The most common type of analysis that has been performed compares historical load ratios either week over week or year over year. However, there are many more factors that need to be taken into account to get an accurate estimation for the impact of lockdowns on regional electricity demand. Historical weather climatology and unadjusted historical load is insufficient to produce a counterfactual estimation of load in the absence of lockdown, as there is significant short-term variability in weather over any given region. As such, models that incorporate either weather actuals or real-time short-term weather forecasts are required. Additionally, for the Northern hemisphere, solar PV production has been increasing over time, which serves as a non-linearity when compared to historical load numbers. In this post, we correct for all these major variables to produce a more accurate estimation of the impact of Covid-19 lockdowns on the electricity demand for 28 regions in Europe, the US and Australia.
From NOAA’s publicly available GHCN weather database, we selected daily average temperatures of about 500 weather stations worldwide. Where available on our internal proprietary operational forecast systems (over Texas and Australia), we used high-resolution weather analysis fields (aka actuals), derived from an ensemble of the HRRR, NAM, ECMWF, and GFS weather models. We then calculated a simple average daily temperature per region, which we fed into the model.
For some plots and analysis we also needed a lockdown date, which we first took from publicly available news feeds. For many regions, these dates didn’t correspond well with the reduction in electricity demand. Many locales took it e.g. upon themselves to begin quarantining before an official government mandate. Also, restrictive measures were taken in several steps, making it difficult to define a single date. The OpenTable datasource (https://www.opentable.com/state-of-industry) was a very valuable guide in determining these “apparent” lockdown dates, as it generally reflects the start dates of de facto lockdowns, but we did not rely on it for all countries as it is not comprehensive enough for our needs. We finally decided to calculate “apparent” lockdown dates as the first date in March 2020 for which the (smoothed) demand reduction percentage increased beyond 2%. For some regions like Italy (locked down on March 9) this date was exactly the same as the official lockdown date, but on average it was half a week earlier and in New York up to 10 days earlier. For the rest of this post, “lockdown” is defined as the apparent lockdown date inferred from the electricity demand reduction and guided by the OpenTable dataset.
We then built a simple linear regression model, trained on 6 months of daily data for January 8 - April 30 2019, and January 8 - February 29 2020, with in total 9 weather and calendar features. We predicted March 1 - April 10 (out of sample, but with actual weather data), as a counterfactual forecast. This allowed us to gain insight into counterfactual forecasts across many regions around the globe to understand what the load would have been in the absence of the pandemic, and the divergence between these counterfactuals and the actual load numbers can reasonably be attributed to the lockdowns.
In the first plot we show an estimation of the maximum weekly reduction of electricity demand per region. The error bars are based on the (normalized mean absolute) error in the model fit, based on the model training up to February 29 2020, indicating the goodness of fit up to that point.
It’s clear that Italy leads the way with almost 25% decrease in electricity demand, followed by Spain at almost 17%. On average an impact of 8.7% with a relatively large standard deviation of 4.9% was found. We can generally correlate the variances with the intensity of the coronavirus outbreak within each region, with higher rates of cases per capita leading to higher rates of demand reduction.
In the second and third plot we show over time the demand actuals and forecast, respectively the percentual forecast error, both annotated with a vertical line for the apparent lockdown date. For some regions the model fit is accurate, demand reduction is sudden and strong, the lockdown date is visually obvious and corresponds with the official lockdown date (Italy). For other regions demand reduction happens more gradually (New York), the fit is noisy (South Korea), a day of week pattern can be spotted (Germany) or you see what happens when a country decides not to lock down (Sweden). Interestingly, for Italy, France, Belgium, Portugal, California and Austria, demand reduction already seems beyond it’s maximum, but we’ll have to wait and see if this is really the case.
For the fourth plot we combined the estimated electricity demand reduction in a scatterplot with the number of Covid-19 cases before and after lockdown and fitted a logarithmic trend on the data after March 1 2020. Though the logx doesn’t cover a complete or perfect fit, for many regions it captures the relatively steep reduction at the first cases, which then flattens when the outbreak matures.
In plot five, aggregating all data from plot 4, we compare the percentage of electricity reduction to the number of Covid-19 cases per 1M population. We aggregated all underlying daily data points also into several bins to expose a clearer view on the average reduction (at 0, 200, 500, 1000 and 2000 cases per 1M). The error bars and band indicate a 95% confidence band of this fit. Roughly speaking, at around 200 per 1M Covid-19 cases you can expect about a 5% demand reduction, increasing towards 10% at 1000 per 1M cases and going towards an average maximum impact of 15% beyond 2000 per 1M cases.
In the sixth plot, for a selection of regions, we show the relationship between the moment of lockdown and the amount of electricity demand reduction, fitted with a (lowess) smoother. Generally, we can see that the more time passes after a lockdown, the more the electricity demand is reduced, which is collinear with the amount of cases in each respective outbreak. We hypothesize that this relationship is actually sigmoidal, as once a full lockdown state is achieved, we should see a stabilization in the demand trend to reflect the new regime. It may be too early to see this relationship for many of the countries below, so it’s worth revisiting this point some weeks from now. According to the smoother the impact for France is already declining for almost two weeks, but this might not hold.
Indicating (again) how much the impact can vary between regions, in the seventh plot we show the data for the two regions in Europe with the highest per 1M Covid-19 numbers (Italy and Spain) and the three regions we investigated in the US (New York, Texas and California). Though towards 3000 per 1M, impacts seem to converge around 15%, around 1500 per 1M cases we could still see a huge spread in demand reduction between 5% and almost 30%, showing regions have reacted very differently under similar situations. As cases all around the world are being measured differently, caution though needs to be taken in the interpretation of the amount of Covid-19 cases.
Based on our research, we deployed preemptive measures for our clients on a case-by-case basis as early as March 17th, adjusting predictions to ensure the imbalance is as low as possible even when demand was significantly reduced. This is a difficult task given the completely out-of-sample nature of this phenomenon. Learning from historical load patterns when adjusted for weather conditions is a standard technique in load forecasting, but regime changes in usage like we’re seeing in these lockdowns cannot be learned from prior examples due to a lack of training data.
Even known pattern changes in electricity demand, as seen in Italy and other countries, can not be used “as-is” for grid regions that might be affected by lockdowns in the future, as different flavors of lock-downs were implemented and their effects differ from one place to another. Regional grids also differ wildly in consumption patterns, and it is not possible to simply train a model on one grid’s data and use that model in a new grid region without incurring significant performance impacts.
To solve this, we implemented a fast-learning methodology that can learn from very recent consistent deviations from known patterns. Once such deviations are detected, and found significant by the system, we predict the most likely effect going forward. This is done by resampling past examples and heavily overweighting very recent ones, and learning the magnitude of deviations over time. To test the efficiency of this mechanism, we’ve simulated biases that mimic changes from other grid regions. Running multiple back-tests we’ve found that such methodology can reduce by 70% on average the error that can be attributed to lock-down impacts. Although it’s very appealing to have such a mechanism in place even with no lockdown in sight, we’ve found that false positives can have a slight negative effect on load predictions when no pattern changes occur. Therefore, we monitor the effect of such calibration in each of the regions we service, and make a case-by-case decision when to apply it in each.
We’ve applied this methodology in several models that service territories in AEMO (Australia) and ERCOT (Texas). To date, the system has not detected a significant impact in ERCOT, but from the four AEMO territories we service, we’ve detected large regime shifts in patterns in New South Wales starting the last week of March. The following graph is an example of our predictions with and without this adjustment for that territory for several days in the start of April. We will continue monitoring the situation, making adjustments as needed when the pandemic begins to taper off and updating the models accordingly.
