Power grid dynamic reactive power cooperative control system and method based on multi-source data fusion

Abstract

The invention discloses a power grid dynamic reactive power cooperative control system and method based on multi-source data fusion, and the system comprises a cloud cooperative control platform and a plurality of edge autonomous control nodes, and the cloud cooperative control platform collects multi-source heterogeneous data of a main network, a distribution network, distributed energy and a load side through a multi-source data sensing module. A power grid panoramic state feature tensor is generated through a data fusion module, a power grid digital twin module constructs and dynamically updates a power grid digital twin based on the feature tensor, and a centralized optimization decision module performs collaborative optimization calculation based on a digital twin model to generate an adjustment instruction of reactive power equipment in each region; and the edge autonomous control node executes a cloud instruction when the communication is normal, and performs local autonomous control when the communication is abnormal. According to the method, power grid panoramic perception is realized through deep fusion of multi-source data, the system can perform local adjustment when communication is abnormal based on the cloud edge collaborative architecture, and the globality, the real-time performance and the reliability of power grid reactive power control are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system collaborative control, in particular to a power grid dynamic reactive power collaborative control system and method based on multi-source data fusion. BACKGROUND

[0002] With the promotion of the "double carbon" goal, the penetration rate of new energy represented by wind power and photovoltaic power in the power grid is continuously increasing, and the operation characteristics of the power grid are showing new features. In the field of power system operation control technology, the existing power grid reactive power and voltage control system mainly adopts a hierarchical and partitioned architecture. The main grid automatic voltage control system relies on traditional data sources such as SCADA, acquires voltage, reactive power and other parameters of key nodes in the power grid, and executes control strategies based on rules or optimization algorithms to adjust the reactive power compensation devices in the main grid substation. The distribution network side acquires distribution network operation data through the distribution automation system and independently controls the reactive power devices of the distribution network lines and transformer areas. At the same time, the distributed energy system also has local reactive power control function, which adjusts the reactive power output of the inverter by monitoring the voltage state of the access point. These systems realize basic control of reactive power and voltage in their respective ranges, forming a multi-level reactive power and voltage control system from the main grid to the distribution network and then to the distributed energy system.

[0003] However, under the background of high proportion of new energy access, real-time perception and collaborative control of power grid operation state face new challenges, and coordination between multi-level control systems cannot meet the demand for voltage stability and optimal operation in complex power grid environment. SUMMARY

[0004] To solve the above problems, the present application provides a power grid dynamic reactive power collaborative control system based on multi-source data fusion, which comprises a cloud collaborative control platform and a plurality of edge autonomous control nodes in communication connection with the cloud collaborative control platform. The cloud collaborative control platform is used for global reactive power collaborative optimization control across regions and levels, comprising: A multi-source data perception module acquires multi-source heterogeneous data from the main grid dispatching side, the distribution network side, the distributed energy side and the load side in real time through a standardized interface. A data fusion module performs synchronization, outlier cleaning and spatio-temporal coordinate alignment operations on the multi-source heterogeneous data through a multi-source data fusion engine, and maps the aligned data into a unified dimension of power grid panoramic state feature tensor through a feature extraction network. A power grid digital twin module constructs and dynamically updates a power grid digital twin based on the power grid panoramic state feature tensor. The power grid digital twin contains power grid topology and electrical parameters, and is used to simulate the dynamic response characteristics of reactive power resources. The centralized optimization decision module, based on the power grid digital twin model, performs collaborative optimization calculations with a preset global optimization objective, generates a reactive resource allocation strategy benchmark for equipment in different regions, and generates adjustment instructions for reactive power equipment in each region based on the reactive resource allocation strategy benchmark, and sends the adjustment instructions to the corresponding edge autonomous control nodes. The edge autonomous control node, deployed in a designated area or site of the power grid, is used for reactive power autonomous control within its region, including: The local data processing module collects and processes real-time operational data for the local area. The edge autonomous decision-making module receives adjustment instructions for reactive power equipment in the local area when the cloud-based collaborative control platform is in normal communication mode; when communication with the cloud-based collaborative control platform is abnormal, it generates adjustment instructions for reactive power equipment in the local area based on locally collected data and preset autonomous control strategies. The instruction execution driver module receives and decomposes the adjustment instructions from the edge autonomous decision-making module, and then sends the instructions to the corresponding reactive power devices in this area.

[0005] Furthermore, in the data fusion module, the feature extraction network is a hybrid neural network composed of a cascaded temporal convolutional network and a self-attention mechanism, wherein: The temporal convolutional network contains three layers of dilated causal convolutional layers, used to extract local temporal dependency features of each data source in the time dimension; The self-attention mechanism receives multi-source temporal feature sequences output by a temporal convolutional network, models the spatiotemporal coupling relationship between different data sources through a learnable cross-source association weight matrix, and outputs the power grid panoramic state feature tensor.

[0006] Furthermore, in the power grid digital twin module, the dynamic update mechanism of the power grid digital twin includes: Every T1 cycle, the newly generated power grid panoramic state feature tensor is input into the power grid digital twin; Based on the real-time topology status and switching action information in the power grid panoramic state feature tensor, changes in the power grid topology are identified, and the network connection relationships and operating topology of the power grid digital twin are updated synchronously. Based on the real-time power output and electrical quantity measurement data of the equipment in the power grid panoramic state feature tensor, an online parameter identification algorithm is used to update the equivalent impedance model and dynamic response delay parameters of each reactive power source and reactive power compensation device in the power grid digital twin in real time. When a load change is detected to be greater than a preset threshold, incremental model retraining is triggered. By utilizing historical operating data within a sliding time window, the power flow calculation kernel parameters of the power grid digital twin are adaptively adjusted and calibrated.

