Energy hosting method based on retrieval enhancement generation and multi-agent collaboration

Abstract

The invention relates to an energy hosting method based on retrieval enhancement generation and multi-agent cooperation, and the method comprises the following steps: S1, carrying out the real-time synchronization of a plurality of time sequence operation data of hosting equipment, carrying out the fitting of a plurality of straight line segments and arc segments, and forming a continuous and smooth alternative curve; s2, the geometric description of the alternative curve is combined with the offline document data to be converted into natural language description, vectorization processing is carried out, and a mixed index knowledge base containing a vector database and a knowledge graph is constructed; s3, receiving a user instruction and disassembling the user instruction into a plurality of subtasks; s4, performing query expansion by utilizing concepts related to query entities in the knowledge graph during vector retrieval through the mixed index knowledge base, and reserving a high-score retrieval result as a high-quality knowledge fragment; and S5, generating an energy equipment control strategy by using the high-quality knowledge fragment, the real-time data of the hosting equipment and the security constraint through a large language model.

Description

Technical Field

[0001] This invention relates to the field of energy management, specifically to an energy management method based on retrieval-enhanced generation and multi-agent collaboration. Background Technology

[0002] Energy management is a professional team that contracts and manages the energy system of an energy-consuming unit, covering equipment operation, maintenance and energy consumption optimization. It can manage the entire chain of photovoltaic energy storage units, chiller units and terminal air conditioning equipment, and at the same time update the technology and equipment of each equipment, thereby achieving the goal of energy saving and saving energy costs.

[0003] Existing energy management methods rely heavily on human experience. Taking chiller group control as an example, load increase / decrease strategies and fault diagnosis depend heavily on the personal experience of operation and maintenance experts, resulting in slow response times and low operational standardization. Furthermore, data silos are severe among various devices. Real-time operating data (such as chiller evaporation pressure / condensation pressure / current percentage), equipment documentation (operation manuals / maintenance guides), historical work orders, and industry standards are scattered across different systems, making it impossible to perform correlation analysis and comprehensive utilization.

[0004] The purpose of this invention is to design an energy management method based on retrieval-enhanced generation and multi-agent collaboration to address the problems existing in the prior art. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention provides an energy management method based on retrieval-enhanced generation and multi-agent collaboration, which can effectively solve at least one of the problems existing in the prior art.

[0006] The technical solution of this invention is: An energy management method based on retrieval-enhanced generation and multi-agent collaboration includes the following steps: S1. Real-time synchronization of multiple time-series operation data of the managed device; screening out abnormal parameters in the time-series operation data whose change process is monotonous and smooth as candidate parameters; performing piecewise linear fitting on the candidate parameters to obtain several straight line segments; performing circular fitting on the data points between adjacent straight line segments to obtain circular arc segments; adjusting the circular arc segments to form a continuous and smooth candidate curve with the corresponding straight line segments; screening out abnormal parameters in the time-series operation data whose change process is monotonous and smooth as candidate parameters includes: calculating the first-order difference of the time-series operation data respectively; if the difference signs tend to be consistent within a preset time window, then calculating the second-order difference of the corresponding time-series operation data; if the absolute value of the second-order difference is less than a preset second-order difference threshold, then it is determined that the change process is monotonous and smooth; screening out the portion of the time-series operation data whose change process is monotonous and smooth that exceeds the alarm threshold or trend threshold as candidate parameters. S2, obtain offline documentation from the managed device, combine the geometric description of the candidate curve with the offline documentation to convert it into a natural language description, perform semantic segmentation and vectorization on the offline documentation and the natural language description, and construct a hybrid index knowledge base containing a vector database and a knowledge graph; S3 receives user commands or system-triggered events, performs intent recognition based on a large language model, and breaks down the target task into multiple sub-tasks; S4. Based on the structured information contained in each subtask, a search is performed in the hybrid index knowledge base. During vector retrieval, the hybrid index knowledge base is used to expand the query by utilizing concepts related to the query entity in the knowledge graph. During knowledge graph query, the hybrid index knowledge base is used to expand the query by utilizing new entities extracted from the fragments recalled by vector retrieval. The two types of search results are filtered and scored, and the search results with high scores are retained as high-quality knowledge fragments. S5, the high-quality knowledge fragments, real-time data from the managed device, and security constraints are used to generate an energy device control strategy through a large language model.

