Power distribution network regulation method, device, system and equipment based on lightweight model and medium

Abstract

The application discloses a power distribution network regulation and control method, device, system, equipment and medium based on a lightweight model, belongs to the field of power distribution network regulation and control, and is suitable for an edge regulation and control node in communication with a terminal sensing device and a cloud management and control center respectively. The method comprises the following steps: acquiring multi-source heterogeneous data based on the terminal sensing device; extracting multi-source feature data from the multi-source heterogeneous data, and transmitting the multi-source feature data to the cloud management and control center, so that the cloud management and control center generates and feeds back model update information according to the multi-source feature data; updating a preset regulation and control model according to the model update information to obtain an updated regulation and control model; calling the updated regulation and control model to determine a regulation and control strategy according to the multi-source heterogeneous data, and performing regulation and control processing on the power distribution network based on the regulation and control strategy. The application can regulate and control through the edge regulation and control node, can shorten the data link of regulation and control, can not only improve the regulation and control rate, but also can adapt to different working conditions of the power distribution network, and improve the regulation and control precision.

Description

Technical Field

[0001] This invention relates to the technical field of distribution network control, and in particular to a distribution network control method, device, system, equipment and medium based on a lightweight model. Background Technology

[0002] With the advancement of the "dual-carbon" strategy, the penetration rate of distributed photovoltaic (PV) power in low-voltage distribution networks continues to increase. However, the volatility and intermittency of PV power output present a significant contradiction with the weak immunity and complex topology of low-voltage distribution networks. To reduce the impact of various fluctuations in PV power output, real-time regulation and control of the distribution network are required based on actual needs.

[0003] One common method for distribution network regulation is to pre-construct a regulation strategy model for the distribution network using an ARIMA time series model or an LSTM deep learning model and historical data of the distribution network, and then integrate this model with the distribution network's cloud platform or back-end platform. Data is collected in real time from various terminal devices in the distribution network. The platform then calls the regulation strategy model to output regulation strategies based on the real-time data, and implements regulation accordingly to reduce fluctuations in photovoltaic power output and load changes.

[0004] However, the above methods have the following technical problems: Since the models are generally deployed on cloud platforms, the data transmission links are long, which leads to extended control time and low control efficiency. Moreover, ARIMA time series models or LSTM deep learning models generally collect single-strategy control, resulting in low control accuracy and difficulty in adapting to different operating conditions of the distribution network. On this basis, the model has a long iteration cycle after deployment, and after a period of use, it will deviate from the actual operating conditions of the distribution network, further reducing the control accuracy. Summary of the Invention

[0005] This invention provides a distribution network control method, device, system, equipment, and medium based on a lightweight model, which can solve one or more technical problems existing in the prior art.

[0006] A first aspect of this invention provides a distribution network control method based on a lightweight model, the method being applicable to edge control nodes that communicate with terminal sensing devices and a cloud-based control center, respectively, the method comprising: Multi-source heterogeneous data is acquired based on terminal sensing devices, wherein the multi-source heterogeneous data is data on the operating status, topology, and environment of the power distribution network; Multi-source feature data is extracted from the multi-source heterogeneous data and transmitted to the cloud management center so that the cloud management center can generate and feed back model update information based on the multi-source feature data; The preset control model is updated according to the model update information to obtain the updated control model. The preset control model is a lightweight cloud model trained by the cloud control center based on historical data of the distribution network and then deployed. The updated control model is invoked to determine the control strategy based on the multi-source heterogeneous data, and the distribution network is controlled based on the control strategy.

[0007] This invention can be controlled through edge control nodes, which can shorten the control data link, improve data transmission efficiency, and increase control speed. On this basis, the model can be updated in the cloud before control is performed, which can adapt to different operating conditions of the power distribution network, reduce control deviation, and improve control accuracy.

[0008] In conjunction with the first aspect, in one implementation, acquiring multi-source heterogeneous data based on the terminal sensing device includes: Multi-source data is acquired from terminal sensing devices, including meteorological data, power data, topology data, and device data. The multi-source data is subjected to feature standardization and optimization processing to obtain optimized data, wherein the optimization processing includes: missing data completion, feature normalization processing and boundary standardization processing; After constructing the topological feature encoding of the optimized data and completing the consistency verification, the optimized data is converted into a standardized feature matrix of a unified dimension to obtain multi-source heterogeneous data.

[0009] In conjunction with the first aspect, in one implementation, the extraction of multi-source feature data from the multi-source heterogeneous data includes: The multi-source feature data is converted into a multi-dimensional joint feature matrix, and a fully connected mapping layer is used to project the multi-dimensional joint feature matrix onto a unified feature space to obtain unified hidden features. Spatial enhancement features and temporal enhancement features are extracted from the unified hidden features, and the spatial enhancement features and temporal enhancement features are fused to obtain fused enhancement features; A preset fusion network is invoked to extract multi-dimensional cross features from the fusion enhancement features to obtain multi-source feature data, wherein the multi-dimensional cross features are the cross relationship between spatial enhancement features and temporal enhancement features in the fusion enhancement features.

[0010] In conjunction with the first aspect, in one implementation, the step of calling the updated control model to determine a control strategy based on the multi-source heterogeneous data, and performing control processing on the distribution network based on the control strategy, includes: The updated control model is invoked to perform photovoltaic power output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic power output prediction parameters, wherein the photovoltaic power output prediction parameters include: short-time photovoltaic power output prediction data and medium- and long-term photovoltaic power output prediction data. A multi-objective optimization function is constructed using the photovoltaic output prediction parameters and the real-time detection parameters of the distribution network, wherein the real-time detection parameters include: real-time operating parameters of the power grid and the status of the energy storage system; The multi-objective optimization function is solved based on preset constraints to obtain the control strategy. The preset constraints include: photovoltaic output constraints, energy storage remaining power constraints, energy storage power constraints, voltage constraints, and line load constraints. The control strategy is converted into control instructions, and the terminal sensing device is controlled to perform corresponding adjustments through the control instructions.

[0011] In conjunction with the first aspect, in one implementation, the step of calling the updated control model to perform photovoltaic output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic output prediction parameters includes: The update control model is used to divide the multi-source heterogeneous data into short-term feature data and medium-to-long-term feature data based on a preset time scale; The feature importance of the short-term feature data and the medium-to-long-term feature data is determined based on the attention mechanism. The feature importance is then used to reorganize the short-term feature data and the medium-to-long-term feature data to obtain short-term reorganized data and medium-to-long-term reorganized data. The updated control model is invoked to perform photovoltaic power output prediction processing based on the short-term and medium-to-long-term reorganization data to obtain photovoltaic power output prediction parameters.

[0012] In conjunction with the first aspect, in one implementation, after the steps of extracting multi-source feature data from the multi-source heterogeneous data and transmitting the multi-source feature data to the cloud management center, the method further includes: Obtain early warning information, which is generated by the cloud-based control center after calling the built-in hybrid model to determine the fault type of the distribution network based on the multi-source feature data; Based on the fault type corresponding to the early warning information, a preset backup control strategy is invoked to perform early warning adjustment processing on the distribution network.

[0013] A second aspect of this invention provides a power distribution network control device based on a lightweight model. The device is suitable for edge control nodes that communicate with terminal sensing devices and a cloud-based control center, respectively. The device includes: The acquisition module is used to acquire multi-source heterogeneous data based on terminal sensing devices, wherein the multi-source heterogeneous data is data on the operating status, topology, and environment of the power distribution network; The extraction module is used to extract multi-source feature data from the multi-source heterogeneous data and transmit the multi-source feature data to the cloud management center, so that the cloud management center can generate and feed back model update information based on the multi-source feature data; The update module is used to update the preset control model according to the model update information to obtain the updated control model. The preset control model is a lightweight cloud model trained by the cloud control center based on historical data of the distribution network and then deployed. The control module is used to call the updated control model to determine the control strategy based on the multi-source heterogeneous data, and to perform control processing on the distribution network based on the control strategy.

[0014] A third aspect of this invention provides a power distribution network control system based on a lightweight model. The system includes: a terminal sensing device, an edge control node, and a cloud control center. The edge control node is communicatively connected to the cloud control center and the terminal sensing device, respectively. The edge control node executes the distribution network control method based on the lightweight model as described above.

[0015] Compared to existing technologies, the present invention provides a distribution network control method, device, system, equipment, and medium based on a lightweight model. Its advantages lie in the following: the present invention is applicable to edge control nodes that communicate with terminal sensing devices and a cloud-based control center. The method includes: acquiring multi-source heterogeneous data based on the terminal sensing devices; extracting multi-source feature data from the multi-source heterogeneous data and transmitting the multi-source feature data to the cloud-based control center, so that the cloud-based control center can generate and feed back model update information based on the multi-source feature data; updating a preset control model according to the model update information to obtain an updated control model; calling the updated control model to determine a control strategy based on the multi-source heterogeneous data, and performing control processing on the distribution network based on the control strategy. The present invention can perform control through edge control nodes, which can shorten the control data link, improve data transmission efficiency, and increase the control rate. Furthermore, by updating the model in the cloud before control, it can adapt to different operating conditions of the distribution network, reduce control deviations, and improve control accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a distribution network control method based on a lightweight model according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a power distribution network control device based on a lightweight model according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a power distribution network control system based on a lightweight model, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of a large-scale model of a cloud-based control center provided in an embodiment of the present invention; Figure 5 This is an interactive flowchart between the edge control node and the cloud control center provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] With the advancement of the "dual-carbon" strategy, the penetration rate of distributed photovoltaic (PV) power in low-voltage distribution networks continues to increase. However, the volatility and intermittency of PV power output present a significant contradiction with the weak immunity and complex topology of low-voltage distribution networks. To reduce the impact of various fluctuations in PV power output, real-time regulation and control of the distribution network are required based on actual needs.

[0019] One common method for distribution network regulation is to pre-construct a regulation strategy model for the distribution network using an ARIMA time series model or an LSTM deep learning model and historical data of the distribution network, and then integrate this model with the distribution network's cloud platform or back-end platform. Data is collected in real time from various terminal devices in the distribution network. The platform then calls the regulation strategy model to output regulation strategies based on the real-time data, and implements regulation accordingly to reduce fluctuations in photovoltaic power output and load changes.

[0020] However, the above methods have the following technical problems: Since the models are generally deployed on cloud platforms, the data transmission links are long, which leads to extended control time and low control efficiency. Moreover, ARIMA time series models or LSTM deep learning models generally collect single-strategy control, resulting in low control accuracy and difficulty in adapting to different operating conditions of the distribution network. On this basis, the model has a long iteration cycle after deployment, and after a period of use, it will deviate from the actual operating conditions of the distribution network, further reducing the control accuracy.

[0021] To address the aforementioned issues, the following detailed embodiments will be used to describe and illustrate a lightweight model-based distribution network control method, device, system, equipment, and medium provided in this application.

