Power distribution network voltage control method based on hybrid expert multi-modal large model and related equipment

Abstract

This application provides a distribution network voltage control method and related equipment based on a hybrid expert multimodal large model, belonging to the field of smart distribution network operation optimization technology. The method includes: acquiring cross-sectional measurement data, time-series data, graph structure data, and textual description data of the distribution network; performing feature extraction and semantic alignment on each type of data to generate a unified-dimensional multimodal feature sequence; inputting the fused multimodal features into a pre-trained large language model using a parameter freezing strategy for semantic understanding and reasoning; and generating coordinated control commands for devices such as on-load tap-changing transformers, grouped switching capacitors, and photovoltaic inverters by dynamically calling multiple specialized expert networks through an output head network based on a hybrid expert architecture using an intelligent gating routing mechanism. This application achieves efficient fusion of multi-source heterogeneous data from the distribution network and full utilization of domain knowledge, significantly improving the accuracy, robustness, and computational efficiency of voltage control.

Description

Technical Field

[0001] This application relates to the field of smart distribution network operation optimization and artificial intelligence model application technology, and in particular to a distribution network voltage control method and related equipment based on a hybrid expert multimodal large model. Background Technology

[0002] As a crucial link connecting the transmission network and users, the voltage quality of the distribution network directly affects the safe and stable operation of the power system and the user's electricity experience. With the large-scale integration of distributed renewable energy sources and the rapid development of flexible resources such as electric vehicles and energy storage systems, the operating conditions of the distribution network are becoming increasingly complex, and voltage fluctuation problems are becoming more and more prominent.

[0003] Traditional distribution network voltage control primarily relies on on-load tap-changing transformers, group-switching capacitor banks, and photovoltaic inverters. On-load tap-changing transformers regulate voltage by adjusting transformer tap positions without interrupting power supply; their operation frequency is limited by mechanical lifespan, typically used for voltage regulation over longer timescales. Group-switching capacitor banks regulate node voltage by connecting or disconnecting capacitor banks; this regulation is discrete and cannot be frequently implemented. Photovoltaic inverters offer rapid response capabilities, achieving voltage control through reactive power regulation or active power reduction. These control devices exhibit significant differences in response time, regulation accuracy, and regulation range, constituting the multi-timescale characteristics of distribution network voltage control.

[0004] Under the new power system context, the distribution network exhibits "dual high" characteristics (high proportion of distributed power sources and high power electronics), which dramatically increases the complexity of voltage control. Specifically, this manifests in the following ways: 1) Distributed power output is random and intermittent, leading to frequent and large voltage fluctuations in the distribution network; 2) The distribution network topology is complex, with numerous nodes and strong coupling relationships between the voltages of each node; 3) Measurement data comes from diverse sources, including SCADA systems, phasor measurement units (PMUs), smart meters, and distribution terminal units (DTUs), and the data types cover various forms such as real-time measurements, historical records, topology information, and text descriptions, exhibiting typical multi-source heterogeneous characteristics; 4) Distribution network operation involves a large amount of textual information such as dispatch logs, equipment ledgers, maintenance records, and operating procedures. This information contains rich domain knowledge and expert experience, but traditional numerical analysis methods are difficult to utilize effectively.

[0005] In recent years, deep learning technology has been widely used in the power system field, providing a new technical approach for distribution network voltage control. However, existing methods mainly have the following limitations: 1) Single data representation method: Traditional methods usually convert distribution network data into numerical vector form for processing, which makes it difficult to fully retain the physical characteristics and semantic information of the data; 2) Difficulty in fusing multi-source heterogeneous data: The method of simply splicing together various types of data often leads to the curse of dimensionality and semantic conflicts; 3) Insufficient utilization of domain knowledge: Traditional data-driven methods are difficult to effectively embed professional knowledge such as voltage-reactive power characteristics, power flow constraints, and equipment operating limits into the model; 4) The contradiction between computational efficiency and accuracy: Increasing model complexity to improve prediction accuracy will lead to a sharp increase in computational overhead.

[0006] Therefore, there is an urgent need for a voltage control method that can effectively integrate multi-source heterogeneous data from the distribution network, make full use of domain knowledge, and achieve collaborative control of multiple devices and multiple time scales. Summary of the Invention

[0007] The main objective of this application is to propose a distribution network voltage control method, electronic device, storage medium, and program product based on a hybrid expert multimodal large model, so as to achieve efficient fusion and unified representation of multi-source heterogeneous data of the distribution network, make intelligent decisions by utilizing the general reasoning ability of the large language model, and realize collaborative voltage control of multiple devices and multiple time scales through a hybrid expert architecture.

[0008] To achieve the above objectives, one aspect of this application proposes a distribution network voltage control method based on a hybrid expert multimodal large model, the method comprising: Acquire raw multimodal data of the power distribution network, including cross-sectional measurement data, time series data, graph structure data, and text description data; Feature extraction and semantic alignment are performed on the original multimodal data to generate a multimodal feature sequence with a unified dimension; After fusing the multimodal feature sequences, they are input into a pre-trained large language model that employs a parameter freezing strategy for semantic understanding and reasoning. The main parameters of the pre-trained large language model are kept frozen, and only the parameters of its input processing layer and output processing layer are updated. The output state of the pre-trained large language model is input into an output head network based on a hybrid expert architecture. Multiple specialized expert networks are dynamically invoked through an intelligent gating routing mechanism to generate and output collaborative control commands for at least one voltage control device in the distribution network.

