Special topic planning (Part 2) | How to achieve deep coupling between big models and industries?

Liu Su : China National Petroleum Corporation is fully committed to exploring intelligent development projects and carrying out the "artificial intelligence + " action with Kunlun Big Model as the core. The R&D team carries out work around the "five ones", namely training a set of domestic leading industry big models, implementing a series of innovative application scenarios, forming a set of high-quality industry data sets, building a centralized and unified AI middle platform, and building a resource-sharing intelligent computing center.

Kunlun Big Model is a big model for the energy industry. The top-level design includes four layers of architecture: L0 general basic big model, L1 industry big model, L2 professional big model, and L3 scenario big model. It covers business areas such as exploration and development, oil refining and chemical industry, oil and gas storage and transportation, product sales, engineering construction, and equipment manufacturing. It has the capabilities of industry knowledge question and answer, object detection, and image and text generation. It can effectively solve business problems in all links of the energy and chemical industry and promote technological innovation and industrial upgrading. Among them, the L1 industry big model training has formed 70 billion parameters of language, 300 million parameters of vision, and 16 billion parameters of multimodality, becoming the first industry big model in the energy and chemical industry that has passed the national filing. In terms of the ability to understand the energy and chemical industry, the scope of visual recognition, and the ability of multimodal fields, the industry big model has been significantly improved compared with the basic model.

The professional big model is formed through pre-training + fine-tuning training based on the industry big model . Its characteristics are that it meets the needs of professional scenarios and has strong professional attributes. It improves the operating efficiency in seismic processing, seismic interpretation, and logging processing and interpretation. In 2025 , we also plan to make real-time big models. At present, we have built more than 20 scene big models, created 4 types of application scenarios, and built an AI middle platform to support three major pipelines. In the future, we will create more industry scenarios and explore the application of new technologies such as intelligent agents.

Zhu Hong : The first multi-modal large model of the power industry in China with a scale of 100 billion yuan, the Guangming Electric Power Large Model, was released on December 19. The development of artificial intelligence by State Grid Corporation of China aims to promote digital transformation and improve the level of power grid operation and management, using a collaborative model of large models and special models. The Guangming Electric Power Large Model is a multi-modal large model with a scale of 100 billion yuan for the power industry, which can provide professional and intelligent services for the entire industry chain including power production, construction, management, operation, scientific research, manufacturing, and services.

The R&D team improves the professional capabilities of large models from three aspects: building a foundation, strengthening knowledge, and practicing thinking. It adds a cross-modal adaptation layer to the architecture, enhances multimodal fusion analysis capabilities, creates samples for the entire industry, improves knowledge understanding and generation capabilities, and uses actual cases to guide logical reasoning and self-optimization.

Guangming Electric Power's big model plays the role of an intelligent expert in more than 600 application scenarios such as power grid planning, power grid operation and maintenance, power regulation, and power supply services , realizing the synergistic empowerment of power and computing power, and assisting in the construction of new power systems and new energy systems. At present, the function of automatically generating power supply and utilization plans has been put into practical application in Jiangsu. In the future, Guangming Electric Power's big model will be committed to the coordinated development with the original model, deepening business collaboration and other fields, and forming a new ecology.

Wang Feng : China Telecom is fully committed to promoting the integration of cloud and intelligence, and building a " 1+1+1+M+N " layout, namely, an intelligent computing cloud base, a general large model base, a multimodal data set base, M large models within enterprises, and N large industry models. In the era of large models, computing power is an important infrastructure, and the development of computing power presents a "point, line, and surface" scale. China urgently needs to break through the "bottleneck" problem of computing power. As a cloud network resource operator, China Telecom provides computing power services for industry large models, and promotes the efficient supply of computing power in accordance with the "network to supplement computing" approach, that is, with user needs as the core, the computing power, storage, algorithm and other resource information of the service nodes are distributed through the network control plane, and the best computing, storage, network and other resources are provided. Distribution, association, transaction and allocation, so as to achieve the optimal configuration and use of the entire network resources.

Chu Zhengyu : Artificial intelligence must be viewed in the context of application scenarios, and the battery big model is an innovative engine for new energy asset management.

In the energy industry chain, energy storage is in the middle. With the rapid increase in China's installed capacity of new energy, the scale of energy storage and power batteries is also growing, and has entered the terawatt-hour era. However, battery safety issues are becoming more and more prominent, such as frequent battery fires, which seriously threaten people's lives and property. The existing battery management system is based on simple logical judgments, such as alarms when the voltage and temperature are greater than a certain value, which easily generates a large number of error messages. It is difficult for operation and maintenance personnel to quickly judge the actual impact on safety and cannot effectively solve the problem of battery fires. In this context, in order to solve the potential safety hazards of batteries, it is urgent to develop a large battery model, aiming to use advanced technology to improve the safety and reliability of battery management.

