AI energy dynamic optimization control system and method based on intelligent agent

Abstract

The invention relates to the technical field of intelligent control of an energy system, and discloses an AI energy dynamic optimization control system and method based on an intelligent agent. The method comprises the following steps: extracting energy network equipment attributes and topology, and constructing a multi-dimensional data cube; and real-time sensing data and a flexible constraint rule are synchronously injected. And driving the intelligent agent to carry out incremental semantic annotation on the data stream, generating an energy event with a space-time mark, and compiling the energy event into a source load feature portrait. And according to the portrait, activating a matched regulation and control mode sub-graph in the strategy knowledge graph, generating a dynamic control parameter, and loading the dynamic control parameter to an actuator. According to the method, real-time deep semantic understanding of the operation state of the energy system and dynamic self-adaptive accurate regulation and control based on historical experience are realized, and the intelligent level and the operation efficiency of complex energy network control are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for energy systems, specifically to an AI-based energy dynamic optimization control system and method based on intelligent agents. Background Technology

[0002] Current dynamic optimization control of energy systems mainly relies on two technical approaches. One is control through monitoring systems combined with preset static rules. The other is to use machine learning models trained on historical data to predict and formulate control strategies.

[0003] Rule-based control methods have fixed logic and cannot understand the complex spatiotemporal relationships and actual semantics of operational states behind the data. Their adaptability is severely insufficient when faced with fluctuations in renewable energy output and random load changes. While data-driven models possess some learning capabilities, their decision-making processes lack interpretability, and the trained model strategies are rigid, making it difficult to flexibly and specifically reuse historically scattered and validated control experiences in different scenarios under new real-time operating conditions. When the system encounters sudden disturbances or complex scenarios with multiple coupled factors, the response accuracy and adaptability of existing methods exhibit significant bottlenecks.

[0004] The intelligent upgrading of energy systems therefore faces two key challenges. First, it requires the immediate and in-depth analysis of massive real-time sensor data streams, not only determining whether data exceeds limits but also identifying complex events characterizing the evolution of the system's operational status. Second, it requires a mechanism that can systematically organize historical control knowledge and, based on a precise characterization of the current system state, quickly match and activate the most relevant historical successful strategy fragments, thereby achieving autonomous evolution of control strategies and experience-driven real-time dynamic optimization. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based energy dynamic optimization control system and method based on intelligent agents to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an AI-based dynamic energy optimization control method based on intelligent agents, the method comprising: An energy system metadata layer is established, in which device attributes and topology associations are extracted from the infrastructure description information of the target energy network; Based on the device attributes and topology associations, a multi-dimensional data cube is constructed. The dimensions of the multi-dimensional data cube include at least the device type dimension, the time granularity dimension, and the physical location dimension. Within the framework of the multi-dimensional data cube, real-time sensor reading streams and a preset set of flexible constraint rules are synchronously injected. Drive an intelligent agent kernel, which performs incremental semantic annotation on the sensor reading stream based on the current state of the multi-dimensional data cube and the flexible constraint rule set, and generates a series of energy events with spatiotemporal markers. The series of energy events with spatiotemporal markers are compiled into an executable source-load feature profile; based on the source-load feature profile, a matching regulation pattern subgraph is activated in the strategy knowledge graph, which stores historical regulation experience. Dynamic control parameters are generated based on the activated control mode subgraph, and the dynamic control parameters are loaded into the underlying actuator of the target energy network.

[0007] Preferably, constructing the multi-dimensional data cube includes: parsing the device attributes and topology associations to identify all controllable units and sensing units in the target energy network; defining the device type, observable physical location, and minimum and maximum allowed data reporting intervals for each controllable unit and sensing unit; initializing a sparse cube data structure with the device type as the first axis, different time granularities from milliseconds to hours as the second axis, and the physical location as the third axis; and pre-allocating a buffer queue for each unit of the cube data structure to accommodate the subsequently injected sensor reading streams.

[0008] Preferably, the synchronous injection of the real-time sensor reading stream and the preset flexible constraint rule set includes: continuously monitoring the raw data packets from the sensor unit, decoding and timestamping the raw data packets; pushing the values ​​of the decoded data into the buffer queue of the corresponding unit in the cube data structure according to the device type and physical location of the data; simultaneously, reading flexible constraint rule sets for different seasons and weather templates from the policy library, wherein the flexible constraint rule sets define boundary conditions and operation suggestions in an if-then logical form; converting the flexible constraint rule sets into a rule tree structure, and anchoring the root node of the rule tree structure to the device type dimension of the cube data structure.

[0009] Preferably, incremental semantic annotation of the sensor reading stream includes: the agent kernel periodically scanning the latest data of each unit buffer queue in the cube data structure; matching the latest data with the rule tree structure, and triggering the annotation of the latest data that satisfies a certain if condition branch when a certain if condition branch is met; the annotation content includes at least the pattern category to which the data belongs, the degree of deviation from the baseline, and the associated potential operation, forming a triplet annotation information; binding the triplet annotation information with the timestamp, device type, and physical location corresponding to the annotated data, and packaging it into an energy event object.