[0007] Furthermore, the alternative parameters include some or all of the following: the cell temperature, SOC, cell differential pressure, and charge / discharge rate of the energy storage battery; the exhaust temperature, condensing pressure, chilled water outlet temperature, and current of the chiller unit; the output power of the photovoltaic inverter; the return air temperature of the terminal air conditioner; and the power factor of the power distribution system.

[0008] Further, piecewise linear fitting is performed on the candidate parameters to obtain several straight line segments, and circular arc fitting is performed on the data points between adjacent straight line segments to obtain circular arc segments. Adjusting the circular arc segments to form a continuous and smooth candidate curve with the corresponding straight line segments includes: The alternative parameters are simplified into polylines connected by several key points, and the length of each polyline is greater than a preset duration, thus obtaining the straight line segment; Least-squares circular arc fitting is performed using transition data points between adjacent straight line segments, and the directions of the preceding and following adjacent straight line segments are used as tangent constraints at the start and end points of the circular arc segment. By adjusting the number of data points involved in the fitting and the range of the data window, a circular arc segment that is continuous and smoothly connected to the adjacent straight line segments is obtained.

[0009] Further, step S2 includes: S2.1, The geometric description of the candidate curve is introduced into the equipment operation rules, parameter constraints and fault mechanisms in the offline document data, and the geometric description is semantically enhanced to generate a professional description that integrates physical meaning and handling suggestions; S2.2, Semantically segment the offline document data slices and the specialized descriptions, then perform vectorization processing and write them into a vector database; S2.3 Extract entity relationships containing devices, parameters, anomalies, causes, measures, and constraints from the offline document data slices and the specialized descriptions to construct a knowledge graph; S2.4 Simultaneously vectorize and extract entity relationships for knowledge units belonging to the same text in the vector database and knowledge graph, and establish bidirectional metadata association between vector database records and knowledge graph nodes, thereby forming a hybrid index knowledge base that combines semantic indexing and structural indexing.

[0010] Further, step S2.1 includes: S2.1.1, The candidate curve is described as a geometric description text including the geometric description by using time-series data textification technology; S2.1.2, retrieve relevant knowledge corresponding to the geometric description text from the offline document data; S2.1.3, the geometric description text and the relevant knowledge are integrated, and the possible causes, handling measures and risk level descriptions of the anomaly are further retrieved from the offline document data and added to generate the professional description.

[0011] Further, step S2.4 includes: The text content of the same knowledge unit is simultaneously vectorized and entity relations are extracted, so that the knowledge unit forms a semantic index in the vector database and a structural index in the knowledge graph. The mapping relationship between semantic index and structural index is established through the knowledge unit identifier, thereby constructing a hybrid index knowledge base that combines semantic index and structural index.

[0012] Further, step S4 includes: The RAG engine searches for the closest text fragment in the vector database based on the structured information contained in each subtask, and then expands the text fragment based on the triples in the knowledge graph corresponding to the text fragment. Based on the structured information contained in each subtask, the closest triple is found in the knowledge graph, and then the triple is expanded based on the text fragments in the vector database corresponding to the triple. A cross-encoder is used to score the relevance of the recalled fragments. The fragments are then filtered based on keywords that are not relevant to the subtask, and the highest-scoring search results are selected as high-quality knowledge fragments.

[0013] Further, step S5 includes: S5.1, based on the parameter values ​​and trends in real-time data, dynamically filter high-quality knowledge fragments and retain knowledge fragments that match the current working conditions; S5.2, based on preset security constraints, limit the boundaries of the selected knowledge fragments and eliminate or correct suggestions that violate physical limits, operating procedures or logical constraints; S5.3 assembles the filtered, eliminated, or corrected knowledge fragments with real-time data summaries and security constraints into structured prompt words, inputs them into a large language model, and generates energy equipment control strategies containing executable control instructions.

[0014] Therefore, the present invention provides the following effects and / or advantages: This application filters multiple time-series operational data, recreates curves by fitting straight lines and arcs to accurately reflect the details of equipment changes reflected in the curves, then forms a bidirectional association between a hybrid index knowledge base using a vector library and a knowledge graph, obtains high-quality knowledge fragments through a collaborative enhancement mechanism of hybrid retrieval, and finally generates a control strategy.

