The discussion on AI in the new energy storage industry has moved beyond the "whether to do it" stage and entered the practical competition of "how to do it and where to draw the boundaries."
How does AI truly integrate into energy storage products, and where do enterprises' real capabilities lie? Expert Say Energy Storage specially organized the "Energy Storage + AI" thematic planning, conducting a sample survey of 13 representative enterprises including Envision, Huayuxin, Hoenergy, Weiheng Intelligent, NR Electric, Yuanxin Energy Storage, GCL Energy Storage, Lising Energy, Saiwei Digital Energy, Guoneng Rixin, Conenergy, Star Charge, and Xingchen New Energy. Combined with the research findings of the "2026 AI + Energy Storage Bidirectional Synergistic Development Research White Paper," we explore the threshold and specific path for building AI capabilities in energy storage products.
The following is the first part of this thematic planning, focusing on four dimensions: product form, functional penetration, path selection, and capability threshold, to comprehensively dissect the AI layout paths of the 13 representative enterprises. In the second part, we will focus on the premium space and implementation of AI in real energy storage stations.
According to the research by Expert Say Energy Storage, when new energy storage enterprises build product AI capabilities, they are not just telling the same story on the surface; instead, their paths have diverged, each with its own entry point, different product forms, and varying focuses on specific AI value functions.
▍ Native AI Accounts for 58.33%, While Built-in Device AI is About 41.67%
About 41.67% of energy storage enterprises adopt built-in device AI, specifically embedding AI capabilities directly into the equipment, allowing the device to complete data analysis and strategy output locally. Representative enterprises include Hoenergy, Huayuxin, Xingchen New Energy, and Saiwei Digital Energy. Another 58.33% of energy storage enterprises emphasize native AI, such as Envision, Weiheng Intelligent, and Guoneng Rixin, taking AI as the product gene. Among them, Huayuxin and Weiheng Intelligent actually occupy both sides—they have both a native perspective and firmly grow AI within the equipment. Lising Energy, on the other hand, places more emphasis on a full-stack energy storage AI system, integrating hardware, software, algorithms, and scenarios across the entire stack.
▍ EMS Has the Highest AI Content, at About 75%
From the perspective of product forms, the closer the link is to decision-making and software, the higher the AI penetration rate. The AI penetration rate in EMS is the highest, at about 75%—EMS undertakes data-intensive tasks such as power forecasting, dispatch optimization, and revenue calculation, naturally falling within the capability range of AI. The overall solution of energy storage system + AI ranks second, accounting for about 66.67%, reflecting that most manufacturers tend to embed AI capabilities into existing product lines. Virtual power plant (VPP) + AI ranks third with 58.33%, as its aggregated dispatch and electricity market trading scenarios are also high-value landing points for AI.
BMS + AI accounts for about 41.67%. Battery management is directly related to safety and lifespan, with stringent requirements for deterministic output. Currently, AI mostly stays at the level of auxiliary analysis. PCS + AI is only 8.33%, the lowest in the entire field. Notably, independent closed-loop AI products account for only about 25%, indicating that the industry as a whole is still in the stage of "AI empowering hardware" rather than "AI redefining products."
▍ AI Value: Charge and Discharge Strategies Become Standard
Looking at the functions realized by enterprise AI in energy storage applications, the coverage shows a gradient of "basic operation first, deep optimization later": charge and discharge strategies cover all surveyed enterprises (100%), which has become an industry standard; operation and maintenance account for about 91.67%, and the two together constitute the foundation of AI applications; electricity trading is about 75%, with revenue-side implementation taking initial shape; intelligent warning is at 83.33%, and lifespan prediction is at 58.33%. This sequence is essentially a difficulty-value trade-off for AI implementation—scenarios with quick results and high fault tolerance go first. As NR Electric said: the output of AI is inherently probabilistic; from input to output, it is almost a "black box" for users, while energy storage systems have strict requirements for response speed and reliability. Therefore, current AI is more suitable for scenarios with higher fault tolerance and no fatal consequences in the short term, such as electricity trading decisions and power forecasting. In terms of implementation paths, the principle of progressive trust should be followed—first complete verification in non-critical scenarios, and then gradually extend to core optimization links. Main Paths and Key Barriers for Building Energy Storage AI Capabilities.