[0010] Preferably, compiling the executable source-load feature profile includes: collecting all energy event objects generated within a time window; clustering the energy event objects according to the physical location dimension to form spatial event clusters; within each spatial event cluster, performing sub-clustering analysis according to the device type dimension and the continuity of time occurrence; extracting the core features of each sub-cluster, the core features including the statistical distribution of event density, dominant mode category, and operation suggestions; encapsulating the core features of each sub-cluster into a profile unit, and the collection of all profile units constitutes the source-load feature profile.

[0011] Preferably, activating the control mode subgraph includes: the strategy knowledge graph consists of nodes and edges, where nodes represent different system states or operation instructions, and edges represent the transition relationships between states or the sequential logic of instructions; semantic similarity calculation is performed between each portrait unit in the source load feature portrait and the state nodes in the strategy knowledge graph; state nodes with similarity exceeding a threshold are selected as matching nodes, and all associated nodes and edges of these matching nodes within three hops in the strategy knowledge graph are extracted; based on the extracted nodes and edges, a temporary control mode subgraph for the current context is constructed.

[0012] Preferably, generating the dynamic control parameters includes: traversing the control mode subgraph and converting it into a partially ordered sequence of operation instructions; specifying the abstract instructions in the sequence of operation instructions into control commands with explicit set values ​​based on the real-time status parameters of the corresponding devices in the multi-dimensional data cube; scheduling the execution time of the control commands based on conflict detection to ensure that no conflicting device operations occur at the same physical location and at the same time; and outputting the final set of dynamic control parameters after scheduling, where each parameter includes a device identifier, control action, set value, and execution time.

[0013] Preferably, loading the dynamic control parameters to the underlying actuator includes: placing the set of dynamic control parameters into a priority distribution queue according to the order of execution time; establishing a security verification gateway, which requires secondary confirmation from the security verification gateway when parameters are retrieved from the head of the priority distribution queue for distribution; the security verification gateway instantaneously queries the latest status of the target device and compares it with the expected pre-state in the control parameters, allowing distribution only when they match; the distributed parameters are encapsulated into a device-recognizable instruction protocol and sent to the corresponding underlying actuator via a communication network.

[0014] Preferably, the method further includes: after the underlying executor executes the control command, it generates an execution feedback message; the intelligent agent kernel receives the execution feedback message and parses out the actual execution effect data from it; the actual execution effect data is injected back into the corresponding unit buffer queue of the multi-dimensional data cube as a new energy event; this process forms a closed loop, which is used to trigger a new round of annotation, profile generation and strategy adjustment.

[0015] Preferably, when the processor executes the computer program, it implements the steps of the AI ​​energy dynamic optimization control method based on the intelligent agent as described in any of the above-mentioned methods.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Incremental semantic annotation of real-time sensor data streams transforms continuous numerical readings into energy events with clear type, time, and location markers. This is accomplished within a multi-dimensional data cube framework, imbuing the raw data with rich contextual semantic information. Traditional methods can only generate isolated out-of-limit alarms, while this technology can identify complex event sequences characterizing the evolution of system operating states. The data is thus transformed into structured knowledge that can be directly understood and reasoned about by machines.

[0017] By compiling and generating source-load feature profiles, matching control pattern subgraphs are activated within a pre-built policy knowledge graph. Essentially, this involves rapidly matching and dynamically associating a precise characterization of the current system state with historically validated and effective control experiences. System decisions no longer rely on a single fixed model or rule base, but can flexibly invoke and combine successful past policy fragments based on real-time scenarios. This enables control strategies to adaptively adjust based on scenario similarity, allowing for more agile and precise generation of control commands when facing new, complex operating conditions or sudden disturbances, thus improving the timeliness and targeting of control. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based energy dynamic optimization control method based on intelligent agents described in this invention. Figure 2 A flowchart for constructing a multi-dimensional data cube; Figure 3 A flowchart for synchronously injecting sensor reading streams and flexible constraint rule sets; Figure 4 A diagram illustrating the multi-time granularity status monitoring characteristics of the Alpha transformer in a substation. Figure 5 To control the distribution of instruction execution states. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides an AI-based dynamic energy optimization control method based on intelligent agents. The method includes: First, establishing an energy system metadata layer, in which device attributes and topology associations are extracted from the infrastructure description information of the target energy network; then, based on the extracted device attributes and topology associations, constructing a multi-dimensional data cube, the dimensions of which include at least device type dimension, time granularity dimension, and physical location dimension; within the framework of this multi-dimensional data cube, synchronously injecting real-time sensor reading streams and a preset set of flexible constraint rules; next, driving an intelligent agent kernel, which performs incremental semantic annotation on the continuously flowing sensor reading streams according to the current state of the multi-dimensional data cube and the set of flexible constraint rules, thereby generating a series of energy events with spatiotemporal tags; further, compiling this series of energy events into an executable source-load feature profile; based on the source-load feature profile, activating a matching control mode subgraph in a pre-constructed policy knowledge graph, the policy knowledge graph storing historical control experience; finally, generating dynamic control parameters based on the activated control mode subgraph, and loading these parameters into the underlying actuators of the target energy network to complete closed-loop control.