[0022] To address the technical problems of long adjustment times and inaccuracies in existing technologies, refer to Figure 1The diagram shows a flowchart of a distribution network control method based on a lightweight model provided by an embodiment of the present invention.

[0023] In one embodiment, the method is applicable to edge control nodes that communicate with terminal sensing devices and cloud control centers respectively.

[0024] In one embodiment, the edge control node is deployed in the distribution concentrator or photovoltaic power station control box. The hardware uses an industrial-grade edge computing gateway (CPU uses ARM Cortex-A72, memory ≥4GB, storage ≥32GB), integrating data acquisition interfaces (RS485, Ethernet, LoRa) and communication modules (PLC, 5G). Core functions: 1. Collect real-time data from terminal devices (sampling frequency ≥ 1Hz) and perform data preprocessing (filtering, deduplication, format conversion); 2. Deploy a lightweight large model inference module (compress cloud model parameters to less than 1 / 5 of their original size through knowledge distillation) to achieve local real-time control (such as active / reactive power regulation of photovoltaic inverters and local energy storage scheduling). 3. When communication with the cloud control center is interrupted, it automatically switches to "local self-governance mode" to maintain basic system operation based on preset rules and a lightweight model; 4. Upload pre-processed real-time data and local control logs to the cloud control center, and receive model update packages and global policies issued by the cloud control center.

[0025] As an example, the distribution network control method based on the lightweight model may include: S11. Acquire multi-source heterogeneous data based on terminal sensing devices, wherein the multi-source heterogeneous data is data on the operating status, topology, and environment of the power distribution network.

[0026] In one embodiment, the terminal sensing device includes a photovoltaic inverter (supporting adjustable active / reactive power, with a communication protocol compatible with GB / T 27930-2023), an energy storage system (equipped with a battery management system (BMS) that supports real-time SOC monitoring and charge / discharge control), a smart meter (supporting load data acquisition with a sampling frequency of ≥15 minutes / time), and a meteorological monitoring terminal (collecting solar irradiance, temperature, humidity, and wind speed with a sampling frequency of ≥1 minute / time).

[0027] Terminal sensing devices can collect equipment operation data and environmental data in real time and upload them to the cloud platform through edge control nodes; terminal sensing devices can also receive control commands issued by edge control nodes and perform operations such as photovoltaic power adjustment and energy storage charging and discharging.

[0028] Based on this, the terminal sensing device also has local fault detection capabilities (such as overcurrent and overvoltage protection) and promptly feeds back fault information to the edge control node.

[0029] After collecting multiple different data, the terminal sensing device can upload them to the edge control node to obtain multi-source heterogeneous data.

[0030] As an example, the acquisition of multi-source heterogeneous data based on the terminal sensing device may include the following sub-steps: S111. Obtain multi-source data from the terminal sensing device, wherein the multi-source data includes: meteorological data, power data, topology data and device data.

[0031] S112. Perform feature standardization and optimization processing on the multi-source data to obtain optimized data, wherein the optimization processing includes: missing data completion, feature normalization processing and boundary standardization processing.

[0032] S113. After constructing the topological feature encoding of the optimized data and completing the consistency verification, the optimized data is converted into a standardized feature matrix of a unified dimension to obtain multi-source heterogeneous data.

[0033] In one embodiment, multi-source data can be acquired from the terminal sensing device, and the multi-source data includes: 1. Meteorological data: solar irradiance (W / m²), temperature (°C), humidity (%), wind speed (m / s), cloud cover (%); 2. Power data: Real-time output of photovoltaic inverter (kW), user-side load (kW), grid voltage (kV), frequency (Hz), and line power flow (kA); 3. Topology data: Low-voltage distribution network line length (km), impedance (Ω), number of nodes, and equipment connection relationships; 4. Equipment data: Photovoltaic inverter efficiency (%), energy storage SOC (%), charge and discharge efficiency (%), equipment operating time (h).

[0034] Z-score standardization and min-max normalization were used to standardize data of different dimensions to eliminate the influence of units; missing data were filled using an attention-based interpolation algorithm with an accuracy of ≥95%.

[0035] In one operating method, the feature standardization process is as follows: Step 1: Data classification and feature recognition.

[0036] First, the input data is classified according to its source and attributes, forming meteorological feature sets, power feature sets, topological feature sets, and equipment feature sets.

[0037] Among them: meteorological characteristics include solar irradiance, temperature, humidity, wind speed, and cloud cover; power characteristics include photovoltaic output, load power, voltage, frequency, and line power flow; topology characteristics include line length, line impedance, number of nodes, and equipment connection relationships; equipment characteristics include inverter efficiency, energy storage SOC, charge and discharge efficiency, and operating time.

[0038] At the same time, a unified label code can be established for each feature to form a standardized feature index table.

[0039] The second step is abnormal data detection and cleaning.

[0040] Before standardization, the raw data undergoes a quality check.

[0041] The mean and standard deviation of a feature are calculated using a sliding time window method, when the following conditions are met: It is then judged as an outlier.

[0042] in: This is the current sampled value; The mean of the window; The standard deviation is the window value.

[0043] For outliers: if the data mutation is caused by a communication error, it is directly removed; if it is a short-term fluctuation, the average of the nearest time is used as a substitute; if there are continuous outliers, they are marked as missing data and enter the completion process.

[0044] After feature standardization, optimization processing can be performed. This optimization processing includes: missing data completion, feature normalization, and boundary standardization.

[0045] To complete missing data, missing data detection can be performed. Specifically, the data integrity rate within each time window can be calculated.

[0046] When data loss occurs due to interrupted data upload, sensor malfunction, or communication packet loss, a missing data marker matrix is ​​automatically generated. in: This indicates that the data is valid; This indicates missing data. This matrix also serves as input to the subsequent attention completion module.

[0047] Next, missing data can be imputed based on an attention mechanism. Specifically, for missing data, an attention interpolation algorithm can be used for imputation.

[0048] First, construct the feature sequences of the current time and historical times, as shown in the following formula: .

[0049] Next, the query vector, key vector, and value vector can be calculated. Specifically, the query vector Q, key vector K, and value vector V can be represented by the following formula: , , .

[0050] Then calculate the attention weights: .

[0051] Finally, the estimated missing location is obtained: .

[0052] in: For feature dimensions; This is the attention weight matrix; These are the completed eigenvalues.

[0053] Optionally, spatial correlation correction can be performed simultaneously by combining data from adjacent node devices, for example, by using the output of neighboring photovoltaic units and data from adjacent weather stations to improve the completion accuracy. The completion accuracy should be no less than 95%.

[0054] After completion, feature normalization can be performed.

[0055] For data with fixed upper and lower limits, max-min normalization is used, as shown in the following formula: ; Applicable to: SOC; humidity; inverter efficiency; cloud cover, etc.

[0056] Normalized data is mapped to: Interval.

[0057] After normalization, Z-score standardization can be performed.

[0058] For data with large fluctuations and no fixed boundaries, Z-Score standardization is used: .

[0059] Applicable to: photovoltaic output; user load; voltage; power flow; irradiance, etc.

[0060] After standardization, the mean of the data is 0 and the standard deviation is 1, which yields optimized data.

[0061] After obtaining the optimized data, topological feature encoding can be performed.

[0062] Specifically, non-numerical topology data such as device connection relationships are processed using graph structure encoding.

[0063] The power distribution network is represented as a diagram: .

[0064] in: A set of nodes; This represents a set of line connections. Information such as line impedance, voltage level, and power flow direction is encoded as edge feature vectors to achieve a digital representation of the topology.

[0065] Then, feature consistency verification can be performed on the optimized data. After standardization, consistency checks can be performed on various features. Specifically, this may include: voltage range verification; SOC boundary verification; power balance verification; and power flow direction verification.

[0066] Optionally, when a feature is found to violate the physical constraints of the power grid, a re-completion or re-standardization process is automatically triggered.

[0067] Finally, after completing the above verification, the optimized data can be converted into a standardized feature matrix with a unified dimension to obtain multi-source heterogeneous data.

[0068] Specifically, the final standardized feature matrix with a unified dimension can be represented by the following formula: ; Subsequently, the multi-source heterogeneous data can be encapsulated into a large model input vector in time series order. The standardized feature vectors maintain a consistent dimension, and the data transmission latency does not exceed 5ms, providing high-quality input data for subsequent cross-feature extraction, photovoltaic power output prediction modules, and multi-objective optimization scheduling.

[0069] In low-voltage distribution networks, photovoltaic inverters, energy storage systems, smart meters, and distribution concentrators belong to different manufacturers, resulting in inconsistent data formats and creating "data silos." Traditional control systems lack a unified data fusion platform, hindering real-time data interaction and collaborative decision-making among multiple devices. For example, real-time output data from photovoltaic inverters cannot be transmitted to the energy storage system in a timely manner, causing a disconnect between energy storage charging and discharging plans and actual photovoltaic output. Distribution concentrators struggle to obtain dynamic changes in user-side loads, making it easy for global scheduling strategies to conflict with local operating conditions, leading to low efficiency in collaborative control.

[0070] This invention breaks down data silos, enhances the collaboration of multiple devices, and enables standardized data interaction among devices such as photovoltaic inverters, energy storage systems, and smart meters. Through the global decision-making capabilities of a large model, it achieves collaborative scheduling of photovoltaic-storage-grid devices, and reduces cross-device collaborative response time to less than 200ms.

[0071] S12. Extract multi-source feature data from the multi-source heterogeneous data and transmit the multi-source feature data to the cloud management center so that the cloud management center can generate and feed back model update information based on the multi-source feature data.

[0072] In one embodiment, multi-source feature data can be extracted from multi-source heterogeneous data. In one operation mode, the edge control node can be embedded with a spatiotemporal attention mechanism to extract the cross features of multi-dimensional data, thereby obtaining multi-source feature data.

[0073] For example, by allocating spatial attention weights, the correlation between "solar irradiance in areas shaded by high-rise buildings and photovoltaic output" can be highlighted; by allocating temporal attention weights, the synergistic relationship between "peak and valley load periods and photovoltaic output" can be strengthened.

[0074] In an optional embodiment, the extraction of multi-source feature data from the multi-source heterogeneous data may include the following sub-steps: S121. The multi-source feature data is converted into a multi-dimensional joint feature matrix, and a fully connected mapping layer is used to project the multi-dimensional joint feature matrix onto a unified feature space to obtain unified hidden features.

[0075] S122. Extract spatial enhancement features and temporal enhancement features from the unified hidden features respectively, and fuse the spatial enhancement features and the temporal enhancement features to obtain fused enhancement features.

[0076] S123. Call the preset fusion network to extract multi-dimensional cross features from the fusion enhancement features to obtain multi-source feature data, wherein the multi-dimensional cross features are the cross relationship between spatial enhancement features and temporal enhancement features in the fusion enhancement features.

