A cross-channel consistency automatic verification method for dual-channel oriented energy storage system algorithm analysis

Abstract

The application discloses a cross-channel consistency automatic verification method for double channels of energy storage system algorithm analysis, and relates to the technical field of software test automation. In view of the technical defects of the consistency verification of the double channels of the energy storage system algorithm analysis in the prior art, such as great difficulty in input construction, complex result extraction dimension, unmatched comparison strategy and difficulty in difference attribution, the application automatically constructs the consistency verification input case matching the input characteristics of the double channels, analyzes the channels and UI display channels, respectively completes the collection of the calculation output and the interface rendering result, completes the result extraction from multiple dimensions of function logic, numerical accuracy and display style, constructs the cross-channel consistency comparison rule based on the extracted result, realizes the automatic positioning and type attribution of the difference points, and finally generates a standardized consistency verification report. The application is mainly used for the algorithm analysis function test of energy storage related equipment and systems.

Description

Technical Field

[0001] This invention relates to the field of software testing automation technology, and in particular to an automated cross-channel consistency verification method for dual-channel algorithm analysis of energy storage systems. Background Technology

[0002] With the continuous expansion of the scale of new energy storage power stations, energy storage web management systems have evolved from single data monitoring platforms into comprehensive systems integrating data analysis and intelligent operation and maintenance. Currently, mainstream energy storage web systems generally adopt a "dual-channel" architecture to provide algorithm analysis services: on the one hand, users can trigger calculations by selecting the site, equipment, algorithm type, and time interval step by step in a structured form interface through the data analysis panel, obtaining structured results including statistical indicators, analysis reports, and visualization charts; on the other hand, users can initiate the same algorithm analysis request through the embedded AI operation and maintenance assistant, asking questions in natural language. After the AI ​​system interprets the user's intent, it calls the same algorithm engine in the backend and returns the analysis results in natural language.

[0003] The core design requirement of this dual-channel architecture is that, regardless of which channel a user initiates the analysis request, as long as the requests are semantically equivalent (i.e., pointing to the same site, equipment, algorithm, and time interval), the results returned by the two channels should be consistent. This consistency is a key indicator of system reliability—if the UI panel displays "average pressure difference consistency 99.94," while the AI ​​assistant replies with "average pressure difference consistency 98.34" for the same request, it indicates a serious data inconsistency defect between the channels.

[0004] However, current consistency verification for such dual-channel systems faces the following technical challenges: (1) High difficulty in input construction: The UI channel requires operating multi-level device tree controls to select parameters layer by layer, while the AI ​​channel requires constructing semantically equivalent natural language text. The two input forms are completely different. Manually constructing equivalent input pairs is very easy to cause semantic deviations, resulting in unreliable verification results. (2) The results extraction dimensions are complex: the analysis results of the UI channel are scattered in three heterogeneous carriers: DOM statistical cards, analysis report text area and Canvas rendered visualization charts, which cannot be extracted in a unified way; the return results of the AI ​​channel are unstructured natural language paragraphs, which need to be parsed to extract the structured fields corresponding to the UI channel before they can be compared. (3) Difficulty in designing comparison strategies: The dual-channel output includes three modalities: numerical statistical indicators, textual analysis conclusions, and image-based visualization charts. Traditional precise string matching methods cannot be applied. Numerical indicators may have reasonable differences in calculation accuracy, conclusion texts may use different wording to express the same meaning, and charts may have rendering differences but the data is consistent. (4) Difficulty in attributing differences: When inconsistencies are found, it is necessary to determine whether the differences are due to the AI’s misunderstanding of the user’s intent (such as incorrect identification of time range), differences in algorithm calculation accuracy, or logical deviations in the conclusion generation. Currently, there is a lack of systematic methods for classifying and adjudicating differences.

[0005] Existing web automation testing methods (including conventional testing methods based on frameworks such as Playwright and Selenium) are mainly geared towards single-channel, single-modal functional verification. They verify functional correctness through precise location of page elements and fixed assertions, but cannot solve the technical problems mentioned above regarding dual-channel, multimodal, semantic equivalence mapping, and tolerance comparison.

[0006] Therefore, there is a need for an automated cross-channel consistency verification method for dual-channel algorithm analysis of energy storage systems that can achieve full-process automation of cross-channel consistency verification, significantly improve the testing efficiency and defect detection capability of energy storage system algorithm analysis functions. Summary of the Invention

[0007] To address the technical shortcomings of existing technologies for consistency verification of dual-channel algorithms in energy storage systems, such as the difficulty in constructing inputs, the complexity of result extraction dimensions, the mismatch of comparison strategies, and the difficulty in attributing differences, this invention proposes an automated cross-channel consistency verification method for dual-channel algorithms in energy storage systems. This method can achieve full-process automation of cross-channel consistency verification, significantly improving the testing efficiency and defect detection capability of energy storage system algorithm analysis functions.

