SVG-APF collaborative optimization scheduling system and method based on AI prediction

Abstract

The invention discloses an SVG-APF collaborative optimization scheduling system and method based on AI prediction, and relates to the technical field of power system scheduling optimization, and the system comprises an AI prediction module, an SVG scheduling module, an APF scheduling module, a collaborative scheduling decision module, and a real-time monitoring and feedback module. And all the modules are deeply linked to form a full-link intelligent closed loop of acquisition, pre-judgment, scheduling, feedback and evolution. According to the invention, the system has the capability of pre-judging the load, the reactive power demand and the harmonic load change in advance, and particularly, the follow-up scheduling is more prospective and robust by outputting confidence and risk level labels, so that the hysteresis problem in the traditional scheduling is avoided, the timeliness and the accuracy of reactive power compensation and harmonic suppression are ensured, and the system is suitable for large-scale popularization and application. The cooperative scheduling decision module can actively identify and solve coupling conflicts between reactive compensation and harmonic suppression, secondary optimization or negative effects possibly caused by independent scheduling are avoided, and optimization of global electric energy quality is achieved.

Description

Technical Field

[0001] This invention relates to the field of power system dispatch optimization technology, and in particular to an SVG-APF collaborative optimization dispatch system and method based on AI prediction. Background Technology

[0002] In modern power systems, with the continuous expansion of distributed renewable energy grid connection, the increasing number of high-end precision loads, and the widespread application of power electronic equipment, power grid operation faces increasingly severe power quality problems such as declining power factor and serious harmonic pollution. Static Synchronous Compensators (SVG), as an advanced reactive power compensation device, can quickly and accurately provide the required reactive power, effectively maintaining grid voltage stability and improving the power factor. Active Power Filters (APFs) are power electronic devices used to suppress harmonic pollution. They can track and generate a compensation current equal in magnitude and opposite in direction to the harmonic current in real time, thereby canceling harmonic components in the power grid.

[0003] Existing dispatching systems typically operate independently. SVG (Static Var Generator) is primarily used for reactive power compensation, while APF (Active Power Filter) is mainly used for harmonic suppression. These two functions may conflict under certain circumstances, and existing dispatching systems struggle to effectively resolve such conflicts, making it difficult to achieve globally optimal power quality control. Furthermore, with increasing grid complexity, the fusion, processing, and effective utilization of multi-source heterogeneous data, as well as the real-time adaptability and self-evolutionary capabilities of dispatching strategies, have become pressing issues that current technologies must address. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-predictive-based SVG-APF collaborative optimization scheduling system and method.

[0005] To achieve the above objectives, this invention adopts the following technical solution: an AI-predictive SVG-APF collaborative optimization scheduling system, comprising an AI prediction module, an SVG scheduling module, an APF scheduling module, a collaborative scheduling decision module, and a real-time monitoring and feedback module. These modules are deeply interconnected to form a fully intelligent closed loop encompassing data acquisition, prediction, scheduling, feedback, and evolution. The AI ​​prediction module outputs load, reactive power demand, and harmonic load prediction results with confidence level and risk level labels. The SVG scheduling module achieves precise reactive power compensation scheduling based on the prediction results, maintaining a stable power factor in the power grid. The APF scheduling module achieves targeted harmonic suppression, controlling the harmonic content of the power grid. The collaborative scheduling decision module resolves the coupling conflict between reactive power compensation and harmonic suppression, achieving global optimization. The real-time monitoring and feedback module collects data across all dimensions and provides hierarchical proactive feedback, supporting algorithm iteration in each module.

[0006] As a further description of the above technical solution: The AI ​​prediction module constructs a prediction system that integrates spatiotemporal causal fusion, cross-domain knowledge transfer, and multimodal feature reasoning. It integrates multi-source data, including historical operation data, real-time power grid data, cross-domain power grid common data, equipment health status data, new energy access fluctuation data, and power grid topology switching records. After preprocessing such as outlier removal, data standardization, feature engineering, intelligent identification of power grid operation scenarios, and cross-domain feature mapping, the prediction results are updated regularly through proactive uncertainty exploration and dynamic correction logic.

[0007] As a further description of the above technical solution: The SVG scheduling module adopts a scheduling algorithm with dual-weight dynamic balancing and three-dimensional balance optimization. First, it allocates the reactive power output ratio by weighting voltage sensitivity and equipment health. Then, it performs three-dimensional balance dynamic optimization based on the urgency of reactive power demand, equipment adjustment capability, and scheduling economic cost. This is combined with a reactive power fluctuation prediction and smoothing algorithm, a timing backtracking correction mechanism, and a dynamic redundancy adaptive allocation strategy.

[0008] As a further description of the above technical solution: The APF scheduling module uses an adaptive Fourier transform combined with a harmonic mode identification algorithm to identify steady-state, transient, and interharmonic harmonic modes and trace key links across the entire chain. It employs a triple mechanism of targeted suppression, dynamic evolution, and online evolution, and designs a cluster load balancing-cooperative suppression algorithm for multi-APF deployment scenarios.

[0009] As a further description of the above technical solution: The collaborative scheduling decision-making module constructs a multi-agent game and hierarchical cross-temporal consensus collaborative framework, which encapsulates SVG, APF, new energy equipment, and energy storage system as independent intelligent agents, and implements hierarchical game according to the main grid layer, distribution network layer, and terminal layer. It resolves conflicts through a prediction-negotiation-evolution system and reserves future adjustment space in combination with medium and long-term forecasts.

[0010] As a further description of the above technical solution: The real-time monitoring and feedback module deploys full-link monitoring units at the power grid bus, critical load side, and equipment output end to monitor all dimensions of parameters, including voltage, current, active power, reactive power, harmonic content, equipment temperature, operating load rate, and switching losses. It adopts Ethernet plus 5G dual-link transmission and uses a multi-source data virtual-real fusion verification algorithm to distinguish between data errors and real fluctuations, matching different feedback strategies according to the anomaly level.

[0011] As a further description of the above technical solution: The AI-predictive SVG-APF collaborative optimization scheduling method includes the following steps: Multi-dimensional data collection and cross-domain adaptation: The real-time monitoring module collects multi-source data, which is then preprocessed, verified through virtual-real fusion, and mapped across domains before being pushed to each module; Cross-domain fusion AI proactive prediction: The AI ​​prediction module calls a cross-domain knowledge base and outputs prediction results labeled with confidence level and risk level; Multi-dimensional balanced SVG precise scheduling: The SVG scheduling module generates scheduling instructions through dual-weight balancing and three-dimensional balancing algorithms; Full-link evolutionary APF targeted suppression: The APF scheduling module initiates dynamic evolution and online evolution suppression strategies through modality recognition and full-link tracing; Hierarchical cross-temporal collaborative decision-making: The collaborative scheduling decision-making module generates globally optimal correction instructions through multi-agent hierarchical game theory; Full-link feedback and algorithm iteration: Each module optimizes algorithm parameters based on feedback data, forming a closed-loop cycle.

[0012] As a further description of the above technical solution: The AI ​​prediction module mines the load transmission relationship of power grid nodes through spatiotemporal correlation, avoids interference from false data by combining causal reasoning, reuses common knowledge of cross-domain power grids, integrates physical, environmental and behavioral multimodal features for reasoning, and simultaneously identifies ambiguous prediction areas and adjusts the data collection frequency of those areas.

[0013] As a further description of the above technical solution: The collaborative scheduling decision module uses AI prediction to identify potential conflicts between reactive power compensation and harmonic suppression in advance. It resolves these conflicts proactively through a negotiation mechanism involving concessions and exchanges of benefits. The conflict types and negotiation results are stored in the strategy library, enabling the self-evolution of conflict resolution strategies.

[0014] As a further description of the above technical solution: The modules communicate with each other using ModbusTCP, MQTT and RESTfulAPI protocols. Data is encapsulated in JSON format, including data generation timestamp, confidence level, source identifier and anomaly marker. It is transmitted through Ethernet plus 5G dual-link redundancy, and uses CRC32 check and TLS1.3 encryption to ensure data reliability and security.

[0015] The present invention has the following beneficial effects: 1. In this invention, by introducing an AI prediction module, the system possesses the ability to predict changes in load, reactive power demand, and harmonic load. In particular, by outputting confidence levels and risk level labels, subsequent scheduling becomes more forward-looking and robust, thus avoiding the lag problem in traditional scheduling and ensuring the timeliness and accuracy of reactive power compensation and harmonic suppression. The collaborative scheduling decision module can proactively identify and resolve the coupling conflict between reactive power compensation and harmonic suppression, avoiding suboptimal or negative impacts that may result from independent scheduling. This achieves global power quality optimization, ensuring stable and economical grid operation. The real-time monitoring and feedback module enables end-to-end data acquisition and hierarchical proactive feedback. The feedback mechanism not only provides data support for the algorithm iteration of each module, but also enables the entire system to continuously learn, adapt, and evolve like a living organism, significantly improving the system's long-term performance and adaptability to complex power grid environments. Through the collaborative optimization of SVG and APF, this system effectively controls the power grid harmonic content while maintaining the stability of the power grid power factor, achieving a comprehensive upgrade of power quality, reducing line losses, extending equipment life, and improving power supply reliability. This system adopts a modular design, with each module communicating through a standard protocol, exhibiting good scalability and compatibility, easy deployment and maintenance, and the ability to adapt to power systems of different scales and complexities. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the present invention; Figure 2 This is a flowchart of the method 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] Reference Figure 1-2This invention provides an embodiment of an AI-predictive SVG-APF collaborative optimization scheduling system, comprising an AI prediction module, an SVG scheduling module, an APF scheduling module, a collaborative scheduling decision module, and a real-time monitoring and feedback module. These modules are deeply interconnected, forming a fully intelligent closed-loop system encompassing data acquisition, prediction, scheduling, feedback, and evolution. The AI ​​prediction module outputs load, reactive power demand, and harmonic load predictions with confidence levels and risk levels. The SVG scheduling module uses these predictions to achieve precise reactive power compensation scheduling, maintaining a stable power factor in the grid. The APF scheduling module implements targeted harmonic suppression, controlling the harmonic content of the grid. The collaborative scheduling decision module resolves the coupling conflict between reactive power compensation and harmonic suppression, achieving global optimization. The real-time monitoring and feedback module collects data across all dimensions and provides hierarchical proactive feedback, supporting algorithm iteration across all modules.

