Do Solar Panels Need to be Cleaned?

It seems like my solar photovoltaic (PV) panel monitoring service, Enphase, started sending more emails recently. Or maybe because I recently wrote a blog (Residential Lighting Efficiency Really Does Make a Difference) during my PV true up period, I probably had solar on the brain. I just received an email recommending that I wash my panels before winter, and it made me start to wonder – my car is parked outside and gets filthy super quick, so do I need to climb up on the roof and clean the panels or hire someone to do it? Living in San Diego, it is true that there’s not much rain and the panels have been up there for four years now without a proper bath. It totally makes sense that general dust, dirt and fire-related ash would make them less efficient, right? But why would I do it before winter when it rains more? How on Earth do you wash them? Will I fall off the roof? All of this is a little counter-intuitive to what the dollar amount on my latest true up indicated, and although I looked at my lighting efficiency, I didn’t really look at the solar production over the years. I never considered that there might also be losses due to grime.

Some research indicates that cleaning your solar panels leads to small improvements in output, yet others say you should clean them twice a year. One site even suggested a 35% loss after two years, but it turns out that all of the cleaning recommendations tended to be from solar panel cleaning companies or from quoting stats via cleaning companies. Then I stumbled on a study by the Jacobs School of Engineering at UCSD that made me feel much better about not having given my panels any attention since they were installed. According to their research, due to the angle that the panels are mounted and being on a roof, they found that rain did a fine job of cleaning the panels as long as there are no bird droppings.

Again, having been part of the forecasting team for so long, I also had to look at the data and graph it:

Surprisingly, there has been a slight annual uptick in production (is that global warming?!). In any case, I definitely agree that my data and the research available are in alignment. I’m good with not cleaning my solar panels. If I wanted to increase my production a smidge during the summer, when there isn’t any rain in sight, I could clean them. But I don’t think it’s worth the effort, and hiring someone definitely would not offset the cost.


Join Our Brown Bag: Community Choice Aggregation Load Forecasting

The last Itron Forecasting Brown Bag of 2020 is on Tuesday, Dec. 8 and is entitled "Community Choice Aggregation Load Forecasting." During this free webinar, Andy Sukenik will present some background on Community Choice Aggregation (CCA) and will discuss tips on how to create forecasts for the short and long term.

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 cannot attend a seminar or if you 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, plus you will receive a link to watch any of this year’s previous webinars.

Register today at www.itron.com/forecastingworkshops.

Itron’s Forecasting group has conducted webinars on a variety of forecasting and load research-based topics for many years. All of our past webinars were recorded and are available in a YouTube library.


October 2020: Trends in Estimated Load Impacts of COVID-19 Mitigation Policies on European and North American Electricity Consumption

March through October 2020

As previously discussed in the first of this series on April 13, as lockdown policies are enacted to help reduce the spread of COVID-19, the Itron Forecasting Team is leveraging publicly available hourly load data for most North American Independent System Operators (ISOs) and a select set of European countries to build a picture of the load impacts by region. To assess the load impact of COVID-19 mitigation strategies, actual loads when many of these policies began are compared to baseline loads without COVID-19 policy impacts.

Across Europe and North America, the biggest estimated load reductions occurred in April with an estimated reduction in average daily load between -12.3% and -7.2%. In recent months, the combination of reduced lockdown restrictions and weather has led to no apparent load impact in Europe. In contrast, North America loads continue to fall below expectations that are not adjusted for prevailing weather.

For a detailed summary of the estimated load impacts, go to the forecasting website to download the latest COVID-19 Load Impact memo.

The Itron Forecasting Team will continue to post updated summary blogs and corresponding memos on the trends.

Subscribe to our blog to be notified of new posts and contact us at forecasting@itron.com if you have further questions.


Residential Lighting Efficiency Really Does Make a Difference!

It’s hard to believe that it’s been 4 years since we installed the solar photovoltaic (PV) panels on our house. As a solar customer in the San Diego Gas & Electric (SDG&E) service territory, we don’t get monthly bills, just one true-up bill on the anniversary date of your start date. We are on a November to November true-up period and just received our fourth true up bill and I was a bit surprised.

