Looking Backward to Move Forward

For most things in life, I look forward to go forward. I keep my eyes on the road ahead. I look before I leap. I aim at my target. But this summer while we were on our family vacation, we rented a row boat and it reminded me that rowing a boat is different. When rowing, I face backward to maximize my power going forward. Forecasting electricity is like rowing a boat, we need to look backward to go forward.

Intuitively, we already do this by using regression models. Regression models look backward, find relationships and project them forward. Although too often, we leave the looking to models instead of our own eyes.

Below is annual historical average use for residential customers over an 11-year period. Leaving the forecast to a regression model with HDD, CDD, and some binary variables, the outlook is close to 11,600 kWh/year (blue dashed line) and flat because the model does not include any trend variables.

But, the last three years show dramatic decline in usage and raises the question of whether the forecast should be closer to 11,000 kWh/year (yellow-dashed line).

Instead of leaving the forecast to the model, we should use our own eyes to identify the underlying pattern in the history. Is average use declining?

The only way to answer this question is to weather normalize historic average use removing the impact of weather variation. Weather normalization will answer the question of why 2016 was so low, why 2014 was so high, and what happened in 2012.

After weather normalizing historic average use, we can answer what happened in 2012, 2014, and 2016.

  • 2012 was a mild weather year
  • 2014 was an extreme weather year
  • 2016 was a mild weather year

Additionally, we can see that average usage has consistently declined from 2009 to 2016.

If we assume the trend stops in 2016, the forecast should be close to 11,400 kWh (purple solid line). If we assume the trend continues, the forecast will be close to 11,000 kWh/year in 2020 (purple dashed line).

Weather normalization is an essential part of the forecaster’s work process. The process allows us to look backwards to ensure that we move forward in the right direction.


Seasonal Sales? Not a Problem.

There is a client who has a reporting issue problem. Let’s call him Ray. His annual sales are to be split into summer and winter sales, but on a seasonal, not calendar basis. In Ray’s world, winter begins in November and ends in April spanning two calendar years. While Ray sounded calm when he called me, I could feel the tear of frustration welling up in his eyes. How can we summarize annual winter sales in MetrixND when the sales fall in two calendar years?

Let’s be more specific. The picture to the right shows the monthly sales. Ray wants to call November 1995 through April 1996 the Winter 1995 sales.

Typically, the MetrixND’s SumAcross function is used to convert monthly data to annual total. The process takes two steps. First, the monthly sales are split into calendar summer and winter sales in a monthly transformation table. Second, the annual sums are calculated using the SumAcross function in an annual transformation table. The steps, transformations, and results are show below.

But, this is not what Ray wants to do. To use MetrixND’s data transformation capabilities, Ray needs to move the January 1996 through April 1996 values into the January 1995 through April 1995 positions as show below. If Ray can do this, then the annual transformation technique works.

The good news is that Ray called and I have a solution.

Using the following transform, I can move January to April sales using the Lead function.

Once I move the data, I can use the SumAcross function, just like before, to summarize annual summer and winter sales leaving Ray very happy.


Remember the 2017 ISO Forecasting Summit

This year’s 11th Annual ISO/RTO/TSO Forecasting Summit was held in San Antonio, Texas from May 16-18. Just like the defenders of the Alamo banded together amid adversity many years ago, forecasters from CAISO, ISO New England, NYISO, PJM, MISO, ERCOT, SPP, AESO, IESO, Tennessee Valley Authority, and Bonneville Power Administration gathered to share insights into some of the most pervasive challenges facing today’s industry.

During the three days, a vast range of rich and thought-provoking topics were discussed. These topics included emerging challenges of solar penetration and trends, and complex modeling issues. These topics are further described below.

  • Emerging Challenges of Solar Penetration. The emerging challenges of solar penetration were highlighted as attendees presented on various approaches used to incorporate behind-the-meter solar generation into both short and long-term load forecasts. An interesting discussion of common practices for managing steadily expanding solar and wind resource markets got everyone engaged, and empirical research exhibiting increased load weather sensitivity caused by increased PV penetration gave everyone pause.
  • Emerging Trends. Companies discussed the potential impacts of time-of-use (TOU) rates, plug-in electric vehicles (PEV), and emerging trends. Some anecdotes about TOU rates offered perspective on incorporating prices into a forecast model.
  • Modeling Issues. Finally, we were all reminded of the importance of getting creative when it comes to building robust forecast models. Attendees demonstrated the value of identifying additional explanatory variables such as solar irradiance, and the challenges of moving from system level forecasts to point-of-delivery forecasts. The consequences of inaccurate weather forecasts and irregular weather events were also a high point.