[0019] In one embodiment, the AI ​​prediction module provides the downstream scheduling module with forward-looking predictions of load, reactive power demand, and harmonic load. More importantly, it constructs proactive intelligence for the system through spatiotemporal causal fusion, cross-domain knowledge transfer, and multimodal feature reasoning, enabling the scheduling strategy to shift from traditional passive response to trend-based proactive planning. By integrating historical operating data, real-time grid data, cross-domain grid common data, equipment health status data, renewable energy access fluctuation data, and grid topology switching records, the AI ​​prediction module builds a multi-source input foundation for grid status, giving the prediction results the ability to be constrained by real physical mechanisms, map load transmission relationships, and adapt to cross-scenario migration. The AI ​​prediction module performs outlier removal, data standardization, feature engineering, intelligent identification of grid operating scenarios, and cross-domain feature mapping preprocessing on the input data. By judging the confidence level of the data transmitted from the real-time monitoring and feedback module, it achieves dynamic adaptation to data quality. Power grid operation scenario identification enables the AI ​​prediction module to automatically switch to an appropriate prediction model logic based on factors such as bus voltage, equipment load factor, and renewable energy fluctuations. Cross-domain feature mapping extracts common patterns across different voltage levels and load types within the power grid, ensuring stable inference capabilities even when local data is scarce or new operating conditions emerge. In the prediction calculation phase, the AI ​​prediction module focuses on spatiotemporal causal fusion. By mining the load transmission paths and temporal correlations between power grid nodes, it clarifies the causal chains of load changes, thereby reducing the error risk associated with relying solely on correlation. The spatiotemporal causal fusion prediction model formula is as follows: , : Target value predicted at time t (which can represent load forecast, reactive power demand forecast, or harmonic load forecast, and is the core forecast data output to the SVG scheduling module and APF scheduling module). Spatiotemporal feature weight coefficients (obtained through data training, reflecting the contribution of spatiotemporal features to the prediction results). : Spatiotemporal feature vector at time t (including load transmission path characteristics and temporal correlation characteristics between power grid nodes, calculated based on historical operating data and real-time power grid data collected by the real-time monitoring module). Causal factor weight coefficient (obtained through causal reasoning verification, reflecting the corrective force of causal relationship on prediction results). : Causal factor at time t (generated by causal inference mechanism, after filtering out false trends and invalid noise, reflecting the real driving factors of power grid changes). The random error term for prediction at time t (with a mean of 0 and a variance of ) The normal distribution represents random fluctuations not explained by spatiotemporal features and causal factors. Based on this, a causal inference mechanism filters out spurious trends and invalid noise, ensuring the prediction results accurately reflect the driving factors of grid changes. Simultaneously, a cross-domain knowledge transfer mechanism reuses common knowledge from cross-scenario grids, abstracting and mapping historical cross-domain grid characteristics to maintain stable predictive performance even when facing uncommon load patterns, strong fluctuations in new energy sources, or short-term structural changes caused by topology switching. Cross-domain feature fusion model: Predictive models that integrate cross-domain knowledge: , The feature vector at time t that incorporates cross-domain knowledge (serves as the core input of the prediction model, taking into account both cross-domain commonalities and local characteristics). Cross-domain feature similarity weight at time t (value range 0-1, calculated by comparing features of the target power grid and the cross-domain power grid, derived from cross-domain feature mapping processing, when local data is scarce). Approaching 1, when local data is sufficient (approaching 0) : Cross-domain power grid feature vector at time t (from the cross-domain power grid common data interface, containing the common reactive power and harmonic characteristics of power grids with different voltage levels and load types). : Local power grid feature vector at time t (obtained by preprocessing local historical operating data, equipment health status data, new energy access fluctuation data, and power grid topology switching records collected by the real-time monitoring module). : Linear prediction function (simplified core prediction mapping relationship, which can be extended by combining neural networks and other models in practice). Feature weight matrix (obtained through joint training of cross-domain knowledge base and local data). Predictive bias term (compensating for systematic biases in the feature mapping process). Multimodal feature inference further integrates physical, environmental, and behavioral features. Through joint inference of the logical relationships between different types of features, the prediction conclusions conform to both data trends and the physical laws of power grid operation. Multimodal feature weighted fusion model: Multimodal fusion prediction model: , : Multimodal fusion feature vector at time t (integrating physical, environmental and behavioral features to make the prediction results conform to the physical laws of power grid operation). These are the normalized weights for physical features, environmental features, and behavioral features, respectively (obtained through feature importance assessment, reflecting the degree of influence of different modal features on prediction). : Physical feature vector at time t (including power grid topology parameters, impedance characteristics, equipment rated parameters, etc., from the power grid basic data of the real-time monitoring module). : Environmental feature vector at time t (including temperature, humidity, extreme weather warnings, etc., from the environmental data acquisition unit of the real-time monitoring module). : Behavioral feature vector at time t (including load change patterns, new energy output fluctuation trends, grid topology switching behavior, etc., extracted from time-series data based on real-time monitoring module). Multimodal fusion prediction function (integrates multimodal features with spatiotemporal and causal features to achieve comprehensive reasoning). Multimodal feature contribution coefficient (determined through training, balancing the influence of multimodal features with spatiotemporal and causal features). Multimodal feature weight matrix (optimized mapping of multimodal fusion features). During prediction optimization, the AI ​​prediction module simultaneously possesses proactive uncertainty exploration and dynamic correction logic. Prediction confidence calculation model: Dynamically correct prediction model: , : Prediction confidence level at time t (value range 0-100%, output to downstream modules as auxiliary decision data; higher confidence level means more reliable prediction results). Standard deviation of the prediction error at time t (calculates the degree of deviation and fluctuation between the predicted value at time t and the actual value at the same historical time). : Confidence score calculation time window length (default is the historical data window corresponding to a 5-minute update cycle, i.e., T=5). Time window The maximum standard deviation of the internal prediction error (used to normalize the confidence level to the 0-1 range). The final predicted value after correction at time t (the final result output to the SVG scheduling module, APF scheduling module, and collaborative scheduling decision module). Error correction coefficient (value range 0-1, learned through error memory, balancing the correction strength of historical errors on the current prediction). : The abnormal precursor classification coefficient at time t (corresponding to the abnormal precursor classification label of the real-time monitoring module, slight = 0.1, moderate = 0.3, severe = 0.5, the higher the abnormal level, the greater the correction force). The prediction error at time t-1 (i.e. , (The actual observation value at time t-1 is from feedback data from the real-time monitoring module). When ambiguous areas appear in the prediction results, the module will proactively increase the data acquisition frequency of the relevant areas to enhance data density and improve prediction reliability. It will also establish a correlation between prediction deviations and actual deviations through an error memory mechanism, used to update model parameters and cross-domain knowledge transfer strategies. The abnormal precursor classification indicators (minor, moderate, severe) pushed by the real-time monitoring and feedback module will also serve as important inputs to the prediction correction rules, enabling the prediction module's calculation logic to enter a high-sensitivity mode in advance when early signs of instability appear in the power grid, thereby improving overall predictive capabilities. The output of the AI ​​prediction module includes not only load predictions, reactive power demand predictions, and harmonic load predictions, but also prediction confidence levels, reactive power fluctuation risk levels, and harmonic mode prediction results. These prediction results are generated separately according to different time scales: short-term and medium-to-long-term. Short-term results are used to drive the immediate adjustment of the SVG scheduling module and the APF scheduling module, while medium-to-long-term prediction results provide cross-temporal and spatial coordination basis for the collaborative scheduling decision module, enabling the system to reserve future adjustment space in advance. In terms of timing requirements, the AI ​​prediction module updates predictions every 5 minutes as a regular rhythm, and triggers immediate updates in scenarios such as sudden load changes, strong fluctuations in new energy sources, or abnormal equipment status, ensuring that its prediction results always maintain high timeliness and foresight. Through a prediction system built on cross-domain fusion, causal reasoning, and multimodal features, the AI ​​prediction module enables the system to move beyond relying solely on current data and instead develop proactive control capabilities based on trends, causal relationships, and cross-domain knowledge. Consequently, the SVG scheduling module and APF scheduling module can proactively construct output strategies and pre-configure pre-charging or suppression parameters under the guidance of prediction results and their confidence levels, thereby improving the stability of reactive power compensation and harmonic suppression. The collaborative scheduling decision module relies on the medium- and long-term trends and risk labels provided by the AI ​​prediction module to proactively identify potential conflicts in the reactive power compensation and harmonic suppression process, thereby constructing cross-temporal and spatial collaborative strategies to maintain the global consistency of the system. Through this proactive prediction mechanism, the entire system forms a self-evolving closed loop of prediction-matching-feedback correction, enabling the system to have high adaptability and robustness in dealing with multiple scenarios, cross-regional operations, and strong fluctuations.

