Can large AI models enable the energy system to replicate the success of intelligent driving?
Header Image |
Autonomous driving isn't about building cars; it's about training a brain that can understand the world. Similarly, the intelligent transformation of the energy system isn't about simply piling up equipment; it's about creating a "digital and intelligent entity" that can self-perceive, weigh options, and make decisions.
Essentially, they follow the same path, enabling large models to not only understand language but also control reality.
However, in the energy sector, it's triggering a silent revolution. The main battlefield isn't in coal, photovoltaics, or wind power, but in algorithms, models, and intelligent agents. It's no longer the dispatchers who control the power grid, but the energy AI behind the scenes that understands physics, can play strategic games, and can evolve.
So, who is training the large energy models? Can it achieve "autonomous decision-making" in complex environments like intelligent driving?
Is it Feasible to Train the Power Grid like Training a Large Model?
According to the latest data compiled by Tianyancha Media, China's power system is expected to reach 2 billion kilowatts by 2030.
At this time, AI isn't just a fancy little tool; it's the "nerve center" of the entire power grid. It can accurately grasp the load pulse, closely monitor the output of wind and solar power, and precisely deploy energy storage and hydropower. It can comprehensively coordinate the entire process of where the electricity comes from, where it goes, and how it's distributed.
Facing the dual fluctuations of peak electricity consumption and low output, it can weigh the pros and cons, make judgments, and mobilize resources within milliseconds, truly achieving power regulation by AI.
The protagonist of the new power system has changed.
In 2025, a large number of large models in the energy industry were launched, not for show, but for real implementation.
On June 28th, it was the China National Energy Group that revealed its "trillion-level" trump card. Its globally first trillion-level large model for the power generation industry, "Qingyuan," was officially released. It not only horizontally covers business lines such as thermal power, hydropower, and wind power, but also vertically connects 75 scenarios such as equipment maintenance, power trading, and safety management, and deploys more than 41 intelligent agents.
▲ Picture: Release of the Globally First Trillion-Level Large Model for the Power Generation Industry, "Qingyuan"
In addition, "Da Watt · Power Control" was released by China Southern Power Grid. It's a large intelligent simulation model for the power system. It can compress the operation simulation of a large power grid, which traditionally takes several days to calculate, to a "second-level response," adding an AI wing to dispatching optimization, accident analysis, and power planning.
In the past power system, everything relied on humans. Dispatchers flattened the load based on experience, engineers checked equipment by listening to sounds and feeling temperatures, and traders set prices based on intuition. Now, all of this is being "taken over" by models, and they learn faster than humans.
How does it achieve this?
The key is "hierarchical modeling." At the bottom layer, there's an L0 general brain that can read images, understand semantics, and recognize trends, laying the foundation for overall dispatch. Moving up, from L1 to L3, it's refined layer by layer according to thermal power, wind power, and nuclear power, directly learning how to predict power, adjust equipment, and ensure safety.
People are no longer training a tool, but cultivating a "superbrain" that can understand the entire energy system. Behind every kilowatt-hour of electricity, it's no longer humans making decisions, but AI making choices after weighing countless costs. This isn't just a technological upgrade; it's an evolution of the wisdom of the energy system.
For example, on a certain wind turbine in a wind farm, "Qingyuan" can access real-time signals such as vibration and temperature, automatically determining whether the blades are about to fatigue and break; in the dispatch center, the large model can combine multi-source data such as meteorology, hydrology, and load changes to automatically generate the optimal power generation and transmission and distribution combination plan.
Of course, behind this kind of intelligence, it doesn't rely on a few experts' guesses, but on the reshaping of the industrial chain. Taking the "Qingyuan" model as an example, it connects the world's largest installed capacity, trillion-level data assets, a thousand-person AI team, and more than a hundred verified business scenarios.
People may underestimate the scale of this transformation, just like no one expected back then that a mobile phone could fit a camera, MP3, phone, and game console in one pocket.
Now, similarly, no one dares to easily believe that an AI model can really handle dispatching, trading, load forecasting, and equipment maintenance at the same time.
But it's really here. The large energy models are connecting all these things, just like autonomous driving, turning the previously fragmented systems into a self-thinking whole.
This isn't just an auxiliary tool; it's a reconstruction of the way of thinking.
Behind Every Decision, the Model is "Weighing the Costs"
AI faces a major test in the energy sector, which is to handle the complex decision-making challenges brought about by the transformation.
In February 2025, the National Development and Reform Commission and the National Energy Administration jointly issued Document No. 136, announcing that new energy has officially entered the market. The fixed electricity prices and full guarantees that new energy assets used to enjoy have been cancelled, and they now face market competition directly. They must participate independently in multi-level power markets such as spot, medium- and long-term, and auxiliary services.
This isn't just a simple mechanism change. Behind it, no decision is easy. Every order placement, every transaction, and every curve may affect profits and even development. If in the past, the competition was about having more stations and better resources, now it's about whether the system can weigh the costs in real-time. And AI has become the core role that can build a closed-loop risk control system in complex and uncertain situations.
