Original Title: AI Authentication Checklist from Over 10 Energy Storage Manufacturers
The biggest consensus in the energy storage circle over the past two years is that almost everyone is working on AI. AI prediction, AI dispatch, AI O&M, AI trading—the stories told in product manuals are all related to AI. As for the biggest divergence, it is about what has actually been achieved: Is the data secure? Is the software strong? Who has real skills, and who is just attaching labels?
The previous article "75% and 8.33%: Energy Storage AI Penetration Map" by Hangjiashuo Energy Storage reviewed the foundation of AI adoption in the industry: which segments are using it and to what extent.
In this issue, we change the questioning approach. Based on in-depth investigations into representative enterprises such as Envision, Zhejiang Hyxi Technology Co., Ltd. (referred to as “HYXI”), Hoenergy, Weiheng Intelligent, RelyEZ Energy, GCL Energy Storage, Lisiner, SAV Digital Power, Sprixin (301162), Star Charge, and Xingchen Xinneng, combined with the research from the "Hangjiashuo Energy Storage: 2026 AI + Energy Storage Two-way Synergistic Development Research White Paper", we provide the industry with an AI authentication checklist for reference.
This article will focus on the following points:
1. How to distinguish real from fake AI?
2. What indicators to look at for software capabilities?
3. How to ensure data quality and security?
4. Key insights from representative enterprises are attached at the end.
How to Distinguish Real from Fake AI?
When all enterprises are shouting "AI empowerment", the most basic question from customers is whether this is a gimmick or real. According to research by Hangjiashuo Energy Storage, most energy storage enterprises frankly state that the divide between real and fake AI does not lie in what model is used, but in reducing AI to three "can-its"—can it speak with evidence, can it enter the business closed-loop, and can it withstand comparison and accountability.
- At the evidence level, everyone repeatedly emphasizes that real AI dares to report real project data, display backtesting reports, present settlement bills, and provide third-party empirical evidence; at the closed-loop level, AI must penetrate prediction, strategy, execution, monitoring, and review, preferably directly connecting to trading declarations and equipment control, rather than just giving suggestions and producing reports; at the boundary and accountability level, one must ask how abnormal scenarios are degraded, how errors are rolled back, and whether these are written into the contract. GCL Energy Storage put it most straightforwardly: "What can produce reports is information; what can be held accountable is capability."
In terms of specific methods, various enterprises have provided different authentication checklists. HYXI proposed "six follow-up questions", comprehensively inquiring from real projects, revenue baselines, control links to abnormal operations and settlement reviews; Envision listed "three dimensions": Is there a three-layer data accumulation at the equipment level + system level + label level in real scenarios? Can it achieve "unity of knowledge and action"? Is there verification from large-scale implemented projects?
Weiheng Intelligent broke down the customers' AI validation standards into five aspects: data boundaries, comparison baselines, closed-loop depth, failure boundaries, and continuous operation evidence; Sprixin is more direct, asking customers to pull historical strategy backtesting reports, look at established algorithm teams, and find stations under continuous trusteeship for over a year for anonymous comparison; Star Charge defined real AI with four phrases: "calculate accurately, dispatch effectively, earn profits, and review clearly".
Their common subtext is: whether AI actually works will be revealed by data; those who are afraid of testing and only dare to talk about concepts are most likely just packaging.
What Indicators to Look at for Software Capabilities?
In the energy storage industry, hardware competition is basically transparent. How many cycles the battery cells last, what the system efficiency is, and what the energy density is are all quantified through actual testing, making the differences clear at a glance. Software capabilities are different. While both promote "AI empowerment", some truly help power stations earn an extra 8% in the spot market, while others just put a large model skin on traditional EMS. To compare software capabilities, a different set of rulers must be used.
As for what specific rulers to use, Hangjiashuo Energy Storage, combined with research and analysis of energy storage enterprises, identifies four main points:
Prediction accuracy is just a ticket to enter, not the endpoint. Merely reporting a general "90% accuracy" is meaningless; performance during peak errors, extreme weather, holidays, and equipment failures is where the gap is widened.
