Power distribution network regulation method based on multi-modal intention perception and light-weight large model

Abstract

The application discloses a power distribution network regulation and control method based on multi-modal intention perception and light-weight large model, and belongs to the technical field of power system operation and control. The method comprises the following steps: obtaining multi-modal power consumption intention information submitted by a user, outputting a unified intention feature vector through a multi-modal fusion encoder, inputting the unified intention feature vector into a light-weight large language model deployed on an edge side, and automatically generating a structured load model representation code of flexible resources; substituting load model parameters into a power distribution network safety constraint verification model to verify compliance, and triggering re-generation when a violation occurs; taking the verified load model as input to solve a rolling optimization regulation and control model, outputting an optimal power regulation instruction sequence, and executing the optimal power regulation instruction sequence; receiving device layer feedback and user multi-round dialogue update information, and triggering strategy re-optimization. The application realizes real-time analysis of unstructured user intention, automatic construction of a load model, and physical safety verification of regulation instructions, and can be deployed at a low cost on an edge side.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control, and in particular to a distribution network control method based on multimodal intent perception and lightweight large model. Background Technology

[0002] With the continuous increase in the penetration rate of distributed renewable energy, a large number of heterogeneous and flexible resources have emerged in the distribution network, including distributed energy storage, electric vehicles, and temperature-controlled loads. Effectively incorporating these resources into the control system is key to achieving the safe, economical, and efficient operation of the distribution network.

[0003] Existing power distribution network flexibility resource control schemes can be summarized into three categories: The first category is rule-based demand response methods. These methods pre-define fixed control rules (such as temperature setpoint adjustments, charge / discharge time window limits, etc.), and execute corresponding control actions when the system state meets the trigger conditions. Their advantages are simple implementation and low computational overhead, but they lack the dynamic adaptability to changes in system state and user demands, resulting in rigid control strategies.

[0004] The second category is price-incentive-based methods. These methods guide users to adjust their electricity consumption behavior through price signals such as time-of-use pricing and real-time pricing. While these methods can achieve peak shaving and valley filling to some extent, the price signals are uniform, making it difficult to achieve refined and personalized control over the electricity preferences of different users. Furthermore, the responsiveness depends on the user's sensitivity to the price signals.

[0005] The third category is based on optimization and reinforcement learning methods. These methods establish a distribution network optimization scheduling model and use mathematical programming or reinforcement learning algorithms to obtain the optimal control strategy. For example, patent CN108011366A proposes a coordinated optimization control method for controllable resources in a distribution network, using a rolling optimization model to determine the optimal control variables for controllable resources; patent CN119029901A proposes a voltage control method considering load-side flexibility resources, establishing a voltage control strategy model with the goal of minimizing network losses and voltage deviation. Theoretically, these methods can obtain globally optimal or near-optimal scheduling schemes, but they all require users to input their electricity demand information in the form of standardized parameters (such as target values ​​for state of charge, upper and lower limits of charging and discharging power, and available time periods), posing a high barrier to entry for ordinary residential users.

[0006] The aforementioned existing technical solutions share the following common defects: (1) It does not support multimodal input such as natural language. Existing methods all require users to submit their electricity usage intentions in the form of structured parameters, which does not match the colloquial expression habits of ordinary residents and greatly limits the breadth of participation of flexible resources. Although there is an invention patent with publication number CN121984989A that attempts to parse user instructions through multimodal modules, its application scenario is limited to the plug-and-play of power grid Internet of Things devices and does not involve the control modeling of distribution network flexible resources.

[0007] (2) Load modeling relies on manual work by experts. Mathematical models of flexible resources (such as the state equation of energy storage and the thermal dynamic equation of temperature-controlled loads) need to be established by professionals one by one according to equipment parameters, which is cumbersome and difficult to scale up. The invention patent with publication number CN115955406B involves a self-management method for power grid models based on an edge computing framework, but it focuses on model version management and synchronization and does not involve automatically generating load model code using large language models.

[0008] (3) There are safety risks in directly generating control commands from large models. If a general large language model is directly used for distribution network control, its output may have an illusion phenomenon—the generated commands are grammatically reasonable but violate the physical operation constraints of the distribution network, posing a threat to the safe operation of the power grid.

[0009] (4) Lack of multi-round closed-loop human-computer interaction capability. Most existing methods are static modes of "one-time setting and long-term execution", and users cannot dynamically adjust the electricity consumption plan according to changes in actual needs in the middle.

[0010] Furthermore, current high-performance large language models generally have more than 10 billion parameters and require hundreds of gigabytes of GPU memory for inference, making them impossible to deploy and run directly on power distribution network edge controllers with limited computing resources.

[0011] Based on this, a distribution network control method based on multimodal intent perception and lightweight large model is proposed. Summary of the Invention

[0012] The purpose of this invention is to provide a distribution network control method based on multimodal intent perception and lightweight large model, which can realize real-time parsing of users' unstructured electricity consumption intent under edge computing power constraints, automatically construct a flexible load model that meets the constraints of safe operation of the distribution network, and drive the safe, interpretable, and personalized dynamic control of flexible resources.