[0009] In some embodiments, the cross-sectional measurement data includes node voltage amplitude and phase angle, branch power, transformer tap position, capacitor bank switching status, and distributed generation output; the time series data includes historical voltage series, load time series, distributed generation output time series, and meteorological data; the graph structure data includes the topological connection relationship of the distribution network, line parameters, and node type labels; and the text description data includes at least one of the following: dispatch log, equipment ledger, fault report, maintenance record, operating procedures, and expert knowledge base.

[0010] In some embodiments, the step of performing feature extraction and semantic alignment on the original multimodal data includes: The cross-sectional measurement data are represented and aligned in a high-dimensional manner using a feedforward neural network. For the time series data, a patch-based segmentation strategy and a multi-head attention mechanism are used for temporal feature extraction and alignment; For the graph structure data, a graph convolutional network is used for topological feature extraction and alignment; For the text description data, feature extraction and alignment are performed using the word segmentation and embedding mechanism of a pre-trained large language model.

[0011] In some embodiments, the processing of the cross-sectional measurement data is implemented through the following mathematical model:

[0012]

[0013] in, For cross-sectional measurement data, For high-dimensional representation of cross-sectional measurement data, The semantic alignment results of the cross-sectional measurement data. and For the learnable weight matrix of the feature extraction network, For the corresponding bias vector For activation function, and These are the weight matrix and bias vector for the semantic alignment layer.

[0014] In some embodiments, the processing of the time series data is implemented using the following mathematical model:

[0015]

[0016] in, For time series data, This represents the hidden features extracted from time-series data using a multi-head attention mechanism. The semantic alignment results for time series data. This represents a multi-head attention operation used to model the temporal dependencies between time-series patches. This indicates the patch length used when segmenting the time series. The learnable weight matrix for the semantic alignment layer. This is the bias vector for the semantic alignment layer.

[0017] In some embodiments, the processing of the graph structure data is implemented through the following mathematical model:

[0018]

[0019] in, For graph-structured data, The original feature matrix, Let GCN_L(·) be the adjacency matrix reflecting the node connectivity, and let GCN_L(·) be an L-layer graph convolutional network. This is the hidden representation of graph-structured data after graph convolution feature extraction. The semantic alignment result for graph-structured data; The learnable weight matrix for the semantic alignment layer. This is the bias vector.

[0020] In some embodiments, the processing of the text description data is implemented through the following mathematical model:

[0021] in, For text description data, TokenEmbedding(·) is the word embedding encoding, and PositionalEmbedding(·) is the position embedding. This describes the semantic alignment results of the text description data.

[0022] In some embodiments, the parameter freezing strategy specifically includes: Freeze layer: Keep the basic weights of the word embedding layer, all Transformer encoder layers, and output layer of the pre-trained large language model unchanged during fine-tuning; Low-rank adaptation layer: A low-rank factorization matrix is ​​introduced into at least one Transformer layer of the pre-trained large language model. Domain adaptation is achieved by optimizing the parameters of the low-rank factorization matrix. Its forward propagation representation is as follows:

[0023] in, For the frozen pre-trained weight matrix, B∈ ^(d×r) and A∈ ^(r×k) is a trainable low-rank matrix, and the rank r is much smaller than the dimensions d and k; Task-specific layer: The modules used for feature extraction and semantic alignment, as well as the output head network of the hybrid expert architecture, are set to be fully parameter-trainable.

[0024] In some embodiments, the hybrid expert architecture output header network includes: Expert network ensemble: comprising multiple specialized expert networks, wherein the multiple specialized expert networks include: 1) On-load tap-changing transformer control expert network, whose inputs include target node voltage deviation, current tap position and load forecast information, and output is tap adjustment range suggestion; 2) A grouped capacitor switching control expert network, whose inputs include node reactive power demand, current capacitor bank status and system topology constraints, and whose output is a switching status suggestion for each capacitor bank; 3) Photovoltaic inverter control expert network, whose inputs include photovoltaic output prediction, current inverter status and local voltage deviation, and outputs reactive power regulation or power factor setpoint. 4) Coordinated control expert network, whose inputs include the current status of each voltage control device, system voltage distribution and optimization objectives, and whose output is a coordinated control strategy adjustment suggestion; A gated network is used to dynamically calculate the activation weights of each expert network in the expert network set based on input features; The expert fusion module is used to perform a weighted summation of the outputs of each expert network in the expert network set based on the activation weights output by the gating network, and generate the final collaborative control command.

[0025] In some embodiments, the mathematical expression for the gating network is:

[0026] Wherein, H_T is the output state of the pre-trained large language model, W_g is the gating weight matrix, and b_g is the bias vector; The mathematical expression for the expert fusion module is:

[0027] Where N is the number of expert networks, Expert_i(·) is the output of the i-th expert network, and G_i(·) is the corresponding activation weight.

[0028] In some embodiments, the gating network is configured as follows: Task-oriented gating: Selecting relevant expert network combinations based on the type of control task; State-aware gating: Adjusting the weights of each expert network according to the current operating status of the distribution network; Time-scale gating: Dynamically adjusts the activation mode of the expert network based on the time urgency of the control action.

[0029] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0030] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0031] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the method described above.