Revenue must be reflected in the settlement bills. "How much can theoretically be earned" does not count. With the same equipment, same electricity price, and same cycle, the additional settleable net revenue compared to the original strategy is the value created by the software. Whether AI can "take action" is the watershed. Those that only produce reports and give suggestions are auxiliary tools; those that can automatically issue commands, roll back, and hold individuals accountable are control systems. Revenue cannot be exchanged for the battery's life. SOC, power, lifespan, and warranty boundaries are hard constraints. Strategies that exchange short-term revenue for excessive cycles will lose more the more they earn.
How to Ensure Data Quality and Security?
As a relevant person in charge at Envision pointed out, from the perspective of system dispatch, control, and O&M, the manifestation of boundaries is data. Currently, many people in the energy industry do not understand AI training, thinking that as long as a model is built, equipment is available, and integration is done, a solution can be created. In reality, this is not the case.
What the model outputs is not just a piece of text, but commands directly connecting energy storage equipment and the grid. Hallucinations are bugs elsewhere, but in energy storage, they could be accidents: overcharging, tripping, and revenue distortion. Therefore, manufacturers who truly step in to empower energy storage products with AI have no less work in quality control and compliance than in modeling. Data is the ingredient, governance is the kitchen, and AI is the chef—if any link fails, what is served on the table is not a dish, but a mine. According to research by Hangjiashuo Energy Storage, the consensus among energy storage manufacturers on data mainly includes the following points:
Dirty Data Does Not Enter the Model
Unify the caliber of time, units, and entity IDs, perform cleaning and completion, anomaly verification, and missing data handling; reject release if it does not pass. The closer the checkpoint is to the front, the better: clean at the equipment end and intercept at the access side. Before going online, historical data must be used for backtesting and reproduction: "If it doesn't match, it is not allowed to go online."
Use Mechanisms to Control Large Model Hallucinations
Key values such as electricity prices and revenue must come from real databases, prediction models, or strategy services; "large models are not allowed to generate them directly." The output of large models must be verified by mechanism models, knowledge bases, and expert systems to prevent "hallucinations" from mixing into dispatch and settlement.
Data Does Not Leave the Domain
Localized deployment and edge closed-loop are standard actions: "data does not leave the station, algorithms run locally, strategies are generated locally, and control is closed-loop locally." It continues to run even when disconnected from the network, and algorithm anomalies automatically switch back to safety rules. Sensitive scenarios support private deployment, and external cooperation uses federated learning to achieve "data is usable but invisible."
AI Must Not Exceed Authority
AI is not a bypass; intelligent calls go through the same authentication and isolation channels as humans. Strategies generated by the model must pass SOC, power, temperature, lifespan, and safety rule verification before entering the execution link; the underlying equipment protection priority is always higher; action-type operations default to manual confirmation, with limits set, rollback capability, and manual takeover.
Every Step Can Be Held Accountable
Least privilege, hierarchical authorization, tenant isolation, and transmission and storage encryption. Every call and every command can be traced to the person, time, and basis, with full-link audit trails—traceable, verifiable, and explainable.
Views of Representative Enterprises
▍HYXI: Intelligence Cannot Be Exchanged for Safety
HYXI believes that the quality of energy storage software should ultimately be verified through three tables: the strategy execution table, the project settlement table, and the equipment health table. The first table proves that the algorithm is truly executed, the second proves that the algorithm creates revenue, and the third proves that the revenue is not obtained at the expense of safety and lifespan. HYXI's basic principle is: intelligence cannot be at the expense of data sovereignty and system safety. Truly mature AI should ensure the safe operation of the energy storage system even without the cloud or when the model is abnormal. They break down data quality and safety into four tasks: how to manage data, how to deploy the system, who can view it, and who can modify it. In terms of actions: equipment time, sampling frequency, unit dimensions, and communication status are first managed uniformly; abnormal data is filtered out, and critical data is cross-verified; the algorithm supports fully localized deployment based on EMS, data does not leave the station, and it can still run when disconnected from the network; collection follows the "minimum necessary" principle, with hierarchical permissions and full-process logging to protect sensitive data; the model does not directly exceed authority to control energy storage equipment, and in case of anomalies, it must allow human takeover, strategy rollback, and degradation—the whole process is auditable.
