Improved Adaboost-based single-three-phase series-parallel micro-grid group island detection method

Abstract

The invention discloses a single three-phase series-parallel micro-grid group island detection method based on improved Adaboost, and the method comprises the following steps: collecting sample data of different electrical characteristic quantities in a single three-phase series-parallel micro-grid group, which are measured at a point of common coupling (PCC); based on partial least squares (PLS), correlation analysis is carried out on various electrical characteristic quantities measured at a common coupling point, strong correlation electrical characteristic quantities are screened out, and a training set and a test set are formed; according to the method, an Adaboost learning device is improved on the basis of a mucus optimization algorithm (SMA) and a GSN weight learning mode, so that the classification capability is improved, the disturbance influence is weakened, and the detection time consumption is shortened. And constructing an island detection model of the microgrid group based on the improved Adaboost learner, and outputting an island detection result according to the real-time test set data at the PCC. According to the method, accurate island detection can be carried out without being affected by disturbance signals and the three-phase unbalance degree of the system, and the method has high anti-interference performance, accuracy and generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of microgrid islanding detection, and in particular to a single-phase and three-phase hybrid microgrid group islanding detection method based on improved Adaboost. Background Art

[0002] Microgrids can effectively coordinate and control distributed power generation devices, and have become one of the main forms of efficient use of distributed power sources. Microgrids can operate in two modes: grid-connected and islanded. The islanded operation mode is divided into planned islanding and unplanned islanding. When unplanned islanding occurs, the voltage and frequency of the microgrid will be out of the control of the main grid and lose stability, which may damage the performance of power equipment and affect stable operation. Therefore, it is crucial to study the islanding detection technology of microgrids, quickly identify the unplanned islanding state, and take measures to change the control mode of the inverter to ensure the reliability and stability of the operation of microgrids and the main grid.

[0003] In the existing technical literature: Literature [1]: "Hybrid island detection method based on genetic algorithm and BP neural network" (Yu Yunjun, Zhong Guoying, Wan Xiaofeng, et al. Hybrid island detection method based on genetic algorithm and BP neural network [J]. Renewable Energy, 2018, 36(5): 701-706.) uses genetic algorithm to improve the traditional BP neural network, and uses the output of the improved BP neural network as the judgment result of the system state. This method can maintain a faster detection speed and a lower false detection rate under different load characteristic scenarios.

[0004] Reference [2]: "Research on the interpretability of island detection and integration of photovoltaic grid-connected systems based on LightGBM algorithm" (Zhu Chunlin, Yu Chengbo. Research on the interpretability of island detection and integration of photovoltaic grid-connected systems based on LightGBM algorithm [J]. Electric Power Automation Equipment, 2023, 43(7): 80-86+150.) proposes an island detection method based on the light gradient boosting machine algorithm. This method performs island detection and analysis through K-fold cross validation and Shapley value additive interpretation method based on decision tree, which improves the credibility of the model.

[0005] However, the above studies did not consider the screening and extraction of multiple electrical feature quantities at the PCC related to the islanding state. Reference [3]: "Islanding Detection Method for Grid-connected Photovoltaic Power Generation System Based on Adaboost Algorithm" (Jia Ke, Xuan Zhenwen, Lin Yaoqi, et al. Islanding Detection Method for Grid-connected Photovoltaic Power Generation System Based on Adaboost Algorithm [J]. Transactions of China Electrotechnical Society, 2018, 33(5): 1106-1113.) proposed an intelligent passive islanding detection method based on Adaboost learner. This method can avoid the problem that traditional detection methods cannot determine the threshold of electrical quantities without introducing interference signals. However, this method still has problems such as being affected by disturbance factors and taking too long to train when there are many sample features in the process of applying Adaboost learner. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides a single-phase three-phase hybrid microgrid group island detection method based on improved Adaboost. The method can perform accurate island detection without being affected by disturbance signals and system three-phase imbalance, and has high anti-interference, accuracy and generalization ability.