[0008] The present invention discloses an automated verification method for cross-channel consistency of dual-channel energy storage system algorithm analysis, comprising the following steps: S1. Definition and automatic assembly of analysis parameters for energy storage device tree hierarchy model: Based on the physical device hierarchy of the energy storage system, a parameter model, an algorithm selection path model, and time series parameter standardization rules for a multi-level device tree are defined, and the three types of defined parameters are automatically assembled into a structured analysis request object. S2, UI analysis channel triggering and unified extraction of multimodal results: Based on the structured analysis request object, the data analysis panel is automatically triggered to perform analysis through the Playwright browser automation framework, including layer-by-layer selection of the device tree, step-by-step positioning and leaf node selection of the algorithm classification tree, and setting of the time interval. After the analysis is triggered, multimodal results are extracted from the three types of heterogeneous carriers and assembled into a unified UI channel result snapshot. S3. Automatic mapping of structured parameters to equivalent natural language requests: Based on the structured analysis request object, semantically equivalent natural language question text is automatically generated through predefined mapping template rules; S4, AI question-answering channel triggering and adaptive result reception: Navigate to the AI ​​Operation and Maintenance Assistant interface, initiate a new dialogue session to ensure no interference from historical context; input the natural language question text into the dialog box and send it; use an adaptive waiting strategy to monitor the AI's response text; S5. Structured parsing of unstructured results from AI responses: The AI ​​response text is subjected to structured parsing, which includes time range parsing, statistical indicator parsing, and conclusion parsing; the extracted results are assembled into an AI channel result snapshot; the data structure of the AI ​​channel result snapshot is aligned with the data structure of the UI channel result snapshot, providing a structural matching basis for subsequent cross-channel comparison; S6. Cross-channel comparison within a time range: Perform string matching and comparison between the standardized time range set in S1 and the time range parsed from the AI ​​response in S5; S7. Comparison of statistical indicator tolerances: The seven statistical indicators of the UI channel extracted by S2 are compared one by one with the corresponding indicators of the AI ​​channel extracted by S5; after comparing the seven indicators one by one, a statistical indicator consistency report is generated, recording the comparison results and the type of difference for each item.

[0009] S8. Conclusion: Semantic slot comparison of text. For the analysis conclusion text returned by the dual channels, the semantic slot model is defined as a triple structure; the corresponding conclusion paragraphs are located from the report text of the UI channel and the response text of the AI ​​channel based on predefined analysis dimension keywords, and evaluation level words and trend direction words are extracted from them; the slot triples extracted from the two channels are compared dimension by dimension, and when the analysis dimension and evaluation level match, the current conclusion is determined to be consistent; S9. Consistency Integration Decision and Difference Classification: Summarize all the comparison results of the S6 time range, the S7 statistical indicators, and the S8 conclusion text, and generate a final consistency judgment according to the preset adjudication rules.

[0010] Furthermore, it also includes: S10, Batch Validation and Regression Execution of Multiple Algorithms: The cross-channel consistency verification process from S1 to S9 is encapsulated into a reusable verification unit, with algorithm type and time interval as parameterized inputs; complete cross-channel verification is performed sequentially for different algorithm types; regression execution is performed on each algorithm under different time windows to detect the stability of the system's consistency performance under different data volumes; the results of all verification rounds are summarized to generate a global cross-channel consistency verification report.

[0011] Further: In S1, the definition of the algorithm selection path model specifically means: classifying the algorithms into a tree structure according to their functions and supporting hierarchical positioning from the classification node to the leaf node of the specific algorithm; The specific standardization rule for defining time series parameters is as follows: the date range input by the user is uniformly converted into the complete timestamp format required by the system; The structured analysis request object includes four fields: site identifier, equipment path, algorithm type, and standardized time interval, which serve as a unified input benchmark for subsequent dual-channel verification.

[0012] Further: In S2, the time interval setting includes an automatic historical month rollback algorithm, which automatically calculates and executes the rollback click count of the calendar control based on the month difference between the target date and the current date; the multimodal result includes: Seven statistical indicators and their corresponding fluctuation coefficients are extracted from the DOM statistics card node, and then the evaluation text is extracted and the chart is captured.

[0013] Furthermore, in S3, the mapping template rules are specifically as follows: the time interval is converted from the standard timestamp format to Chinese date expression, the device path is converted from the hierarchical identifier to the concatenation of device names with numbers, and the algorithm type identifier is converted to the full name of the algorithm; at the same time, semantic equivalence variants are handled, including the mapping of the standard writing of device capacity units and the compatibility of multiple expressions of date format, to ensure that the generated natural language text and structured parameters are strictly equivalent in semantics, and to avoid the AI ​​channel's misunderstanding due to differences in expression.

[0014] Furthermore, in S4, the adaptive waiting strategy is specifically as follows: poll the response content area at fixed intervals, and when the length and hash value of the text content obtained by multiple consecutive polls no longer change, it is determined that the AI ​​response has been completed.

[0015] Furthermore: In S5, The time range parsing is performed by extracting the values ​​of the "start time" and "end time" fields through regular expression matching; The statistical indicator analysis involves extracting seven statistical indicator values ​​item by item from a list-type DOM structure. The conclusion analysis extracts the analysis conclusions from the analysis report text according to a preset semantic slot pattern. The analysis conclusions include four categories: consistency evaluation, time series trend evaluation, data fluctuation evaluation, and operational status evaluation.

[0016] Furthermore, in S6, the specific process of comparing the time ranges includes: If the start and end times are exactly the same, it is determined that the AI ​​has correctly understood the time window of the analysis request, and the time dimension verification is passed; if they are inconsistent, it is marked as a "channel understanding deviation" difference, the deviation details are recorded, and subsequent indicator and conclusion comparisons are skipped.