[0015] This application effectively filters out data suitable for alternating combinations of straight lines and circular arcs by calculating the first and second differences of time-series running data. Then, a suitable algorithm converts this curve into a curve with alternating straight lines and circular arcs. Using the geometric description of the circular arcs as an intermediary language between the curve and offline documentation, it effectively reflects more detailed changes in the time-series running data. This application also establishes a bidirectional association between a vector database and a knowledge graph, linking vector data with corresponding data in the knowledge graph and vice versa. In subsequent mixed retrieval processes, it proactively expands the vector data or knowledge graph data, providing knowledge graph data not present in the vector database, thus expanding the query results and obtaining more complete and high-quality knowledge fragments.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0017] It should be understood that the above summary and the following detailed description of the invention are exemplary and explanatory, and are intended to provide further explanation of the invention as claimed. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating one embodiment of the present invention. Detailed Implementation

[0019] To facilitate understanding by those skilled in the art, the present invention will now be described in further detail with reference to the embodiments: refer to Figure 1 An energy management method based on retrieval-enhanced generation and multi-agent collaboration, characterized by the following steps: S1. Real-time synchronization of multiple time-series operation data of the managed device, filtering out abnormal parameters in the time-series operation data whose change process is monotonous and smooth as candidate parameters, performing piecewise linear fitting on the candidate parameters to obtain several straight line segments, performing circular arc fitting on the data points between adjacent straight line segments to obtain circular arc segments, and adjusting the circular arc segments to form a continuous and smooth candidate curve with the corresponding straight line segments. S2, obtain offline documentation from the managed device, combine the geometric description of the candidate curve with the offline documentation to convert it into a natural language description, perform semantic segmentation and vectorization on the offline documentation and the natural language description, and construct a hybrid index knowledge base containing a vector database and a knowledge graph; S3 receives user commands or system trigger events, performs intent recognition based on a large language model, and breaks down the target task into multiple sub-tasks. This step can receive commands from user interaction terminals (Web / App / voice) or trigger events from external systems (work orders / alarms). The intent recognition module determines the task type, and complex tasks are broken down into a sequence of atomic sub-tasks using thought chain technology. For example, upon receiving the command "Optimize tomorrow's energy storage charging and discharging strategy," it is broken down into six sub-tasks: weather retrieval, electricity price query, status acquisition, plan generation, security review, and command issuance.

[0020] S4. Based on the structured information contained in each subtask, a search is performed in the hybrid index knowledge base. During vector retrieval, the hybrid index knowledge base is used to expand the query by utilizing concepts related to the query entity in the knowledge graph. During knowledge graph query, the hybrid index knowledge base is used to expand the query by utilizing new entities extracted from the fragments recalled by vector retrieval. The two types of search results are filtered and scored, and the search results with high scores are retained as high-quality knowledge fragments. S5, the high-quality knowledge fragments, real-time data from the managed device, and security constraints are used to generate an energy device control strategy through a large language model.

[0021] The steps involved in this application and their specific descriptions are as follows.

[0022] Furthermore, the alternative parameters include some or all of the following: the cell temperature, SOC, cell differential pressure, and charge / discharge rate of the energy storage battery; the exhaust temperature, condensing pressure, chilled water outlet temperature, and current of the chiller unit; the output power of the photovoltaic inverter; the return air temperature of the terminal air conditioner; and the power factor of the power distribution system.

[0023] In this embodiment, the temperature of a single cell in the energy storage battery reflects the balance between heat generation and dissipation due to the battery's internal resistance. The heat generation rate varies with the square of the current, while the heat dissipation rate varies linearly with the temperature difference. During constant current charging and discharging, the state of charge (SOC) changes linearly. When power is limited or the strategy is adjusted, the current changes smoothly, and the SOC curve transitions smoothly. The single-cell voltage difference is caused by battery inconsistency, and the voltage difference change is affected by the equalization circuit. The equalization current is usually small and continuously adjusted. The charge / discharge rate is continuously adjusted by the BMS based on temperature, SOC, and power demand, and the adjustment process is smoothed by the control algorithm.

[0024] The exhaust temperature of a chiller unit is related to its compression ratio and intake superheat, and is affected by the cooling water temperature and condenser cleanliness. It has the characteristics of large thermal inertia and slow temperature change. The condensing pressure change is constrained by the refrigerant circulation inertia and changes continuously. The chilled water outlet temperature is smoothed by the thermal inertia of the water system. The current change is smoothed by the motor inertia and control algorithm. When the output power of a photovoltaic inverter is normal, the power changes smoothly with the irradiance. When the irradiance changes smoothly, the process of the power approaching the limit value is controlled by temperature feedback and shows a smooth deceleration.

