A method for dynamic load power balancing control of microgrids

Abstract

This application discloses a dynamic load power balancing control method for microgrids, relating to the field of power system automation and intelligent control technology. It includes: acquiring the power fluctuation amplitude, interruption duration, and priority dynamic decay rate of the load to form an initial feature set; quantifying the coupling degree between power fluctuation and interruption duration using fuzzy logic to determine a time-period fluctuation index; when the time-period fluctuation index exceeds the limit, calculating a weight adjustment coefficient by fusing multiple features using a neural network; searching for a target power supply sequence for high-weight loads using a genetic algorithm; updating load priorities through secondary fuzzy inference based on real-time power fluctuations on the supply side; calculating a response delay correction factor using a neural network; and adjusting the target power supply sequence to obtain the final power allocation scheme. This invention integrates multiple algorithms to achieve closed-loop control, improving the power balance and dynamic adaptability of the microgrid, ensuring stable power supply, and is applicable to microgrid load control scenarios.

Description

Technical Field

[0001] This application relates to the field of power system automation and intelligent control technology, and in particular to a method for dynamic load power balancing regulation of microgrids. Background Technology

[0002] As an important component of distributed energy systems, microgrids play a crucial role in ensuring reliable power supply in remote areas, industrial parks, and urban building clusters. Their importance lies in their ability to effectively integrate various renewable energy sources and achieve efficient utilization of local energy, thereby reducing dependence on the main power grid and improving energy efficiency.

[0003] Current microgrid load regulation methods often struggle to balance various actual demands when facing power supply fluctuations. Many solutions prioritize certain fixed load categories while neglecting the varying degrees of urgency among different loads in specific operating scenarios, leading to significant discrepancies between regulation results and actual on-site power demand. Other methods, while considering the time-of-use characteristics of loads, fail to adequately assess the drastic changes in load power and the system's ability to withstand power outages. This can result in situations where critical loads are excessively reduced or non-critical loads consume excessive resources during emergencies. These shortcomings make it difficult for microgrids to achieve truly fair and efficient power allocation during dynamic periods of power shortages.

[0004] The core technical challenge of dynamic load control in microgrids lies in accurately characterizing and quantifying the comprehensive importance of each load in real time. The importance of a load depends not only on its preset level label but also on multiple factors, including the sensitivity of the power consumption period, the magnitude of power demand fluctuations, and the tolerance for power outages. These factors are interconnected and interdependent: when a load is in a time-sensitive period, large power fluctuations and low outage tolerance can significantly impact the overall system stability; conversely, if the system cannot comprehensively weigh these different characteristics, it is prone to making inappropriate power cut-off decisions during power shortages. For example, during the evening peak hours of an industrial microgrid, the air conditioning system's power suddenly surges, while a critical experimental device has a highly sensitive power consumption period and extremely low outage tolerance. If the control system allocates power based solely on power size or simple priority, it may mistakenly cut off the experimental device's power, leading to production interruptions or data loss. Meanwhile, the air conditioning load, although high in power, has a relatively controllable impact from an outage. This decision-making error directly exposes the operational contradictions arising from the lack of a scientific and comprehensive evaluation mechanism.

[0005] Therefore, in dynamic scenarios of tight power supply, how to construct a mechanism that can calculate the comprehensive priority of each load in real time based on multi-dimensional characteristics such as load importance level, power consumption time sensitivity, power fluctuation amplitude and response delay tolerance, and rationally determine the power supply sequence and power allocation ratio accordingly, has become a key issue in realizing dynamic load power balance control of microgrids. Summary of the Invention

[0006] To address the aforementioned problems, in a first aspect, this application provides a method for dynamic load power balancing control of a microgrid, the method comprising the following steps:

[0007] S1. Obtain the power surge amplitude, interruption duration, and priority dynamic decay rate of each load in the microgrid to obtain the initial load feature set;

[0008] This step involves acquiring the power fluctuation magnitude, interruption duration, and priority dynamic decay rate of each load within the microgrid to form an initial load characteristic set. The priority dynamic decay rate reflects the degree of change in load power supply priority, thus upgrading load characteristics from static parameters to dynamic performance indicators.

[0009] S2. Based on the initial load feature set, fuzzy logic is used to calculate the degree of coupling between the power mutation amplitude and the interruption duration, and the time-period fluctuation index of each load is determined based on the degree of coupling.

[0010] Based on the initial load feature set, a fuzzy logic algorithm is used to quantify the degree of coupling between the power fluctuation amplitude and the duration of the interruption. Then, based on this membership value, the time-period fluctuation index of each load is determined, thereby achieving a quantitative characterization of the comprehensive fluctuation characteristics of the load during different power consumption periods.

[0011] S3. When the load period fluctuation index exceeds the preset fluctuation threshold, extract the priority dynamic decay rate and interruption tolerance graded quantization value of the load, generate a fusion feature vector through neural network fusion, calculate the weight adjustment coefficient in combination with the period fluctuation index, and obtain the adjusted weight set based on the weight adjustment coefficient.

[0012] The trigger condition is that the time-period fluctuation index exceeds a preset fluctuation threshold. This preset threshold is determined by statistically analyzing the historical time-period fluctuation indices of the microgrid, summing the average value to 0.5-1.5 times the standard deviation. For abnormal loads requiring key regulation, two core features are extracted: their priority dynamic attenuation rate and interruption tolerance grading quantification value. These features are then deeply fused using a neural network to generate a fused feature vector. This vector, combined with the load's time-period fluctuation index, is used to quantify and calculate adjustment coefficients, enabling dynamic correction of load weights. This ensures that the weights align with the actual load importance and fluctuation characteristics during specific time periods, providing a basis for subsequent key load selection.

[0013] S4. Extract the loads whose weights exceed a preset weight threshold from the adjusted weight set, and use a genetic algorithm to search for the power supply sequence based on the interruption duration of the loads exceeding the preset weight threshold to obtain the target power supply sequence;

[0014] By selecting high-weight loads from the adjusted weights, a genetic algorithm is used to search for the optimal power supply sequence based on the interruption duration, resulting in a standardized target power supply sequence that meets the power balance requirements of the microgrid. This provides a basic execution framework for subsequent dynamic correction, which not only improves the optimization targeting and computing efficiency, but also achieves the global optimal solution for the power supply sequence through the genetic algorithm, effectively improving the overall power load balancing effect of the microgrid.

[0015] S5. Obtain real-time power fluctuation data from the target power supply sequence, call the preset second fuzzy inference rule library to perform fuzzy logic calculation on the real-time power fluctuation data, interruption duration and priority dynamic decay rate to obtain the priority adjustment parameters of each load, and then combine the initial weight of each load to obtain the updated priority value.