[0013] To achieve the above objectives, this invention provides a distribution network control method based on multimodal intent perception and a lightweight large model, comprising the following steps: Step 1: Obtain the multimodal electricity consumption intention information submitted by the user, and perform fusion processing on the electricity consumption intention information through a multimodal fusion encoder to output a unified intention feature vector; Step 2: Input the unified intent feature vector into the lightweight large language model deployed on the edge of the distribution network. After parsing the intent feature vector, the lightweight large language model automatically generates the structured load model representation code of the distribution network flexibility resources. The structured load model representation code contains the optimization objective function and operating constraints of the flexibility resources. Step 3: Substitute the load model parameters in the structured load model representation code into the distribution network security constraint verification model for compliance verification; if the verification passes, proceed to Step 4; if the verification fails, the violation information is fed back to the lightweight large language model, triggering the regeneration of the structured load model representation code, until the verification passes or the preset retry limit is reached; Step 4: Using the load model that has passed the safety constraint verification as input, solve the rolling optimization control model of the distribution network and output the optimal power control command sequence for each flexibility resource; Step 5: Send the optimal power control command sequence to each flexible resource device for execution, and receive real-time operating status feedback from the device layer; in response to the updated power demand information submitted by the user through multiple rounds of dialogue, re-execute steps 1 to 4.

[0014] Preferably, in step 1, the modality includes at least one of natural language text, voice commands, and sensor states; the multimodal fusion encoder employs a cross-modal attention mechanism to weightedly fuse the feature vectors of each modality, and the fusion method is as follows: ; in, for Dimensional unified intent feature vector, For modal sets, For the first The original input data for each modality, For the first Feature encoding functions corresponding to each mode For the first Attention weights for each modality .

[0015] Preferably, in step 2, before parsing the intent feature vector, the lightweight large language model uses a two-stage collaborative fine-tuning method for domain adaptation: Phase 1: Supervised fine-tuning based on low-rank matrix factorization is employed. Efficient parameter fine-tuning is achieved by superimposing a low-rank increment matrix onto the frozen original weight matrix of the pre-trained model. The parameter update method is expressed as follows: ; in, This is the original weight matrix of the pre-trained model; It is a low-rank incremental weight matrix; This is the effective weight matrix after fine-tuning; and Given two trainable low-rank matrices, For matrix transpose; The rank of the low-rank decomposition; This is the scaling factor; In this stage, the cross-entropy loss function is used as the optimization objective; The second stage employs a reinforcement fine-tuning approach based on group-relative strategy optimization. This involves using modeling accuracy as a reward signal to optimize the model's inference ability. The value network is replaced by within-group relative scoring, and its optimization objective function... for: ; in, These are weight parameters; The number of candidate outputs generated for the same input; For the first A sequence of word terms for each candidate output; For the first The length of the word sequence of each candidate output; This is the original relative advantage function; To truncate the relative advantage function; The weighting coefficients of the KL divergence regularization term control the degree to which the strategy deviates from the reference model; For the current strategy With reference model strategy KL divergence between them; This represents the expectation on the training dataset.

[0016] Preferably, in the second stage, the reward function for enhanced fine-tuning consists of two parts: a reward for compliance with format requirements and a reward for the correctness of the adjustment results, expressed as: ; in, For the first The total reward value of each candidate output; The reward component is based on format rules; The reward component is based on the correctness of the regulatory results; The format compliance bonus is used to evaluate whether the structured load model representation code conforms to the preset format specifications, and its calculation method is as follows: ; in, This is the format reward value corresponding to the correct output format. This is the format penalty value corresponding to the output format error; The set of standard start and end fields required by the structured load model code; The structured representation content output by the large model; The parentheses represent Allen Iverson. The value is 1 if the condition inside the parentheses is true, and 0 otherwise. The reward for the correctness of the control results is used to evaluate whether the optimization model, which takes the structured load model representation code as input, is solvable and whether the control instructions obtained from the solution are correct. Its calculation method is as follows: ; in, This is the reward value when the model is solvable and the control instructions are correct. This is the reward value when the model is solvable but the control command is incorrect. To optimize the penalty value corresponding to when the model is unsolvable; To generate the penalty value corresponding to when the code fails to execute; , , , These are the sets of candidate output results that satisfy the above four conditions.

[0017] Preferably, in step 3, the distribution network security constraint verification model is constructed based on the DistFlow simplified power flow equation, expressed as: ; ; ; in, , They are respectively Time-of-day branch Flow to Node The active and reactive power; , They are nodes The active and reactive power demand of the load; , They are nodes The flexibility of access and the output of active and reactive power of resources; For nodes The set of downstream child nodes; , They are respectively Time period nodes and nodes The voltage amplitude; , Branch roads The resistance and reactance parameters; Its safety constraints include at least node voltage magnitude constraints and branch current thermal stability constraints.