[0032] Compared with the prior art, this application has the following advantages and beneficial effects: 1) Efficient fusion of multimodal data: Dedicated feature extraction modules were designed for different types of distribution network data. The semantic alignment module maps the features of multimodal data to a semantic space compatible with a large language model, realizing efficient mapping of multimodal data to the semantic space of a large language model. This enables the model to simultaneously utilize the real-time nature of cross-sectional measurement data, the historical trends of time series data, the topological relationships of graph structure data, and the domain knowledge of text description data.

[0033] 2) High computational efficiency and strong generalization ability: By adopting a parameter freezing strategy, only the parameters of the input and output processing layers are updated, maintaining the general reasoning ability of the large language model while achieving accurate adaptation to the professional field of distribution network voltage control. The introduction of a low-rank adaptation mechanism further reduces the number of trainable parameters, significantly reduces the GPU memory required for training, and improves the training speed by about 2-5 times. At the same time, it effectively prevents catastrophic forgetting and improves the generalization ability and stability of the model.

[0034] 3) Multi-device, multi-timescale collaborative control: A hybrid expert architecture is applied to the design of the large language model output head. By constructing specialized expert networks for different control devices (OLTC, SC, PV inverters, etc.), multi-task distributed collaborative output is achieved. The MoE architecture dynamically allocates computing resources through an intelligent gating mechanism, which significantly reduces computing overhead while maintaining the model's expressive power, solving the technical challenge of a single output head being unable to adapt to multi-device, multi-timescale collaborative control.

[0035] 4) Significantly Improved Voltage Control Performance: Through computational verification on the IEEE 33-bus test system, the proposed method outperforms traditional OLTC / SC, LSTM, GCN, MLP, and LSTM+GCN methods in all three typical scenarios: load fluctuation, photovoltaic output fluctuation, and multiple faults. Specifically, it surpasses traditional methods in terms of voltage limit exceedance rate, average voltage deviation, voltage standard deviation, number of equipment actions, control cost, limit exceedance recovery rate, and control success rate. Particularly in high-penetration distributed energy scenarios, the voltage limit exceedance rate is only 2.4%, a reduction of approximately 83% compared to traditional methods. In fault scenarios, the limit exceedance recovery rate reaches 96.8%, with an average recovery time of only 15.8 seconds. Attached Figure Description

[0036] Figure 1 This is a flowchart of a distribution network voltage control method based on a hybrid expert multimodal large model provided in an embodiment of this application; Figure 2 This is a flowchart of the power distribution network voltage control method in the embodiments of this application; Figure 3 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0038] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0039] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0040] 1) An on-load tap changer is a power transformer that can adjust the output voltage by changing the winding turns ratio while operating under load (without power interruption). Its core feature is the on-load tap changer (OLTC) that ensures uninterrupted power supply during voltage regulation.

[0041] 2) Grouped capacitor switching is a device used for reactive power compensation in power systems. It involves dividing parallel capacitors into several groups according to their capacity and automatically or manually switching (connecting or disconnecting) one or more groups based on load changes. This achieves dynamic adjustment of the system's reactive power, thereby improving the power factor, reducing line losses, and improving voltage quality. SC is an abbreviation for Series Capacitor Installation, which belongs to reactive power compensation equipment in power systems. 3) PV inverter (Photovoltaic Inverter), also known as photovoltaic inverter or solar inverter, is the core power electronic device in a photovoltaic power generation system. Its main function is to convert the variable direct current (DC) generated by the solar panel into alternating current that meets the grid standards or the needs of the electrical equipment.

[0042] 4) MoE (Mixture of Experts) is a neural network architecture whose core idea is to improve model efficiency and performance through a "divide and conquer" approach. When processing each input, it activates only a small subset of specialized subnetworks (called "experts") instead of using all parameters, thus achieving high parameter capacity and low computational overhead.

[0043] In recent years, deep learning technology has been widely used in the power system field, providing a new technical approach for distribution network voltage control. However, existing methods mainly have the following limitations: 1) Limited data representation methods. Traditional methods typically convert distribution network data into numerical vectors for processing, making it difficult to fully preserve the physical characteristics and semantic information of the data. For example, directly flattening time-series voltage data into vectors will destroy its temporal correlation; encoding topology information as node number indices will ignore its graph structure essence.

[0044] 2) Difficulty in fusing multi-source heterogeneous data. Multi-source heterogeneous data in power distribution networks exhibit significant differences in structure, sampling frequency, and semantic level, making effective fusion difficult within a unified framework. Simply splicing together different types of data often leads to the curse of dimensionality and semantic conflicts.

[0045] 3) Insufficient utilization of domain knowledge. Distribution network voltage control involves a large amount of specialized knowledge, such as voltage-reactive power characteristics, power flow constraints, and equipment operating limits. Traditional data-driven methods struggle to effectively embed this domain knowledge into the model, resulting in limited generalization ability and interpretability of the model.

[0046] 4) The contradiction between computational efficiency and accuracy. Increasing model complexity to improve prediction accuracy leads to a sharp increase in computational overhead, making it difficult to meet the timeliness requirements of real-time control of distribution networks.

[0047] Large Language Models (LLMs), as a major technological breakthrough in the field of artificial intelligence, demonstrate powerful semantic understanding, knowledge fusion, context learning, and generative decision-making capabilities, providing a new technical path for distribution network voltage control. However, there are currently no mature technical solutions for effectively applying LLMs to distribution network voltage control and solving key technical problems such as multimodal data fusion, domain adaptation, and multi-task collaborative output. Therefore, this application provides a distribution network voltage control method, electronic device, storage medium, and program product based on a hybrid expert multimodal large model to achieve efficient fusion of multi-source heterogeneous data from the distribution network, accurate domain adaptation of the large language model, and collaborative optimization control of various control devices.