It’s amazing what happens when you get some of the most brilliant minds in the industry in one room.  I’m already looking forward to next year.


Cotati, CA – Transitioning to AMI: Removing the Guesswork

Cotati, California is a bedroom community of 7,500 residents located in Sonoma County about 45 miles north of San Francisco. As a progressive, forward thinking municipality, the city has a long standing ethic of water efficiency and in 2011 contracted with WaterSmart Software to deploy a water report program to better educate residence on their water spend and ways to become increasingly efficient to save money. Along with this engagement program WaterSmart provided Cotati with a customer portal where end-users could view their consumption and better understand their water use. While this was a great benefit to customers, the city was still manually reading water meters every other month. This meant that information that was presented in the water reports and portal was only updated every 60 days, which limited the level of engagement with end-use customers.

The city remained interested in increasing customer engagement and further improving water efficiency as a historic five year drought struck California in the summer of 2011. In June of 2014 the city signed a performance contract with Siemens Industry, Inc. which financed a series of upgrades to city infrastructure including the water metering system. A performance contract pays for a group of upgrades for a city and the 'loan' for the capital upgrades is repaid from savings the city realizes from those efficiency investments. The performance contractor, Siemens in this case, guarantees a certain level of savings so the city knows exactly how much money will be repaid over the course of the multi-year agreement. While the performance contract also provided financing for a city lighting upgrade, the portion allocated to water system improvements had 3 primary objectives:

  • Minimize 'unaccounted for water'
  • Provide customers with better information
  • Offer proactive management of the water distribution system

Cotati chose to implement Advanced Metering Infrastructure (AMI) and selected Itron, Inc. as the vendor for the water meter and network upgrades. Deployment of the AMI network was completed in mid-2016 and the city then worked with Itron to integrate the new available hourly interval data with the WaterSmart platform. This gave the city and residents powerful new capabilities including notifications for possible leaks and more detailed consumption analytics to help the city more effectively manage resources. In a recent interview, Cotati Public Works Director, Craig Scott, talked about his experience with the new Itron AMI system and the WaterSmart platform.

One of the key takeaways for Mr. Scott, was how the WaterSmart customer portal acted as a bridge to the city's new AMI system . The portal had been deployed for several years prior to the AMI roll-out, but as hourly consumption data became available from the Itron system, that information was seamlessly offered to end-use customers through the portal. "We had the WaterSmart portal, but information was one data point every two months. With the Itron technology we kept that portal but now the information is more real-time," explained Mr. Scott. "What we saw was a tremendous increase in the number of customers that signed up for that feature." The availability of hourly data enabled the WaterSmart system to detect unusual usage patterns and alert customers to possible leaks. The more frequent data feed also provided an incentive for more customers to register for access to the consumption information.

The combination of the technologies is helping the utility reduce costs by avoiding the need to manually read meters while getting real-time visibility into possible system leaks. On the customer side, users are more engaged, registering for leak notifications and self-resolving many of their own questions, thus helping reduce call volumes for the city's customer service representatives. "It's hard to explain to people what happens when you start using new technology." said Mr. Scott of the new system. "An example is the shift going from a single monitor to a dual monitor with your desktop computer - you're singing the praises of how did I ever live with one monitor - with the AMI system, it's like that."

Ultimately, reducing costs and improving service quality is the goal of every water supplier, and the City of Cotati is a great example of how regular information, proactive outreach, and technology can help demonstrate the value of services provided by the utility. "We can show the data which is a huge motivator for people," concludes Mr. Scott. "They can see the reasoning behind it and it's not just guesswork."

To learn more about Cotati and their use of the Itron and WaterSmart technologies, take a look at our customer testimonial video: Bridge to AMI: Removing the Guesswork.


What Itron Idea Labs Brings to Itron and the Industry

Nine out of 10 Fortune 500 companies in 1955 disappeared from the list by 2016. The average tenure of a Fortune 500 company went from 33 years in 1965 to 18 years in 2012, and it is forecast to shrink to 14 years by 2026. The accelerating rate of technology development and disruptive innovation that displaces incumbents are driving this trend. 