[0020] In one embodiment, the SVG dispatch module provides continuous and accurate reactive power compensation to the power grid through proactive responses to reactive power demand. Based on reactive power demand forecasts, load forecasts, forecast confidence levels, and reactive power fluctuation risk levels provided by the AI ​​prediction module, this module employs a dispatch algorithm comprised of dual-weighted dynamic balancing and three-dimensional balance optimization. This enables the power grid's reactive power compensation strategy to shift from passive response to proactive regulation. During system operation, the SVG dispatch module is responsible not only for matching the real-time reactive power demand of each node but also for regulating grid voltage levels, suppressing reactive power fluctuations, reducing operating losses, and preventing reactive power deficiency or excess. Thus, it plays a crucial role in stabilizing voltage, ensuring equipment health, and improving operational economy within the entire collaborative system. In terms of specific technical implementation, the SVG dispatch module first allocates the reactive power output ratio through a dual-weighted mechanism of voltage sensitivity and equipment health, allowing dispatch commands to balance the voltage demand of each node with the health status of the SVG equipment itself. When a node has high voltage sensitivity, the module increases the reactive power compensation priority for that node. Conversely, the lower the equipment health, the more restricted the equipment load. This dual-weight allocation mechanism ensures that the reactive power output distribution better matches the actual needs of the power grid and prevents equipment degradation due to overload. Combining operating temperature, load factor, and switching losses from the real-time monitoring and feedback module, the SVG scheduling module can correct the equipment health assessment in real time, ensuring the accuracy of the weight allocation. Dual-weight dynamic balancing reactive power output weight calculation for SVG equipment: Calculation of reactive power output ratio for a single SVG device: Actual reactive power output command for a single SVG device: , : No. The overall weight of each SVG device (integrating voltage sensitivity and device health to determine the priority of the device in reactive power distribution) Weighting coefficients for voltage sensitivity and equipment health (satisfying) Obtained through offline training of the system, by default , (It can be dynamically adjusted according to the power grid voltage stability requirements). : No. The voltage sensitivity of the power grid node corresponding to each SVG device (value range 0-1, "node voltage deviation" collected by the real-time monitoring and feedback module) "and SVG reactive power regulation" The ratio calculation of ", i.e. A higher sensitivity indicates that the voltage at that node is more sensitive to changes in reactive power, and reactive power should be allocated preferentially. : No. The health status of each SVG device (value range 0-1, based on the device operating temperature collected by the real-time monitoring and feedback module) Load factor Switching losses Weighted calculation, i.e. , For rated temperature, For the maximum allowable temperature, To maximize allowable switching losses, a lower health level indicates that the equipment needs to have its load limited. : No. The reactive power output ratio of each SVG device (determines the distribution ratio of total reactive power demand among multiple SVGs, ensuring that reactive power output matches node demand and device status). Total number of SVG devices (statistics are compiled by the real-time monitoring and feedback module based on the power grid topology and capacity configuration). : No. The actual reactive power output command of the SVG device (the final execution value sent to the device must meet the device's rated capacity constraint) , , (The minimum / maximum reactive power output capability of the SVG is derived from the real-time monitoring module). Total reactive power demand of the power grid (calculated by combining the "reactive power demand forecast value" provided by the AI ​​prediction module with the "current reactive power deficit" monitored in real time, i.e.) , For AI short-term predictions, (For real-time reactive power deviation). Based on dual-weight balancing, the SVG scheduling module further employs a three-dimensional balance optimization algorithm considering the urgency of reactive power demand, equipment adjustment capability, and scheduling economic cost. Standardized calculation of the three-dimensional indicators: Three-dimensional equilibrium comprehensive optimization objective: , Standardized value for reactive power demand urgency (range 0-1, reflecting the urgency of the power grid's current and short-term reactive power demand; higher values ​​require priority response). The AI ​​prediction module provides "short-term (15 minutes - 2 hours) reactive power demand forecast" (core input, reflecting prediction-driven proactive adjustment). The AI ​​prediction module outputs a "reactive power fluctuation risk level coefficient" (minor risk). Medium risk Serious risks (The higher the risk, the greater the weight of urgency). : Upper limit of total reactive power regulation capacity of the power grid (all SVG devices) The sum (from the real-time monitoring module). Standardized value of equipment regulation capacity (range 0-1, reflecting the current available regulation margin of SVG; the higher the value, the more reactive power the equipment can handle). : No. Adjustment margin of the SVG (Static Varistor) The current output collected by the real-time monitoring module (Calculated with rated capacity). Standardized value of scheduling economic cost (range 0-1, reflecting the economy of scheduling; the higher the value, the lower the operating loss and aging cost). : No. The real-time operating loss of the SVG (calculated from voltage and current data collected by the real-time monitoring module, i.e.) , For SVG output current, (equivalent resistance of the equipment) : No. The aging cost coefficient of the SVG (positively correlated with the equipment's operating years and the number of switching operations, fitted by the equipment health status data recorded by the real-time monitoring module, unit: yuan / time). : No. The adjustment frequency of the SVG (unit: times / minute, calculated by the real-time monitoring module based on the number of instruction adjustments in the last 10 minutes; the more frequent the adjustment, the higher the aging cost). : No. Rated operating losses of each SVG (factory parameters of the equipment, used as a benchmark value for economic cost). Maximum aging cost factor (aging cost at the end of the equipment's lifespan, used as a standardized benchmark). : No. Three-dimensional equilibrium integrated target value of SVG (system selection) The largest equipment is prioritized for reactive power tasks to achieve a global balance between urgency, adjustability, and economy. Weighting coefficients of the three-dimensional indicators (default) , , The collaborative scheduling decision module can dynamically adjust the power grid based on the current power grid target, such as increasing the power grid capacity when the voltage is unstable. This algorithm dynamically assesses the urgency of reactive power demand, the current available adjustment margin of equipment, and the energy consumption and loss costs of scheduling execution, ensuring that the reactive power compensation process always maintains a reasonable balance between efficiency, cost, and equipment lifespan. The urgency of reactive power demand is jointly determined by the short-term prediction value of the AI ​​prediction module and the reactive power fluctuation risk level. The equipment adjustment capability is updated in real time based on the operating status data provided by the real-time monitoring and feedback module, while the economic cost of scheduling is estimated by comprehensively considering operating losses, adjustment frequency, and equipment aging. Through a three-dimensional balance optimization strategy, the SVG scheduling module can automatically select the most suitable adjustment path under complex operating conditions, avoiding scheduling imbalances caused by a single factor and significantly improving the accuracy and stability of reactive power compensation. To ensure the continuity and smoothness of the reactive power compensation process, the SVG scheduling module combines a reactive power fluctuation prediction and smoothing algorithm to generate a stepped output curve, preventing abrupt changes in scheduling actions and reducing disturbances to the grid voltage. Reactive power fluctuation prediction and smoothing algorithm: Calculation of smoothed output value at future moments: , : No. Taiwan SVG in the future Smooth output instructions at different times ( For the current moment, The default time step. (minutes, matching the 5-minute update cycle of the AI ​​prediction module) : No. Taiwan SVG at the current moment The actual output command (benchmark value, from the result of the dual-weight / three-dimensional balancing algorithm). Smoothing coefficient (range 0.2-0.5, dynamically adjusted by prediction confidence): AI prediction confidence. hour , hour , hour The confidence level comes from the output of the AI ​​prediction module, avoiding sudden changes in output due to prediction errors. The AI ​​prediction module provides the "first" The corresponding node of the SVG is in "Reactive power demand forecast at any given time" (core input, reflecting forward-looking planning) : Smoothing time window length (default) This corresponds to a 5-minute prediction update cycle, meaning the output curve for the next 5 time steps is planned in advance to achieve pre-charging and smooth adjustment. The algorithm analyzes short-term trends provided by the AI ​​prediction module to construct a progressively changing reactive power output curve in advance. This allows the SVG to pre-charge before demand increases and reduce output power before demand decreases, minimizing the deviation between actual operation and the predicted target. Simultaneously, the module introduces a time-series backtracking correction mechanism. By continuously comparing the execution results with the target output, the accumulated impact of the deviation is gradually corrected, ensuring the scheduling strategy remains accurate through continuous iteration. In the time-series backtracking correction mechanism, the corrected output instruction at the current moment is: Calculation of prediction deviation at the previous moment: , : No. Taiwan SVG at the current moment The corrected output instructions (eliminate historical deviation accumulation and improve scheduling accuracy). Deviation correction coefficient (range 0.1-0.3, dynamically adjusted based on deviation stability feedback from the real-time monitoring module; small deviation fluctuations are corrected accordingly). Large deviation fluctuations (To avoid over-correction leading to new volatility) : No. Taiwan SVG at the previous moment The prediction deviation (reflects the difference between historical scheduling instructions and actual operation; a positive deviation indicates that the previous output was insufficient and the current output needs to be increased; a negative deviation indicates that the output needs to be reduced). : No. Taiwan SVG at the previous moment The actual output value (calculated from the voltage and current data collected by the real-time monitoring module, i.e.) , (This refers to the power factor angle, ensuring data accuracy and reliability). Under conditions of significant system dynamic changes, the SVG scheduling module reserves cross-device coordination space for the system through a dynamic redundancy adaptive allocation strategy. Actual redundancy capacity calculation: , : No. Taiwan SVG at the current moment The actual redundancy capacity (reserved emergency adjustment capacity to cope with sudden reactive power fluctuations or cross-device coordination). : No. The baseline redundancy capacity of the SVG (by default, 15% of the device's rated capacity, i.e.) (Determined by equipment parameters and power grid safety requirements) Current moment The reactive power fluctuation risk level coefficient (consistent with Formula 4, derived from the AI ​​prediction module: slight risk) Medium risk Serious risks ), Logical connection: When AI predicts "reactive power demand eases" (risk level is mild and When this is the case, the redundancy factor can be reduced to... The released redundant capacity is allocated to APF (Active Power Controller) or new energy equipment through the collaborative scheduling decision module, achieving cross-equipment capability complementarity. When the reactive power fluctuation risk level increases, the module will proactively increase redundant capacity to cope with possible sudden demand; when the AI ​​prediction module predicts that reactive power demand will ease in a certain period, the module can reduce redundant resource occupation, leaving more regulation potential for the APF scheduling module or new energy and energy storage equipment that operate in tandem with it. With the support of this dynamic redundancy management mechanism, the SVG (Static Var Generator) scheduling module can maintain high-frequency synchronization with the APF scheduling module and provide the collaborative scheduling decision module with equipment health status assessment, regulation margin, and execution feedback, enabling the entire system to form complementary regulation capabilities at the cross-equipment level. In the collaborative operation of the entire system, the SVG scheduling module is not only the executor of reactive power compensation, but also a data feedback provider and an important participant in collaborative decision-making. Its output of real-time reactive power compensation data, grid impedance change prediction, tiered output change information, and scheduling execution feedback are used by the APF scheduling module and the collaborative scheduling decision module for harmonic suppression strategy adjustment and cross-equipment conflict resolution. When the APF scheduling module detects that the harmonic variation trend is affected by the SVG output rhythm, the SVG scheduling module can adjust its output rhythm in a timely manner under a 5ms-level data synchronization mechanism, avoiding scheduling interference between the two. Through this high-frequency, low-latency output and feedback mechanism, the SVG scheduling module constructs a key execution link that both responds to the real-time state of the power grid and closely coordinates with the strategy, ensuring the coordination and consistency of reactive power compensation and harmonic suppression throughout the system. Through bidirectional data linkage with the AI ​​prediction module, APF scheduling module, and collaborative scheduling decision module, the SVG scheduling module transforms reactive power compensation from an instantaneous response to global planning, from single-device behavior to cross-level collaboration, and from result-oriented to prediction-driven, achieving a comprehensive improvement in the accuracy, stability, and self-evolutionary capability of the power grid's reactive power regulation.