[0048] like Figure 1 As shown, this embodiment provides a distribution network voltage control method based on a hybrid expert multimodal large model, including the following steps: Step S1: Obtain the raw multimodal data of the distribution network. The raw multimodal data includes cross-sectional measurement data, time series data, graph structure data, and text description data.

[0049] Step S2: Perform feature extraction and semantic alignment on the original multimodal data to generate a multimodal feature sequence with a unified dimension. Specifically, this includes: The cross-sectional measurement data are represented and aligned in a high-dimensional manner using a feedforward neural network. For the time series data, a patch-based segmentation strategy and a multi-head attention mechanism are used for temporal feature extraction and alignment; For the graph structure data, a graph convolutional network is used for topological feature extraction and alignment; For the text description data, feature extraction and alignment are performed using the word segmentation and embedding mechanism of a pre-trained large language model.

[0050] Step S3: After fusing the multimodal feature sequences, input them into a pre-trained large language model employing a parameter freezing strategy for semantic understanding and reasoning. The main parameters of the pre-trained large language model are kept frozen; only the parameters of its input and output processing layers are updated.

[0051] Step S4: Input the output state of the pre-trained large language model into an output head network based on a hybrid expert architecture, dynamically call multiple specialized expert networks through an intelligent gating routing mechanism, and generate and output collaborative control instructions for at least one voltage control device in the distribution network.

[0052] The solutions of the embodiments of this application will be described in detail and explained below with reference to specific application examples.

[0053] This embodiment provides a distribution network voltage control method based on a hybrid expert multimodal large model. For example... Figure 2 As shown, the method includes the following steps: (1) Construct a multimodal data classification system related to voltage control in power distribution networks.

[0054] The multi-source data related to voltage control are divided into four categories: cross-sectional measurement data, time series data, graph structure data, and text description data. The characteristic attributes of each type of data are analyzed, and the technical difficulties of integrating each type of data with the large language model are identified.

[0055] In this embodiment, the multimodal data system involved in distribution network voltage control is divided into four categories: cross-sectional measurement data, time series data, graph structure data, and text description data, as detailed below: 1) Cross-sectional measurement data is an instantaneous snapshot of the distribution network's operating status, carrying complete operational information of the system at a specific moment. In voltage control scenarios, cross-sectional measurement data mainly includes: voltage amplitude and phase angle at each node, active power, reactive power and apparent power of branches, transformer tap positions, capacitor bank switching status, and distributed generation output. This type of data has high-dimensional sparsity (the number of distribution network nodes is large, but the distribution of measurement devices is limited) and strong physical constraints (node ​​voltages are constrained by power balance equations and equipment limits).

[0056] 2) Time series data records the dynamic trajectory of key parameters of the distribution network over time, providing important information support for voltage trend analysis and forecasting. In voltage control scenarios, time series data mainly includes: historical voltage sequences, load (total active and reactive power) time series, distributed generation output time series (photovoltaic, wind power), electricity price sequences, and meteorological data (temperature, irradiance, wind speed, etc.). This type of data exhibits multi-timescale characteristics, coupling features of periodicity and randomness, strong temporal correlation, and spatiotemporal coupling features of multiple variables.

[0057] 3) Graph data is a mathematical abstraction of the distribution network topology, describing the connections and network structure between various nodes (generators, transformers, loads, etc.) in the system. In voltage control scenarios, graph data mainly includes: distribution network topology connections, line parameters (impedance, power flow limits), transformer parameters (turns ratio, capacity), node type labels (PV nodes, PQ nodes, slack nodes), and equipment operating status. This type of data reflects the hierarchical nature of the network topology, the heterogeneity of nodes and edges, and strong spatial correlations.

[0058] 4) Textual description data provides crucial semantic support for intelligent analysis of distribution networks. In voltage control scenarios, textual description data mainly includes: dispatch logs and operation records, equipment ledgers and parameter descriptions, fault reports and event descriptions, maintenance records and schedules, operating procedures and specifications, expert experience and knowledge bases, etc. This type of data is characterized by a high concentration of domain-specific vocabulary (power industry terminology), significant structural heterogeneity (documents from different sources have different formats), and strong multimodal correlation (textual descriptions and numerical data corroborate each other).

[0059] (2) Design an adaptive feature extraction and unified representation method for multimodal data of power distribution network.

[0060] For cross-sectional measurement data, a feedforward neural network is used to achieve high-dimensional representation. For time-series data, a patch-based segmentation strategy and a multi-head attention mechanism are employed. For graph-structured data, a graph convolutional network is used for topological feature extraction. For textual description data, the word segmentation and embedding mechanisms of a pre-trained large language model are utilized. A semantic alignment module maps various heterogeneous features to a semantic space compatible with the large language model.

[0061] The core idea of ​​the unified representation framework for multi-source heterogeneous data constructed in this embodiment is to design a dedicated neural network architecture for each data modality to replace the input processing layer of the large language model, so as to fully capture the inherent feature patterns of various types of data, and at the same time achieve effective mapping to the semantic space of the large language model.

[0062] The unified representation process for various types of data can be described by a mathematical expression as follows:

[0063] Where X_m represents the m-th class of multi-source heterogeneous data (m∈{1,2,3,4} corresponds to cross-sectional measurement data, time series data, graph structure data, and text description data, respectively); T_m is the output result of the m-th class of data after unified representation processing; FeatureCapture_m(·) represents the unique feature extraction module for each type of data; Align_m(·) is the alignment transformation module for the semantic space of a large language model, which is responsible for mapping the extracted features to a unified dimension compatible with the representation space of the main computation layer of Transformer.