[0021] Example 1: See Figure 2 Constructing a multi-dimensional data cube includes the following steps: parsing the device attributes and topology associations obtained from the energy system metadata layer to identify all controllable units and sensing units in the target energy network; defining the device type, observable physical location, and minimum and maximum allowed data reporting intervals for each controllable unit and sensing unit; initializing a sparse cube data structure with device type as the first axis, different time granularities from milliseconds to hours as the second axis, and physical location as the third axis; and pre-allocating a buffer queue for each unit of the cube data structure to accommodate subsequent injected sensor reading streams.

[0022] In practical implementation, parsing device attributes and topology associations aims to clearly identify all controllable and state-aware units constituting the target energy network from structured descriptive information. In some embodiments, controllable units refer to devices capable of receiving commands and changing their operating state, such as circuit breakers, adjustable transformers, and energy storage converters; sensing units include voltage transformers, current transformers, power sensors, and temperature sensors. For each identified controllable and sensing unit, a set of metadata for constructing a data cube is defined. The defined metadata includes the device type to which the unit belongs, its observable location coordinates or area code in physical space, and the minimum and maximum time intervals allowed for the unit when reporting data.

[0023] The first dimension axis is the defined set of device types, encompassing all device categories in the network. The second dimension axis uses a pre-defined time granularity ranging from milliseconds, seconds, minutes to hours. The third dimension axis uses the set of physical location identifiers for all units. Based on these three dimensions, a sparse multidimensional array is initialized as the core cube data structure. It can be understood that during initialization, most units within this cube data structure do not store valid data, reserving logical space only for subsequent data injection. After the cube data structure initialization is complete, an independent buffer queue is pre-allocated to each logical unit in the data structure. This buffer queue is a data structure used to temporarily store subsequently injected sensor reading stream data units in a first-in, first-out (FIFO) order.

[0024] In some embodiments, the physical location dimension can be characterized using a global geographic coordinate system, spatial coding in a building information model, or node numbers in a power system wiring diagram. The temporal granularity dimension is determined based on the timeliness requirements of the specific control scenario; high-frequency control scenarios are associated with millisecond-level granularity, while scenarios used for trend analysis are associated with hourly-level granularity. In a specific implementation, the cube data structure is implemented in memory as key-value pairs, where the key is a composite key composed of a device type identifier, a time granularity identifier, and a physical location identifier, and the value is a corresponding buffer queue reference pointer. Optionally, to improve access efficiency, a multi-level index structure can be established, firstly by performing a first-level hash partition according to physical location, and then creating a second-level index within each partition according to device type.

[0025] Example 2: See Figure 3The synchronous injection of real-time sensor reading streams and preset flexible constraint rule sets is achieved through the following process: continuously listening to raw data packets from sensor units, decoding and timestamping these raw data packets; pushing the values ​​of the decoded data into the buffer queue of the corresponding cell in the cube data structure according to the device type and physical location; simultaneously, reading flexible constraint rule sets for different seasons and weather templates from the policy library, the flexible constraint rule sets defining boundary conditions and operation suggestions in an if-then logical form; converting the flexible constraint rule sets into a rule tree structure, and anchoring the root node of the rule tree structure to the device type dimension of the cube data structure. The process of incremental semantic annotation of sensor reading streams is as follows: The agent kernel periodically scans the latest data in the buffer queues of each cell in the cube data structure; the latest data is matched with the rule tree structure, and when a certain if condition branch is met, the annotation of the latest data that meets the condition branch is triggered; the annotation content includes at least the pattern category to which the data belongs, the degree of deviation from the baseline, and the associated potential operation, forming a triplet annotation information; the triplet annotation information is bound to the timestamp, device type, and physical location corresponding to the annotated data, and packaged into an energy event object.

[0026] In practice, the synchronization injection process begins with continuous monitoring of data packets from physical sensing units. Specifically, data acquisition services deployed at the edge or central server continuously monitor raw data packets sent by sensing units via message queues or dedicated communication protocols. The raw data packets are then decoded, converting binary or text-formatted packets into structured data objects containing device identifiers, measurement types, numerical values, and timestamps according to predefined communication protocols. These structured data objects are then timestamped, calibrating the time information within the packets to a system clock based on Coordinated Universal Time (UTC). Based on the device identifier carried in the decoded data object, the corresponding device type and physical location coordinates are mapped. Using these coordinates, the numerical portion of the data object is pushed into the end of the buffer queue of the corresponding logical unit within the cube data structure. In some embodiments, the original message reported by the sensing unit is "ID:SENSOR_001,V:235.6,T:1625097600123". After decoding, the device identifier "SENSOR_001" is obtained. By querying the metadata, it is determined that the device type of "SENSOR_001" is "voltage sensor" and the physical location is "bus A". Then, the value 235.6 is pushed into the buffer queue of the logic unit corresponding to the cube data structure [device type = "voltage sensor", time granularity = "second", physical location = "bus A"].