▍Envision: Treating Self-Investment as a Data Reserve
When Envision talks about data, its confidence comes from "stepping into the arena personally". "If a project is not invested in and executed by the enterprise itself, the first-hand data the enterprise can obtain is limited." Therefore, it invested in Chifeng Green Hydrogen and Ammonia, Baotou Green Alloy, and took on high-energy-consuming scenarios like Nike and Starbucks. The wind and solar prediction data accumulated from over a decade of wind turbine business and the real operation data from its own stations—"these data that general customers do not disclose externally are the moat". At the same time, Envision is trying to move data away from the stage of manual labeling; once successful, it may leave the industry half a step behind. For software capabilities and the authenticity of AI, Envision's answer points to the same term: Physical AI, emphasizing that AI must be deeply coupled with the physical world, rather than relying solely on large language models. It treats energy storage as "embodied intelligence" for R&D; from product design to market trading, and then to long-term O&M, AI is the underlying logic rather than a tool. The hard indicators are right there: Envision's AI energy storage system improves the full lifecycle IRR by 4-8%, the electricity price prediction accuracy at key trading nodes exceeds 90%, the "Tianji" meteorological large model pushes the wind and solar power generation curve prediction to over 95%, the EN 12.5MWh system energy efficiency exceeds 92%, and grid connection efficiency improves by over 50%. The most convincing battle is in Inner Mongolia: a 12.8GWh energy storage cluster connected to the electricity spot market passed the grid's "three charges and three discharges" verification at once—in Envision's own words, "this is the best test of the authenticity of AI capabilities."
▍Hoenergy: Turning "Performance" into "Measurable Items"
Hoenergy breaks down data issues into four characters: "source, edge, pipe, and use". At the source, edge computing power is used at the equipment end to eliminate burrs and fill in missing data; a smarter move is to use "mechanism models" to control "data models", directly intercepting outputs that contradict electrochemical laws to prevent AI hallucinations. In transmission, high-grade encryption chips plus private communication protocols prevent tampering throughout the process. In use, it emphasizes "data is usable but invisible"—when cooperating externally to optimize algorithms, federated learning is used to only exchange model gradient parameters, and not a single piece of raw data leaves the premises. How to compare software? Hoenergy provided five practical hard indicators: response time, strategy revenue, SOC/SOH estimation accuracy, strategy robustness, and OTA success rate. Applied to products, Hoenergy Energy Storage's COSMOS 2.0 digital system integrates an AI dynamic strategy engine, multi-dimensional data prediction, high-speed communication, and digital visual monitoring functions. It can integrate load data, electricity price fluctuations, equipment health status, etc., in real time, and generate optimal charging and discharging strategies that balance safety and economy through AI algorithms. It supports multiple revenue models such as peak-valley arbitrage and demand management, and has ultra-fast response capabilities with millisecond-level control and second-level data collection, accurately matching grid dispatch needs. Going further up, there are the D-Galaxy cloud platform, virtual power plant management platform, and power trading auxiliary decision-making system, forming the entire chain of "intelligent prediction - dynamic decision - precise dispatch - trustworthy trading".
