Reduction in Carbon Emissions — A Machine Learning Framework
This post is inspired by the Drivetrain approach of obtaining value from algorithmic predictions. Machine learning and Artificial Intelligence must be converted into actionable outcomes that can drive businesses and nations to prosper sustainably. All thoughts echoed here are personal and welcome for any constructive criticism.
Part I : Background
Before we delve deeper into machine learning frameworks it is necessary to understand some jargon.
To reduce carbon emissions from nations, it is necessary to find the right balance between different policies. A clear example of two policies that could be in conflict would be that of renewable energy subsidies and pollution emission credits.
Renewable energy subsidies are incentives offered by the government of a nation to enable larger scale adoption of such technologies. This is called technology learning. When a new technology enters the market, naturally it is an incumbent to others that have existed from before in terms of price. This is because, over time, businesses have learned how to make them cheaper and incorporate economies of scale. This is clearly illustrated below in the fall of Lithium-ion battery prices. Subsidies are necessary to allow for market proliferation of new technologies.
Pollution Emission credits are a fixed set of emission allowances offered by a government and distributed among the emitting industries/companies of a nation. Any company/industry that emits more than the allowances it has been allotted will be penalized severely. If Company A performs well; has cut its emissions by 20% and has an excess of emission credits, it can sell them to, say, Company B which has overstepped its allowances. Such a mechanism is also termed cap and trade. As such, there is an incentive for companies to cut down their emissions or invest into renewable energy technologies.
However, when there is no right balance between the above two policies, the right objective of carbon emission reduction may not be achieved. As an example scenario, if renewable energy subsidies are too high, some companies will invest in renewable energy technologies reducing emission by a certain amount. Additionally, this will disincentivize pollution emission credit purchase; the price of pollution credits goes down. This may allow some companies to pollute a lot more because it is easy to purchase allowances as they are priced low. Consequently, carbon emissions may increase disproportionately and lead to the nation defaulting on its targets.
What happened? The objective that we initially set out to achieve (which was reduction in carbon emissions) was driven by unbalanced levers (pollution emission credits and renewable energy subsidies).
How does one price them optimally? This is where we can bring in machine learning.
Part II : Lever Optimization
The Drivetrain method of producing actionable outcomes from machine learning relies on a systematic procedure:
Step 1 — Define Objective
Maximize reduction in carbon emissions.
Step 2 — Define Levers of Action
Levers of action or variables that can drive carbon emissions include number and price of pollution emission credits, amount and value of renewable energy subsidies, price of carbon tax, and regulatory approaches (regulatory approaches differ from pollution emission credits in that there is no trade mechanism involved: the government decides a particular industry/company cannot emit more than a certain quantity/quality of emissions).
Step 3 — Data Crawling
We can crawl lots of data. How much does each industry emit? What is the quality of emissions? Where do the emissions come from? Are there forests around which can absorb these CO2 emissions? How much budget does an industry allot for reducing emissions? How much is actually utilized in reducing emissions or investing into renewable energy technologies? What is the growth pattern of the renewable energy industries? How does the price of the renewable energy raw materials affect the cleantech industry? How does the proliferation of cleantech affect the profits of an emitting industry? How much does the budget dedicated to emission reduction eat into the profits of the industry?
The data collection seems daunting but it is all out there. There needs to be a careful curation of data and organization of the same.
There is one more issue: the current data present is for a business as usual scenario for a set of pre-decided values of the levers. However, we would like to have causal data. What happens if we have different values of the levers, how does the industry react? The ideal method to generate such a data would be to randomize the lever values and collect responses of the industry’s emission reduction or increase. Of course, this is not a smart way to generate such data for large-scale regions such as a nation. What can be done is to run small-scale experiments in a location to generate such data.
Step 4 — Predictive Modelling & Simulator
Assuming all of the data necessary is present, we can build predictive models for the business as usual scenario. Since we have a set of randomized experiments as well, we could simulate some scenarios. What happens if we increase the renewable energy subsidy price and decrease the number of pollution emission credits available or vice-versa?
Step 5 — Optimization
For optimization, we play on the different levers. We pull on it iteratively and observe the reduction in carbon emissions in the simulated system. We do this until we find a maximum possible reduction in carbon emissions.
Each company can also implement internal carbon pricing. And the above methodology can be applied to streamline their businesses and reach optimum spending on resources while maximizing profit sustainably.