[0025] The return air temperature of the terminal air conditioner is determined by the balance between the indoor load and the air conditioning cooling capacity. The indoor load changes continuously, and the air conditioner starts, stops, or adjusts its frequency continuously.

[0026] The power factor of the power distribution system is regulated by reactive power compensation devices. The response of the compensation devices is continuous or stepped switching, but the overall response is smooth.

[0027] Overall, the abnormal development of these parameters is a unidirectional process, such as a continuous increase in temperature, a continuous decrease in SOC, and a continuous increase in pressure. Furthermore, these parameters exhibit inertia, meaning their rate of change cannot change abruptly. They are prone to abnormal acceleration or deceleration within abnormal ranges, such as thermal runaway. The parameters naturally conform to the geometric premise of fitting with straight line segments or circular arc segments in subsequent steps, enabling the precise expression of their change process using the fewest possible geometric units, thereby generating a high-quality textual description.

[0028] Further, step S1 includes: S1.1 Calculate the first difference of the time series running data respectively. If the difference signs tend to be consistent within the preset time window, calculate the second difference of the corresponding time series running data. If the absolute value of the second difference is less than the preset second difference threshold, it is determined that the change process is monotonous and smooth. Select the part of the time series running data with monotonous and smooth change process that exceeds the alarm threshold or trend threshold as the candidate parameter. In this step, the first-order difference is used to determine whether the time-series data is monotonically increasing or decreasing within the segmented time period, and the second-order difference is used to determine whether the time-series data has no obvious oscillations and the local curvature changes are gradual. For exceeding alarm thresholds or trend thresholds, refer to the device parameters in the offline documentation of the managed device.

[0029] S1.2, simplify the alternative parameters into a polyline connected by several key points, and the length of each polyline is greater than the preset duration, to obtain the straight line segment; In this step, the curve can be directly calculated using the RDP algorithm to obtain multiple key points, which are then connected to form multiple straight line segments. The preset duration can be set empirically, for example, it should last at least 5 seconds.

[0030] S1.3, perform least squares circular arc fitting using transition data points between adjacent straight line segments, and use the directions of adjacent straight line segments as tangent constraints at the start and end points of the circular arc segment. By adjusting the number of data points involved in the fitting and the range of the data window, a circular arc segment that is continuous and smoothly connected to the adjacent straight line segments is obtained. In this step, by adjusting the number of data points, time window position, and window length involved in the arc fitting between adjacent straight line segments, the fitted arc segment is made to simultaneously meet the fitting error requirement and the tangent continuity requirement. If the window is too large, it may pull in some points of the straight line segment, causing arc deformation and inaccurate fitting; if the window is too small, it may only use a small portion of the data in the transition zone, and the fitted arc may not accurately reflect the physical bending process, and the connection with the two end straight line segments may not be smooth enough. Therefore, it is necessary to adjust the number of data points involved in the fitting and the range of the data window. The intersection of two straight lines can be used as the center, and several points around it can be used to form a window. Arc fitting is performed on windows with different numbers and positions of data. After calculating the fitting error, tangent deviation, and trend preservation error, the smallest data window that simultaneously meets the error constraint conditions is selected as the target window. The error constraint is dynamically adjusted according to the parameter type and current value of the candidate parameters. For example, a larger error can be allowed in the normal range to make the curve transition more smoothly, while for positions that change rapidly and are close to the alarm threshold or trend threshold, a smaller error is needed to make the curve better reflect the changes at that point.

[0031] At this point, a combination of straight line segments and circular arc segments can be obtained, and the corresponding geometric descriptions can be intuitively seen, such as the length and slope of the straight line, and the length, radius, and center of the circular arc segment. It can accurately describe the time and shape of anomalies, and accurately describe nonlinear anomalies such as accelerated rise and decelerated fall of curves, while adapting to curves with different noise levels and rates of change.