▍RelyEZ Energy: Preventing Large Models from Generating Key Data
RelyEZ Energy Storage puts data authenticity first, not allowing large models to directly generate key data such as electricity prices and revenue; every number from the AI must come from real databases, prediction models, or strategy services. Data quality emphasizes unified caliber, time alignment, anomaly verification, and lineage tracking; predictions must be compared with actual values, and strategies must be reviewed against settlement revenue, with errors quantified in the open. The agent architecture adopts "mandatory toolchain invocation + result verification", retaining the source, parameters, and invocation records for key results to ensure traceability, verifiability, and explainability. However, RelyEZ Energy believes that data authenticity is just a ticket to enter, not a moat. What truly determines the value of energy storage AI is whether market understanding, algorithm solving, asset constraints, trading operations, and settlement reviews can be unified into long-term net revenue. Based on RelyEZ OS, RelyEZ Energy has built the EnergyNexusAI / RelyEZ Energy AI Agent Middle Platform, which uniformly undertakes electricity price and supply-demand prediction, trading strategy optimization, revenue review, and Yuanxin Zhiling, running through the closed-loop of "prediction - strategy - execution - review - optimization", and has formed energy storage industry Agent capabilities in regions such as Yunnan and West Inner Mongolia. The core view of RelyEZ Energy is: prediction accuracy is only an intermediate indicator; ultimately, software value should be tested by risk-adjusted long-term net revenue. Specifically, five aspects: look at revenue, look at risk, look at asset constraints, look at closed-loop, and look at generalization and engineering capabilities. In the energy storage AI battle, the winner is not the enterprise with the most labels, but the one that runs through the decision-making closed-loop and continuously delivers long-term net revenue.
▍Weiheng Intelligent: Restoring AI from a Functional Noun to a Reproducible Business Closed-Loop
Weiheng Intelligent believes that the core of distinguishing real from fake AI lies in whether the supplier can restore AI from a "functional noun" to a reproducible business closed-loop and be willing to bear acceptable results. If an "AI capability" only has a chat entrance, promotional videos, and theoretical optimal revenue, without baselines, on-site continuous data, abnormal scenarios, degradation plans, and third-party reviewable records, customers should remain cautious. In terms of software capabilities, Weiheng Intelligent's attitude is not to pile up indicators but to look at scenarios: for prediction scenarios, look at MAE and WAPE, and evaluate them by region, season, and prediction time interval respectively; for strategy scenarios, look at the actual revenue improvement rate relative to manual scheduling or fixed strategies, confirming that the charging and discharging strategies do not break through SOC, power, grid connection, and warranty constraints. For actual implementation, look at online rate, command success rate, continuous operation time, etc.; for fault analysis and O&M, look at fault diagnosis accuracy, false and missed alarms, early warning lead time, and fault localization time. Compare under the same station, same cycle, same electricity price, and same equipment constraints, and finally accept with on-site operation records, customer bills, and market settlement results. In data governance, unify equipment, station, and time calibers, using entity IDs and UTC time as primary keys; before data enters the model, it must pass checks for completeness, physical consistency, and outliers, and reject release if it does not pass. Data centers in China, Australia, Europe, and North America are isolated from each other, with local storage and encrypted transmission, and raw data is not mixed across regions. It is also one of the few that makes algorithm compliance explicit—the "WHES AI Intelligent Customer Service Generative Synthetic Algorithm" has completed algorithm safety self-assessment and is only used for information query, knowledge Q&A, and auxiliary analysis.
▍GCL Energy Storage: What Can Be Held Accountable Is Capability
GCL Energy Storage summarizes data quality into "four checkpoints": intercepting abnormal messages at the access side, unifying the caliber and filling in missing points in the time-series database, performing three-level reconciliation of "raw value/statistical value/settlement value" on the business side, and backtesting with historical data before going online—"if it doesn't match, it is not allowed to go online", spoken resolutely. For security, there are "three iron rules": AI is not a bypass, all intelligent calls go through exactly the same authentication channels as humans, "AI cannot obtain any data beyond the current user's authorization scope"; full-link audit, every command can be traced to the person, time, and basis; data can stay within the domain, and inference runtime and knowledge bases support self-hosted private deployment. For software evaluation, GCL Energy Storage provided seven indicators, including data availability rate, strategy empirical nature, control closed-loop depth, observability and traceability, openness, iteration pace, and generalization cost. For distinguishing real from fake AI, GCL Energy Storage suggests not looking at PPTs but looking at the site, such as taking real data from the power station to ask an unprogrammed question; checking if the numbers match the accounts, conducting backtesting reports and actual trading comparisons; whether it can take action, and if so, how to roll back errors, who is responsible, and whether it is written into the contract. Additionally, one can switch to a low-privilege account and ask again to see if it leaks data it shouldn't; take this AI to a project with another electricity price mechanism and see how long it takes to run. In general, what can produce reports is information; what can be held accountable is capability.