[0021] In one embodiment, the APF scheduling module, based on the harmonic load predictions, harmonic mode prediction results, and prediction confidence levels from the AI ​​prediction module, transforms harmonic management from a traditional passive response to proactive control. The APF scheduling module, through adaptive Fourier transform combined with a harmonic mode identification algorithm, can distinguish between different types of harmonic modes, such as steady-state harmonics, transient harmonics, and interharmonics. This allows it to accurately identify harmonic sources, determine harmonic propagation characteristics, and trace key links even under complex operating conditions. In actual operation, the APF scheduling module first uses adaptive Fourier transform to analyze real-time harmonic monitoring data from the real-time monitoring and feedback module. By dynamically adjusting the analysis window and analysis parameters, the harmonic spectrum extraction can adapt to changes in the power grid's operating state. Dynamic window length adjustment: Harmonic amplitude and phase extraction: , , The analysis window length (unit: sampling points) of the adaptive Fourier transform at time t, dynamically matching the power grid operating state. : Reference window length (default is 512 sampling points, determined by the power grid fundamental frequency of 50Hz and the sampling frequency of 10kHz to ensure the accuracy of fundamental frequency analysis). Window adjustment factor (default) (Through offline testing and calibration, the control window's variation with fluctuations is measured). Standard deviation of grid current fluctuation at time t (instantaneous current data collected by the real-time monitoring and feedback module) Calculations reflect current stability; the greater the fluctuation, the better. The larger the window, the longer the window. The smaller the value, the faster the transient response. : time t The amplitude of the subharmonic (unit: A) is the core basis for the APF current suppression command. The value should be between 2 and 50 (covering the 2-50th harmonic monitoring range required by the document). : time t The instantaneous current of the power grid at each sampling point (collected by the current sensor of the real-time monitoring and feedback module, with a sampling frequency of 10kHz to ensure high-frequency harmonic capture). : time t The phase of the subharmonic (unit: rad) is used for phase synchronization of the APF suppression current to avoid introducing new fluctuations during suppression. The complex argument function is used to extract phase information from the Fourier transform result. Based on the analytically obtained spectral data, the harmonic mode recognition algorithm further identifies the characteristics of different modes and improves recognition accuracy by combining the mode prediction results provided by the AI ​​prediction module. Harmonic mode recognition algorithm: Modal feature fusion: Modal membership calculation: , The fused harmonic mode feature vector at time t (combining measured and predicted features to improve recognition accuracy, in line with the document's requirement to "improve recognition accuracy by combining modal prediction results provided by the AI ​​prediction module"). Weight coefficients of measured features and AI-predicted features (satisfying) ,default , AI prediction confidence hour Increased to 0.5). : Measured modal eigenvectors at time t (including harmonic amplitude fluctuation coefficients) Harmonic duration Calculated from harmonic data by the real-time monitoring module, steady-state harmonics have small fluctuations and long durations, while transient harmonics have the opposite characteristics. The modal feature vector output by the AI ​​prediction module at time t (from the "harmonic mode prediction result", which provides prior information on the modal type in advance). : time t Membership degree of each mode (value range 0-1, the mode with the highest membership degree is the current dominant mode). Represents steady-state harmonics, Represents transient harmonics, (representing interharmonic waves) : No. The baseline feature vectors of each mode (obtained offline through historical harmonic data, such as steady-state harmonics) The fluctuation coefficient is ≤0.05 and the duration is ≥10s. : No. The characteristic variance of a modality (the degree of dispersion of the baseline feature, by default) , , ), Euclidean distance is used to calculate the similarity between the fused features and the baseline features. After identifying the dominant harmonic mode, the module analyzes the entire process of harmonic generation, propagation, and attenuation through a full-link tracing mechanism, and locates the key links with the greatest impact on the power grid. This tracing mechanism, based on the trend of harmonic amplitude and phase changes and the response characteristics of power grid nodes, enables the APF scheduling module to accurately identify the location of harmonic sources, main propagation paths, and sensitive nodes that are easily affected, providing a reliable basis for subsequent targeted suppression. In the execution of harmonic mitigation strategies, the APF scheduling module adopts a triple mechanism of targeted suppression, dynamic evolution, and online evolution. The targeted suppression mechanism is based on the identified harmonic modes and key links, and directly applies the suppression strategy to the harmonic source, harmonic propagation path, or affected nodes, making the suppression behavior highly accurate and targeted. Single APF targeted suppression current command: , : time t Taiwan APF targets the first Targeted suppression current command for subharmonics (complex form, including amplitude and phase, to ensure accurate harmonic cancellation). : time t The grid impedance corresponding to the second harmonic (unit: Ω, harmonic voltage collected by the real-time monitoring and feedback module) Current Calculation, i.e. (This reflects the impedance characteristics of the harmonic propagation path). : time t The equivalent output impedance of the APF (unit: Ω, determined by the APF device parameters and operating status, default value is 0.05Ω to ensure matching suppression capability). : SVG at time t for the first The suppression correction coefficient of the APF (range 0.9-1.1, calculated from the "real-time reactive power output data" provided by the SVG scheduling module, and the SVG output change rate) hour This compensates for the impact of SVG output on harmonics, in accordance with the document's requirement to "adjust suppression parameters based on real-time reactive power output data from the SVG scheduling module". The phase compensation term for suppressing the current ensures that the APF output current is out of phase with the harmonic current, achieving precise cancellation. Based on this, the dynamic evolution mechanism automatically adjusts the suppression parameters according to the real-time monitoring and feedback module's grid impedance changes and the SVG dispatch module's real-time reactive power output data. This allows the suppression strategy to adapt to dynamic changes in the grid, preventing a decrease in harmonic suppression effectiveness due to changes in grid impedance or SVG output rhythm. Dynamic correction of the PI controller proportional coefficient: Dynamic correction of the integral coefficient of the PI controller: , : time t Taiwan APF targets the first The proportional and integral coefficients of the PI controller for subharmonics (dynamically adjusted to adapt to changes in the power grid). : No. Taiwan APF targets the first Reference values ​​for the PI parameters of the second harmonic (determined through offline debugging, such as...) , (to ensure steady-state suppression accuracy). Parameter adjustment factor (default) , (The influence of control impedance and SVG output on PI parameters). : time t The rate of change of subharmonic grid impedance (unit: Ω / s, calculated from impedance data of the real-time monitoring module, reflecting the dynamic changes in grid impedance). The rate of change of SVG reactive power output at time t (unit: kvar / s, provided by the SVG scheduling module, reflecting the rhythmic change of SVG output, conforming to the document requirement of "adjusting suppression parameters according to grid impedance changes and SVG output"). The online evolution mechanism, by periodically trying new parameter combinations, evaluating their performance, and retaining the optimal strategy, enables the APF scheduling module to have self-optimization capabilities, thereby maintaining high harmonic suppression performance in long-term operation. Fitness calculation for new parameter combinations: , : time t Taiwan APF targets the first The fitness of new parameter combinations for subharmonics (value range 0-1, the higher the fitness, the better the parameter, and the parameter is retained). Weighting coefficients for THD improvement, response time, and loss (satisfying) ,default , , Prioritize ensuring THD compliance. The maximum permissible THD value for the power grid (document requires ≤2%), therefore... ), : time t Taiwan APF Suppression The actual THD value after subharmonics (collected by the real-time monitoring and feedback module, reflecting the suppression effect). Target THD value (default value is 1.5%, which is higher than the document requirement of 2%, leaving a margin). : Baseline suppression response time (default value is 10ms, which is the upper limit of the design response of APF). : time t Taiwan APF targets the first The actual suppression response time of subharmonics (collected by the real-time monitoring module, the time required from harmonic exceeding the standard to THD meeting the standard). APF baseline operating loss (determined by equipment rated parameters, such as...) ), : time t Taiwan APF Suppression The actual operating loss during harmonics (calculated from voltage and current data collected by the real-time monitoring module, reflecting economic efficiency). The APF scheduling module and the SVG scheduling module maintain high-frequency data synchronization, enabling the two modules to avoid harmonic mitigation failure due to mutual interference in scheduling behavior. Real-time reactive power output data, grid impedance change prediction data, and output rhythm change notifications from the SVG scheduling module are important inputs to the APF scheduling module. When a step-like change in SVG output may lead to an increase in a certain frequency harmonic, the APF scheduling module can generate harmonic suppression strategy adjustments within a millisecond response time, avoiding harmonic surges by correcting suppression parameters in advance. If necessary, the APF scheduling module can also issue harmonic suppression conflict warnings to the SVG scheduling module, enabling the SVG to adjust its output rhythm and provide a more suitable suppression environment for the APF, thereby ensuring that reactive power compensation and harmonic mitigation do not conflict with each other. For multi-APF deployment scenarios, the APF scheduling module achieves multi-device collaboration through a cluster load balancing-cooperative suppression algorithm. Single APF load priority calculation: Single APF suppression current distribution ratio Single APF final suppression current command: , : time t The load priority of each APF (value range 0-1, higher priority means the device is currently under heavier load and should be assigned fewer tasks). : time t The operating temperature of the APF (unit: °C) is collected by the real-time monitoring module. , (Higher temperature, higher priority) : time t The operating load factor of the APF (range 0-1, calculated from the actual output current and rated current collected by the real-time monitoring module) is... (Higher load rate, higher priority) : time t The inhibitory efficacy of the APF (value range 0-1, determined by...) Calculation, i.e. Higher efficiency means lower priority, which aligns with the document's requirement to "prioritize equipment with lighter loads and higher suppression efficiency to handle complex harmonic suppression." : Minimum and maximum suppression effectiveness (default) , ), : time t Taiwan APF targets the first The current distribution ratio for suppressing subharmonics (to ensure load balance across multiple APFs). Total number of devices in the APF cluster (statistics from the real-time monitoring module). : At time t, the APF cluster targets the first The total suppression current requirement for subharmonics (the sum of all APF targeted suppression currents calculated by Equation 5, i.e.) ), : time t The final suppression current command of each APF (Advanced Harmonic Filter) balances targeting accuracy and load balancing to avoid overload or resource idleness in a single unit. Based on the operating temperature, load rate, and current suppression efficiency of each APF device, the algorithm automatically allocates suppression tasks through real-time comparison of device status. Devices with lighter loads, lower temperatures, or better health conditions undertake more harmonic suppression tasks, preventing overload of a single device or resource idleness in some devices. Through a 2ms-level synchronization mechanism within the cluster, the APF scheduling module ensures consistency in harmonic suppression targets across multiple APF units, avoiding harmonic bounce and harmonic phase superposition caused by differences in execution between devices, thereby improving the overall suppression effect. At the system coordination level, the APF scheduling module is a crucial data provider for the collaborative scheduling decision module. Its output, including actual THD values, harmonic suppression rates, device load rates, suppression capacity redundancy, and scheduling execution feedback, is used by the collaborative scheduling decision module for conflict identification, priority adjustment, and cross-device adjustment capability matching. When the collaborative scheduling decision module identifies a potential conflict between harmonic suppression and reactive power compensation, it constructs a negotiation strategy based on data provided by both the APF and SVG to drive the entire system to achieve a globally optimal scheduling result. Through deep data linkage with the AI ​​prediction module, SVG scheduling module, and collaborative scheduling decision module, the APF scheduling module integrates harmonic mitigation into the overall system's collaborative optimization, enabling the power grid to maintain stable harmonic levels and high-quality operation under various operating conditions, with multiple devices, and at different time periods.

