Let us show you how we can help create smarter, more sustainable facilities.
We look forward to speaking with you.

Loading...

M&V Modeling with Limited Data

In an earlier blog post, we detailed how Verdigris helped Juan, the General Manager of the Orchard Hotel Group in San Francisco, perform measurement and verification (M&V) of a retrofit on his cooling tower. The old 2-step fan drive was upgraded to a variable frequency drive (VFD). The upgrade was based on a quote with generous savings numbers, and an expected positive ROI within 1.3 years. Verdigris was tasked with answering the classic question of whether or not a building upgrade was worth the money. Since we engaged with Juan only one month prior to the upgrade, our main concern became: how do we provide a solid M&V analysis with only one month of data?

Before doing a formal M&V analysis, we did quick back-of-the-envelope calculations to get some preliminary results. The device-level submetered energy data for the cooling tower already revealed a dramatic difference. The monitoring from June 1st through July 6th, prior to the retrofit, showed a total consumption of 699 kWh. After making a broad assumption that cooling tower use in the summer can be used as a daily average, an annual consumption of 7,087 kWh can be extrapolated for the year. In other words, the annual consumption of the old 2-step drive extrapolated from ground truth data was about 20% that of the contractor’s estimate of 35,600 kWh.

The quote claimed the VFD would save more than two times as much energy as what the old drive was already consuming in total. Clearly, more analysis was required.

The Linear Regression Model

Having only a single summer month meant the seasonality of cooling tower use throughout the year would be hard to model. To boost model performance in light of having only a short time range of data, we tested adding in more features for prediction than just outdoor air temperature, which is the typical bread and butter for HVAC M&V work.

Features used in the final model:

  • Outdoor air dry bulb temperature
  • Outdoor air dewpoint temperature
  • Chiller energy consumption, from Verdigris sensors
  • Corresponding time of day, polarized onto a unit circle
  • Corresponding day of week, polarized onto a unit circle

 

Weather data was obtained via Forecast.io’s Dark Sky API, as well as observations from the San Francisco International Airport weather station. Time of day and day of week were polarized, such that the time periods at the end of a day/week, in numerical value, were closer to the timer periods at the start of the day/week. In other words, hour 23:00 was very similar to hour 00:00, rather than being 23 increments above it. This helps the model treat times of day/week more similarly, the closer they are in the real world, rather than based on a numerical representation.

Orchard-Hotel-MV.png

Together, these features were used to predict cooling tower fan energy consumption on an hourly basis. Since the training data set was warmer summer weather, it led the model to some negative energy consumption during colder days in winter. A lower bound of zero was set to override any negative values, effectively telling the model to turn the fan off instead of assuming it could run in reverse. While we had up to minutely energy data from our sensors to use in the analysis, it was downsampled to hourly data to match the weather data.

The Final Results

Value

Quote

Back of the
envelope/

Measured

Model/
Measured

Annual energy, no retrofit (kWh)

35,600

7,087

7,053

Annual energy, with retrofit (kWh)

11,200

3,132

3,132

Annual energy savings (kWh)

24,400

3,955

3,921

Annual savings ($)

$4,148

$672

$667

Simple Payback Period (years)

1.3

7.9

7.9



There are a few factors to the model that could cause these lower than expected numbers. Cooling tower consumption from June 2nd, 2015 to July 6th, 2015 was used as the baseline to predict power for July 7th, 2015 to June 5th, 2016. This is 335 days, so roughly 11 months. The average energy consumption per day was used to extrapolate out the rest of the year. Granted, this missing date range (6/5-7/6) is likely to be warmer weather. Also, the Fall of 2015 was significantly warmer than the Spring of 2016. To be more conservative, we can multiply the 6 months starting from 7/7-1/5 by 2 to recreate a complete year that includes the warmer fall weather twice. This leads to:

4331 kWh/year in savings by upgrading to a VFD, or ~$740/year with their same base rate, which is still a far cry from the quoted values. Even with the conservative estimate, it is still only 18% of the promised savings, and gives a 7.2 year payback.

Potential Improvements

There are a number of ways we could further tweak this model, including additional features (adding in occupancy data for the hotel), manipulating existing features (adjusting the dewpoint from SFO to better match the dry bulb temperatures at the hotel), or changing the cost structure (moving away from the $0.17/kWh base rate, and accounting for time-of-use and demand charges). Regardless, the current analysis is sufficient at demonstrating the large difference in ROI, and how important monitoring can be to making capital improvement decisions.