Itron Idea Labs’ main goal is to create disruptive innovation. We start with a business idea, validate the business model, then grow the business and develop the product. After the project obtains revenues, we transition it to a business unit. 

Each project is run by an EIR (Entrepreneur-in-Residence) who is responsible for just that -- to first prove the need/value of an idea and then grow the business. We use lean startup and customer development techniques that apply to new market opportunities. These are all state-of-the-art methods used by startup companies and top incubators. Many companies have attempted to crack the nut of organic growth and disruptive innovation, but few have succeeded.

Itron Idea Labs started as an experiment and we learned from our mistakes. Through trials, iterations, failures, learnings and pivots, we are now at the point where we know how to create disruptive innovation.

An example is Itron Grid Connectivity. It started with the idea that electric utilities may need to know physical and electrical connectivity of meters to transformers and meters to feeders. Through customer discovery, we learned that the real pain point was the electrical, not physical connectivity. The next step was to identify the target customers and the value of the solution. That led to creating a revenue model that reflected that value. Then we started looking at possible solutions to that problem. We developed several solutions and we tried them with several customers. Today, we have a product based on sophisticated machine learning that determines electrical connectivity with nearly 100 percent accuracy. That product will create great value to our customers.

There are exciting times ahead of us, where we’ll create new businesses, products and opportunities. That will allow us to evolve the company to new phases and to keep up with the exponential rate of technology.


Common Challenges at the Energy Forecasters Meeting

Turning to her left, then right, Grace finds herself surrounded by strangers – men and women from across the United States; there’s even one from Canada. The prompt is simple, “What are the big issues you are facing?” Hesitantly, the first person begins to share his list. Suddenly, the monologue turns to dialogue with each person commiserating, and then adding insights to the issues.

Each year, Itron’s Annual Energy Forecasters Meeting begins with the “hot” issues round table session. This year’s 15th annual meeting in Chicago was no different. The 60 forecasters laid out classic challenges and emerging issues creating an environment of collaboration.

The major themes from the round table session include disruptive and new technologies, growth challenges, and emerging business issues. Below are some highlights of what was discussed.

Disruptive and New Technologies

As expected, the disruption caused by solar technology led the discussion. However, solar was not the only technology warranting attention. Several companies added electric vehicles, batteries, and combined heat and power to the list. Each technology discussion brought a wave of forecasting concerns ranging from market penetration and modelling techniques to load shape implications.

The advancement of AMI technology emerged as a positive challenge. With enough historical data to begin analysis, companies discussed the potential to improve unbilled calculations, revenue variance analysis, and accuracy, as well as explore end-use metering applications.

Growth Challenges

Since 2008, many companies have experienced slow load growth and/or declining average usage. While some companies struggle to explain the stagnation, others pointed to energy efficiency changes, price effects, and DSM programs. The additional effects of the Clean Power Plan, housing stock changes, and industrial automation also entered the discussion.

Slow growth also renews focus on understanding weather volatility. Growth no longer hides errant budget variance or weather normalization reporting.  Participants weighted the effect of model accuracy and normal weather definitions impact.  Amid the range of concerns, Itron conducted an informal survey of normal weather definitions allowing participants to discuss definitional impacts.

Emerging Business Issues

With 38 companies participating, the range of emerging issues was broad reflecting the changing nature of the electric and gas industries. Four major business issues highlight the discussion. First, pricing changes, including bill complexity and TOU rates, imply future consumption changes.  Second, variations among economic vendor forecasts opens the discussion on the best economic vendors. Third, flat sales are creating budget constraints requiring operating and decision making with limited resources. Finally, retail choice is emerging in some regions changing the nature of forecasting.

With the realization that these people share common challenges, Grace relaxes and finds that she is no longer surrounded by strangers, but colleagues. Fearlessly, she enters the dialogue, listens to ideas, and makes connections.   At the end of the hour, Grace sits back and quietly says to herself, “This is going to be a good conference.”


Free Forecasting Brown Bag Webinar

Now in our 11th year of hosting Itron’s popular, free hour-long internet seminars on a range of forecasting topics, we take great pride in helping our customers around the world understand the complexities of energy forecasting and talking about the latest forecasting issues and solutions. Take advantage of Itron’s expertise and experience to gain insightful knowledge and help you to improve your forecasts.

Participation is free, but prior registration is required. Each seminar lasts approximately one hour, allowing 45 minutes for the presentation and 15 minutes for questions. Seminars start at noon Pacific-time.  If you can’t attend a seminar or missed one, don’t worry! Your registration ensures that a link to the recording will be sent to you automatically approximately one week after the seminar date.