[0064] 2.1) Characterization of cross-sectional measurement data For cross-sectional measurement data ∈ ^(1×n), this invention uses a feedforward neural network to achieve high-dimensional representation. The specific transformation process is described as follows:

[0065]

[0066] in, ∈ ^(1×d_F) represents the high-dimensional representation of the cross-sectional measurement data; ∈ ^(1×d_model) represents the semantic alignment result of the cross-sectional measurement data; and This is the learnable weight matrix for the feature extraction network; and It is the bias vector; d is the activation function; d_model is the hidden dimension of the large language model.

[0067] Through this linear transformation, the original n-dimensional measurement data is mapped to a d_model-dimensional feature space that is compatible with the input space of the large language model, and can be directly read by the Transformer main computation layer of the large language model.

[0068] 2.2) Time Series Data Representation For time series data ∈ ^(l×1), this invention employs a patch-based segmentation strategy. Non-overlapping segments are created based on time intervals of patch length p, resulting in... ∈ ^(n_patch×p), where n_patch = l / p represents the number of patches after partitioning.

[0069] The core idea of ​​patching is to transform continuous time series into discrete sequence units, conceptually corresponding to the concept of tokens in natural language processing. Specifically, each time patch can be regarded as a basic semantic unit of time-series data, carrying complete time-series information within a specific time period.

[0070] Based on the patching strategy described above, the feature extraction process for time series data employs a multi-head attention mechanism to uncover temporal dependencies and distribution patterns in the patched sequences:

[0071] Here, H2 represents the time-series hidden state representation after multi-head attention processing.

[0072] Subsequently, a linear transformation is used through a semantic alignment layer to align the hidden state H2 from the patch feature space to the standard feature space dimension of the large language model:

[0073] in, ∈ ^(n_ts×d_model) represents the high-dimensional representation of the distribution network time series data.

[0074] 2.3) Graph Data Representation For distribution network diagram data This invention selects Graph Convolutional Networks (GCNs) as the feature extractor. The selection of GCNs is based on the following two considerations: First, graph structures possess natural permutation equivariance, meaning that reordering nodes does not alter the topological nature of the graph. GCN naturally satisfies this property through its symmetric normalized adjacency matrix operations.

[0075] Secondly, the distribution network topology typically exhibits sparsity and local connectivity characteristics, which are precisely matched by the GCN's layer-by-layer neighborhood aggregation mechanism.

[0076] The feature extraction process for graph data achieves iterative updates of node features through a multi-layered GCN:

[0077] By stacking L layers of GCN, a node feature representation that incorporates the global topology is finally obtained:

[0078] Subsequently, a semantic alignment layer is used to align the feature representation of each node from the graph feature space to the standard feature space of the large language model:

[0079] in, ∈ ^(n_G×d_model) represents the high-dimensional representation of the distribution network diagram data.

[0080] 2.4) Textual Description Data Representation For text description data, this invention utilizes the lexicalization and embedding mechanisms inherent in large language models. For text data X4 of length L_t, it is first converted into a token sequence through a lexicalization process, then converted into a dense vector representation through a word embedding encoder, and finally combined with positional embeddings to obtain the final representation of the text data.

[0081] in, ∈ ^(n_t×d_model) represents the high-dimensional representation of the text description data.

[0082] Based on the aforementioned unified representation framework for multi-source heterogeneous data, all types of data in the power distribution network can be transformed into a unified representation form with dimension d_model through corresponding feature extraction and semantic alignment processes. For any type of data, its representation result T can be directly used as the input sequence of a large language model.

[0083] Since the high-dimensional tensors obtained after unified representation of various data maintain consistency in feature dimensions (the second dimension is always d_model), cross-modal splicing can be directly performed to achieve integrated modeling of multi-source information. For example, in scenarios where distribution network topology information, real-time measurement data, historical voltage sequences, and dispatch text information are simultaneously input, a fused representation can be obtained:

[0084] in, This is a fusion representation of multi-source heterogeneous data.

[0085] (3) Construct a voltage control framework for pre-trained large language models based on supervised fine-tuning.

[0086] By adopting a parameter freezing strategy, the general reasoning ability of the pre-trained large language model is maintained, and only the parameters of the input processing layer and the output processing layer are updated, so as to achieve accurate adaptation to the professional field of distribution network voltage control.

[0087] The huge number of parameters in large language models poses a severe challenge to traditional full-parameter fine-tuning methods. Full-parameter fine-tuning requires updating all model parameters on the distribution network voltage control dataset, resulting in huge computational overhead and easily leading to the problem of catastrophic forgetting - while the model adapts to new tasks, it will lose the general language understanding and world knowledge accumulated in the pre-training stage.

[0088] To address this issue, the present invention adopts a parameter freezing strategy, only updating the trainable parameters in the input processing layer and the output processing layer, and at the same time introducing the Low-Rank Adaptation (LoRA) mechanism in the Transformer layer of the large language model to achieve efficient domain adaptation while keeping the main body parameters of the model unchanged.