[0022] In one embodiment, the collaborative scheduling decision module constructs a collaborative framework based on multi-agent game theory and hierarchical cross-temporal consensus. This framework establishes a coordination mechanism between different devices, levels, and time scales, transforming the system's decision-making logic from independent responses by a single module to a global optimization decision-making process involving collaborative participation from all modules. Since reactive power compensation and harmonic suppression often result in target conflicts or mutual interference in actual power grid operation, the collaborative scheduling decision module uses a mechanism of prediction, negotiation, and evolution to resolve potential conflicts before they affect system stability, ensuring the reliability and consistency of operational commands. The collaborative scheduling decision module first encapsulates SVG, APF, new energy equipment, and energy storage systems as independent agents, enabling each agent to propose local demands based on its own operating status, regulation capacity, and equipment health. To avoid global instability that may result from direct competition among agents, the module employs a hierarchical game theory mechanism, stratifying devices into main grid, distribution grid, and terminal layers. The local payoff function for an agent within a layer (taking a specific agent in the main grid layer as an example) is defined. For example, (Can be SVG / APF / new energy / energy storage) Optimal resource allocation ratio for agents within a layer (based on maximizing benefits): Cross-layer global consensus resource allocation (mainnet layer) Distribution network layer Terminal layer ): , : t time layer Inner agent Local benefits (values ​​0-1, considering comprehensive adjustment capabilities, health, and economic efficiency; higher benefits lead to priority resource allocation). Weights of adjustment ability, health, and economy (total = 1, default) , , (The collaborative module can be dynamically adjusted.) : agent at time t Actual adjustment capability (SVG takes) (Adjusting redundancy, from the SVG scheduling module:) ); APF take (Suppression performance, from the APF scheduling module:) ); New energy sources (Reactive power regulation potential, derived from the real-time monitoring module) : t time layer The sum of the maximum regulatory capacity of the intra-layer agents (SVG is the sum of the maximum regulatory capacity of the intra-layer agents) The sum of APF is the sum of the values ​​within the layer. (=1) sum, statistics from the real-time monitoring module). : intelligent agent at time t Health (SVG derived from voltage sensitivity) With operating temperature ( APF comes from load factor. With temperature (Provided by the real-time monitoring module) : agent at time t The operating losses (SVG takes the switching losses, APF takes the suppression losses, and the real-time monitoring module collects and calculates them). : t time layer The maximum permissible total loss of the internal intelligent agent (equipment rated parameters, called by the real-time monitoring module). : t time layer Inner agent The local optimal resource allocation ratio (ensuring that resources within the layer are given priority to agents with high returns). Hierarchical weights (mainnet layer) Distribution network layer Terminal layer Based on the importance of the power grid hierarchy, it conforms to the logic of "prioritizing the main grid". : agent at time t The global resource allocation ratio (integrating local optima within the fusion layer and cross-layer weights to achieve global consensus). Within each layer, agents reach a locally optimal coordination solution through game theory, while cross-layers reach a global decision result through consensus negotiation. This layered approach effectively reduces the system's decision-making complexity, enabling cross-device and cross-regional adjustment needs to converge to an executable global scheduling solution with higher efficiency. To ensure the system maintains consistency under complex operating conditions, the collaborative scheduling decision module relies on the full prediction data from the AI ​​prediction module. By predicting confidence levels, risk levels, and medium- to long-term operating trends, this module can identify potential future reactive power compensation pressures, harmonic exceedance risks, or the impact of renewable energy fluctuations, thus reserving future adjustment space in the current scheduling strategy. SVG medium- to long-term adjustment space reservation: Long-term inhibition space reserve in APF: , : time t Taiwan's SVG has reserved adjustment space for medium and long-term demand (ensuring that there is still adjustment capability during peak reactive power demand in the future, avoiding timing discrepancies). : Reserved coefficient at time t (based on AI-predicted risk level: minor risk) ,medium ,serious , (From AI prediction module) The medium- to long-term (2-24h) reactive power demand forecast output by the AI ​​prediction module at time t (a core input for cross-temporal and spatial collaboration, reflecting "forward-looking planning"). The short-term (15min-2h) reactive power demand forecast output by the AI ​​prediction module at time t (used as a benchmark for medium- and long-term demand). : time t Taiwan's APF (Advanced Harmonic Filtering) provides a suppression margin for medium- and long-term harmonic demands (to avoid insufficient suppression when harmonic loads increase in the future). The predicted THD value of the medium- and long-term harmonics (APF) output by the AI ​​prediction module at time t (APF reserved). This cross-temporal and spatial collaborative logic enables the system's scheduling to no longer rely solely on the current state, but to make forward-looking plans based on future operating trends, avoiding insufficient adjustment capabilities or strategy conflicts caused by time-series discrepancies. In terms of conflict handling, the collaborative scheduling decision module constructs a prediction-negotiation-evolution system. Conflict risk calculation (reactive power-harmonic coupling conflict as an example): Adjustment of SVG reactive power target after conflict negotiation: Adjustment amount of APF harmonic suppression target after conflict negotiation: , : Conflict risk level at time t (value ranges from 0 to 1) (Assessed as "high risk of conflict," requiring the initiation of negotiations) Weights of reactive power deviation and THD deviation (total = 1, default) , When the voltage is unstable Increased to 0.6). The deviation between the reactive power demand predicted by AI at time t and the total regulation capacity of SVG ( , For the short-term (15min-2h) reactive power demand of the AI ​​prediction module, (total SVG adjustment redundancy). Total SVG regulation capability at time t (within layer) (The sum, provided by the SVG scheduling module). The deviation between the harmonic load predicted by AI at time t and the total suppression capability of APF ( , For AI prediction module harmonic THD, (This represents the current total inhibitory effect of APF). : Target THD value at time t (document requires ≤2%, therefore...) ), : The prediction confidence of the AI ​​prediction module at time t (values ​​range from 0 to 1; the lower the confidence, the greater the risk of conflict, which is consistent with the logic that "the probability of conflict is high when the prediction is inaccurate"). : after negotiation at time t Reactive power target adjustment of SVG (based on local target) Concessions were made, taking into account the harmonic requirements of the APF. SVG intelligent agent at time t The negotiation concession coefficient (values ​​0-0.3, health level) hour To avoid device overload (from the SVG scheduling module). : The trade-off coefficient between SVG and APF at time t (values ​​range from 0 to 0.2, representing the APF's inhibitory efficacy). hour Prioritize fulfilling APF requirements (from the APF scheduling module). The increase in harmonic suppression demand of APF at time t ( (Provided by the APF scheduling module) : APF negotiation concession coefficient, and SVG benefit exchange coefficient (logic same as SVG). Based on APF health status, Based on SVG reactive power gap), The incremental reactive power compensation demand of SVG at time t ( (Provided by the SVG scheduling module). When the prediction data provided by the AI ​​prediction module indicates that reactive power compensation and harmonic suppression may conflict at a certain time, this module will initiate a negotiation mechanism in advance. Through concessions and exchanges of interests, the goals of each agent will be adjusted without affecting global stability. The types of conflicts handled and the negotiation results will be stored in the policy library, enabling the system to reuse effective solutions in future operating conditions. This will form a self-evolving ability for conflict resolution strategies, making the collaborative process increasingly accurate and efficient. Policy evolution fitness evaluation algorithm (conflict handling policy library update): Conflict handling policy fitness: , : time t The fitness of each conflict resolution strategy (values ​​0-1, Retained in the policy library to achieve "self-evolution") Weights for conflict resolution effectiveness, system indicators, and economic efficiency (total = 1, default) , , ), : t-time strategy Conflict resolution rate ( , The better the conflict risk assessment in the next cycle after strategy implementation, the better the resolution effect. The closer to 1), : t-time strategy System indicator compliance rate ( , This is the actual power factor (provided by the SVG scheduling module). (Document Requirements) : t-time strategy Economic indicators ( , Total system loss (provided by the real-time monitoring module). (For maximum allowable losses). Regarding the generation of scheduling instructions, the collaborative scheduling decision module integrates the operating status, equipment health, redundancy capacity adjustment, and instruction execution feedback from the SVG scheduling module, APF scheduling module, and real-time monitoring and feedback module. Through hierarchical consensus and cross-agent negotiation, it forms the globally optimal correction instruction. SVG global reactive power correction instruction: APF Global Harmonic Suppression Correction Command: , The coordination module at time t sends the message to the first... The global reactive power correction instruction of the Taiwan SVG (final execution instruction, integrating global allocation, conflict negotiation, and cross-temporal reservation). Total reactive power demand of the power grid at time t ( , (This refers to the real-time reactive power deficit, which originates from the real-time monitoring module). The coordination module at time t sends the message to the first... Taiwan APF's global harmonic suppression correction command (target THD value, must be ≤2%). Total harmonic suppression requirement of the power grid at time t ( , (For real-time THD, sourced from the real-time monitoring module). It outputs reactive power output correction instructions, cross-device resource call instructions, and target weight dynamic adjustment suggestions to the SVG scheduling module, enabling the SVG to execute scheduling tasks in a way that better aligns with global objectives in the next time period. Similarly, it outputs harmonic suppression correction instructions, multi-APF cluster load allocation instructions, and suppression priority adjustment suggestions to the APF scheduling module, ensuring that harmonic mitigation remains consistent with the reactive power compensation process and avoiding harmonic bounce or insufficient suppression due to scheduling differences. The collaborative scheduling decision module simultaneously performs decision correction and algorithm iteration functions throughout the system. After receiving execution feedback from the APF and SVG scheduling modules, it updates the prediction deviation correction suggestions and algorithm evolution parameter adjustment instructions based on the feedback deviation, and pushes these adjustments to the AI ​​prediction module, enabling the prediction model to be optimized in the next cycle. Through this closed-loop iterative mechanism, a stable information flow, strategy flow, and feedback flow are formed among the various modules of the system, ensuring the adaptive and self-evolving capabilities of the entire system in long-term operation. The collaborative scheduling decision module transforms the potential contradiction between reactive power compensation and harmonic suppression into a coordinated overall goal, enabling the entire system to maintain a consistent, efficient, and stable operating state under complex power grid conditions.

