​Zhou Fei: Discussion on the R&D path and application mode of energy and power large model

By Zhou Fei

( Senior Engineer at China Electric Power Research Institute, Project Leader of Key Technology Research on Heterogeneous Fusion Brain-Like Computing for Power, Executive Leader of Distributed Computing Scheduling and Collaborative Training and Reasoning Project for Power Artificial Intelligence Large Model)

Generative artificial intelligence (AI) big models have become a research hotspot in the current third wave of AI, representing the latest development trend of AI technology and will lead three revolutions: the human-computer interaction revolution, in which natural language communication between humans and machines becomes more natural and efficient; the cognitive collaboration revolution, in which big models use their own cognitive abilities to work with humans to solve complex problems; and the computing paradigm revolution, which will promote the transformation of AI computing systems from those dominated by central processing units (CPUs) to those dominated by graphics processing units (GPUs).

In the field of energy and power, generative AI technology is developing rapidly, and multiple industry models have emerged. For example, the Guangming Electric Power Model, the first multi-modal industry model with a capacity of 100 billion yuan in China, can provide a "super brain" for the safe and stable operation of the power grid, promote the consumption of new energy, and provide good power supply services; the Kunlun Model supports the professional massive data modeling needs of oil and gas new energy, refining and new materials, etc.

From the perspective of "the essence of the world is mathematics", all problems can be described by a high-dimensional nonlinear set of mathematical equations. From the grand energy, social and power complex systems to the microscopic devices, they can all be described and calculated using mathematical thinking.

People also expect to find a tool through artificial intelligence technology to efficiently describe complex high-dimensional nonlinear systems, and then find an analytical method to observe, predict and make decisions. The current reality is that the new generation of large language models based on neural network technology still rely on statistical probability to be equivalent to or approximate any complex high-dimensional nonlinear dynamic system. This determines that the large language model only has statistical correlation, lacks logical causality, and does not have the ability of logical reasoning. 

The current research and development technology route adopted by a large number of industry big models is: select a general large model as the base model, and then use a large amount of industry common data for incremental pre-training to make the general model adapt to industry characteristics and master professional knowledge in the field. Finally, industry experts annotate the model output results, and then combine multiple rounds of expert feedback reinforcement learning to further improve the model's professionalism and accuracy in industry tasks.

Although this R&D path can quickly build large models that adapt to industry needs, it has some shortcomings: incremental pre-training relies on a large amount of high-quality industry data, and the collection, cleaning and labeling of industry data are expensive; expert feedback learning requires the participation of a large number of professionals, which is time-consuming and highly subjective; the physical properties of energy and electricity are not deeply embedded in the large model construction process, resulting in the training and inference results cannot fully meet the requirements of trusted artificial intelligence.

Over the past year or so, while serving as the project leader for brain-like computing and computing power networks for future industries, I have gradually come to realize that the traditional AI research and development path, namely the training method based on large amounts of data, large amounts of parameters, and large amounts of computing, can no longer meet the performance requirements of actual applications, especially in the power industry, a field with high timeliness and high precision requirements. The traditional path faces bottlenecks.

At present, the application modes of large models in the industry mainly include web page access, interface call and private deployment. Web page access is more common, and users access large models online through browsers; interface calls integrate large model functions into the business system of enterprises through open application programming interfaces (APIs); private deployment deploys large models to the local environment of enterprises to ensure data privacy compliance, but has performance requirements for local computing and reasoning hardware.

The shortcomings of the above three application modes are: the web page access method is weakly related to the specific business process of the enterprise, and it exists more as an independent tool; the role of interface calls is often limited to specific task processing, and it is impossible to truly embed business flows and information flows; private deployment has high costs and technical barriers, and currently remains at the level of auxiliary tools. In short, the three modes are more of an upgrade of traditional human-computer interaction, and it is difficult to achieve true cognitive collaboration and computing paradigm change, and the model potential has not yet been fully released.

In general, the application of large models is still at the stage of personal (To C) application. Real industry (To B) applications require models to solve business problems. It cannot be just question-and-answer, or assist decision-making through serial operation. It should be a serial-parallel or mutually collaborative operation mode, and ultimately achieve a leap from a human-machine collaborative interface to a cognitive collaborative function. 

The root of the problem is that the big model achieves business cognition but cannot be put into action. Therefore, the above three application modes fail to achieve participation in all business links and embedding in the entire business architecture.

my country's power system is the world's largest, most complex, and fastest-transforming man-made giant system. There are many challenges in terms of power balance, system regulation, equipment operation and maintenance, and personnel operations. It is necessary to use advanced artificial intelligence technology to improve the intelligence level of state perception, operation cognition, and control decision-making. "Tailor-made" training of large models in the energy field will find a feasible new path for the integration of data and reality.

The big model needs to be integrated into all business links of the power system. An effective solution is to combine the embodied intelligence that has emerged in recent years with intelligent agents, build an embodied intelligent agent for power, and add an "actionable body" to the big model.

The core of the power embodied intelligent body is to rely on the cloud-edge computing power base, use large models and small business models to build intelligent bodies, and use embodied evolution technology to guide the autonomous learning of intelligent bodies. The small models in the intelligent body are responsible for performing specific tasks, and the large model is responsible for scheduling and arranging the small models, and initiating cloud-edge co-evolution with the cloud-side industry large model as needed. The embodied intelligent body can realize cross-level and cross-scenario data, knowledge, and model collaboration, and realize closed-loop evolution from data perception to analysis, decision-making, planning, and action, ensuring real-time perception, accurate analysis, scientific planning, and rapid decision-making of power business.