[0077] In one embodiment, the cross-feature extraction layer is used to mine the deep correlations between meteorological data, power data, topology data, and equipment data. By introducing a spatiotemporal attention mechanism, dynamic weight allocation and collaborative modeling among features of different dimensions are achieved. The specific steps are as follows: Step 1: Feature vector construction.

[0078] The system receives multi-source heterogeneous data output from the feature normalization layer. This multi-source heterogeneous data is a normalized feature matrix, which can be represented by the following formula: ; Among them: meteorological feature vector Electricity feature vector Topological feature vectors Device status feature vector .

[0079] Alignment is performed according to a unified timestamp to form a multidimensional joint feature matrix: As input to the cross-feature extraction module.

[0080] The second step is feature mapping and encoding.

[0081] To improve the comparability between different types of features, a fully connected mapping layer can be used to project data of different dimensions onto a unified feature space to obtain a unified hidden feature. Specifically, the unified hidden feature H can be expressed as follows: .

[0082] in: The input feature matrix; This is the mapping weight matrix; These are bias parameters; The hidden features are after unified encoding.

[0083] After encoding, all feature dimensions are unified to 256 or 512.

[0084] The third step is spatial attention feature extraction.

[0085] The spatial attention module primarily explores the relationships between different devices and different regions at the same time.

[0086] First, construct the query vector Q, key vector K, and value vector V, as shown in the following formula: ; ; ; Then calculate the association weights between each node: ; in: This is the spatial attention weight matrix; The feature dimension is then used to obtain spatially enhanced features: This process can automatically identify important spatial factors affecting photovoltaic output, such as: the relationship between the shading area of ​​high-rise buildings and photovoltaic power generation; the irradiance differences between different photovoltaic units; the relationship between the location of distribution network nodes and voltage distribution; and the relationship between the location of energy storage devices and power flow changes.

[0087] For regions with high relevance, greater attention weight can be automatically assigned.

[0088] Step 4: Extraction of temporal attention features.

[0089] The Time Attention module is used to uncover time dependencies in historical runtime data.

[0090] Continuous time windows can be constructed: Extracting feature sequences from consecutive time points: Then calculate the temporal attention weights: Obtain time-enhanced features: This process can automatically learn: peak and valley load variation patterns; daily cycle patterns of photovoltaic output; weather change trends; user load variation trends; and energy storage SOC variation patterns.

[0091] For example, during the midday high-irradiance period, the weight of irradiance characteristics can be automatically increased; during the evening peak load period, the weight of load change characteristics can be increased.

[0092] Step 5: Fusion of spatiotemporal cross-features.

[0093] Spatial augmentation features and temporal augmentation features can be fused to obtain fused augmentation features, as shown in the following formula: ; Then, a fusion network is used to further extract multi-dimensional cross features to obtain multi-source feature data. Specifically, this can be illustrated as follows: ; Among them, multidimensional cross features are the cross relationship between spatial enhancement features and temporal enhancement features in the fusion enhancement features.

[0094] This stage can establish the following typical correlations: solar irradiance × cloud cover × photovoltaic output; user load × energy storage SOC × grid power flow; voltage fluctuation × node topology × photovoltaic access capacity; weather change × grid load × energy storage dispatch demand. This achieves deep integration between meteorological, power, equipment, and topology information.

[0095] Optionally, important features can be filtered.

[0096] Specifically, feature importance can be calculated based on the magnitude of attention weights: .

[0097] When the importance of a feature is below a set threshold, its influence is automatically reduced; when it is above the threshold, its weight is increased.

[0098] For example, cloud movement speed, real-time irradiance, and peak-valley load variations typically receive higher weights, while redundant features that contribute less to prediction are automatically downweighted.

[0099] Finally, the output multi-source feature data can be a high-dimensional feature vector. Specifically, the multi-source feature data can be represented by the following formula: ; Its dimensions are no less than 1024. The output results simultaneously retain: spatial correlation information; temporal evolution information; and cross-correlation information of multi-source data; and serve as the unified input for subsequent photovoltaic power output prediction modules, multi-objective optimization scheduling modules, and fault diagnosis modules, and the total delay of feature extraction and transmission does not exceed 5ms.

[0100] After obtaining multi-source feature data, the data can be transmitted to the cloud control center, allowing the center to update the model based on the data and generate model update information. This updated information is then fed back to the edge control nodes.

[0101] In one optional operation mode, edge control nodes upload preprocessed runtime data to the cloud control center in real time. The cloud control center builds a dynamically updated dataset (with ≥10GB of new data added monthly). The cloud control center uses incremental learning technology to fine-tune the complete model based on the new data, avoiding the high cost of full training, with a training cycle of ≤24h. The cloud control center distributes the updated lightweight model package (≤100MB) to the edge control nodes, which automatically complete the model replacement without interrupting the control service (downtime ≤10s). The edge control nodes compare the control effects before and after the update (such as prediction accuracy and abandonment rate) and provide feedback on the verification results to the cloud control center, forming an iterative closed loop of "data accumulation - model training - effect verification".

[0102] S13. Update the preset control model according to the model update information to obtain the updated control model, wherein the preset control model is a lightweight version of the cloud model trained by the cloud control center based on historical data of the distribution network and then deployed.

[0103] In one embodiment, the edge control node can update the preset control model according to the model update information to obtain the updated control model. The preset control model is a lightweight cloud model trained by the cloud control center based on historical data of the distribution network and then deployed.

[0104] This invention adopts a large-scale model development paradigm of "pre-training + fine-tuning". The basic model is pre-trained based on power system mechanism data and public datasets, and then fine-tuned using low-voltage distribution network data from specific regions to adapt to the target scenario. The core modules of the large model include a multi-dimensional feature fusion module, a photovoltaic power output prediction module, a multi-objective optimization scheduling module, and a fault diagnosis module.

[0105] The mechanisms of power systems are mainly embedded into large models through four methods: mechanism constraint embedding, topology encoding, physical parameter characterization, and loss function constraint.

[0106] Specifically, firstly, physical parameters such as node connections, line impedance, transformer capacity, photovoltaic inverter capacity, and energy storage SOC range of the low-voltage distribution network are transformed into structured features and used as part of the input to the large model. Secondly, the distribution network topology is represented as a graph structure, with nodes representing users, photovoltaic systems, energy storage, or transformers, and edges representing line connections. Line impedance, voltage, current, and power flow direction are used as edge features input to the model. Thirdly, power flow calculation results are introduced as supervision or correction information during model training and inference to ensure that the control strategies output by the model satisfy basic physical laws such as voltage, current, and power balance. Finally, penalty terms for voltage exceeding limits, line overload, power imbalance, and energy storage SOC exceeding limits are added to the loss function to constrain the model from generating control commands that violate the safety boundaries of power system operation.

[0107] In this way, the large model does not simply rely on data fitting, but learns from historical operating patterns while being constrained by power flow constraints, topological constraints, and equipment physical characteristics of the distribution network, thereby improving the interpretability, security, and engineering feasibility of the control results.

[0108] The large, fully-fledged model (parameter scale ≥ 100 million) trained in the cloud-based control center is lightweighted to obtain a pre-defined control model. This lightweighting process can employ a combination of three techniques: knowledge distillation, quantization compression, and pruning. 1. Knowledge Distillation: Using a complete large model as the "teacher model," train "student models" with a parameter scale of ≤20 million, enabling the student models to learn the reasoning logic of the teacher model. After lightweighting, the model's accuracy retention rate is ≥95%. 2. Quantization and Compression: The model parameters are quantized from 32-bit floating-point numbers (FP32) to 16-bit floating-point numbers (FP16) or 8-bit integers (INT8), reducing the model storage volume by more than 75% and increasing the inference speed by more than 3 times; 3. Pruning: Remove redundant neurons and connections in the model while retaining the core inference path. After pruning, the computational load of the model is reduced by more than 60%, and it is adapted to the computational power of edge control nodes (it can run smoothly on ARM Cortex-A72).

[0109] The lightweight model is deployed at the edge control node and is responsible for real-time control tasks, such as short-term photovoltaic output prediction (5min-1h), local energy storage scheduling, and preliminary equipment fault diagnosis. The inference latency is ≤30ms, which meets the real-time requirements. The complete model is deployed in the cloud control center, responsible for medium and long-term forecasting (1 day to 7 days), cross-regional collaborative scheduling (such as inter-regional energy storage complementarity), and complex fault diagnosis (such as multi-device linkage failure). The inference results are sent to the edge control nodes via 5G. Edge control nodes combine local inference results with instructions from the cloud control center to optimize control strategies. For example, after the cloud control center issues a cross-regional energy storage dispatch plan, the edge control nodes adjust the charging and discharging power according to the actual state of their local energy storage SOC to ensure the feasibility of the strategy.

[0110] Existing control models are mostly deployed in cloud-based control centers, relying on large-scale data transmission and centralized computing. Low-voltage distribution network terminal equipment (such as edge controllers) has limited computing power and cannot handle the inference tasks of complex models. This results in control commands having to go through a long chain of "terminal data acquisition - cloud computing - command issuance," with latency often exceeding 500ms, making it unable to cope with short-term emergency situations such as sudden changes in photovoltaic output and load shocks. Furthermore, traditional model iterations require manual retraining and deployment, and cannot be dynamically updated based on real-time data, causing model adaptability to gradually decrease over time.

[0111] This invention adopts a deployment strategy of "lightweight edge inference + full cloud training" to compress the core modules of the large model to edge devices to meet the real-time control requirements of short-term emergency situations; the cloud control center realizes model iterative updates through global data training to ensure long-term performance optimization, and the model iteration cycle is shortened from "monthly" to "daily".

[0112] Moreover, through the continuous learning capabilities of large models, a knowledge base for photovoltaic control faults in low-voltage distribution networks is established, enabling accurate diagnosis of hidden faults (such as energy storage capacity decay and abnormal line losses), with a diagnostic accuracy rate exceeding 90%. At the same time, backup control strategies are automatically triggered, reducing fault handling time to less than 10 minutes and ensuring stable grid operation.

[0113] S14. Call the updated control model to determine the control strategy based on the multi-source heterogeneous data, and perform control processing on the distribution network based on the control strategy.

[0114] After the update is completed, the updated control model can be called to determine the control strategy based on multi-source heterogeneous data, and the control strategy can be used to control the various terminal devices of the distribution network so that the output and power of the distribution network can meet the existing real-time status of the distribution network.

[0115] In one embodiment, the step of calling the updated control model to determine the control strategy based on the multi-source heterogeneous data and performing control processing on the distribution network based on the control strategy may include the following sub-steps: S141. The updated control model is invoked to perform photovoltaic power output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic power output prediction parameters, wherein the photovoltaic power output prediction parameters include: short-term photovoltaic power output prediction data and medium- and long-term photovoltaic power output prediction data.

[0116] S142. Construct a multi-objective optimization function using the photovoltaic output prediction parameters and the real-time detection parameters of the distribution network, wherein the real-time detection parameters include: real-time operating parameters of the power grid and the status of the energy storage system.