Our second seminar of the year is next Tuesday, May 23 entitled "Modeling Humidity Effects." For more information on this Brown Bag and other forecasting events, go to http://www.itron.com/forecastingworkshops.


Itron Riva Streetlights: Sunglasses On!

You might want to wear your sunglasses while reading the following...

It all started in January 2017, at the Itron Sales Conference, held in Palm Springs, California. Philip Mezey, Itron President and Chief Executive Officer, told the enthusiastic gathering how Itron was leading the industry in three areas: Electricity, Gas, and Water.

Philip then mentioned a fourth area where Itron would become a industry leader: Streetlights. Throughout the world, streetlights are being upgraded to offer edge intelligence and control, offering network-based functionality that is similar to that offered by Itron's existing line of electric, gas, and water meters.

Itron Riva Dev Mini board

Right after the January meeting, a team at Itron Idea Labs got to work. The team integrated the brand new Itron Riva Dev Mini board (from another Itron Idea Labs project) with a Neptun Streetlight Controller.  Neptun is a third party vendor of streetlights, and was eager to partner with Itron to get their streetlights onto the Itron Riva ACT (Adaptive Communications Technology) network.

It is now May 2017, and the finished product is headed to TECO, an Itron utility customer, for a pilot. Importantly, the testing will show interoperability between Itron Riva Meters and Itron Riva Streetlights. The Itron Riva Streetlights are also being sent to Avista, another Itron utility customer, for testing in their streetlight pilot -- they too will test the interoperability between Itron Riva Meters and Itron Riva Streetlights.

Itron Riva streetlight
How could the Itron Idea Labs team accomplish so much so quickly? The answer is by building upon existing Itron Riva technologies!
Itron Riva iSOM

First, the streetlight hardware builds upon Itron Riva iSOM (Itron System on a Module), the same module used in Itron Riva electric meters, Itron Riva SmartNICs, and Itron Riva CAMs (CGR ACT Modules). This module contains a powerful ARM-based processor (same type of processor as used in your cell phone), 128MB of RAM, and 256 MB of flash storage.

Second, the firmware builds upon Itron Riva MUSE (Meter Usage Software Environment), an Linux-based environment offering utility-strength robustness and reliability. A key benefit of MUSE is that Itron Riva applications can run directly on the device -- edge intelligence in action!

Itron Riva Dev Mini inside the streetlight

Third, the networking stack builds upon Itron Riva ACT (Adaptive Communications Technology), the same networking stack used by Itron Riva electric, gas and water meters. The result? Itron Riva Streetlight can use the same set of networking software as Itron Riva Meters --- for example, the Cisco Field Network Director (FND) can be used with Itron Riva Streetlight for network management.

Itron Riva Streetlights registered in Field Network Director (FND)

In summary, the Itron competitive advantage is clear: Itron Riva is compatible with Electricity, Gas, Water, and... (sunglasses on!) Streetlights!

Read our full press release: Itron Releases Next Generation Itron Riva™ Development Kit for Faster Path to Innovation.


A Grand View

The view from the 96th floor of the John Hancock Center is amazing. From here, cars are merely dots between the straight rows of lights and people are absent. Even the contours formed by the cluster of high-rise buildings pale in comparison to the largeness of Lake Michigan and the vastness of the city.

Itron’s 15th Annual Energy Forecasting Meeting provided a similar perspective pushing aside our small daily challenges to see the grand view of the energy forecasting world. This year’s meeting was held in Chicago from April 26-28 where 60 attendees from 38 companies spent three days discussing the implication of the economy, new technologies, prices, energy efficiency, and normal weather on the electric and gas forecasting world.

The View. The broadest pictures of the electric industry were covered by Mark Quan and Mike Russo (Itron), Steve Cochran (Moody’s Analytics) and Erin Boedecker (Energy Information Administration). Mark and Mike stepped back and showed historical growth of the industry and preliminary projections based on Itron’s latest Benchmarking and Trend survey. Steve presented the current state of the U.S. economy and forward-looking risks, and Erin provided details about the EIA’s latest forecast for the residential and commercial sectors which go through 2050. These presentations painted a picture of the horizon and direction for the electric and gas industries.

