How Effective was the UK Carbon Tax? A Machine Learning Approach to - PowerPoint PPT Presentation

How Effective was the UK Carbon Tax? A Machine Learning Approach to Policy Evaluation Jan Abrell, Mirjam Kosch, Sebastian Rausch IAEE, 25.8.2019

Two main questions 1. What was the impact of the UK carbon price support on emissions? 2. How can we use machine learning for policy evaluation in the absence of a control group? 2

Low CO 2 price … 175 35 Carbon emissions [Mio. t] 150 30 Carbon price [€/t] 125 25 100 20 75 15 50 10 25 5 0 0 09 10 11 12 13 14 15 16 EUA CPS Emissions 3

Low CO 2 price leads to introduction of UK carbon tax  Carbon price support (CPS) 175 35 Carbon emissions [Mio. t] introduced in 2013 by UK government 150 30 Carbon price [€/t] 125 25  Tax on electricity sector emissions 100 20  Varies by year 75 15 50 10 25 5 0 0 09 10 11 12 13 14 15 16 EUA CPS Emissions 4

Low CO 2 price leads to introduction of UK carbon tax  Carbon price support (CPS) 175 35 Carbon emissions [Mio. t] introduced in 2013 by UK government 150 30 Carbon price [€/t] 125 25  Tax on electricity sector emissions 100 20  Varies by year 75 15 50 10  What was the impact of the CPS on 25 5  coal and gas generation? 0 0 09 10 11 12 13 14 15 16  emissions? EUA CPS Emissions  What were the abatement costs? Sources: EEX (2017), Hirst (2017), EC (2016) 5

Coal-to-gas switch Impact of CPS on power market? Marginal cost c [€/MWh] d 𝑞 𝐷𝑃 2 𝑞 𝐷𝑃 2 Gas Coal Nuclear/ Hydro Installed capacity k [MW] 6

Coal-to-gas switch – and other reasons for lower emissions Impact of CPS on power market? Other reasons for lower emissions?  More renewables  Coal-to-gas switch  Lower demand  More imports  Less fossil capacity Marginal cost c [€/MWh] d  How to isolate effect of CPS? 𝑞 𝐷𝑃 2 𝑞 𝐷𝑃 2 Gas Coal Nuclear/ Hydro Installed capacity k [MW] 7

How would emissions have evolved without CPS? ?  Methodological challenge: No control group  Methodological Approach 1. Predict unobserved counterfactual (using machine learning) 2. Treatment effect: Difference between observed and «no policy» counterfactual 8

Literature and contributions Literature  Impact of fuel and carbon prices on electricity sector emissions  Empirical studies: Martin et al., 2016; McGuiness & Ellerman 2008; Martin et al. 2014; Jaraite and Di Maria, 2015; Cullen & Mansur 2017; Leroutier, 2019  Simulation studies: Delarue et al. 2008, 2010  Machine learning for policy evaluation Burlig et al. 2019; (Cicala 2017) Contributions  Ex-post assessment of carbon price impacts in electricity sector and how they depend on fuel prices  Program evaluation in the absence of a control group using machine learning 9

ҧ ҧ ҧ ҧ Methodological Approach in a Nutshell Proposed procedure Cottam Coal Power Plant (1) Theoretical model 1 𝑧 𝑗𝑢 = 𝑔 𝑗 𝑦 𝑗𝑢 , 𝑨 𝑢 + 𝜗 𝑗𝑢 , 1,6 Monthly Generation [TWh] CPS 2 ; 𝜗 𝑗𝑢 ⊥ 𝑦 𝑗𝑢 , 𝑨 𝑢 1,4 𝜗 𝑗𝑢 ~ 0, 𝜏 𝜗 1,2 𝑦 𝑗𝑢 controls 𝑨 𝑢 treatment variable 1,0 0,8 𝒜 ത 𝜺 𝒋𝒖 (2) Train prediction model f 2 0,6  Machine Learning approach 0,4 (3) Counterfactual prediction 3 0,2 𝑨 = 𝑔 𝑧 𝑗𝑢 𝑗 𝑦 𝑗𝑢 , 𝑨 𝑢 = ഥ 𝑨 𝑢 0,0 𝑨 𝑢 counterfactual treatment feb.11 jul.12 nov.13 apr.15 avg.16 obs pred noCPS (4) Derive treatment effect 4 𝑨 = 𝑧 𝑗𝑢 − 𝑧 𝑗𝑢 𝑨 𝜀 𝑗𝑢 10