[0007] Furthermore, in the centralized optimization decision-making module, the collaborative optimization calculation employs the multi-timescale collaborative optimization algorithm, which includes: The first optimization layer performs minute-level global rolling optimization: with the comprehensive optimization objectives of reducing network losses, improving node voltage qualification rate, and reducing the total number of reactive power equipment operations, the model predictive control method is used to perform periodic rolling solutions based on the power grid digital twin and ultra-short-term forecast data to generate a reactive power resource allocation strategy benchmark. The second optimization layer executes second-level local trigger control: when the real-time voltage deviation is detected to exceed the preset threshold, a pre-trained multi-agent deep reinforcement learning model is launched. The multi-agent deep reinforcement learning model takes the distribution network feeder area as the agent unit. Each agent takes the local and adjacent areas' grid panoramic state feature tensor as the joint state input, takes the adjustment of local reactive power equipment as the action space, and takes minimizing the regional voltage deviation integral, reducing cross-regional reactive power interaction penalties, and reducing equipment action costs as the composite reward function to make collaborative decisions and output the adjustment instructions for reactive power equipment in each area.

[0008] Another aspect of the present invention provides a dynamic reactive power coordinated control method for power grids based on multi-source data fusion, wherein the method is applied to the control system described in any one of the above claims, characterized in that it includes: Cloud-based collaborative control platform execution: S1.1. Collect multi-source heterogeneous operation data in real time from the main grid dispatching system, distribution network automation system, distributed energy monitoring system and smart electricity terminals through standardized communication interfaces; S1.2. The data fusion module performs time synchronization, outlier identification and cleaning, and spatiotemporal coordinate alignment on the multi-source heterogeneous operating data, and generates a unified-dimensional power grid panoramic state feature tensor through the feature extraction network. S1.3 Based on the power grid panoramic state feature tensor, and combined with the power grid topology and electrical parameters, construct and dynamically update the power grid digital twin; S1.4. Through the centralized optimization decision module, a multi-time-scale collaborative optimization algorithm is used to perform collaborative optimization calculations: At the minute-level optimization layer, based on the power grid digital twin and ultra-short-term forecast data, a reactive power resource allocation strategy benchmark for the next N control cycles is generated. At the second-level triggering layer, based on the reactive resource allocation strategy benchmark, the pre-trained multi-agent deep reinforcement learning model is invoked to output the adjustment instructions for reactive equipment in each region. After the adjustment commands of reactive power equipment in each area are verified for safety and encapsulated, they are sent to the edge autonomous control nodes of the corresponding partitions. Execution of edge autonomous control nodes: S2.1. Collect local operating data in real time, including bus voltage, line power flow, reactive power equipment operating status and switch position, and perform preprocessing operations. S2.2 Real-time monitoring of the link quality between the cloud-based collaborative control platform and the current communication status is determined as "normal" or "abnormal" based on RTT latency, packet loss rate, and acknowledgment success rate. S2.3, Perform dual-mode switching control through the edge autonomous decision-making module: When the communication status is "normal", it receives adjustment instructions from the cloud and sends them to the instruction execution driver module; When the communication status is "interrupted", switch to local autonomous control mode, load the preset autonomous control strategy, and solve the reactive power equipment adjustment command of this area that meets the voltage qualification rate and equipment safety constraints based on real-time voltage deviation, reactive power deficit and equipment action priority. S2.4 The instruction execution driver module decomposes the received regional adjustment instructions, sorts them according to priority, and sends them to drive the corresponding devices to perform reactive power adjustment.

[0009] The beneficial effects of this application are: This application provides a dynamic reactive power coordinated control system for power grids based on multi-source data fusion. The system deploys a cloud-based control center for cross-regional and cross-level global reactive power coordinated optimization control. It receives multi-source heterogeneous data from the cloud, and through a feature extraction network, maps the aligned heterogeneous data into a unified-dimensional panoramic state feature tensor of the power grid, constructing a digital twin of the power grid. A centralized optimization decision module generates adjustment commands for reactive power equipment in each region and distributes these commands to the corresponding edge autonomous control nodes, avoiding control command conflicts that may occur in traditional hierarchical control systems. Edge autonomous control nodes are deployed in designated power grid zones or sites to perform autonomous reactive power control within their respective regions. Under normal communication with the cloud-based coordinated control platform, these nodes receive adjustment commands for reactive power equipment in their region. When communication with the cloud-based coordinated control platform fails, they generate adjustment commands for reactive power equipment in their region based on locally collected data and a preset autonomous control strategy, and distribute these commands to the corresponding reactive power equipment within the region. This ensures the continuity of control during communication failures and improves the system's reliability and robustness. Through the collaborative work of the cloud-based collaborative control platform and the edge autonomous control nodes, unified and coordinated control of reactive power resources across the entire network, from the main grid to the distribution network and then to distributed energy, has been achieved, effectively solving the problem of grid voltage stability and optimized operation under the background of high proportion of new energy access. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall structure of the dynamic reactive power coordinated control system for power grids based on multi-source data fusion provided in this application.