[0018] Preferably, the violation information includes at least one of the following categories: The first type of violation information: physical power flow limit violation information directly captured by the distribution network safety constraint verification model, including at least the node number of the violation node, the violation period, the offset of the current voltage amplitude from the upper or lower limit of the voltage constraint, and / or the branch number of the violation branch, the violation period, the offset of the current amplitude from the upper limit of the thermal stability current carrying capacity. The second type of violation information: information that the optimization model is unsolvable or unexecutable due to the structured load model characterization code itself, including at least one of the following: contradictory constraints, syntax errors, or missing parameters; The third type of violation information: Violation information caused by power commands or timing constraints in the structured load model characterization code exceeding the physical or logical limits of the equipment, including at least one of the following: power commands exceeding the rated power limit of the equipment or logical contradictions in timing constraints.

[0019] Preferably, in step 4, the distribution network rolling optimization control model aims to minimize the operating cost of the distribution network. The operating cost includes at least one or more of the following: the cost of purchasing and selling electricity between the distribution network and the upper-level power grid, the equivalent operating cost of flexible resources, and the generation cost of distributed generators.

[0020] Preferably, in step 4, the objective function is: ; in, For the distribution network in the control time domain Total operating costs within the facility; for The purchase and sale price of electricity between the time-of-use distribution network and the upstream power grid; for Active power purchased by the distribution network from the upper-level power grid during a given time period; The equivalent operating cost coefficient per unit charge / discharge power for flexible loads; A set of flexible load nodes participating in regulation; , They are nodes Flexible load at Charging and discharging power during a given period; A set of distributed generator nodes; , , The first Cost coefficients for the quadratic, primary, and constant terms of a distributed generator; For the first Distributed generators in Power generation during a given time period; For the first A generator in The start / stop status for a time period is set to 1 to indicate operation and 0 to indicate shutdown. To adjust the time step.

[0021] Preferably, the flexible resources include at least one of distributed energy storage, temperature-controlled loads, and electric vehicles.

[0022] Preferably, in step 5, the real-time operating status fed back by the equipment layer includes the actual power, state of charge, and / or indoor temperature of each flexible resource.

[0023] Therefore, the distribution network control method based on multimodal intent perception and lightweight large model of the present invention has the following beneficial effects: (1) Lowering the threshold for user participation and expanding the scope of participation in flexible resources: This invention supports unstructured inputs such as natural language and voice through a multimodal fusion perception framework, breaking through the fundamental limitation of traditional distribution network control methods that require users to submit their electricity usage intentions in the form of structured parameters. It expands the scope of users participating in the flexible control of the distribution network from technical personnel with professional knowledge to ordinary residents, significantly improving the breadth of participation and response depth of flexible resources.

[0024] (2) Achieve lightweight deployment on the edge side and break through the bottleneck of computing power constraints: The two-stage collaborative fine-tuning method (LoRA supervised fine-tuning + GRPO enhanced fine-tuning) proposed in this invention is designed with differentiated training for the characteristics of the automatic modeling task of flexible load in distribution network. This enables the lightweight model with small parameter level to accurately complete the load modeling task and can be deployed locally at low cost on consumer-grade GPUs, solving the fundamental obstacle that high parameter large model cannot run on the edge side.

[0025] (3) Ensuring the physical security of control commands from the mechanism level: This invention uses a generation-verification-correction closed-loop mechanism to take the physical security constraints such as distribution network node voltage and branch current as hard thresholds for control commands. Any load model that does not meet the security constraints will be intercepted and triggered to be regenerated. This eliminates the potential threat to the safe operation of the power grid from the mechanism level of the large language model illusion phenomenon, and ensures the physical feasibility of all final issued commands.

[0026] (4) Support for multi-round interaction and continuous adaptive optimization: The present invention constructs a complete human-machine closed-loop feedback link of "user → system → device → user". Users can add, modify or cancel their electricity demand at any time through multi-round dialogue. The system automatically triggers the whole process to be re-executed, realizing the continuous adaptive optimization of the control strategy.

[0027] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0028] Figure 1 This is a flowchart of a power distribution network control method based on multimodal intent perception and lightweight large model according to the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.

[0031] This invention provides a distribution network control method based on multimodal intent perception and a lightweight large model, such as... Figure 1 As shown, it includes the following steps: Step 1: Multimodal User Intent Acquisition and Fusion The system acquires multimodal electricity consumption intention information submitted by the user, including at least one of natural language text, voice commands, and sensor status; it then fuses the electricity consumption intention information through a multimodal fusion encoder to output a unified intention feature vector.

[0032] The multimodal fusion encoder employs a cross-modal attention mechanism to weightedly fuse the feature vectors of each modality. The fusion method is as follows: ; in, for A unified intent feature vector is used as the input to the downstream lightweight large model; It is a set of modalities, corresponding to three types of information sources: natural language, speech, and sensors; For the first The original input data for each modality, For the first Feature encoding functions corresponding to each mode For the first Attention weights for each modality Attention weight It is obtained through adaptive computation using a cross-modal attention mechanism.

[0033] Step 2: Automatic modeling of large, lightweight models at the edges: This invention employs a lightweight large language model with a unified intent feature vector input deployed at the edge of the distribution network. For the task of automatic modeling of flexible loads in the distribution network, a two-stage collaborative training strategy of supervised fine-tuning and reinforcement fine-tuning is designed to efficiently adapt the basic large language model with a small number of parameters to the domain.