[0117] S143. Solve the multi-objective optimization function based on preset constraints to obtain the control strategy, wherein the preset constraints include: photovoltaic output constraints, energy storage remaining power constraints, energy storage power constraints, voltage constraints, and line load constraints.

[0118] S144. The control strategy is converted into a control command, and the terminal sensing device is controlled to perform the corresponding adjustment through the control command.

[0119] In one embodiment, the updated control model can be invoked to perform photovoltaic power output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic power output prediction parameters, wherein the photovoltaic power output prediction parameters include: short-time photovoltaic power output prediction data and medium- and long-term photovoltaic power output prediction data.

[0120] Short-term photovoltaic power output forecast data can be photovoltaic power output forecast data on a short-term (5min - 2h) scale, while medium- and long-term photovoltaic power output forecast data can be photovoltaic power output forecast data on a medium- and long-term (1 day - 7 days) scale.

[0121] Next, a "photovoltaic-storage-grid" coordinated scheduling strategy can be generated based on the photovoltaic output prediction parameters and the real-time detection parameters of the distribution network (real-time grid operation parameters and energy storage status).

[0122] In one embodiment, a "photovoltaic-storage-grid" collaborative scheduling model can be constructed based on photovoltaic output prediction parameters, real-time grid operation parameters, and energy storage system status. Through steps such as operation status perception, power balance analysis, operation scenario identification, multi-objective optimization solution, and strategy generation, collaborative control between the photovoltaic system, energy storage system, and distribution network can be achieved.

[0123] Specifically, the process of generating and controlling strategies can include the following steps: The first step is to perceive the operational status.

[0124] First, obtain real-time operating data from the photovoltaic side, energy storage side, and grid side.

[0125] Specifically: the photovoltaic side includes current photovoltaic output, predicted photovoltaic output, inverter status, and curtailment rate; the energy storage side includes SOC, charging and discharging power, battery temperature, and state of health (SOH); and the grid side includes node voltage, line current, line load factor, power flow distribution, and user load demand.

[0126] The distribution network state vector is then constructed, as shown in the following formula: ; in: Represents the distribution network state vector; Indicates the projected output of photovoltaic power; Indicates the state of charge of the energy storage; Indicates load demand; Indicates node voltage; Indicates the line current; This indicates the health status of energy storage.

[0127] This state vector serves as the input for the subsequent scheduling decision module.

[0128] The second step is power balance analysis.

[0129] The power balance relationship for future periods is calculated based on photovoltaic forecasts and load demand: ; in: Indicates the power balance value; This indicates the predicted photovoltaic output; Indicates the energy storage charging and discharging power; This indicates the power demand of the load.

[0130] when: This indicates that the power generation capacity of the distribution network exceeds the load demand, and there is surplus power.

[0131] when: This indicates that the power generation capacity of the distribution network is insufficient, and it is necessary to call upon the energy storage system or the upstream power grid for compensation.

[0132] when: This indicates that the supply and demand of the power distribution network are basically balanced.

[0133] Power balance analysis can help identify supply and demand relationships in advance.

[0134] The third step is to run scene recognition.

[0135] The current operating scenario can be automatically identified based on the power balance analysis results.

[0136] If the following conditions are met: If so, it is determined to be a scenario with surplus photovoltaic power.

[0137] At this time, surplus photovoltaic power can be used to charge the energy storage system.

[0138] If the energy storage SOC reaches its limit, then photovoltaic power generation will be limited or power will be fed back to the upstream grid.

[0139] If the following conditions are met: If the supply and demand are balanced, then the current operating state can be maintained, with only minor fluctuations adjusted using the energy storage system.

[0140] If the following conditions are met: If the photovoltaic (PV) capacity is insufficient, then the scenario is considered a photovoltaic (PV) shortage. In this case, the energy storage system can be prioritized for discharge to compensate for the load shortfall; if the energy storage capacity is insufficient, electricity can be purchased from the upstream power grid.

[0141] The fourth step is to construct a multi-objective optimization model.

[0142] To achieve coordinated operation between photovoltaic systems, energy storage systems, and the power grid, a multi-objective optimization function can be constructed, as shown in the following equation: ; in: This represents the comprehensive optimization objective function; Indicates the light rejection rate; Indicates line loss; Indicates energy storage lifespan loss; Indicates the cost of purchasing electricity; This represents the weighting coefficient.

[0143] The optimization objectives include: minimizing solar curtailment rate; minimizing line losses; minimizing energy storage lifetime loss; and minimizing grid purchase costs.

[0144] Achieving a comprehensive balance between economy and safety through multi-objective optimization.

[0145] The fifth step is to establish operational constraints.

[0146] During the optimization process, operational constraints can be established simultaneously. These preset constraints include: photovoltaic output constraints, remaining energy storage capacity constraints, energy storage power constraints, voltage constraints, and line load constraints.

[0147] The photovoltaic output constraint conditions can be expressed as follows: ; in: This indicates the maximum power output of photovoltaic power generation.

[0148] The SOC constraint for energy storage can be expressed as follows: ; in: Indicates the minimum permissible state of charge for energy storage; This indicates the highest permissible state of charge for energy storage.

[0149] The energy storage power constraint can be expressed as follows: ; in: Indicates the maximum discharge power; This indicates the maximum charging power.

[0150] The voltage constraint condition can be expressed as follows: ; in: Indicates the minimum allowable voltage of the node; This indicates the maximum allowable voltage at the node.

[0151] The line load constraint can be expressed as follows: ; in: Indicates the real-time load rate of the line; This indicates the maximum allowable load rate of the line.

[0152] The above constraints ensure that the dispatching strategy meets the requirements for safe operation of the power grid.

[0153] Step 6: Solve the collaborative scheduling strategy.

[0154] A reinforcement learning algorithm combined with a multi-objective optimization algorithm is used to solve the scheduling problem.

[0155] The distribution network state vector can be first... Input scheduling model.

[0156] Then, a set of scheduling actions is generated based on the current state: ; The overall benefit after each action is executed can be calculated separately, and the scheduling scheme with the greatest benefit can be selected.

[0157] For example: during the peak photovoltaic power generation period at noon, the energy storage charging strategy is prioritized; during the evening peak load period, the energy storage discharging strategy is prioritized; when the voltage exceeds the limit, the reactive power compensation and inverter regulation strategy is implemented; when the line is overloaded, the load transfer strategy is implemented.

[0158] The optimal scheduling result is obtained through continuous iterative optimization.

[0159] Step 7: Generate a collaborative scheduling strategy.

[0160] After optimization, the final "photovoltaic-storage-grid" coordinated scheduling strategy can be output, resulting in a control strategy that can include photovoltaic-side strategy, energy storage-side strategy, and grid-side strategy.

[0161] The photovoltaic-side strategy can be as follows: ; in: This indicates the optimized photovoltaic output.

[0162] Energy storage strategies can be as follows: ; The grid-side strategy can be as follows: ; Ultimately, a control strategy is formed between the photovoltaic system, the energy storage system, and the power distribution network. This control strategy can be a collaborative control scheme.

[0163] Step 8: Execute feedback and dynamic correction.

[0164] Specifically, the control strategy can be converted into control commands, and these commands can be used to control the terminal sensing devices to perform corresponding adjustments. The dispatch commands can be sent to photovoltaic inverters, energy storage controllers, and distribution network control equipment for execution.

[0165] During execution, continuously monitor: photovoltaic output deviation; load changes; energy storage SOC changes; voltage fluctuations; and line load rate changes.

[0166] When the following conditions are met: ; in: Indicates actual photovoltaic output; This indicates the predicted photovoltaic output; This indicates the allowable error threshold.

[0167] It can automatically re-execute the scheduling optimization process and generate new scheduling strategies.

[0168] Ultimately, a closed-loop control mechanism is formed: prediction → scheduling → execution → feedback → re-optimization, realizing real-time coordinated control between photovoltaic systems, energy storage systems and distribution networks.

[0169] It should be noted that the constraint definition clarifies the hard and soft constraints of the low-voltage distribution network, as follows: Hard constraints: grid voltage deviation ≤ ±7% (compliant with GB / T 12325 - 2020), frequency deviation ≤ ±0.2Hz (compliant with GB / T 15945 - 2018), line current ≤ 1.2 times rated current, energy storage SOC range 20% - 80%; Soft constraints: curtailment rate ≤ 5%, energy storage charge / discharge cycle count ≤ 90% of design life, grid line loss rate ≤ 8%.

[0170] For constructing a multi-objective optimization function, a weighted optimization function can be constructed with the objectives of "minimizing the curtailment rate (F1), minimizing the grid line loss rate (F2), and maximizing the energy storage lifetime (F3)".

[0171] Specifically, it is possible Based on the photovoltaic output forecast results, real-time grid operating parameters, and energy storage operating status, the curtailment rate, grid line loss rate, and energy storage lifetime loss index can be calculated respectively, and used as input factors for a multi-objective optimization function.

[0172] (1) Construction of the target factor F1 for light absorption rate.

[0173] The photovoltaic power output forecast for the future scheduling cycle can be obtained first.

[0174] Let the theoretical power generation of photovoltaic power during the predicted period be: ; in: This indicates the predicted photovoltaic power generation.

[0175] Simultaneously obtain the actual photovoltaic power generation allowed to be received after the scheduling plan is executed: ; in: This indicates the actual amount of photovoltaic power consumed.

[0176] Due to insufficient load demand, limited grid capacity, or insufficient energy storage capacity, some photovoltaic power cannot be utilized, resulting in curtailed photovoltaic power. ; in: This indicates the amount of electricity wasted.

[0177] Further calculation of the light curtailment rate: ; in: This represents the target factor for light rejection rate.

[0178] The lower the curtailment rate, the stronger the photovoltaic absorption capacity; therefore, the optimization objective is: ; That is, to minimize the light rejection rate.

[0179] (2) Construction of the target factor F2 for power grid line loss rate.

[0180] Line losses can be calculated based on real-time power flow distribution, node voltage, and line current.

[0181] For the i-th line in the distribution network, its line loss is: ; in: Indicates the line current; Indicates the line resistance; This indicates the power loss of the line.

[0182] The total loss is obtained after calculating the losses of all lines: ; in: Indicates the number of lines; This represents the total line loss.

[0183] Further calculation of line loss rate: ; in: This represents the target factor for line loss rate; This indicates the total load power.

[0184] The lower the line loss rate, the more reasonable the power flow distribution and the higher the power grid operating efficiency. Therefore, the optimization objective is: ; This means achieving the lowest possible power grid line loss rate.

[0185] (3) Construction of energy storage lifetime target factor F3.

[0186] Battery life loss can be calculated based on the charge and discharge status of the energy storage system.

[0187] First, calculate the charge / discharge depth within the current scheduling period: ; in: Indicates the depth of discharge; This indicates the state of charge of the energy storage.