[0007] The technical solution adopted by the present invention is:

[0008] The island detection method of single-phase and three-phase hybrid microgrid based on improved Adaboost includes the following steps:

[0009] Step 1: Collect sample data of different electrical characteristic quantities in the single-phase and three-phase hybrid microgrid group measured at the point of common coupling (PCC);

[0010] Step 2: Based on the partial least squares (PLS) method, a correlation analysis is performed on various electrical characteristic quantities measured at the common coupling point, and strongly correlated electrical characteristic quantities are screened out to form a training set and a test set;

[0011] Step 3: Improve the Adaboost learner based on the Slime Mould Algorithm (SMA) and GSN weight learning method to improve classification ability, reduce the impact of disturbances and shorten detection time.

[0012] Step 4: Build an island detection model for the microgrid based on the improved Adaboost learner, and output the island detection results according to the real-time test set data at the PCC.

[0013] In step 1, the island electrical characteristic quantity is caused by the occurrence of the island effect and is affected by the system and its operating state. The island electrical characteristic quantities considered in the present invention include voltage, voltage change rate over time, voltage total harmonic distortion rate, current, current change rate over time, current total harmonic distortion rate, frequency, frequency change rate over time, impedance, impedance change rate over time, active power, active power change rate over time, reactive power, reactive power change rate over time and power factor. Therefore, it is necessary to collect sample data of the above electrical characteristic quantities in step 1. Sample data of different electrical characteristic quantities are shown in Table 1.

[0014] Table 1 Basic island electrical characteristics of single-phase and three-phase hybrid microgrids

[0015]

[0016]

[0017] In step 2, in the intelligent island detection method, the selection of island electrical characteristic quantities will directly affect the accuracy and speed of island detection. In machine learning, the more characteristic quantities the training model contains, the longer the training time is and the higher the calculation requirements are. Therefore, the present invention sorts the importance of the basic island electrical characteristic quantities of the microgrid group based on the PLS method, and selects the electrical characteristic quantities with the variable projection importance (VIP) index exceeding 1 as the strongly correlated electrical characteristic quantities for the microgrid group island detection;

[0018] The PLS method is used to process the collected basic electrical characteristic data and establish a multivariate linear regression empirical expression:

[0019]

[0020] Where: X is the data set consisting of basic island electrical characteristic quantities; is the kth electrical characteristic corresponding to the PCC of the first microgrid; is the kth electrical characteristic corresponding to the PCC of the second microgrid; The i-th microgrid PCC corresponds to the k-th electrical characteristic quantity; The nth microgrid PCC corresponds to the kth electrical characteristic quantity.

[0021] Y is the data set of island status; y1 is the island status of the first microgrid, y1=0 is non-island, y1=1 is island; y2 is the island status of the second microgrid, y2=0 is non-island, y2=1 is island; j is the island state of the jth microgrid, y j = 0 means non-island, y j=1 for an isolated island; y n is the island state of the nth microgrid, y n = 0 means non-island, y n =1 is an isolated island.

[0022] is the predictor variable, which is the result predicted by the multiple linear regression model; is the first principal component variable obtained by PLS method; is the second principal component variable obtained by PLS method; is the mth principal component variable obtained by PLS method; a1, a2, a m are the coefficients of the first, second, and mth variables in the empirical expression of multiple linear regression, indicating that the variables and proportion among all variables;

[0023] a i is the coefficient of the empirical expression of multiple linear regression; m is the total number of electrical characteristics of the main components extracted; r i k For variables The corresponding correlation coefficient; ω ih is the axis ω h The i-th principal component coefficient component.

[0024] Based on the above formula, the variable projection importance VIP value is calculated. The specific formula is as follows:

[0025]

[0026]

[0027] Where: For variables Corresponding VIP value; Rdy is the hth extracted electrical characteristic component for the dependent variable y in the island state data set Y j The explanatory power of CRdy is m principal component electrical characteristic quantities to y j The cumulative explanatory power of r(y j ;t h ) is the jth dependent variable y j The hth extracted main component electrical characteristic component t h The correlation coefficient between them; n is the number of samples.