I wrote a PV blog after our first year of the solar journey, but here’s a quick overview. We live in San Diego in a 1,500 square foot house and my son likes to keep our house at freezer-like conditions. In 2015, our annual electric bill was about $2,700 (9,000 kWh). Summer bills were in the $400-500 range (800-900 kWh) and always hit tier 2 rates. We ended up with a 7.25 kW system with 25 panels and were able to get into the NEM Successor Tariff (Schedule NEM-ST, NEM-ST or NEM 2.0) where non-bypassable charges are assessed. There was a nominal interconnection fee and we were grandfathered into the tiered rates for 5 years after our system went live. We have 1 more year before we need to move to the TOU rates, which will probably prompt another set of blogs.

In 2017, our first solar year, we owed $48 and we added an electric vehicle (EV) to the mix in December. In 2018, we owed $258 which included a $500 SDG&E EV climate credit (EVCC) and I started using a free EV charging station by the office in October. In 2019, I was doing the majority of my EV charging at work and receiving an $850 EVCC from SDG&E, and we ended up with an $800 credit overall. The latest bill had no EVCC and was only $101.26.

I’m not complaining, but I was surprised it wasn’t higher. I have been working from home since March, so there is additional electric load in general; my son added a larger TV to his room that seems to be on all of the time, in addition to his computers; almost all of my EV charging has been at home, albeit less driving in general due to the pandemic; and, it was a really hot summer in San Diego so the air conditioner ran way more than normal, hitting tier 2 rates a few times.

Working with Itron’s Energy Forecasting Group (EFG) residential sector and end-use data for so long made me a little curious about our usage. What had changed? During our remodeling project toward the end of last year, we installed a bunch of new LED canned lights in our main room, replacing a beam of incandescent lighting that is now hardly used. The adjacent kitchen lights were almost always turned on.

This is our beam of incandescent lighting, and you can see one of the new LED lights on the ceiling.

When you add up all those cute little incandescents, it turns out that the beam uses 663 watts. Adding in the old kitchen lights and using a conservative estimate of 6 hours of use per day, translates into roughly 1,600 kWh/year. Holy cow! They looked so cool when we put them up and I knew they would suck a bunch of energy, but I didn’t do the math. That’s more than 4 new refrigerators! Our new set up has 20 new LED canned lights which are rarely all on, but if they were, that would only be 438 kWh/year. Our tier 1 rate is 28 cents per kWh, so that is a minimum of $310 savings for the year.

It is good timing on this analysis because one of our next home projects is to replace our original single-paned windows. The residential geek in me was planning to replace them with overpriced, super high energy-efficient low U-Factor ones with whatever gas inside. I already have a few quotes but I am now reconsidering the efficiency level needed because just switching to more energy efficient lighting has brought us pretty close to break even with the EV. The ENERGY STAR program was founded in 1991, so ANY new window will be an improvement in efficiency from the current windows that were installed in 1986.


Watch: Short-term Load Forecasts that Account for COVID-19 Mitigation Policies

During last week’s virtual Itron Utility Week, did you miss the forecasting team’s presentation from Frank Monforte entitled Short-term Load Forecasts that Account for COVID-19 Mitigation Policies? Frank discussed learning how to improve day-ahead and intraday forecasts in the face of existing lockdowns and the future re-opening of economies. His presentation ties in with the continued COVID-19 mitigation blog series that you may have been following throughout the last few months.

If you missed this discussion, you can still register to watch the session and any of the other great Itron Utility Week presentations available on demand.

Be sure to subscribe to our blog to be notified of new posts. Contact us at forecasting@itron.com if you have further questions.


Through September 2020: Trends in Estimated Load Impacts of COVID-19 Mitigation Policies on European and North American Electricity Consumption

As previously discussed in the first of this blog series on April 13, as lockdown policies are enacted to help reduce the spread of the coronavirus disease (COVID-19), the Itron Forecasting Team is leveraging publicly available hourly load data for most North American Independent System Operators (ISOs) and a select set of European countries to build a picture of the load impacts by region. To assess the load impact of COVID-19 mitigation strategies, actual loads when many of these policies began are compared to baseline loads without COVID-19 policy impacts.

Across Europe and North America, the biggest estimated load reductions occurred in April with an estimated reduction in average daily load between -12.3% and -7.2%.  In recent months, the combination of the relaxing of lockdown restrictions and hot weather has led to no apparent load impact in North America, but a continued small impact in Europe.

For a detailed summary of the estimated load impacts, go to the forecasting website to download the latest COVID-19 Load Impact memo.

The Itron Forecasting Team will continue to post updated summary blogs and corresponding memos on the trends.

Subscribe to our blog to be notified of new posts and contact us at forecasting@itron.com if you have further questions.

