/Chaowentong/December 28, 2012, Guangzhou News - The third phase of cooperation between Baichuan Plan and Qingteng, the interviewee is Xia Jiantao, founder and CEO of Quanying Technology.
For a thermal power plant, coal is the most important resource and the largest source of cost. 70% of the money of the power plant is spent on the purchase of coal. When the price of coal continues to rise (nearly three times in recent years), the demand for energy conservation and consumption reduction in the power plant becomes more urgent.
In a larger sense, China uses more than 4 billion tons of coal for thermal power production every year. If the energy efficiency is improved by 1%, it will save up to 40 million tons of coal, and the coal price will be nearly 60 billion yuan. Calculated by 2.7 tons of carbon dioxide equivalent produced by each ton of coal combustion, it can reduce nearly 100 million tons of greenhouse gases every year.
Thermal power enterprises put energy conservation and consumption reduction in an important position, but due to the development of existing soft and hard technologies, the effect of energy efficiency improvement is poor. The thermoelectric boiler equipment in China is the world leader with superb technology. With decades of tireless transformation by experts in thermal power plants, the combustion efficiency of power plant boilers has averaged more than 95%, and the energy efficiency has approached the absolute peak. Experts have tried their best to improve even 0.5%.
Xia Jiantao thought of another way to create an industrial Internet platform for thermal power production, train artificial intelligence models using big data, and combine dozens of such digital twin models for different equipment operations to precisely control the whole process of thermal power production. After more than 4 years of practice in more than 60 power plants, this action has achieved remarkable results. According to the measured data, the thermoelectric intelligent system invented by Xia Jiantao and his colleagues can bring about an average energy efficiency improvement of 2% to 5% for each power plant every year, which can be converted into 10,000 to 20,000 tons of coal.
Xia Jiantao is the founder and CEO of Quanying Technology. He completed his undergraduate, postgraduate and doctoral studies in Northwestern Polytechnical University and his post doctoral research in Nanyang University of Technology in Singapore. His research direction is AI. However, when he graduated, AI was in its third low ebb and had not yet mature applications in the industrial field. After returning to China, Xia Jiantao has been engaged in technology research and development for more than ten years in Delta Group, a world-famous power electronics and industrial automation group, providing industrial enterprises with software and hardware solutions for automation control, power electronics and visual operation. In 2012, Xia Jiantao learned about the industrial Internet for the first time during his visit to GE, which gave birth to the idea of using industrial Internet technology to serve industrial enterprises. In 2016, Xia Jiantao left Delta and established Allying Technology with several big data, artificial intelligence experts from IBM and Microsoft and experts from the thermal power industry.
Up to now, Quanying has a business layout in the major thermal power production provinces and cities in China, with more than 60 online customers. However, the company is still unable to be responsible for its own profits and losses. Xia Jiantao revealed that when the number of online customers reaches 100, the company can enter into a break even state. Tencent Investment, Bohua Capital, Kaihui Energy Fund, Hillhouse Venture Capital, Mingshi Capital, Linear Capital and Songhe Capital have invested in Quanying.
At the end of November, we talked with Xia Jiantao of Tsinghua Qingteng Future Science and Technology School (Phase IV) about the advantages and difficulties of digitalization in the thermal power industry. In response to past exploration and future challenges, as well as the particularity of industrial AI, Xia Jiantao sincerely shared his own experience and thinking based on his own practice.
The interview text below has been edited and modified.
one
Qingteng: What is thermoelectricity?
Xia Jiantao: By burning coal in boilers, water is turned into high-temperature and high-pressure steam, which is used to drive steam turbines to generate electricity. This is called thermoelectric production process.
Some large industrial enterprises, such as petrochemical, biopharmaceutical or paper-making enterprises, build their own thermoelectric units in the plant area to supply power for the plant area in order to reduce the cost of electricity. In addition, in the process of industrial production, some still need heat energy, such as oil smelting, pharmaceutical fermentation, and the waste heat after power generation can be used. This is cogeneration.
In addition, the heating in the north is also the waste steam after power generation, which is changed into hot water after heat replacement and flows to each household through the pipe network. Therefore, heat and power are applied in three sectors: thermal power generation, cogeneration of power generation and heat energy comprehensive utilization, and urban central heating.
Qingteng: When you started your business in 2016, you decided to build a standardized industrial Internet platform instead of personalized customization. Why?
Xia Jiantao: One of the reasons is that I don't think entrepreneurial enterprises have the ability to customize products for enterprises in terms of scale and manpower.
But which industries can be standardized? I have been in Delta for more than ten years, and I still know about all walks of life in industry. To standardize products, we should first narrow down our product functions and focus enough.
Second, what we want to solve is the core pain point of an industry, not many problems.