[0188] The lifespan of energy storage batteries is closely related to the depth of charge and discharge and the number of charge and discharge cycles.

[0189] Let the life loss per cycle be: ; in: Indicates the life loss per cycle; Indicates the energy storage charging and discharging power; , This represents the lifespan loss coefficient.

[0190] Further calculation of cumulative lifetime loss: ; in: Indicates cumulative lifespan loss; Indicates the length of the scheduling cycle.

[0191] Define the energy storage lifetime target factor: ; in: This indicates the energy storage lifespan loss index.

[0192] Since less lifespan degradation leads to longer battery life, the optimization objective is: ; This means achieving the longest possible energy storage lifespan.

[0193] (4) Form a set of multi-objective optimization factors After completing the above index calculations, a multi-objective optimization factor set can be constructed: .

[0194] in: Indicates the target factor for light rejection rate; This represents the target factor for power grid line loss rate; This represents the energy storage lifetime target factor.

[0195] The three target factors mentioned above represent the operational performance from three dimensions: new energy absorption capacity, grid operation efficiency, and energy storage health status, and serve as input parameters for the subsequent multi-objective optimization function construction module.

[0196] After constructing the target factors F1 for curtailment rate, F2 for grid line loss rate, and F3 for energy storage lifetime, a comprehensive optimization objective function can be established: ; in: This represents the comprehensive optimization objective function value; Indicates the target factor for light rejection rate; This represents the target factor for power grid line loss rate; This represents the target factor for energy storage lifetime loss. , , This represents the corresponding target weight coefficient.

[0197] Each weight satisfies: ; The weight parameters can be dynamically adjusted according to different operating scenarios.

[0198] For example: improving photovoltaic penetration in high-penetration scenarios Improvement in heavy-load scenarios of power distribution networks Improvement in scenarios with frequent charging and discharging of energy storage .

[0199] The optimization objective is: ; That is, to minimize the comprehensive objective function while satisfying the constraints of photovoltaic power output, energy storage operation, voltage, and line capacity.

[0200] When F reaches its minimum value, it indicates that the current dispatch scheme is in terms of: photovoltaic absorption capacity; grid operating efficiency; and energy storage health status. The three aspects achieve the best overall result.

[0201] When solving for decision variables, the actual solution during the optimization process is not F itself, but the scheduling decision variables: ; in: This indicates that photovoltaic power generation contributes energy. Indicates the energy storage charging and discharging power; This indicates the reactive power compensation amount of the inverter; This indicates the transferable load capacity.

[0202] The above decision variables can be continuously adjusted using reinforcement learning algorithms or multi-objective optimization algorithms to continuously reduce the objective function F.

[0203] Solution results output The optimal decision vector is finally obtained: ; in: This represents the optimal scheduling result; This represents the optimal photovoltaic power output scheme; This represents the optimal energy storage charging and discharging scheme; This represents the optimal reactive power compensation scheme. This represents the optimal load transfer scheme.

[0204] Scheduling policy generation The final "optical-storage-grid" coordinated scheduling strategy can be generated based on the optimal decision vector: ; Specifically, this includes: photovoltaic power generation regulation; energy storage charging and discharging planning; node voltage regulation; power flow optimization control; and load transfer control.

[0205] The final output is sent to the scheduling and execution module.

[0206] The reinforcement learning solution employs the Deep Deterministic Policy Gradient (DDPG) algorithm to solve the optimization function. The agent's state space consists of "predicted photovoltaic output, grid voltage, frequency, and energy storage SOC," while its action space comprises "active / reactive power regulation of the photovoltaic inverter and charging / discharging power of the energy storage." The reward function is designed based on the optimization function and constraints. Through continuous interaction with the environment, the agent iteratively optimizes the strategy, achieving a control response time ≤50ms.

[0207] For the output of the scheduling strategy, specific control instructions can be generated, including the active power adjustment step size of the photovoltaic inverter (≥1kW / step), reactive power compensation amount (kVar), and energy storage charging and discharging power (kW), which are then sent to the terminal equipment through the edge control node.

[0208] After completing the multi-objective optimization solution, the optimal decision vector can be obtained: ; in: This represents the optimal photovoltaic power output scheme; This indicates the optimal energy storage charging and discharging power; This represents the optimal reactive power compensation amount; This represents the optimal load transfer capacity.

[0209] Specific scheduling instructions can be generated based on the optimal decision vector and then distributed to each terminal device for execution through the edge control node.

[0210] (1) Generation of active power regulation commands for photovoltaic inverters: First, you can obtain the actual output power of the current photovoltaic inverter: ; And the target output power obtained by optimization: ; Calculate the power regulation: ; in: Indicates the amount of active power adjustment; This indicates the current actual photovoltaic output; This indicates the output of the optimized target.

[0211] Then, the adjustment step size is generated based on the inverter's allowable adjustment accuracy: ; in: Indicates the step size for a single adjustment; Indicates the number of adjustments.

[0212] The minimum adjustment step can be set to be no less than 1kW / step.

[0213] For example: Current output: ; Target output: ; but: ; It can generate 50 × 1kW load reduction commands or 10 × 5kW load reduction commands to avoid sudden changes in inverter power.

[0214] Finally, the inverter's active power regulation command is generated: .

[0215] (2) Generation of reactive power compensation instructions: The reactive power compensation target can be obtained from the optimization solution: ; And the current reactive power output of the inverter: ; Calculate reactive power adjustment: ; in: This indicates the amount of reactive power compensation adjustment.

[0216] For example: Current reactive power output: ; Optimization goal: ; but: ; It can generate: .

[0217] The data is then sent to the inverter for execution. Node voltage stability control is achieved through reactive power compensation.

[0218] (3) Generation of energy storage charging and discharging commands: The current charging and discharging power of the energy storage can be obtained: ; And the optimized solution results: ; Calculate the power adjustment amount: ; in: This indicates the amount of energy storage power adjustment.

[0219] like: This indicates that a charging operation is being performed. like: This indicates that a discharge operation is being performed.

[0220] For example: Current energy storage capacity: ; Optimization results: ; Then it generates: ; like: ; Then it generates: ; It can also verify SOC constraints: ; Avoid overcharging and over-discharging.

[0221] (4) Generation of load transfer instructions: For adjustable loads, the load transfer capacity can be determined based on the optimization results: Generate a load adjustment plan.

[0222] For example: ; Then a load transfer command can be sent to the demand response terminal: or This enables optimized load configuration.

[0223] (5) Instruction encapsulation and edge delivery: Various control commands can be uniformly encapsulated into scheduling messages: ; in: Photovoltaic control commands; This is a reactive power compensation instruction; For energy storage control commands; This is a load control command.

[0224] Then, protocol conversion and command issuance are performed through edge control nodes.

[0225] Supports: Modbus TCP; IEC 61850; MQTT; OPC UA; Industrial communication protocols, etc.

[0226] Edge control nodes send scheduling instructions to: photovoltaic inverters; energy storage converters (PCS); smart switches; reactive power compensation devices; and demand response terminals to complete the scheduling execution.

[0227] After the terminal device completes the execution, it provides real-time feedback: ; in: Indicates actual photovoltaic output; Indicates the actual reactive power output; Indicates the actual energy storage capacity; Indicates the state of charge of the energy storage; This represents the node voltage.

[0228] The feedback results can be compared with the target value.

[0229] If the following conditions are met: If so, the optimized scheduling process will be restarted.

[0230] The final result is a closed-loop scheduling and control mechanism: optimization solution → instruction generation → edge distribution → device execution → status feedback → further optimization.

[0231] Traditional control systems rely on preset rules for fault handling, and can only identify common, simple faults (such as photovoltaic inverter offline). They lack the ability to diagnose "hidden faults" (such as energy storage battery capacity decay and increased losses due to line insulation aging). When a non-preset fault occurs in the power grid, backup strategies cannot be automatically triggered, requiring manual intervention. Fault handling time often exceeds one hour, seriously affecting the reliability of power supply. In contrast, this invention, based on the multi-task processing capabilities of a large model, constructs a multi-objective optimization function that minimizes solar curtailment rate, minimizes grid losses, and maximizes energy storage lifetime. It dynamically generates photovoltaic inverter regulation and energy storage scheduling strategies, reducing the control response lag time to less than 50ms, and adapting to dynamic operating conditions such as photovoltaic output fluctuations and load changes.

[0232] In one embodiment, the step of calling the updated control model to perform photovoltaic output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic output prediction parameters may include the following sub-steps: S1411. Using the update control model, the multi-source heterogeneous data is divided into short-term feature data and medium-to-long-term feature data based on a preset time scale.

[0233] S1412. Determine the feature importance of the short-term feature data and the medium-to-long-term feature data based on the attention mechanism, and use the feature importance to perform feature recombination on the short-term feature data and the medium-to-long-term feature data to obtain short-term recombined data and medium-to-long-term recombined data.

[0234] S1413. The updated control model is invoked to perform photovoltaic power output prediction processing based on the short-term recombination data and the medium-to-long-term recombination data to obtain photovoltaic power output prediction parameters.

[0235] In one embodiment, photovoltaic power output prediction is achieved on short-term (5 min - 2 h) and medium-to-long-term (1 day - 7 days) scales based on the fused feature vectors.

[0236] The updated control model can include a short-term forecasting submodule and a medium- to long-term forecasting submodule.

[0237] The short-term prediction submodule can adopt a "convolutional neural network (CNN) + gated recurrent unit (GRU)" structure. The CNN extracts spatial features such as solar irradiance and cloud movement speed, while the GRU captures the temporal correlation of historical output and load changes. The prediction step size is 5 minutes, and the prediction window is 2 hours. By introducing an attention mechanism, features under special meteorological conditions such as cloudy or overcast skies are given higher weights, and the short-term prediction error is ≤10%.

[0238] The medium- and long-term forecasting submodule can adopt a "Transformer encoder + fully connected layer" structure. The Transformer encoder uses a self-attention mechanism to capture the correlation between long-term meteorological trends (such as seasonal changes and weather system movement) and photovoltaic output. The forecast step size is 1 hour, and the forecast window is 7 days. Combining regional historical meteorological data and photovoltaic output data, the medium- and long-term forecasting error is ≤15%.

[0239] The fused feature vector is first differentiated according to the prediction time scale, and then input into the short-term prediction submodule and the medium-to-long-term prediction submodule respectively. In other words, after the multi-dimensional feature fusion module outputs a unified fused feature vector, it can filter and reorganize the features according to the time scale, feature sensitivity and prediction task requirements to form a short-term prediction feature set and a medium-to-long-term prediction feature set.

[0240] Specifically, after the multi-dimensional feature fusion module outputs a unified fused feature vector, it is not directly input into the prediction module. Instead, it first enters the time-scale feature separation layer, where features are filtered, reorganized, and weighted according to the time range of the prediction task, forming short-term prediction feature sets and medium-to-long-term prediction feature sets, respectively. The specific process is as follows: The first step is to fuse the feature inputs.