[0017] Furthermore: In S7, the specific process of comparing the tolerances of the statistical indicators includes: The raw data format for each indicator is "principal value / volatility coefficient". First, the principal value and volatility coefficient are separated. If the principal values ​​are completely consistent, the indicator passes. If there is a difference in the principal values, the relative deviation rate is calculated. When the relative deviation rate is less than the preset tolerance threshold, it is judged as "calculation accuracy difference", which means it is acceptable. Otherwise, it is judged as "indicator inconsistency", which means it is not acceptable. At the same time, the volatility coefficient is compared. If there is an order of magnitude difference in the volatility coefficient assessment of the same indicator between the two channels, it is additionally marked as "data stability assessment divergence".

[0018] Furthermore: In S9, the specific process of the consistency determination includes: When the time range is consistent, all indicators pass the comparison, and all conclusion slots match, the decision is "completely consistent"; when the time range is consistent but there are indicator deviations within the tolerance range or individual conclusion slots do not match, the decision is "basically consistent" and a difference detail is attached; when the time range is inconsistent or there are indicator deviations exceeding the tolerance or multiple conclusion slots do not match, the decision is "inconsistent" and a difference classification report is output.

[0019] The beneficial effects of this invention are: 1. A cross-channel consistency automated verification method for a dual-channel architecture combining structured analysis panel and AI natural language question answering is proposed for energy storage systems, filling the technical gap that existing testing methods cannot cover dual-channel output consistency verification. 2. By using the tree-level parameter model of energy storage devices and the equivalence mapping rules from structured parameters to natural language, the technical challenge of automatically constructing dual-channel semantically equivalent inputs is solved, ensuring the credibility of the verification input; 3. By using a multimodal unified extraction method that extracts DOM statistics, captures Canvas 2D charts, and parses report text, as well as a structured parsing method for AI responses, the alignment and comparison problem of heterogeneous output results was solved; 4. By employing numerical tolerance comparison, semantic slot extraction, and multi-factor comprehensive adjudication strategies, the technical challenge of inconsistent determination of multimodal results has been solved, enabling the differentiation between acceptable differences in computational precision and unacceptable system defects. 5. Through parameterized encapsulation and batch regression execution mechanisms, it achieves full-coverage consistency verification across multiple algorithm types and time windows, demonstrating good scalability and regression testing capabilities. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of the cross-channel consistency automated verification method; Figure 2 A schematic diagram of the tree-level parameter model and structured request assembly for energy storage devices; Figure 3 A schematic diagram illustrating the extraction of multimodal results and the construction of a unified snapshot for UI analysis channels; Figure 4 A schematic diagram illustrating the mapping rules from structured parameters to equivalent natural language requests; Figure 5 This is a schematic diagram of a multi-dimensional tolerance consistency comparison and comprehensive adjudication strategy. Detailed Implementation

[0021] The following are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. The embodiments described below are only for explaining the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention should be determined by the scope of the claims. The embodiments of the present invention are described in detail below. In order to facilitate the description of the present invention and simplify the description, the technical terms used in the specification of the present invention should be interpreted broadly, including but not limited to conventional alternatives not mentioned in this application, and including both direct and indirect implementation methods.

[0022] Example 1 Combination Figures 1-5 This embodiment describes an automated verification method for cross-channel consistency in dual-channel energy storage system algorithm analysis, comprising the following steps: S1. Definition and automatic assembly of analysis parameters for energy storage device tree hierarchy model: like Figure 2 As shown, based on the physical device hierarchy of the energy storage system ("site → container → cluster → individual cell"), a multi-level device tree parameter model is defined; an algorithm selection path model is defined, classifying algorithms by function into a tree structure and supporting step-by-step positioning from the classification node to the leaf node of the specific algorithm; a time-series parameter standardization rule is defined, uniformly converting the user-input date range into the complete timestamp format required by the system; the above three types of parameters are automatically assembled into a structured analysis request object, containing four fields: site identifier, device path, algorithm type, and standardized time range, serving as the unified input benchmark for subsequent dual-channel verification.

[0023] S2, UI analysis channel triggering and unified extraction of multimodal results: Based on the structured analysis request object generated by S1, the data analysis panel, the algorithm classification tree for hierarchical positioning and leaf node selection, and the time interval setting operation are driven by the Playwright browser automation framework. The time interval setting includes an automatic historical month rollback algorithm, which automatically calculates and executes the rollback click count of the calendar control based on the month difference between the target date and the current date. like Figure 3 As shown, after the analysis is triggered, the results are extracted from the three types of heterogeneous carriers and assembled into a unified UI channel result snapshot: Seven statistical indicators (mean, standard deviation, minimum, maximum, median, first quartile Q1, and third quartile Q3) and their corresponding fluctuation coefficients are extracted from the DOM statistics card nodes. Consistency evaluation conclusion text, trend evaluation, fluctuation evaluation, and operational status evaluation are extracted from the analysis report text area. JavaScript code is injected into the browser through the page.evaluate() method provided by Playwright to obtain the underlying data of the chart instance for the two-dimensional chart elements rendered by Canvas. Time-value coordinate pairs of key nodes (extreme points, first and last points, and equally spaced sampling points) are extracted to form the chart coordinate sequence.

[0024] S3. Automatic mapping of structured parameters to equivalent natural language requests: Based on the structured analysis request object described in S1, semantically equivalent natural language question text is automatically generated through predefined mapping template rules; The mapping template rules include: The time interval was changed from the standard timestamp format to Chinese date representation; the device path was changed from hierarchical identifier to concatenated device names with numbers; and the algorithm type identifier was changed to the full name of the algorithm. Simultaneously, it handles semantically equivalent variants, including the mapping of standardized notation of device capacity units and compatibility with various expressions of date formats, ensuring that the generated natural language text is strictly semantically equivalent to the structured parameters, and avoiding misunderstandings in the AI ​​channel due to differences in expression.