[0027] Simultaneously, another parallel process involves reading a pre-defined set of flexible constraint rules from a persistently stored policy library. This set of rules is a logical collection predefined for different operating scenarios; for example, the rule set for a high-temperature summer scenario differs from that for a low-temperature winter scenario, and the rule set for a high-power photovoltaic scenario on a sunny day differs from that for a rainy day scenario. The flexible constraint rule set defines the boundary conditions and suggested operations for system operation in an "if-then" logical form. A complete rule is presented in the form of "if {conditional expression}, then {operation suggestion}". These textual or structured rule entries are then converted into an internal data representation called a rule tree structure. A rule tree structure is a tree-like data structure where the root node represents a top-level device type or rule category, child nodes branching from the root node represent different conditional judgment branches, and leaf nodes are associated with specific operation suggestions. The conversion process can be understood as including syntax parsing, conditional expression compilation, and tree structure construction. Anchoring the root node of the rule tree structure to the device type dimension of the cube data structure means that each type of device has its associated rule subtree. When processing data of this type of device, the corresponding rule subtree will be quickly located and loaded for matching calculation.

[0028] After the cube data structure continuously receives sensor readings and the rule tree structure is loaded, the agent kernel begins to execute incremental semantic annotation. The agent kernel periodically scans the head or the latest data record of each logical unit's buffer queue in the cube data structure at fixed time intervals or in an event-driven manner. The latest data obtained from the scan is matched against the loaded rule tree structure. The matching process starts from the root node of the rule tree structure and evaluates whether the data satisfies the conditional expression of the current node level by level. When a conditional branch is reached and the latest data satisfies all the "if" conditions of this branch, a semantic annotation operation is triggered for that latest data. The annotation operation generates at least the pattern category to which the data belongs, the degree of deviation from the baseline, and the associated potential operation. These contents together constitute a triplet annotation information. It can be understood that the pattern category is used to qualitatively describe the system state represented by the data, such as "overvoltage," "load surge," and "normal fluctuation"; the degree of deviation from the baseline is used to quantify the magnitude of data anomalies or significant features; and the associated potential operation is directly derived from the content of the "then" part of the rule. In some embodiments, for a voltage sensor value of 235.6 volts, the rule "If voltage > 233.0 volts and duration > 5 seconds, then mode category = 'overvoltage', reactive power compensation is recommended" is matched. Assuming this data has consistently met the conditions, the generated triplet annotation information is: {Mode category: "overvoltage", deviation from baseline:} The associated potential operation is "Initiating reactive power compensation". This includes the degree of deviation from the baseline. It can be calculated using the formula:

[0029] Where: symbol The per-unit value represents the calculated deviation, with the symbol... Represents the latest voltage measurement obtained from the buffer queue, symbol This represents a preset voltage reference value, such as a rated voltage of 230 volts. Finally, this triplet annotation information is bound to the timestamp, device type, and physical location carried by the annotated raw data, encapsulating it into an energy event object with a unified structure. The energy event object is then output to the downstream processing flow. Optionally, the binding process assigns a unique event identifier to the energy event object and stores the tuple {timestamp, device type, physical location, triplet annotation information} as the event content.

[0030] Example 3: The steps for compiling an executable source-load feature profile include: collecting all energy event objects generated within a time window; clustering the energy event objects according to the physical location dimension to form spatial event clusters; within each spatial event cluster, performing sub-clustering analysis according to the device type dimension and the continuity of time occurrence; extracting the core features of each sub-cluster, including event density, dominant mode category, and statistical distribution of operation suggestions; encapsulating the core features of each sub-cluster into a profile unit, and the set of all profile units constitutes the source-load feature profile. The process of activating the control mode subgraph is as follows: the strategy knowledge graph consists of nodes and edges, where nodes represent different system states or operation instructions, and edges represent the transition relationship between states or the sequential logic of instructions; calculating the semantic similarity between each profile unit in the source-load feature profile and the state nodes in the strategy knowledge graph; selecting state nodes with similarity exceeding a threshold as matching nodes, and extracting all associated nodes and edges of these matching nodes within three hops in the strategy knowledge graph; and constructing a temporary control mode subgraph for the current context based on the extracted nodes and edges.

[0031] In practice, the process of compiling an executable source-load feature profile begins with collecting all energy event objects generated by the agent kernel within a preset time window. The time window length can be configured to different scales, such as five minutes or fifteen minutes, to balance the real-time nature of control with the macro-level nature of decision-making. The collected energy event objects are then clustered according to their physical location, grouping events occurring under the same or adjacent physical location identifiers to form multiple spatial event clusters. Within each spatial event cluster, further sub-clustering analysis is performed according to device type and temporal continuity, aggregating events of the same type of device that occur consecutively or nearby in the time series into a sub-cluster. The core features of each sub-cluster are extracted, including the event density per unit time, the statistically dominant pattern category distribution, and the statistical distribution of operational suggestions carried by all events. Each sub-cluster's core features are encapsulated into an independent profile unit, and the collection of all profile units constitutes the source-load feature profile representing the current state and characteristics of the energy network.