[0016] Based on the target power supply sequence, real-time power fluctuation data from the power supply side is obtained. By fusing real-time power fluctuation data, interruption duration and priority dynamic decay rate through fuzzy logic, the power supply priority value of each load is dynamically updated. This allows the power supply priority to adapt to the real-time operating conditions of the power supply side and the dynamic changes in load priority. This provides an accurate and dynamic basis for determining the deviation of the subsequent power supply sequence, ensuring the real-time performance and dynamic adaptability of power regulation.

[0017] S6. If there is a deviation between the updated priority value and the interruption tolerance level quantization value, the actual response delay measurement value of the load with the deviation is calculated by the preset neural network and the difference between the expected interruption tolerance time limit to obtain the response delay correction factor.

[0018] For loads that deviate from the updated priority value and the interruption tolerance classification quantification value, a neural network is used to calculate the difference between the actual response delay measurement value and the expected interruption tolerance time limit, generating a response delay correction factor. The response delay correction factor quantitatively characterizes the degree of deviation between the actual power supply response of the load and the theoretical tolerance requirement, providing a quantitative correction basis for subsequent precise adjustment of the load power supply time, making up for the adaptation deviation between theoretical control parameters and actual operating conditions, and improving the accuracy and implementation of power control.

[0019] S7. Adjust the power supply timing of the loads in the target power supply sequence according to the response delay correction factor to obtain the final power allocation scheme.

[0020] Based on a response delay correction factor, the theoretical power supply timing of the target power supply sequence is adjusted to offset the actual system response delay, such as deviations caused by command issuance and device switching time, thus avoiding load interruption timeouts due to delays. The target power supply sequence is transformed into a final power allocation scheme with precise power supply timing, resolving the issue that while the theoretical sequence is feasible, its actual execution is inaccurate.

[0021] The microgrid dynamic load power balancing control method disclosed in this application constructs a load power balancing control system adapted to the dynamic changes in microgrid operating conditions through a collaborative design involving multi-feature extraction, fuzzy logic coupling analysis, neural network fusion optimization, genetic algorithm global search, and dynamic correction throughout the entire process. This system solves technical problems in traditional microgrid load control, such as single feature analysis, static and fixed weights, local optima in power supply sequence optimization, and poor adaptability of control schemes to actual operating conditions. It achieves precise, dynamic, and efficient load power balancing control in microgrids. Specifically, fuzzy logic overcomes the limitations of linear calculation, accurately quantifying the coupling relationships between multiple load features and achieving dynamic fusion of real-time data from the power supply side and load features. A neural network performs deep nonlinear fusion of multi-dimensional features, generating a fused feature vector that better matches the actual load operating conditions, while also achieving precise quantitative calculation of response delay deviation. A genetic algorithm performs a global search for the power supply sequence of the core control load, overcoming the limitations of local optima in traditional algorithms and ensuring the global optimality of the power supply sequence. The collaborative application of multiple algorithms makes the quantitative analysis and optimization of each stage more aligned with the nonlinear and multi-constraint characteristics of microgrids, improving the accuracy of control. Attached Figure Description

[0022] Figure 1 This is a flowchart of a microgrid dynamic load power balancing control method according to this application. Detailed Implementation

[0023] To further understand the content of this application, a detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. The following detailed description, in conjunction with the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative of the relevant invention and not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0024] like Figure 1 This embodiment of a microgrid dynamic load power balancing control method may specifically include:

[0025] S1. Obtain the power fluctuation amplitude, interruption duration and priority dynamic decay rate of each load in the microgrid to obtain the initial load feature set, wherein the priority dynamic decay rate reflects the degree of change in the load power supply priority;

[0026] In this embodiment, a microgrid sensor network is deployed to collect power data of each load, and the power fluctuation amplitude of each load is calculated by the time-domain change of the power data. The power supply status data of each load is collected by the sensor network, and the start time and recovery time of the power interruption of each load are counted. The time difference is calculated to obtain the interruption duration of each load. Based on the basic power supply priority of the load, the priority dynamic decay rate of each load is calculated by a preset priority dynamic decay model. The three feature datasets of power fluctuation amplitude, interruption duration, and priority dynamic decay rate are subjected to data standardization processing to obtain an initial load feature set.

[0027] Specifically, a microgrid sensor network is deployed to cover all loads to be monitored. A sampling frequency adapted to the microgrid operating conditions is set, typically 1-10Hz, to continuously collect real-time power data of each load. Time-domain analysis is performed on the collected real-time power data to calculate the power change rate. A power mutation judgment threshold is set according to the microgrid's rated power and load type. The mutation moments when the power change rate exceeds the judgment threshold are selected, and the stable power before and after the mutation is recorded. The absolute difference between the two is calculated to obtain the power mutation amplitude of each load.

[0028] Using the aforementioned sensor network, such as during critical periods of industrial production and core power supply periods for residential loads, power supply status data for each load can be collected. This data can be represented by binary identifiers: 1 for power supply and 0 for interruption. The power supply status data is parsed to capture the interruption start time when the power supply status changes from 1 to 0, and the recovery time when it changes from 0 to 1. The time difference between the interruption start time and the recovery time is calculated to obtain the interruption duration for each load. If there is no power supply interruption, the interruption duration is recorded as 0.

[0029] The priority dynamic decay rate is a coefficient between 0 and 1, used to characterize the degree of priority decay caused by power outages. The longer the interruption lasts, the smaller the decay rate, and the lower the final dynamic priority. It is calculated using a priority dynamic decay model, which is as follows: Where λ is the priority dynamic attenuation rate, k is the preset attenuation coefficient calculated based on historical microgrid operation data, and T is the duration of the intermediate section. To determine the maximum tolerable interruption duration, when T=0 (no interruption), λ=1, and the priority does not decrease; when T= When λ=e -k The priority is reduced to a preset ratio; when T→∞, λ→0, and the priority is completely reduced.

[0030] The obtained power fluctuation amplitude, interruption duration, and priority dynamic decay rate are summarized to form the original feature dataset. The original feature data is processed by a standardization algorithm to eliminate the dimensional differences between different features and make all feature values ​​uniformly mapped to the range of 0-1. The standardized feature data is organized to form the initial load feature set corresponding to each load in the microgrid.

[0031] S2. Based on the initial load feature set, fuzzy logic is used to calculate the degree of coupling between the power fluctuation amplitude and the interruption duration, and the time-period fluctuation index of each load is determined based on the degree of coupling.