[0028] All training sample data and detection sample data of the present invention are collected from the PCC. Starting from the time of island occurrence, the average value of electrical characteristic quantities within three consecutive cycles is collected. 70% of the samples are randomly selected from the collected data set as the training sample set to train the single-phase three-phase hybrid microgrid group island classification model based on improved Adaboost; the remaining 30% of the samples are grouped as the detection sample set to test the classification accuracy of the trained island classification model.

[0029] The step 3 specifically includes:

[0030] (1): Slime Mould Algorithm (SMA):

[0031] Although the traditional Adaboost learner can improve the utilization rate of each weak classifier, it is still easily affected by disturbance factors and redundant features. Therefore, the present invention improves the Adaboost learner based on SMA to improve its ability to classify samples and reduce the impact of disturbance factors;

[0032] The Adaboost learner trains the weak classifier G(x,l,ψ,δ) to reduce its classification error rate e m Reduced to the minimum, where the most important parameters are the bias parameter ψ and the threshold parameter δ, x is the test sample; l(x) is the function for calculating the eigenvalue.

[0033] The present invention effectively optimizes the bias parameter ψ and the threshold parameter δ based on the slime mold optimization algorithm SMA which has dynamic search capability and is not prone to falling into local optimality;

[0034] First, the parameters of the slime mold optimization algorithm SMA are initialized with the classification error rate e of the Adaboost learner. m To optimize the objective function of the slime mold optimization algorithm SMA, the Adaboost learner is trained, including the following contents:

[0035] The objective function is the classification error rate e of the Adaboost learner m , whose expression is

[0036] Where: is the weight of the i-th data in the m-th round of iterative training; G m (x i ) is the current weak classifier for the i-th sample x i The prediction result of y i is the true label of the i-th sample.

[0037] Secondly, the objective function is used to calculate the fitness and sort the population. Based on the sorting results, the population position is updated, which includes the following:

[0038] The objective function is used to calculate the fitness value of each individual, and all individuals in the population are sorted according to the fitness value. The position of each individual is updated according to the sorting result, and a new population position is obtained. For each individual i, its fitness value is its corresponding classification error rate, that is,

[0039] Next, calculate the population fitness value, update the optimal fitness value and the best individual position of the slime mold, including the following:

[0040] The set of individual fitness values ​​is the population fitness value. After calculating the fitness value of each individual, find the individual with the smallest fitness value from the set, and use its fitness value and position as the optimal fitness value and optimal individual position of the current population.

[0041]

[0042] Where: x best is the best individual position in the current population; f best is the fitness value of the best individual position in the current population.

[0043] Then, after the optimal conditions are met, the best individual position is output, and the global optimal parameter optimization result of the Adaboost learner is obtained, which specifically includes the following contents:

[0044] If the number of iterations reaches the maximum value, the algorithm is considered to have converged, and the position of the current best individual is output. The parameters in the best individual position are the global optimal parameters of the Adaboost learner.

[0045] T=500

[0046] X best =[δ * ,ψ * ]

[0047] Where: T is the number of iterations; X best is the optimal individual position vector output; δ * , * are the optimal threshold parameter and the optimal bias parameter respectively.

[0048] Finally, the optimized parameters are input into the Adaboost learner to obtain the trained highest performance binary classifier, which specifically includes the following contents:

[0049] The global optimal parameters obtained above are input into the Adaboost learner to retrain the model. After the training is completed, the trained binary classifier with the highest performance is output.

[0050] G m (x i )=G(x,l,ψ * ,δ * )

[0051]

[0052]

[0053] Where: x is the test sample, l(x) is the function for calculating the eigenvalue; ψ * is the optimal bias parameter; δ * is the optimal threshold parameter; is the weight of the i-th data in the m-th round of iterative training; G m (x i ) is the current weak classifier for the i-th sample x i The prediction result of y i is the true label of the i-th sample; e m is the classification error rate of the mth round of iterative training; α m is the weight of the weak classifier in the mth round of iterative training; y(x) is the final strong classifier; M is the number of iterations; sign is the sign function.

