Information Reliability Essential for Use of Smart Grid DER Behind the Meter
Written by Linda M. Zeger
As Distributed Energy Resources (DER) behind the meter (BTM) increase, the need to carefully manage both the demand-supply balance and increased utilization of local feeders and transformers is highlighted; this management requires accurate and timely information. Reliability issues in grid data could degrade this information and result in the propagation of inaccuracies, including into smart grid control. Poor data quality or inaccurate forecasts could potentially contribute to unnecessary green-house gas (GHG) emissions and financial expenditures, as well as adverse grid events.
Successful leveraging of real-time information, as well as forecasts of how both load and DER supply fluctuate, is essential. Reliability of this information and its usage depends on many smart grid components, including sensor systems, communication networks, and data quality processes. Several types of information reliability challenges, as well as improvement approaches, are outlined below.
Missing and Noisy Data
If assumptions upon which missing data imputation  is based do not hold, biases can result in inferences that use such imputations. For example, if data used to predict load is missing from some locations due to network communications issues, and that missing data is imputed or estimated from locations without communication issues, then inferences made using those imputations may be biased if locations with and without communication issues differ in their load profile associations. Furthermore, estimating a missing data point with a single value, such as a median, neglects variability .
While some data is inherently noisy, as from sensor measurements, some commonly used privacy schemes intentionally add noise to data. Automation of data pre-processing may itself potentially introduce new errors into the data. These sources of noise and errors may add to measurement errors, which may reduce forecast accuracy.
Model Applicability Depends on Inclusion of Key Variables
Forecasts of load or DER derived from data that does not include key factors, such as customer types or local conditions, can give poor predictions. Moreover, models and forecasts can become outdated if they do not incorporate the evolution of key factors, for example, the arrival of new EVs or changes in various energy management systems.
If data or models developed from early DER BTM adopting customers is used to predict the load of new customers with insufficient data, this use can create biases if newer DER BTM customers differ from earlier adopters in ways associated with their load that is not represented by collected variables. Similarly, predetermined sample data sets can also introduce biases as they may not be representative of future conditions, such as use of real-time pricing signals.
Likewise, the evaluation of any smart grid component, such as a DER management system or communication system, is challenging because the data collected depends on numerous local conditions. Ideally, a randomized controlled trial (RCT) would control for all the factors. In practice, it would be difficult to implement a RCT in the smart grid. Hence, results of any trial may not pertain to other conditions, locations, or customers.
Forecasts of DER and local distribution loads, especially BTM, are challenging due to a dearth of BTM data, large fluctuations in load, and extremely local weather variations . The inputs upon which these forecasts are based may also have data quality issues and depend on weather or price forecasts, which themselves have inaccuracies.
Forecasts of load that depend on BTM DER may in turn be used as inputs to energy management systems, pricing, planning, and other smart grid applications . The resulting impacts of errors and uncertainties in load forecasts on GHG emissions, grid reliability, resilience, and cost, are largely unknown .
Evaluation of load forecasts is often based on averages such as mean absolute percentage error, which poorly reflect peaks in fluctuating data, and even evaluation  of peak forecasts is often unreliable. Quantile regression, a probabilistic forecasting method which can express uncertainty in forecasts, is used to some extent in load forecasting . However, widespread robust evaluation of quantile regression and other probabilistic forecasting methods is lacking .
Communication & Cybersecurity
Reliable, rapid communication of data, forecasts, and control information is essential to operation of the smart grid. Since EVs may serve as a prime large BTM flexible DER, there is an increasing need for communication with EVs, including when EVs are not at charging stations. Communication with EVs is even more urgent during extreme weather when regularly used charging stations may have less availability, or EVs may be needed to supply power.
Challenges to smart grid communication include limited bandwidth , cost, and for some wireless systems: spectral interference, signal degradation from weather, and coverage issues such as obstructions including indoor garages, which may contain charging stations. Communications may also be disrupted from extreme weather communication infrastructure damage or power outages; such disruption may hinder reliable information exchange with EVs at critical times.
Information can also be corrupted by cybersecurity attacks on either communication networks, stored data, or on the training  and use of artificial intelligence (AI) algorithms.
Improving Information Reliability
To handle potential errors introduced by use of imputed or estimated values for missing data, determination should be made as to whether data missingness  is random: if it is, the uncertainty associated with imputation should be incorporated into uses of the data. Otherwise, information from the pattern of missingness should be utilized.
Additional metrics to quantify data and forecast quality should be designed , incorporated efficiently into smart grid standards, and input to algorithms that use the data and forecasts. These metrics should quantify potential uncertainties, for example, through use of prediction intervals , for forecasts. The metrics should also quantify biases, and include not only average forecast performance, but also forecast performance under conditions that strain the grid. Evaluation of the impact of forecast inaccuracies should be performed . For example, a comparison could be made of how a system performs using predicted load at a prediction interval center when the actual load is at one end of the interval. Reliability of prediction intervals themselves should be robustly validated against actual data, as should additional probabilistic and point forecasting methods .
Use of multiple diverse communication networks, including for example, Wi-Fi, cellular, FM radio, and satellite communication, can increase resilience. Proactive and opportunistic charging and discharging of EVs not only over time, but also over space, could mitigate subsequent disruption of information and power. For example, before a predicted extreme event, the smart grid should initiate communication with EVs to incentivize early charging at favorable locations and discharging for upcoming potential outages. Finally, communication demand can be reduced by storing and processing more data locally, which current technology development is addressing.
More actual or synthesized data is needed, particularly from extreme weather and other unusual conditions. In addition, supplementary information  can be obtained from use of physics constraints, as well as from outside sensor systems. Forecasting models should be monitored and upgraded as grid conditions evolve. Defense against the range of cybersecurity threats, including attacks on training and use of AI algorithms , should be incorporated at the start of planning.
- R. Little, D. Rubin, “Statistical Analysis with Missing Data”, 1987.
- “Addressing Solar Data Challenges for Utilities and Power Systems”, DOE/NREL Workshop, June 2022.
- S. Haben, et al., ``Review of Low Voltage Load Forecasting: Methods, Applications, and Recommendations'', Applied Energy, December 2021.
- A. Ilyas, et al., “Adversarial Examples are Not Bugs. They are Features”, NeurIPS 2019.
This article was edited by Vijay Sood.
Linda M. Zeger leads the data strategy for smart grid projects, including data quality improvement and performance evaluation of communication and artificial intelligence components. She has contributed a variety of novel performance improvement capabilities and analytics techniques to a range of systems including communication and sensor networks. Her experience includes consulting, research, smart grid edge standards development, program management, and instruction. She is the holder of a number of patents and she has given presentations to a wide range of audiences. She is the author of numerous published papers in journals and conferences across several disciplines. She holds a Ph.D. in physics from Harvard University and an A.B. in physics from Princeton University.
She is the founder and principal consultant of Auroral LLC, which provides advisory services in data science strategy, system performance assessment, experimental design, adaptable energy management, and communication networks. She has also held positions with MIT Lincoln Laboratory, Lucent Technologies Bell Laboratories, Princeton University, and Florida Polytechnic University, and she completed a postdoctoral fellowship in statistics at Educational Testing Service.