▍ Three Paths Each Have Their Costs, and the "Key Technology Chain" is the True Watershed
From the perspective of the construction paths of AI capabilities in energy storage enterprises, they are mainly divided into self-developed algorithms, purchasing mature software, and ecological cooperation. These three paths have their own advantages and disadvantages, specifically:
The advantage of self-developed algorithms is end-to-end controllability, the ability to fit their own equipment, and the ability to precipitate differentiated capabilities; the cost is heavy investment, long cycles, and extreme dependence on compound talents.
The advantage of purchasing mature software is fast launch, controllable initial costs, and the ability to quickly make up for functional shortcomings; the shortcomings are also highly similar—homogenization, black box, supplier lock-in, and difficulty in deeply entering the equipment control logic.
Ecological cooperation can quickly make up for shortcomings by borrowing external capability resources such as meteorology, virtual power plants, electricity trading, computing power platforms, and general models; the risk lies in the high degree of dependence, and it is necessary to clearly define data ownership, control permissions, fault responsibilities, and exit mechanisms in advance.
According to the research by Expert Say Energy Storage, most enterprises base their path selection for AI capabilities on the following principle: "Self-develop core decision chains/control chains/safety boundaries, and cooperate for general foundations and external scarce resources." Whether it is "self-developed or outsourced," what enterprises care about most is how long the "key technology chain" that must be held in their own hands is. The position of this line determines the divergence of each company's route.
▍ Thresholds for AI in Energy Storage: The Difficulty Lies Not in Algorithms
Expert Say Energy Storage has learned through in-depth market research that the true threshold for AI in current energy storage system products is not the size of the model, but turning "algorithms into a set of real capabilities that can run stably in energy storage stations." For enterprises, the most difficult part of building AI capabilities is: crossing from "prediction," "read-only," and "suggestions" to "executable, closed-loop, and accountable actions." Single-point algorithms can be bought, but the ability to integrate hardware, software, algorithms, and scenarios across the full stack cannot be bought.
According to Huayuxin, energy storage enterprises need to cross at least six major thresholds to build AI capabilities:
1. Data Foundation: "Dirty data is more fatal than weak algorithms." It requires long-term, continuous, high-frequency, and governed multi-dimensional data. The reflection of real working conditions in the data is more important than the data scale.
2. Mechanism and Safety Boundaries: AI outputs must meet rigid constraints such as SOC, power, efficiency, temperature, lifespan, grid specifications, grid-connected/off-grid logic, and equipment warranty. Models detached from mechanisms and equipment boundaries, no matter how high their offline indicators, cannot enter the production environment.
3. Leap from Prediction to Decision: Generation/load/electricity price forecasting is only the input. Enterprises also need to turn it into executable charge and discharge strategies, and balance revenue, safety, lifespan, backup, and user goals. A large amount of industry capability stops at "more accurate prediction" and has not solved "how to make decisions and execute after prediction."
4. Cloud-Edge-Device Engineering: Cloud, edge, and device must have a reasonable division of labor: the cloud handles training and global optimization, the edge handles local prediction and offline operation, and the device handles fast execution and underlying protection. If the division of labor is unclear, AI will drag down the system due to network fluctuations or service anomalies.
5. Cross-scenario Generalization: Running through in a pilot station ≠ being replicable to different countries, electricity price mechanisms, equipment types, and loads; it must automatically identify scenario differences and adapt.
6. Long-term Operation and Iteration: Energy storage AI needs continuous correction through long-term operation. If there is a lack of real power stations, market trading, and continuous operation experience, algorithms can easily stay at the laboratory model or demonstration level.