[0032] Further, step S2 includes: S2.1, The geometric description of the alternative curve is introduced into the device operation rules, parameter constraints and fault mechanisms in the offline document, and the geometric description is semantically enhanced to generate a professional description that integrates physical meaning and handling suggestions; The purpose of this step is to semantically enhance the geometric description of the candidate curves by combining it with offline documentation. Using the rules, thresholds, mechanisms, and measures in the offline documentation, engineering semantic enhancement is applied to the geometric feature description, forming corresponding relevant statements as knowledge. Specifically: Further, step S2.1 includes: S2.1.1, The candidate curve is described as a geometric description text including the geometric description by using time-series data textification technology; In this step, the geometric description can include the equipment to which the curve belongs, parameter name, time range, slope of several straight line segments, radius of circular arc segments, arc length, bending direction, inflection point location, peak value, final value, duration, etc. Time-series data textification technology can describe the candidate curves through geometric features, such as "The exhaust temperature of chiller No. 3 increases monotonically from 14:10 to 14:40; the first segment is a low-slope straight line segment, the middle segment has an upward convex circular arc transition, the latter segment turns into a high-slope straight line segment, and the final value reaches 98℃." as the geometric description text.

[0033] S2.1.2, retrieve relevant knowledge corresponding to the geometric description text from the offline document data; For example, regarding the statement "exhaust temperature rises monotonically and approaches 98°C", relevant information can be found in the documentation, such as "exhaust temperature exceeding 95°C may pose a high-temperature risk", "continuous rise in exhaust temperature is usually related to insufficient heat dissipation on the condenser side", and "under high-temperature conditions, it is recommended to prioritize load reduction and check the cooling system".

[0034] S2.1.3, the geometric description text and the relevant knowledge are integrated, and the possible causes, handling measures and risk level descriptions of the anomaly are further retrieved from the offline document data and added to generate the professional description.

[0035] In this step, the relevant knowledge in S2.1.2 is further expanded by adding several related enhancements. Then, the geometric description text and the relevant knowledge are integrated to obtain a more professional description, such as "The exhaust temperature of chiller No. 3 has been rising continuously for the past 30 minutes, and the rate of temperature rise in the final stage has increased, reaching a final value of 98°C, which exceeds the 95°C safety threshold specified in the equipment manual, indicating that the equipment is in a high-temperature risk state; combined with the offline operation and maintenance documents, the possible causes are insufficient condenser heat dissipation or excessive load. It is recommended to reduce the load first and check the cooling circuit."

[0036] S2.2, Semantically segment the offline document data slices and the specialized descriptions, then perform vectorization processing and write them into a vector database; Because the entire instruction manual or work order is too long to be directly searched, semantic segmentation is necessary first. This step can be done by segmenting using semantic units. For example, the sentence "The exhaust temperature of chiller No. 3 has been rising continuously for the past 30 minutes, with an increased rate of temperature rise at the end, reaching a final value of 98℃, exceeding the 95℃ safety threshold specified in the equipment manual, indicating that the equipment is in a high-temperature risk state; combined with offline maintenance documentation, the possible cause is insufficient condenser heat dissipation or excessive load, and it is recommended to prioritize reducing the load and checking the cooling circuit" can be semantically segmented into segment 1: "The exhaust temperature has been rising continuously for the past 30 minutes, with an increased rate of temperature rise at the end", segment 2: "Exceeds the 95℃ safety threshold specified in the equipment manual", segment 3: "The possible cause is insufficient condenser heat dissipation or excessive load", and segment 4: "It is recommended to prioritize reducing the load and checking the cooling circuit". Then, a vector representation is generated for each document segment using a text embedding model and written into a vector database.

[0037] S2.3 Extract entity relationships containing devices, parameters, anomalies, causes, measures, and constraints from the offline document data slices and the specialized descriptions to construct a knowledge graph; In this step, the implicit device-parameter-anomaly-cause-measure-constraint relationships in the offline document data slices and the specialized descriptions are transformed into explicit, reasonable structured relationships. For example, after entity extraction, the following is obtained: Equipment: No. 3 chiller, energy storage battery cluster; Parameters: exhaust temperature, condensing pressure, SOC, current; Abnormalities: High temperature risk, high pressure risk, excessive pressure difference; Causes: Insufficient heat dissipation, excessive load, insufficient cooling water flow; Measures: Reduce load, start equalization, and check the cooling system; Constraints: Temperature ≤ 95℃, SOC ≥ 10%, Pressure difference ≤ 150mV; Then, the relationships between entities are identified and triples are formed. For example, the text "When the exhaust temperature exceeds 95℃, there is a risk of high temperature, which may be caused by insufficient heat dissipation of the condenser. It is recommended to reduce the load" is generated into the triple "High temperature risk, possible cause, insufficient heat dissipation of the condenser". Finally, the triples are written into the graph database to obtain a knowledge graph.