Guo Xin : Accurate forecasting of renewable energy is crucial to the stable operation of the power grid. In order to improve data assimilation capabilities and forecasting accuracy, the German Meteorological Service has developed a large artificial intelligence model that successfully solves the intelligent connection of different types of data (such as ground measurements, radar, satellite data, etc.) in time and space, allowing data assimilation and weather forecasting to be completed in the same model. This model not only significantly improves the accuracy and efficiency of renewable energy forecasting, but also significantly reduces the cost of forecasting services.

With the advancement of weather forecasting technology, artificial intelligence has shown great potential in the field of operation and dispatch, and has brought about profound changes. The German power grid generally adopts a double-circuit design to ensure that when one line fails, the other line can take over the power transmission, which makes the effective transmission capacity of the power grid only 50% of the theoretical capacity . However, with the continuous increase in the proportion of new energy, how to improve the transmission capacity of existing lines has become a problem that needs to be solved urgently. To meet this challenge, German power grid companies are testing the "healing dispatch" method, which is to optimize the operation of the power grid with the help of "network boosters" equipped with batteries. In practical applications, this method not only enhances the ability of the power grid to respond to emergencies, but also reduces operating costs, thereby alleviating the high cost pressure brought about by power transformation. At the same time, it will also become an important research direction of artificial intelligence in the field of power grid dispatching.

Wang Gang : Artificial intelligence and energy are inherently strongly related, and exploring the " AI+ green electricity " cooperation model is of great significance.

In the process of cooperation with energy companies, two application modes should be adopted. The first is "application first, small steps and fast running", relying on the application development platform of large model intelligent bodies to create a series of intelligent body assistants for knowledge, data, and processes. Identify high-value application scenarios in the energy industry, such as equipment inspection and maintenance, safety inspections, virtual duty, technical supervision, smart teams, "two tickets" generation and other core production scenarios for continuous empowerment. The second is top-level planning and overall planning. Relying on the group-level artificial intelligence foundation, create a centralized model training environment, gather high-quality data sample sets, precipitate and manage common algorithm models and intelligent body applications, and provide distributed reasoning capabilities for industrial units and regional companies.

Currently, iFLYTEK has cooperated with a number of energy companies. During the cooperation process, it has summarized a set of multi-dimensional methodologies including building computing power, managing data, training models, ensuring safety, and precise operations. It has built a large model system for the energy industry at the L0~L3 level, and implemented a new paradigm of "artificial intelligence + energy " that combines large and small collaborations and general and specialized capabilities .

Wu Wenjun : Deeply promoting the development of generative artificial intelligence technology so that it can be widely applied is the current focus of the artificial intelligence field.

Although the current large-scale model reasoning capabilities have been significantly improved, such as progress in mathematics, code, and logical reasoning, the existing language models have certain limitations and are still far from deep integration with the industry and practical applications.

Improving the logical reasoning ability of large models is crucial to the future development of artificial intelligence and the realization of industry applications. Pure data-driven can only construct "parrot-like" repetitive intelligent entities, so purely relying on expanding the scale to achieve an absolutely universal intelligent model is not a feasible technical route; facing a large number of vertical fields, such as complex giant system scenarios such as power grids, it is necessary to take the "combination of general and specialized" technical route to form a comprehensive system architecture that includes reasoning, memory, coordination and other capabilities, and realize the widespread implementation and empowerment of large models in the industrial field as soon as possible. This will become the main research direction in the next 5 to 10 years. In the field of industrial manufacturing, large models can be used as the basis to build professional models for various industrial tasks, form intelligent entities in various industry fields, and achieve the adaptation of large models to the manufacturing industry.

Dong Jian : Artificial intelligence standards are the key to supporting the high-quality development of the industry. At the policy level, both domestic and international efforts are being made to develop artificial intelligence. China focuses on high-quality development and empowering the real economy, the United States emphasizes technological development, and the European Union focuses on regulation. At the industrial level, the artificial intelligence industry is gaining popularity, and global technology companies are actively developing artificial intelligence, with the industry scale growth rate entering the fast lane. At the standard level, the importance of standardization of artificial intelligence at home and abroad is unprecedented.

Standards are an important technical foundation and institutional guarantee for promoting the development of the artificial intelligence industry. They play an important role in promoting technological innovation, ensuring security, promoting industrial collaboration and international cooperation. In 2024 , the Ministry of Industry and Information Technology, the National Standards Committee and other four ministries and commissions jointly issued the "Guidelines for the Construction of a National Comprehensive Standardization System for the Artificial Intelligence Industry ( 2024 Edition)". China is also actively participating in the formulation of international standards in the field of artificial intelligence, while promoting the construction of domestic artificial intelligence standardization, such as artificial intelligence management system standards, data management capability maturity standards, large model evaluation standards, and artificial intelligence computing equipment evaluation methods. In addition, the development of embodied intelligence standards, artificial intelligence security governance frameworks, and large model filings are also being promoted.