▍ Representative Enterprises
Huayuxin
Huayuxin has made "self-developed key links, open cooperation for general capabilities, and no outsourcing of core data and equipment execution rights" its principles. Centered on the HYXI Muse AI Native Energy Autonomy System, and focusing on "revenue traction, easy implementation, and safety bottom line," it has built three professional Agents: Revenue, Operation, and Guardian, responsible for generation and consumption forecasting and dynamic dispatch, simplifying user operation processes, and full-link anomaly identification and automatic handling, respectively.
Huayuxin is continuously evolving towards "full-domain self-development," ensuring that the key technology chains affecting energy storage safety and revenue are autonomous and controllable, including: generation and consumption forecasting and data processing; charge and discharge dispatch and revenue optimization; SOC, power, lifespan, and safety constraints; local strategy generation for EMS; equipment command issuance and execution verification; operation effect evaluation and model iteration; AI assistant, knowledge base, and strategy interpretation capabilities.
In Huayuxin's view, the core competition in the energy storage industry in the future is not "whether the most advanced algorithms are called," but whether the enterprise has mastered the complete responsibility chain from data to equipment actions. The most difficult part of energy storage AI is not training a certain model, but establishing a long-term closed loop of perception—prediction—decision—execution—verification—re-learning. Its core barrier is a composite barrier formed by equipment data, energy mechanisms, algorithm capabilities, control systems, operation feedback, and project experience.
Envision
From the very beginning of doing energy storage, Envision proposed that "the essence of energy storage is trading." Based on this, it chose the route of full-stack self-development of software and hardware, independently controlling everything from battery cells to PCS, BMS, EMS, and SCADA, in order to thoroughly understand the system performance boundaries and let AI and equipment collaborate to the extreme. Envision has been deeply engaged in new energy for nearly 20 years. In its early days, it developed the EnOS intelligent IoT operating system, which currently has access to over 1005GW of new energy equipment data. On this basis, it has created the "Tianji" meteorological large model and the "Tianshu" energy large model.
Its data barriers are mainly divided into three layers: equipment-level real operation data, system-level collaborative data, and the experience layer where engineers combine physical mechanisms with data results for labeling. The most difficult part of building AI is not the algorithm, but the deep accumulation of data and systems. Even harder is "the unity of knowledge and action"—whether the trading decisions of the large model can be accurately and timely executed by the physical system. Envision's ultimate goal is to build an Energy AI Agent, taking all constraints as the Harness, and each equipment and load as a Skill, uniformly scheduled and integrated by the large model.
Hoenergy
Hoenergy focuses on the digital and intelligent upgrade of energy storage. With AI algorithms, big data, and cloud computing as the core technology foundation, it deeply cultivates the integrated innovation of "Energy Storage + AI," achieving breakthrough results in multiple dimensions such as intelligent system R&D, operation strategy optimization, safety protection upgrade, and market scenario implementation, building an integrated intelligent energy storage system of technology, products, applications, and ecology.
To build good AI capabilities, energy storage enterprises need to cross three hurdles: data, engineering, and safety. Hoenergy's BMS 2.0 system, relying on AI multi-physics field coupling modeling and high-dimensional data feature extraction technology, breaks through the limitations of passive monitoring in traditional battery management systems, building a multi-level intelligent protection system of cell-level anomaly warning and system-level situation awareness. The system can dynamically monitor the battery operation status in real time, accurately diagnose potential battery faults and thermal runaway risks through AI algorithms, and form a closed-loop safety control mode of "dynamic monitoring—intelligent diagnosis—active intervention."
Weiheng Intelligent
To build good AI capabilities, the key for energy storage enterprises is not how large the model is, but whether it can solve on-site problems. First, the data must be continuous, real, and usable; second, it is necessary to truly understand the battery, equipment, and electricity price rules, and no strategy can touch the boundaries of safety, grid connection, and warranty; finally, the software system must be stable, able to degrade and roll back when encountering network disconnection, data anomalies, or equipment status changes.