[0038] S2.4 Simultaneously vectorize and extract entity relationships for knowledge units belonging to the same text in the vector database and knowledge graph, and establish bidirectional metadata association between vector database records and knowledge graph nodes, thereby forming a hybrid index knowledge base that combines semantic indexing and structural indexing.

[0039] The core objective of this step is to put the text semantics corresponding to the knowledge units of the same text data source into the vector database, the corresponding entity relations into the knowledge graph, and enable the same knowledge unit to find each other on both sides.

[0040] Specific steps S2.4 include: The text content of the same knowledge unit is simultaneously vectorized and entity relations are extracted, so that the knowledge unit forms a semantic index in the vector database and a structural index in the knowledge graph. The mapping relationship between semantic index and structural index is established through the knowledge unit identifier, thereby constructing a hybrid index knowledge base that combines semantic index and structural index.

[0041] The goal of this step is to ensure that the same knowledge unit can be retrieved in both the vector database and the knowledge graph. An association field is added to the semantic index record in the corresponding vector database to store a list of all graph node identifiers related to that knowledge unit; similarly, an association field is added to the structural index node of the knowledge graph to store a list of all vector record identifiers related to that knowledge unit.

[0042] For example, the specialized description "The exhaust temperature of chiller No. 3 continues to rise to 98℃, exceeding the 95℃ threshold, possibly caused by insufficient condenser heat dissipation; load reduction is recommended." corresponds to vector data ID K001 and knowledge graph data ID V001. Therefore, V001 is added after ID K001 in the vector data, and K001 is added after ID V001 in the knowledge graph data, thus establishing a bidirectional mapping and forming a unified knowledge base combining semantic and structural indexes. This hybrid index knowledge base supports the collaboration of two retrieval modes: during vector retrieval, the identifiers of the associated graph nodes stored in the semantic index records can be used to jump to the knowledge graph to obtain relevant logical relationships, reasoning chains, and constraints for query expansion and result enhancement; during knowledge graph query, the identifiers of the associated vector records stored in the graph nodes can be used to jump to the vector database to obtain the corresponding complete text description for supplementing semantic details and context. Through the aforementioned bidirectional association, the semantic index and structural index of the same knowledge unit form a mutual index, enabling the semantic retrieval results to quickly locate relevant entities and their relationships in the graph, and the graph query results to quickly obtain complete semantic description text.

[0043] Step S4 includes: The RAG engine searches for the closest text fragment in the vector database based on the structured information contained in each subtask, and then expands the text fragment based on the triples in the knowledge graph corresponding to the text fragment. Based on the structured information contained in each subtask, the closest triple is found in the knowledge graph, and then the triple is expanded based on the text fragments in the vector database corresponding to the triple. A cross-encoder is used to score the relevance of the recalled fragments. The fragments are then filtered based on keywords that are not relevant to the subtask, and the highest-scoring search results are selected as high-quality knowledge fragments.

[0044] In this step, the vector database and knowledge graph were linked as described in the previous steps. After finding the closest text fragment in the vector database, relevant content can be extracted based on the ID of the corresponding vector data and its corresponding ID in the knowledge graph, thus expanding the vector data. Similarly, the data retrieved from the knowledge graph is expanded. For example, the subtask text "Reasons for high exhaust temperature of chiller No. 3" is input, generating a query vector. A similarity search is performed in the vector database, recalling multiple semantically similar candidate fragments. One candidate fragment reads, "The exhaust temperature of chiller No. 3 continues to rise to 98℃, exceeding the 95℃ threshold, possibly caused by insufficient condenser heat dissipation; load reduction is recommended." The system performs named entity recognition on this fragment, extracting "insufficient condenser heat dissipation" as a new entity. Subsequently, a query is performed in the knowledge graph using this entity as a condition to obtain other related concepts, such as "insufficient cooling water flow." "Insufficient cooling water flow" is added as an expanded term to the original query, and vector retrieval is performed again to obtain more related fragments, such as "Exhaust temperature also rises when cooling water flow is insufficient." Next, the system obtains a structured list of causes based on the above statements: "Insufficient condenser heat dissipation" (probability 0.75), "Insufficient refrigerant" (probability 0.20), and "Insufficient cooling water flow" (probability 0.15). For each cause name, the system performs a keyword search in the vector database to obtain detailed descriptive text. For example, for "Insufficient condenser heat dissipation," the recall manual excerpt states: "Insufficient condenser heat dissipation is usually caused by fan failure or clogged heat sinks; the fan speed and heat sink cleanliness should be checked."