[0089] LoRA is a parameter-efficient fine-tuning technique. Its core idea is to add a low-rank decomposition matrix beside the weight matrix of the pre-trained model and indirectly update the model weights by optimizing these low-rank matrices. Suppose the weight matrix of a certain Transformer layer of the pre-trained model is W ∈ ^ (d×k), LoRA represents the weight update as:

[0090] where B ∈ ^ (d×r), A ∈ ^ (r×k), and r << min(d, k) is the low-rank dimension. The original forward propagation becomes:

[0091] During the fine-tuning process, W remains frozen and only the matrices A and B are trained. The mathematical principle of this design is based on the following observation: the change matrix of the weights of the pre-trained language model during fine-tuning usually has a low intrinsic rank, that is, the parameter update direction is distributed in a low-dimensional subspace in the high-dimensional parameter space. Therefore, the core direction of parameter updates can be efficiently captured through low-rank approximation.

[0092] Frozen layer: Keep all the original weights W of the pre-trained large language model unchanged, including the word embedding layer, the main body of the Transformer encoder, and the base weights of the output layer. These parameters have been pre-trained on a large-scale general corpus and contain rich language knowledge and world common sense, which have important knowledge transfer value for the distribution network voltage control task.

[0093] LoRA adaptation layer: Introduce LoRA modules in the Transformer layer of the large language model and only train the parameters of these low-rank matrices. In the supervised fine-tuning stage, optimize the A and B matrices through gradient descent so that the model can learn the specific knowledge and decision-making patterns in the distribution network voltage control field.

[0094] Task-specific layers: The input processing layer (feature extraction network and semantic alignment module) and the output processing layer (expert network and gating network of the MoE architecture) participate in the training of all parameters. These layers are responsible for converting multi-source heterogeneous data of the distribution network into a model-processable representation and mapping the model output to specific control actions.

[0095] Compared to full parameter fine-tuning, the partial parameter freezing and LoRA strategy adopted in this embodiment has the following advantages: the number of parameters is greatly reduced, and the GPU memory required for training is reduced from hundreds of GB in full parameter fine-tuning to several GB; the training speed is increased by about 2-5 times; catastrophic forgetting is effectively prevented, and pre-trained knowledge is preserved; low-rank constraints have a regularization effect, which improves the model's generalization ability on domain data.

[0096] (4) Design a voltage control output head network based on a hybrid expert architecture.

[0097] By constructing specialized expert networks for different control devices and time scales, and using intelligent gating routing mechanisms, multi-task distributed collaborative output can be achieved.

[0098] Voltage control in power distribution networks involves various control devices such as OLTC, SC, and PV inverters. These devices differ significantly in response time, regulation accuracy, and operating mechanisms, resulting in typical multi-tasking characteristics. Traditional single-output head networks struggle to effectively handle such complex control tasks involving multiple devices and multiple time scales. Therefore, this embodiment designs a voltage control output head network based on a hybrid expert architecture. By constructing specialized expert networks tailored to different control devices and time scales, it achieves distributed collaborative output for multiple tasks.

[0099] The core idea of ​​hybrid expert architecture is to decompose complex tasks into multiple subtasks, each handled by a dedicated network of "experts," with dynamic task allocation achieved through a gating mechanism. In the distribution network voltage control scenario, the MoE architecture designed in this embodiment includes the following key components: 4.1) Expert Network Assemblies: Construct N specialized expert networks, each expert focusing on the control logic of a specific type of equipment. Based on the characteristics of distribution network voltage control, expert networks are divided into the following categories: OLTC Control Expert: Responsible for adjusting the tap position of the on-load tap-changing transformer. The inputs to this expert network include target node voltage deviation, current tap position, and load forecast information; the output is a tap adjustment suggestion. The OLTC expert network employs a two-layer feedforward neural network structure with a hidden layer dimension of 256 using the ReLU activation function, and a softmax classification layer as the output layer, outputting the activation probability of each tap position.

[0100] The SC (Signal Control) expert network is responsible for optimizing the switching states of capacitor banks in groups. The inputs to this expert network include node reactive power demand, current capacitor bank status, and system topology constraints; the output is a suggested switching state for each capacitor bank. The SC expert network employs a hybrid structure, including a time-series processing layer (used to capture trends in reactive power demand) and a decision output layer. The output layer uses a sigmoid activation function to output the switching probabilities for each capacitor bank.

[0101] PV Inverter Expert: Responsible for reactive power regulation of the photovoltaic inverter. The inputs to this expert network include photovoltaic output prediction, current inverter status, and local voltage deviation; the output is reactive power regulation or power factor setpoint. The PV inverter expert network employs a multi-head attention structure to fuse photovoltaic output timing information and voltage deviation information. The output layer uses a tanh activation function to output continuous reactive power regulation.

[0102] The Coordination and Control Expert is responsible for the coordinated optimization of multiple devices, considering the coupling relationships and action priorities between devices. The inputs to this expert network include the current state of each device, system voltage distribution, and optimization objectives; the output is a suggestion for adjusting the coordination and control strategy. The Coordination and Control Expert uses a graph attention network structure to model the interaction relationships between devices.

[0103] 4.2) Gating Network: The gating network is responsible for dynamically allocating the weights of each expert network based on the input characteristics. The structural design of the gating network takes into account the actual needs of power distribution network voltage control. Task-oriented gating: Select relevant expert network combinations based on the type of control task (such as emergency voltage support, normal voltage regulation, distributed energy fluctuation suppression, etc.); State-aware gating: Adjusting expert weights based on the current operating status of the distribution network (such as voltage deviation magnitude, trend of change, adjustable resource margin, etc.); Time-scale gating: The expert activation mode is dynamically adjusted according to the time urgency of the control (such as millisecond-level emergency response, second-level dynamic adjustment, and minute-level steady-state optimization).