[0241] Receive the fused feature vector output from the cross-feature extraction layer: ; This includes: meteorological features, power operation features, distribution network topology features, equipment operation features, and spatiotemporal cross-features. All features are stored uniformly in the Feature Pool.

[0242] The second step is to establish a feature time sensitivity evaluation model.

[0243] Since different features have varying degrees of influence on predictions at different time scales, we can first calculate the time sensitivity index for each feature: ; in: The time sensitivity of the i-th feature; Characteristic change; The time interval is specified. Higher time sensitivity indicates that the feature changes faster and is more suitable for short-term prediction.

[0244] The third step is short-term feature screening.

[0245] For prediction tasks ranging from 5 minutes to 2 hours, real-time changing features can be prioritized.

[0246] Set a short-term filtering threshold: Features that meet the conditions are included in the short-time feature set: ; This mainly includes: current solar irradiance; cloud cover; cloud movement speed; photovoltaic output curve for the past 2 hours; load change curve for the past 2 hours; voltage fluctuation; power flow change; energy storage SOC change rate; and real-time temperature change rate.

[0247] This type of characteristic has strong instantaneous fluctuation properties and can reflect the trend of photovoltaic power output changes in the next few minutes to hours.

[0248] The fourth step is to filter features over a medium to long term.

[0249] For prediction tasks ranging from 1 to 7 days, trend and periodic features can be prioritized.

[0250] Set a trend contribution indicator: ; in: This is a historical photovoltaic power output sequence; This is the correlation coefficient function.

[0251] when: ; Then it enters the medium-to-long-term feature set: ; This mainly includes: historical weather trends; changes in sunshine duration; seasonal variation factors; monthly average temperature changes; historical load curves; holiday load patterns; regional photovoltaic power generation patterns; long-cycle energy storage operation status; and historical statistical indicators of distribution network operation.

[0252] These characteristics can reflect trends over the next few days.

[0253] The fifth step is to redistribute attention weights.

[0254] To avoid information loss caused by simple filtering, attention mechanisms can be further utilized to calculate feature importance: ; in: ; Obtain the weights of each feature.

[0255] For short-term forecasting tasks: increase the weight of real-time irradiance; increase the weight of cloud movement; increase the weight of load fluctuations.

[0256] For medium- and long-term forecasting tasks: increase the weight of weather trends; increase the weight of seasonal characteristics; increase the weight of historical output patterns.

[0257] Step 6: Feature recombination.

[0258] The filtered features can be reorganized separately: Short-time prediction input vector: ; Medium- to long-term prediction input vector: ; in: For short-term attention weights; This represents the attention weights for medium to long-term periods.

[0259] Step 7: Form a dual prediction channel.

[0260] Ultimately, two independent prediction paths are formed: Short-term forecast channel: ; Used to predict photovoltaic output in the next 5 minutes to 2 hours.

[0261] Medium- to long-term forecast channel: ; Used to predict photovoltaic power output for the next 1 to 7 days.

[0262] Step 8: Output the prediction results.

[0263] The two prediction submodules output the following: and medium- and long-term photovoltaic power output forecast data The photovoltaic power output prediction parameters were obtained.

[0264] Optionally, after obtaining short-term and medium-to-long-term photovoltaic power output prediction data, Kalman filtering smoothing and error feedback correction can be performed to finally form photovoltaic power output prediction parameters.

[0265] For the short-term prediction submodule, features with strong real-time characteristics can be selected, including current solar irradiance, cloud cover, recent photovoltaic (PV) output, load changes, and voltage fluctuations. First, a CNN extracts local spatial features, such as the relationship between cloud movement, shading changes, and abrupt changes in PV output. Then, the feature sequence output by the CNN is input into a GRU, which captures the variation patterns of PV output within a continuous time window. Finally, a fully connected layer outputs the predicted PV output for the next 5 minutes to 2 hours.

[0266] For the medium- and long-term forecasting submodule, features with strong periodicity and trends can be selected, including historical meteorological trends, sunshine cycles, seasonal factors, historical load curves, and regional photovoltaic power output patterns. These features are first input into a Transformer encoder, which uses a self-attention mechanism to extract the correlation between long-term meteorological changes, load changes, and photovoltaic power output. Then, a fully connected layer maps the encoding results to output photovoltaic power output forecasts for the next 1 to 7 days.

[0267] The outputs of the two prediction submodules will enter the prediction result optimization layer, where they will be smoothed by Kalman filtering and compared with the actual photovoltaic output to form feedback data for subsequent model fine-tuning.

[0268] For the prediction result optimization layer, the prediction results are smoothed by the Kalman filter algorithm to eliminate the influence of random noise; at the same time, the predicted value is compared with the actual output value, the error is calculated and fed back to the large model fine-tuning module to dynamically optimize the model parameters, and the prediction accuracy is improved by ≥2% per month.

[0269] In one embodiment, after the steps of extracting multi-source feature data from the multi-source heterogeneous data and transmitting the multi-source feature data to the cloud management center, the method may further include the following steps: S21. Obtain early warning information, which is generated by the cloud control center after calling the built-in hybrid model to determine the fault type of the distribution network based on the multi-source feature data.

[0270] S22. Based on the fault type corresponding to the early warning information, invoke the preset backup control strategy to perform early warning adjustment processing on the distribution network.

[0271] By leveraging the continuous learning capabilities of large models, accurate diagnosis and adaptive handling of photovoltaic control-related faults in low-voltage distribution networks can be achieved, specifically as follows: 1. Fault Feature Library Construction: Collect common fault data of low-voltage distribution networks (such as photovoltaic inverter overcurrent, energy storage battery capacity decay, line overload, and communication interruption), extract fault features (such as sudden current change, voltage drop, and abnormal SOC change), and construct a fault feature library with ≥20 initial fault types, supporting dynamic addition.

[0272] 2. Fault Diagnosis Reasoning: A hybrid model of "Convolutional Neural Network (CNN) + Support Vector Machine (SVM)" is adopted. CNN extracts fault features from real-time data, and SVM classifies fault types, achieving a diagnostic accuracy of ≥90%. For latent faults (such as slow capacity decay of energy storage batteries), early warning is achieved by comparing the deviation trend between historical data and real-time data (warning time ≥24h).

[0273] The early warning of latent faults is mainly achieved through the process of "establishing historical benchmarks - calculating real-time deviations - trend analysis - threshold judgment - early warning output".

[0274] First, a health baseline model for equipment can be established based on historical normal operation data. For example, normal operation baselines can be established for energy storage batteries, including SOC changes, charge / discharge efficiency, capacity retention rate, and temperature rise curves; normal operation baselines can be established for lines, including voltage, current, line loss rate, and temperature changes. These baselines can be established separately for different seasons, different load levels, and different photovoltaic output ranges to avoid misjudging normal operating condition changes as faults.

[0275] Secondly, real-time equipment operating data can be collected and compared with historical benchmark data under the corresponding operating conditions to calculate the deviation value. For example, if the SOC of the energy storage battery decreases significantly faster than historical levels under the same charging and discharging power, or if the line loss rate of the line continues to increase under the same load, it indicates that the equipment may have hidden problems such as capacity decay, poor contact, or line aging.

[0276] Furthermore, continuous time window analysis can be performed on the deviation values ​​to not only identify single-point anomalies but also determine whether the deviation has a continuous increasing trend. For example, if an energy storage unit experiences a decrease in charging and discharging efficiency for several consecutive days, and the decrease exceeds a set threshold, it is judged as a risk of capacity degradation; if a line experiences a gradual increase in voltage drop and a continuous increase in line loss rate under similar load conditions, it is judged as a risk of line aging or abnormal contact.

[0277] Finally, tiered early warnings can be issued based on the magnitude, duration, and rate of increase of the deviation. A general warning is issued when the deviation exceeds the first-level threshold; a severe warning is issued when the deviation continues to increase and exceeds the second-level threshold, triggering backup control strategies such as reducing the charging and discharging load of the energy storage unit, adjusting photovoltaic output, or shifting some load. In this way, abnormal trends can be detected before a visible equipment failure occurs, enabling early warning.

[0278] 3. Adaptive handling strategy: Pre-set backup control strategies for different fault types: Photovoltaic inverter failure: Automatically switch to backup inverter, or adjust the output of surrounding photovoltaic units to make up for the power shortfall; Energy storage system failure: Reduce the charging and discharging load of the energy storage unit and call on other energy storage units to cooperate; Line overload: Reduce the photovoltaic output within the power supply range of the overloaded line, or guide the load to other lines; Communication interruption: The edge control node switches to local autonomous mode and maintains basic control based on the lightweight model.

[0279] Furthermore, when multiple faults are detected, instead of simply executing all backup strategies simultaneously, the faults are first prioritized based on their severity, scope of impact, and threat level to power grid security, and then executed in a tiered and coordinated manner. Specifically, faults can be categorized into three levels: safety-level, stability-level, and economic-level.

[0280] Safety-level faults have the highest priority, such as severe line overload, voltage exceeding limits, inverter overcurrent, and energy storage thermal runaway. These faults may directly affect equipment safety and power supply safety, and protective control measures can be implemented immediately, such as reducing photovoltaic output in the relevant area, disconnecting faulty equipment, and limiting the charging and discharging power of energy storage.

[0281] Stability-level faults have the next lowest priority, such as communication interruptions, local voltage fluctuations, and abnormal energy storage SOC. These faults mainly affect the stable operation of the distribution network and will switch to local autonomous mode or adjust local control strategies after ensuring that safety-level faults are handled.

[0282] Economic-grade faults have the lowest priority, such as slight increases in curtailment rate, high line loss rate, and excessive number of energy storage cycles. These issues mainly affect operational efficiency and will be optimized and adjusted after safety and stability are guaranteed.

[0283] If there is no conflict between multiple faults, the corresponding strategies can be executed in parallel; if there is a conflict between multiple strategies, the decision is made according to the principle of "grid safety first, power supply reliability second, and economic optimization last". For example, when line overload and increased curtailment rate occur at the same time, the line load is reduced first, and even if curtailment increases in a short period of time, the safety of the line and equipment must be ensured first.

[0284] Record the time, type, handling process and results of the fault, and upload them to the cloud management center for fine-tuning of the large model. The monthly fault diagnosis accuracy rate is improved by ≥3%.

[0285] In this embodiment, the present invention provides a distribution network control method based on a lightweight model. Its advantages are as follows: the present invention is applicable to edge control nodes that communicate with terminal sensing devices and a cloud-based control center. The method includes: acquiring multi-source heterogeneous data based on the terminal sensing devices; extracting multi-source feature data from the multi-source heterogeneous data and transmitting the multi-source feature data to the cloud-based control center, so that the cloud-based control center can generate and feed back model update information based on the multi-source feature data; updating a preset control model according to the model update information to obtain an updated control model; calling the updated control model to determine a control strategy based on the multi-source heterogeneous data, and performing control processing on the distribution network based on the control strategy. The present invention can perform control through edge control nodes, which can shorten the control data link, improve data transmission efficiency, and increase the control rate. Furthermore, by updating the model in the cloud before control, it can adapt to different operating conditions of the distribution network, reduce control deviations, and improve control accuracy.

[0286] This invention also provides a distribution network control device based on a lightweight model, see [link to relevant documentation]. Figure 3 The diagram shows a schematic of the structure of a power distribution network control device based on a lightweight model according to an embodiment of the present invention.

[0287] The device is suitable for edge control nodes that communicate with terminal sensing devices and cloud control centers respectively. As an example, the distribution network control device based on a lightweight model may include: The acquisition module 201 is used to acquire multi-source heterogeneous data based on the terminal sensing device, wherein the multi-source heterogeneous data is data on the operating status, topology and environment of the power distribution network; Extraction module 202 is used to extract multi-source feature data from the multi-source heterogeneous data and transmit the multi-source feature data to the cloud management center, so that the cloud management center can generate and feed back model update information based on the multi-source feature data; The update module 203 is used to update the preset control model according to the model update information to obtain the updated control model, wherein the preset control model is a lightweight cloud model trained by the cloud control center based on historical data of the distribution network and then deployed. The control module 204 is used to call the updated control model to determine the control strategy based on the multi-source heterogeneous data, and to perform control processing on the distribution network based on the control strategy.

[0288] Optionally, acquiring multi-source heterogeneous data based on the terminal sensing device includes: Multi-source data is acquired from terminal sensing devices, including meteorological data, power data, topology data, and device data. The multi-source data is subjected to feature standardization and optimization processing to obtain optimized data, wherein the optimization processing includes: missing data completion, feature normalization processing and boundary standardization processing; After constructing the topological feature encoding of the optimized data and completing the consistency verification, the optimized data is converted into a standardized feature matrix of a unified dimension to obtain multi-source heterogeneous data.

[0289] Optionally, the step of extracting multi-source feature data from the multi-source heterogeneous data includes: The multi-source feature data is converted into a multi-dimensional joint feature matrix, and a fully connected mapping layer is used to project the multi-dimensional joint feature matrix onto a unified feature space to obtain unified hidden features. Spatial enhancement features and temporal enhancement features are extracted from the unified hidden features, and the spatial enhancement features and temporal enhancement features are fused to obtain fused enhancement features; A preset fusion network is invoked to extract multi-dimensional cross features from the fusion enhancement features to obtain multi-source feature data, wherein the multi-dimensional cross features are the cross relationship between spatial enhancement features and temporal enhancement features in the fusion enhancement features.

[0290] Optionally, the step of calling the updated control model to determine the control strategy based on the multi-source heterogeneous data, and performing control processing on the distribution network based on the control strategy, includes: The updated control model is invoked to perform photovoltaic power output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic power output prediction parameters, wherein the photovoltaic power output prediction parameters include: short-time photovoltaic power output prediction data and medium- and long-term photovoltaic power output prediction data. A multi-objective optimization function is constructed using the photovoltaic output prediction parameters and the real-time detection parameters of the distribution network, wherein the real-time detection parameters include: real-time operating parameters of the power grid and the status of the energy storage system; The multi-objective optimization function is solved based on preset constraints to obtain the control strategy. The preset constraints include: photovoltaic output constraints, energy storage remaining power constraints, energy storage power constraints, voltage constraints, and line load constraints. The control strategy is converted into control instructions, and the terminal sensing device is controlled to perform corresponding adjustments through the control instructions.

[0291] Optionally, the step of calling the updated control model to perform photovoltaic output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic output prediction parameters includes: The update control model is used to divide the multi-source heterogeneous data into short-term feature data and medium-to-long-term feature data based on a preset time scale; The feature importance of the short-term feature data and the medium-to-long-term feature data is determined based on the attention mechanism. The feature importance is then used to reorganize the short-term feature data and the medium-to-long-term feature data to obtain short-term reorganized data and medium-to-long-term reorganized data. The updated control model is invoked to perform photovoltaic power output prediction processing based on the short-term and medium-to-long-term reorganization data to obtain photovoltaic power output prediction parameters.

[0292] Optionally, the device further includes: The information module is used to obtain early warning information after the steps of extracting multi-source feature data from the multi-source heterogeneous data and transmitting the multi-source feature data to the cloud management and control center. The early warning information is generated by the cloud management and control center after calling the built-in hybrid model to determine the fault type of the distribution network based on the multi-source feature data. The early warning module is used to invoke a preset backup control strategy to perform early warning adjustment processing on the distribution network based on the fault type corresponding to the early warning information.

[0293] This invention also provides a distribution network control system based on a lightweight model, see [link to relevant documentation]. Figure 3 The diagram shows a schematic representation of a distribution network control system based on a lightweight model, according to an embodiment of the present invention.

[0294] As an example, the distribution network control system based on the lightweight model may include: terminal sensing devices, edge control nodes, and a cloud control center, wherein the edge control nodes are communicatively connected to the cloud control center and the terminal sensing devices, respectively. The edge control node executes the distribution network control method based on the lightweight model as described in the above embodiments.

[0295] The cloud-based control center is deployed on a cloud computing platform, and its hardware includes a high-performance server cluster (CPUs are Intel Xeon Platinum series, GPUs are NVIDIA A100), a data storage array (supporting PB-level data storage, with read and write speeds ≥10GB / s), and a security encryption module.

[0296] Specifically, the core functions of the cloud-based management and control center are as follows: 1. Store historical data of the entire low-voltage distribution network (photovoltaic output, load data, meteorological data, equipment status data) for a time span of ≥3 years; 2. Run the complete large model (pre-training + fine-tuning), responsible for global data training, model iteration and update, and cross-regional collaborative optimization; 3. Receive real-time data uploaded by edge control nodes and issue global control strategies (such as cross-regional energy storage dispatch instructions). 4. Monitor the system's operating status and generate operation and maintenance reports and fault warnings.

[0297] Edge control nodes can be deployed in the control boxes of transformer substations or photovoltaic power stations. The hardware adopts an industrial-grade edge computing gateway (CPU adopts ARM Cortex-A72, memory ≥4GB, storage ≥32GB), and integrates data acquisition interfaces (RS485, Ethernet, LoRa) and communication modules (PLC, 5G).

[0298] The core functions of the edge control node are as follows: 1. Collect real-time data from terminal devices (sampling frequency ≥ 1Hz) and perform data preprocessing (filtering, deduplication, format conversion); 2. Deploy a lightweight large model inference module (compress cloud model parameters to less than 1 / 5 of their original size through knowledge distillation) to achieve local real-time control (such as active / reactive power regulation of photovoltaic inverters and local energy storage scheduling). 3. When communication with the cloud control center is interrupted, it automatically switches to "local self-governance mode" to maintain basic system operation based on preset rules and a lightweight model; 4. Upload pre-processed real-time data and local control logs to the cloud control center, and receive model update packages and global policies issued by the cloud control center.

[0299] In one embodiment, the terminal sensing device may include a photovoltaic inverter (supporting adjustable active / reactive power, with communication protocol compatible with GB / T 27930-2023), an energy storage system (equipped with a battery management system (BMS) that supports real-time SOC monitoring and charge / discharge control), a smart meter (supporting load data acquisition with a sampling frequency of ≥15 minutes / time), and a meteorological monitoring terminal (collecting solar irradiance, temperature, humidity, and wind speed with a sampling frequency of ≥1 minute / time).

[0300] The core functions of the terminal sensing device are as follows: 1. Collect equipment operation data and environmental data in real time and upload them to the system through edge control nodes; 2. Receive control commands from edge control nodes and execute operations such as photovoltaic power adjustment and energy storage charging and discharging; 3. Possess local fault detection capabilities (such as overcurrent and overvoltage protection) and promptly report fault information to edge control nodes.

[0301] In one embodiment, the data interaction process between the edge sensing device, the edge control node, and the cloud control center is as follows: 1. The terminal sensing device transmits real-time data to the edge control node via RS485 / Ethernet / LoRa, with a data transmission latency of ≤10ms; 2. After preprocessing the data, the edge control node uses it for local large model inference to generate control commands and send them to the terminal device. On the other hand, it uploads data to the cloud control center via PLC / 5G. The upload cycle is configurable (real-time data ≤ 1 minute / time, historical data ≤ 1 hour / time). 3. The cloud-based control center trains a large model based on data from the entire region, generates global control strategies and model update packages, and distributes them to edge control nodes via 5G; 4. After receiving instructions from the cloud control center, the edge control node optimizes the control strategy based on local operating conditions to achieve collaborative decision-making of "global cloud optimization + local edge adaptation".

[0302] This invention adopts a large model development paradigm of "pre-training + fine-tuning". The basic model is pre-trained based on power system mechanism data and public datasets, and then fine-tuned using low-voltage distribution network data of specific regions to adapt to the target scenario.

[0303] Reference Figure 4 The diagram shows a schematic structural diagram of a large model of a cloud-based management and control center provided in an embodiment of the present invention.

[0304] The core modules of the large model include a multi-dimensional feature fusion module, a photovoltaic power output prediction module, a multi-objective optimization scheduling module, and a fault diagnosis module, the specific design of which is as follows: Figure 4 As shown.

[0305] The power system mechanism is embedded into the large model primarily through four methods: mechanism constraint embedding, topology encoding, physical parameter characterization, and loss function constraints. Specifically, firstly, physical parameters such as node connections, line impedance, transformer capacity, photovoltaic inverter capacity, and energy storage SOC range of the low-voltage distribution network are transformed into structured features and used as part of the large model input. Secondly, the distribution network topology is represented as a graph structure, with nodes representing users, photovoltaic systems, energy storage, or transformers, and edges representing line connections. Line impedance, voltage, current, and power flow direction are input as edge features into the model. Thirdly, power flow calculation results are introduced during model training and inference as supervisory or corrective information, ensuring that the model's output control strategy satisfies fundamental physical laws such as voltage, current, and power balance. Finally, penalty terms for voltage exceeding limits, line overload, power imbalance, and energy storage SOC exceeding limits are added to the loss function to constrain the model from generating control commands that violate the power system's operational safety boundaries.

[0306] In this way, the large model does not simply rely on data fitting, but learns from historical operating patterns while being constrained by power flow constraints, topological constraints, and equipment physical characteristics of the distribution network, thereby improving the interpretability, security, and engineering feasibility of the control results.

[0307] The control operation of the model can refer to the distribution network control method based on the lightweight model described in the above embodiments.

[0308] Reference Figure 5 The diagram illustrates the interaction flowchart between the edge control node and the cloud control center provided in an embodiment of the present invention.

[0309] To balance the need for real-time control with the requirements of global optimization, this invention adopts a deployment strategy of "lightweight edge + complete cloud" and designs a dynamic iteration mechanism, such as... Figure 5 As shown.