The Contour. The current challenges of the industry shape the horizon. These challenges include the penetration of AMI data, behind-the-meter technology such as solar and batteries, and changing weather patterns. Andy Sukenik, Mike Russo, and William Marin (Itron) discussed solar penetration, solar shape modeling, and battery technology. Kristin Larson (Storm Geo) showed alternative climate normal calculations, and Dennis Kelter (ComEd) addressed the uses of AMI data.

The Details. Within the broad view and the contours of the industry, several attendees addressed specific issues and techniques useful in our current situation. Bo Xing (Salt River Project), Abdul Razack (Nevada Power), and Reynaldo Guerra (CPS Energy) showed modeling techniques including peak calibration, model selection tests, and incremental change techniques. Andrew Trachsell (IESO) and Chad Burnett (AEP) discussed time-of-use pricing and price elasticities, and Markus Leuker (DTE) showed the power of daily tracking and weather normalization with AMI data.

With the broad array of topics and multiple perspectives, attendees found the discussion challenging and informative. When reflecting on the experience, Nicole Fan (Alectra Utilities) said, “The meeting was a great success; the topics have been expanded so much including regulations, pricing, economics and new technologies. I enjoyed it a lot.”

I agree. The view is amazing.

 

 


Itron + Scholarship America = Dream

In the next five to 10 years, more than 500,000 employees in the energy industry are expected to retire , according to the Committee on Energy and Natural Resources. The average age of an employee in the energy industry is more than 50 years old. Our incoming workforce will need the science, technology, engineering and mathematics (STEM) skills necessary to fill these jobs. That is why it is important to help the next generation gain the skills and education they need to work in the energy industry.

This year, Itron is proud to once again partner with Scholarship America, a non-profit that helps further education through scholarships and additional support to make postsecondary success possible for all. This partnership, which began six years ago, helps support STEM students with scholarship and educational aid.

As a part of this partnership, Itron donates the funds for items that are raffled at the AGA Operations Conference and Biennial Exhibition. This year, the prizes are a Microsoft Surface Pro 4, a DJI Phantom 3 drone and two Alaska cruise tickets . The funds raised through the raffle benefit STEM recipients of the Dream Award—a renewable scholarship that students can begin receiving as a sophomore.

Itron is proud to help inspire and create the next generation of resourcefulness through this partnership.

Join us for a Facebook Live from AGA with Marian Marchese, director of strategic intelligence at Scholarship America, on Tuesday, May 2 at 6:00 p.m. EDT to learn more. Winner(s) will be drawn at the Itron booth (#601) on Wednesday, May 3 at 6:30 p.m. EDT.

You can also learn more about Itron’s partnership with Scholarship America here.


An Exciting Time for the Gas Industry

Next week, the natural gas industry will convene in Orlando for the American Gas Association Operations Conference to share knowledge, ideas and practices for safe, efficient and reliable delivery of gas.

In the Itron booth (#601) this year, we will be demonstrating ways we are bringing intelligence to utilities, including new ways to enhance safety and perform maintenance more efficiently. I invite you to come and see hands-on demos of how we are improving system integrity with our gas quality, cathodic protection and pressure management solutions.

In addition to our latest technology and solutions, our booth will also feature key partners who are bringing new capabilities to our OpenWay Riva™ solution , as well as innovations from Itron Idea Labs. The objective of Itron Idea Labs is to discover new opportunities for effective energy management utilizing Itron Riva™ technology. Stop by to learn how Itron Idea Labs accelerates innovation by engaging with our customers early and often, delivering new and strategic solutions at a transformative pace.

Look for my article in the upcoming issue of American Gas to learn more about the gas utility of the future.


Behind-The-Meter Solar Generation and Real-Time Load Forecasting – Part 3

In my blog posting “Behind-The-Meter Solar Generation and Real-Time Load Forecasting,” I presented four load forecasting challenges that have arisen as the result of deep penetration of behind-the-meter (BTM) distributed energy resources (e.g. solar generation), time-of-use rates, demand response programs, BTM storage, and electric vehicle charging.  These four challenges are:

  • Challenge 1.  Growing Disconnect between Measured Load and Demand for Electricity Services
  • Challenge 2.  The Relationship between Measured Load and Weather is Becoming Cloudy
  • Challenge 3.  Increased Load Forecast Errors and Error Volatility
  • Challenge 4.  Constructing Load Forecast Confidence Bounds.