[0054] (2): GSN weight learning method:

[0055] The present invention fully considers the impact of sample weight changes on algorithm performance, and improves the original normalization method of Adaboost learner based on GSN weight learning method; GSN is a weight learning method formed by integrating batch normalization (BatchNormalization, BN), instance normalization (Instance Normalization, IN), and group normalization (GroupNormalization, GN). This weight learning method can achieve the effect of improving accuracy and reducing computational redundancy. The specific expression is as follows:

[0056]

[0057] Where: is the weight of the i-th sample in the m+1-th round of iterative training; is the weight of the i-th sample in the m-th round of iterative training; α m is the weight of the weak classifier in the mth round of iterative training; y i is the true label of the i-th sample; Gm (x i ) is the prediction result of the weak classifier for the i-th sample in the m-th round of iterative training;

[0058] γ is the initialization learnable scaling factor; ζ is the initialization learnable translation factor; υ is the hyperparameter; μ GSN , are the feature mean and feature variance obtained by GSN processing; μ BN , are the mean and variance of the features with respect to BN; μ IN , are the mean and variance of the features with respect to IN; μ GN , are the mean and variance of the features with respect to GN respectively; are weight coefficients, representing the contribution of BN, IN and GN to the special mean and special variance respectively.

[0059] The weight learning method formula in the above formula is:

[0060]

[0061] Where: is the weight value; Ω = {BN, IN, GN} is the set of normalization methods; q is the selected normalization method type; θ q is a control parameter, whose value is greater than 0; ln(1+θ q ) is the parameter θ q The natural logarithm transformation of .

[0062] Among them, the use of logarithmic function can effectively avoid the phenomenon that the weights of correctly classified samples in the original Adaboost learner decrease in a logarithmic trend.

[0063] Based on the GSN weight learning method, the original normalization method of the Adaboost learner is improved, specifically as follows:

[0064] The batch normalization, instance normalization and group normalization methods are combined to form GSN, which is introduced into the Adaboost learner to replace the original normalization method.

[0065] The weight of the weak classification is adjusted according to the following formula for the sample weight of the next round of iterative training:

[0066]

[0067] Where: H (m+1) is the sample weight matrix for the m+1th round of iterative training; are the weights of sample 1, sample 2, and sample n in the m+1th round of iterative training; is the weight of the i-th sample in the m+1-th round of iterative training; is the weight of the i-th sample in the m-th round of iterative training; α m is the weight of the weak classifier in the mth round of iterative training; y i is the true label of the i-th sample; G m (x i ) is the prediction result of the weak classifier for the i-th sample in the m-th round of iterative training; γ is the initialization learnable scaling factor; ζ is the initialization learnable translation factor; υ is the hyperparameter; μ GSN , are the feature mean and feature variance obtained after GSN processing.

[0068] In step 4, based on the strong classifier generated by the improved Adaboost, an islanding detection model for the microgrid is constructed, and the islanding detection result is output according to the threshold of the electrical characteristic quantity of the islanding detection.

[0069] The island detection model of the microgrid group is constructed through the improved Adaboost learner, which is used to monitor and identify in real time whether the microgrid group is disconnected from the main grid and enters the island state. The model collects the electrical characteristic data (voltage amplitude, frequency and phase angle changes) at the PCC of the microgrid group and uses the improved Adaboost algorithm to analyze and classify these data. When it is detected that the electrical characteristic quantity exceeds the preset threshold, the model will output the island detection result, indicating that the microgrid has entered the island state. The island detection model of the microgrid group implements the island detection method flow, such as Figure 3 shown.

[0070] The island detection result is output based on the threshold of the electrical characteristic quantity of the island detection. The details are as follows:

[0071] The threshold is set according to the difference in electrical characteristic quantities in normal operation and island state. The present invention selects voltage, phase angle change and frequency as the electrical characteristic quantities to be monitored. The voltage threshold is set to a fluctuation range of ±10%, that is, [198V, 242V]; the phase angle change threshold is set to [-5°, +5°]; the frequency threshold is set to a fluctuation range of ±1%, that is, [49.5Hz, 50.5Hz]. When one or more electrical characteristic quantities exceed these set thresholds, it can be determined that the microgrid is in an island state.

