be used. For example, when predicting the power load price (LMP), would it be accurate to use the 11th-hour price to predict the 12th hour? Similarly, Monday at 11 am might be a good indication of Tuesday 11 am. However, Sunday at 11 am might be less relevant when looking at weekdays. Knowing what information was the most reliable and valuable was essential to deciding what and when to plug into the models.
Machine learning offers a choice of several models. Rather than focusing on training one model at a time, we evaluated multiple popular models in parallel, including support vector machine (SVM), regressions, ensemble/random forest, neural network (2 layer) and Gaussian. This required only small amounts of additional effort, making it a far more efficient approach.
Both these steps became iterative: identify training data, which was a good representative sample to begin training the model; then run the models, enabling them to develop intuition based on the factors affecting accuracy. With each output, we could adjust variables until we achieved the desired model accuracy.
Scaling into production
Before deploying to production, we will need to put this model through a set of ‘risk governance’ tests. This will include gauging sensitivity with different variations of inputs as well as extreme inputs to ensure they are valid and the model does not go out of control at the boundary conditions or extreme conditions while in production. Once the models begin generating the intended result, using a scalable platform will make it significantly easier to deploy for production usage.
The future of analytics—Amplifying human intelligence
Analytics are evolving at a rapid pace, with new and innovative approaches continually raising the bar. In the case of congestion trading, the stakes are high and opportunities are fleeting, so finding new ways of generating accurate predictions is vital.
Our initial work in rethinking the technology infrastructure helped us ensure the best possible framework was in place. We were then able to refine our data strategy, which included deploying iterative approaches to hone the accuracy of our models.
However, for each organization, the challenges and specifics of each stage will vary. In a future paper, Sapient will consider each stage in more detail, explore approaches that can be applied across the sector and share the results of the proof of concept. After all, as more and more firms succeed in increasing their analytical maturity, success will depend on who can amplify their human intelligence by training their machines to be the smartest.
Masud Haq, Senior Vice President, Head of Digital Strategy for Energy and Commodities
Masud Haq leads Sapient’s Digital Strategy for the energy and commodities industries. Since joining Sapient in 1994, he has worked with several major oil companies, merchant energy companies, gas and liquids pipeline companies, retail energy companies and investment banks. He was instrumental in propelling Sapient into a leadership role in the gas and power market in the United States following deregulation in the mid-1990s through 2006. From 2007 to 2016, he ran the North American energy business for Sapient, establishing the company’s Houston and Calgary offices. In his current role, he is responsible for setting strategy and leading key digital business transformation engagements. Prior to joining Sapient, Masud was a research assistant at the MIT Center for Theoretical Physics, where he also studied physics and philosophy. He has published in the Annals of Physics (August 1993, H.M. Haq and M. Crescimanno).
Rajiv Gupta, Director, Business Consulting for Energy and Commodities
Rajiv brings 17 years of experience delivering transformational projects and digital empowerment in the energy domain across wholesale/retail power, LNG, midstream/downstream crude oil and refined products. Specializing in fit-for-purpose solutions across start-up, mid-size and large multi-national companies, Rajiv helps to envision and implement mission-critical projects to support mergers, business transformation and time-to-market business strategies.