Climpact: A User Study of Perceived Carbon Footprint Victor Kristof 1 , Robin Zbinden 1 , Valentin Quelquejay-Leclère 1 , Blanche Dalimier 2 , Alexis Barrou 2 , Edouard Cattin 2 , Lucas Maystre 1 , Jérôme Payet 2 , Matthias Grossglauser 1 , Patrick Thiran 1 victor.kristof@epfl.ch 1 I nformation and N etwork Dy namics Lab (indy.ep fl .ch) Data Science in Climate and Climate Impact Research 2 CYCLECO Life Cycle Assessment (cycleco.eu) August 21, 2020
High impact, poorly documented (Except for car-related actions) Moderate impact, well documented Low impact, well documented 3
This Work Are people actually poorly educated about the impact of their actions? Understand how people perceive the carbon footprint of their actions fl ying How much does emit? eating meat lighting a house Except for experts , it is very di ffi cult to estimate our absolute carbon footprint To make decisions about daily actions, we need the relative carbon footprint 4
Psychometrics [Thurstone 1927] 5
Psychometrics [Thurstone 1927] 5
Ranking from Pairwise Comparisons Di ffi cult task Easy computation 2300 kgCO 2 40 kgCO 2 800 kgCO 2 kgCO 2 Easy task 100 times more 10 times more Di ffi cult computation… 2 times more …made possible via a statistical model of pairwise comparisons! 6
Statistical Model of Comparisons Let be a set of actions and let be a triplet encoding that action has an impact ratio of over . 풜 y ∈ R >0 M ( i , j , y ) i j ( ) , , Information: relative order of magnitude y = 100 Given some parameters representing the « log » carbon footprint of actions and , we posit w i , w j ∈ R i j y = ˜ w i = exp w i where w i = log ˜ w i = 100 w j ˜ exp w j with , assuming ϵ ∼ 풩 (0, σ 2 n ) comparisons are noisy y = exp w i log( ⋅ ) 0 log y = w i − w j + ϵ = x ⊺ w + ϵ … exp w j 1 i comparison vector in R M x = … selecting the pairs of actions log y ∼ 풩 ( x ⊺ w , σ 2 n ) -1 Reminder: j … 0 100 times more We estimate the global perception from relative comparisons 7
Estimating the Global Perception For a dataset of independent triplets, the likelihood of the model is N Hyperparameters prior mean μ ∈ R M N ∏ p ( y i | x ⊺ i w , σ 2 n ) = 풩 ( Xw , σ 2 p ( y | X , w ) = n I ) prior covariance Σ p ∈ R M × M i =1 Assuming a Gaussian prior for the parameters , we compute the posterior distribution as w ∼ 풩 ( μ , Σ p ) Σ = ( σ − 2 − 1 p ) n X ⊺ X + Σ − 1 used for active learning Σ p ( w | X , y ) = p ( y | X , w ) p ( w ) = 풩 ( ¯ w , Σ ) w = Σ ( σ − 2 p ( y | X ) p μ ) n X ⊺ y + Σ − 1 gives the perception ¯ exp ¯ w exp ¯ w kgCO 2 8
Enriching the Model: Perception Bias Users: Actions: We want to capture the perception bias of users and actions into the model - Age - Category - Gender - Source of energy - Education - Duration exp ( w i + ∑ k b ( u ) ik ) y = exp w i , where the bias depends on user and on action y = b ( u ) ik ∈ R u i exp ( w j + ∑ k b ( u ) jk ) exp w j Example: exp ( ∑ k b jk + ∑ k b jk + ∑ k b jk + ∑ k b jk ) « Transport » + + if the user is a female participant u y = exp ( ∑ k b jk + ∑ k b jk + ∑ k b jk + ∑ k b jk ) + « Housing » + These assumptions enable fl exibility and interpretability of the model! 9
Active Learning We can use the covariance matrix of the posterior distribution of the model to (smartly) select pairs of actions . Σ = ( σ − 2 − 1 p ) Recall: where n X ⊺ X + Σ − 1 p ( w | X , y ) = 풩 ( ¯ used for active learning w , Σ ), Σ As proposed by [MacKay* 1992], we want to select the pair of actions that is maximally informative about the values that the model parameters should take. This is obtained by maximizing the total information gain : w Δ S = S N − S N +1 = 1 0 2 log ( 1 + σ 2 n x ⊺ Σ N x ) , where Σ N = [ σ 2 ij ] M i , j =1 … 1 i Entropy of multivariate Gaussian i.e. , all possible comparisons x = … -1 j To maximize , we maximize for all possible in our dataset. We seek, therefore, to fi nd x ⊺ Σ N x Δ S x … 0 ( i ⋆ , j ⋆ ) = argmax { σ 2 ii + σ 2 jj − 2 σ 2 Very fast to compute for our model! ij } i , j We can actively select the next pair of actions * Yes, the same MacKay who wrote the book Sustainable Energy – Without The Hot Air ! 10