[0032] In this embodiment, the power fluctuation amplitude and interruption duration of each load are extracted from the initial load feature set to construct a dual-feature basic data pair specific to each load, focusing on two types of core operating condition features for coupling analysis. Next, fuzzy domains are defined for each type of feature, fuzzy linguistic variables are defined, and suitable membership functions are selected to perform membership division on the dual-feature basic data pair, generating fuzzy sets corresponding to power fluctuation amplitude and interruption duration. Subsequently, a first fuzzy inference rule base containing full combination association rules of the two types of fuzzy linguistic variables is constructed. The two types of fuzzy sets of each load are matched with the rule base to calculate the coupling degree membership value between them. Then, the interaction influence level of each load is determined based on the coupling degree membership value, further refining the degree of difference in feature coupling. Finally, through weighted fusion, combining the coupling degree membership value and the interaction influence level, the time-period fluctuation index of each load is calculated, completing the comprehensive quantification of load fluctuation characteristics.

[0033] Specifically, based on the initial load feature set, for each load within the microgrid, the power surge amplitude and interruption duration are extracted as individual core feature parameters. The power surge amplitude is a quantified value used to characterize the intensity of power surges in the load, and the interruption duration is also a quantified value used to characterize the duration of power outages during sensitive power supply periods. These two parameters are bound to a unique load identifier, forming a unique dual-feature basic data pair for each load, collectively constituting a dual-feature dataset for the microgrid loads.

[0034] The basic definitions of fuzzy logic are completed for the two parameters respectively. The fuzzy universes are defined as follows: Based on the load power characteristics and power interruption patterns of the actual microgrid operation, continuous numerical universes are defined for the power surge amplitude and interruption duration, such as a power surge amplitude universe of 0-100kW and an interruption duration universe of 0-60min; simultaneously, an output universe is defined for the coupling degree membership value. The range of the universes can be dynamically adjusted according to the actual microgrid scenario; fuzzy linguistic variables are assigned to each universe to fuzzify the quantified values; the fuzzy linguistic variables for the power surge amplitude are set as small, medium, and large, and the fuzzy linguistic variables for the interruption duration are set as small, medium, and large. The linguistic variables are set to short, medium, and long, and the fuzzy linguistic variables for coupling degree membership values ​​are set to low, medium, and high. Membership degrees are divided, and commonly used membership functions in engineering are selected, such as triangular membership functions and trapezoidal membership functions. Triangular membership functions are preferred, taking into account both computational efficiency and accuracy. The fuzzy universe of discourse of each feature is mapped to the corresponding fuzzy linguistic variable to complete the construction of fuzzy sets for power mutation amplitude and interruption duration, realizing the membership degree division from quantized values ​​to fuzzy sets. That is, each feature quantization value can correspond to the membership degree value of each fuzzy linguistic variable.

[0035] Based on two input features of fuzzy linguistic variables, a fuzzy inference rule base with full combinatorial association is constructed, namely the first fuzzy inference rule base. This first fuzzy inference rule base covers all combinations of the two input feature fuzzy linguistic variables, with no rule omissions. The logical relationships of the rules are formulated based on the actual engineering laws of microgrid load power supply. The core principle is that the larger the power fluctuation amplitude and the longer the interruption duration, the higher the coupling degree between the two. For example, for nine full combinations of small, medium, and large power fluctuation amplitudes and short, medium, and long interruption durations, the corresponding coupling degree membership values ​​of the fuzzy linguistic variables are defined as low, medium, and high, respectively, forming nine basic fuzzy inference rules. For example, Rule 1: If the power fluctuation amplitude during peak electricity consumption is small and the interruption duration during the sensitive period is short, then the coupling degree is low; Rule 2: If the power fluctuation amplitude during peak electricity consumption is small and the interruption duration during the sensitive period is medium, then the coupling degree is low; Rule 3: If the power fluctuation amplitude during peak electricity consumption is small and the interruption duration during the sensitive period is long, then the coupling degree is medium; 4. If the power fluctuation during peak electricity consumption is moderate and the interruption duration during the sensitive period is short, the coupling degree is low. 5. If the power fluctuation during peak electricity consumption is moderate and the interruption duration during the sensitive period is moderate, the coupling degree is moderate. 6. If the power fluctuation during peak electricity consumption is moderate and the interruption duration during the sensitive period is long, the coupling degree is high. 7. If the power fluctuation during peak electricity consumption is large and the interruption duration during the sensitive period is short, the coupling degree is moderate. 8. If the power fluctuation during peak electricity consumption is large and the interruption duration during the sensitive period is moderate, the coupling degree is high.

[0036] Rule 9: If the power fluctuation during peak electricity consumption is large and the interruption duration during sensitive periods is long, then the coupling degree is high.

[0037] For each load in the microgrid, fuzzy inference and defuzzification operations are performed on the dual-feature basic data pair to obtain the specific coupling degree membership value. The quantified values ​​of the power surge amplitude and interruption duration of a single load are substituted into the corresponding fuzzy set to obtain its membership value under each fuzzy linguistic variable. The membership value is matched one by one with the first fuzzy inference rule base to filter out the inference rules that meet the conditions, determine the fuzzy linguistic variable of coupling degree corresponding to the load, and convert the fuzzy linguistic variable result of coupling degree into a continuous quantized value between 0 and 1. This value is the coupling degree membership value between the power surge amplitude and the interruption duration. The larger the value, the stronger the coupling relationship between the two features.

[0038] The interaction impact level between the power surge amplitude and the interruption duration of each load is determined based on the coupling degree membership value. This interaction impact level characterizes the degree of influence of the interaction between the two features on the power supply stability of the microgrid. Based on the output universe of discourse of the coupling degree membership value, it is divided into three continuous and non-overlapping numerical intervals, corresponding to low, medium, and high interaction impact levels, respectively. The division criteria must conform to the actual operation of the microgrid. The coupling degree membership value of each load is substituted into the above division criteria to match the corresponding interaction impact level. A unique and corresponding level weight is assigned to each interaction impact level, and the level weight is positively correlated with the interaction impact level. The coupling degree membership value of each load and the corresponding interaction impact level weight are weighted and fused to obtain the final time-period fluctuation index. This time-period fluctuation index quantifies the coupling effect of the load due to power surge and interruption duration in different power supply periods. The weighted fusion formula is determined as: Time-period fluctuation index = Coupling degree membership value × Corresponding interaction impact level weight. For example: Select a specific load within a microgrid: unique load identifier: L001, power fluctuation range: 40kW (universe of discourse: 0–100kW), interruption duration: 25min (universe of discourse: 0–60min); substitute the dual-feature quantization value of load L001 into a preset triangular membership function to calculate the membership degree of each fuzzy linguistic variable, where the preset triangular membership function is specifically as follows:

[0039]

[0040] Here, a, b, and c are three key parameters that define the shape and position of the triangle, corresponding to the left endpoint, vertex, and right endpoint of the triangle, respectively, while x is the input value.

