Effect of COVID-19 on electricity demand, consumption and price
Shekhar Hazari
17/10/2020
1. Introduction
Access to electricity is a key factor in understanding human wellbeing, poverty and economic development of a country. It is no surprise that remote villages which still do not have access to electricity are called ‘backward’. Many researchers have tried to find the causal effect of electricity demand and consumption on a country’s economic development. In one such research paper “Does energy consumption cause economic growth?: Evidence from a systematic study of over 100 countries” by ‘Jaruwan Chontanawat’, ‘Lester C. Hunt’ and ‘Richard Pierse’ they state “Energy causes GDP to some extent. It is found that a greater proportion of OECD/developed countries show energy to GDP causality than that for non-OECD/developing countries.”
We know that COVID-19 has affected lives of many people. It has slowed down economic development all over the world, but to get real sense of economic effects of such once in a lifetime events it is not practical to wait for a period of 2-3 months. To mitigate this problem of identifying triggers which indicate upcoming slowdown in economy, it is desired that you look into some high frequency data like electricity.
Since it is hard to store electricity commercially, the demand and price keep fluctuating within as little as 5 minutes. The connection of electricity to economy makes it a very good variable to keep track of for anomalies.
In this report, I have analysed data from Australia and New Zealand to check if there is any anomaly or triggers in electricity market (consumption, demand and price). And lastly, whether we can detect these triggers next time they happen.
2. Data
Electricity demand, consumption and price vary greatly with Weather condition like daily temperature, rainfall, solar exposure etc. An example is shown below, which shows that electricity demand increases when temperature is lower or higher than usual and can be fitted using a quadratic equation. 
Other factors which may affect price of electricity can be weekly cycles for example electricity demand decreases drastically on weekends therefore the price of electricity is lesser on the weekends. 
The data for New Zealand and Australia has been collected from different sources. A brief description of the two data is given below:
2.1 New Zealand Data
The data for New Zealand’s residential electricity consumption has been collected from Electricity Market Information’s website. It consists of monthly data at Regional council level from January 2018 to July 2020. The weather data was collected from NIWA’s official website. It consists of monthly statistics like total rainfall, Avg max temperature etc.
The chosen dependent variable for this data was ‘Electricity consumption’.
2.2 Australia Data
The data for Australia’s 5 states’ wholesale demand of electricity was collected from Australian Energy Market Operator’s website. The data contains wholesale prices for electricity distributors across 5 states of Australia. The data was originally at the frequency of 5 minutes which was aggregated to day level. The data contains 2 important variables, ‘Total demand’ and ‘RRP’ (Dispatch price). One of these variables can be used to predict the other. In my case, I have predicted RRP using ‘total demand’. The weather data for Australia was collected from Australia Government’s Bureau of Meteorology’s website. It consists of daily min-max tempratures, solar exposure, rainfall etc.
The chosen dependent variable for this data was ‘RRP’ (dispatch price).
3. Problem Statement
Analyse the data to look for potential triggers that can help the government and businesses to identify a shift from usual behaviour which creates imbalance in the pattern of electricity consumption/pricing (which in turn can help identify imbalances in economy).
4. General Approach to solve the problem
To begin with, I went with Bayesian approach to fitting the parameters on the data. The reason for choosing the Bayesian approach to fit the model parameters rather than any point estimate algorithm is that I want to have a deeper analysis of the model parameters itself. So that, rather than noticing a shift in values of point estimates of parameters in pre-covid and post-covid era and then being left with no meaningful way to interpret the difference in the point estimates of the parameters, I want to know the significance of difference of the parameter values between the two models. Instead of saying that the point estimate of a parameter for a given predictor differs between the two models, I want to look at the parameter distribution and conclude how much does their distributions differ to determine if that is significant or not, which is not that easy to do with point estimates of these parameters.
The Bayesian algorithm I would be using is Gibbs Sampling algorithm to simulate a Monte Carlo Markov Chain (MCMC) model to uncover the distributions of the parameters of the models fitted on pre-covid data and post-covid data.
4.1 New Zealand’s monthly electricity consumption analysis
After finding out significant change in electricity consumption in New Zealand, the following analysis was done: I simulated a linear regression model for the this data in which the distribution of predicted value of electricity consumption follows a normal distribution with a constant variance (a parameter to the model with an inverse gamma prior distribution) and the mean of which is given by a linear combination of observed predictor values, and each of the linear parameters themselves follow a normal distribution. The priors for each parameter are taken such that they form a conjugate pair with the response variable distribution in isolation for convenient sampling. As I had no significant prior beliefs about the distribution of parameters hence I used a very non-informative priors for all the parameters.
4.1.1 Interesting Observations
Some key transitions were observed in the parameter distributions of a few variables when comparing them in pre-covid and post-covid models. The first of them being the parameter associated with total coverage. Interestingly it changed from being positively correlated to response variable to being negatively correlated. But of course, just the point estimates of these parameters doesn’t give away much information. So I plot the distributions for the same and observe the result: 
To summarize, in pre-covid era, the model shows that, bigger the electricity coverage, more likely the area to consume higher power, which make sense from a normal behavioural point of view. But in post-covid era this dynamic change, as with more total coverage the less likely it is to consume more power
Another significant change was in the parameter attached to mean temperature. The distribution shifts for the same are as follows: 
4.2 Australia’s daily dispatch price for wholesale market
Analysis similar to New Zealand’s was done but this time the data was at daily level. The model here is again a regression model, although some of the predictors are not linearly related to the response variable, but in quadratic order. The choice of using a quadratic regression function for some predictors was based off the scatter plot of these predictors with response variable which showed a quadratic trend line rather than a linear trend line. This changes the interpretation of some parameters and what they signify so I will be explicitly mentioning that in my post-modelling analysis of any such parameter
4.2.1 Interesting observations
The biggest change is again observed in a parameter that is similar to “total coverage” parameter in previous model, i. e., the total demand parameter. This parameter changes behaviour from being positively impacting electricity pricing to negatively affecting it in post-covid era. The probability that b1 was negative in pre-covid model was 1, and probability that it was positive in post-covid model posterior distribution was 0.99. Clearly a stark difference. The distribution for the same is as follows: 
To solidify the consistency in these transitions, the only other parameter that changes behaviour drastically here is the one associated with maximum daily temperature. The posterior probability that the parameter b3 of pre-covid model was greater than post-covid model came out to be 0.99. The posterior distribution for the same looks as follows: 
This is again, in line with what we observed on the monthly dataset in New Zealand in which the parameter associated with temperature shifted drastically
Monitoring Transition Triggers
Based on all the analysis above on two separate datasets, one can infer the following:
In pre-pandemic times, a higher electricity consumption/price is positively related to the demand/coverage of electricity in a region. Post-pandemic this behavior changes drastically by being negatively related. This information can be used to monitor triggers to economic instability. If one observes a drastic shift in correlation of electricity prices with the total coverage percentage/demand (from positive to negative) over a sustained period (like a week or fortnight) then it may signal an economic imbalance like the one caused by covid-19.
In pre-pandemic times, the affect of temperature on the demand/price is starkly different from post-covid times. One can use this observation to monitor any drastic aberrations of the correlation between temperature and consumption of electricity to forecast an economic imbalance.