Third, the industry should have a standardized physical foundation and data.
Under the guidance of this idea, we took stock of two major industries in the industry. One is discrete industry, such as 3C manufacturing, garment manufacturing, mechanical processing, etc. The biggest problem in these industries is that the underlying digital foundation is not very good. The equipment either has no data or the data is not easy to get. The process is also controlled by people, so there is no way to provide standardized solutions. When I was in Delta, I led a team of 250 people to upgrade the automation of production lines in the group. I found that our factories are different, and we need to customize everything we do for ourselves.
The other direction is process industry. Because the core problem we want to solve is the intelligence of the production process, let's look at the process. If the process is physically standard, it can be standardized. We looked at petrochemical, coal chemical, salt chemical, paper making, thermal power and other industries, and found that the process of thermal power industry is the most standardized. Although it also has six processes, these six processes cover all coal-fired power generation processes.
Although there is a lot of demand for intelligent thermal power plants, the core is energy conservation. 70% of the cost of thermal power plants is coal, and the price of coal has increased several times from 500 to 600 years ago to 1561 tons now, resulting in the suppression of the operating efficiency of power plants.
Therefore, we hope to use artificial intelligence technology to reconstruct the thermal power production process in the digital space, and calculate the control parameters of the power plant every hour and every second, so as to maximize the operating efficiency of the power plant.
Qingteng: Do customers accept the SaaS model?
Xia Jiantao: The thermal power plant does not have the conditions for privatization deployment. First of all, to achieve intelligent control of the thermal power production process, it is necessary to build a digital twin model of the entire power plant. A power plant has as many as 40000 or 50000 data points as 2000 or 3000 data points. The amount of calculation is very large, and traditional computers cannot bear such a task. Second, the power plant is full of boiler experts, steam turbine experts, automation experts, and there is no expert who knows IDC operation and maintenance, nor computer experts, so privatization delivery is impossible. Third, thermoelectric production is a dynamic process. It is impossible for us to adjust the parameters at one time and deliver them to customers for use. For example, the heat conduction characteristics of the boiler will deposit dust and coke with the increase of the boiler service time, and the heat conduction efficiency will change, which requires the system to learn the dynamic parameters and output reasonable control decisions.
We adopt the cloud+edge+end system. Data encryption goes to the cloud, model training is done, and specific control calculation is done for the model at the edge. Therefore, this system needs to be deployed in the cloud.
In addition, since the production of power plants is dynamic, services are needed, whether they are computing services of artificial intelligence models or expert services. But our services are provided to customers through the back end of the cloud platform, and we can't follow the one-time charging model.
Qingteng: What's the charge?
Xia Jiantao: According to the capacity of the boiler unit, 1 million to 5 million yuan of SaaS service fee will be paid every year from small to large. For SaaS services that charge more than one year, if you want to continue charging, you must create customer perceived value, that is, it can be calculated. We have calculated that the average annual energy efficiency improvement for each power plant is 2% to 5%, which is about 10000 tons to 20000 tons of coal saving, and the cash converted is 15 million to 30 million yuan.
Qingteng: Is the customer willing to pay because the AI effect is clear and calculable?
Xia Jiantao: These can be calculated. Coal can be weighed, electricity can be measured, and steam consumption can also be measured. Customers can see whether this system is useful or not. In addition, it can also reduce carbon dioxide emissions. According to 2.7 tons of carbon dioxide equivalent emissions of 1 ton of coal, it can reduce carbon emissions of 30000 to 40000 tons. After the "3060" target came out, the energy industry was the first to bear the brunt and become the key target of emission reduction. At present, China's carbon price is 57 yuan/ton (40 euros/ton in Europe), and the excess part will be paid.
Qingteng: Is there another reason why customers can accept AI, that is, the previous energy conservation and efficiency enhancement methods have been used to the extreme?
Xia Jiantao: Yes, our country is now the largest manufacturer of boiler equipment in the world. Our boiler efficiency is more than 95%, which has reached the limit of this equipment. Unless there is significant progress in materials, such as graphene, but that is expensive, it is difficult to improve significantly.
In addition, for decades, experts in power plants have been thinking about efficiency improvement every day. They have changed it every day, with an annual cost of 45 million yuan. They have changed everything they can think of, but now it is impossible to improve energy efficiency by 1%. So when we told potential customers that we could improve the energy efficiency of the thermal power system by 3% to 5% on average, they all thought we were liars. When we opened up the market in Zhejiang that year, 9 out of 10 customers thought we were liars.
Another reason is that thermal power plants and industrial enterprises that need combined heat and power supply are mostly concentrated in third and fourth tier cities, and young people are unwilling to go to work in such places. In fact, thermal power plants are also facing severe talent difficulties.
Qingteng: Looking back, what's good about choosing the thermal power industry?