Zhou Fei : The current research and development of large models in the industry generally follows the step-by-step research and development path of "base model construction-domain knowledge injection-cognitive ability enhancement": First, build a cognitive center based on a general base model with hundreds of billions of parameters to provide a basic cognitive framework for industry applications; second, through incremental pre-training of energy and power industry data, achieve deep adaptation of industry language features; finally, adopt an expert-guided supervised fine-tuning strategy to improve the model's knowledge representation accuracy and logical reasoning ability in specific scenarios. This paradigm performs well in static business scenarios, but when dealing with the high-dimensional nonlinear and strongly time-varying dynamic characteristics of new power systems, the traditional path shows significant limitations in dynamic algorithm updates, physical deep embedding, and edge real-time reasoning. Based on this, it is recommended to focus on the core needs of high-dimensional representation, real-time observation, accurate prediction, dynamic decision-making, and other core needs of the complex dynamic characteristics of new power systems based on the first principles, and reshape the research and development and application paradigm of energy and power large models from the underlying mathematical principles. The specific implementation includes the following two aspects:

First, we will innovate the model architecture and build a neural-symbolic hybrid computing framework that integrates numerical calculation of physical equations, neural network data-driven, and symbolic logic and mathematical deduction, to achieve native expression of complex dynamic characteristics at the model architecture level, and effectively solve the pain points of the new power system, such as "explosion of computing dimensions, insufficient model accuracy, and delayed decision response."

The second is to innovate the training paradigm, create a research and development path of "native training-cognitive distillation-reinforcement learning", build a three-level collaborative architecture of "industry model-professional model-operation model", and form a closed-loop evolutionary link of perception, cognition, decision-making, and execution to ensure real-time perception, precise deduction, scientific decision-making, and rapid response for energy and power business.

Zeng Zhenyu : The rise of big models has brought about a major change in the interaction mode of business software. In the past, complex interaction modes and processes required professionals to remember a large number of operation methods. Now, with the help of big models, the interaction is gradually shifting to more humanized methods such as natural language, pictures, and voice. The communication between people and machines is becoming more and more like communication between people. Behind this is the support of large language models or multimodal models with tens or even hundreds of billions of words.

With the continuous development of technology, big models are expected to be deeply applied in more industries and fields, further promoting industrial upgrading and innovative development. But at the same time, we also need to pay attention to the interpretability and data security issues faced in the development of big models. Through continuous technological innovation and research, we will continuously improve and optimize big model technology to better serve social and industrial development.

Wang Guan : General large language models represented by GPT have limitations such as inaccurate numerical calculations, loose logical reasoning, and inability to conduct online autonomous learning and iteration.

Based on the probabilistic system, the general large model with Transformer as the main architecture learns the statistical laws of language rather than mathematical formulas or numerical calculation methods. The large model learns relationships and associations through data, relying on probabilities rather than logical steps. Arithmetic requires step-by-step reasoning and precise steps, which are difficult for these models that are mainly used to recognize language patterns to handle. Therefore, even if the model has seen many examples of arithmetic problems, it may only make approximations or guesses based on the learned patterns without understanding what "calculation" is.

The large model is not designed with special arithmetic logic. In a model based on a probability system, the knowledge learned and the output content essentially come from the probability distribution, which means that the probability of the correct answer or the best answer is less than 100% , which is completely different from the symbolic system.

To solve these problems, the integration of symbolic system and probabilistic system should be introduced, and the large model needs to make adjustments in the underlying iteration mechanism to achieve accurate calculation, rigorous logical reasoning and online learning. The core of online learning is the ability to interact with the environment and human users, that is, interactive reinforcement learning technology, which updates the model in real time through the changes in the environment of each user individual and intelligent agent. As NVIDIA's Huang Renxun said, "reasoning is training" is the core technological breakthrough for the future large model to achieve industrial implementation.

Lu Jixiang : NARI Group has been deeply involved in the research and development of Guangming Electric Power's large model. Guangming Electric Power's large model integrates professional knowledge and provides new ways for the application of the power industry. Power system parameters and instructions require precise values. Inspired by the achievements of the 2024 Nobel Prize in Chemistry winners, it is possible to consider building a specialized model for power grid operation, providing accurate calculation services through intelligent agents, and breaking through the technical difficulties of the power system. By continuously collecting and organizing data, improving model architecture, integrating power physics laws and professional knowledge, adding physical formulas to the objective function, adding neural networks and attention mechanisms when building models, etc., it is expected to provide technical support for the intelligent transformation of power systems in the future.

Lai Shaoming : The rise of artificial intelligence technology has brought about tremendous changes to the public opinion ecology, media landscape, and communication methods. Big models can replace a large number of general and basic content production, shorten the cycle, improve the efficiency of editing and compiling, and provide ideas and creativity in specific fields such as creative planning. Moreover, artificial intelligence has also brought about profound changes in content dissemination and interaction. After more than a year of practice, Yingda Media Group has extensively applied artificial intelligence big model technology in the production of new media content; in the future, based on the massive data of the media resource library and the big model of the Guangming Electric Power Industry of the State Grid Corporation of China, it will create a Yingda Media vertical big model consisting of "general model + power knowledge + scenario application ", and on this basis, empower various AI tools to accurately serve the entire process of planning, collecting, editing and publishing, and promote the transformation and upgrading of news dissemination and knowledge services in the energy field.