[0011] Figure 2 This is a schematic diagram of the cloud-based collaborative control platform provided in this application.

[0012] Figure 3 This is a schematic diagram of the edge autonomous control node provided in this application. Detailed Implementation

[0013] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0014] Example 1 Please see Figures 1 to 3 The power grid dynamic reactive power coordinated control system based on multi-source data fusion includes a cloud-based coordinated control platform and multiple edge autonomous control nodes that communicate with the cloud-based coordinated control platform. A cloud-based collaborative control platform is used for cross-regional and cross-level global reactive power collaborative optimization control, including: The multi-source data sensing module collects multi-source heterogeneous data from the main grid dispatch side, distribution network side, distributed energy side and load side in real time through standardized interfaces; The data fusion module performs synchronization, outlier cleaning, and spatiotemporal coordinate alignment operations on multi-source heterogeneous data through a multi-source data fusion engine, and maps the aligned data into a unified-dimensional power grid panoramic state feature tensor through a feature extraction network. The power grid digital twin module constructs and dynamically updates a power grid digital twin based on the power grid panoramic state feature tensor. The power grid digital twin contains power grid topology and electrical parameters, which are used to simulate the dynamic response characteristics of reactive power resources. The centralized optimization decision module, based on the power grid digital twin model, performs collaborative optimization calculations with a preset global optimization objective, generates reactive resource allocation strategy benchmarks for equipment in different regions, and generates adjustment instructions for reactive equipment in each region based on the reactive resource allocation strategy benchmarks, and sends the adjustment instructions to the corresponding edge autonomous control nodes. Edge autonomous control nodes, deployed in designated power grid zones or sites, are used for reactive power autonomous control within their respective areas, including: The local data processing module collects and processes real-time operational data for the local area. The edge autonomous decision-making module receives adjustment instructions for reactive power equipment in the local area when the cloud-based collaborative control platform is in normal communication mode; when communication with the cloud-based collaborative control platform is abnormal, it generates adjustment instructions for reactive power equipment in the local area based on locally collected data and preset autonomous control strategies. The instruction execution driver module receives and decomposes the adjustment instructions from the edge autonomous decision-making module, and then distributes the instructions to the corresponding reactive power devices in this area.

[0015] Specifically, this embodiment constructs a hybrid elastic architecture of "cloud-edge" collaboration, achieving an organic unity of global optimization capabilities and local robust response capabilities through layered decoupling. The cloud-based collaborative control platform, acting as the global cognition and decision-making center, undertakes the task of coordinating reactive power resources across regions and levels; the edge autonomous control nodes, as local execution and emergency response units, possess independent closed-loop control capabilities. The two establish a highly reliable, low-latency bidirectional link through a standardized communication protocol, supporting command issuance, status feedback, and event reporting. This architecture avoids the single-point failure risk and communication bottlenecks of purely centralized systems while overcoming the shortcomings of purely distributed systems that lack a global perspective and collaborative constraints, providing a dynamic reactive power collaborative control paradigm for highly uncertain power grids that combines economy, security, and resilience.

[0016] As a preferred embodiment, in the data fusion module, the feature extraction network is a hybrid neural network composed of a temporal convolutional network and a self-attention mechanism cascaded together, wherein: Temporal convolutional networks consist of three layers of dilated causal convolutional layers, used to extract local temporal dependency features of each data source in the time dimension; The self-attention mechanism receives multi-source temporal feature sequences output by a temporal convolutional network, models the spatiotemporal coupling relationship between different data sources through a learnable cross-source association weight matrix, and outputs a panoramic state feature tensor of the power grid.

[0017] Specifically, the hybrid structure of cascaded temporal convolutional networks and self-attention mechanisms in this application can solve the technical problems of insufficient modeling capabilities for multi-source heterogeneous data in terms of temporal dynamics, nonlinear coupling between sources, and long-range dependencies, wherein: The temporal convolutional network comprises three dilated causal convolutional layers, each with a kernel size of 3. The first layer has a dilation rate of 1, the second 2, and the third 4, ensuring the receptive field covers at least 15 consecutive time steps. This allows it to capture the time scales corresponding to typical voltage fluctuation events. Normalization and a GELU activation function are applied after convolution, and residual connections are introduced to ensure training stability. Deployed in the data fusion module of a cloud-based collaborative control platform, the input is synchronized and cleaned temporal data streams from various sources, and the output is a sequence of independently encoded local temporal features from each data source.

[0018] The self-attention mechanism is a multi-head self-attention submodule based on the Transformer architecture. The input is a concatenated tensor of multi-source temporal feature sequences output by a temporal convolutional network. Specifically, after alignment by time step, feature vectors from four data sources—main grid, distribution grid, distributed energy, and load—are stacked along the channel dimension and jointly input. Linear projection generates three sets of vectors: query, key, and value. After scaling and dot product operations, a learnable cross-source correlation weight matrix is ​​calculated. This weight matrix explicitly models physically interpretable relationships such as the dynamic response lag between "voltage drop at a distribution grid feeder" and "reactive power output change of adjacent distributed photovoltaic inverters," and the coupling strength between "main grid bus frequency offset" and "reactive power regulation response delay of load-side intelligent terminals." After weighted summation and linear mapping, a unified-dimensional grid panoramic state feature tensor is output. This tensor retains the original temporal evolution trajectory of each source and embeds inter-source spatiotemporal dependency priors, which can be directly input into the grid digital twin module for state mapping. The temporal convolutional network and the self-attention mechanism are cascaded. The former is responsible for the deep extraction of the internal temporal features of a single source, while the latter is responsible for the explicit modeling of the external correlations of multiple sources. The output of the former is a necessary prerequisite for the input of the latter. The two cannot be interchanged in order, nor can they run in parallel or independently. At the bottom layer, it focuses on "how each source changes over time", and at the top layer, it focuses on "how each source influences each other". Thus, without relying on manually constructed coupling rules, it automatically learns the dynamic coupling law of "source-grid-load-storage" that is widely present in the operation of the power grid.

[0019] As a preferred embodiment, the dynamic update mechanism of the power grid digital twin module includes: Every T1 cycle, the newly generated power grid panoramic state feature tensor is input into the power grid digital twin; Based on the real-time topology status and switching action information in the power grid panoramic state feature tensor, changes in the power grid topology are identified, and the network connection relationships and operating topology of the power grid digital twin are updated synchronously. Based on the real-time power output and electrical quantity measurement data of equipment in the power grid panoramic state feature tensor, an online parameter identification algorithm is adopted to update the equivalent impedance model and dynamic response delay parameters of each reactive power source and reactive power compensation device in the power grid digital twin in real time. When a load change is detected to be greater than a preset threshold, incremental model retraining is triggered. Historical operating data within a sliding time window is used to adaptively adjust and calibrate the power flow calculation kernel parameters of the power grid digital twin.

[0020] Specifically, this application achieves this through the synergistic effect of four levels: periodic data injection, real-time topology mapping, online parameter correction, and kernel adaptive calibration. This enables the digital twin to not only statically reflect the initial network structure but also dynamically characterize typical time-varying behaviors such as switching operations, equipment aging, operating condition migration, and load jumps. This provides a low-latency, high-confidence simulation support environment for centralized optimization decision-making.

[0021] The specific update steps are as follows: The system sets a fixed time step update cycle, and every T1 cycle, it acquires the latest power grid panoramic state feature tensor output by the data fusion module. The switch quantity and topology identifier fields are extracted from the tensor and generated into the current topology adjacency table through logical parsing. Then, the graph structure update engine is called to compare with the topology snapshot of the previous period, identify change types such as new / broken branches, isolated areas, and closed ring networks, and reconstruct the network connection relationship graph and running topology layer inside the digital twin in real time.

[0022] Equivalent circuit models of each reactive power source and reactive power compensation device are established. Using time-series data such as port voltage, output current, active power output, and reactive power commands synchronously acquired in the tensor, dynamic response parameters are estimated online using the recursive least squares method. The updated parameters are directly written into the digital twin component library and participate in subsequent power flow calculations and transient simulations.

[0023] The system continuously monitors the total load power sequence of each feeder in the tensor and calculates its standard deviation within the sliding window. with the mean The ratio (i.e., coefficient of variation) When the CV exceeds the threshold θ, it is determined to be a load mutation event. At this time, a sliding time window of data of 900 seconds, from 600 seconds before the mutation to 300 seconds after the mutation, is automatically extracted as an incremental training set. Gradient descent fine-tuning is performed only on the 3-5 most sensitive parameters with the greatest impact in the power flow calculation kernel, rather than retraining the entire model. This process is completed in a lightweight training sandbox dedicated to the digital twin, with a single session taking less than 800 milliseconds, and convergence verification is automatically performed after training to be effective.

[0024] Through the above technical solutions, this application has achieved systematic modeling and adaptation of dynamic characteristics such as frequent switching of grid topology, time-varying reactive power equipment parameters, and random load jumps under high proportion of new energy access.

[0025] As a preferred embodiment, in the centralized optimization decision module, the collaborative optimization calculation adopts a multi-timescale collaborative optimization algorithm, which includes: The first optimization layer performs minute-level global rolling optimization: with the comprehensive optimization objectives of reducing network losses, improving node voltage qualification rate and reducing the total number of reactive power equipment operations, it uses model predictive control method to perform periodic rolling solutions based on the power grid digital twin and ultra-short-term forecast data to generate a reactive power resource allocation strategy benchmark. The second optimization layer executes second-level local trigger control: when the real-time voltage deviation is detected to exceed the preset threshold, a pre-trained multi-agent deep reinforcement learning model is launched. This multi-agent deep reinforcement learning model takes the distribution network feeder area as the agent unit. Each agent takes the local and adjacent areas' grid panoramic state feature tensor as the joint state input, adjusts the local reactive power equipment as the action space, and uses the composite reward function of minimizing the regional voltage deviation integral, reducing cross-regional reactive power interaction penalties, and reducing equipment action costs to make collaborative decisions and output adjustment instructions for reactive power equipment in each area.

[0026] Specifically, this application constructs a two-layer closed-loop decision-making mechanism of "steady-state economic optimization + transient adaptive regulation" by integrating the forward planning capability of model predictive control with the rapid response capability of multi-agent deep reinforcement learning.

