SOLVER: Load Forecasting
Inaccurate hourly load forecasts due to variations in demand, blind spots in the distribution grid and data challenges – all leading to waste.
A comprehensive predictive data science approach, combining data quality advances and powerful new models drawing on real-time data.
Understanding the amount of power needed at any node in the distribution grid, at any hour of the day, is critical to successful grid operations. An oversupply creates waste, an undersupply causes outages. Getting the balance right on a 24x7x365 basis is not only important for ensuring commercial and residential customers are provided the power they require, but also for directing and re-directing load to successfully maintain the grid.
Gaining visibility into demand, and forecasting it accurately on an hourly basis, is not easy. While many utilities are increasing the use of sensors across their networks to gather real-time data and have deployed smart meters (or are in the process of doing so), there remain many blind spots in the network. Also, predicting usage has grown more complicated with the rise of on-site generation resources like solar and storage.
To account for these uncertainties, utilities often overcompensate. When, for instance, it’s time to conduct scheduled maintenance, most utilities base their switching plans on forecasts using conservative winter or summer peak load numbers, because they don’t know what the actual load will be at a particular hour on a particular day, and there is little room for error – they can’t fall short.
Knowing there had to be a better way, a major utility turned to predictive data science. Based on earlier successes working with the TROVE Platform to produce precise premise-level load forecasts for its Demand Response programs, the utility asked the predictive data analytics firm to aim its modeling magic higher up the distribution grid.
Precision Forecasts: Filling in the Blind Spots
Having already transformed the Demand Response program, TROVE knew it had strong models for forecasting demand at the network’s edge. Now it needed to bubble up that bottom-level forecast, parsing the smart-meter data the utility collected while incorporating it with the SCADA and other sensor data being captured at the sub-station level to produce an accurate hourly forecast between the substation and the meter.
Adding to the challenge, the data science effort had to account for the fact that a) smart-meters weren’t universally deployed (not everyone had them) and b) many meters read “zero,” meaning users may have produced more energy than they demanded that day, or simply that the meters had failed to report.
Fortunately, TROVE had the technology, team and experience to tackle this problem and deliver stellar Phase I results. Working with historical data, TROVE applied its Load Forecasting Solver™, which combines unique new machine-learning models and adept data-cleansing and data-enhancement techniques, to generate the needed hourly forecasts.
The solver has provided the utility a new baseline for forecasting accuracy and gained the utility’s approval for a Phase II live deployment. But to get there, TROVE had to prove the new forecasts worked. The TROVE Science Squad teamed with the utility to test the models along three key metrics: accuracy, stability and confidence:
Accuracy: This metric examines the accuracy of the models on average. In this case, the TROVE predictive models produced hourly, device-level forecasts that were, on average, only 13% off when measured day-ahead – twice as accurate as the utility’s internal, smart-meter-based models.
Stability: Yes, an average forecasting error of 13% vs. 26% is good, but if the models occasionally produce a forecast that is “way off,” and the utility executes plans based on such a forecast, there could be harmful consequences, including outages. To get in front of such problem scenarios, TROVE introduced a stability metric, measuring how often the forecast error remained below 50%. Within these parameters, the TROVE models were 98.3% stable, meaning 1.7% of their forecasts had error above 50%, – 10 times more stable than the smart-meter loading model the utility had developed.
Confidence: Here, TROVE got even more creative, building a model that supervised the others, flagging TROVE’s forecasts – as they were generated – whenever they could lead to a poor result. Understanding that 1.7% of the TROVE models’ forecasts were above 50% error, the confidence model flagged 99.35% of those as they occurred, meaning that an operator using the solver would have been told nearly 100% of the time when a forecast could lead to error, enabling the operator to take appropriate corrective measures proactively.
With increased forecasting accuracy, stability and confidence, the utility is looking forward to matching supply and demand more efficiently, simplifying switching plans, and empowering its people to put data to work to better meet its customers’ energy needs.
TROVE’s Predictive Data Science Platform is built on four key pillars: Datasynthesis™, Solvers™, Frictionless Functionality™ and Science Squad™. Here’s how each factors into making data useful at the utility: