How Much Does Wind Power Reduce $\text {CO}_{2}$ Emissions? Evidence from the Irish Single Electricity Market

Abstract

This paper evaluates the effect of wind generation on $\text {CO}_{2}$ emissions using 2008–2012 historical data for the Irish Single Electricity Market. Wind generation displaces $\text {CO}_{2}$ emissions, as expected, in line with the average system emissions. Over the whole period, wind generation avoided about 8.8 million tons of $\text {CO}_{2}$, equivalent to about 12% of total system emissions. To understand what drives the level of abatement we evaluate the results by technology and determine that wind generation has similar effects on total emissions from CCGT and coal plants, due to the higher carbon content of coal. Each MWh of wind, however, replaces more generation from CCGTs than from coal plants, in proportion to their generation. We also test the hypothesis that as wind displaces baseload plants it pushes them to generate less efficiently, but find no evidence of a strong negative effect of wind on CCGT or coal plant efficiency. Finally wind displaces about 2.5% fewer emissions when the pumped storage plant is on outage, suggesting that wind is more effective when paired with a flexible system.

Appendices

Appendix A: Availability Data

We use data on plant-level availability from SEMO. We clean the data so that plant availability is:

• 1. Never larger than maximum capacity (allowing for 10% tolerance);

• 2. Never 0 when the plant is actually generating;

• 3. Interpolated from non-missing data when it is missing.

Some of the data is missing and some is registered as 0 even when a plant is generating, which can occur for a couple of reasons: a. availability is registered as 0 for system operation reasons. For example a thermal plant that is associated with a windfarm location is defined as unavailable according to the SEM. b. Some of the data might also be registered as 0 when data providers (plant operators) enter 0 instead of missing.

EirGrid publishes monthly availability for Republic of Ireland (ROI) plants (www.eirgrid.com/operations/systemperformancedata/availabilityreports/#d.en.797). We make sure that where information on availability is missing, the interpolated version is compatible with the EirGrid availability reports.

Appendix B: Daily Fixed Effects

In this section we report the results when the equations include daily fixed effects (and therefore exclude variables collected at the daily level, such as fuel and $\text {CO}_{2}$ prices). This specification does not change the results significantly with respect to the results reported in the main text. In particular, the average effect of wind on system emissions is −0.49 tonnes of $\text {CO}_{2}$ displaced, compared to −0.48 with month-year fixed effects.

As in the results reported in the main analysis, we implement fixed effects by taking the first difference with respect to daily averaged. To account for the remaining autocorrelation, we also impose an AR(1) specification on the residuals.

Table 6 Effect of wind on system $\text {CO}_{2}$ emissions (tonnes) with daily fixed effects

Tables 6 and 7 show the results of the specification with daily fixed effects for CCGT and coal emissions and generation. The results do not significantly differ from those reported in the main text. Table 8 reports the results on the analysis of the effects of wind on average efficiency of CCGT and coal plants with daily fixed effects.

Table 7 Effect on generation (MWh), by technology with daily fixed effects)

Table 8 Effect on efficiency (%), by technology with daily fixed effects)

Appendix C: Other Generation Technologies

Combustion turbine, oil, distillate and OCGT technologies each account for less than 2% of demand during our period, on average. There are a large number of periods when these plants do not generate. Even when they do generate, the distribution of emissions is far from normal. This leads to challenges in identifying a robust specification to estimate the effect of wind on the emissions of plants using these technologies. In particular, specifications that do not account for the large number of zeroes in the dependent variable can lead to biased coefficients.

The following tables show the results of three different specifications: a simple OLS, an OLS with AR(1) residuals and a two-step hurdle model, that helps us address the abundance of zeroes and the highly skewed distribution of non-zero values. In the hurdle model the first step estimates the probability that each technology generates a positive amount of electricity and the second estimates the model conditional on there being positive generation, as shown by the following equation:

$$\begin{aligned} \left\{ \begin{array}{rl} Prob(Emissions>0|\mathbf {X}) = F(\mathbf {X})&{} \\ Emissions_{t}= G(\beta \mathbf {X'})+ \mathbf {\epsilon } &{} \text{ if } \quad Emissions > 0 \end{array} \right. \end{aligned}$$

(4)

where X is the matrix of explanatory variables and $\epsilon$ is the error vector. X includes wind generation, load, the availability of other plants, net imports, fuel and $\text {CO}_{2}$ prices, outages at Turlough Hill, Moyle and their interactions with wind generation.

In the first part, the probability of generating and therefore emitting $\text {CO}_{2}$ is captured by a probit. The second part is modelled with a Poisson distribution with overdispersion, to account for the right skewness of the distribution of oil, distillate, CT and OCGT plants.¹⁰

All other control variables for both the first and the second steps are the same as in Eq. 3, including month-year fixed effects. We include wind and load in levels instead of quartiles, since the Poisson specification is itself non linear, implying a non-linear effect of wind and load even when they are included in levels.

For the OLS and AR(1) specifications, the standard errors of the marginal effects are calculated using the delta method.¹¹

The following tables report the estimated coefficients for wind and load (in levels) for the hurdle model, the OLS and the OLS with AR(1) specifications. For the hurdle model, we report the estimates for the second step. Complete results are available from the authors upon request.

Table 9 Wind effect on tonnes $\text {CO}_{2}$, other technologies, 2008–2012

Table 10 Load effect on tonnes $\text {CO}_{2}$, other technologies, 2008–2012

We explore the effect on generation also for hydro, which is not associated with measured $\text {CO}_{2}$ emissions. We do not study the effect on pumped storage, where generation decisions are more complex as they have to address both when to generate and when to recharge and therefore use electricity.

Table 11 Wind effect on generation (MWh), other technologies, 2008–2012

Table 12 Load effect on generation (MWh), other technologies, 2008–2012

Tables 9, 10, 11 and 12 above show that across the specifications the estimated coefficients are fairly similar. For distillate and oil, the technologies with the lowest number of periods with positive generation, the two-part model shows a larger effect of wind and load, as expected. The second step of the hurdle model excludes periods with 0 generation. For all these technologies marginally increasing wind generation has a similar effect to marginally decreasing load. The largest effect of wind is on CT and oil plant generation and emissions.

When we examine the effect on generation of hydro, we see that wind has a smaller effect than load. The effect of wind on hydro generation is not estimated robustly. The two-step model suggests no effect, in line with expectations, whereas the OLS and OLS with AR(1) suggest respectively a positive and a negative effect. One explanation is that to model hydro generation more accurately, we would have to be consider additional variables, as it depends for example on rainfall.