[0027] The first optimization layer performs minute-level global rolling optimization with a fixed time window period and the current moment as the starting point. Based on the topology, component parameters, and dynamic response characteristics represented by the power grid digital twin, and combined with ultra-short-term power prediction data for the next 15–60 minutes, a constrained nonlinear optimization problem is constructed. The objective function of this problem is in the form of a weighted combination, with weight coefficients that can be adjusted online. Their physical meanings correspond to: reducing the active power loss of the entire network, improving the voltage qualification rate of key nodes, and reducing the total number of reactive power equipment operations. These three factors together constitute a comprehensive representation of the system's economy, safety, and equipment lifespan. The solution method adopts a model predictive control framework, that is, only the first control step in the optimization result is executed each time, and the latest state and prediction data are collected again in the next cycle. The optimization problem is updated and solved repeatedly in a rolling manner, thereby continuously adapting to the time-varying characteristics of the system. The output is a reactive power resource allocation strategy benchmark for all regions of the entire network.

[0028] The second optimization layer executes second-level local trigger control. It doesn't operate according to a fixed cycle, but rather uses real-time voltage deviation as the event-driven signal. When the absolute value of the difference between the measured voltage amplitude and the reference voltage at any key monitoring node continuously exceeds a preset threshold, it is determined to be a voltage over-limit event requiring emergency intervention, and this layer is immediately activated. Based on the reactive power resource allocation strategy benchmark, a pre-trained multi-agent deep reinforcement learning model is loaded. Each agent strictly corresponds to a geographically or logically defined distribution network feeder area. Its observation space is jointly composed of two parts: one is the overall state feature tensor of the power grid in this area, and the other is the corresponding feature tensor of adjacent feeder areas that have an electrical coupling relationship with this area, thus forming a time-sensitive... The system employs a joint state representation of spatial local sensing capabilities. The action space can be discrete or continuous, corresponding to the set of adjustment commands for all controllable reactive power devices within the region. The reward function is composite, consisting of a weighted sum of three terms: the first term is the regional voltage deviation integral, reflecting the voltage recovery speed and steady-state accuracy; the second term is the cross-regional reactive power interaction penalty term, defined as the sum of the absolute values ​​of the net reactive power transferred from this region to adjacent regions, used to suppress long-distance ineffective reactive power flow; the third term is the device action cost term, assigned different weights according to the device type to avoid frequent adjustments leading to device fatigue. The output is a directly executable second-level adjustment command, which does not need to wait for the next round of rolling optimization and can be immediately driven by the edge node after local parsing to drive the device response.

[0029] The reactive resource allocation strategy benchmark generated in the first optimization layer serves as the initial action preference and value function reference benchmark for each agent in the second optimization layer, ensuring that MARL decisions are always anchored near the global economic optimal trajectory and avoiding local aggressive adjustments that deviate from the overall system objective. The real-time control results of the second optimization layer serve as feedback signals, which are used to correct the ultra-short-term prediction error model and digital twin parameter identification module on which the first optimization layer relies, forming a two-way closed loop of "optimization guiding control and control verifying optimization".

[0030] As a preferred embodiment, the edge autonomous decision-making module is configured as follows: It receives reactive power equipment adjustment instructions from the centralized optimization decision module as the first instruction source, and receives real-time operating data from the local data processing module as the second instruction source. When communicating normally with the cloud-based collaborative control platform, the first instruction source is sent to the instruction execution driver module; When communication with the cloud-based collaborative control platform fails, the system switches to a local autonomous control mode based on a second command source. It loads a preset autonomous control strategy, which performs local reactive power optimization calculations based on the real-time voltage deviation, reactive power deficit, and preset equipment action priorities and voltage safety constraints of the region, generating adjustment commands for reactive power equipment in the region.

[0031] Specifically, the edge autonomous decision-making module is the core logic unit deployed in the edge autonomous control nodes of designated power grid zones or sites, used to achieve dynamic trade-offs and seamless switching between "centralized guidance and distributed autonomy". This module constructs a disaster-tolerant local closed-loop control capability through a triple structure: a dual-instruction source input mechanism, a communication status awareness mechanism, and a strategy loading and local optimization solution mechanism. When communication is reliable, it strictly adheres to the global collaborative strategy, ensuring the overall network economy and consistency. Under abnormal conditions such as communication interruptions or timeouts, it can independently complete reactive power regulation decisions that conform to the safety boundaries of the local area without relying on the cloud, thereby avoiding local voltage instability, equipment malfunctions, or control vacuums, significantly improving system robustness and power supply reliability.