[0032] The process of activating the regulation mode subgraph relies on a pre-generated and stored policy knowledge graph. This graph consists of nodes and edges. Nodes represent different system operating states or abstract operational instructions, while edges represent transition relationships between states or sequential execution logic between instructions. Semantic similarity is calculated between each portrait unit in the source payload feature profile and the state nodes in the policy knowledge graph. This calculation compares the spatial distance between the core feature vector of the portrait unit and the feature vector associated with the state node. In some embodiments, cosine similarity or embedding vector similarity based on a pre-trained language model can be used. The feature vector of the portrait unit is composed of its event density, the encoding of the dominant mode category, and the statistics of the operation suggestion distribution. Policy knowledge graph state nodes whose semantic similarity exceeds a preset threshold are selected as matching nodes. Starting from these matching nodes, all associated nodes in the policy knowledge graph within three hops of the matching nodes and the edges connecting these nodes are extracted. A three-hop distance means that in the graph structure of the knowledge graph, all nodes reachable from the matching node via a maximum of three edges are within the extraction range. Based on the extracted nodes and edges, a temporary subgraph of regulation patterns is constructed specifically for the context described by the current source load feature profile.

[0033] In some embodiments, for a profile unit describing "continuous overvoltage events and load fluctuation events in bus A area during the afternoon," its core feature vector has high similarity to a state node named "regional voltage spikes accompanied by load disturbances" in the strategy knowledge graph, and this state node is therefore selected as a matching node. Starting from this matching node, adjacent operation instruction nodes such as "starting parallel reactors," "adjusting transformer taps," and "checking capacitor bank status" are extracted from the knowledge graph, along with edges connecting these nodes such as "preconditions," "execution order," and "mutual exclusion relationships." These elements together constitute a subgraph of the control mode for handling regional voltage spikes. It can be understood that semantic similarity calculation is a crucial bridge connecting real-time generated profiles and historical experience knowledge; its accuracy directly affects the applicability of the activated subgraph. Similarity calculation can be performed using the following formula:

[0034] Where: symbol The symbol represents the calculated semantic similarity value. Represents the core feature vector extracted and normalized from the image unit, symbol The feature vectors pre-stored in the state nodes of the strategy knowledge graph represent the features of the state nodes. The maximum of the norms of vector P and vector S is used as a normalization factor in this formula to scale the distance metric to a relative range. In this way, source load feature profiles are effectively mapped to specific regions of the policy knowledge graph, thereby activating the most relevant historical regulatory experience fragments and forming a regulatory pattern subgraph that can guide current control.

[0035] Example 4: The process of generating dynamic control parameters is as follows: traverse the control mode subgraph and transform it into a partially ordered sequence of operation instructions; concretize the abstract instructions in the sequence of operation instructions into control commands with explicit set values ​​based on the real-time status parameters of the corresponding devices in the multi-dimensional data cube; schedule the execution times of these control commands based on conflict detection to ensure that no conflicting device operations occur at the same physical location and at the same time; after scheduling, output the final set of dynamic control parameters, each parameter containing device identifier, control action, set value, and execution time.

[0036] In practical implementation, the control mode subgraph is traversed according to the dependency and order relationships defined by the edges in the subgraph. A graph traversal algorithm is used to visit each operation instruction node and state node in the subgraph, and the access path is transformed into a partially ordered sequence of operation instructions. In some embodiments, the control mode subgraph includes "reduce bus A voltage" as the target node, and is connected to operation instruction nodes such as "adjust transformer T1 tap", "connect capacitor bank C1", and "reduce generator G1 reactive power output". The edges between the nodes indicate that "adjust transformer T1 tap" must be executed after "check transformer T1 current tap", while "connect capacitor bank C1" and "reduce generator G1 reactive power output" can be executed in parallel. The operation instruction sequence generated after traversal reflects this partially ordered relationship.

[0037] Transforming abstract instructions in an operational instruction sequence into control commands with specific setpoints requires querying real-time status parameters in a multi-dimensional data cube. In practice, for an abstract instruction like "adjust transformer T1 tap," the agent kernel locates the logical unit in the cube data structure corresponding to the device type "on-load tap-changing transformer," the physical location "substation Alpha," and the time granularity "seconds." It then retrieves the latest tap position value, winding temperature, and load rate status parameters of that transformer from its buffer queue. Based on these real-time status parameters and the objective of the abstract instruction, the specific control setpoint is calculated. For example, if the objective is to reduce the voltage by 2%, and the current tap is 5, with each tap changing the voltage by 0.625%, then the concretized control command would be "adjust transformer T1 tap from 5 to 2."