Xia Jiantao: In addition to the three standards mentioned above, there is also a guiding principle that the industry must be large enough and not too monopolistic. Otherwise, customers are all large group enterprises, and we are bound to become a project oriented company. The thermoelectricity industry is very good. China has the largest thermoelectricity scale in the world. Every year, 4 billion tons of coal is used for power generation, accounting for more than 60% of the total power generation. If our system is used, the economic value of the coal saved will be huge.
Moreover, the number of customers is very large. The five groups together have more than 2000 power plants, the large and small cogeneration enterprises together have 4-5000, private and state-owned, and the types of customers are very rich.
two
Qingteng: You just mentioned the degree of digitalization when selecting an important standard of the industry. The advantages of the thermal power industry are relatively outstanding?
Xia Jiantao: Generally speaking, the digitalization of China's process industry is better than that of discrete industry. Power generation is a high-speed reaction process, and it is very difficult to control by personnel, so the power plant has started the process of digitalization for a long time. Now all equipment has sensors, such as furnace temperature, furnace pressure, flue gas and oxygen, steam flow and turbine speed in the boiler. A series of sensors have been used. With sensors and data, automatic control systems will be used. The DCS distributed control system we mentioned was first used in power plants, and now the power plants have been 100% controlled by DCS. It is just said that today's electric production process is still judged by the data on the DCS panel of experts, giving control instructions. The boiler expert gives instructions to the boiler, the steam turbine expert gives control instructions to the steam turbine, the heat network expert gives control instructions to the heat, and the power grid gives control instructions to the power grid. Therefore, what we see is that the power plant is controlled by one operation team to achieve thermal power production.
Qingteng: The Quanying system is mainly used to help the operation team realize the thermal power production control?
Xia Jiantao: Yes, we are the intelligent control of the production process, turning the process controlled by human experience and skills into an artificial intelligence model to make decisions, give control instructions, and accurately adjust the operation process of the power plant.
Qingteng: Does AI rely on the experience of these experts in the process of learning?
Xia Jiantao: Our AI model integrates three kinds of knowledge systems. One is the industrial mechanism of thermoelectricity, the boiler combustion principle, the steam turbine work principle, the thermodynamics of the heat network, the transmission process of the electromagnetic wave of the power grid, etc. These are scientific descriptions of the electrical production process. All power plants are the same, and can be learned from the "textbook"; The second type is expert knowledge, that is, professional experience in power generation process production, such as parameter settings of various equipment. After decades of development, people have found out some common good experience; Another is the expertise of customers. There are no two power plants in the world that are exactly the same. Each power plant has its own characteristics.
Our artificial intelligence model does not learn from human experience, but from the whole process and the operating characteristics of equipment. As each power plant is different, our deployment process takes an average of four to five months. On top of the standard model, each power plant should reassemble the model according to its physical structure, which is a bit like building Lego bricks.
Qingteng: The digital foundation of the power plant is good, and the demand for cost reduction and efficiency increase is also strong. Does that mean that customers have a high degree of acceptance of the Quanying system?
Xia Jiantao: The experts in the power plant are all the boiler experts, power generation experts, steam turbine experts, and automatic control experts we just mentioned. Such experts are traditional production managers. They are relatively deficient in knowledge of intelligence, cloud computing, big data, and artificial intelligence, which is a cognitive challenge. Therefore, the important task of our early marketing is to tell customers about this knowledge and repeat it. With the first case, customers can come to visit.
Customers in the thermal power industry are very cautious about the use of digital tools. They must demonstrate such a system repeatedly, visit the implemented cases, and have in-depth exchanges with experts.
In particular, we control the core production process control. If the control is not good, the machine may be shut down if it is light, and a malignant accident may occur if it is heavy. In addition, some cogeneration enterprises may have hundreds of enterprises in the downstream. If the power plant suddenly shuts down, the production of all downstream plants will have to stop. This loss is very serious.
Therefore, the intelligence of industry is different from that of consumer industry. The former requires 100% accuracy. The functions of our products are also very simple. There are not so many fancy functions. The first thing is to ensure safe and stable operation.
Qingteng: Don't pursue to use the most advanced and cutting-edge algorithm?
Xia Jiantao: The algorithm is very advanced. Without advanced algorithms, energy efficiency cannot be achieved. The power plant is already a high lean production scenario.
The computer face recognition, voice recognition and deep learning that you know a lot are useless in the field of industrial production control. A typical feature of industry is small data rather than big data (relative to the consumption field). Let's talk about the construction process of a device model, which is actually a time series of small data. The construction of this model is very challenging. I did my doctoral thesis in 1998 to study small sample machine learning algorithms. How to extract rules from small samples and get results. In fact, this is a very advanced algorithm system.