[0023] In one embodiment, the real-time monitoring and feedback module continuously monitors parameters such as voltage, current, active power, reactive power, harmonic content, equipment temperature, operating load rate, and switching losses through end-to-end monitoring units deployed on the power grid bus, critical load side, and equipment output. This enables the system to acquire the real operating status of the power grid and equipment within millisecond timescales. A dual-link transmission method using Ethernet and 5G ensures that this high-frequency collected data can be stably and with low latency transmitted to the AI ​​prediction module, SVG scheduling module, APF scheduling module, and collaborative scheduling decision module, thus providing a solid data foundation for the entire system's scheduling logic. The real-time monitoring and feedback module is not only responsible for raw data acquisition but also for verifying and correcting data reliability. Through a multi-source data virtual-real fusion verification algorithm, the real-time monitoring data of the physical power grid is compared with the virtual prediction data of the AI ​​prediction module to identify acquisition errors, temporary equipment failures, and real fluctuations in the power grid. The virtual-real data deviation rate is calculated (taking voltage parameters as an example; other parameters such as current and THD follow the same logic): Data confidence calculation: Deviation type determination (output "acquisition error / equipment failure / actual fluctuation"): , Parameters at time t The virtual-to-real deviation rate (dimensionless, reflecting the relative deviation between physical monitoring data and AI prediction data) It can represent all dimensions of parameters monitored by modules such as voltage, current, and THD. Parameters at time t The physical monitoring values ​​(collected by monitoring units deployed on the power grid bus and load side, such as the effective voltage value) (such as harmonic THD values, etc.) Parameters at time t The AI ​​virtual prediction values ​​(from the AI ​​prediction module, such as short-term voltage prediction values ​​and THD prediction values, which meet the cross-module data linkage requirements of "virtual-real fusion verification") Parameters at time t The rated / reference values ​​(such as grid rated voltage 10kV, THD reference value 2%, from equipment parameters or grid standards). Parameters at time t The data confidence level (value 0.1-1.0, output to the AI ​​prediction module and SVG / APF scheduling module to guide their data usage strategy; for example, when the confidence level is <0.5, the SVG scheduling module will prioritize referring to the historical best parameters). Confidence adjustment factor ( , (Through offline testing and calibration, the rate of confidence decay with deviation is controlled). Deviation rate threshold ( , , This is the upper limit of the "acceptable deviation". (This is the lower limit for "serious deviation," set based on the accuracy requirements of power grid operation.) Physical monitoring data at time t The standard deviation of the fluctuation (reflecting the stability of the collected data, calculated from the sampled data within the most recent 100ms) (for fluctuation threshold) Physical monitoring data at time t The rate of change (reflects whether the equipment is stationary; for example, when the equipment is faulty, the voltage / current may be constant, and the rate of change will approach 0). AI prediction data at time t The self-bias rate (derived from the "prediction confidence" output by the AI ​​prediction module) Through a virtual-real fusion verification process, the system can determine the reliability of data and assign a confidence level to each data point to guide the subsequent use strategies of each module. If a deviation occurs at a monitoring point, the verification logic will distinguish whether it is a measurement error or an abnormal fluctuation in the power grid, and the anomaly identifier will be synchronously pushed to each scheduling module to reduce the risk of misjudgment during decision-making. To further shorten the response latency to real-time status, the real-time monitoring and feedback module sets up edge computing nodes at key monitoring points. Edge computing nodes can complete preliminary data preprocessing and real-time analysis locally, and feed back the processing results to each module with a millisecond delay. Through this localized processing method, the system can take adjustment measures in the shortest possible time in scenarios such as rapid changes in harmonics, node voltage fluctuations, and sudden changes in new energy output, reducing the probability of voltage deviation propagation or harmonic amplification. The real-time monitoring and feedback module has also established an anomaly precursor hierarchical identification system. Through trend analysis and anomaly pattern recognition of the collected data, anomalies in the power grid or equipment are divided into three levels: minor, moderate, and severe, and different feedback strategies are triggered to different modules according to the level. Anomaly characteristic value calculation (fusion parameter fluctuation amplitude and trend): Anomaly level membership calculation (output "Slight / Moderate / Severe"): , , Parameters at time t Abnormal characteristic values ​​(dimensionless, combining fluctuation amplitude and trend of change; the larger the value, the higher the degree of abnormality). Weights of fluctuation range and rate of change ( , (Prioritize parameter fluctuation range) Anomaly analysis time window (default) This matches the module's "millisecond-level response" requirement, ensuring real-time capture of early signs of anomalies. Parameters at time t For the Membership degree of level anomalies (values ​​0-1, Represents mild, Representing the middle, (Representing seriousness) : No. Baseline feature values ​​for major anomalies (obtained through offline training, such as minor anomalies) ,medium ,serious ), : No. Characteristic variance of level anomalies ( , , (This reflects the degree of dispersion of the benchmark characteristics). Parameters at time t The final anomaly level is output to the corresponding module: minor → AI prediction module optimizes the algorithm; moderate → SVG / APF scheduling module fine-tunes parameters; severe → collaborative scheduling decision module provides emergency coordination, conforming to the "tiered proactive feedback" logic. When the anomaly is at the minor level, the module pushes relevant information to the AI ​​prediction module as a basis for algorithm optimization, enabling it to improve predictive adaptability in the next cycle. When the anomaly reaches the moderate level, it triggers a request for parameter fine-tuning to the SVG and APF scheduling modules, allowing the scheduling strategy to proactively compensate for unstable factors. When the anomaly is identified as severe, it immediately sends an emergency signal to the collaborative scheduling decision module, prompting the system to enter emergency coordination mode, enabling multiple devices to adjust uniformly in a short time to ensure the safe operation of the power grid. At the data feedback level, the real-time monitoring and feedback module not only transmits the collected results to other modules but also receives detailed scheduling instructions from the SVG and APF scheduling modules and compares them with real-time monitoring data. This feedback comparison mechanism enables the system to determine the execution deviation of scheduling instructions and promptly transmit deviation information to support functions such as time-series backtracking correction, dynamic evolution, and strategy evolution. Through this scheduling-feedback bidirectional loop, the real-time monitoring and feedback module enables the behavior of the SVG scheduling module and the APF scheduling module to be reflected in the system status in real time and accurately, thereby enhancing the accuracy and stability of overall regulation. The real-time monitoring and feedback module also undertakes the function of safeguarding the stability of the data transmission link. Because it uses Ethernet plus 5G dual-link redundancy for data transmission, when the primary link fails, the system can switch to the backup link in a very short time, ensuring that data is not lost or interrupted. Primary Link (Ethernet) Status Assessment Index: Link switching trigger conditions: , The status index of the main link (Ethernet) at time t (values ​​range from 0 to 1, taking into account both transmission success rate and latency; the lower the value, the worse the link status). The number of successfully transmitted data packets on the main link at time t (statistics based on the CRC32 check result of the module; a successful check indicates success). : The total number of data packets transmitted on the main link at time t (including successful and failed data packets, reflecting the link load). : Average transmission delay of the main link at time t (unit: ms, required for modules ≤ 5ms, the greater the delay, the greater the impact on scheduling accuracy). Delay penalty coefficient (default) (This amplifies the negative impact of latency on link status). : Link switching command (1 = switch, 0 = do not switch, conforms to the "Ethernet + 5G dual link redundancy" design, and the switchover is completed within 50ms when the main link fails). Link state threshold (default) If the value is lower than this, it is considered a main link failure. Last link handover time (unit: ms). Switch cooldown time (default) To avoid data interruptions caused by frequent switching, a retransmission mechanism, CRC32 checksum, and TLS 1.3 encryption ensure the integrity, security, and availability of data during transmission, enabling the entire system to maintain high-reliability communication quality even in complex environments. The retransmission mechanism ensures that data is not lost due to momentary link fluctuations, network packet loss, or brief interference during Ethernet and 5G dual-link transmission. When pushing monitoring data to other modules, the real-time monitoring and feedback module sets a transmission timeout threshold for each data packet. If a data packet does not receive an acknowledgment within the set time limit, the module immediately initiates retransmission logic. This allows the system to maintain data integrity during link jitter or momentary network quality degradation, preventing prediction or scheduling failures due to undelivered reactive power, voltage, and harmonic data. The number of retransmissions is limited to three to ensure that the system does not get bogged down in repeated requests, consuming bandwidth or causing communication blockage between modules when consecutive transmission failures occur. If retransmission fails after three attempts, the real-time monitoring and feedback module will issue a transmission anomaly alarm to the collaborative scheduling decision module, enabling the system to implement degradation strategies as quickly as possible and ensuring that downstream modules do not make incorrect decisions due to missing data. This retransmission mechanism allows the real-time monitoring and feedback module to maintain reliable delivery of monitoring data even in complex power grid environments, wireless interference scenarios, or cross-site data link fluctuations, ensuring the continuity and stability of the entire system's scheduling link. Retransmission trigger determination: Retransmission count control (to avoid infinite retransmissions): , Retransmit the trigger signal at time t (1 = trigger, 0 = do not trigger, based on both timeout and verification results). : The sending timestamp of the data packet at time t (accurate to milliseconds, compared with the receiving timestamp to determine if a timeout has occurred). Transmission timeout threshold (default) (This conforms to the timeout settings of the module's "retransmission mechanism"). : The CRC32 check result of the data packet at time t (1 = check passed, 0 = check failed, the module calculates the data packet check code and compares it with the sender). : The cumulative number of data packets retransmitted at time t (initially 0, incremented by 1 after each retransmission). Maximum number of retransmissions (default) (This meets the document's requirement of "retransmission count ≤ 3 times"). The final retransmission instruction at time t (if retransmission fails, a "transmission anomaly alarm" is pushed to the collaborative scheduling decision module to ensure timely system degradation) is executed. The real-time monitoring and feedback module uses CRC32 checksum to identify whether data packets are corrupted during transmission. Each data packet collected by the monitoring unit and to be transmitted includes a CRC32 checksum, calculated by the sender based on the original data content. When the data packet arrives at the receiving module, the receiver recalculates the CRC32 checksum based on the received data and compares it with the attached checksum. This effectively identifies bit flips or data loss caused by noise, electromagnetic interference, link jitter, etc., enabling the system to reject erroneous data from entering the AI ​​prediction module, SVG scheduling module, or APF scheduling module, preventing erroneous data from affecting prediction accuracy or scheduling execution. CRC32 checksum has extremely low overhead, making it suitable for high-frequency data transmission scenarios, allowing the real-time monitoring and feedback module to quickly complete checksum processing even at millisecond-level data acquisition frequencies. In case of checksum failure, it can be linked with the retransmission mechanism to immediately trigger data retransmission, ensuring that the data ultimately obtained by downstream modules is completely reliable. The combination of CRC32 checksum and retransmission mechanism provides robust "detectable-recoverable" data transmission, forming the most fundamental and critical integrity guarantee in the real-time data link. The real-time monitoring and feedback module employs TLS 1.3 encryption to protect the data link throughout the entire process, ensuring that collected data is not stolen, tampered with, or forged during transmission. TLS 1.3, as a mainstream secure transmission protocol, provides encrypted encapsulation capabilities for monitoring data, device status data, and cross-domain feature mapping data, ensuring that data cannot be read or modified by unauthorized entities even after passing through public networks or long-distance transmission links. Encryption mechanisms protect all monitoring data and prevent the leakage of sensitive device status information such as "switching losses, operating temperature, and aging degree" during transmission. A handshake mechanism ensures the authentication of both communicating parties, preventing unauthorized nodes from impersonating system modules to inject false data into other modules and eliminating the risk of malicious tampering with the scheduling link. The provided data integrity protection mechanism, together with CRC32 checksum, forms a dual guarantee, enabling the receiving end to simultaneously verify both encryption integrity and physical layer integrity, thus elevating the system's data reliability to a higher level. Through TLS 1.3 encryption, the real-time monitoring and feedback module not only ensures the confidentiality and tamper-proofness of data during transmission, but also enables the system to have a high level of security to cope with cross-domain data transmission, remote site access, and multi-device communication scenarios.