[0025] S4, AI question-answering channel triggering and adaptive result reception: Use Playwright to navigate the browser to the AI ​​operations assistant dialogue interface, initiate a new dialogue session to ensure no interference from historical context; input the natural language question text generated by S3 into the dialog box and send it; An adaptive waiting strategy based on content stability is used to monitor the AI ​​response status. The AI ​​response is determined to be complete when the length and hash value of the text content obtained from multiple consecutive polls no longer change. This replaces the traditional fixed timeout waiting method and effectively adapts to the high uncertainty of AI response time.

[0026] S5. Structured parsing of unstructured results from AI responses: Structured field extraction is performed on the AI ​​natural language response text obtained by S4: The time range parsing is performed by extracting the values ​​of the "start time" and "end time" fields through regular expression matching; The statistical indicator analysis involves extracting seven statistical indicator values ​​item by item from a list-type DOM structure. The conclusion analysis extracts the analysis conclusions from the analysis report text according to a preset semantic slot pattern. The analysis conclusions include four categories: consistency evaluation, time series trend evaluation, data fluctuation evaluation, and operational status evaluation. The extracted results are assembled into an AI channel result snapshot, whose data structure is aligned with the UI channel result snapshot of S2, providing a structural matching basis for subsequent cross-channel comparison.

[0027] S6. Precise cross-channel comparison within a time range: The standardized time interval set in S1 is compared with the time range parsed from the AI ​​response in S5 using precise string matching. If the start and end times are completely consistent, it is determined that the AI ​​has correctly understood the time window of the analysis request, and the time dimension verification is passed. If they are inconsistent, they are marked as "channel understanding deviation" differences, the deviation details are recorded, and subsequent comparisons of indicators and conclusions are skipped (because different time windows mean that the two channels are not analyzing the same batch of data, and subsequent comparisons are meaningless).

[0028] S7. Comparison of statistical indicator tolerances: The seven statistical indicators extracted from the UI channel in S2 are compared one by one with the corresponding indicators extracted from the AI ​​channel in S5. The original data format of each indicator is "principal value (fluctuation coefficient)". First, the principal value and the fluctuation coefficient are separated. The principal value comparison uses a four-decimal-place precision matching. If the principal values ​​are completely consistent, the item passes. If there is a difference in the principal values, the relative deviation rate is calculated. When the relative deviation rate is less than the preset tolerance threshold, it is judged as "calculation precision difference" (acceptable). Otherwise, it is judged as "indicator inconsistency" (unacceptable). At the same time, the fluctuation coefficient is compared. If there is an order of magnitude difference in the evaluation of the fluctuation coefficient of the same indicator between the two channels, it is additionally marked as "data stability evaluation divergence". After comparing the seven indicators one by one, a statistical indicator consistency report is generated, recording the comparison results and difference types of each item.

[0029] S8. Conclusion: Semantic slot comparison of text. For the analysis conclusion text returned by the dual channels, the semantic slot model is defined as a triple structure {analysis dimension, evaluation level, trend direction}. Based on predefined analysis dimension keywords (such as "consistency average", "time series trend", "data fluctuation range", "system operating status"), the corresponding conclusion paragraphs are located from the report text of the UI channel and the response text of the AI ​​channel. Evaluation level words (such as "very high", "good", "average", "poor") and trend direction words (such as "stable", "rising", "falling", "fluctuating") are extracted from these. The slot triples extracted from the two channels are compared dimension by dimension. When both the analysis dimension and evaluation level match, the conclusion is considered consistent. The text descriptions of the two channels are not required to be completely identical; only semantic slot matching is required.

[0030] S9. Consistency Integration Decision and Difference Classification: Summarize all results from S6 time range comparison, S7 statistical indicator comparison, and S8 conclusion text comparison, and generate a final consistency judgment according to the preset adjudication rules: when the time range is consistent, all indicator comparisons pass, and all conclusion slots match, the judgment is "completely consistent"; when the time range is consistent but there are indicator deviations within the tolerance range or individual conclusion slots do not match, the judgment is "basically consistent" and a detail of the differences is attached; when the time range is inconsistent, or there are indicator deviations exceeding the tolerance or multiple conclusion slots do not match, the judgment is "inconsistent" and a difference classification report is output; the difference classification includes three categories: channel understanding deviation (AI misidentifies request parameters), calculation accuracy difference (numerical deviation is within an acceptable range), and conclusion semantic deviation (inconsistent evaluation level or trend judgment).

[0031] S10, Batch Validation and Regression Execution of Multiple Algorithms: like Figure 2 As shown, the cross-channel consistency verification process from S1 to S9 is encapsulated into a reusable verification unit, with algorithm type and time interval as parameterized inputs. Complete cross-channel verification is performed sequentially for different algorithm types (including comprehensive consistency analysis, differential pressure consistency analysis, temperature difference consistency analysis, module differential pressure analysis, parallel differential pressure analysis, etc.). Regression execution is performed on each algorithm under different time windows to detect the stability of the system's consistency performance under different data volumes. The results of all verification rounds are summarized to generate a global cross-channel consistency verification report, which includes the adjudication results, difference distribution statistics, and trend analysis of each algorithm for each time window.