[0032] Example 2 A dynamic reactive power coordinated control method for power grids based on multi-source data fusion is applied to the control system described above, including: Cloud-based collaborative control platform execution: S1.1. Collect multi-source heterogeneous operation data in real time from the main grid dispatching system, distribution network automation system, distributed energy monitoring system and smart electricity terminals through standardized communication interfaces; S1.2. The data fusion module performs time synchronization, outlier identification and cleaning, and spatiotemporal coordinate alignment on the multi-source heterogeneous operating data, and generates a unified-dimensional power grid panoramic state feature tensor through the feature extraction network. S1.3 Based on the power grid panoramic state feature tensor, and combined with the power grid topology and electrical parameters, construct and dynamically update the power grid digital twin; S1.4. Through the centralized optimization decision module, a multi-time-scale collaborative optimization algorithm is used to perform collaborative optimization calculations: At the minute-level optimization layer, based on the power grid digital twin and ultra-short-term forecast data, a reactive power resource allocation strategy benchmark for the next N control cycles is generated. At the second-level triggering layer, based on the reactive resource allocation strategy benchmark, the pre-trained multi-agent deep reinforcement learning model is invoked to output the adjustment instructions for reactive equipment in each region. After the adjustment commands of reactive power equipment in each area are verified for safety and encapsulated, they are sent to the edge autonomous control nodes of the corresponding partitions. Execution of edge autonomous control nodes: S2.1. Collect local operating data in real time, including bus voltage, line power flow, reactive power equipment operating status and switch position, and perform preprocessing operations. S2.2 Real-time monitoring of the link quality between the cloud-based collaborative control platform and the current communication status is determined as "normal" or "abnormal" based on RTT latency, packet loss rate, and acknowledgment success rate. S2.3, Perform dual-mode switching control through the edge autonomous decision-making module: When the communication status is "normal", it receives adjustment instructions from the cloud and sends them to the instruction execution driver module; When the communication status is "interrupted", switch to local autonomous control mode, load the preset autonomous control strategy, and solve the reactive power equipment adjustment command of this area that meets the voltage qualification rate and equipment safety constraints based on real-time voltage deviation, reactive power deficit and equipment action priority. S2.4 The instruction execution driver module decomposes the received regional adjustment instructions, sorts them according to priority, and sends them to drive the corresponding devices to perform reactive power adjustment.

[0033] Specifically, in a new power distribution network demonstration area, a cloud-based collaborative control platform and 18 edge autonomous control nodes are deployed. The communication adopts a dual-channel redundant architecture of 5G slicing and power private network. The demonstration area includes two 110kV substations, 18 10kV feeders, and a distributed photovoltaic installed capacity of 42MW. The load peak-valley difference rate is 65%, and there are typical voltage over-limit and reactive power backfeed problems.

[0034] The cloud-based collaborative control platform performs the following: S1.1 Real-time sensing of multi-source heterogeneous data: On the main grid dispatch side: through the IEC 61970 CIM model interface, the 220 / 110kV bus voltage phasor (PMU data), main transformer load rate and cross-sectional power flow are acquired every second; On the distribution network side: the 10kV bus three-phase voltage / current, switch position, and feeder head power are collected every 2 seconds via the IEC 61850 GOOSE / SV protocol; On the distributed energy side: the active / reactive power output, DC side voltage, and MPPT status of the photovoltaic inverter are accessed every 5 seconds via the Modbus-TCP protocol. On the load side: Through HPLC carrier communication of smart meters, the total reactive power of the distribution area and the voltage qualification rate of typical users are collected every 10 seconds.

[0035] All data is aligned with a unified timestamp (IEEE 1588v2 time synchronization) before entering the data fusion module.

[0036] S1.2 Generation of Power Grid Panoramic State Feature Tensor: The data fusion module removes consecutive missing values ​​caused by communication interruptions; it triggers sliding window anomaly detection for sudden changes in photovoltaic output, resamples data from different sampling frequencies to a unified 1Hz time granularity, and aligns them spatially with the electrical topology according to geographical coordinates; the feature extraction network takes a 128-dimensional × 60-step multi-source time series matrix (containing 21 types of features such as voltage, power, and switching status) as input, extracts local time series patterns through a three-layer dilated convolution of TCN, and calculates the spatiotemporal attention weights between each source through the Attention layer, outputting a three-dimensional tensor (time step × spatial node × feature channel) with dimensions of [60, 182, 64], which is the power grid panoramic state feature tensor.

[0037] S1.3 Construction and Dynamic Updating of Lightweight Digital Twin of Power Grid: Based on the topology switch states in the tensor, topology changes are automatically identified, and graph theory algorithms are called to reconstruct the network connectivity matrix. Based on the real-time reactive power output of SVG and the corresponding bus voltage response delay in the tensor, its equivalent impedance is identified online, and the dynamic model parameters of SVG in the digital twin are updated. If the power output fluctuation rate of the photovoltaic cluster exceeds 25% / min, incremental training is triggered: take the historical data within the sliding window of the most recent 30 minutes, and fine-tune the node admittance matrix correction coefficient in the power flow calculation kernel on the edge side lightweight training platform.

[0038] S1.4 Multi-timescale collaborative optimization and instruction issuance: Minute-level optimization: Using the next 15 minutes as the prediction domain, and inputting ultra-short-term load forecasts and photovoltaic power forecasts, solve a constrained nonlinear programming problem on a digital twin. in, To provide power to the kth reactive power unit, For the action indication function, the weights Output 15 sets of equipment output command sequences, and take the first set as the reference command for the current cycle.

[0039] Second-level triggering: When the voltage at the end of feeder 7 is below the 0.93 pu threshold and remains below it for 220 ms, the MARL model is triggered—the feeder agent receives its own tensor splicing state and that of feeders 6 and 8, and outputs an action vector. The SVG command is +3.2Mvar (voltage boost), the SVC remains unchanged (to avoid frequent operation), and the inverter switches from PQ mode to Q(U) mode (reactive power support).

[0040] The safety verification module checks whether all instructions meet the following conditions: SVG overload rate < 105%, capacitor switching interval > 60s, and voltage sensitivity matrix condition number < 10. 4After verification, it is encapsulated as an IEC 61850-8-1 MMS message and sent to the feeder 7 edge node via the 5G slicing channel.