[0072] The present invention provides a single-phase three-phase hybrid microgrid group island detection method based on improved Adaboost, and the technical effects are as follows:

[0073] 1) In step 1 of the present invention, the electrical characteristic data collected at the PCC covers a variety of electrical parameters of the microgrid group, such as voltage, current, power, etc., which can fully understand the operating status of the microgrid. These detailed electrical parameter data accurately reflect the actual operating status of the microgrid, which helps to improve the accuracy and reliability of island detection.

[0074] 2) In step 2 of the present invention, the PLS method is used to analyze the correlation between electrical features and islanding states, which can effectively screen out the features with the strongest correlation with islanding detection from a large number of features. This method reduces the dimension of the data and improves the quality of training and test samples, thereby improving the accuracy and efficiency of the model for islanding state detection.

[0075] 3) In step 3 of the present invention, by introducing the SMA algorithm and the GSN weight learning method into the Adaboost learner, the performance of the model can be significantly improved. Among them, the SMA algorithm can improve the model classification ability and reduce the impact of disturbances, and the GSN weight learning method can shorten the model detection time.

[0076] 4) In step 4 of the present invention, by integrating multiple optimized weak classifiers, the improved Adaboost algorithm can significantly improve classification accuracy and robustness, especially when dealing with complex and changeable electrical characteristics. At the same time, the improvement of the sample weight update mechanism enables the model to more effectively handle noise and abnormal data, shorten the detection time and improve the detection efficiency. Therefore, the island detection model constructed using the improved Adaboost strong classifier can significantly improve the accuracy and response speed of island detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of the island detection method proposed by the present invention.

[0078] Figure 2(a) shows the island detection result when the three-phase unbalance is 2%;

[0079] Figure 2(b) shows the island detection result when the three-phase unbalance is 8%;

[0080] Figure 2(c) shows the island detection result when the three-phase unbalance is 15%.

[0081] Figure 3 Flowchart of island detection method implemented for island detection model of microgrid cluster. DETAILED DESCRIPTION

[0082] The island detection method of single-phase and three-phase hybrid microgrid based on improved Adaboost includes the following steps:

[0083] S1: Collect sample data of different electrical characteristics of single-phase and three-phase hybrid microgrid groups measured at the point of common coupling (PCC);

[0084] S2: The collected sample data are processed using the Partial Least Squares (PLS) method to screen out a set of strongly correlated electrical feature quantities, and form a training sample set and a test sample set;

[0085] S 3: Initialize the parameters of the Slime Mould Algorithm (SMA), the training sample set and its corresponding weights, and calculate the current classification error rate;

[0086] S4: Train the model with the classification error rate as the objective function, update the population position, and calculate the fitness value of the current individual;

[0087] S5: Update the best fitness value and the best individual position according to the current fitness value and the historical best fitness value, and judge whether the optimal conditions are met. If so, output the global best optimization parameters of the Adaboost learner, otherwise repeat S4;

[0088] S6: Input the optimized parameters into the Adaboost learner, and use the Group Switchable Normalization (GSN) weight learning method to update the weights of the original training sample set to obtain the best strong binary classifier;

[0089] The GSN weight learning method is used to update the weights of the training samples, and then all weak classifiers are linearly weighted with their corresponding weights after the update to obtain the best strong binary classifier. The GSN weight learning method can effectively update the weights of the training samples in the Adaboost learner, thereby improving the performance of the classifier.

[0090] S7: Use the improved Adaboost learner to train the generated set of strongly correlated electrical feature quantities to obtain the best strong classifier;

[0091] Firstly, the electrical characteristic data at the PCC are collected. Then, the PLS method is used to analyze the correlation between these electrical characteristics and the islanding state of the single-phase and three-phase hybrid microgrid. The set of electrical characteristics that are strongly correlated with islanding detection is extracted and the training and test data sets are constructed. Then, the Adaboost learner is improved based on the SMA algorithm and the GSN weight learning method, and the model is trained using the training data set to finally obtain the best strong classifier.