[0032] Example 2 This embodiment is described in conjunction with Example 1. This embodiment uses the "consistency analysis" algorithm family of a certain energy storage web management system as the verification object. The system provides two algorithm analysis entry points: the "data prediction analysis" panel (UI channel) and the "Xuanwei Operation and Maintenance Assistant" AI dialogue channel. It is necessary to verify the consistency of the output of the two channels for the same analysis request.

[0033] S1, such as Figure 4 As shown, the overall cross-channel consistency verification process is as follows: Initiate verification by loading the list of algorithms to be verified and the time window configuration. Automatically assemble the structured analysis request object based on the configuration parameters.

[0034] Taking "differential pressure consistency analysis" as an example, the assembly results are as follows: { station: "YunChu 95MW / 380MWH First Grid-Side Energy Storage Power Station" container: "#1 Standard Box" cluster: "#1 cluster", Algorithm: "Differential Pressure Consistency Analysis" timeRange: { start: "2025-08-17 00:00:00", end: "2025-08-19 23:59:59"} } S2. In the UI analysis channel, analysis is automatically triggered using the Playwright driver based on the aforementioned request object. The specific process is as follows: (1) Select the equipment tree layer by layer: First, select the station as “YunChu 95MW / 380MWH First Grid-Side Energy Storage Power Station” in the data dimension drop-down box. Then, click the expand icon of the “#1 Standard Box” node in the equipment tree control container, expand its child node cluster, and then click to select “#1 Cluster”.

[0035] (2) Algorithm classification tree location: In the algorithm drop-down box, first locate the "consistency analysis" classification node and click to expand, then select the leaf node algorithm as "pressure difference consistency analysis".

[0036] (3) Time interval setting (including automatic rollback algorithm for historical months): Calculate the difference between the target month and the current month: month_diff = (current year - target year) × 12 + (current month - target month). After opening the start time calendar control, click the "last month" button month_diff several times to roll back the calendar to the target month, and then select the target date. The end time is set in the same way.

[0037] (4) Trigger analysis and wait for results: Click the “Start Analysis” button and wait for the algorithm to complete the calculation.

[0038] (5) Extraction of multimodal results: Statistical indicator extraction: Click "View Details" to extract the following seven statistical indicators and their corresponding fluctuation coefficients from the pop-up details dialog box by locating the DOM statistics card node: Average value: 99.9377 (0.0006); Standard deviation: 0.0524 (0.0005); Minimum value: 99.8313 (0.0002); Maximum value: 99.9836 (0.0017); Median: 99.9696(0.0003); First quartile (Q1): 99.8723 (0.0002); Third quartile (Q3): 99.9755 (0.0013); Evaluation text extraction: Extract evaluation text from the analysis report area, such as "The average pressure difference consistency is very high, which is an excellent performance."

[0039] Chart Capture: For a 2D chart element rendered on a Canvas page, select key node data and inject JavaScript to obtain the underlying data of the chart instance, extracting the time-value coordinate pairs of the key nodes. Specifically, for the 2D chart element rendered on the Canvas, inject JavaScript code into the browser using the `page.evaluate()` method provided by Playwright to obtain the underlying data of the chart instance, extract the time-value coordinate pairs of the key nodes, and form a chart coordinate sequence; the key nodes include extreme points, first and last points, and equally spaced sampling points.

[0040] (6) Assemble the above results into a UI channel result snapshot object: ui_snapshot={ stats:[ {name:"Average", value:99.9377, coefficient:0.0006} {name:"Standard deviation", value:0.0524, coefficient:0.0005} ... ] conclusions:[ "The average differential pressure consistency is very high, demonstrating excellent performance." "Time series trend: The overall change in differential pressure is stable, and the system stability remains good." "Data volatility (IQR level): The box height (IQR) is small, the data dispersion is low, and the overall consistency is good." "The system is operating excellently, and the differential pressure consistency control is outstanding." ] charts:[{key_points:[{x,y,label},...]}] timeRange: {start:"2025-08-17 00:00:00", end:"2025-08-19 23:59:59"} } S3. Based on the structured analysis request object, automatically generate semantically equivalent natural language question text using predefined mapping template rules; such as Figure 4 As shown, the mapping rules are as follows: Time interval conversion: "2025-08-17" → "August 17, 2025"; Equipment path splicing: "#1 standard box" + "#1 cluster" → "#1 standard box #1 cluster"; Site name mapping: "YunChu 95MW / 380MWH First Grid-Side Energy Storage Power Station" → "YunChu 95MW / 380MWH First Grid-Side Energy Storage Site" (wherein: in the site name used by the AI ​​channel, "MWh" is written as "MWH" and "power station" is written as "site". The mapping rules automatically handle such equivalent variants). Algorithm type mapping: "Differential Pressure Consistency Analysis" → "Differential Pressure Consistency Analysis" (direct pass-through); The final generated natural language query text is: “Perform differential pressure consistency analysis on the #1 standard box and #1 cluster of the YunChu 95MW / 380MWH grid-side energy storage site from August 17, 2025 to August 19, 2025”; S4. Using the Playwright browser automation framework, navigate the browser to the AI ​​operations and maintenance assistant dialogue interface, click to start a new dialogue, enter the natural language question text, and send it. An adaptive waiting strategy is adopted: poll the text content of the AI ​​response area every 3 seconds, record the text length and content hash value, and determine that the response is complete when the length and hash value remain unchanged for 3 consecutive polls.

[0041] S5. Structure the results from the AI's response text. The parsing process is as follows: Time range parsing: Using regular expressions to match "start time:(". "?)" and "End time:(". The "?" mode extracts the time range for AI recognition.