[0041] Execution of edge autonomous control nodes: S2.1 Local Data Acquisition and Preprocessing: The edge node of feeder 7 is connected to the local SVG controller, SVC monitoring unit and 3 photovoltaic inverters via RS485. Every 200ms, it collects: SVG DC bus voltage, SVC thyristor firing angle, inverter grid connection point voltage / reactive power, and feeder end voltage. Preprocessing includes: voltage signals are filtered by a 50Hz notch filter to eliminate harmonic interference; and data during inverter communication interruptions are completed using forward padding and linear interpolation.

[0042] S2.2 Real-time communication status determination: Dual-channel heartbeat monitoring: The average RTT for the 5G channel is 28ms (standard deviation ±5ms), with a packet loss rate of 0.02%; the average RTT for the private network channel is 15ms, with a packet loss rate of 0.005%; the overall assessment is "normal". If both channels are interrupted for ≥10 seconds, it is determined as an "interrupt".

[0043] S2.3, Autonomous Decision-Making for Dual-Mode Switching: When communication is normal: Receive the SVG+3.2Mvar command sent from the cloud, verify the local SVG capacity, generate the actual execution command +3.2Mvar, and add it to the device action queue (priority: SVG>SVC>inverter). When communication is interrupted: Open-loop autonomous mode is activated—the local LSTM voltage prediction model is called, predicting that the terminal voltage will drop to 0.918 pu. The preset rule base is checked: "Voltage < 0.92 pu and photovoltaic output > 6MW" → SVG full power generation (+5.0Mvar) + inverter Q(U) mode (slope -30kVar / V) is activated; after safety verification (SVG temperature rise rate < 1.5℃ / min), the final instruction is generated.

[0044] S2.4 Instruction Decomposition and Device Drivers: The instruction execution driver module decomposes +5.0Mvar into: SVG adjustment instruction (Modbus address 40001 writes 32000) and inverter reactive power slope setting (CAN frame ID 0x1A2 sends 0x001E0000). Controlled by timing: First send the SVG command (delay 100ms to ensure hardware response), then send the inverter command (to avoid reactive power superposition impact); after the drive is completed, return the execution status code (0x01=success, 0x02=timeout) to the edge autonomous decision module for archiving.

[0045] Using the method described in this application, the voltage qualification rate was increased from 92.3% to 99.82%, the average daily number of reactive power equipment operations across the entire network decreased by 41.6%, and during a 120-second communication interruption, the voltage deviation at the end of the line was controlled within ±0.012 pu in the local autonomous mode, without triggering the upper-level protection.

[0046] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented 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. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0047] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), 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.

[0048] 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.

[0049] 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 1The steps of the function specified in one or more boxes.

[0050] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0051] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A dynamic reactive power coordinated control system for power grids based on multi-source data fusion, characterized in that, The system includes: a cloud-based collaborative control platform, and multiple edge autonomous control nodes that are communicatively connected to the cloud-based collaborative control platform; The cloud-based collaborative control platform is used for cross-regional and cross-level global reactive power collaborative optimization control, including: The multi-source data sensing module collects multi-source heterogeneous data from the main grid dispatch side, distribution network side, distributed energy side and load side in real time through standardized interfaces; The data fusion module performs synchronization, outlier cleaning, and spatiotemporal coordinate alignment operations on the multi-source heterogeneous data through a multi-source data fusion engine, and maps the aligned data into a unified-dimensional power grid panoramic state feature tensor through a feature extraction network. The power grid digital twin module constructs and dynamically updates a power grid digital twin based on the power grid panoramic state feature tensor; the power grid digital twin includes power grid topology and electrical parameters, used to simulate the dynamic response characteristics of reactive power resources; The centralized optimization decision module, based on the power grid digital twin model, performs collaborative optimization calculations with a preset global optimization objective, generates a reactive resource allocation strategy benchmark for equipment in different regions, and generates adjustment instructions for reactive power equipment in each region based on the reactive resource allocation strategy benchmark, and sends the adjustment instructions to the corresponding edge autonomous control nodes. The edge autonomous control node, deployed in a designated area or site of the power grid, is used for reactive power autonomous control within its region, including: The local data processing module collects and processes real-time operational data for the local area. The edge autonomous decision-making module receives adjustment instructions for reactive power equipment in the local area when the cloud-based collaborative control platform is in normal communication mode; when communication with the cloud-based collaborative control platform is abnormal, it generates adjustment instructions for reactive power equipment in the local area based on locally collected data and preset autonomous control strategies. The instruction execution driver module receives and decomposes the adjustment instructions from the edge autonomous decision-making module, and sends the instructions to the corresponding reactive power devices in this area.

2. The power grid dynamic reactive power coordinated control system based on multi-source data fusion according to claim 1, characterized in that: In the data fusion module, the feature extraction network is a hybrid neural network composed of a temporal convolutional network and a self-attention mechanism cascaded together, wherein: The temporal convolutional network contains three dilated causal convolutional layers, used to extract local temporal dependency features of each data source in the time dimension; The self-attention mechanism receives multi-source temporal feature sequences output by a temporal convolutional network, models the spatiotemporal coupling relationship between different data sources through a learnable cross-source association weight matrix, and outputs the power grid panoramic state feature tensor.

