Huawei's Intelligent Agent Safeguards Electricity Usage for Households in Xiangtan

As a traditional industrial city in Hunan Province, Xiangtan has a massive industrial load volume and a significant peak-valley load difference. With distributed PV, energy storage systems (ESS), and charging piles being connected to urban and rural distribution networks on a large scale, the power flow in transformer districts has become bidirectional and voltage fluctuations frequent. As a result, the operating conditions of distribution networks are growing increasingly complex, and the traditional manual screening and remediation model for transformer district governance can no longer adapt to the current reality of the local power grid.

State Grid Xiangtan Power Supply Company (Xiangtan Company), in partnership with Huawei, has adopted a joint innovation model combining transformer district governance and AI to build an intelligent diagnosis agent for transformer districts. This creates a 24/7 digital and intelligent diagnosis expert for transformer districts, transforming outstanding field practices and expert experience into AI algorithms to achieve AI-driven automation, effectively reducing the workload and fundamentally improving efficiency.

An Urgent Need for Transformation in Traditional Transformer District Governance

In the past, system loss management in transformer districts at Xiangtan Company relied heavily on manual analysis. Constrained by personnel capacity, and with various system rules typically applicable only to specific scenarios, the company could analyze a maximum of 50 transformer districts per day. Consequently, analysis had to be prioritized sequentially, making it impossible to cover all anomalous transformer districts. A large number of hidden hazards, such as illegal electricity usage, could not be handled in a timely manner, causing economic losses to Xiangtan Company. Analysts' varying levels of experience also led to inconsistent precision and quality in work orders. O&M personnel frequently had to visit each household door-to-door. Frequent trips to sites had become a routine part of their work.

Regarding short-circuit and leakage faults in transformer districts, on-site O&M personnel mainly relied on expert experience to conduct manual investigations step by step. The fault handling process was time-consuming and inefficient, taking several days or even three to four weeks for complex faults. Especially for new employees who lacked experience in fault analysis, misjudgment and omission were likely to occur, leading to varying quality in fault handling. The fault handling process for short-circuit and leakage faults fell into a predicament exacerbated by difficult onsite location and prolonged troubleshooting time.

Faced with this situation, Xiangtan Company decided to build an intelligent diagnosis agent for transformer districts to improve the level and efficiency of transformer district governance. The company invited Huawei as an important technical partner for in-depth cooperation, and chose Xiangtan as a pilot site to build an intelligent marketing R&D base, promoting the in-depth integration of AI and marketing services.

Building an Intelligent Diagnosis Agent for Transformer districts Based on the Deep Diagnosis Technology Powered by Altas Computing Power

Early in 2026, Xiangtan Company partnered with Huawei to jointly innovate the intelligent diagnosis agent for transformer districts, and established a dedicated team focusing on AI and marketing. The two parties worked together at the Spark Innovation Studio of the customer service center of Xiangtan Company.

After three months of joint R&D, the intelligent system loss diagnosis and short circuit and leakage diagnosis functions of the agent were launched in quick succession. During the three-month pilot period, the agent has assisted on-site personnel in managing 53 transformer districts with exceptional system loss and guided them in the prompt investigation of 14 short-circuit and leakage hazards in transformer districts.

These pilot verifications demonstrate that the agent has excellent capabilities in system loss diagnosis and short-circuit and leakage fault location. It can help on-site personnel accurately and quickly locate faults, effectively improve the quality of transformer district governance, and significantly enhance the efficiency and accuracy of governance.

Different from traditional rule-based calculation and time series model training and inference technologies, this agent uses deep diagnosis based on Altas large-scale expert parallelism (EP), greatly improving the generalization and explainability. The fault diagnosis accuracy in transformer districts has increased to 80%, reaching the level of diagnosis by experts in the system loss team. The key technologies are as follows:

1. Enhanced diagnosis and reasoning of LLMs: The system loss diagnosis methodology used by on-site personnel is extracted and summarized, and "taught" to AI. LLMs are enabled with strict logical reasoning capabilities through the use of strong and weak evidence chains, counterfactual reasoning, feature matrix verification, and confidence calculation. The agent collaboration in blackboard mode supports multi-role supervision and multi-channel voting, reducing hallucinations and improving reasoning accuracy.

2. Data feature engineering. Based on measurement data such as household meter voltage, current, power, and meter readings, high-order features are extracted through data feature engineering from multiple dimensions, such as space-time, value, group, event, and behavior. This transforms raw data into standardized inputs that LLMs can comprehend and learn from. In addition, information is filtered and optimized to help LLMs better extract key insights, transitioning diagnosis from relying on expertise to being based on data, making agent diagnosis more accurate and efficient.

3. Expert skill training model. Based on the DeepAgent framework of openJiuwen, the agent is constructed with expert thinking, diagnosis manuals, and question sets through prompts. An iterative mechanism of question answering, reflection, and improvement is established to implement low-threshold expert skill training. This converts expert experience into digital skills and helps enhance the skill of new employees.

4. Altas acceleration for batch tasks. Considering the service characteristics of batch diagnosis at night and manual review during the day, DeepSeek V4 is deployed in large-scale EP mode. In addition, multi-turn conversations, cache optimization, day/night prefill & decode (PD) adjustment, and parallel diagnosis for over 30 transformer districts are implemented. The diagnosis time for a single transformer district takes only about 20 minutes.

During the development of the intelligent diagnosis agent for transformer districts, Huawei Altas computing power fully leverages the advantages of large-scale EP networking and optimized configurations. This enables in-depth diagnosis to achieve better inference precision and performance in terms of large-parameter models, long-chain thinking, and parallel inference, laying a super-performance computing foundation for the accurate analysis and location of transformer district faults such as system loss, short circuits, and leakage.

Dual-Value Realization: Technology for Supporting Livelihoods

Since put into production, the agent has demonstrated remarkable value in two aspects.

1. Efficiency value: Reduced workload and improved efficiency. In the past, experts had to manually analyze system loss and locate short circuits and leakage, a time-consuming and labor-intensive task. Now, the agent can complete comprehensive analysis of about 700 high-loss transformer districts in a single day, boosting the diagnosis efficiency by 13 times compared with manual methods. The workload of onsite troubleshooting in transformer districts has been cut by over 80%, greatly reducing labor and time costs.

2. Management value: Risk identification for safe electricity use. The agent helps accurately and efficiently detect illegal electricity usage that affects safe electricity use, and address short-circuit and leakage hazards that pose safety risks. This helps the power company protect customer electricity safety, uphold fair electricity usage order, and safeguard the stable operation of the power grid.

Embarking on a New Chapter of Transformer District O&M with AI

In the future, Xiangtan Company will partner with Huawei to continuously deepen the application of the intelligent diagnosis agent for transformer districts, further improving the diagnosis accuracy. Acting as a "general practitioner" for fault diagnosis and root cause analysis in transformer districts, the agent will ensure the safety of distribution networks around the clock. With digital intelligence, the agent will help power grids illuminate thousands of households.