[0310] The large, fully-fledged model (parameter scale ≥ 100 million) trained in the cloud control center is lightweighted using a combination of knowledge distillation, quantization compression, and pruning techniques. 1. Knowledge Distillation: Using a complete large model as the "teacher model," train "student models" with a parameter scale of ≤20 million, enabling the student models to learn the reasoning logic of the teacher model. After lightweighting, the model's accuracy retention rate is ≥95%. 2. Quantization and Compression: The model parameters are quantized from 32-bit floating-point numbers (FP32) to 16-bit floating-point numbers (FP16) or 8-bit integers (INT8), reducing the model storage volume by more than 75% and increasing the inference speed by more than 3 times; 3. Pruning: Remove redundant neurons and connections in the model while retaining the core inference path. After pruning, the computational load of the model is reduced by more than 60%, and it is adapted to the computational power of edge control nodes (it can run smoothly on ARM Cortex-A72).

[0311] In one embodiment, edge control nodes and cloud control centers can perform collaborative reasoning.

[0312] 1. Edge Inference: The lightweight model is deployed at the edge control node and is responsible for real-time control tasks, such as short-term photovoltaic output prediction (5min - 1h), local energy storage scheduling, and preliminary equipment fault diagnosis. The inference latency is ≤30ms, which meets the real-time requirements. 2. Cloud-based inference: The complete model is deployed in the cloud control center, responsible for medium- and long-term forecasting (1 day to 7 days), cross-regional collaborative scheduling (such as inter-regional energy storage complementarity), and complex fault diagnosis (such as multi-device linkage failure). The inference results are distributed to the edge control nodes via 5G. 3. Collaborative Decision Making: Edge control nodes combine local inference results with instructions from the cloud control center to optimize control strategies. For example, after the cloud control center issues a cross-regional energy storage dispatch plan, the edge control nodes adjust the charging and discharging power according to the actual state of their local energy storage SOC to ensure the feasibility of the strategy.

[0313] In another alternative embodiment, since the edge control node deploys a lightweight model, after receiving the data uploaded by the terminal, it refers to... Figure 5 Edge control nodes can generate control strategies based on data uploaded by terminals, and then implement control strategies accordingly, thereby improving control efficiency.

[0314] This invention constructs a closed-loop control architecture of "data perception - large model analysis - intelligent decision-making - execution feedback". Based on a pre-trained large model (integrating power system mechanisms and deep learning), it realizes integrated optimization of photovoltaic output prediction, load matching, energy storage scheduling and grid status assessment, breaking through the limitations of traditional modular control and improving the overall system synergy.

[0315] By embedding a spatiotemporal attention mechanism and a meteorological-electric cross-feature extraction module within a large model framework, the system integrates meteorological data such as solar irradiance, cloud movement speed, temperature and humidity, as well as power data such as historical photovoltaic output and grid load fluctuations in real time. This solves the problem of low photovoltaic output prediction accuracy on short-term (5min-2h) and medium-to-long-term (1-7 days) scales, reducing the prediction error rate by more than 30% compared to traditional models.

[0316] Furthermore, by employing model lightweighting techniques (knowledge distillation and quantization compression), the core inference module of the large model is deployed on edge nodes (such as distribution concentrators and local controllers of photovoltaic inverters) to achieve millisecond-level local control response; the complete large model is deployed in the cloud to be responsible for global data training, model iterative updates, and cross-distribution collaborative optimization, balancing real-time performance with global optimization requirements.

[0317] During regulation, based on the photovoltaic output prediction results output by the large model, the real-time operating parameters of the power grid (voltage, frequency, power flow distribution) and the status of the energy storage system (SOC, charging and discharging efficiency), a multi-objective optimization function is constructed with the goal of "minimizing curtailment rate, minimizing grid loss, and maximizing energy storage life". The active / reactive power regulation strategy of the photovoltaic inverter and the energy storage charging and discharging plan are dynamically generated through the reinforcement learning module built into the large model to adapt to the complex operating conditions of the low-voltage distribution network.

[0318] In addition, the large model continuously learns the characteristic data of common faults in low-voltage distribution networks (such as photovoltaic inverter malfunctions, energy storage battery failures, and line overloads) to establish a fault diagnosis knowledge base. When a grid anomaly is detected, it automatically triggers backup control strategies, such as switching backup photovoltaic units and adjusting the energy storage charging and discharging mode, to ensure the stability of system operation.

[0319] Those skilled in the art will understand that, for ease of description and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0320] Furthermore, this application also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the power distribution network control method based on the lightweight model as described in the above embodiments.

[0321] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer-executable program, which is used to cause a computer to execute the power distribution network control method based on the lightweight model as described in the above embodiments.

[0322] In the description of the embodiments of the present invention, it should be noted that the terms "above," "below," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. When an element such as a layer, region, or substrate is referred to as being "above" or "on top of" another element, it may be directly on the other element, or there may be an intermediate element. Conversely, when an element is referred to as being "directly on" or "above" another element, there is no intermediate element. It should also be understood that when an element is referred to as being "below" or "under" another element, it may be directly below or under the other element, or there may be an intermediate element. Conversely, when an element is referred to as being "directly below" or "under" another element, there is no intermediate element. Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0323] Those skilled in the art will understand that embodiments of this application may also include computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0324] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), devices, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0325] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0326] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0327] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A distribution network control method based on a lightweight model, characterized in that, The method is applicable to edge control nodes that communicate with both terminal sensing devices and cloud-based control centers, and the method includes: Multi-source heterogeneous data is acquired based on terminal sensing devices, wherein the multi-source heterogeneous data is data on the operating status, topology, and environment of the power distribution network; Multi-source feature data is extracted from the multi-source heterogeneous data and transmitted to the cloud management center so that the cloud management center can generate and feed back model update information based on the multi-source feature data; The preset control model is updated according to the model update information to obtain the updated control model. The preset control model is a lightweight cloud model trained by the cloud control center based on historical data of the distribution network and then deployed. The updated control model is invoked to determine the control strategy based on the multi-source heterogeneous data, and the distribution network is controlled based on the control strategy.

2. The distribution network control method based on a lightweight model according to claim 1, characterized in that, The acquisition of multi-source heterogeneous data based on terminal sensing devices includes: Multi-source data is acquired from terminal sensing devices, including meteorological data, power data, topology data, and device data. The multi-source data is subjected to feature standardization and optimization processing to obtain optimized data, wherein the optimization processing includes: missing data completion, feature normalization processing and boundary standardization processing; After constructing the topological feature encoding of the optimized data and completing the consistency verification, the optimized data is converted into a standardized feature matrix of a unified dimension to obtain multi-source heterogeneous data.

3. The distribution network control method based on a lightweight model according to claim 1, characterized in that, The step of extracting multi-source feature data from the multi-source heterogeneous data includes: The multi-source feature data is converted into a multi-dimensional joint feature matrix, and a fully connected mapping layer is used to project the multi-dimensional joint feature matrix onto a unified feature space to obtain unified hidden features. Spatial enhancement features and temporal enhancement features are extracted from the unified hidden features, and the spatial enhancement features and temporal enhancement features are fused to obtain fused enhancement features; A preset fusion network is invoked to extract multi-dimensional cross features from the fusion enhancement features to obtain multi-source feature data, wherein the multi-dimensional cross features are the cross relationship between spatial enhancement features and temporal enhancement features in the fusion enhancement features.

4. The distribution network control method based on a lightweight model according to claim 1, characterized in that, The step of calling the updated control model to determine the control strategy based on the multi-source heterogeneous data, and performing control processing on the distribution network based on the control strategy, includes: The updated control model is invoked to perform photovoltaic power output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic power output prediction parameters, wherein the photovoltaic power output prediction parameters include: short-time photovoltaic power output prediction data and medium- and long-term photovoltaic power output prediction data. A multi-objective optimization function is constructed using the photovoltaic output prediction parameters and the real-time detection parameters of the distribution network, wherein the real-time detection parameters include: real-time operating parameters of the power grid and the status of the energy storage system; The multi-objective optimization function is solved based on preset constraints to obtain the control strategy. The preset constraints include: photovoltaic output constraints, energy storage remaining power constraints, energy storage power constraints, voltage constraints, and line load constraints. The control strategy is converted into control instructions, and the terminal sensing device is controlled to perform corresponding adjustments through the control instructions.

5. The distribution network control method based on a lightweight model according to claim 4, characterized in that, The step of calling the updated control model to perform photovoltaic output prediction processing based on the multi-source heterogeneous data to obtain photovoltaic output prediction parameters includes: The update control model is used to divide the multi-source heterogeneous data into short-term feature data and medium-to-long-term feature data based on a preset time scale; The feature importance of the short-term feature data and the medium-to-long-term feature data is determined based on the attention mechanism. The feature importance is then used to reorganize the short-term feature data and the medium-to-long-term feature data to obtain short-term reorganized data and medium-to-long-term reorganized data. The updated control model is invoked to perform photovoltaic power output prediction processing based on the short-term and medium-to-long-term reorganization data to obtain photovoltaic power output prediction parameters.

6. The distribution network control method based on a lightweight model according to any one of claims 1-5, characterized in that, After the steps of extracting multi-source feature data from the multi-source heterogeneous data and transmitting the multi-source feature data to the cloud management center, the method further includes: Obtain early warning information, which is generated by the cloud-based control center after calling the built-in hybrid model to determine the fault type of the distribution network based on the multi-source feature data; Based on the fault type corresponding to the early warning information, a preset backup control strategy is invoked to perform early warning adjustment processing on the distribution network.

7. A distribution network control device based on a lightweight model, characterized in that, The device is suitable for edge control nodes that communicate with terminal sensing devices and cloud control centers respectively, and the device includes: The acquisition module is used to acquire multi-source heterogeneous data based on terminal sensing devices, wherein the multi-source heterogeneous data is data on the operating status, topology, and environment of the power distribution network; The extraction module is used to extract multi-source feature data from the multi-source heterogeneous data and transmit the multi-source feature data to the cloud management center, so that the cloud management center can generate and feed back model update information based on the multi-source feature data; The update module is used to update the preset control model according to the model update information to obtain the updated control model. The preset control model is a lightweight cloud model trained by the cloud control center based on historical data of the distribution network and then deployed. The control module is used to call the updated control model to determine the control strategy based on the multi-source heterogeneous data, and to perform control processing on the distribution network based on the control strategy.

8. A distribution network control system based on a lightweight model, characterized in that, The system includes: a terminal sensing device, an edge control node, and a cloud control center, wherein the edge control node is communicatively connected to the cloud control center and the terminal sensing device, respectively. The edge control node executes the distribution network control method based on the lightweight model as described in any one of claims 1-6.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the power distribution network control method based on a lightweight model as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the power distribution network control method based on a lightweight model as described in any one of claims 1-6.

Images