[0041] The membership degrees for power surge amplitude (0-100kW) and interruption duration are calculated based on the preset triangular membership function: {small: 0, medium: 0.67, large: 0}. The membership degrees for interruption duration are also calculated: {short: 0, medium: 0.75, long: 0}. These are then matched with the first fuzzy inference rule base to determine the coupling degree as "medium". The centroid method is used to defuzzify the coupling degree "medium", normalizing it to a continuous value of 0.5 (0-1). Based on the coupling degree membership value and the interaction influence level table, the interaction influence level is determined to be medium, with a corresponding weight of 0.5. The weighted fusion formula is: Time Period Fluctuation Index = Coupling Degree Membership Value × Corresponding Interaction Influence Level Weight. The calculated Time Period Fluctuation Index is 0.5 × 0.5 = 0.25. The Time Period Fluctuation Index for load L001 is 0.25. This value quantifies the impact of the coupling effect between power surge and power interruption of load L001 on the stability of the microgrid power supply as moderate to low.

[0042] S3. When the load period fluctuation index exceeds the preset fluctuation threshold, extract the priority dynamic decay rate and interruption tolerance graded quantization value of the load, generate a fusion feature vector through neural network fusion, calculate the adjustment coefficient in combination with the period fluctuation index, and obtain the adjusted weight set based on the adjustment coefficient;

[0043] In this embodiment, the time-period fluctuation index of a single load is compared with a preset fluctuation threshold. If the time-period fluctuation index of the load is greater than or equal to the preset fluctuation threshold, it is determined to be a load to be regulated. For each selected load to be regulated, the priority dynamic decay rate and interruption tolerance graded quantization value are extracted, fused and normalized to obtain a fused feature base set. The fused feature base set is input into a preset neural network model to generate a high-dimensional fused feature vector. The fused feature comprehensive value is calculated based on the high-dimensional fused feature vector. Based on the fused feature comprehensive value and the time-period fluctuation index of the load, the final weight adjustment coefficient is obtained. Based on the initial weight set of the microgrid load, the initial weight of each load to be regulated is dynamically corrected through the weight adjustment coefficient to obtain the adjusted weight set.

[0044] The core logic involves threshold comparison, feature extraction, neural network fusion, calculation of adjustment coefficients, and generation of an adjusted weight set. Specifically, the time-period fluctuation index of a single load is compared with a preset fluctuation threshold. If the load's time-period fluctuation index is greater than or equal to the preset fluctuation threshold, it is identified as a load to be regulated, and its time-period fluctuation characteristics have a significant impact on the power balance of the microgrid, requiring dynamic weight adjustment. If it is less than the preset fluctuation threshold, it is identified as a regular load, maintains its initial weight, and does not participate in the weight adjustment in this step.

[0045] For each selected load to be regulated, two core quantitative parameters are precisely extracted from the initial load feature set and the microgrid load feature database: priority dynamic decay rate and interruption tolerance graded quantization value. The priority dynamic decay rate characterizes the degree to which the load's power supply priority changes over time or operating status; the interruption tolerance graded quantization value characterizes the load's tolerance to power outages. These two core quantitative parameters are bound to the unique identifier of the load to be regulated, forming a fusion feature base set. The two core quantitative parameters in the fusion feature base set are then normalized. The normalized fusion feature base set is input into a pre-defined neural network model. Through the nonlinear mapping capability of the neural network, deep fusion of the two basic features is achieved, generating a high-dimensional fusion feature vector, uncovering hidden correlations between features, and improving the accuracy of weight adjustment. The generation of the interruption tolerance graded quantization value includes: obtaining the interruption sensitivity characteristics of each load based on its core attributes within the microgrid, where the core attributes include load purpose, power level, degree of loss after interruption, recovery difficulty, and maximum allowable interruption duration; classifying interruption tolerance levels based on the load's interruption sensitivity characteristics, and normalizing and quantizing these interruption tolerance levels to obtain the interruption tolerance graded quantization value.

[0046] The neural network model uses a fully connected neural network (FCNN), with an input layer, hidden layer, and output layer structure. Specifically, the input layer has a dimension of 2, corresponding to the normalized priority dynamic decay rate and interruption tolerance graded quantization values; the hidden layer has 2 layers, with the number of neurons determined according to the scale of the load to be regulated, such as 16 / 32, and the activation function is ReLU; the output layer has a preset dimension, such as 8 / 16, which can be adjusted according to the regulation precision, and the activation function is Sigmoid. The neural network model is a pre-trained mature model, and the training set is a historical operational measurement dataset of the microgrid, covering all operating conditions such as different seasons, peak / off-peak electricity consumption, and ample / short power supply. The data is taken from the microgrid historical dispatch database and consists of standardized data of the load to be regulated in each historical period. The system includes: ① Priority dynamic attenuation rate: real-time attenuation coefficient of each load under power outage and power fluctuation scenarios, with a value of 0-1; ② Interruption tolerance graded quantification value: quantification value of each load based on its purpose, interruption loss, and recovery difficulty, with a value of 0-1; The sample size is determined according to the load scale, with no less than 5000 samples for a single microgrid scenario, divided into training set, validation set, and test set, accounting for 75%, 15%, and 10% respectively; Preprocessing adopts the min-max normalization method to map all input feature data to the 0-1 interval, and removes abnormal samples with data deviations exceeding ±30%, ensuring that the accuracy of the neural network model meets the engineering control requirements. The two-dimensional feature vectors of the loads in the fusion feature base set are successively substituted into the neural network model to perform forward inference calculation; the high-dimensional fusion feature vector corresponding to each load to be controlled is output, realizing deep fusion of basic features. The vector represents the load after deep fusion of multiple features. The high-dimensional fusion feature vector is a high-dimensional refined feature generated by mining potential correlations between two basic features: priority dynamic decay rate and interruption tolerance grading quantization values. Each dimension corresponds to a specific control characteristic of the load, including but not limited to: time-period fluctuation adaptability, interruption priority weight, power fluctuation sensitivity, control response adaptability, and power restoration priority. This comprehensively represents the load's overall control attributes and compensates for the lack of low-dimensional feature information. The vector dimension can be 8 or 16, which can be flexibly adjusted according to the control precision. Each dimension's value is activated by a Sigmoid function and takes a value of 0-1. The larger the value, the more significant the load's control characteristic in that dimension. For example, in a 16-dimensional vector, the first dimension represents the load importance, the second dimension represents the interruption tolerance sensitivity, the third dimension represents the power fluctuation impact, and so on. The values ​​of each dimension are not coupled or interfered with, and independently represent the specific characteristics.