[0034] Specifically, the lightweight large language model employs a two-stage collaborative fine-tuning method for domain adaptation: Phase 1: Supervised fine-tuning based on low-rank matrix factorization (LoRA): By superimposing a low-rank increment matrix on the frozen original weight matrix of the pre-trained model, domain adaptation of the model is achieved with very few trainable parameters. The parameter update method is expressed as follows: ; in, This is the original weight matrix of the pre-trained model; It is a low-rank incremental weight matrix; This is the effective weight matrix after fine-tuning; and Given two trainable low-rank matrices, For matrix Transpose of; The rank of the low-rank decomposition ( ), controls the number of trainable parameters; This is the scaling factor, used to adjust the magnitude of the impact of the low-rank increment on the original weights.

[0035] In this stage, the cross-entropy loss function is used as the optimization objective, expressed as: ; in, This represents the value of the cross-entropy loss function; The total length of the word sequence generated by the model; For the first One generated word element; In the first All word sequences generated before each word; For model parameters At that time, in the generated sequence Generating lexical elements under conditions The probability of.

[0036] Phase Two: Enhancement Fine-Tuning Based on Group Relative Policy Optimization (GRPO) Building upon supervised fine-tuning, the GRPO algorithm is introduced to reinforce and fine-tune logically dependent data. This algorithm uses modeling correctness as a reward signal to optimize the model's inference ability. GRPO significantly reduces training computational overhead by replacing the value network with within-group relative ratings. Its optimization objective function... for: ; in, These are weight parameters; The number of candidate outputs generated for the same input; For the first A sequence of word terms for each candidate output; For the first The length of the word sequence of each candidate output; This is the original relative advantage function (the product of the strategy ratio and the advantage estimate). To truncate the relative advantage function (the product of the policy ratio after pruning and the advantage estimate); The weighting coefficients of the KL divergence regularization term control the degree to which the strategy deviates from the reference model; For the current strategy With reference model strategy KL divergence between them; This represents the expectation on the training dataset.

[0037] The relative advantage function for each candidate output is calculated as follows: ; ; in, For the current strategy under the conditions The next generation each word element The probability of; This represents the probability of generating the word under the same conditions using the old strategy (the strategy used during sampling); To accommodate the dynamic load requirements of user input; For the first The nth candidate output The entire sequence of lexical elements preceding each lexical element; For the first The candidate output of the th... The odds estimate at each word element; As a truncation function, it restricts the policy ratio to a certain value. Within the range; To prune boundary hyperparameters, control the single-step policy update magnitude.

[0038] The advantage function of each candidate output is obtained by normalizing the relative reward within the group: ; in, For the first The advantage function value of each candidate output reflects its quality relative to other outputs in the group; For the first The total reward value obtained by scoring each candidate output using the reward model; In the same group The average of the reward values ​​of each candidate output; In the same group The standard deviation of the reward values ​​of each candidate output.

[0039] When assigning control tasks to the distribution network, the reward function consists of two parts: a format compliance reward and a control result correctness reward, expressed as: ; in, For the first The total reward value of each candidate output; The reward component is based on format rules; The reward component is based on the correctness of the regulatory results; The format compliance bonus is used to evaluate whether the structured load model representation code conforms to the preset format specifications, and its calculation method is as follows: ; in, This is the format reward value corresponding to the correct output format. This is the format penalty value corresponding to the output format error; The set of standard start and end fields required by the structured load model code; The structured representation content output by the large model; The parentheses represent Iverson. The value is 1 if the condition inside the parentheses is true, and 0 otherwise.

[0040] The reward for the correctness of the control results is used to evaluate whether the optimization model, which takes the structured load model representation code as input, is solvable and whether the control instructions obtained from the solution are correct. Its calculation method is as follows: ; in, This is the reward value when the model is solvable and the control instructions are correct. This is the reward value when the model is solvable but the control command is incorrect. To optimize the penalty value corresponding to when the model is unsolvable; To generate the penalty value corresponding to when the code fails to execute; , , , These are the sets of candidate output results that satisfy the above four conditions.

[0041] After two stages of fine-tuning, the lightweight large language model analyzes the intention feature vector and automatically generates the structured load model representation code of the distribution network flexibility resources. The structured load model representation code contains the optimization objective function and operating constraints of the flexibility resources.

[0042] The flexible resources include at least one of distributed energy storage (ES), temperature-controlled loads (TCL), and electric vehicles (EV). The constraints for each type of resource are as follows.

[0043] (1) Constraints of distributed energy storage (ES): The charge state transition equation for energy storage is: ; in, for The state of charge (per unit) of the stored energy at the end of the time period; , These are the charging efficiency and discharging efficiency of energy storage, respectively. for Time-of-use energy storage charging power (MW) for Time-of-use energy storage discharge power (MW). , ; Rated energy storage capacity (MWh); To adjust the time step (h).

[0044] The operating boundary constraints for energy storage power and state of charge are: ; , ; in, , These are the lower and upper limits of the state of charge, respectively; Maximum permissible charging power (MW); The maximum permissible discharge power (MW).