[0038] Scheduling the execution times of all control commands generated after specification based on conflict detection is a necessary step. Conflict detection aims to identify and resolve conflicting equipment operations occurring at the same physical location and time point, such as simultaneously sending "open" and "close" commands to the same circuit breaker, or issuing conflicting voltage regulation commands to equipment associated with the same busbar. The scheduling algorithm analyzes the equipment identifier, physical location, action type, and expected execution time of all control commands to construct a time-resource occupancy model. In specific implementation, the specific method for constructing the time-resource occupancy model is as follows: the model defines each physical location and each controllable equipment identifier under that location as an independent logical resource, and maintains a timeline extending from the current moment into the future for each logical resource. The model traverses the set of control commands to be scheduled, maps each command to the corresponding logical resource based on its equipment identifier and physical location, and marks the expected execution time and estimated duration of the command as a time occupancy interval on the timeline of that resource. Simultaneously, based on predefined rules in the strategy knowledge graph or the physical characteristics of the equipment, the model identifies the mutual exclusion relationships between different command action types. For example, "open" and "close" commands for the same circuit breaker are absolutely mutually exclusive in time, or multiple voltage regulation commands associated with the same bus segment are considered conflicting within a short period. By traversing the timelines of all resources and checking the overlap and mutual exclusion relationships of time occupancy intervals, the model constructs a global, visualized resource occupancy status graph. Optionally, the scheduling algorithm can employ a priority-based time window offset method to reallocate execution time points for commands with resource conflicts, ensuring the safety and logical consistency of system operation. After scheduling, the final set of dynamic control parameters is output. Each parameter entry in the dynamic control parameter set contains a clear equipment identifier, the control action to be executed, the specific set value, and the execution time point determined by scheduling.

[0039] In some embodiments, the initial set of dynamic control parameters generated from the control mode subgraph contains commands that overlap in time or conflict logically. Through conflict detection and scheduling, a final set of parameters that can be safely issued is generated. Refer to Table 1 for a simplified example of before-and-after scheduling comparison.

[0040] Table 1: Example Table Comparing Control Command Scheduling Before and After

[0041] It is understandable that the original execution time of the command for "Capacitor Bank C1" in Table 1 is the same as that of the command for "Transformer T1". However, the scheduling algorithm detected that the system transient process during the adjustment of "Transformer T1" is not suitable for immediate switching of capacitors, so it is postponed to T+2 seconds. The disconnection command for "Circuit Breaker CB-A" is scheduled after all regulatory commands have been executed to ensure the continuity of power supply during the adjustment process. The scheduling decision is based on a priority evaluation function, which can be calculated based on the formula:

[0042] Where: symbol The scheduling priority score representing the command, symbol and These are weighting coefficients, with the following symbol: The symbol represents the urgency quantification value of the command. This represents the criticality quantification value of the devices affected by the command. The scheduling algorithm scores based on priority. The system determines conflict relationships, allocates appropriate execution time windows for commands, and ultimately outputs a set of dynamic control parameters that has undergone time sorting and logical verification, making it a final instruction set that can be directly issued to the executor.

[0043] See Figure 4 In the multi-time-granularity status monitoring of transformers in substation Alpha, the dynamic correlation between voltage, temperature, and load rate is presented based on sensor data at different time granularities. Specifically, each state variable is presented in millisecond, second, minute, and hourly time dimensions: voltage (kV) shows a linear decreasing trend from a high of 10.20 in the millisecond timescale to 9.80 in the hourly timescale; temperature (°C) gradually increases from 65 in the millisecond timescale to 70 in the minute timescale, then falls back to 69 in the hourly timescale; load rate (%) rapidly increases from 75 in the millisecond timescale to 82 in the minute timescale, then falls back to 78 in the hourly timescale. The coupling characteristics of these three variables are as follows: at the minute-level time granularity, the decrease in voltage and the peak values ​​of temperature and load rate form a synchronous inflection point, reflecting the high-load operating state of the transformer at this time scale. During parameter presentation, the fineness of the time granularity directly affects the sensitivity of state variable fluctuations. For example, at the millisecond level, each state variable is at its initial baseline level, while at the minute level, it corresponds to the high dynamic range of system operation.

[0044] Example 5: Loading dynamic control parameters to the underlying actuator includes: placing the set of dynamic control parameters into a priority distribution queue according to the order of execution time; establishing a security verification gateway, which confirms parameters twice when they are retrieved from the head of the priority distribution queue; the security verification gateway instantly queries the latest status of the target device and compares it with the expected pre-state in the control parameters, allowing distribution only if they match; the distributed parameters are encapsulated into a device-recognizable instruction protocol and sent to the corresponding underlying actuator via the communication network. After executing the control command, the underlying actuator generates an execution feedback message; the intelligent agent kernel receives the execution feedback message and parses the actual execution effect data from it; the actual execution effect data is injected back into the corresponding unit buffer queue of the multi-dimensional data cube as a new energy event; this process forms a closed loop, used to trigger a new round of annotation, profile generation, and strategy adjustment.