For Weiheng Intelligent, we are close to EMS and terminal scenarios, allowing algorithms, software, hardware, and operation and maintenance teams to iterate together, continuously precipitating on-site experience into replicable product capabilities. Weiheng Intelligent has formed a full-stack energy AI capability matrix covering the entire life cycle of energy storage, covering data intelligence, predictive decision-making, equipment control, asset health, and operation and maintenance. Commercial and industrial, as well as residential, adopt independent business services and share the capability core; the cloud is responsible for prediction, diagnosis, reporting, and cross-station analysis, while the edge is responsible for real-time data, alarms, execution auditing, and control assistance after approval. Corresponding algorithms are also specially developed for operation and maintenance and batteries, and EMS integration and gradual agent-ization are realized.
NR Electric
The core feature of AI technology is probabilistic output, giving the "most probable result," not the "deterministic result." However, the high reliability requirement of energy storage system control is deterministic. Overload, over-temperature, thermal runaway, etc., must make correct actions at the millisecond level, and any "probably correct" speculation is unacceptable. At the same time, the reasoning process of AI is a black box for users. After probabilistic output, it cannot provide users with clear correction experience, and even AI's own correction cannot provide users with a clear logic of deterministic improvement.
Therefore, AI is more suitable for fields with high fault tolerance and no disastrous consequences in the short term, such as trading decisions, power forecasting, etc. For early warning of equipment operation and maintenance, improving warning efficiency still relies on human or non-AI decision logic; its application should follow the path of progressively increasing trust, verifying the reliability of AI in non-critical scenarios, accumulating data and user confidence, and then gradually expanding to more core optimization scenarios. It will still take a long process to reach the stage of real-time safety control. At this stage, claiming to use AI technology for real-time control in products is either a gimmick or a result of sacrificing product reliability and performance.
Yuanxin Energy Storage
Currently, the integration of energy storage and AI mainly has three paths: self-developed core algorithms, purchasing mature tools, and ecological cooperation. For Yuanxin Energy Storage, the three paths are layered and combined around core capabilities: core capabilities that determine asset revenue—such as electricity price and supply-demand forecasting, trading strategy optimization, volume and price reporting suggestions, strategy review, and revenue attribution—require mastering core algorithms, rule modeling, and continuous iteration capabilities; general large models, basic computing power, and development tools are reasonably purchased to improve R&D efficiency; meteorological information, market rules, etc., are supplemented through ecological cooperation. The core principle is: key capabilities are autonomous and controllable, the general foundation does not reinvent the wheel, and external cooperation cannot cut off the closed loop of data, strategy, execution, and review.
Based on this architecture, Yuanxin has built the EnergyNexusAI intelligent agent middle platform, as the "decision-making brain" of the software ecology, uniformly bearing capabilities such as electricity price prediction, model governance, trading strategy optimization, strategy center, revenue review, and Yuanxin Zhiling, forming a business closed loop of "prediction—strategy—execution—review—optimization." Among them, Yuanxin Zhiling adopts a multi-agent orchestration architecture based on LangGraph, connecting real databases, knowledge bases, prediction models, and EMS strategy services through controlled tool calls, ensuring that key decision results are traceable, verifiable, and interpretable.
GCL Energy Storage
The most difficult step is not to make AI talk, but to make AI dare to act: crossing from "read-only Q&A" to "executable actions" requires simultaneously solving unauthorized access, misoperation, idempotence, rollback, and responsibility attribution. If this step cannot be crossed, AI will always just be a chat box on a dashboard.
Another underestimated sticking point is the modernization level of the platform architecture. Old technology stacks cannot be connected to the contemporary AI ecology, and no matter how good the capabilities are, there is nowhere to land them. GCL Energy Storage launched the DuLink intelligent energy platform, a full-stack self-developed commercial and industrial energy storage cloud platform, from equipment access, time-series data foundation to the AI capability layer, all belonging to the same architecture. It has three characteristics: AI and the platform have equal rights and the same source, sharing authentication, tenant isolation, and audit channels, with no dedicated backdoors; reasoning and knowledge bases support private self-hosting, and data does not leave the domain; capabilities are opened according to risk grading—from read-only Q&A to diagnostic suggestions to controlled automatic dispatch, each level is equipped with auditing and rollback.