[0024] In one embodiment, the present invention provides a method for applying the above-mentioned AI prediction-based SVG-APF cooperative optimization scheduling system, which includes the following steps: Step 1: Comprehensive Data Acquisition and Cross-Domain Adaptation. In this step, the real-time monitoring and feedback module continuously collects comprehensive data, including voltage, current, power, harmonic content, and equipment operating status. After preprocessing such as outlier removal and data standardization, this data undergoes multi-source data fusion verification to ensure data quality and authenticity. Simultaneously, through cross-domain feature mapping technology, common knowledge and features learned from historical data and other similar power grids are applied to the current power grid, completing data integration and adaptation. Finally, the processed data is pushed to the AI ​​prediction module, SVG scheduling module, APF scheduling module, and collaborative scheduling decision module.

[0025] Step Two: Cross-Domain Integrated AI Proactive Prediction. Upon receiving comprehensive data, the AI ​​prediction module utilizes its constructed cross-domain knowledge base and prediction model. Employing a prediction system combining spatiotemporal causal fusion, cross-domain knowledge transfer, and multimodal feature reasoning, it generates predictions for future load, reactive power demand, and harmonic load. These predictions include not only quantitative values ​​but also confidence levels (indicating the reliability of the prediction) and risk level labels, allowing subsequent decision-making modules to conduct risk assessments and adjust strategies based on this information. During the prediction process, the AI ​​prediction module also mines load transmission relationships between grid nodes through spatiotemporal correlation analysis, effectively mitigating prediction biases caused by false data interference using causal reasoning methods. This module also reuses common knowledge from across the power grid, integrating multimodal features such as physical (e.g., line parameters), environmental (e.g., meteorological conditions), and behavioral (e.g., user electricity consumption habits) features for reasoning to improve the comprehensiveness and accuracy of the predictions. Furthermore, the module can simultaneously identify ambiguous areas in the prediction results and dynamically adjust the data collection frequency for those areas to obtain more information for refining the predictions.

[0026] Step 3: Multi-dimensional Balanced SVG Precise Scheduling. Based on the reactive power demand prediction results provided by the AI ​​prediction module, the SVG scheduling module activates its dual-weight dynamic balancing and three-dimensional balance optimization scheduling algorithm. It considers the dual-weight allocation of voltage sensitivity and equipment health, as well as the three-dimensional balance of reactive power demand urgency, equipment adjustment capability, and scheduling economic cost. Combined with reactive power fluctuation prediction and smoothing algorithm, time-series backtracking correction mechanism, and dynamic redundancy adaptive allocation strategy, it generates instructions to perform precise reactive power compensation scheduling for SVG equipment to maintain the power factor of the power grid within a reasonable range.