[0104] The mathematical expression for a gated network is:

[0105] Where W_g ∈ ^(N×d_model) is the gated weight matrix, b_g ∈ ^(N×1) is the bias vector, G(H_T) ∈ ^(N×1) represents the activation weights of N expert networks, satisfying... .

[0106] 4.3) Expert Fusion Strategy: The final voltage control output is obtained by a weighted combination of the various expert networks.

[0107] Here, Expert_i(H_T) is the output of the i-th expert network, and G_i(H_T) is the corresponding gating weight. Through weighted fusion, the professional knowledge of each expert network is effectively integrated to form a coordinated and unified control strategy.

[0108] Case Analysis A comparative analysis of voltage control performance was conducted on the IEEE 33-bus test system. The core objective of voltage control is to maintain the voltage at each node of the system within the allowable range (0.95-1.05 pu) by adjusting control variables such as the OLTC tap position, SC switching status, and PV inverter reactive power output, while meeting various constraints.

[0109] The system configuration is as follows: 1) Reference voltage: 12.66kV 2) Base capacity: 100MVA 3) Total active power load: 3715kW 4) Total reactive load: 2300 kVar 5) OLTC transformer: 1 unit (at node 1, ±8 adjustable range) 6) Segmented capacitors: 2 groups (at nodes 8 and 28, each 100kVar) 7) Photovoltaic power generation: 5 connections (nodes 8, 12, 18, 25, and 30, 200kW each) The control constraints are as follows: 1) Node voltage limit: 0.95-1.05 pu 2) OLTC tap adjustment range: ±8 levels 3) Capacitor switching status: 0 or 1 4) PV inverter reactive power regulation range: 0-50kVar / unit 5) Equipment operation limit: no more than 5 times per hour. Voltage control performance is compared for the following typical operating scenarios: Scenario 1: Load Fluctuation Scenario The simulation of weekday load fluctuations was conducted, with 96 time points set (24 hours, one control decision point every 15 minutes), and the load fluctuating randomly between 80% and 120% of the baseline value.

[0110] Scenario 2: Fluctuation in photovoltaic output Based on Scenario 1, the photovoltaic output fluctuation (considering actual factors such as cloudy skies and shadows) is superimposed, and the photovoltaic output varies between 0-100%, with a penetration rate of 26.9%.

[0111] Scenario 3: Multiple Fault Scenario Under normal load conditions, the voltage recovery control after a branch is disconnected or a device fails is simulated to test the robustness of the method.

[0112] Table 1 shows a comparison of the voltage control performance of each method under scenario 1 (load fluctuation): Table 1

[0113] Table 2 shows a comparison of the voltage control performance of each method under scenario 2 (photovoltaic power output fluctuation): Table 2

[0114] Table 3 shows a comparison of the voltage control performance of each method under scenario 3 (multiple faults): Table 3

[0115] The method described in this application demonstrates optimal voltage control performance in three typical scenarios: Scenario 1: The voltage over-limit rate is only 1.8%, which is 78.8% lower than the traditional OLTC / SC method and 56.1% lower than the GCN method. This is due to the semantic alignment and fusion of multi-source heterogeneous data, which enables the model to accurately predict load change trends and make control decisions in advance.

[0116] Scenario 2: In a high-penetration distributed energy scenario, the voltage exceedance rate is only 2.4%, significantly lower than other methods (14.2% for traditional methods and 6.8% for GCN methods). The intermittent fluctuations in photovoltaic output pose a severe challenge to voltage control. This invention captures the variation patterns of photovoltaic output through time-series data, understands the spatial propagation mechanism of voltage through graph structure data, and injects expert knowledge such as "inverter reactive power regulation should be prioritized during periods of high photovoltaic power generation" into textual description data, thus achieving stable voltage control in high-penetration scenarios.

[0117] Scenario 3: The over-limit recovery rate reached 96.8%, the control success rate reached 95.2%, and the average recovery time was only 15.8 seconds. This verifies the rapid response and reliable control capabilities of the MoE architecture in fault scenarios. Different expert networks handle the control of OLTC, SC, and PV inverters respectively, and the gating network quickly routes to relevant experts according to the fault type, realizing collaborative emergency control in fault situations.

[0118] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0119] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0120] Please see Figure 3 , Figure 3 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 302 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 302 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 302 and is called and executed by the processor 301 using the methods described above in the embodiments of this application. Input / output interface 303 is used to implement information input and output; The communication interface 304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 305 transmits information between various components of the device (e.g., processor 301, memory 302, input / output interface 303, and communication interface 304); The processor 301, memory 302, input / output interface 303, and communication interface 304 are connected to each other within the device via bus 305.

[0121] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0122] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0123] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0124] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0125] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented in the embodiments of this program product are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. The executable computer program code or "code" used to perform the various embodiments can be written in high-level programming languages ​​such as C, C++, Python, Smalltalk, Java, JavaScript, Visual Basic, Structured Query Language (e.g., Transact-SQL), Perl, or in various other programming languages.