[0045] In practical implementation, the process of loading dynamic control parameters to the underlying actuator first sorts each parameter entry in the dynamic control parameter set according to the order of their execution time. The sorted parameter entries are then placed sequentially into a data structure called a priority distribution queue. This queue ensures that at any given time, the parameter entry with the earliest execution time is always at the head of the queue, waiting to be retrieved. An independent security verification gateway module is established. When a parameter entry is retrieved from the head of the priority distribution queue and is ready to be sent to the underlying actuator, it must undergo secondary confirmation by the security verification gateway. The security verification gateway instantly queries the latest state of the target device at the current moment. This query is performed by directly accessing the buffer queue of the corresponding device logic unit in the multi-dimensional data cube to obtain the latest sensor reading and comparing the latest state with the expected preceding state recorded in the control parameter entry. Only when the latest state obtained in real-time is completely consistent with the expected preceding state in the control parameter entry will the security verification gateway allow the control parameter entry to continue the distribution process. If the states are inconsistent, the security verification gateway intercepts the parameter entry and triggers an exception handling process, while simultaneously sending an alarm event to the agent kernel. The issued parameter entries are encapsulated into specific instruction protocols that the underlying actuators can recognize. The instruction protocols follow the communication protocols specified by the equipment manufacturer or industry standards. The encapsulated instruction messages are sent to the corresponding underlying actuators through wired or wireless communication networks.

[0046] In some embodiments, a control parameter entry to be issued contains the following content: {Device Identifier: "Circuit Breaker CB-A", Control Action: "Closed", Setpoint: N / A, Execution Time: T+5 seconds}. When the system time reaches T+5 seconds, this entry is retrieved from the head of the priority distribution queue and sent to the security verification gateway. The security verification gateway immediately queries the current status of the device "Circuit Breaker CB-A" and obtains its latest status from the cube data structure as "Open" with the adjacent current sensor reading as "0A". This is consistent with the expected preceding state of the control parameter entry, "Open and No Current", thus the verification passes. The verified parameter entry is then encapsulated into a "Write Coil" command frame conforming to the Modbus protocol and sent via industrial Ethernet to the "Circuit Breaker CB-A" intelligent controller installed in the substation. It can be understood that the introduction of the security verification gateway is a key security protection mechanism, and its core logic can be formally expressed by a binary decision function, as shown in the formula:

[0047] Where: symbol This represents the output decision result of the security verification gateway. A value of 1 indicates that the transmission is allowed, and a value of 0 indicates that the transmission is rejected. Represents the actual state vector of the target device obtained through instantaneous query; symbol This represents the expected preceding state vector carried in the control parameter entry. The symbol " "" indicates that the two state vectors are perfectly matched across all predefined key dimensions.

[0048] After receiving and executing control commands, the underlying actuator generates an execution feedback message containing execution result information. This feedback message is typically transmitted back to the central control system via the same communication link. The agent kernel receives and parses the execution feedback message, extracting the actual execution effect data. This data includes whether the command was successfully received, executed, the new device state after execution, and related measurement values. In some embodiments, after the circuit breaker "CB-A" executes a "close" command, the feedback message transmitted by its intelligent controller includes "Command executed, current state: closed, measured current: 150A". The agent kernel treats this actual execution effect data as a new, special energy event object, which carries the "execution feedback" mode category and related state data. Subsequently, the agent kernel injects this energy event object back into the buffer queue of the corresponding logical unit in the multi-dimensional data cube. Specifically, it maps the device identifier "Circuit Breaker CB-A" to its device type and physical location, and then pushes the state data "Closed" and current data "150A" as part of the sensor reading stream into the buffer queue of the corresponding unit in the cube data structure. Optionally, the back-injected data will have an event source tag to distinguish it from ordinary sensor readings. This back-injection process allows the actual control effect to re-enter the multi-dimensional data cube framework, thus forming a complete closed loop from perception, decision-making, execution to re-perception. It can be understood that the injection of feedback data will trigger a new round of incremental semantic annotation by the agent kernel. Based on the comparison between the actual effect and the expected effect, new annotations such as "control successful" or "control deviation" will be generated. The energy events generated by these new annotations will participate in compiling the source-load feature profile of the next time window, thereby activating different regulation mode subgraphs or fine-tuning strategy knowledge graphs, realizing the self-adjustment and continuous improvement of the dynamic optimization control method.

[0049] See Figure 5 In an AI-based energy dynamic optimization control system based on intelligent agents, the distribution of states (success, failure, and timeout) of different underlying actuators (VF-1 inverter, CT-1 / CT-2 contactors, and CB-A / CB-B circuit breakers) during the control command execution phase is presented. Specifically, the command execution of each device is predominantly in the "success" state: CT-1 and CB-A each have 12 successful commands, CT-2 has 11 successful commands, and VF-1 inverter and CB-B each have 9 successful commands. The proportion of failure and timeout states is extremely low, with only VF-1 showing 1 timeout and 2 failure commands, CT-2 showing 1 failure command, and CB-B showing 3 failure commands. This distribution result intuitively reflects the reliability of the current control parameter issuance and execution process, and also provides a quantitative basis for the actual execution effect of optimizing the control mode in the strategy knowledge graph.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based dynamic energy optimization control method based on intelligent agents, characterized in that, The method includes: An energy system metadata layer is established, in which device attributes and topology associations are extracted from the infrastructure description information of the target energy network; Based on the device attributes and topology associations, a multi-dimensional data cube is constructed. The dimensions of the multi-dimensional data cube include at least the device type dimension, the time granularity dimension, and the physical location dimension. Within the framework of the multi-dimensional data cube, real-time sensor reading streams and a preset set of flexible constraint rules are synchronously injected. Drive an intelligent agent kernel, which performs incremental semantic annotation on the sensor reading stream based on the current state of the multi-dimensional data cube and the flexible constraint rule set, and generates a series of energy events with spatiotemporal tags. The series of energy events with spatiotemporal markers are compiled into an executable source-load feature profile; based on the source-load feature profile, a matching regulation pattern subgraph is activated in the strategy knowledge graph, which stores historical regulation experience. Dynamic control parameters are generated based on the activated control mode subgraph, and the dynamic control parameters are loaded into the underlying actuator of the target energy network.

2. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 1, characterized in that, Constructing the multi-dimensional data cube includes: parsing the device attributes and topology associations to identify all controllable units and sensing units in the target energy network; defining the device type, observable physical location, and minimum and maximum allowed data reporting intervals for each controllable unit and sensing unit; initializing a sparse cube data structure with the device type as the first axis, different time granularities from milliseconds to hours as the second axis, and the physical location as the third axis; and pre-allocating a buffer queue for each unit of the cube data structure to accommodate the subsequently injected sensor reading streams.

3. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 2, characterized in that, The synchronous injection of the real-time sensor reading stream and the preset flexible constraint rule set includes: continuously monitoring the raw data packets from the sensor unit, decoding and timestamping the raw data packets; pushing the values ​​of the decoded data into the buffer queue of the corresponding unit in the cube data structure according to the device type and physical location of the data; simultaneously reading flexible constraint rule sets for different seasons and weather templates from the policy library, wherein the flexible constraint rule sets define boundary conditions and operation suggestions in an if-then logical form; converting the flexible constraint rule sets into a rule tree structure, and anchoring the root node of the rule tree structure to the device type dimension of the cube data structure.

4. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 3, characterized in that, Incremental semantic annotation of the sensor reading stream includes: the agent kernel periodically scanning the latest data in the buffer queues of each unit in the cube data structure; matching the latest data with the rule tree structure, and triggering the annotation of the latest data that satisfies a certain if condition branch when a certain if condition branch is met; the annotation content includes at least the pattern category to which the data belongs, the degree of deviation from the baseline, and the associated potential operation, forming a triplet annotation information; binding the triplet annotation information with the timestamp, device type, and physical location corresponding to the annotated data, and packaging it into an energy event object.

5. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 4, characterized in that, Compiling the executable source-load feature profile includes: collecting all energy event objects generated within a time window; clustering the energy event objects according to the physical location dimension to form spatial event clusters; within each spatial event cluster, performing sub-clustering analysis according to the device type dimension and the continuity of time occurrence; extracting the core features of each sub-cluster, the core features including the statistical distribution of event density, dominant mode category, and operation suggestions; encapsulating the core features of each sub-cluster into a profile unit, and the collection of all profile units constitutes the source-load feature profile.

6. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 5, characterized in that, Activating the control mode subgraph includes: the strategy knowledge graph consists of nodes and edges, where nodes represent different system states or operation instructions, and edges represent the transition relationships between states or the sequential logic of instructions; semantic similarity is calculated between each portrait unit in the source load feature portrait and the state nodes in the strategy knowledge graph; state nodes with similarity exceeding a threshold are selected as matching nodes, and all associated nodes and edges of these matching nodes within three hops in the strategy knowledge graph are extracted; based on the extracted nodes and edges, a temporary control mode subgraph for the current context is constructed.

7. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 6, characterized in that, Generating the dynamic control parameters includes: traversing the control mode subgraph and converting it into a partially ordered sequence of operation instructions; specifying the abstract instructions in the sequence of operation instructions into control commands with explicit set values ​​based on the real-time status parameters of the corresponding devices in the multi-dimensional data cube; scheduling the execution time of the control commands based on conflict detection to ensure that no conflicting device operations occur at the same physical location and at the same time; and outputting the final set of dynamic control parameters after scheduling, where each parameter includes a device identifier, control action, set value, and execution time.

8. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 7, characterized in that, Loading the dynamic control parameters to the underlying actuator includes: placing the set of dynamic control parameters into a priority distribution queue according to the order of execution time; establishing a security verification gateway, which requires secondary confirmation from the security verification gateway when parameters are retrieved from the head of the priority distribution queue for distribution; the security verification gateway instantaneously queries the latest status of the target device and compares it with the expected pre-state in the control parameters, allowing distribution only if they match; the distributed parameters are encapsulated into a device-recognizable instruction protocol and sent to the corresponding underlying actuator via a communication network.

9. The AI-based dynamic energy optimization control method based on intelligent agents as described in claim 1, characterized in that, The method further includes: after the underlying executor executes the control command, it generates an execution feedback message; the intelligent agent kernel receives the execution feedback message and parses out the actual execution effect data from it; the actual execution effect data is injected back into the corresponding unit buffer queue of the multi-dimensional data cube as a new energy event; this process forms a closed loop, which is used to trigger a new round of annotation, profile generation and strategy adjustment.

10. An AI-based energy dynamic optimization control system based on intelligent agents, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the AI ​​energy dynamic optimization control method based on an intelligent agent as described in any one of claims 1 to 9.

Images