DER load forecasting: A powerful tool for DSM and situational awareness

In early 2018, the EnergyHub team set out to build a better DER load forecasting solution for our utility customers. After careful investigation of existing approaches and methodologies, it was clear there was a gap when it came to accurately predicting the load shape, magnitude, and timing of DER and connected device load activity. So we built our own solution, leveraging our expertise in advanced machine learning and access to grid-edge device data. We launched our forecasting feature with our Mercury DERMS user base in 2018, and now that we have completed a year of forecasts, we want to share a look at what we learned and how our solution performed.

But first, here’s some quick context on how EnergyHub’s forecasting approach is different from conventional approaches. Typical utility operational forecasts, whether generated internally or by a third party, are infrequently updated and offer insight at the system- or substation-level only. EnergyHub’s DER load forecasting, by contrast, offers utilities a targeted prediction of load from grid-edge devices, based on real-time customer behavior. While this represents only a subset of utility load, it correlates strongly to peak demand, and the DER forecast provides utilities with a valuable real-time update to their operational forecast, which otherwise wouldn’t be available.

EnergyHub has solved for a unique challenge: accurately and continuously predicting the load shape of behind-the-meter devices whose data the utility doesn’t even normally have access to. In the Mercury DERMS platform, EnergyHub designed its DER load forecasting capability to provide utility operators insight into projected DER activity, updated in real-time and bundled upwards from the grid edge. Behind the scenes, EnergyHub’s forecasting models leverage sophisticated machine learning techniques, rather than conventional regression-based approaches. Forecasts are continually updated based on device state and power consumption, as well as external data feeds such as weather, and made available via intuitive graphical displays at 15-minute intervals. Download our load forecasting fact sheet to learn more about our approach.

Validating use-cases

As part of our iterative development process, EnergyHub worked closely with utility users on uses they found most compelling for Mercury’s load forecasting capability. The insights we gained helped further validate the key use cases for DER forecasting.

Informing demand-side management strategies

Each utility undertakes a complex decision-making process when deciding whether or not to call a demand response event — relying on ISO-level forecasted and real-time load, weather, and even pricing data feeds to try to predict the peaks across their territory, as well as the stability and cost implications. EnergyHub’s unique DER load forecasting capability helped support those decision-making processes in tandem with, and often more effectively than, other options. In particular, it enabled more surgical targeting of stressed load pockets and provided unique visibility into the coincidence of DER peaks with overall system peaks.

Many decisions on calling demand response events are made day-ahead, so the ability to anticipate the timing and magnitude of peak load from DERs under management was helpful to teams across the utility, such as demand-side management and network operations. Furthermore, as scheduled demand response event times approached, users were able to see updated forecasts of load during the predicted peak period, helping them confirm or modify their demand response strategy.

Improving situational awareness

The second key use case for load forecasting focuses on improving utility operators’ situational awareness. As discussed earlier, today’s utility forecasts are limited in temporal and spatial resolution. Our clients confirmed that being able to visualize expected load activity at 15-minute intervals helps shed light on what’s actually happening at the grid edge, and driving their peak, especially on high load days.

Load forecasts were also useful when sliced based on devices’ locations in the distribution network. In EnergyHub’s Mercury DERMS platform, utilities can automatically, locationally organize devices based on provided grid topography data from a utility’s GIS system, enabling load forecasts that are generated and presented at substation, feeder, and circuit levels. These higher spatial-resolution forecasts allowed utilities to zero in on expected load in problematic load pockets.

Forecast accuracy evaluation methodology

Of course, load forecasts only deliver these benefits if they are accurate when it counts. With a summer’s worth of forecasts behind us, we can give a comprehensive picture of how our forecasts measure up to reality.

Our metrics for evaluation are grounded in the goal of improving utilities’ situational awareness. The timing of each utility’s system peak varies, but the 3 p.m. -7 p.m. window is a typical timeframe. This is also when demand response events are typically called. Accuracy during this peak afternoon window therefore provides the most value to utilities.

To evaluate how well forecasting performed for a program on a given day, we looked back at the forecasts that Mercury produced in the morning (9 a.m.) and afternoon (3 p.m.) of that day, as well as the day-ahead forecast. We graded each of these forecasts on the average relative error between the forecasted load curve and the realized load curve over the 3 p.m. – 7 p.m. window. An accuracy score of 80 percent (for instance) means that over the 3 p.m.- 7 p.m. window, the forecasted load curve deviated from the realized curve by an average of 20 percent. Note that accuracy involves correctly forecasting not only the total load average load during this window (the average height of the load curve), but also the shape of the load curve.

There’s more to accuracy than getting the average total load correct. A forecast can correctly predict the average total load over the 3 – 7 p.m. window while incorrectly forecasting the load shape, and in particular missing the timing of the load peak (left). The Mercury DERMS DER load forecasting tool has shown to accurately predict total average load and the shape of the load curve (right).

To measure forecasting accuracy for the summer, we rolled up these daily accuracy scores into a season-level number by taking their median over all weekdays in the summer, excluding days with demand response events. This yielded season-wide accuracy numbers for day-ahead, morning, and afternoon forecasts. Since forecasting accuracy is most valuable on days with significant load, we also produced separate accuracy numbers for the subset of days that had significant cooling load (peak daily load of at least 30 percent duty cycle).

Results and Takeaways

There are a few main takeaways from the 2018 season of forecasting results:

• Forecasting accuracy is even higher on days with significant load; afternoon forecasts are 93.4% accurate
• Accuracy is high across the board; in particular, afternoon forecasts are 90.6% accurate
• Accuracy increases as the day progresses and more day-of load and temperature data is incorporated into forecasting

Conclusion

It has been a great first year for the usage and performance of the Mercury DERMS DER load forecasting capability. It’s clear that clients are benefitting today and that our forecasting accuracy is impressive — especially for the unique and challenging nature of continually forecasting the shape, timing, and magnitude of DER load activity. As utilities and their customers continue to install connected DERs in record numbers, utilities will continue to benefit from a clearer picture of how devices under management constitute — and drive — peak load.

With our core forecasting capability in place, we continue to invest in expanding our forecasting capability to solve for additional utility needs. This includes providing flexibility forecasts, a natural extension of load forecast, which allows utility users to make informed predictions for expected and available load flexibility across multiple grid services. Another application is for continuous utility-wide or system-level forecasts — requiring a different set of data inputs but ultimately relying on the core machine learning intelligence we’ve built. We also plan to explore how our forecasting data can assist utility network operations in a more automated fashion, such as through integrations with Advanced Distribution Management Systems (ADMS).