[0027] Step 4: End-to-End Evolutionary APF Targeted Suppression. Next, the APF scheduling module receives harmonic load prediction results and real-time monitoring data from the AI ​​prediction module. It then uses adaptive Fourier transform combined with a harmonic mode identification algorithm (identifying steady-state, transient, and interharmonics) to perform end-to-end tracing to determine the harmonic source. This module then initiates a triple mechanism of targeted suppression, dynamic evolution, and online evolution to specifically suppress harmonics in the power grid. In multi-APF deployment scenarios, this step also executes a cluster load balancing-cooperative suppression algorithm to ensure that each APF works collaboratively, maximizing the harmonic suppression effect.

[0028] Step 5: Layered, Spatiotemporal Collaborative Decision-Making. In this step, the collaborative scheduling decision-making module encapsulates SVG, APF, new energy equipment, and energy storage systems as independent intelligent agents. Through a multi-agent game framework, layered game theory is implemented at the main grid, distribution network, and terminal layers. This module, combined with results from the AI ​​prediction module, identifies potential conflicts between reactive power compensation and harmonic suppression in advance. When conflicts occur, they are resolved and consensus reached through prediction, negotiation (using mechanisms such as concessions and interest exchanges), and an evolutionary system. This generates globally optimal correction instructions, guiding SVG and APF to perform higher-level collaborative operations and optimizing the overall power quality of the grid. The conflict type and negotiation results are stored in a strategy library, enabling the self-evolution of conflict resolution strategies. Simultaneously, this step utilizes medium- and long-term forecast results to reserve sufficient space for future grid regulation.

[0029] Step 6: End-to-End Feedback and Algorithm Iteration. The real-time monitoring and feedback module collects system operation data and scheduling effect data, including full-dimensional grid parameters and the actual operating performance and power consumption of SVG / APF devices. This data undergoes multi-source data fusion verification and different feedback strategies are matched according to the anomaly level. This real-time feedback data is sent back to each module for evaluation, parameter adjustment, and iterative optimization of its internal algorithm models. For example, the AI ​​prediction module uses actual operating data to correct its prediction model, the SVG and APF scheduling modules adjust their scheduling strategies based on the actual compensation and suppression effects, and the collaborative scheduling decision module optimizes its negotiation mechanism based on the success rate of conflict resolution, thus forming a continuously learning and evolving intelligent closed loop.

[0030] In one embodiment, standardized, highly reliable communication protocols are used for data exchange between the modules constituting the system and between the system and external devices. Data transmission protocols include ModbusTCP, MQTT (Message Queuing Telemetry Transport), and RESTfulAPI (Representative State Transfer Application Programming Interface) to adapt to communication needs in different scenarios. ModbusTCP is commonly used for data exchange between industrial control devices, MQTT is suitable for lightweight, publish / subscribe messaging, and RESTfulAPI provides a flexible network service interface. All transmitted data is encapsulated in JSON (JavaScript Object Notation) format, a lightweight data exchange format that is easy for humans to read and write, and also easy for machines to parse and generate. The encapsulated data includes not only the actual data value but also a timestamp of data generation (ensuring data timeliness), confidence level (derived from the reliability assessment of the AI ​​prediction module), source identifier (distinguishing the source device or module of the data), and anomaly marker (indicating whether the data contains anomalies or alarm information). To ensure the reliability and security of data transmission, data communication employs Ethernet plus 5G dual-link redundant transmission, meaning that if one communication method fails, it can immediately switch to another, avoiding single points of failure. Meanwhile, all data is protected by CRC32 checksum and TLS1.3 encryption during transmission. CRC32 ensures that no bit errors occur during data transmission, while TLS1.3 provides strong end-to-end encryption to ensure the confidentiality and integrity of the data.

[0031] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AI-predictive SVG-APF collaborative optimization scheduling system, characterized in that: The system comprises an AI prediction module, an SVG scheduling module, an APF scheduling module, a collaborative scheduling decision module, and a real-time monitoring and feedback module. These modules work in close collaboration to form a fully integrated intelligent closed loop encompassing data acquisition, prediction, scheduling, feedback, and evolution. The AI ​​prediction module outputs load, reactive power demand, and harmonic load predictions with confidence levels and risk levels. The SVG scheduling module uses these predictions to achieve precise reactive power compensation scheduling, maintaining a stable power factor in the grid. The APF scheduling module targets and suppresses harmonics, controlling the harmonic content of the grid. The collaborative scheduling decision module resolves the coupling conflict between reactive power compensation and harmonic suppression, achieving global optimization. The real-time monitoring and feedback module collects data across all dimensions and provides hierarchical proactive feedback to support algorithm iteration across all modules.

2. The AI-predictive-based SVG-APF collaborative optimization scheduling system according to claim 1, characterized in that: The AI ​​prediction module constructs a prediction system that integrates spatiotemporal causal fusion, cross-domain knowledge transfer, and multimodal feature reasoning. It integrates multi-source data, including historical operation data, real-time power grid data, cross-domain power grid common data, equipment health status data, new energy access fluctuation data, and power grid topology switching records. After preprocessing such as outlier removal, data standardization, feature engineering, intelligent identification of power grid operation scenarios, and cross-domain feature mapping, the prediction results are updated regularly through proactive uncertainty exploration and dynamic correction logic.

3. The AI-predictive-based SVG-APF collaborative optimization scheduling system according to claim 1, characterized in that: The SVG scheduling module adopts a scheduling algorithm with dual-weight dynamic balancing and three-dimensional balance optimization. First, it allocates the reactive power output ratio by weighting voltage sensitivity and equipment health. Then, it performs three-dimensional balance dynamic optimization based on the urgency of reactive power demand, equipment adjustment capability, and scheduling economic cost. This is combined with a reactive power fluctuation prediction and smoothing algorithm, a timing backtracking correction mechanism, and a dynamic redundancy adaptive allocation strategy.

4. The AI-predictive-based SVG-APF collaborative optimization scheduling system according to claim 1, characterized in that: The APF scheduling module uses an adaptive Fourier transform combined with a harmonic mode identification algorithm to identify steady-state, transient, and interharmonic harmonic modes and trace key links across the entire chain. It employs a triple mechanism of targeted suppression, dynamic evolution, and online evolution, and designs a cluster load balancing-cooperative suppression algorithm for multi-APF deployment scenarios.

5. The AI-predictive-based SVG-APF collaborative optimization scheduling system according to claim 1, characterized in that: The collaborative scheduling decision-making module constructs a multi-agent game and hierarchical cross-temporal consensus collaborative framework, which encapsulates SVG, APF, new energy equipment, and energy storage system as independent intelligent agents, and implements hierarchical game according to the main grid layer, distribution network layer, and terminal layer. It resolves conflicts through a prediction-negotiation-evolution system and reserves future adjustment space in combination with medium and long-term forecasts.

6. The AI-predictive-based SVG-APF collaborative optimization scheduling system according to claim 1, characterized in that: The real-time monitoring and feedback module deploys full-link monitoring units at the power grid bus, critical load side, and equipment output end to monitor all dimensions of parameters, including voltage, current, active power, reactive power, harmonic content, equipment temperature, operating load rate, and switching losses. It adopts Ethernet plus 5G dual-link transmission and uses a multi-source data virtual-real fusion verification algorithm to distinguish between data errors and real fluctuations, matching different feedback strategies according to the anomaly level.

7. A method applied to the AI-predictive-based SVG-APF collaborative optimization scheduling system as described in claims 1-6, characterized in that: Includes the following steps: Full-dimensional data collection and cross-domain adaptation: The real-time monitoring module collects data from multiple sources, and after preprocessing, virtual-real fusion verification and cross-domain feature mapping, it is pushed to each module; Cross-domain integrated AI proactive prediction: The AI ​​prediction module calls a cross-domain knowledge base and outputs prediction results with confidence level and risk level labels; Multidimensional Balanced SVG Precise Scheduling: The SVG scheduling module generates scheduling instructions through dual-weight balancing and three-dimensional balancing algorithms; Full-link evolutionary APF targeted suppression: The APF scheduling module initiates dynamic evolution and online evolutionary suppression strategies through modality recognition and full-link tracing; Hierarchical cross-temporal collaborative decision-making: The collaborative scheduling decision-making module generates globally optimal correction instructions through hierarchical game theory among multiple agents; End-to-end feedback and algorithm iteration: Each module optimizes algorithm parameters based on feedback data, forming a closed loop.

8. The AI-predictive-based SVG-APF cooperative optimization scheduling method according to claim 7, characterized in that: The AI ​​prediction module mines the load transmission relationship of power grid nodes through spatiotemporal correlation, avoids interference from false data by combining causal reasoning, reuses common knowledge of cross-domain power grids, integrates physical, environmental and behavioral multimodal features for reasoning, and simultaneously identifies ambiguous prediction areas and adjusts the data collection frequency of those areas.

9. The AI-predictive-based SVG-APF cooperative optimization scheduling method according to claim 7, characterized in that: The collaborative scheduling decision module uses AI prediction to identify potential conflicts between reactive power compensation and harmonic suppression in advance. It resolves these conflicts proactively through a negotiation mechanism involving concessions and exchanges of benefits. The conflict types and negotiation results are stored in the strategy library, enabling the self-evolution of conflict resolution strategies.

10. The SVG-APF collaborative optimization scheduling system and method based on AI prediction according to claim 7, characterized in that: The modules communicate with each other using ModbusTCP, MQTT and RESTfulAPI protocols. Data is encapsulated in JSON format, including data generation timestamp, confidence level, source identifier and anomaly marker. It is transmitted through Ethernet plus 5G dual-link redundancy, and uses CRC32 check and TLS1.3 encryption to ensure data reliability and security.

Images