From a load forecasting perspective, these challenges are leading to an evolution in the way we develop models and forecasts.  In my second blog on this topic, I introduced the steps that Itron is taking to address the first three challenges.  In this blog, I introduce an approach for constructing load forecast confidence bounds that leverages a modeling technique applied widely in financial markets.

Why do we need confidence bounds?  From the perspective of scheduling and dispatching generation, the confidence bounds provide valuable input to the decisions around how much generation should be held in reserve in case observed load conditions deviate from forecasted load conditions.  The concept of spinning reserves addresses the operational reality that the unexpected is expected.  But holding excess generation in reserve comes at a cost.  In the ideal world, these costs would be minimized if the unexpected could be accurately forecasted.  As load forecasters, we may not be able to forecast all the unexpected events that could occur (e.g. drivers taking out transformers, squirrels jumping upon sagging transmission lines), but we can quantify the load forecast uncertainty.

What are the sources of load forecast uncertainty?  First, the very loads that the load forecast models are built on are subject to measurement error.  In many cases, what a system operator sees as load is a result of detailed calculation rather than the aggregate sum of metered demand.  Specifically, load is computed as Total Generation + Net Imports - Transmission and Distribution Losses.  In some systems, this calculation is straightforward and results in a solid estimate of load.  In other systems, the network topology is complex and forming a clean definition of what points are “in” the system versus what points are part of another system is subject to redefinition.  You might be surprised by the number of times we have been called in to fix a model only to discover that the load forecast deviations were due to a redefinition of load that was driven by the addition of a new transmission line to the network topology.

Second, with deep penetration of BTM solar generation what we measure as load is a proxy for the demand for electricity services.  Because BTM solar generation is volatile, this volatility is translated to an increased volatility of measured load.  This in turn increases load forecast uncertainty because the target we are trying to hit is bouncing around more frequently and in growing amplitude.

The third source of load forecast uncertainty is weather forecast uncertainty.  This will not come as a surprise to anyone that makes a living based on vagaries of weather forecasts.  Weather forecast error comes in a couple flavors.  First, most weather forecasts under forecast hot temperatures and over forecast cold temperatures.  Essentially, most weather forecasts tend to ride in between the two extremes.  Some of this is a result of a weather vendor taking a weighted average of several high level model forecasts, where the high level model forecasts reflect different weather scenarios.  The process of averaging pulls the point forecast away from the extremes and toward the center of the weather scenarios.  Second, because weather forecasts are a derivative of a model there arises the possibility that the forecasts have a systematic model bias at certain times of a day.  For example, I have seen the case where the weather forecast was consistently 2 degrees warmer at 6am and 7am regardless of the rest of the weather forecast.  This consistent over forecast is a derivative of the weather forecast model.  Finally, there are catastrophic weather forecast error which is when all the high level weather models simply get it wrong by 5 degrees (F) or more.  Another flavor of a catastrophic weather forecast error is missing the timing of a weather front as it moves through a service territory.  Catastrophic weather forecast error is the hardest to control for, but unfortunately these are the days that are remembered by management because those are the days the load forecast was way wrong.

The final source of load forecast uncertainty is the load forecast framework which includes both the load forecast models and any manual interventions that are applied to the model forecasts.  Despite what people think, statistical models are subject to error.  Simply put, there are too many things that drive the demand for electrical services to model individually.  The art of load forecast model development is to capture systematic load patterns with the goal that what is left unexplained has the signature of random noise.  Large well behaved systems are easier to model because the law of large numbers works in our favor.  The random, non-systematic behavior of millions of households and businesses when aggregated together tends to smooth out to repeatable and predictable load patterns.  In these cases, what is left unexplained tends to be random noise.  Further, that noise as a percentage of total load tends to be less than 0.50%.  When load forecast models are fitted to smaller systems or to loads with more granular geographic specificity the law of large numbers breaks down.  As a result, the ratio of noise to repeatable load patterns goes up.  This, in turn leads to greater load forecast uncertainty.

What makes good confidence bounds?  Good confidence bounds should quantify the load forecast uncertainty with respect to each source of load forecast uncertainty.  Further, good confidence bounds should be sensitive to the forecast horizon.  For example, if the forecast horizon is five minutes from now and we are utilizing a highly autoregressive model, the main source of load forecast error will be measurement error rather than temperature forecast error.  In contrast if the forecast horizon is day-ahead then the temperature forecast error grows in importance.  Intuitively, we expect the confidence bounds to widen the further into the forecast horizon, resulting in what would look like cone shape bounds.  We also expect that the confidence bounds would be time-of-day, day-of-the-week, and potentially season dependent.