▍Xingchen Xinneng: Letting AI Enter the Real Business Closed-Loop of Energy Storage Operations
Starting from key data such as equipment status, operation efficiency, available capacity, fault information, market prices, and trading results, Xingchen Xinneng has established a management process covering data collection, caliber unification, verification and cleaning, and business applications, improving the consistency and availability of data among different equipment and systems. On this basis, AI integrates market predictions, power station status, dispatch requirements, and operation constraints to deduce different schemes and generate strategies; strategies involving trading declarations and equipment operation also need to undergo simulation verification, risk verification, and necessary manual review. In the view of Xingchen Xinneng, judging the value of AI and software cannot only look at interfaces, the number of functions, or technical parameters, but also whether it truly improves business indicators such as prediction accuracy, equipment availability rate, strategy execution efficiency, O&M costs, and trading revenue. AI should not stay at the auxiliary level of displaying data, generating reports, and providing suggestions, but enter the complete business closed-loop of prediction, decision-making, verification, declaration, execution, and review updates, connecting power station operation management, power trading, O&M execution, and full lifecycle management, continuously improving the operational efficiency and long-term value of energy storage assets.
▍SAV Digital Power: Evaluating with a Quantifiable KPI System
Hardware can be quantified through actual testing; software cannot be "weighed by the catty", but can be evaluated using a quantifiable KPI system: electricity price/load prediction accuracy, system comprehensive efficiency, capacity release improvement, system lifespan improvement, O&M cost reduction, and system reliability/online rate. The evaluation mantra: do not look at DEMOs but empirical evidence, do not look at functions but indicators, do not look at promises but data. They also cited the judgment of a technical person in charge of a European energy storage investment institution: when selecting a supplier, first look at the grid-forming capability simulation report and revenue simulation results, with the quotation ranked last. In SAV Digital Power's words, this sentence accurately summarizes the qualitative change in customer selection logic in 2026. SAV Digital Power's data guarantee is a standard layered approach: at the collection layer, edge nodes perform real-time preprocessing, and alarm data is parsed bit by bit; at the transmission layer, ciphertext transmission has also won the European EN18031 network security certification; at the governance layer, metadata standards are established for key points such as SOC, power, temperature, and loss; the application layer is the most critical—"AI model output is double-verified by knowledge base + expert system to suppress 'large model hallucinations'". After these four layers, dirty data cannot get in, and nonsense cannot get out.
▍Lisiner: Energy Storage Evolves from Equipment-type Assets to Operating Assets
For Lisiner, the main focus is on two issues: whether data can truly reflect the status of energy assets, and whether these data can be used safely and compliantly. Lisiner emphasizes that real data can be used to steadily improve model capabilities, but never at the expense of user safety and privacy. In terms of data quality, Lisiner's trump card is the cumulative shipment of nearly 10GWh and the operating power station scale of 3.5GWh—data such as equipment status, station operation, electricity prices, and trading are all derived from real scenarios, and these data can be continuously verified by actual operation results. Lisiner believes that good software essentially needs to complete two leaps: from "prediction" to "decision-making", and then from "decision-making" to "trustworthy execution". Currently, energy storage is evolving from equipment assets to operating assets. Under the same installed capacity conditions, the gap between earning more or less is widened by this set of software capabilities of prediction-decision-execution.
