When the proportion of new energy installed capacity exceeds 50%, the traditional power dispatch mode is facing unprecedented challenges. AI is becoming the 'brain' of the new power system.
1、 The core challenge of the new power system
As of early 2026, the installed capacity of wind and photovoltaic power generation in China has exceeded 1.2 billion kilowatts, and new energy has become the largest power source. However, the intermittency and volatility of new energy bring enormous pressure to power grid scheduling - phenomena such as "extreme heat without wind" and "late peak without light" occur frequently, and traditional physical model-based scheduling methods are becoming increasingly difficult to adapt to.
The core contradiction of the new power system can be summarized as follows: the uncertainty on the supply side increases dramatically, the demand side requires faster response speed, and the safety and stability constraints of the system cannot be relaxed.
2、 AI load forecasting: from hourly to minute level
Traditional load forecasting relies on statistical regression and physical models, resulting in significant errors when faced with complex scenarios such as extreme weather and holidays. The emergence of deep learning models, especially time series models (LSTM, Transformer) and graph neural networks (GNN), has raised the accuracy of load forecasting to new heights.
State Grid Jiangsu Electric Power has introduced a time series Transformer model into its provincial load forecasting system, integrating multidimensional information such as meteorological data, historical loads, and holiday characteristics to achieve ultra short term load forecasting for the next 72 hours. After the system was launched, the accuracy of daily load forecasting increased to over 98.5%, and the prediction error in extreme weather scenarios decreased by about 40%.
In terms of new energy power prediction, the AI power prediction project jointly developed by Huawei Cloud and Huaneng Group uses a multimodal fusion model to reduce the root mean square error of wind power prediction by 25%.
3、 AI intelligent scheduling: from experience driven to data-driven
Power dispatching has long relied on the experience judgment and offline simulation calculation of dispatchers. AI dispatching decision system is changing this mode.
The "AI Dispatcher" system piloted by Southern Power Grid in the Guangdong region is based on a deep reinforcement learning framework, which takes the power grid topology structure, real-time measurement data, and maintenance plan as inputs to output the optimal scheduling strategy. In the pilot area, the response speed of AI scheduling decisions has been shortened from minutes to seconds, and the decision accuracy in N-1 fault scenarios exceeds 95%.
The distributed resource aggregation and regulation platform deployed by State Grid Shanghai Electric Power has achieved minute level observability, measurability, and controllability of over 100000 distributed photovoltaics by clustering, predicting, and coordinating massive distributed resources through AI algorithms.
4、 Collaborative optimization of AI+electricity market
By the end of 2025, China's electricity spot market will have been basically established nationwide. The application of AI in the electricity market is shifting from auxiliary analysis to trading decision-making.
Power generation companies use reinforcement learning models to optimize their pricing strategies, maximizing power generation revenue while meeting grid safety constraints. Electricity sales companies use Transformer models for daily and real-time market price forecasting, assisting in purchasing decisions and risk management.
It is worth noting that the integrated model of "source grid load storage" is giving rise to new AI application scenarios - through AI algorithms, the generation side, grid side, load side, and energy storage side are modeled and optimized in a unified manner to achieve optimal operation of the entire power system.
5、 Outlook and Challenges
The prospect of AI empowering new power systems is broad, but there are still several challenges to be faced: interpretability - power dispatch is crucial to grid safety, and the "black box" problem of AI decision-making requires breakthroughs in interpretable AI technology; Data quality - There are issues such as missing and delayed data collection in the power grid, which affect the effectiveness of model training; Small sample learning - there are very few historical samples in extreme working conditions, and the ability to learn from small samples needs to be improved; Real time requirements - Power grid control requires millisecond level response, which places extremely high demands on model inference speed.
Overall, AI is moving from an "auxiliary tool" to a "core component" in the power system. With the continuous breakthroughs of big model technology in time series prediction, multimodal fusion, and intelligent decision-making, a new type of power system that is smarter, more reliable, and greener is accelerating.
This article is a comprehensive compilation of technical reports and industry information publicly released by State Grid, Southern Power Grid, and Huaneng Group.