[0047] Specifically, the fully connected neural network (FCNN) is a pre-trained, mature engineering model adapted to the load feature fusion and weight adjustment requirements of microgrids. Its core training parameters and configuration are as follows:

[0048] (1) Construction of training dataset

[0049] The training set uses historical operational data of microgrids. The sample size is determined according to the model fitting accuracy requirements. The sample data is standardized labeled data, and each sample contains input features and output labels: the input features are two-dimensional vectors, corresponding to the normalized priority dynamic decay rate and interruption tolerance graded quantization values, which match the dimension of the model input layer; the output labels are high-dimensional fused feature vectors, with dimensions consistent with the preset dimensions (8 / 16 dimensions) of the model output layer. The label values ​​are determined by back-calculation using historical load regulation data of microgrids and the optimal solution for power balance, covering load feature data under different load scales, different electricity consumption periods, and different power fluctuation scenarios. At the same time, a validation set and a test set are divided, with the validation set and the test set accounting for 15% and 10% respectively, to ensure the model's generalization ability. Outliers and missing values ​​in the samples are removed, and the samples are put into model training after 0-1 normalization preprocessing.

[0050] (2) Selection of loss function

[0051] For the core tasks of high-dimensional feature fusion and regression fitting in the model, the mean squared error (MSE) loss function is selected as the model training loss function, which balances regression accuracy and training stability. The loss function formula is as follows: ,in Here, n represents the loss value, and n is the number of samples in a single training batch. The predicted high-dimensional fusion feature vector output by the model. The high-dimensional fusion feature vector is the real label corresponding to the sample. By minimizing the loss value, the weights and biases of the model neurons are optimized to improve the feature fusion accuracy until the loss value converges to the preset threshold that meets the engineering control accuracy requirements.

[0052] (3) Training algorithm and optimizer configuration

[0053] The Adam adaptive moment estimation algorithm was selected as the model training optimization algorithm. This algorithm has fast convergence speed, strong stability, and is suitable for high-dimensional feature fitting scenarios. It does not require manual adjustment of the learning rate. The core hyperparameters are set as follows: learning rate is set to 0.001, batch size is set to 16 / 32 according to the load scale and computing power configuration, and the number of iterations is set to 100-200. At the same time, an early stopping mechanism is set. When the validation set loss value does not decrease for 10 consecutive iterations, the training is terminated in advance to prevent the model from overfitting. The ReLU function is used as the activation function of the hidden layer to alleviate the gradient vanishing problem, and the Sigmoid function is used as the activation function of the output layer to map the output value to the 0-1 interval, which meets the numerical range requirements of the high-dimensional fused feature vector.

[0054] (4) Model training and inference

[0055] The preprocessed training and validation sets are input into the constructed FCNN model. Iterative training is performed based on the aforementioned loss function and training algorithm to optimize and tune the model weights and bias parameters. After training, the model accuracy is verified using a test set; a feature fusion error ≤2% meets the engineering control requirements. The two-dimensional load feature vectors from the fusion feature base set are then substituted one by one into the trained FCNN model, and forward inference calculations are performed to output a high-dimensional fusion feature vector corresponding to each load to be controlled. This achieves deep nonlinear fusion of basic features, uncovers potential correlations between features, and improves the accuracy of subsequent weight adjustments. The fusion feature comprehensive value is calculated by summing the generated high-dimensional fusion feature vectors. This, combined with the load's time-period fluctuation index, yields the final weight adjustment coefficient. Based on the initial weight set of the microgrid load, the initial weights of each load to be controlled are dynamically corrected using the weight adjustment coefficient, while maintaining the weights of regular loads unchanged, ultimately generating the adjusted weight set. For example, the fusion feature comprehensive value is obtained by summing the high-dimensional fusion feature vectors of each load to be controlled element-wise, using the following formula: Where n is the dimension of the high-dimensional vector. Let F be the i-th high-dimensional fusion feature vector, and F be the fusion feature comprehensive value. The fusion feature comprehensive value is obtained by normalization. The weight adjustment coefficient is calculated using a weighted fusion method, and the formula is as follows: Where K is the weight adjustment coefficient. This is the normalized fusion feature composite value. The normalized time-period fluctuation index is represented by α and β, respectively. α and β are the weighting coefficients of the normalized fusion feature composite value and the normalized time-period fluctuation index, which can be adjusted according to engineering requirements. For each load to be controlled, the initial weights are dynamically adjusted according to the formula: ,in, For the adjusted weights, These are the initial weights.

[0056] S4. Extract loads whose weights exceed a preset weight threshold from the adjusted weight set, and use a genetic algorithm to search for the power supply sequence based on the power fluctuation amplitude of the load and the duration of the interruption, to obtain the target power supply sequence;

[0057] In this embodiment, loads with weights exceeding a preset weight are obtained from the adjusted weight set to form a high-load set. The interruption duration of the loads in the high-load set is determined, and a genetic algorithm is used to search for a power supply sequence based on the interruption duration to obtain a preliminary power supply sequence. The preset weight is determined by statistically analyzing historical data of the adjusted weights of each load in the microgrid, taking the 85%-95th percentile or average value and the standard deviation as a weight.

[0058] Determine whether the preliminary power supply sequence meets the current available power constraint. If not, remove loads from the preliminary power supply sequence whose correlation threshold between power fluctuation amplitude and interruption duration exceeds a preset correlation threshold to obtain the target power supply sequence.

[0059] Specifically, the adjusted weight value of a single load is compared with the preset weight threshold. If the adjusted weight value of the load is greater than or equal to the preset weight threshold, it is determined to be a high-weight load and needs to be included in the power supply sequence optimization range. If it is less than the preset weight threshold, it is determined to be a low-weight load, and power is allocated according to the default power supply rules. It does not participate in the power supply sequence search in this step. The 85%-95% quantile of the historical adjusted weight data is taken as the preset weight threshold, with the 90% quantile preferred to balance the screening accuracy and universality.

[0060] For each load in the selected high-load set, the interruption duration required for power supply sequence optimization is extracted from the initial load feature set, providing complete data support for the genetic algorithm search and subsequent constraint verification. Using the interruption duration optimization objective of the high-load set, the core parameters of the genetic algorithm are configured, and the fitness function is defined to complete the pre-search environment initialization, ensuring the efficiency and optimization effect of the algorithm search. A classic genetic algorithm is selected, configured with engineering parameters adapted to microgrid load regulation. Core parameters include: population size, determined by the number of loads in the high-load set, typically 2-5 times the number of loads; number of iterations, set according to optimization accuracy requirements, typically 50-200 generations; crossover probability set to 0.7-0.9 to better ensure population diversity; mutation probability set to 0.01-0.1 to better avoid local bias in the algorithm; and integer encoding, where each gene locus corresponds to a load ID within the high-load set, chromosome length equal to the number of loads in the high-load set, and the chromosome sequence representing the power supply order. The core optimization objective is to minimize the total interruption duration of all loads in the high-load set, while also considering load weight priority. A fitness function is constructed by combining the optimization objective and load weights, with the optimization objective being: Where n is the total number of loads in the high-weight load set, and i is the load number. Let i be the weight value of the i-th load. The interrupt duration of the i-th load, weighted The higher the load, the longer the interruption time. The larger the proportion of a high-weight load in the total, the more the algorithm will prioritize minimizing the interruption duration of high-weight loads (i.e., prioritizing power supply) when minimizing F(S). This achieves both "minimizing the total interruption duration" and "maximizing the priority of high-weight loads." The genetic algorithm needs to transform the "minimization objective" into "maximizing fitness," therefore the fitness function is defined as: ,in For minimal offset, a higher fitness(S) value indicates a better power supply sequence.