First, look at accuracy—from minute-level to trend-based advance identification of problems; on the O&M side, it is whether abnormal identification and fault localization are accurate enough; on the operation side, it is whether predictions of electricity prices, loads, battery status, etc., can support subsequent strategy optimization. If the accuracy is not enough, the rest is a castle in the air;
Second, look at response speed—whether the digital twin can quickly locate faults and match solutions when equipment is abnormal; whether strategies can be quickly recalculated and issued when the market changes. Response speed determines whether software can move from auxiliary analysis into asset operation;
Third, the hardest one, look at the gain. Lisiner's intelligent strategy integrates electricity prices, power, demand, and battery degradation, using the MILP model and HiGHS solver to calculate dispatchable charging and discharging strategies. Compared with fixed or manual strategies, configuration efficiency increases by 80%, and revenue increases by 8%-15%.
▍Star Charge: Calculate Accurately, Dispatch Effectively, Earn Profits, and Review Clearly
Star Charge's confidence lies in its scale and certifications. On the security side, it has obtained MLPS Level 3, ISO 27001, ISO 27701, TISAX, SOC 2, and IEC 62443 all the way, with a CyberVadis third-party assessment score of 935/1000—the industry average is only 654; cross-border data goes through compliant gateways, stored locally or only flowing to regions with high protection levels, with private clouds additionally configured for customers with high requirements. The foundation on the quality side is the trillion-level real operation data accumulated from over 80,000 operating stations, plus full-process MES traceability in manufacturing, with a price prediction accuracy of over 97.8%. To distinguish the good from the bad in energy storage software capabilities, the key is to see whether it can truly be transformed into quantifiable operational value and system reliability, rather than a feature list. Specifically, this includes looking at revenue, response, reliability, etc. The company's platform has a price prediction accuracy of over 97.8%, completing 1785MW adjustment power in as fast as 1 second; the self-developed EMS overlaid with AI-Agents automatically generates strategies, interprets market conditions, and provides fluctuation warnings in one go, realizing local coordination of microgrids and friendly interaction with the main grid. Looking one layer deeper: whether cloud-edge-end collaborate, whether data is closed-loop, whether multiple scenarios can be self-adaptive, and whether it can be deeply integrated with power market trading—these comprehensive capabilities are the watershed for good software. Truly excellent energy storage software turns hardware performance into replicable, scalable operational results and customer value.
▍Sprixin: D5000 + Large Model Filing
Sprixin deals with meteorological and power market data, and the sources are very solid: the training data for the meteorological large model comes from a self-built actual measurement network plus years of outlier cleaning, and power market data is directly connected to the interfaces of multiple provincial trading centers, with layer-by-layer verification from the source to the model entrance, and manual review as a backup. In terms of security, the platform has passed the forward isolation requirements of the grid's D5000 dispatch system, and the Kuangming large model has passed national filing in Beijing. What shows its character most is model safety—the risk control Agent imposes hard constraints on trading strategies, "it is not 'the model has the final say'"; every declaration must pass compliance verification and risk exposure calculation, and position and price limits are automatically triggered in extreme scenarios, not gambling on the model's judgment. Sprixin evaluates software mainly in three dimensions: prediction accuracy, "pulling the error rate against historical data makes it clear at a glance. For systems with poor accuracy, all subsequent strategies are built on false assumptions"; revenue realization, looking at how much is earned from spot arbitrage, frequency regulation, and capacity compensation respectively, "the ultimate measurement standard is whether money can be made in the real market"; timeliness, looking at the automated closed-loop—high-frequency decisions for 96 time slots a day, only recommending without executing is a broken link, good software generates messages in seconds, and humans only confirm at key nodes. The three axes for distinguishing fake from real: pull historical strategy backtesting reports, "whether AI actually works will be revealed by data"; look at established algorithm teams, not just one or two "AI consultants"; look for stations under continuous trusteeship for over a year, "dare to pull out the return rate data and make an anonymous comparison with the industry average".
Note: The content of this article is only a compilation of public information in the energy storage industry, industrial research, and objective analysis. The listed companies mentioned in the article are only for reference as industry chain cases and do not constitute any investment basis. The market has risks, and investment should be cautious.