[0045] (2) Temperature-controlled load (TCL) constraints: The dynamic equation for indoor temperature of temperature-controlled loads, represented by air conditioners, is as follows: ; in, for Indoor temperature during the time period (°C); for Outdoor ambient temperature (°C) during the time period; Building thermal resistance (°C / kW); Building heat capacity (kWh / ℃); The coefficient of performance (COP) for air conditioning cooling (heating); for Air conditioning power consumption during different time periods (kW); It is a natural constant.

[0046] The comfort constraints and power constraints for temperature-controlled loads are as follows: ; ; in, , These are the lower and upper limits (°C) of the indoor temperature acceptable to users, respectively. for The time-based air conditioner on / off status variable, where 1 indicates on and 0 indicates off; The rated power of the air conditioner (kW).

[0047] (3) Constraints on electric vehicles (EVs): The state transition equation for the electric vehicle charge is: ; in, For electric vehicles State of charge during a given time period; , These refer to the charging efficiency and discharging efficiency of electric vehicles, respectively. for Time-of-use charging power (kW); for Power discharged to the grid during a given time period (kW); This refers to the battery's rated capacity (kWh).

[0048] Electric vehicles during network connection time The internal operating constraints are: ; ; in, , These are the lower and upper limits for operation under charged conditions, respectively. , These are the times when the electric vehicle arrives at and leaves the charging station, respectively. This is the target value for the minimum state of charge when leaving the site, as required by the user.

[0049] Step 3: Verification and correction of distribution network security constraints: The load model parameters in the structured load model representation code are substituted into the distribution network security constraint verification model for compliance verification. The distribution network security constraint verification model is constructed based on the DistFlow simplified power flow equation. In the distribution network branches, nodes are set. For upstream nodes, node For a downstream node, its branch power flow equation is: ; ; ; in, , They are respectively Time-of-day branch Flow to Node Active and reactive power (MW, MVar); , They are nodes The active and reactive power demand of the load (MW, MVar); , They are nodes The flexibility of access includes the output of active and reactive power (MW, MVar). For nodes The set of downstream child nodes; , They are respectively Time period nodes and nodes Voltage amplitude (pu); , Branch roads The resistance and reactance parameters (Ω).

[0050] Its safety constraints include at least node voltage magnitude constraints and branch current thermal stability constraints: , ; , ; In the formula, , Do not specify the lower and upper limits (pu) of the node voltage amplitude; For the set of all nodes in the distribution network; For the set of control periods; for Time-of-day branch The current amplitude (A); branch road Upper limit of thermally stable current carrying capacity (A); This is the set of all branches in the distribution network.

[0051] If the verification passes, proceed to step 4; if the verification fails, the violation information will be fed back to the lightweight large language model, triggering the regeneration of the structured load model representation code.

[0052] Violations include at least one of the following categories: The first type of violation information: physical power flow limit violation information directly captured by the distribution network safety constraint verification model, including at least the node number of the violation node, the violation period, the offset of the current voltage amplitude from the upper or lower limit of the voltage constraint, and / or the branch number of the violation branch, the violation period, the offset of the current amplitude from the upper limit of the thermal stability current carrying capacity. The second type of violation information: information that the optimization model is unsolvable or unexecutable due to the structured load model characterization code itself, including at least one of the following: contradictory constraints, syntax errors, or missing parameters; The third type of violation information: Violation information caused by power commands or timing constraints in the structured load model characterization code exceeding the physical or logical limits of the equipment, including at least one of the following: power commands exceeding the rated power limit of the equipment or logical contradictions in timing constraints.

[0053] Repeat the above verification and regeneration process until verification is successful or the preset maximum number of retries is reached.

[0054] Step 4: Solving for the optimal control command: Using the load model that has passed safety constraint verification as input, the rolling optimization control model of the distribution network is solved. The objective function is to minimize the operating cost of the distribution network: ; in, For the distribution network in the control time domain Total operating cost (RMB); for The purchase and sale price of electricity between the distribution network and the upstream power grid during a given time period (RMB / MWh); for The active power (MW) purchased by the distribution network from the upper-level power grid during a given time period; a positive value indicates purchased electricity, and a negative value indicates sold electricity. Equivalent operating cost factor per unit charge / discharge power for flexible loads (RMB / MWh); A set of flexible load nodes participating in regulation; , They are nodes Flexible load at Charging and discharging power (MW) over a period of time; A set of distributed generator nodes; , , The first Cost coefficients (yuan / MW) for the quadratic, primary, and constant terms of a distributed generator 2 h, yuan / MWh, yuan); For the first Distributed generators in Power generation during a given time period; For the first A generator in The start / stop status for a time period is set to 1 to indicate operation and 0 to indicate shutdown. To adjust the time step.

[0055] After solving, the optimal power control command sequence for each flexibility resource is output.

[0056] Step 5: Human-machine closed-loop feedback and dynamic strategy update: The optimal power control command sequence is sent to each flexible resource device for execution, and the system receives real-time operating status feedback from the device layer. The real-time operating status includes the actual power, state of charge, and / or indoor temperature of each flexible resource. Users can submit updated electricity demand information at any time through multi-turn dialogue. In response to the updated information, the system re-executes steps 1 to 4 to achieve continuous adaptive optimization of the control strategy.