Automatically generating real-time dynamic reports of multi-scenario information and making real-time judgments on whether to clear and whether to hedge can all give full play to AI's capabilities, not to mention that it can obtain real-time information on power station output, market node prices, equipment operating status, and meteorology.
The establishment of this entire mechanism is based on the full-chain connection between SCADA (Supervisory Control and Data Acquisition), trading algorithm systems, and risk engines. In the future, energy traders may not be the ones who understand the market best, but the ones who trust the models the most.
AI can integrate data such as meteorology, water conditions, equipment status, and market prices into a unified decision-making system, transforming those judgments in human experience like "this order is cost-effective" and "the equipment should be stabilized" into mathematical modeling and strategy optimization.
It knows that bidding at a low price now may be in vain, and bidding two hours later may bring greater profits. So, it no longer blindly pursues "quick entry," but "cost-effectiveness."
In equipment maintenance, the large model no longer waits for a breakdown to handle it, but offers suggestions: "This bearing can still last for 28 hours, but there's an 80% probability of a slight abnormality if it exceeds that." So, maintenance no longer relies on experience to distinguish abnormalities by sound.
In dispatch operation, AI isn't just an auxiliary reference, but actively offers suggestions or even executes automatically.
An intelligent simulation model like "Da Watt · Power Control" can run tens of millions of data simulations within seconds, predicting the safety risks, load balancing capabilities, and economic costs of different dispatch plans.
According to the comprehensive information from Tianyancha, the model is a scientific decision-making basis formed through massive data deduction, multiple rounds of verification, and continuous feedback. For example, in new energy power prediction, the AI model introduces time series modeling + deep learning + physical fusion modeling, and even injects prior knowledge, allowing the model to know in advance that sand and dust weather will affect photovoltaic efficiency and wind shear will disrupt wind power output.
The costs weighed by the model aren't just electricity costs and efficiency, but also safety, sustainability, and user experience.
This is exactly the sign that the energy system has entered the era of intelligent decision-making.
The Battle for the Energy Brain is a New Battlefield for AI "New Infrastructure"
If in the past, the competition among large models was about who could be more like a general artificial large model, then in the energy industry, the competition among large models has changed dimensions.
Who can create "dispatch rights" for the industry? Who can gain the "right to think" in the energy system?
This is a real new infrastructure competition.
According to the comprehensive information from Tianyancha, today's large models aren't just smart heads in the corpus; they've become the central control brains of the entire energy system. It's not about being able to explain clearly, but about calculating accurately, adjusting effectively, and saving money. Whoever can have such an energy brain will have the right to speak in the future of the entire industry.
And whoever writes this new "protocol" first will master it first. Don't forget, the speed of large models depends on the combination of computing power, data, and algorithms. And the energy industry has a natural advantage: an extremely large-scale, highly complex, and deeply closed-loop data system.
A data structure of a national-level energy group is itself a natural AI experimental field.
So, in essence, the energy brain competition isn't about which model is faster or which output is more accurate, but about truly mastering a complete set of modeling paradigms and operating models. This is the underlying control power, the leading power at the algorithm level of the future energy system.
Moreover, technology defines scenarios, and scenarios feed back to standards.
Once a company successfully runs through more than 50 scenarios, deploys hundreds of intelligent agents, and achieves a real-time dispatch closed-loop, it may master the future reference standards of the industry. This first-mover advantage will ultimately evolve into the rule construction of the algorithm framework.
It's not about building factories or towers, but about building algorithms, standards, and protocols.
Of course, the energy system of the future won't lack electricity, but "smart electricity."
Just like intelligent driving isn't about the engine of the car, but about who controls the closed-loop of the perception system, decision-making system, and execution system, the core of the future energy system isn't about how much is generated and supplied, but about who understands better how to adjust, save, and coordinate.
Whoever can train the next super large energy model can build a higher-dimensional energy order.
And this is exactly the sign that large models have truly entered the core area of productivity.
So, what's the endgame?
It means that every power plant, every photovoltaic panel, and every wind turbine will no longer be an information island, but will be connected to a thinking energy network; it means that intelligent dispatch, operation and maintenance, trading, and safety will all run in parallel, and humans will just ask questions and AI will answer at any time; it means that energy AI will become a part of the infrastructure like water, electricity, and road networks.
Compared with the high-speed development of intelligent driving, the trials here are deeper and slower. Because every judgment must be worthy of the stability of the entire power grid.
If the first wave of AI's explosion was about understanding language, the second wave may be about understanding energy. And what's really worth imagining is that when it's embedded in the entire energy network, what people may face isn't just an AI that can answer questions, but a smart base that can actively support the light of cities.
This article is from the WeChat official account "Zero State LT" (ID: LingTai_LT), author: Lin Feixue, published by 36Kr with authorization.