Estimated Daily Average Energy Impact Wedge


Reducing the Potential for Generation Redispatch

The pan-European electricity market enables competitive wholesale and retail exchange of electricity while ensuring the security of the electric grid. The market design supports both day-ahead and intraday electricity exchanges that require accurate electricity consumption and generation forecasts to ensure safe and cost-effective grid operations. The day-ahead market establishes an initial schedule of generation resources required to meet the anticipated day-ahead electricity consumption. The intraday market provides for electricity exchanges to clear current day energy imbalances. As a last resort, Transmission System Operators (TSOs) redispatch generation to relieve congestion bottlenecks and maintain system security.

Over the period of 2015 through 2019, the cost of Germany’s generation redispatch averaged a little under €1 billion per year. To help manage the cost of redispatch in the face of the rapid deployment of renewable generation, Germany’s Redispatch 2.0 guidelines allow German TSOs to throttle down grid-connected wind, solar, and combined heat and power generation if other redispatch options are more costly. To assess all possible redispatch options, German Distribution System Operators (DSOs) are required to submit to their respective Transmission System Operator (TSO) day-ahead and intraday forecasts of generation (> 100 KW) and consumption within their distribution system. The DSO forecasts are critical inputs to network models that redispatch grid and non-grid connected generation (for example, centralized dispatched load control) to meet consumption subject to transmission and distribution system operating constraints.

Day-ahead and intraday consumption forecasts are developed generally using techniques from two broad classes of load forecast frameworks:

  • Equation-based Approaches: Under this approach, parametrized equations are used to predict loads as a function of calendar (e.g., day-of-the-week, season, holidays, special event days), economic (e.g., operating schedules, employment levels), weather (e.g., temperature, humidity, wind speed, precipitation) and solar (sunrise/sunset times, observance of daylight savings, solar irradiance) conditions. Historical load data are combined with historical measurements of explanatory variables or features to estimate the parameters of the equations. Once the parameters are estimated, forecasted calendar, economic, weather and solar conditions are passed into the equations to form the load forecast. Techniques that fall into this class are multivariate regression, advanced neural networks, and support vector regression, among others.
  • Classification-based Approaches: Under this approach, historical load data are classified based on calendar, economic, weather and solar conditions. Days with similar conditions as the forecast day are then averaged to form the load forecast. Techniques that fall into this class are Gradient Boost, Random Forests, and Like Day Lookup, among others.

A useful way of thinking of a load forecast is to look at it as an average of historical load data under similar calendar, economic, weather and solar conditions as the day being forecasted. From the perspective of an event like COVID-19, the challenge is that the historical data upon which the load forecasts are constructed are not under the same economic conditions as those prevailing under lockdown mandates imposed by countries to curb the spread of COVID-19. Although lockdown policies vary across Europe, in general, the policies have led to school closures and reduced operations or closures of non-essential businesses. Many other businesses have a large portion of their employees working remotely from home. The net effect is a shift of weekday electricity consumption from the nonresidential sector to the residential sector. For DSOs and TSOs operating in countries with lockdown mandates in place, there has been an evolution of the system load shape toward a residential load pattern. Day-ahead and intraday load forecast frameworks that do not adapt to the shift in load patterns will realize an erosion of forecast performance. This adds pressure on system operators to continually evaluate their generation dispatch plans to ensure all energy imbalances are cleared.

Dr. Frank Monforte, Director of Forecasting Solutions at Itron, recently wrote a white paper that presents a framework for improving day-ahead and intraday consumption forecasts in the face of existing COVID-19 lockdown mitigation policies, as well as when the lockdown policies are relaxed to allow for a re-opening of Europe’s economies. Historical hourly load data for 10 countries spanning the pan-European electric grid is used to demonstrate how load consumption patterns have changed because of COVID-19 lockdowns. The analysis of the data suggests how existing load forecasting frameworks can be adapted to prevent erosion of forecast performance. Download a copy today!


If It Were Not for the Data, Everything Would Be Great

I recently developed a set of 15-minute load forecast models when I noticed something odd in the pattern of the MAPEs (Mean Absolute Percent Errors) from interval to interval. After unsuccessfully trying a few of my usual tactics to resolve the problem, I presented the following figure to one of my colleagues:

Not wanting to let the proverbial cat out of the bag myself, I said, “Do you see what I see?” After a moment of reflection, he indeed saw what I saw: the MAPE of every fourth interval jumps up. That is not cool at all.

Here is a bit of background. This set of models is highly auto-regressive. That is, each model uses lagged loads from the prior four intervals as driver variables (as well as the typical calendar and weather variables). The model for 1 a.m., for instance, depends upon loads from 12:45 a.m., 12:30 a.m., 12:15 a.m. and midnight. This is a common approach for very short-term forecasting. Not to put too fine a point on it, but I have used this same approach many times without incident.

In an effort to identify and to ameliorate the problem, I tried a few things. I had thought that the weather data might not be aligned correctly with the load data. I removed all of the coincident weather variables and re-worked the models so they all have the same specification (except for the relevant lagged-loads) with the same time-of-day weather variables. When that failed to resolve the problem, I was sufficiently convinced that the weather was not the issue. I also tried manipulating the lagged intervals, excluding the fourth lag (i.e., the 1-hour back) variable and other similar inclusions and exclusions of various lags. That too proved to be like an empty apple bag – entirely fruitless.

This is not the data for an individual customer or even a small municipality. In fact, the data is for a transmission grid, accounting for the electricity consumption of millions of customers. After some deeper contemplation, consultation and reflection, we decided that this is not the way real load data behaves! There is no reason that models at the top of the hour should be systematically worse than models from the :45- minute mark or the :15- minute mark. This data was clearly not measured, but rather, it was constructed. Further, it was constructed in such a way as to create a bias in the resulting values. This is not obvious when observing the load data itself or even when viewing the ramp-rates (i.e. the interval-to-interval deltas).

We dug deep into our mental archives for a solution, where we found our old friends Savitzky and Golay. (Let me tell you: those guys knew how to rock!) I will leave it to you to review their work. Essentially, their idea is to apply a smoothing algorithm to the data. In this case, we applied a 5-period centered polynomial weighted average. Our goal is to identify the ‘signal’ in the data and not to be distracted by the “noise.”

The following figure displays one day of 15-minute data that is both smoothed (in blue) and raw (in red). A few observations are circled to illustrate that the smoothed data cuts through the center of the noise in the raw data, thereby creating a less volatile series.

The 5-period centered smoothing algorithm is generalized as:

Where:

  • i = interval. In this case, there are 96 intervals per day because the data has a 15-minute frequency.
  • -3, 12, and 17 are the “convolution coefficients” as derived by Savitsky and Golay.

The following is an example for the 12 p.m. interval:

I will spare you the details of the mechanical processes we used in MetrixND and MetrixIDR to perform these calculations and I will cut directly to the climactic fight scene. We ran the identical models, but this time the data had first been smoothed. The following figure presents the MAPEs from the original models in blue and the new models with smoothed data in orange. There are two salient points:

  1. The dramatically higher MAPEs at the top of the hour are largely eliminated by smoothing the data.
  2. The models with the smoothed data are systematically better than the original models, as evidenced by the orange bars being lower than the blue bars for the same intervals. In fact, the average MAPE for the 96 intervals was reduced from 0.50% to 0.27%. Cutting the average MAPE in half is no small feat.


Many of the changes we make in models tend to have marginal impacts. Further, the changes often have the effect of improving some intervals while degrading other intervals. I rarely see changes in models that are both dramatic and systematic. When those two criteria are met, I declare victory.

What can we learn from this? First, we can use the models themselves to inform us about the data. Remember, it was the pattern in the MAPEs that led us down this road. Second, we cannot necessarily trust the data itself. The numbers are not provided by a ‘data fairy’ who magically and lovingly delivers pristine data to us under our pillows while we sleep. In most cases, we do not know the story behind the collection or calculation of the data. We must be vigilant and skeptical. Further, we can use the tools at our disposal to address many of the issues that we face.


Through August 2020: Trends in Estimated Load Impacts of COVID-19 Mitigation Policies on European and North American Electricity Consumption

March through August 2020

As discussed in the first of this blog series on April 13, lockdown policies were enacted globally to help reduce the spread of COVID-19. The Itron Forecasting Team is leveraging publicly available hourly load data for most North American Independent System Operators (ISOs) and a select set of European countries to build a picture of the load impacts by region as a result of these lockdown policies. To assess the load impact of COVID-19 mitigation strategies, actual loads when many of these policies began are compared to baseline loads without COVID-19 policy impacts.

Across Europe and North America, the biggest estimated load reductions occurred in April with an estimated reduction in average daily load between -12.3% and -7.2%. In recent months, the combination of the relaxing of lockdown restrictions and hot weather has led to no apparent load impact in North America, but a continued small impact in Europe.

For a detailed summary of the estimated load impacts, go to the forecasting website to download the latest COVID-19 Load Impact memo.

