The figure above is a time-series of electricity consumption in each country. with an overlay of a SARIMAX model predicting energy consumption in 2019 vs 2020, having been trained on the previous 4 years worth of data. The appendix outlines the accuracy of these models.The objective of this plot and table is to illustrate the predictability of electricity consumption data during a normal year vs a pandemic enduced year. The change is clearly picked up in the drastic changes in accuracy.
Electricity consumption can be attributed to weather, seasonality, business cycles & base load. Weather and seasonality are present moment indicators unless a weather forecast can be added to the model. The temperature causes one to use heating and cooling measures. Business cycles and base load (level of business activity) can be presumptiously reactive to conditions unlike weather and seasonality which are reactive in the moment. Hence, electricity consumption attributed to business cycles and base load can be a forward looking indicator into economic activity.
t-SNE (t-distributed Stochastic Neighbor Embedding) is something called nonlinear dimensionality reduction. This is an algorithm which allows us to separate data that cannot be separated by any straight line. Illustrated below are 3 years worth of Electricity Demand Data from Victoria. 2018 in red, 2019 in blue & 2020 in orange. Hovering over the data, we are able to see what factors such as RRP of electricity, week, day & hour impact the cluster formed. Points which appear further away are indicative of outliers. Interestingly, March 2020 appears to be the only occurance where the pattern breaks.