[0061] The quality of chromosomes (power supply sequences) is quantified, and the basic constraints of the genetic algorithm are set as follows: chromosomes have no duplicate gene loci and the power supply order of the load is unique. The algorithm search environment is then set up. The high-load optimization feature set is input into the initialized genetic algorithm environment, and the interruption duration is used as the core optimization indicator. The algorithm performs selection, crossover, and mutation iteration operations to search for a preliminary power supply sequence that meets the optimization objective.

[0062] The current available power value is obtained from the real-time monitoring system of the microgrid power supply side. Based on the power consumption characteristic parameters of each load in the preliminary power supply sequence, the total power demand value corresponding to the sequence is calculated. The calculation logic is as follows: the rated operating power of each load is accumulated according to the power supply sequence, and the total power demand value is compared with the current available power value to determine the constraint satisfaction: if the total power demand value is less than or equal to the current available power value, the preliminary power supply sequence satisfies the power constraint and is directly designated as the target power supply sequence; if the total power demand value is greater than the current available power value, the preliminary power supply sequence does not satisfy the power constraint. In the case where the preliminary power supply sequence does not satisfy the current available power constraint, the abnormal loads in the sequence whose characteristic correlation exceeds the threshold are removed based on the preset correlation threshold, the total power demand is reduced, so that the corrected power supply sequence satisfies the power constraint, and finally the target power supply sequence is generated.

[0063] The preset correlation threshold is obtained by calibration using a correlation coefficient based on the duration of load interruption. This correlation coefficient is an empirical calibration coefficient, adjusted according to the actual microgrid scenario. The specific calculation formula for the preset correlation threshold is as follows:

[0064]

[0065] in, The preset correlation threshold is α, where α is the correlation coefficient. This is a core characteristic value of the duration of outages based on historical data statistics of microgrids. This represents the maximum tolerable interruption duration.

[0066] S5. Obtain real-time power fluctuation data of the current power supply side from the target power supply sequence, and re-integrate the power mutation amplitude and the priority dynamic decay rate through the fuzzy logic to calculate the updated priority value of each load in the target power supply sequence;

[0067] In this embodiment, real-time power fluctuation data of the power supply side corresponding to the target power supply sequence is collected. The real-time power fluctuation data includes at least the real-time power monitoring value, power change rate, power fluctuation amplitude, power fluctuation period, and the deviation between the steady-state power value and the actual value during the time period. The power mutation amplitude and priority dynamic decay rate of each load in the target power supply sequence are extracted from the initial load feature set. The fuzzy universes of the power mutation amplitude and priority dynamic decay rate are defined respectively, and the fuzzy subsets and membership functions corresponding to each fuzzy universe are determined. A preset second fuzzy inference rule library is called to perform fuzzy fusion inference calculation on the real-time power fluctuation data and the power mutation amplitude and priority dynamic decay rate to obtain the priority adjustment parameters of each load. Based on the priority adjustment parameters of each load and its initial weight in the target power supply sequence, the updated priority value of each load in the target power supply sequence is obtained through weighted fusion calculation.

[0068] Specifically, real-time power fluctuation data matching the target power supply sequence is collected through real-time monitoring sensors or data acquisition terminals on the power supply side of the microgrid. The real-time power fluctuation data includes at least the real-time power monitoring value, power change rate, power fluctuation amplitude, power fluctuation period, and the deviation between the steady-state power value and the actual value for a given period. The collected real-time power fluctuation data is cleaned to remove missing values ​​and outliers, and intermittent data is completed by linear interpolation. Then, all the completed real-time power fluctuation data is normalized to 0-1 to eliminate differences in dimensions and numerical ranges, forming a standardized real-time power fluctuation dataset.

[0069] From the initial load feature set, the power mutation amplitude and priority dynamic decay rate of each load in the target power supply sequence are extracted in a targeted manner. The extracted two types of core feature parameters, namely the power mutation amplitude and priority dynamic decay rate, are normalized respectively. The two types of normalized feature parameters are bound with the unique identifier of the corresponding load and the initial weight of the load in the target power supply sequence to construct a load feature inference set.

[0070] The fuzzy universes of discourse for power surge amplitude and priority dynamic decay rate are both defined as [0,1]. Three-level fuzzy subsets ({small, medium, large}) are defined for power surge amplitude, {slow, medium, fast} for priority dynamic decay rate, and {weak, medium, strong} for real-time power fluctuation data, representing different quantification levels of the corresponding features. A triangular membership function is used to construct a corresponding membership function for each fuzzy subset, mapping the precise values ​​within the universe of discourse to membership values ​​in the [0,1] interval. A preset second fuzzy inference rule library is retrieved from the microgrid fuzzy control rule database. This second fuzzy inference rule library takes standardized real-time power fluctuation data, power surge amplitude, and priority dynamic decay rate as input, and load priority adjustment parameters as output. The rules adopt a multi-condition single-output inference form of "if A and B and C, then D", covering the full combination association of power surge amplitude, priority dynamic decay rate, and core power fluctuation features. Specific rules are shown in Table 1 below.

[0071] Rule Number Real-time power fluctuation data (A) Power fluctuation amplitude (B) Priority dynamic decay rate (C) Load priority adjustment parameter (D) 1 weak Small slow Small 2 weak Small middle Small 3 weak Small quick very small 4 weak middle slow middle 5 weak middle middle Small 6 weak middle quick Small 7 weak big slow middle 8 weak big middle middle 9 weak big quick Small 10 middle Small slow middle 11 middle Small middle middle 12 middle Small quick Small 13 middle middle slow big 14 middle middle middle middle 15 middle middle quick middle 16 middle big slow big 17 middle big middle big 18 middle big quick middle 19 powerful Small slow big 20 powerful Small middle middle 21 powerful Small quick middle 22 powerful middle slow Very large 23 powerful middle middle big 24 powerful middle quick big 25 powerful big slow Very large 26 powerful big middle Very large 27 powerful big quick big

[0072] Table 1. Second Fuzzy Inference Rule Base

[0073] Substituting the normalized power fluctuation amplitude and normalized priority dynamic attenuation rate of each load within the target power supply sequence into the corresponding membership function yields the membership value of each load under the corresponding fuzzy subset. Simultaneously, standardized real-time power fluctuation data is mapped to the corresponding fuzzy subset membership value, completing the fuzzification of all input features. The membership values ​​of all input features are matched against the second fuzzy inference rule base one by one, and inference rules that meet the conditions are selected. The results of the matched rules are synthesized using the larger-smaller method to obtain the comprehensive membership distribution of the fuzzy subset of priority adjustment parameters. The centroid method is used to defuzzify the comprehensive membership distribution to obtain the priority adjustment parameters for each load. The priority adjustment parameters are precise quantized values ​​in the 0-1 range, and are calculated for each load to form a load priority adjustment parameter set.