[0126] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0127] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0129] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0130] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0131] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0133] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0134] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0135] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0136] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A power distribution network voltage control method based on a hybrid expert multimodal large model, characterized in that, The method includes the following steps: Acquire raw multimodal data of the power distribution network, including cross-sectional measurement data, time series data, graph structure data, and text description data; Feature extraction and semantic alignment are performed on the original multimodal data to generate a multimodal feature sequence with a unified dimension; After fusing the multimodal feature sequences, they are input into a pre-trained large language model that employs a parameter freezing strategy for semantic understanding and reasoning. The main parameters of the pre-trained large language model are kept frozen, and only the parameters of its input processing layer and output processing layer are updated. The output state of the pre-trained large language model is input into an output head network based on a hybrid expert architecture. Multiple specialized expert networks are dynamically invoked through an intelligent gating routing mechanism to generate and output collaborative control commands for at least one voltage control device in the distribution network.

2. The method of claim 1, wherein, The cross-sectional measurement data includes node voltage amplitude and phase angle, branch power, transformer tap position, capacitor bank switching status, and distributed generation output; the time series data includes historical voltage series, load time series, distributed generation output time series, and meteorological data; the graph structure data includes the topological connection relationship of the distribution network, line parameters, and node type labels; the text description data includes at least one of the following: dispatch log, equipment ledger, fault report, maintenance record, operating procedures, and expert knowledge base.

3. The method of claim 1, wherein, The step of performing feature extraction and semantic alignment on the original multimodal data includes: The cross-sectional measurement data are represented and aligned in a high-dimensional manner using a feedforward neural network. For the time series data, a patch-based segmentation strategy and a multi-head attention mechanism are used for temporal feature extraction and alignment; For the graph structure data, a graph convolutional network is used for topological feature extraction and alignment; For the text description data, feature extraction and alignment are performed using the word segmentation and embedding mechanism of a pre-trained large language model.

4. The method of claim 3, wherein, The processing of the cross-sectional measurement data is achieved through the following mathematical model: in, For cross-sectional measurement data, For high-dimensional representation of cross-sectional measurement data, The semantic alignment results of the cross-sectional measurement data. and For the learnable weight matrix of the feature extraction network, For the corresponding bias vector For activation function, and These are the weight matrix and bias vector for the semantic alignment layer; The processing of the time series data is implemented through the following mathematical model: in, For time series data, This represents the hidden features extracted from time-series data using a multi-head attention mechanism. The semantic alignment results for time series data, This represents a multi-head attention operation used to model temporal dependencies between time-series patches. This indicates the patch length used when segmenting the time series. The learnable weight matrix for the semantic alignment layer. This is the bias vector for the semantic alignment layer; The processing of the graph structure data is achieved through the following mathematical model: in, For graph-structured data, The original feature matrix, Let GCN_L(·) be the adjacency matrix reflecting the node connectivity relationships, and let GCN_L(·) be an L-layer graph convolutional network. This is the hidden representation of graph-structured data after graph convolution feature extraction. The semantic alignment result for graph-structured data; The learnable weight matrix for the semantic alignment layer. It is the bias vector; The processing of the text description data is implemented through the following mathematical model: wherein, is the text description data, TokenEmbedding(·) is the word embedding encoding, PositionalEmbedding(·) is the position embedding, is the semantic alignment result of the text description data.

5. The method of claim 1, wherein, The parameter freezing strategy specifically includes: Freeze layer: Keep the basic weights of the word embedding layer, all Transformer encoder layers, and output layer of the pre-trained large language model unchanged during fine-tuning; Low-rank adaptation layer: A low-rank decomposition matrix is ​​introduced into at least one Transformer layer of the pre-trained large language model, and domain adaptation is achieved by optimizing the parameters of the low-rank decomposition matrix; Task-specific layer: The modules used for feature extraction and semantic alignment, as well as the output head network of the hybrid expert architecture, are set to be fully parameter-trainable.

6. The method of claim 1, wherein, The hybrid expert architecture output header network includes: Expert network ensemble: comprising multiple specialized expert networks, wherein the multiple specialized expert networks include: 1) On-load tap-changing transformer control expert network, whose inputs include target node voltage deviation, current tap position and load forecast information, and output is tap adjustment range suggestion; 2) A grouped capacitor switching control expert network, whose inputs include node reactive power demand, current capacitor bank status and system topology constraints, and whose output is a switching status suggestion for each capacitor bank; 3) Photovoltaic inverter control expert network, whose inputs include photovoltaic output prediction, current inverter status and local voltage deviation, and outputs reactive power regulation or power factor setpoint. 4) Coordinated control expert network, whose inputs include the current status of each voltage control device, system voltage distribution and optimization objectives, and whose output is a coordinated control strategy adjustment suggestion; A gated network is used to dynamically calculate the activation weights of each expert network in the expert network set based on input features; The expert fusion module is used to perform a weighted summation of the outputs of each expert network in the expert network set based on the activation weights output by the gating network, and generate the final collaborative control command.

7. The method of claim 6, wherein, The mathematical expression for the gated network is: Wherein, H_T is the output state of the pre-trained large language model, W_g is the gating weight matrix, and b_g is the bias vector; The mathematical expression for the expert fusion module is: Where N is the number of expert networks, Expert_i(·) is the output of the i-th expert network, and G_i(·) is the corresponding activation weight.

8. The method of claim 6, wherein, The gating network is configured as follows: Task-oriented gating: Selecting relevant expert network combinations based on the type of control task; State-aware gating: Adjusting the weights of each expert network according to the current operating status of the distribution network; Time-scale gating: Dynamically adjusts the activation mode of the expert network based on the time urgency of the control action.

9. An electronic device, comprising: The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images