3. The power grid dynamic reactive power coordinated control system based on multi-source data fusion according to claim 1, characterized in that: In the power grid digital twin module, the dynamic update mechanism of the power grid digital twin includes: Every T1 cycle, the newly generated power grid panoramic state feature tensor is input into the power grid digital twin; Based on the real-time topology status and switching action information in the power grid panoramic state feature tensor, changes in the power grid topology are identified, and the network connection relationships and operating topology of the power grid digital twin are updated synchronously. Based on the real-time power output and electrical quantity measurement data of the equipment in the power grid panoramic state feature tensor, an online parameter identification algorithm is used to update the equivalent impedance model and dynamic response delay parameters of each reactive power source and reactive power compensation device in the power grid digital twin in real time. When a load change is detected to be greater than a preset threshold, incremental model retraining is triggered. By utilizing historical operating data within a sliding time window, the power flow calculation kernel parameters of the power grid digital twin are adaptively adjusted and calibrated.

4. The power grid dynamic reactive power coordinated control system based on multi-source data fusion according to claim 1, characterized in that: In the centralized optimization decision-making module, the collaborative optimization calculation employs the multi-timescale collaborative optimization algorithm, which includes: The first optimization layer performs minute-level global rolling optimization: with the comprehensive optimization objectives of reducing network losses, improving node voltage qualification rate, and reducing the total number of reactive power equipment operations, the model predictive control method is used to perform periodic rolling solutions based on the power grid digital twin and ultra-short-term forecast data to generate a reactive power resource allocation strategy benchmark. The second optimization layer executes second-level local trigger control: when the real-time voltage deviation is detected to exceed the preset threshold, a pre-trained multi-agent deep reinforcement learning model is launched. The multi-agent deep reinforcement learning model takes the distribution network feeder area as the agent unit. Each agent takes the local and adjacent areas' grid panoramic state feature tensor as the joint state input, takes the adjustment of local reactive power equipment as the action space, and takes minimizing the regional voltage deviation integral, reducing cross-regional reactive power interaction penalties, and reducing equipment action costs as the composite reward function to make collaborative decisions and output the adjustment instructions for reactive power equipment in each area.

5. The power grid dynamic reactive power coordinated control system based on multi-source data fusion according to claim 1, characterized in that: The specific configuration of the edge autonomous decision-making module is as follows: The system receives reactive power equipment adjustment instructions from the centralized optimization decision module as the first instruction source and receives real-time operating data from the local data processing module as the second instruction source. When in normal communication with the cloud-based collaborative control platform, the first instruction source is sent to the instruction execution driver module; When communication with the cloud-based collaborative control platform fails, the system switches to a local autonomous control mode based on the second instruction source, loads a preset autonomous control strategy, and performs local reactive power optimization calculations based on the real-time voltage deviation, reactive power deficit, and preset equipment action priorities and voltage safety constraints of the region to generate adjustment instructions for reactive power equipment in the region.

6. A method for dynamic reactive power coordinated control of power grids based on multi-source data fusion, wherein the method applies the control system described in any one of claims 1-5, characterized in that, include: Cloud-based collaborative control platform execution: S1.

1. Collect multi-source heterogeneous operation data in real time from the main grid dispatching system, distribution network automation system, distributed energy monitoring system and smart electricity terminals through standardized communication interfaces; S1.

2. The data fusion module performs time synchronization, outlier identification and cleaning, and spatiotemporal coordinate alignment on the multi-source heterogeneous operating data, and generates a unified-dimensional power grid panoramic state feature tensor through the feature extraction network. S1.3 Based on the power grid panoramic state feature tensor, and combined with the power grid topology and electrical parameters, construct and dynamically update the power grid digital twin; S1.

4. Through the centralized optimization decision module, a multi-time-scale collaborative optimization algorithm is used to perform collaborative optimization calculations: At the minute-level optimization layer, based on the power grid digital twin and ultra-short-term forecast data, a reactive power resource allocation strategy benchmark for the next N control cycles is generated. At the second-level triggering layer, based on the reactive resource allocation strategy benchmark, the pre-trained multi-agent deep reinforcement learning model is invoked to output the adjustment instructions for reactive equipment in each region. After the adjustment commands of reactive power equipment in each area are verified for safety and encapsulated, they are sent to the edge autonomous control nodes of the corresponding partitions. Execution of edge autonomous control nodes: S2.

1. Collect local operating data in real time, including bus voltage, line power flow, reactive power equipment operating status and switch position, and perform preprocessing operations. S2.2 Real-time monitoring of the link quality between the cloud-based collaborative control platform and the current communication status is determined as "normal" or "abnormal" based on RTT latency, packet loss rate, and acknowledgment success rate. S2.3, Perform dual-mode switching control through the edge autonomous decision-making module: When the communication status is "normal", it receives adjustment instructions from the cloud and sends them to the instruction execution driver module; When the communication status is "interrupted", switch to local autonomous control mode, load the preset autonomous control strategy, and solve the reactive power equipment adjustment command of this area that meets the voltage qualification rate and equipment safety constraints based on real-time voltage deviation, reactive power deficit and equipment action priority. S2.4 The instruction execution driver module decomposes the received regional adjustment instructions, sorts them according to priority, and sends them to drive the corresponding devices to perform reactive power adjustment.

Images