Dataset of Actions Take the train on a 1000-km round-trip Light your house with incandescent bulbs The train is a high-speed train with 360 seats. The seat- Incandescent bulbs emit CO2 because they consume occupancy rate is 55% (average rate for these types of electricity to generate light. The electricity is consumed trains). We count the CO2 emissions per passenger. from a grid with average CO2 rate. Carbon footprint: 17 kgCO2-equivalent Carbon footprint: 239 kgCO2-equivalent Eat eggs and dairy products for one year Fly in fi rst class for a 12000-km round-trip The production of eggs and dairy products (milk, cheese, The plane is a standard aircraft for long-distance fl ights etc.) emits CO2 because of water and land consumption, with 390 seats. The seat-occupancy rate is close to 100%. animal methane, and fossil fuel consumption for We count the CO2 emissions per passenger. Passengers transportation and heating. We consider an average citizen fl ying in fi rst class use more space than passengers in consuming 50 kg of eggs and dairy products per year. economy. Carbon footprint: 100 kgCO2-equivalent Carbon footprint: 9000 kgCO2-equivalent A total of 18 actions covering 3 categories (housing, transportation, and food) New dataset of 50+ actions covering 5 categories (goods and services) 11
New Actions These features can be integrated into the model to capture action biases ! Round-trip in train from Lausanne to Zurich The train is an SBB long-distance IC train. The seat occupancy rate is 28 % (392 passengers). SBB trains run on electricity. They have a service life of 40 years. The travel distance is 348 km. Emissions include rail construction/dismantling, train maintenance, SBB's HV power generation, train station and train construction/dismantling. Emissions are in kg of CO2 eq for one passenger. Station Train Carbon footprint: 2.35 kgCO2-equivalent 8% 8% Abrasion Perimeter: Electricity 2% - Production & dismantlement of train 9% - Production & dismantlement of tracks - Electricity source - Maintenance Maintenance - Train station 3% Functional Unit: Ensure the transportation of people in train Methodology: Bottom-Up LCA Data: Ecoinvent Database Because of steel and concrete required to lay tracks Tracks 71% 12
Climpact.ch: Collecting the Data 13
Climpact.ch: Collecting the Data 14
Climpact.ch: Collecting the Data 15
Results Number of answers: 3102 Number of users: 246 Age of 3/4 of users: 16-25 Log scale! 16
Results Number of answers: 3102 Number of users: 246 Fly in fi rst class on a 12000-km round-trip (411%) Age of 3/4 of users: 16-25 Take the train on a 1000-km round-trip (205%) Heat your house with an oil furnace (682%) Fly in economy for a 800-km round-trip (147%) Fly in economy on a 12000-km round-trip (122%) Dry your clothes with a dryer (620%) Eat local and seasonal fruits and vegetables (200%) Take the bus on a 1000-km round-trip (155%) Log scale! Low-impact (over-estimated) Medium-impact actions High-impact (under-estimated) 16
exp ( ∑ k b jk + ∑ k b jk ) Gender Bias Number of answers: 2905 143 + y = exp ( ∑ k b jk + ∑ k b jk ) Number of users: 239 + 96 Age of 3/4 of users: 16-25 912 kgCO 2 -eq. 2526 kgCO 2 -eq. 965 kgCO 2 -eq. 9000 kgCO 2 -eq. 381 kgCO 2 -eq. 270 kgCO 2 -eq. 17
Limitations and Ongoing Work The model is currently rather simple Include more features (derived from new actions) Active learning is equivalent to uniform selection Make the data collection even more e ffi cient The online platform is very basic Integrate the new actions and visualization tools Data collected over a small, biased population Open the platform to the general public Can we move from active learning to active teaching… ? 18
Thank you! https://climpact.ch (But please don’t share it further!) Read our paper: https://infoscience.epfl.ch/record/275472 Or scan this code Connect with me on Twitter! Or reach out by email! @VictorKristof victor.kristof@epfl.ch indy.ep fl .ch
Recommend
More recommend
