This article was first published from the WeChat public account: HyperAI super nerve
Content overview: ScienceAI, as a technology hotspot in the past two years, has attracted widespread attention and discussion in the industry. This article will revolve around a paper from ScienceAdvances to introduce how to use machine learning to predict the amine emissions of coal-fired power plants.
Key words: AI for Science chemical engineering amine emission
According to a report released by the International Energy Agency, global energy-related CO2 emissions in 2021 will increase by 6% compared with 2020, reaching 36.3 billion tons, a record high.
Among them, the largest increase in CO2 emissions was in the power generation and heat supply industries, with an increase of more than 900 million tons, accounting for 46% of the increase in global CO2 emissions. It is urgent to control and reduce CO2 emissions in power generation and heating industries.
Annual change in CO2 emissions by sector in 2021
Blue indicates annual change, red dots indicate net change
View the full report on global CO2 emissions in 2021
Carbon Capture: Reducing Greenhouse Gas Emissions and Turning Waste into Wealth
In "Research on the Development Path of Carbon Peak and Carbon Neutrality in China's Power Industry", industry experts gave three measures to reduce CO2 emissions in the power industry:
1. Vigorously develop low-carbon power sources such as wind power, hydropower, and nuclear power, and abandon high-carbon power sources such as coal power and oil power
2. For coal-fired power plants, use natural gas, straw, biomass and other low-carbon fuels instead of coal for power generation
3. Use carbon capture technology to capture and utilize CO2 emitted by coal-fired power plants
Among them, carbon capture has attracted the attention of commercial companies, energy companies, and power industry research institutes because of its small transformation range, large imagination space, and ability to turn waste into treasure .
Carbon capture equipment in power plants
Carbon capture refers to the use of CO2 to react with amines to capture and compress the CO2 released by power plants into the atmosphere, and store it in depleted oil fields, natural gas fields, or other safe underground places for subsequent oil extraction, smelting, automobiles, etc. Industrial use.
However, in the process of CO2 reacting with amines, it will also produce amine emissions that endanger public health and ecosystems. Effective monitoring and prediction of amine emissions from different power plants has become a major difficulty in carbon capture.
Recently, a research team composed of EPFL and Heriot-Watt University has developed a machine learning method that can more accurately predict the emissions of harmful gases such as amines during carbon capture based on past data from power plants. The paper is currently published on ScienceAdvances.
Detailed Paper: Using Machine Learning to Solve Chemistry Problems
1. Pilot workshop test
Carbon capture plants are complex because process models typically focus on capturing steady-state operations. However, the design and operation of current and future power plants need to take into account the increasing share of renewable energy generation, which is intermittent and irregular, and therefore also needs to take into account the dynamics and multiplicity of operations outside the steady-state . variable behavior.
In order to simulate the intermittent operation of future power plants, researchers conducted a series of pressure tests on the capture device of the pilot plant of the Niederaußem power plant in Germany, trying to find out the relationship between the intermittent operation of the power plant and the amine emissions.
Simplified schematic diagram of the pilot plant for post-combustion carbon capture in Niederaußem
Although the experiment has accumulated a large amount of data capturing the behavior of the plant, it cannot be used to qualitatively predict future amine emissions, because in addition to the stress test, there is another variable in the experimental process - the intervention of power plant professionals to ensure that the plant during the experiment. safe operation.
2. Get the data set
In the pilot factory experiment, researchers collected data every 5 minutes and accumulated a huge amount of data. How to convert these data into data sets that can be used by machine learning models has become the focus of research.
The researchers' method is to represent the time-dependent process and emission data as an image (data matrix), create a prediction model based on this, and then use machine learning technology for pattern recognition to predict amine emissions.
In this notation, the plant defines a state eigenvector x(t) at a given time t, where the p elements represent process variables (such as flue gas temperature and water wash temperature).
Take the factory state vectors of t time stamps to get a t × p matrix. This matrix can be seen as a "picture" connected to the future emission curve y(t).
Data representation diagram
The data used in this experiment can be thought of as an "image" where:
width = length of input sequence (T)
height = number of parameters p
Color = value of parameter xj at a certain time ti
Next, patterns in the plant's historical imagery are linked to specific future emissions. To this end, the researchers used a gradient-boosted decision tree model to combine rows describing different parameters and emissions into one long vector. Train the model with a quantile loss to obtain an uncertainty estimate.
When evaluating uncertainty, the researchers used a temporal convolutional neural network (temporal convolutional neural network) that supports Monte Carlo dropout, and presented the results obtained with this model in note S8.
With this dataset, a machine learning model can be developed for data analysis with the help of data science methods.
3. Insights into amine emissions from machine learning
Next, the machine learning model can be used to make the following predictions:
1. Future emissions (real-time): Based on historical & current operations and emissions, predict what the emissions will be in the next x hours
2. Causal impact analysis of data: To measure the impact of a specific pressure test on amine emissions, a baseline is required to provide amine emissions without pressure tests
3. Reduce amine emissions: Use the model to predict the emissions under "hypothetical" situations, such as whether lowering the washing temperature will affect the emissions
Using Machine Learning Models to Predict
Amine emissions for next 2 minutes, 1 hour, 2 hours
A post-90s student with a Ph.D., who has been deeply involved in chemistry for 8 years
The paper was published by a research team co-led by Professor Berend Smit from the Faculty of Basic Sciences at EPFL and Professor Susana Garcia from the Center for Carbon Solutions Research at Heriot-Watt University in Scotland.
Among them, the student who developed a machine learning method to transform the amine emission problem into a pattern recognition problem is Kevin Maik Jablonka, a post-90s doctoral student in Professor Smit's group.
The first author of the paper, Kevin Maik Jablonka
Kevin studied chemistry at the Technical University of Munich, Germany. After graduating in 2017, Kevin entered the Swiss Federal Institute of Technology in Lausanne to continue his master's and doctoral studies, and continued his studies in the field of chemistry.
From 2014 to 2022, Kevin spent 8 years building a deep understanding of chemistry and chemical engineering. During this period, he also integrated chemical research and artificial intelligence through the study of applied data science and machine learning, and improved chemical engineering . The efficiency and accuracy of field research is a proper post-90s student.
As many senior figures in the chemical field have said, machine learning may have a greater impact in the fields of chemistry and process engineering than computer vision.
In CV application scenarios, the basic features of images learned by the model are often closely related to the way the human brain perceives images, such as target detection and face recognition.
However, in industrial scenarios, humans often lack understanding of the basic mechanisms, but through machine learning, researchers have discovered the basic rules of mapping from parameters to target observations, and made predictions for hitherto unpredictable phenomena.
In the case of predicting amine emissions from power plants, machine learning surpasses traditional methods and is believed to provide a new perspective on complex chemical processes that has the potential to revolutionize the way coal-fired power plants operate in the future.
Artificial intelligence will be more applied to basic scientific research to provide power, improve efficiency, and accelerate the implementation of scientific research results. How do you see the development of the second half of AI for Science? What breakthroughs will it bring, and what challenges will it face? Welcome to leave a message to share your views and opinions~