[0057] Example This embodiment uses a residential community's power distribution network as an application scenario and verifies the aforementioned method. The community has 200 households, and the power distribution network is connected to the upstream power grid via a single 10kV feeder. The rated capacity of the distribution transformer is 2×800kVA. Specific flexible resources connected within the community include: one distributed energy storage system with a rated power of 100kW and a rated capacity of 200kWh, connected to the 380V bus on the low-voltage side; approximately 60 electric vehicles, equipped with 30 7kW AC charging piles and 6 60kW DC fast charging piles, all with V2G bidirectional charging and discharging capabilities; approximately 160 temperature-controlled loads (air conditioning), with a total installed capacity of approximately 480kW; and approximately 150kWp of distributed photovoltaic power, operating in a self-consumption, surplus power grid-connected mode. The power distribution network is connected to the upstream power grid via a feeder and employs rolling optimization control, with the control time step set to [missing information]. minutes (0.25 hours).

[0058] like Figure 1 As shown, the specific implementation process of the present invention is as follows: Step 1: The superior power grid issues overall control instructions: Based on the current system requirements for frequency stability and peak shaving / valley filling, the upper-level power grid issues an overall power regulation command to the distribution network's flexible resource management platform, specifying the target power range that must be met at the feeder outlet during the control period. For example, based on the next day's load forecast and renewable energy output forecast, the upper-level power grid dispatch center issues an overall power regulation command for the period from 09:00 to 11:00 the following day to the flexible resource management platform of the residential community's distribution network at 23:00 on the same day. This period falls during the weekday morning peak, and the upper-level power grid predicts that the baseline load at the feeder outlet is approximately 2.8MW. To alleviate the pressure of heavy load on the main transformer, the distribution network is required to purchase no more than 2.2MW of active power from the upper-level power grid during the control period. That is, the upper limit of the feeder outlet power target range is 2.2MW, and there is no lower limit (allowing power backflow during periods of high distributed photovoltaic power generation).

[0059] Step 2: Multimodal User Intent Collection and Fusion The system collects electricity usage intent information submitted by users of various flexible resources within the area via natural language, voice, or smart home sensors.

[0060] For example, before leaving get off work, an electric vehicle user sent a voice command via a mobile app: "It's too hot. Please turn on the air conditioner in advance so that the room temperature will be 24 degrees Celsius when I get home at 5:00 PM. Also, the car was just plugged into the charging station and needs to be fully charged by 8:00 AM tomorrow. I'll charge it tonight to save money." The multimodal fusion encoder extracts features from the above voice commands, extracting key semantic information (target indoor temperature 24℃, expected arrival time 17:00, electric vehicle departure time 08:00, and target minimum state of charge at departure). Meanwhile, the smart home sensors reported the initial indoor temperature as... The current state of charge of the electric vehicle is The multimodal fusion encoder uses a cross-modal attention mechanism to weightedly fuse speech features and sensor features, outputting a unified intent feature vector.

[0061] Step 3: Automatic modeling of large, lightweight models at the edges: The unified intent feature vector output from step 2 is input into a lightweight large language model deployed on the cell edge gateway.

[0062] This lightweight large language model has completed two-stage fine-tuning in the offline phase: Phase 1 (LoRA Supervised Fine-Tuning): Historical data on distribution network flexibility resource modeling (including physical parameters, operational constraints, and optimized model code for various resources) is collected to construct a training dataset containing input-output pairs. Using cross-entropy loss as the optimization objective, LoRA fine-tuning is performed on the pre-trained model to enable it to generate structured load model code.

[0063] Phase Two (GRPO Reinforcement Fine-tuning): Building upon supervised fine-tuning, reinforcement fine-tuning is performed using format compliance and correctness of adjustment results as reward signals. For each input intent, the model generates multiple candidate outputs (e.g., Each candidate is assigned a format compliance reward (whether the code contains standard start and end fields) and a regulation result correctness reward (whether the generated optimization model is solvable and whether the solution result is correct). The advantage function is calculated by normalizing the relative reward within the group to optimize the fine-grained reasoning ability of the model.

[0064] After two stages of fine-tuning, the lightweight large model, upon receiving the feature vector from step 2, automatically identifies the user's demand type—the user simultaneously involves comfort settings for temperature-controlled loads (air conditioning) and charging scheduling for electric vehicles. The model automatically generates the corresponding structured load model representation code: For air conditioning load, the model is based on the initial indoor temperature. C. Target Temperature The expected arrival time is 17:00, taking into account the building's thermal parameters (thermal resistance). Heat capacity The system automatically constructs dynamic equations and temperature comfort constraints for indoor temperature (including air conditioning energy efficiency ratio), and calculates the power sequence boundary required to reduce the temperature from 30°C to 24°C.

[0065] For electric vehicles, the model identifies the time of grid entry (current), the time of grid exit (08:00 the next day), the current state of charge (20%), and the target state of charge for grid exit (100%), and automatically constructs the state of charge transition equation constraints, upper and lower limit constraints of the state of charge, and the minimum state of charge target constraint when leaving the grid.

[0066] The model's final output is structured code that can be directly parsed and executed by the controllable platform, containing complete optimization objectives and constraints for the two types of resources mentioned above.