[0074] A weighted fusion rule is established for priority adjustment parameters and the initial weight of the target power supply sequence. This fusion rule is based on the initial weight and dynamically adjusted using the priority adjustment parameters. Two fusion weight coefficients are calibrated: the initial weight coefficient and the priority adjustment parameter weight coefficient. Their specific values ​​are determined based on the weight bias requirements of microgrid power balance control. Simultaneously, the weighted fusion formula for updating the priority value is determined as follows: ,in, The initial weight of the load in the target power supply sequence, Δw is the priority adjustment parameter of the load, α is the initial weight coefficient, and β is the priority adjustment parameter weight coefficient.

[0075] S6. If the update priority value deviates from the interruption tolerance grading quantization value, the neural network is used to calculate the difference between the actual response delay measurement value of the load with the deviation and the expected interruption tolerance time limit to obtain the response delay correction factor.

[0076] In this embodiment, the update priority value of the load is compared with the interruption tolerance classification quantization value. If there is a deviation between the two values, the current load is marked as a deviation load. The difference between the actual response delay value of the deviation load and the preset expected interruption tolerance time limit is obtained. The response delay correction factor is calculated based on the difference. The preset expected interruption tolerance time limit is determined by statistical analysis of the historical interruption tolerance data of each load in the microgrid, and the maximum interruption duration allowed for stable operation of the load is taken.

[0077] Specifically, the update priority value is compared with the interruption tolerance level quantization value to mark the deviating load. The update priority value set of the target power supply sequence load is retrieved, and the update priority value of each load is extracted. At the same time, the interruption tolerance level quantization value corresponding to each load in the target power supply sequence is extracted from the initial load feature set. The update priority value and interruption tolerance level quantization value of the same load are compared one by one. A preset deviation judgment threshold is set, and the absolute deviation value between the two is calculated. If the absolute deviation value is found, it is determined that there is a deviation between the update priority value and the interruption tolerance level quantization value of the load, and the load is marked as a deviating load. The preset deviation judgment threshold is determined by statistically analyzing the difference between the update priority value and the interruption tolerance level quantization value in historical operating data, and taking the sum of the average value and 0.5-1.5 times the standard deviation.

[0078] For the marked deviation loads, two core data types are collected: actual response delay measurement and expected interruption tolerance time limit. The actual response delay measurement value Dact is the average of the actual response delay of the load during the most recent 1-3 power outages extracted from the microgrid real-time monitoring sensors / data acquisition terminals. If there is no historical interruption data, the average response delay of the same type of load is used as a substitute. The expected interruption tolerance time limit Dexp is the difference between the actual response delay and the expected interruption tolerance time limit ΔD, calculated from the expected interruption tolerance time limit preset for the deviation load extracted from the microgrid load interruption tolerance parameter library. ΔD = Dact − Dexp. If ΔD is greater than 0, it indicates that the actual delay exceeds the expectation and requires strong correction. If ΔD ≤ 0, it indicates that the actual delay meets the standard and requires weak correction or no correction. The 0-1 normalization preprocessing of ΔD yields the preprocessed ΔDnorm. A pre-trained BP neural network (backpropagation neural network) is then called, with the preprocessed ΔDnorm as input and the response delay correction factor γ as output. The BP neural network structure is as follows: the input layer contains 1 neuron with ΔDnorm as input, the hidden layer contains 3-5 neurons with the sigmoid activation function, and the output layer contains 1 neuron with the response delay correction factor γ as output.

[0079] S7. Adjust the power supply timing of the corresponding load in the target power supply sequence according to the response delay correction factor to obtain the final power allocation scheme.

[0080] In this embodiment, for each deviation load, the adjusted power supply time is calculated using a response delay correction factor. The adjusted power supply times of all deviation loads are integrated with the initial power supply times of non-deviation loads to form a preliminary adjusted power supply sequence. This preliminary adjusted power supply sequence is subject to dual verification: first, a rationality verification, checking whether there is overlap in the power supply times of each load and whether the power supply order matches the priority of the target power supply sequence. If there is overlap in the power supply times, the power supply times are finely adjusted sequentially from high to low according to the load update priority value to ensure that there is no conflict in the power supply period; second, a power constraint adaptability verification, retrieving the current available power value, calculating the total power demand value corresponding to the preliminary adjusted power supply sequence, and verifying whether the total power demand value is less than or equal to the current available power value to ensure that the adjusted sequence adapts to the power supply capability of the power supply side.

[0081] If the initial adjusted power supply sequence fails the rationality check or power constraint adaptability check, conflict optimization adjustment is performed: For the problem of overlapping power supply times, loads are sorted according to their update priority values. High-priority loads retain the adjusted power supply times first, while low-priority loads have their power supply times fine-tuned. The fine-tuning range does not exceed the preset fine-tuning threshold and does not change the adjustment direction. The preset fine-tuning threshold is determined based on the historical dispatch data of the microgrid and is the maximum time step allowed for the offset of the load power supply time, with a value of 1-3 dispatch cycles. For the problem of excessive power demand, the power supply times of the deviating loads are further optimized, and the delay range of low-priority deviating loads is increased until the total power demand value meets the current available power constraint. After the verification is passed, the final power supply sequence is determined, and the final power supply time, power supply order, and power supply duration of each load are specified.

[0082] Based on the final power supply sequence, the core parameters of each load are integrated, including the unique identifier of the load, the final power supply time, the power supply duration, the updated priority value, the adjusted weight, the rated power supply power, and the tiered quantitative value of the interruption tolerance. At the same time, the current available power value, total power demand value, deviation load adjustment details and verification report of the power supply side are supplemented to form a complete final power allocation scheme. The final power allocation scheme needs to clarify the execution standard of power allocation, the priority ranking of load power supply, the scheduling rules of power supply time and the triggering conditions for subsequent dynamic adjustment, and output them synchronously to the microgrid power allocation control system to guide the implementation of dynamic load power balance control of the microgrid.

[0083] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this application will not describe the various possible combinations separately. Furthermore, various different embodiments of this application can also be arbitrarily combined, as long as they do not violate the spirit of this application, and they should also be regarded as the content disclosed in this application.