Lising Energy
Truly entering the operation of energy assets requires crossing three thresholds: data foundation, scenario closed loop, and full-stack collaboration. Among them, the most core and hardest to replicate barrier is full-stack collaboration. Energy storage AI spans multiple disciplines and fields; if any link is disconnected, it is ultimately difficult to form an effective chain. Lising Energy connects the new energy full-ecosystem service closed loop centered on "R&D, intelligent manufacturing, investment, construction, and operation." Its YuanTelu® energy storage products have been widely put into operation in projects in many countries in Europe, Asia, and Africa. The self-developed Xingneng Cloud Platform covers equipment monitoring, multi-energy collaboration, electricity trading, and VPP. The company further connects the energy side and the computing power side through "Xinghe," the world's first distributed computing-electricity collaborative solution.
For Lising Energy, AI is a capability built on its own equipment, real assets, and continuous operation data. The true barrier of energy storage AI is not a certain algorithm, but whether it can integrate hardware, software, algorithms, and scenarios across the full stack.
Saiwei Digital Energy
Saiwei Digital Energy judges the path of AI capability construction based on two main criteria: whether it is good at it itself, and whether it constitutes the future core competitiveness of commercial and industrial energy storage. If both are met, it will invest funds in self-development and layout. Based on this logic, Saiwei adopts a dual-wheel drive of "core self-development + ecological cooperation"—in terms of self-development, relying on Jack Group's 30 years of electromechanical integration technology accumulation and more than 10 years of industrial Internet experience, it has developed the intelligent charge and discharge strategy of the "Blue Whale Control Energy" system, and the self-developed AI-EMS has achieved an increase in power station revenue of more than 5%. External data capabilities such as meteorology and electricity market interfaces are made up through ecological cooperation. Saiwei's edge-device smartEMS700 intelligent controller, combined with cloud-edge collaborative AI algorithms, realizes intelligent optimization of charge and discharge strategies, with a maximum system efficiency of 92.9%.
Energy storage enterprises need to cross two thresholds to build AI capabilities: first, data quality and credibility are the cornerstone, providing a solid foundation for later AI decision-making; second, safety and operation boundaries—to make the system smarter and reduce manual intervention, it must be built on a reliable safety operation boundary.
Xingchen New Energy
The focus of building AI capabilities does not lie in whether all modules are independently developed by the enterprise, but in whether it has mastered the power station data, decision logic, and the complete closed loop from strategy to execution. The most difficult part is not a single algorithm, but organizing equipment cognition, power systems, market trading, and asset operation capabilities into a sustainably iterable system.
Based on this, Xingchen New Energy adopts the path of "independent construction of core capabilities and ecological synergy of professional capabilities," building three major platforms: the power station intelligent operation and maintenance platform continuously collects equipment status and operation efficiency data; the AI intelligent storage trading cloud platform combines electricity prices, loads, meteorology, and market clearing information to complete strategy generation and trading declaration; the power asset full life cycle management platform coordinates project construction, operation revenue, and long-term value evaluation from an asset perspective. The three are not independent tools, but form a business chain of "status perception—predictive analysis—strategy generation—operation and trading execution—result feedback—strategy iteration." Xingchen New Energy's understanding of "Energy Storage + AI" is not to add an algorithm module to the energy storage system, but to let AI truly enter the operation decision-making and revenue realization process of the power station, promoting the energy storage power station to gradually evolve from passive execution of dispatch instructions to autonomous perception, dynamic decision-making, and continuous optimization.
Star Charge
The core barrier and most difficult point for energy storage enterprises to build AI capabilities is not a single algorithm or data interface, but the deep closed loop of "massive self-owned real operation data + full-stack self-developed algorithms + practical experience in the electricity market + cloud-edge-device collaborative execution." Only by connecting the equipment layer, operation layer, and trading layer can the prediction accuracy be truly transformed into settleable revenue and replicable scaled capabilities. Star Charge is continuously crossing these thresholds and building differentiated competitiveness through full-stack self-development and the "cloud-pipe-device" collaborative model.