[0067] Step 4: Verification and Correction of Distribution Network Safety Constraints Substitute the load model parameters output from step 3 into the DistFlow simplified power flow equation, and verify one by one whether the voltage amplitude of each node in the distribution network meets the constraints and whether the current of each branch meets the thermal stability constraints.

[0068] Verification revealed that the current in branch L12 was 1.28 pu at 17:15, exceeding the thermal stability current carrying capacity limit of 1.0 pu. The system reported a violation to the large model: "The current in branch L12 exceeded the limit at 17:15, with a current value of 1.28 pu and a limit of 1.0 pu, exceeding the limit by 22%. The reason is that multiple electric vehicles are charging simultaneously during this period. It is recommended to reduce the charging power of the electric vehicles connected to node 2 by at least 22%, or to shift the charging time of some vehicles to after 22:00." After receiving the above violation information, the lightweight large model adjusted the maximum single-pile charging power of the electric vehicles connected to node 2 from 7kW to 5.5kW when regenerating the code, and postponed the charging start time of 3 vehicles to 22:00. After re-substituting the data for verification, the branch current dropped to 0.95 pu, meeting the thermal stability constraints.

[0069] In this embodiment, the maximum number of retries is set to 3. If the verification still fails after 3 attempts, the system will automatically switch to conservative control mode (charging at the lowest power).

[0070] Step 5: Solving and issuing the optimal control command: Using the load model that has passed the safety constraint verification as input, the rolling optimization control model of the distribution network is solved.

[0071] The optimization objective is to minimize the total operating cost of the distribution network in the control time domain, including: The cost of purchasing and selling electricity between the distribution network and the upstream power grid (calculated based on time-of-use pricing); Equivalent operating costs of flexible resources (energy storage charging and discharging depreciation costs, electric vehicle battery degradation costs, etc.); The generation cost (including operation and maintenance costs) of distributed generators.

[0072] After solving this optimization problem, the optimal power control command sequence for each flexibility resource in the future control time domain is output. This command sequence is then sent to each device for execution via a standardized communication protocol (such as IEC 61850 or Modbus). Step 6: Human-Machine Closed-Loop Feedback and Dynamic Strategy Update: Each flexible resource device reports its operating status in real time (actual power and state of charge of energy storage, actual power and indoor temperature of air conditioning, actual power and state of charge of electric vehicles, etc.), forming a closed-loop status.

[0073] If users have new electricity needs during the regulation process, they can submit updated requests via the mobile app using natural language: "My plans have changed. I need to go out at 10 pm tonight. Please fully charge the battery before then." After the system parses the update request, it automatically re-executes steps 3 to 5: the lightweight large model updates the off-grid time constraint of electric vehicles (adjusting it from 08:00 to 22:00), regenerates the load model code, re-solves the optimization problem after safety verification, outputs the updated control command sequence and issues it for execution, so as to realize the continuous adaptive optimization of the control strategy.

[0074] Therefore, this invention presents a distribution network control method based on multimodal intent perception and a lightweight large model. It supports multimodal inputs such as natural language, voice, and sensors, lowering the technical threshold for users to participate in distribution network flexibility control and expanding the scope of participation from professionals to ordinary residential users. Under edge computing power constraints, it achieves automatic modeling of flexible loads using a lightweight large language model with a small number of parameters, eliminating the need for manual expert intervention. Through a generation-verification-correction closed-loop mechanism, it ensures that all final control commands meet physical safety constraints such as distribution network node voltage and branch current, eliminating the potential threat of large model illusion to the safe operation of the power grid. It supports multi-round human-machine dialogue closed-loop feedback, allowing users to update their electricity demand at any time, and the system automatically triggers re-optimization of the control strategy. This method can be widely applied in the field of distribution network flexibility resource control, especially suitable for scenarios such as residential communities, commercial buildings, and industrial parks that include heterogeneous flexibility resources such as distributed energy storage, electric vehicles, and temperature-controlled loads. The method of this invention can be implemented through software programming and deployed on distribution network edge computing devices (such as edge gateways and smart terminals), demonstrating clear industrial applicability.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A distribution network control method based on multimodal intent perception and lightweight large model, characterized in that, Includes the following steps: Step 1: Obtain the multimodal electricity consumption intention information submitted by the user, and perform fusion processing on the electricity consumption intention information through a multimodal fusion encoder to output a unified intention feature vector; Step 2: Input the unified intent feature vector into the lightweight large language model deployed on the edge of the distribution network. After parsing the intent feature vector, the lightweight large language model automatically generates the structured load model representation code of the distribution network flexibility resources. The structured load model representation code contains the optimization objective function and operating constraints of the flexibility resources. Step 3: Substitute the load model parameters from the structured load model representation code into the distribution network security constraint verification model for compliance verification; If the verification passes, proceed to step 4; If the verification fails, the violation information will be fed back to the lightweight large language model, triggering the regeneration of the structured load model representation code until the verification passes or the preset retry limit is reached; Step 4: Using the load model that has passed the safety constraint verification as input, solve the rolling optimization control model of the distribution network and output the optimal power control command sequence for each flexibility resource; Step 5: Send the optimal power control command sequence to each flexible resource device for execution, and receive real-time operating status feedback from the device layer; in response to the updated power demand information submitted by the user through multiple rounds of dialogue, re-execute steps 1 to 4.

2. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 1, characterized in that, In step 1, the modality includes at least one of natural language text, voice commands, and sensor states; the multimodal fusion encoder uses a cross-modal attention mechanism to weightedly fuse the feature vectors of each modality, and the fusion method is as follows: ； in, for Dimensional unified intent feature vector, For modal sets, For the first The original input data for each modality, For the first Feature encoding functions corresponding to each mode For the first Attention weights for each modality .

3. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 2, characterized in that, In step 2, before parsing the intent feature vector, the lightweight large language model uses a two-stage collaborative fine-tuning method for domain adaptation: Phase 1: Supervised fine-tuning based on low-rank matrix factorization is employed. Efficient parameter fine-tuning is achieved by superimposing a low-rank increment matrix onto the frozen original weight matrix of the pre-trained model. The parameter update method is expressed as follows: ； in, This is the original weight matrix of the pre-trained model; It is a low-rank incremental weight matrix; This is the effective weight matrix after fine-tuning; and Given two trainable low-rank matrices, For matrix Transpose of; The rank of the low-rank decomposition; This is the scaling factor; In this stage, the cross-entropy loss function is used as the optimization objective; The second stage employs a reinforcement fine-tuning approach based on group-relative strategy optimization. This involves using modeling accuracy as a reward signal to optimize the model's inference ability. The value network is replaced by within-group relative scoring, and its optimization objective function... for: ； in, These are weight parameters; The number of candidate outputs generated for the same input; For the first A sequence of word terms for each candidate output; For the first The length of the word sequence of each candidate output; This is the original relative advantage function; To truncate the relative advantage function; The weighting coefficients of the KL divergence regularization term control the degree to which the strategy deviates from the reference model; For the current strategy With reference model strategy KL divergence between them; This represents the expectation on the training dataset.

4. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 3, characterized in that: In the second phase, the reward function for enhanced fine-tuning consists of two parts: a reward for compliance with formatting and a reward for the correctness of the adjustment results, expressed as: ; in, For the first The total reward value of each candidate output; The reward component is based on format rules; The reward component is based on the correctness of the regulatory results; The format compliance bonus is used to evaluate whether the structured load model representation code conforms to the preset format specifications, and its calculation method is as follows: ； in, This is the format reward value corresponding to the correct output format. This is the format penalty value corresponding to the output format error; The set of standard start and end fields required by the structured load model code; The structured representation content output by the large model; The parentheses represent Allen Iverson. The value is 1 if the condition inside the parentheses is true, and 0 otherwise. The reward for the correctness of the control results is used to evaluate whether the optimization model, which takes the structured load model representation code as input, is solvable and whether the control instructions obtained from the solution are correct. Its calculation method is as follows: ； in, This is the reward value when the model is solvable and the control instructions are correct. This is the reward value when the model is solvable but the control command is incorrect. To optimize the penalty value corresponding to when the model is unsolvable; To generate the penalty value corresponding to when the code fails to execute; , , , These are the sets of candidate output results that satisfy the above four conditions.

5. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 4, characterized in that: In step 3, the distribution network security constraint verification model is constructed based on the DistFlow simplified power flow equation, and is expressed as: ； ； ； in, , They are respectively Time-of-day branch Flow to Node The active and reactive power; , They are nodes The active and reactive power demand of the load; , They are nodes The flexibility of access and the output of active and reactive power of resources; For nodes The set of downstream child nodes; , They are respectively Time period nodes and nodes The voltage amplitude; , Branch roads The resistance and reactance parameters; Its safety constraints include at least node voltage magnitude constraints and branch current thermal stability constraints.

6. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 5, characterized in that: In step 4, the distribution network rolling optimization control model aims to minimize the operating cost of the distribution network. The operating cost includes at least one or more of the following: the cost of purchasing and selling electricity between the distribution network and the upper-level power grid, the equivalent operating cost of flexible resources, and the generation cost of distributed generators.

7. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 6, characterized in that: In step 4, the objective function is: ； in, For the distribution network in the control time domain Total operating costs within the facility; for The purchase and sale price of electricity between the time-of-use distribution network and the upstream power grid; for Active power purchased by the distribution network from the upper-level power grid during a given time period; The equivalent operating cost coefficient per unit charge / discharge power for flexible loads; A set of flexible load nodes participating in regulation; , They are nodes Flexible load at Charging and discharging power during a given period; A set of distributed generator nodes; , , The first Cost coefficients for the quadratic, primary, and constant terms of a distributed generator; For the first Distributed generators in Power generation during a given time period; For the first A generator in The start / stop status for a time period is set to 1 to indicate operation and 0 to indicate shutdown. To adjust the time step.

8. The distribution network control method based on multimodal intent perception and lightweight large model according to claim 7, characterized in that: Flexible resources include at least one of distributed energy storage, temperature-controlled loads, and electric vehicles.

9. A distribution network control method based on multimodal intent perception and lightweight large model according to claim 8, characterized in that: In step 5, the real-time operating status feedback from the equipment layer includes the actual power, state of charge, and / or indoor temperature of each flexibility resource.

Images