Claims

1. A method for dynamic load power balancing control of a microgrid, characterized in that, The method includes the following steps: S1. Obtain the power surge amplitude, interruption duration, and priority dynamic decay rate of each load in the microgrid to obtain the initial load feature set; S2. Based on the initial load feature set, fuzzy logic is used to calculate the degree of coupling between the power mutation amplitude and the interruption duration, and the time-period fluctuation index of each load is determined based on the degree of coupling. S3. When the load’s time-period fluctuation index exceeds the preset fluctuation threshold, extract the priority dynamic decay rate and interruption tolerance graded quantization value of the load, generate a fusion feature vector through neural network fusion, calculate the weight adjustment coefficient in combination with the time-period fluctuation index, and obtain the adjusted weight set based on the weight adjustment coefficient. S4. Extract the loads whose weights exceed a preset weight threshold from the adjusted weight set, and use a genetic algorithm to search for the power supply sequence based on the interruption duration of the loads exceeding the preset weight threshold to obtain the target power supply sequence; S5. Obtain real-time power fluctuation data from the target power supply sequence, call the preset second fuzzy inference rule library to perform fuzzy logic calculation on the real-time power fluctuation data, interruption duration and priority dynamic decay rate to obtain the priority adjustment parameters of each load, and then combine the initial weight of each load to obtain the updated priority value. S6. If there is a deviation between the updated priority value and the interruption tolerance level quantization value, the actual response delay measurement value of the load with the deviation is calculated by the preset neural network and the difference between the expected interruption tolerance time limit to obtain the response delay correction factor. S7. Adjust the power supply timing of the loads in the target power supply sequence according to the response delay correction factor to obtain the final power allocation scheme.

2. The method as described in claim 1, characterized in that, Step S1 includes: Power data of each load is collected by deploying a microgrid sensor network, and the power fluctuation amplitude of each load is calculated by the time-domain change of the power data. By using a sensor network to collect power supply status data of each load, the start and recovery times of power supply interruption for each load are statistically analyzed, and the time difference is calculated to obtain the interruption duration of each load. Based on the basic power supply priority of the load, the priority dynamic decay rate of each load is calculated through a preset priority dynamic decay model; The three feature datasets of power fluctuation amplitude, interruption duration, and priority dynamic decay rate are standardized to obtain the initial load feature set.

3. The method as described in claim 1, characterized in that, Step S2 includes: From the initial load feature set, the power fluctuation amplitude and interruption duration of each load are extracted to form a dual-feature basic data pair for each load; The fuzzy domains are defined for the power mutation amplitude and the interruption duration, respectively. The corresponding fuzzy linguistic variables are defined, and the membership function is selected to perform membership division on the dual-feature basic data pair to obtain the fuzzy set of power mutation amplitude and the fuzzy set of interruption duration. Construct a first fuzzy inference rule base, which contains full combination association rules of the fuzzy linguistic variable of power mutation magnitude and the fuzzy linguistic variable of interruption duration; The fuzzy sets of power fluctuation amplitude and interruption duration of each load are matched with the first fuzzy inference rule base to obtain the degree of coupling between the power fluctuation amplitude and interruption duration of each load. The interaction impact level of each load is determined based on the coupling degree membership value, and the interaction impact level between the power change amplitude and the interruption duration of each load is obtained. The time-period fluctuation index of each load is obtained by combining the coupling degree membership value with the interaction influence level.

4. The method as described in claim 1, characterized in that, Step S3 includes: The time-period fluctuation index of a single load is compared with the preset fluctuation threshold. If the time-period fluctuation index of a load is greater than or equal to the preset fluctuation threshold, it is determined to be a load to be regulated. For each selected load to be regulated, the priority dynamic decay rate and interruption tolerance graded quantization values ​​are extracted and normalized to obtain the fusion feature base set; The normalized fusion feature base set is input into a preset neural network model to generate a high-dimensional fusion feature vector, and the fusion feature comprehensive value is calculated based on the high-dimensional fusion feature vector. Based on the integrated characteristic value and the time-period fluctuation index of the load, the final weight adjustment coefficient is obtained. Based on the initial weight set of the microgrid load, the initial weight of each load to be regulated is dynamically corrected through the weight adjustment coefficient to obtain the adjusted weight set.

5. The method as described in claim 4, characterized in that, The generation of the interruption tolerance grading quantization value includes: Based on the core attributes of each load in the microgrid, the interruption sensitivity characteristics of each load are obtained. The core attributes include load purpose, power level, degree of loss after interruption, recovery difficulty, and maximum allowable interruption duration. Based on the load interruption sensitivity characteristics, interruption tolerance levels are divided, and the interruption tolerance levels are normalized and quantized to obtain interruption tolerance grade quantization values.

6. The method as described in claim 1, characterized in that, Step S4 includes: The loads whose weights exceed the preset weights are obtained from the adjusted weight set to form a high load set. The interruption duration of the loads in the high load set is determined, and a genetic algorithm is used to search for the power supply order based on the interruption duration to obtain a preliminary power supply sequence. Determine whether the preliminary power supply sequence meets the current available power constraint. If not, remove loads from the preliminary power supply sequence whose correlation threshold between power fluctuation amplitude and interruption duration exceeds a preset correlation threshold to obtain the target power supply sequence.

7. The method as described in claim 6, characterized in that, The preset correlation threshold is based on the duration of load interruption and is obtained through correlation coefficient calibration. The correlation coefficient is an empirical calibration coefficient that is adjusted according to the actual scenario of the microgrid.

8. The method as described in claim 1, characterized in that, Step S5 includes: Collect real-time power fluctuation data on the power supply side corresponding to the target power supply sequence. The real-time power fluctuation data includes real-time power monitoring value, power change rate, power fluctuation amplitude, power fluctuation period, and deviation between the steady-state power value and the actual value for the time period. Extract the power fluctuation amplitude and priority dynamic attenuation rate of each load from the initial load feature set; The fuzzy universes of discourse for the power mutation amplitude, the priority dynamic decay rate and the real-time power fluctuation data are defined respectively, and the fuzzy subsets and membership functions corresponding to each fuzzy universe are determined. The preset second fuzzy inference rule library is invoked to perform fuzzy logic calculations on the real-time power fluctuation data, the power mutation amplitude, and the priority dynamic attenuation rate to obtain the priority adjustment parameters for each load. Based on the priority adjustment parameters and the initial weights of each load in the target power supply sequence, the updated priority values ​​of each load in the target power supply sequence are calculated.

9. The method as described in claim 1, characterized in that, Step S6 includes: The update priority value of the load is compared with the interruption tolerance level quantization value. If there is a deviation between the two, the current load is marked as a deviation load. Obtain the difference between the actual response delay value of the deviation load and the preset expected interruption tolerance time limit, and calculate the response delay correction factor based on the difference.

Images