[0092] S8: Based on the strong classifier generated by improved Adaboost, an island detection model for microgrid clusters is constructed.

[0093] Figure 1 This is a flow chart of the island detection method proposed in the present invention. The present invention discloses a single-phase three-phase hybrid microgrid group island detection method based on improved Adaboost. First, sample data of different electrical characteristic quantities in the single-phase three-phase hybrid microgrid group measured at the PCC are collected. Secondly, based on PLS, the correlation analysis and extraction of various electrical characteristic quantities measured at the PCC are carried out, the key characteristic quantities are screened out, and the training set and the test set are formed. Then, the Adaboost learner is improved based on the SMA and GSN weight learning methods to improve the classification ability, reduce the impact of disturbances and shorten the detection time. Finally, the improved Adaboost learner is used to train the generated set of strongly correlated electrical characteristic quantities to obtain the best strong classifier, and the island detection model of the microgrid group is constructed based on the best strong classifier, and the island detection result is output according to the real-time test set data at the PCC.

[0094] Figure 2(a), Figure 2(b), and Figure 2(c) show the island detection results under different three-phase imbalance. The island detection signal of 0 indicates that it is not detected as an island, and 1 indicates that it is detected as an island. When the single-phase three-phase hybrid microgrid suddenly enters the island state at 0.35s, after a certain period of time, the grid-connected circuit breaker is disconnected after being detected as an island state, the system forms a power island, and the grid-connected current becomes 0. After that, the PCC point voltage begins to decay until it reaches 0. Figure 2(a) to Figure 2(c) It can be seen that the method proposed in the present invention can obtain relatively accurate islanding detection results when the three-phase unbalance is 2%, 8% and 15% respectively.

[0095] Table 2 Detection time of different island detection methods

[0096]

[0097] Table 2 shows the detection time of different island detection methods. When the three-phase imbalance is 2%, the island detection time of the method proposed in the present invention is shortened by 17.39%, 13.64% and 11.63% respectively compared with the comparison method 1, comparison method 2 and comparison method 3. When the three-phase imbalance is 8%, the island detection time of the method proposed in the present invention is shortened by 17.20%, 15.58% and 12.16% respectively compared with the comparison method 1, comparison method 2 and comparison method 3. When the three-phase imbalance is 15%, the island detection time of the method proposed in the present invention is shortened by 11.05%, 9.55% and 8.52% respectively compared with the comparison method 1, comparison method 2 and comparison method 3. When the three-phase imbalance degree increases from 2% to 15%, the time spans of the four island detection methods are 0.047s, 0.049s, 0.052s and 0.057s respectively, and the island detection time span of the method proposed in the present invention is reduced by 4.08%, 9.62% and 17.54% respectively compared with the comparison method 1, the comparison method 2 and the comparison method 3. Therefore, compared with the comparison method 1, the comparison method 2 and the comparison method 3, the detection time of the method proposed in the present invention under different three-phase imbalance conditions is the shortest, the influence of the three-phase imbalance is the smallest, and the time consumption is the smallest.

Claims

1. The island detection method of single-phase and three-phase hybrid microgrid based on improved Adaboost is characterized by The following steps are involved: Step 1: Collect sample data of different electrical characteristic quantities in the single-phase and three-phase hybrid microgrid group measured at the common coupling point PCC; Step 2: Based on the partial least squares method, the various electrical characteristic quantities measured at the common coupling point are subjected to correlation analysis, and the strongly correlated electrical characteristic quantities are screened out to form a training set and a test set; Step 3: Improve the Adaboost learner based on the slime mold optimization algorithm and GSN weight learning method; Step 4: Build an island detection model for the microgrid based on the improved Adaboost learner, and output the island detection results according to the real-time test set data at the PCC.

2. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 1 is characterized by: In step 1, the island electrical characteristic quantities considered include voltage, voltage change rate over time, voltage total harmonic distortion rate, current, current change rate over time, current total harmonic distortion rate, frequency, frequency change rate over time, impedance, impedance change rate over time, active power, active power change rate over time, reactive power, reactive power change rate over time and power factor.

3. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 1 is characterized by: In the step 2, the basic island electrical characteristics of the microgrid group are ranked by importance based on the PLS method, and the electrical characteristics with variable projection importance (VIP) index exceeding 1 are screened out as strongly correlated electrical characteristics for microgrid group island detection.

4. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 3 is characterized by: The PLS method is used to process the collected basic electrical characteristic data and establish a multivariate linear regression empirical expression: Y={y1,y2,…,y j ,…,y n } Where: X is the data set consisting of basic island electrical characteristic quantities; is the kth electrical characteristic corresponding to the PCC of the first microgrid; is the kth electrical characteristic corresponding to the PCC of the second microgrid; The i-th microgrid PCC corresponds to the k-th electrical characteristic quantity; The nth microgrid PCC corresponds to the kth electrical characteristic quantity; Y is the data set of island status; y1 is the island status of the first microgrid, y1=0 is non-island, y1=1 is island; y2 is the island status of the second microgrid, y2=0 is non-island, y2=1 is island; j is the island state of the jth microgrid, y j = 0 means non-island, y j =1 for an isolated island; y n is the island state of the nth microgrid, y n = 0 means non-island, y n =1 for an isolated island; is the predictor variable, which is the result predicted by the multiple linear regression model; is the first principal component variable obtained by PLS method; is the second principal component variable obtained by PLS method; is the mth principal component variable obtained by PLS method; a1, a2, a m are the coefficients of the first, second, and mth variables in the empirical expression of multiple linear regression, indicating that the variables and proportion among all variables; a i is the coefficient of the empirical expression of multiple linear regression; m is the total number of electrical characteristics of the main components extracted; For variables The corresponding correlation coefficient; ω ih is the axis ω h The i-th principal component coefficient component; Based on the above formula, the variable projection importance VIP value is calculated. The specific formula is as follows: Where: For variables Corresponding VIP value; Rdy is the hth extracted electrical characteristic component for the dependent variable y in the island state data set Y j The explanatory power of CRdy is m principal component electrical characteristic quantities to y j The cumulative explanatory power of r(y j ;t h ) is the jth dependent variable y j The hth extracted main component electrical characteristic component t h The correlation coefficient between them; n is the number of samples.

5. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 1 is characterized by: In step 3, the Adaboost learner trains the weak classifier G(x, l, ψ, δ) to reduce its classification error rate e m To the minimum, among which the most important parameters are the bias parameter ψ and the threshold parameter δ, and x is the test sample; the slime mold optimization algorithm SMA effectively optimizes the bias parameter ψ and the threshold parameter δ.

6. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 5 is characterized by: First, the parameters of the slime mold optimization algorithm SMA are initialized with the classification error rate e of the Adaboost learner. m To optimize the objective function of the slime mold optimization algorithm SMA, the Adaboost learner is trained, including: The objective function is the classification error rate e of the Adaboost learner m , whose expression is Where: is the weight of the i-th data in the m-th round of iterative training; G m (x i ) is the current weak classifier for the i-th sample x i The prediction result of y i is the true label of the i-th sample; Secondly, the objective function is used to calculate the fitness and sort the population. Based on the sorting results, the population position is updated, which includes the following: The objective function is used to calculate the fitness value of each individual, and all individuals in the population are sorted according to the fitness value. The position of each individual is updated according to the sorting result, that is, a new population position is obtained; for each individual i, its fitness value is its corresponding classification error rate, that is, Next, calculate the population fitness value, update the optimal fitness value and the optimal individual position of the slime mold, including: The set of individual fitness values ​​is the population fitness value. After calculating the fitness value of each individual, find the individual with the smallest fitness value from the set, and use its fitness value and position as the optimal fitness value and optimal individual position of the current population. Where: x best is the best individual position in the current population; f best is the fitness value of the best individual position in the current population; Then, after the optimal conditions are met, the best individual position is output, and the global optimal parameter optimization result of the Adaboost learner is obtained, including: If the number of iterations reaches the maximum value, the algorithm is considered to have converged, and the position of the current best individual is output. The parameters in the best individual position are the global optimal parameters of the Adaboost learner. T＝500 X best =[δ * ,ψ * ] Where: T is the number of iterations; X best is the optimal individual position vector output; δ * , * are the optimal threshold parameter and the optimal bias parameter respectively; Finally, the optimized parameters are input into the Adaboost learner to obtain the trained highest performance binary classifier, including: G m (x i )=G(x,l,ψ * ,d * ) Where: x is the test sample, l(x) is the function for calculating the eigenvalue; ψ * is the optimal bias parameter; δ * is the optimal threshold parameter; is the weight of the i-th data in the m-th round of iterative training; G m (x i ) is the current weak classifier for the i-th sample x i The prediction result of y i is the true label of the i-th sample; e m is the classification error rate of the mth round of iterative training; α m is the weight of the weak classifier in the mth round of iterative training; y(x) is the final strong classifier; M is the number of iterations; sign is the sign function.

7. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 6 is characterized by: GSN is a novel method that integrates batch normalization (BN), instance normalization (Instance Normalization (IN) and group normalization (GN) are three normalization-based weight learning methods. The specific expressions are as follows: Where: is the weight of the i-th sample in the m+1-th round of iterative training; is the weight of the i-th sample in the m-th round of iterative training; α m is the weight of the weak classifier in the mth round of iterative training; y i is the true label of the i-th sample; G m (x i ) is the prediction result of the weak classifier for the i-th sample in the m-th round of iterative training; γ is the initialization learnable scaling factor; ζ is the initialization learnable translation factor; υ is the hyperparameter; μ GSN , are the feature mean and feature variance obtained by GSN processing; μ BN , are the mean and variance of the features with respect to BN; μ IN , are the mean and variance of the features with respect to IN; μ GN , are the mean and variance of the features with respect to GN respectively; are weight coefficients, representing the contribution of BN, IN and GN to the special mean and special variance respectively; The weight learning method formula in the above formula is: Where: is the weight value; Ω = {BN, IN, GN} is the set of normalization method types; q is the selected normalization method type; θ q is a control parameter, whose value is greater than 0; ln(1+θ q ) is the parameter θ q The natural logarithm transformation of .

8. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 7 is characterized by: The original normalization method of the Adaboost learner is improved based on the GSN weight learning method, specifically including: adjusting the sample weights of the next round of iterative training for the weights of the weak classification according to the following formula: Where: H (m+1) is the sample weight matrix for the m+1th round of iterative training; are the weights of sample 1, sample 2, and sample n in the m+1th round of iterative training; is the weight of the i-th sample in the m+1-th round of iterative training; is the weight of the i-th sample in the m-th round of iterative training; α m is the weight of the weak classifier in the mth round of iterative training; y i is the true label of the i-th sample; G m (x i ) is the prediction result of the weak classifier for the i-th sample in the m-th round of iterative training; γ is the initialization learnable scaling factor; ζ is the initialization learnable translation factor; υ is the hyperparameter; μ GSN , are the feature mean and feature variance obtained after GSN processing.

9. The single-phase and three-phase hybrid microgrid island detection method based on improved Adaboost according to claim 1 is characterized by: In step 4, based on the strong classifier generated by the improved Adaboost, an island detection model of the microgrid is constructed, and the island detection result is output according to the threshold of the electrical characteristic quantity of the island detection; the details are as follows: The threshold is set according to the difference in electrical characteristic quantities under normal operation and island state, and voltage, phase angle change and frequency are selected as the monitored electrical characteristic quantities; the voltage threshold is set to a fluctuation range of ±10%, that is, [198V, 242V]; the phase angle change threshold is set to [-5°, +5°]; the frequency threshold is set to a fluctuation range of ±1%, that is, [49.5Hz, 50.5Hz]; when one or more electrical characteristic quantities exceed these set thresholds, it can be determined that the microgrid is in an island state.

Images