(1) Theoretical Model: Short-run Electricity Market 1 Marginal cost c [€/MWh] d 𝑞 𝐷𝑃 2 Generation Capacity 𝑞 𝐷𝑃 2 𝑧 𝑗𝑢 = 𝑔 𝑗 (𝐸 𝑢 , 𝑑 𝑗𝑢 , 𝐿 𝑗𝑢 , 𝑑 −𝑗𝑢 , 𝐿 −𝑗𝑢 ) Gas Demand Marginal cost Coal Nuclear/ Hydro Installed capacity k [MW] 12

(2) Train prediction model with data 2 Hourly generation Hourly available of each unit capacity 𝑧 𝑗𝑢 = 𝑔 𝑗 (𝑠 𝑢 , 𝑢𝑓𝑛𝑞 𝑢 , 𝐸 𝑢 , 𝐿 𝑗𝑢 , 𝐿 −𝑗𝑢 , 𝝔 𝒖 ) 𝑧 𝑗𝑢 = 𝑔 𝑗 (𝐸 𝑢 , 𝑑 𝑗𝑢 , 𝐿 𝑗𝑢 , 𝑑 −𝑗𝑢 , 𝐿 −𝑗𝑢 ) Hourly marginal Hourly demand cost per unit Two challenges 1. Marginal cost not observed 2. Little variation in CPS prices 𝑕𝑏𝑡 , 𝑞 𝑢 𝑑𝑝𝑏𝑚 , 𝑞 𝑢 𝐹𝑉𝐵 , 𝑞 𝑢 𝐷𝑄𝑇 , 𝑢𝑓𝑛𝑞 𝑢 )  Use carbon price inclusive fuel price ratio as 𝑑 𝑗𝑢 = 𝑔 𝑗 (𝑞 𝑢 treatment variable 𝑑𝑝𝑏𝑚 + 𝜄 𝑑𝑝𝑏𝑚 𝑞 𝑢 𝐹𝑉𝐵 + 𝑞 𝑢 𝐷𝑄𝑇 ) 𝑠 𝑢 := (𝑞 𝑢 Daily fuel and Daily mean 𝑕𝑏𝑡 + 𝜄 𝑕𝑏𝑡 𝑞 𝑢 𝐹𝑉𝐵 + 𝑞 𝑢 𝐷𝑄𝑇 ) (𝑞 𝑢 carbon prices temperature Sources: ELEXON (2017), EIKON (2017) 14

(2) Train prediction model with data 2 Cottam Coal Power Plant 1,6  Estimate ෡ 𝑔 𝑗 from input data using Monthly Generation [TWh] CPS 1,4 machine learning 1,2 𝑧 𝑗𝑢 = መ ො 𝑔 𝑗 (𝑠 𝑢 , 𝐸 𝑢 , 𝐿 𝑗𝑢 , 𝐿 −𝑗𝑢 , 𝑢𝑓𝑛𝑞 𝑢 , 𝝔 𝒖 ) 1,0  0,8 In our case: 0,6 LASSO (penalized OLS) 0,4 0,2 0,0 feb.11 jul.12 nov.13 apr.15 avg.16 obs pred 15

(3) Counterfactual prediction 3 𝑑𝑝𝑏𝑚 + 𝜄 𝑑𝑝𝑏𝑚 𝑞 𝑢 𝐹𝑉𝐵 + 𝑞 𝑢 𝐷𝑄𝑇 ) 𝑠 𝑢 := (𝑞 𝑢 What would have happened without the CPS? 𝑕𝑏𝑡 + 𝜄 𝑕𝑏𝑡 𝑞 𝑢 𝐹𝑉𝐵 + 𝑞 𝑢 𝐷𝑄𝑇 ) (𝑞 𝑢 𝑜𝑝𝐷𝑄𝑇 = መ 𝑧 𝑗𝑢 ො 𝑔 𝑗 (𝑠 𝑢 (𝐷𝑄𝑇 = 0), 𝐸 𝑢 , 𝐿 𝑗𝑢 , 𝐿 −𝑗𝑢 , 𝑢𝑓𝑛𝑞 𝑢 , 𝝔 𝒖 ) Set CPS to zero for counterfactual: Cottam Coal Power Plant 1,6 Monthly Generation [TWh] CPS 2014 CPS 2015 CPS 2016 CPS 2013 1,4  Cheaper coal 1,2  More coal (and less gas) 1,0 generation 0,8 0,6 0,4 0,2 0,0 feb.11 nov.13 avg.16 pred noCPS 16