[0042] Statistical indicator analysis: Extract the indicator text item by item from the list DOM structure in the reply, such as "Average value: 99.9377 (0.0006)", and separate the main value and the fluctuation coefficient.

[0043] Conclusion Analysis: The conclusion paragraphs were located and extracted from the text based on keywords in four dimensions: "consistency average", "time series trend", "data fluctuation range" and "system operation status".

[0044] Assemble the parsed results into an AI channel result snapshot object: ai_snapshot={ stats:[ {name:"Average", value:99.9377, coefficient:0.0006}, ... ], conclusions:[ "The average differential pressure consistency is very high, demonstrating excellent performance." "Time series trend: The overall change in differential pressure is stable, and the system stability remains good." ... ], timeRange:{start:"2025-08-17 00:00:00", end:"2025-08-19 23:59:59"} } S6. Cross-channel comparison within a time range; The standardized time range ui_snapshot.timeRange set in S1 is compared with the time range ai_snapshot.timeRange parsed from the AI ​​response in S5. If the start and end times are exactly the same, it is determined that the AI ​​correctly understood the time window of the analysis request, and the time dimension verification is passed. If they are inconsistent, it is marked as a "channel comprehension deviation" difference, the deviation details are recorded, and subsequent comparisons of indicators and conclusions are skipped. This is because different time windows mean that the two channels are not analyzing the same batch of data, and subsequent comparisons are meaningless.

[0045] Comparing ui_snapshot.timeRange and ai_snapshot.timeRange: Start time: "2025-08-17 00:00:00" == "2025-08-17 00:00:00" → Consistent; End time: "2025-08-19 23:59:59" == "2025-08-19 23:59:59" → Consistent; The time dimension verification passed; we will continue with subsequent comparisons.

[0046] S7. Comparison of Statistical Indicator Tolerances. Compare each of the seven statistical indicators: Take the "average" as an example: UI channel: Main value = 99.9377, volatility coefficient = 0.0006; AI Channel: Main value = 99.9377, volatility coefficient = 0.0006; The principal value comparison uses a four-decimal-digit precision matching. If the principal value matches precisely (four decimal places are the same), then this item passes. Consistent volatility coefficients lead to consistent data stability assessments; After all seven statistical indicators passed the comparison, the statistical indicator consistency report was "all consistent".

[0047] S8. Semantic slot comparison of the conclusion text. Taking "average conclusion" as an example: UI channel conclusion: "The average differential pressure consistency is very high, demonstrating excellent performance." AI channel conclusion: "The average differential pressure consistency is very high, demonstrating excellent performance." Slot extraction: UI → {Dimension: "Average Pressure Difference Consistency", Rating: "Very High", Trend: None}; AI → {Dimension: "Average Pressure Difference Consistency", Rating: "Very High", Trend: None}; If all three slots match, then the conclusion is consistent.

[0048] Taking "Time Series Trend Conclusions" as an example: UI Channel: "Time series trend: Overall pressure difference consistency changes steadily, and system stability remains good." AI Channel: "Time series trend: Overall pressure difference consistency changes steadily, and system stability remains good." Slot extraction: UI → {Dimension: "Time Series Trend", Rating: "Good", Trend: "Stable"}; AI → {Dimension: "Time Series Trend", Rating: "Good", Trend: "Stable"}; If all three slots match, then the conclusion is consistent. After all four conclusions are matched in their respective slots, the conclusion consistency report will show "All consistent".

[0049] S9, Consistency Comprehensive Decision.

[0050] Time range: consistent; Statistical indicators: All consistent; Conclusion text: All consistent; Final verdict: Completely consistent.

[0051] S10. Switch to the next algorithm (e.g., "Temperature Difference Consistency Analysis"), and repeat S1 to S9. Perform cross-channel verification of multiple algorithms, such as comprehensive consistency analysis, pressure difference consistency analysis, and temperature difference consistency analysis, in sequence.

[0052] S11. Summarize the cross-channel verification results of all algorithms and generate a global consistency verification report, including: The ruling results for each algorithm type (completely consistent / mostly consistent / inconsistent); Statistics on the distribution of differences (the number and proportion of each type of difference); Comparison of verification results across different time windows.

[0053] S12, End verification.

[0054] In S2, the automatic rollback algorithm for historical months in the time interval setting is as follows: S201. Obtain the year Y_now and month M_now of the current system date.

[0055] S202. Obtain the year Y_target and month M_target of the target analysis date.

[0056] S203. Calculate the monthly difference month_diff=(Y_now-Y_target)×12+(M_now-M_target).

[0057] S204. Open the calendar control and repeatedly execute the "Click last month" operation month_diff times, waiting 500ms after each operation to ensure the calendar animation is complete.

[0058] S205. After locating the target month on the calendar, select the target date number.

[0059] This algorithm ensures that the target date can be automatically and accurately located regardless of the distance between the current date and the target date, without the need for manual intervention.

[0060] The adaptive waiting strategy in S4 is as follows: S401. After sending a natural language question, initialize the polling timer and the stable counter stable_count=0.

[0061] S402. Wait for the basic response time (e.g., 3 seconds) to allow the AI ​​to start generating a response.

[0062] S403. Obtain the text content of the current reply area, and calculate the text length L_current and the content hash value H_current.