Together, this suggests taking a modeling approach to developing the confidence bounds.  It turns out that Dr. Robert F. Engle who won a Nobel Prize in Economics for his seminal paper "Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation". Econometrica. 50 (4): 987–1008. 1982 describes a modeling approach that we can leverage to forecast the desired load forecast error variances.  Dr. Engle’s Autoregressive Conditional Heteroscedasticity (ARCH) model can be written generically as follows:

 

 

 

 

 

 

 

This may not look like anything special, but what if we construct the  by taking the sequence of hour-ahead load forecast errors and squaring them.  Now, I can set up a simple regression model of the hour-ahead forecast error variance as:

Here, we have a one period autoregressive model where the current hour ahead load forecast error variance is a function of the prior hour, hour ahead load forecast error variance.  Of course, we could extend the autoregressive terms by including the prior two, three, four, five, …, up to k hour values of the hour ahead load forecast error variances.  This is the AR part of ARCH.

The Conditional Heteroscedasticity (CH) part of ARCH can contain explanatory variables that would allow the forecast error variance to vary by the time-of-day, day-of-the-week, season and potentially the weather forecast.  To illustrate, let me build out an ARCH model for the one-hour ahead load forecast error variance for the load forecasts made at 7am for 8am.  The ARCH model could look like the following:

Like any regression model, the specification of the ARCH model is wide open.  With some experimentation, the right set of autoregressive and heteroscedastic terms can be decided upon.  The result will be a model that can be used to forecast the load forecast error variance for the one hour ahead forecast of 8am loads that was made at 7am.  In practice, this means there will be 24, load forecast error variance models for each forecast horizon (say 1 hour ahead through to 24 hours ahead).

To incorporate Weather Forecast Uncertainty, we can extend the model as follows:

Here, we added the estimated Weather Forecast Error Variance as another explanatory variable in the model.  In this particular case, the Weather Forecast Error Variance would be based on the one-hour weather forecast errors.  The Weather Forecast Error Variance itself would be derived from a second ARCH model of weather forecast error variances.

In a similar fashion, we could extend the ARCH model to include forecasts of BTM solar generation and the associated BTM solar generation forecast error variance.  An example would be as follows:

As you can see, there are no limits to the variables that you can include in the ARCH model.  The ARCH framework provides a general framework for forecasting load forecast error variances that allow us to go beyond simple +/- two model standard errors to functions of forecasted weather and solar generation.

The challenge we face is deciding on the right set of explanatory variables to include in our ARCH models.  This is where experimentation is required.  I recommend starting with the heteroscedastic pieces: day-of-the-week, season, weather, BTM solar, predicted weather error variances, and predicted BTM solar error variances.  Once you have found a set of explanatory variables that explain the heteroscedastic error variance, you can start layer in the autoregressive terms.

With the deep penetration of BTM solar generation, battery storage, and other distributed energy resources, what is measured as load is no longer a perfect measure of the demand for electricity services.  In my first blog, I identified a list of load forecasting challenges that result from this erosion of measuring the demand for electricity services.  From a load forecasting perspective, these challenges are leading to an evolution in the way we develop models and forecasts.  In my second blog, I presented solutions for improving short-term load forecast performance.  In this third blog, I introduce a general modeling framework for constructing load forecast confidence bounds that incorporates weather and BTM solar generation forecast uncertainty.

I want to leave you with one final thought.  The salient operating feature of the new distributed energy resources (i.e. Solar PV and Other Generation, Storage, TOU Pricing, and Demand Response) is their variability.  If all these technologies worked on pre-defined or predictable operating schedules, the forecasting challenge would be similar to incorporating long-run trends in energy efficient appliances which lower loads, but do not add load variability.  It is the added load variability that comes with the new distributed energy resources that is causing the erosion in load forecast performance.  As a result, the load forecast modeling problem extends beyond capturing average load swings to include capturing the volatility around the average loads.  Put simply, it is no longer sufficient to forecast the mean, but we need to also forecast the variance.  I believe, in time, the forecast of the load variance will be more important from a system operations point of view than the point forecasts operators get today.  This may seem like a small step, but like most things in life, big things come from small steps.  “That’s one small step for man, one giant leap for mankind”, Neil Armstrong circa 20 July, 1969.

Download and read the full white paper, here.