Star Charge takes the Energy Cloud as the core and adopts the "cloud-pipe-device" collaborative model to promote the integration of the charging network, microgrid, and virtual power plant. Core AI capabilities are reflected in three major platforms: the charging cloud platform realizes one-station-one-policy site selection, adaptive price adjustment, and operation efficiency improvement through AI; the Taiyi trading platform uses AI-Agents to realize strategy generation, trend interpretation, and volatility warning, calling top prediction models to optimize trading strategies; the Star Charge virtual power plant platform realizes a price prediction accuracy of over 97.8%, adjusts 1785MW of power in the fastest 1 second, and breaks through 4700MW in adjustable load aggregation. At the same time, it self-develops EMS and the Xingneng SaaS platform, covering real-time monitoring, demand management, and full life cycle operation.
Conenergy
For energy storage BMS enterprises, the core barrier to building AI capabilities lies in the deep understanding of battery physical/chemical characteristics and the coupling of massive real operation data—AI algorithms are just tools, and the model that can accurately describe the "birth, aging, illness, and death" laws of the battery is the soul. The most difficult point is the leap from "data" to "knowledge": how to transform the raw signals of voltage and temperature collected by the BMS into operable operation and maintenance instructions through AI, and be trusted by customers. This is not only a technical issue but also a trust mechanism issue.
Relying on its self-developed BIDS (Battery Intelligent Diagnostic System), Conenergy has transformed energy storage asset management from experience-based "guessing" to data-based "calculating," successfully helping energy storage projects get out of the operation quagmire.
Guoneng Rixin
Energy storage AI needs to cross three progressively deepening thresholds: the accumulation cycle of vertical data is long and cannot be made up by assault; scenario cognition tests the deep understanding of the game logic of the electricity market, which is not a pure algorithm problem; and the most difficult is the engineering closed loop—running through "prediction → strategy → risk control → execution → review" in a real production environment, involving system stability, grid safety compliance, and trading timeliness, requiring verification of at least two to three complete trading cycles. The core decision chain must be self-developed, and the edge side can introduce ecology.
With the Kuangming large model as the foundation, Guoneng Rixin has built an independent energy storage full-stack operation system covering four layers: prediction, dispatch, trading, and operation and maintenance. The prediction layer provides the data foundation for power and electricity price prediction; the dispatch layer completes SOC optimization and safety boundary guard; the trading layer is driven by multi-Agent collaboration to automate the closed loop of multi-market combination strategies, upgrading the revenue from a single peak-valley arbitrage to a diversified structure; the operation and maintenance layer realizes real-time control of 3000+ nodes and revenue attribution analysis. The whole system has been deployed and verified in multiple provinces, with accuracy and revenue leading the industry.
Expert Say Energy Storage Summary:
The Essence of the Energy Storage AI Competition is a Contest of Control Power and Implementation Power
Combining the research situation of the above-mentioned multiple enterprises, the conclusion is actually very clear: the AI competition in the energy storage industry is essentially not an arms race of algorithm parameters, but a contest of "control power" and "implementation power." In terms of paths, enterprises unanimously hold the core chains involving safety and revenue in their own hands; in terms of thresholds, they are stuck at the last pass of "letting AI truly act and be reliable." And the true watershed lies in: whoever first runs the loop of "prediction—decision—execution" into a closed loop in real energy storage stations and generates settleable revenue, is the one who truly crosses the threshold of energy storage AI—otherwise, no matter how advanced the algorithm is, it is just staying in the PPT.
[Acknowledgments: Envision, Huayuxin, Hoenergy, Weiheng Intelligent, NR Electric, Yuanxin Energy Storage, GCL Energy Storage, Lising Energy, Saiwei Digital Energy, Guoneng Rixin, Conenergy, Star Charge, and Xingchen New Energy for their contributions to this article.]
Note: The content of this article is only for sorting out public information in the energy storage industry, industrial research, and objective analysis. The listed companies mentioned in the text are only for reference as industry chain cases and do not constitute any investment basis. The market has risks, and investment should be cautious.