(4) Derive Treatment Effect 4 Cottam Coal Power Plant 1,6 𝐷𝑄𝑇 = ො Monthly Generation [TWh] መ 𝑜𝑝𝐷𝑄𝑇 CPS 𝜀 𝑗𝑢 𝑧 𝑗𝑢 − ො 𝑧 𝑗𝑢 1,4 1,2 Why not: 1,0 Observed – Counterfactual? 𝒜 ത 0,8 𝜺 𝒋𝒖  prediction errors lead to 0,6 biased estimate of treatment 0,4  eliminate bias by comparing 0,2 0,0 predictions feb.11 jul.12 nov.13 apr.15 avg.16 obs pred noCPS 17

Results 18

Impact of CPS on coal and gas generation  Coal (gas) generation decreased (increased) by 45 TWh  Generation impacts robust to inclusion of fixed effects  Generation impacts sum up to zero 19

CPS reduces emissions – at relatively low cost  Abatement: Δ𝐹 𝑗 = σ 𝑢 𝑓 𝑗 መ 𝜀 𝑗𝑢  Technical abatement cost: Change in fuel cost 14 70 Avg. abatement: 24.2 Mt (6.2%) 12 60 Abatement Cost [€/t] Abatement [Mt CO 2 ] 18.2 €/t Avg. cost: 10 50 8 40  What drives the impact? 6 30  Level of CPS 4 20 2 10  Coal-to-gas price ratio 0 0 2013 2014 2015 2016 Abatement Cost 20

Summary 21

Summary 1. What was the impact of the UK carbon price support on emissions? ?  Between 2013 and 2016, CPS lead to an emission reduction of around 6% at average cost of 18.2€/t. Cottam Coal Power Plant 1.6 Monthly Generation [TWh] CPS 1.4 2. How can we use machine learning for 1.2 1.0 policy evaluation in the absence of a ത 0.8 𝒜 𝜺 𝒋𝒖 0.6 control group? 0.4 0.2 0.0  Estimate unobserved counterfactual. Feb 11 Jul 12 Nov 13 Apr 15 Aug 16 obs pred noCPS 22

Backup Slides 23

When does the approach work? Cottam Coal Power Plant 1,6  Monthly Generation [TWh] Prediction errors CPS 1,4 independent of treatment 1,2  Observed prediction errors 1,0 do not depend on treatment 0,8 level 0,6  Do not predict “too far” out 0,4 of sample (covariate 0,2 overlap; positivity) 0,0 Pr 𝑠 𝑢 𝐿 𝑗 , 𝑢𝑓𝑛𝑞 𝑢 , 𝐸 𝑢 > 0 feb.11 jul.12 nov.13 apr.15 avg.16 obs pred noCPS  Independence of observed covariates 𝑕𝑏𝑡 , 𝑞 𝑢 𝑑𝑝𝑏𝑚 , 𝑞 𝑢 𝐹𝑉𝐵 , 𝐿 𝑗𝑢 , 𝑢𝑓𝑛𝑞 𝑢 , 𝐸 𝑢 ⊥ 𝑞 𝑢 𝐷𝑄𝑇 𝑞 𝑢  Conditional independence of unobserved covariates (h it ) 𝑕𝑏𝑡 , 𝑞 𝑢 𝐷𝑄𝑇 | 𝑞 𝑢 𝑑𝑝𝑏𝑚 , 𝑞 𝑢 𝐹𝑉𝐵 , 𝐿 𝑗𝑢 , 𝑢𝑓𝑛𝑞 𝑢 , 𝐸 𝑢 h it ⊥ 𝑞 𝑢 24

The Impact of Fuel Prices on Abatement Low r Intermediate r High r 𝑞 𝑑𝑝𝑏𝑚 < 𝑞 𝑕𝑏𝑡 𝑞 𝑑𝑝𝑏𝑚 > 𝑞 𝑕𝑏𝑡 𝑞 𝑑𝑝𝑏𝑚 ~𝑞 𝑕𝑏𝑡  Higher tax does not necessarily imply higher abatement 25

The Impact of Fuel Prices on Abatement Low r Intermediate r High r 𝑞 𝑑𝑝𝑏𝑚 < 𝑞 𝑕𝑏𝑡 𝑞 𝑑𝑝𝑏𝑚 > 𝑞 𝑕𝑏𝑡 𝑞 𝑑𝑝𝑏𝑚 ~𝑞 𝑕𝑏𝑡 High abatement potential Decreasing abatement No abatement potential potential High technical cost Moderate technical cost Zero technical cost Low abatement High Abatement Low abatement 26