Abstract
The abrupt outbreak of the COVID-19 pandemic was the most significant event in 2020, which had profound and lasting impacts across the world. Studies on energy markets observed a decline in energy demand and changes in energy consumption behaviors during COVID-19. However, as an essential part of system operation, how the load forecasting performs amid COVID-19 is not well understood. This paper aims to bridge the research gap by systematically evaluating models and features that can be used to improve the load forecasting performance amid COVID-19. Using real-world data from the New York Independent System Operator, our analysis employs three deep learning models and adopts both novel COVID-related features as well as classical weather-related features. We also propose simulating the stay-at-home situation with pre-stay-at-home weekend data and demonstrate its effectiveness in improving load forecasting accuracy during COVID-19.
Introduction
Accurate load forecasting is the backbone of power grid stability. Utilities must maintain an "operating reserve"—extra generating capacity—to handle sudden spikes in demand. If they underestimate demand, the grid risks failure; if they overestimate, they waste massive amounts of money and fuel.
The COVID-19 pandemic created a "black swan" event for energy markets. As offices closed and people stayed home, industrial demand plummeted while residential demand surged. Traditional models, which rely on historical patterns where weekdays and weekends are distinct, suddenly saw errors spike as daily routines shifted overnight.
Methods
The researchers evaluated three deep learning architectures against a traditional statistical benchmark (ARIMA):
- FCDNN (Fully Connected Deep Neural Network)
- LSTM (Long Short-Term Memory)
- GRU (Gated Recurrent Unit)
They employed a rolling forecast technique, where the model is updated weekly with the most recent data to adapt to the evolving pandemic situation. The most novel approach was the Stay-at-Home Simulation. Recognizing that a lockdown makes every day look somewhat like a weekend, they proposed training models specifically on pre-pandemic weekend data to predict pandemic-era load, rather than using standard historical weekday data.
The study focused on the New York Independent System Operator (NYISO) market, utilizing hourly load and weather data, Google mobility reports, and daily COVID-19 infection statistics.
Key Results and Insights
The "weekend proxy" proved highly effective. Because stay-at-home orders meant people weren't commuting to offices, weekday energy patterns began to resemble pre-pandemic weekends. Using old weekend data as a training set for new weekdays significantly improved accuracy.
Key findings include:
- Benchmark Failure: In the early weeks of the stay-at-home order, standard models had a Mean Absolute Percentage Error (MAPE) of 5.32%.
- The "Weekend" Fix: By switching the training set to pre-pandemic weekend data, the error for the LSTM model dropped to 3.85%, a significant improvement in utility-scale forecasting.
- Mobility vs. Case Counts: Mobility data (how much people actually moved) was a much more useful auxiliary feature. In contrast, COVID case numbers were actually "baleful" (harmful) to model accuracy, as they are indirect, lagging indicators of behavior. Adding COVID case data increased errors, in some cases exceeding 8%.
Conclusion
While traditional ARIMA models failed during the volatility of 2020, deep learning models (FCDNN, LSTM, GRU) performed comparably well, with LSTM slightly edging out the others. The authors noted that their rolling forecast technique was vulnerable to "distortion" around week 7 of the lockdown, likely as the models struggled to reconcile the initial shock with the gradual stabilization of behavior. Future work includes testing these methods across different energy markets and investigating the specific causes of forecast distortion.
Summary updated