[0063] S404: Compare with the length L_last and hash value H_last of the previous poll: If L_current == L_last and H_current == H_last, then increment stable_count by 1; Otherwise, stable_count is reset to 0, and L_last = L_current and H_last = H_current are updated.

[0064] S405: Determine if stable_count has reached a stable threshold (e.g., remains unchanged for 3 consecutive times): If the condition is met, the AI ​​response is deemed complete, and the process proceeds to S5. If the polling interval is not reached, return to S403 after waiting for the polling interval (e.g., 3 seconds).

[0065] S406: If the total polling time exceeds the maximum waiting limit (e.g., 5 minutes), it is determined that the AI ​​response has timed out, the exception is recorded and the current verification is terminated.

[0066] Compared to a fixed timeout, this strategy can adapt to the actual generation time of the AI ​​response—shorter waiting time for simple questions with quick responses and longer waiting time for complex questions with slow responses. This avoids both response truncation caused by too short a fixed waiting time and inefficiency caused by too long a fixed waiting time.

[0067] Detailed rules for semantic slot matching are as follows: Figure 5 As shown, the detailed rules for semantic slot matching in S9 are as follows: S901: A list of predefined analysis dimension keywords, including: "consistency average", "time series trend", "data volatility / IQR level", and "system operating status".

[0068] S902: For each dimension keyword, locate the sentence or paragraph containing that keyword in the conclusion text of both channels.

[0069] S903: Extract rating words from the located paragraphs. The predefined rating dictionary includes positive ratings: "very high", "very high", "high", "good", "excellent", "outstanding"; and negative ratings: "lower", "low", "poor", "insufficient", "abnormal".

[0070] S904: Extract trend direction words from the located paragraph. The predefined trend dictionary includes: "stable", "up", "down", "fluctuation", and "drift".

[0071] S905: Assemble the slot triples for this dimension: {Dimension, Rating, Trend Direction}. If a field does not appear in the text, it is marked as "Not Mentioned".

[0072] S906: Dimensionally compare slot triples of two channels: Dimension fields must match exactly (to ensure that the same conclusion is being compared). The evaluation level is matched after normalization of synonyms (e.g., "excellent" and "superior" are normalized to the same level). Trend direction is matched using synonym normalization (e.g., "stable" and "steady" are normalized to the same trend); S907: Records the comparison results (match / non-match / partial match) for each dimension and the specific fields of the non-match.

[0073] This method does not require the textual expressions of the two channels to be exactly the same. Instead, it achieves correct matching of semantically equivalent but worded conclusions through semantic slot extraction and synonym normalization, effectively reducing the false judgment rate.

[0074] Test example: The testers prepared to verify the configuration, specifying the energy storage site to be verified as "YunChu 95MW / 380MWH First Grid-Side Energy Storage Power Station", the equipment as "#1 Standard Box #1 Cluster", the list of algorithms to be verified as [Comprehensive Consistency Analysis, Pressure Difference Consistency Analysis, Temperature Difference Consistency Analysis], and the time window as August 17 to August 19, 2025.

[0075] After the system starts, it first performs cross-channel verification on "Comprehensive Consistency Analysis": select the site in the UI panel, expand the equipment tree and select cluster #1 under standard box #1, locate the "Consistency Analysis" category in the algorithm tree and select "Comprehensive Consistency Analysis", set the time interval and trigger the analysis; extract seven indicators such as average value 99.9377 (0.0006) and standard deviation 0.0524 (0.0005) and four analysis conclusions; then automatically generate an equivalent question "Conduct comprehensive consistency analysis on cluster #1 of standard box #1 of the first grid-side energy storage site of YunChu 95MW / 380MWH from August 17, 2025 to August 19, 2025", and send it to the AI ​​operation and maintenance assistant; after the AI ​​replies, it automatically parses its statistical indicators and conclusions; after multi-dimensional comparison, the time range is consistent, the seven indicators are completely matched, and the semantic slots of the four conclusions are all matched, and the decision is "completely consistent".

[0076] The system automatically switches to "pressure difference consistency analysis" and "temperature difference consistency analysis" to repeat the above process, and finally generates a global verification report. All three algorithms pass cross-channel consistency verification within this time window.