The Itron Forecasting Team will continue to post updated summary blogs and corresponding memos on these trends.

Subscribe to our blog to be notified of new posts, and contact us at forecasting@itron.com if you have further questions.


Annual Energy Forecasting Survey Results

Itron’s forecasting group has compiled the results of their Annual Energy Survey to assess industry growth expectations and forecast accuracy for electricity and natural gas.

Survey participants will receive the summary report, however, our next Brown Bag event on Sept. 15 is your chance to hear about the results.

To register for this session, go to www.itron.com/forecastingworkhops.


The Logit Function: A Tool for New Things

With emerging technologies such as electric vehicles and photovoltaics, developing a forecast is challenging due to the lack of history. For many, using a Bass Diffusion Model is the solution. While this model is widely used for forecasting new product sales, it is rarely seen in the electric industry due to the infrequency of new technologies.

The Bass Diffusion Model is shown below:

Where:

  • “q” is the Coefficient of Imitation. The Coefficient of Imitation is the likelihood that someone will start using the product because of internal influence such as “word-of-mouth.”
  • “P” is the Coefficient of Innovation. The Coefficient of Innovation is the likelihood that someone will start using the product because of external influences such as media coverage.
  • “m” is the Market Potential. The Market Potential is the theoretical total number who will use the product.

While the formula may be intimidating, the result is simple. The equation creates an S-shaped forecast with the parameterization controlling the speed of adoption and technology saturation.

While MetrixND doesn’t have a function called “Bass Diffusion,” that doesn’t mean that you can’t create it in a transformation variable. The picture above was created using the following parameters from a microwave oven adoption model (http://www.bus.iastate.edu/zjiang/research/vbm_ijrm.pdf).

  • p = 0.00071
  • q = 0.3444
  • m = 1

However, MetrixND provides a similar result in using the “Logit” function. This function creates an S-shaped curved through two data points using the syntax below.

LOGIT (Year1, Period1, Value1, Year2, Period2, Value2)

In this function, the S-shaped curve goes through Value1 in Year1, Period1 and Value2 in Year2, Period2 where the numerical values for Value1 and Value2 are greater than 0.0 and less than 1.0. The logit function is defined by the equation below.

Once again, if the formula is too intimidating, just look at the results. Using the following parameterization, the result is almost identical to the Bass Diffusion Model shown above.


Using a Bass Diffusion Model or Logit function creates S-shaped curves that replicate real-world technology adoption patterns.

Either of these curves may be used in a regression model to calibrate the shape to historical adoption patterns. By exploring different curve parameterizations, a regression model can be created that fits the historic technology adoption and projects the future adoption along a classic adoption curve.

Below are the Bass and Logit curves applied to historic photovoltaic adoption from 2008 through 2016 for one service territory. Both trends fit the historic data and present an adoption patterns that accelerates in the future. The slight difference in the forecast results from the variation in the mathematical equations.


The next time you forecast a new technology such as electric vehicles or photovoltaics, consider using the Logit function.


Aug. 5 Update: Latest Trends in Estimated Load Impacts of COVID-19 Mitigation Policies on North American and European Electricity Consumption

This data is for March through July 2020

As discussed in the first of this blog series on April 13, the Itron Forecasting Team is leveraging publicly available hourly load data for most North American Independent System Operators (ISOs) to build a picture of the load impacts by region as lockdown policies are enacted to help reduce the spread of COVID-19.

In addition, the Itron Forecasting Team is now including the Southwest Power Pool (SPP) and a select set of European countries to build a picture of the load impacts by region. To assess the load impact of COVID-19 mitigation strategies, actual loads when many of these policies began are compared to baseline loads without COVID-19 policy impacts.

Across Europe and North America, the biggest estimated load reductions occurred in April with an estimated reduction in average daily load between -7.2% and -12.2%. Average daily consumption began to return to normal levels starting in June and continuing through July.

For a detailed summary of the estimated load impacts, go to the forecasting website to download the latest COVID-19 Load Impact memo.

The Itron Forecasting Team will continue to post updated summary blogs and corresponding memos on these trends.

Subscribe to our blog to be notified of new posts and contact us at forecasting@itron.com if you have further questions.


DON'T MISS A POST!
I agree to have my personal information transfered to AWeber ( more information )
Opt in to receive notifications when a blog post is published. Don't miss the thought leadership, insight and news from Itron.
We hate spam. Your email address will not be sold or shared with anyone else.