[0045] Then, a cross-encoder (such as BGE-reranker) is used to accurately score the relevance of each candidate result to the original subtask query. The cross-encoder takes the query and candidate text as input and outputs a relevance score of 0 to 1.

[0046] Next, keyword filtering is performed to remove irrelevant content. For example, in the subtask "Query energy storage charging and discharging strategy," if the candidate text contains words like "maintenance" or "fault," these will be removed since maintenance operations are generally not included in charging and discharging strategies. The remaining content is then ranked, and the top few with the highest relevance scores are retained.

[0047] Further, step S5 includes: S5.1, based on the parameter values ​​and trends in real-time data, dynamically filter high-quality knowledge fragments and retain knowledge fragments that match the current working conditions; S5.2, based on preset security constraints, limit the boundaries of the selected knowledge fragments and eliminate or correct suggestions that violate physical limits, operating procedures or logical constraints; S5.3 assembles the filtered, eliminated, or corrected knowledge fragments with real-time data summaries and security constraints into structured prompt words, inputs them into a large language model, and generates energy equipment control strategies containing executable control instructions.

[0048] In this step, S5.1 first feeds back the changing trends and conditions of the real-time data to the high-quality knowledge fragments. For example, if one of the high-quality knowledge fragments originally described the operating strategy when the temperature continues to rise, but the real-time data shows that the temperature is stabilizing, then this fragment is deleted. Then, S5.2 sets boundary limits for these high-quality knowledge fragments. For example, if one of the operating strategies describing the continuous rise in temperature indicates that power needs to be reduced immediately, but the reduction range is not specified, it needs to be refined. Therefore, based on preset safety constraints, it is modified to reduce the power to 0.3°C.

[0049] Finally, these high-quality knowledge fragments, after being filtered and corrected, along with assembly prompts, are input into a large language model. For example, the system command is input: "You are an energy storage system operation and maintenance expert. Please generate a response strategy based on the following information." The model also provides corresponding equipment information, real-time data, knowledge fragments, safety constraints, etc., ultimately outputting a language description strategy such as "The current temperature rise rate is 2.3℃ / min, which exceeds the emergency threshold of 2℃ / min. The temperature rise must be controlled within 5 minutes, otherwise an emergency shutdown will be triggered."

[0050] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0051] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0052] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0053] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0054] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

Claims

1. An energy management method based on retrieval-enhanced generation and multi-agent collaboration, characterized in that: Includes the following steps: S1. Real-time synchronization of multiple time-series operation data of the managed device; screening out abnormal parameters in the time-series operation data whose change process is monotonous and smooth as candidate parameters; performing piecewise linear fitting on the candidate parameters to obtain several straight line segments; performing circular fitting on the data points between adjacent straight line segments to obtain circular arc segments; adjusting the circular arc segments to form a continuous and smooth candidate curve with the corresponding straight line segments; screening out abnormal parameters in the time-series operation data whose change process is monotonous and smooth as candidate parameters includes: calculating the first-order difference of the time-series operation data respectively; if the difference signs tend to be consistent within a preset time window, then calculating the second-order difference of the corresponding time-series operation data; if the absolute value of the second-order difference is less than a preset second-order difference threshold, then it is determined that the change process is monotonous and smooth; screening out the portion of the time-series operation data whose change process is monotonous and smooth that exceeds the alarm threshold or trend threshold as candidate parameters. S2, obtain offline documentation from the managed device, combine the geometric description of the candidate curve with the offline documentation to convert it into a natural language description, perform semantic segmentation and vectorization on the offline documentation and the natural language description, and construct a hybrid index knowledge base containing a vector database and a knowledge graph; S3 receives user commands or system-triggered events, performs intent recognition based on a large language model, and breaks down the target task into multiple sub-tasks; S4. Based on the structured information contained in each subtask, a search is performed in the hybrid index knowledge base. During vector retrieval, the hybrid index knowledge base is used to expand the query by utilizing concepts related to the query entity in the knowledge graph. During knowledge graph query, the hybrid index knowledge base is used to expand the query by utilizing new entities extracted from the fragments recalled by vector retrieval. The two types of search results are filtered and scored, and the search results with high scores are retained as high-quality knowledge fragments. S5, the high-quality knowledge fragments, real-time data from the managed device, and security constraints are used to generate an energy device control strategy through a large language model.

2. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 1, characterized in that: The optional parameters include some or all of the following: the cell temperature, SOC, cell differential pressure, and charge / discharge rate of the energy storage battery; the exhaust temperature, condensing pressure, chilled water outlet temperature, and current of the chiller unit; the output power of the photovoltaic inverter; the return air temperature of the terminal air conditioner; and the power factor of the power distribution system.

3. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 1, characterized in that: Piecewise linear fitting is performed on the candidate parameters to obtain several straight line segments. Circular fitting is then performed on the data points between adjacent straight line segments to obtain circular arc segments. Adjusting these circular arc segments to form continuous and smooth candidate curves with the corresponding straight line segments includes: The alternative parameters are simplified into polylines connected by several key points, and the length of each polyline is greater than a preset duration, thus obtaining the straight line segment; Least-squares circular arc fitting is performed using transition data points between adjacent straight line segments, and the directions of the preceding and following adjacent straight line segments are used as tangent constraints at the start and end points of the circular arc segment. By adjusting the number of data points involved in the fitting and the range of the data window, a circular arc segment that is continuous and smoothly connected to the adjacent straight line segments is obtained.

4. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 1, characterized in that: Step S2 includes: S2.1, The geometric description of the candidate curve is introduced into the equipment operation rules, parameter constraints and fault mechanisms in the offline document data, and the geometric description is semantically enhanced to generate a professional description that integrates physical meaning and handling suggestions; S2.2, Semantically segment the offline document data slices and the specialized descriptions, then perform vectorization processing and write them into a vector database; S2.3 Extract entity relationships containing devices, parameters, anomalies, causes, measures, and constraints from the offline document data slices and the specialized descriptions to construct a knowledge graph; S2.4 Simultaneously vectorize and extract entity relationships for knowledge units belonging to the same text in the vector database and knowledge graph, and establish bidirectional metadata association between vector database records and knowledge graph nodes, thereby forming a hybrid index knowledge base that combines semantic indexing and structural indexing.

5. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 4, characterized in that: Step S2.1 includes: S2.1.1, The candidate curve is described as a geometric description text including the geometric description by using time-series data textification technology; S2.1.2, retrieve relevant knowledge corresponding to the geometric description text from the offline document data; S2.1.3, the geometric description text and the relevant knowledge are integrated, and the possible causes, handling measures and risk level descriptions of the anomaly are further retrieved from the offline document data and added to generate the professional description.

6. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 4, characterized in that: Step S2.4 includes: The text content of the same knowledge unit is simultaneously vectorized and entity relations are extracted, so that the knowledge unit forms a semantic index in the vector database and a structural index in the knowledge graph. The mapping relationship between semantic index and structural index is established through the knowledge unit identifier, thereby constructing a hybrid index knowledge base that combines semantic index and structural index.

7. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 1, characterized in that: Step S4 includes: The RAG engine searches for the closest text fragment in the vector database based on the structured information contained in each subtask, and then expands the text fragment based on the triples in the knowledge graph corresponding to the text fragment. Based on the structured information contained in each subtask, the closest triple is found in the knowledge graph, and then the triple is expanded based on the text fragments in the vector database corresponding to the triple. A cross-encoder is used to score the relevance of the recalled fragments. The fragments are then filtered based on keywords that are not relevant to the subtask, and the highest-scoring search results are selected as high-quality knowledge fragments.

8. The energy management method based on retrieval enhancement generation and multi-agent collaboration according to claim 1, characterized in that: Step S5 includes: S5.1, based on the parameter values ​​and trends in real-time data, dynamically filter high-quality knowledge fragments and retain knowledge fragments that match the current working conditions; S5.2, based on preset security constraints, limit the boundaries of the selected knowledge fragments and eliminate or correct suggestions that violate physical limits, operating procedures or logical constraints; S5.3 assembles the filtered, eliminated, or corrected knowledge fragments with real-time data summaries and security constraints into structured prompt words, inputs them into a large language model, and generates energy equipment control strategies containing executable control instructions.

Images