Claims

1. An automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems, characterized in that, Includes the following steps: S1. Definition and automatic assembly of analysis parameters for energy storage device tree hierarchy model: Based on the physical device hierarchy of the energy storage system, a parameter model, an algorithm selection path model, and time-series parameter standardization rules for a multi-level device tree are defined, and the three types of defined parameters are automatically assembled into a structured analysis request object. S2, UI analysis channel triggering and unified extraction of multimodal results: Based on the structured analysis request object, the data analysis panel is automatically triggered to perform analysis through the Playwright browser automation framework, including layer-by-layer selection of the device tree, step-by-step location of the algorithm classification tree and selection of leaf nodes, and setting of time intervals. After the analysis is triggered, multimodal results are extracted from the three types of heterogeneous carriers and assembled into a unified UI channel result snapshot. S3. Automatic mapping of structured parameters to equivalent natural language requests: Based on the structured analysis request object, semantically equivalent natural language question text is automatically generated through predefined mapping template rules; S4, AI question-answering channel triggering and adaptive result reception: Navigate to the AI ​​Operation and Maintenance Assistant interface, initiate a new dialogue session to ensure no historical context interference; input the natural language question text into the dialog box and send it; use an adaptive waiting strategy to monitor the AI's response text; S5. Structured parsing of unstructured results from AI responses: The AI ​​response text is subjected to structured parsing, which includes time range parsing, statistical indicator parsing, and conclusion parsing; the extracted results are assembled into an AI channel result snapshot; the data structure of the AI ​​channel result snapshot is aligned with the data structure of the UI channel result snapshot, providing a structural matching basis for subsequent cross-channel comparison; S6. Cross-channel comparison within a time range: Perform string matching and comparison between the standardized time range set in S1 and the time range parsed from the AI ​​response in S5; S7. Comparison of statistical indicator tolerances: The seven statistical indicators of the UI channel extracted by S2 are compared one by one with the corresponding indicators of the AI ​​channel extracted by S5; after comparing the seven indicators one by one, a statistical indicator consistency report is generated, recording the comparison results and the type of difference for each item. S8. Conclusion: Semantic slot comparison of text. For the analysis conclusion text returned by the dual channels, the semantic slot model is defined as a triple structure; the corresponding conclusion paragraphs are located from the report text of the UI channel and the response text of the AI ​​channel based on predefined analysis dimension keywords, and evaluation level words and trend direction words are extracted from them; the slot triples extracted from the two channels are compared dimension by dimension, and when the analysis dimension and evaluation level match, the current conclusion is determined to be consistent; S9. Consistency Integration Decision and Difference Classification: Summarize all the comparison results of the S6 time range, the S7 statistical indicators, and the S8 conclusion text, and generate a final consistency judgment according to the preset adjudication rules.

2. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, Also includes: S10, Batch Validation and Regression Execution of Multiple Algorithms: The cross-channel consistency verification process from S1 to S9 is encapsulated into a reusable verification unit, with algorithm type and time interval as parameterized inputs; complete cross-channel verification is performed sequentially for different algorithm types; regression execution is performed on each algorithm under different time windows to detect the stability of the system's consistency performance under different data volumes; the results of all verification rounds are summarized to generate a global cross-channel consistency verification report.

3. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S1, the definition of the algorithm selection path model specifically means: classifying the algorithms by function into a tree structure and supporting hierarchical positioning from the classification node to the leaf node of the specific algorithm; The specific standardization rule for defining time series parameters is as follows: the date range input by the user is uniformly converted into the complete timestamp format required by the system; The structured analysis request object includes four fields: site identifier, equipment path, algorithm type, and standardized time interval, which serve as a unified input benchmark for subsequent dual-channel verification.

4. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S2, the time interval setting includes an automatic rollback algorithm for historical months, which automatically calculates and executes the rollback click count of the calendar control based on the month difference between the target date and the current date; The multimodal results include: Seven statistical indicators and their corresponding fluctuation coefficients are extracted from the DOM statistics card node, and then the evaluation text is extracted and the chart is captured.

5. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S3, the mapping template rules are as follows: the time interval is converted from the standard timestamp format to Chinese date expression, the device path is converted from the hierarchical identifier to the concatenation of device names with numbers, and the algorithm type identifier is converted to the full name of the algorithm; at the same time, semantic equivalence variants are handled, including the mapping of the standard writing of device capacity units and the compatibility of multiple expressions of date format, to ensure that the generated natural language text and structured parameters are strictly equivalent in semantics, and to avoid the AI ​​channel's misunderstanding due to differences in expression.

6. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S4, the adaptive waiting strategy is specifically as follows: poll the response content area at fixed intervals, and when the length and hash value of the text content obtained by multiple consecutive polls no longer change, it is determined that the AI ​​response has been completed.

7. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S5, The time range parsing is performed by extracting the values ​​of the "start time" and "end time" fields through regular expression matching; The statistical indicator analysis involves extracting seven statistical indicator values ​​item by item from a list-type DOM structure. The conclusion analysis extracts the analysis conclusions from the analysis report text according to a preset semantic slot pattern. The analysis conclusions include four categories: consistency evaluation, time series trend evaluation, data fluctuation evaluation, and operational status evaluation.

8. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S6, the specific process of comparing the time ranges includes: If the start and end times are exactly the same, it is determined that the AI ​​has correctly understood the time window of the analysis request, and the time dimension verification is passed; if they are inconsistent, it is marked as a "channel comprehension deviation" difference, the deviation details are recorded, and subsequent comparison of indicators and conclusions is skipped.

9. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S7, the specific process of comparing the tolerances of the statistical indicators includes: The raw data format for each indicator is "principal value / volatility coefficient". First, the principal value and volatility coefficient are separated. If the principal values ​​are completely consistent, the indicator passes. If there is a difference in the principal values, the relative deviation rate is calculated. When the relative deviation rate is less than the preset tolerance threshold, it is judged as "calculation accuracy difference", which means it is acceptable. Otherwise, it is judged as "indicator inconsistency", which means it is not acceptable. At the same time, the volatility coefficient is compared. If there is an order of magnitude difference in the volatility coefficient assessment of the same indicator between the two channels, it is additionally marked as "data stability assessment divergence".

10. The automated verification method for cross-channel consistency of dual-channel algorithms in energy storage systems according to claim 1, characterized in that, In S9, the specific process of consistency determination includes: When the time range is consistent, all indicators pass the comparison, and all conclusion slots match, the decision is "completely consistent"; when the time range is consistent but there are indicator deviations within the tolerance range or individual conclusion slots do not match, the decision is "basically consistent" and a difference detail is attached; when the time range is inconsistent or there are indicator deviations exceeding the tolerance or multiple conclusion slots do not match, the decision is "inconsistent" and a difference classification report is output.

Images