An island detection method, device and medium for a photovoltaic grid-connected system based on a stacking structure

Abstract

The application provides an island detection method, device and medium for a photovoltaic grid-connected system based on a Stacking stack structure, the method comprising the following steps: step 1: constructing a simulation model of a four-machine parallel photovoltaic grid-connected system, and collecting voltage, electric frequency and other electrical characteristic quantity training samples at a grid-connected point; step 2: taking the training samples as input vectors, performing k-fold cross-validation training on four base models of a first layer of Stacking respectively, and obtaining island detection models respectively based on the training samples; step 3: integrating output vectors of the four base models, combining the output vectors to construct a training set of a next round of meta-models, obtaining a training and verification set of a second layer of meta-models, training and predicting the meta-models, and obtaining a final island detection model. The model can jointly make decisions on the state of the photovoltaic grid-connected system from different algorithm logics, and finally can realize more reliable detection effects than a single excellent island detection model, and can provide continuous and high-accuracy state detection output results of the photovoltaic grid-connected system.

Description

Technical Field

[0001] This invention relates to the field of islanding detection technology for photovoltaic grid-connected systems, specifically to an islanding detection method for photovoltaic grid-connected systems based on a stacking strategy that integrates multiple heterogeneous models. Background Technology

[0002] With the rapid development of distributed photovoltaic (PV) power generation in recent years, the integration of large-scale distributed PV power generation will bring various challenges to the original power system. This also means that the coordinated operation of PV grid-connected systems requires higher standards for the safety and accuracy of the power system. Specifically, for PV grid-connected systems, planned or unplanned power outages on the main grid side can easily lead to the PV power generation system becoming an "island" between itself and local loads. If this operating state is not detected in time, it will pose a significant safety hazard to equipment maintenance personnel and the stable operation of grid-connected equipment. Therefore, PV grid-connected systems must have timely and effective islanding identification capabilities. Furthermore, the increasingly complex operating states of modern power systems will present new challenges to traditional islanding identification methods.

[0003] In recent years, various strategies utilizing ensemble models have been proposed to address the tasks required by power systems. Among these, intelligent island detection methods that integrate multiple weak classifiers of the same type using ensemble learning strategies have demonstrated superior classification performance. Specifically, the Random Forest (RF) island detection model, employing a depth-infinite random growth strategy, shows significant effectiveness in reflecting anomalous features but suffers from poor generalization. The Light Gradient Boosting Machine (LightGBM) classification model, with depth limitations and regularization optimization, performs better overall but is insensitive to the electrical characteristics of abrupt changes and easily overlooks potential operational characteristics brought about by system mutations. Therefore, different intelligent classification models offer different data analysis perspectives, while single island detection models often struggle to handle grid-connected operation data with varying characteristics, making it difficult to comprehensively analyze the operating status of the grid-connected system from different angles. Although such methods offer excellent island detection performance and can obtain the island's pulse signal within 2 seconds of its occurrence, they are still effective. However, after an islanding event occurs, although the detection signal of the island can be obtained, the subsequent output signal may be unstable. For systems that need to obtain the status signal of the grid-connected system after an islanding event, it is necessary to improve the output stability of this type of method. Summary of the Invention

[0004] In the islanding status detection of photovoltaic grid-connected systems, for applications that require a continuous and stable output status detection signal, the confidence of a single intelligent islanding detection method is not high, and there is a lack of strategies for efficiently integrating multiple different intelligent islanding detection models.

[0005] The first aspect of this invention proposes a method for constructing multiple heterogeneous island detection models based on a stacking structure that integrates RF, LightGBM, Catboost, XGBoost, and Bayesian classification algorithms. This enables the models to make joint decisions on the state of the photovoltaic grid-connected system from different algorithmic logics, ultimately achieving a more reliable detection effect than a single superior island detection model, and providing continuous and highly accurate state detection output results for the photovoltaic grid-connected system.

[0006] The specific scheme of the Stacking multi-origin island detection model constructed in this invention includes the following steps:

[0007] Step 1: Establish a simulation model of a four-unit parallel photovoltaic grid-connected system and collect training data samples and test samples;

[0008] Step 2: Using the collected training sample set as the input vector, train the four base models of the first layer of Stacking with k-fold cross-validation to obtain the island detection model built based on the training samples. At the same time, the output vectors obtained after cross-validation of each base model are combined to construct the training set of the next round of meta-model.

[0009] Step 3: Integrate the output vectors obtained after k-fold cross-validation of the four base models to obtain the training and validation set of the second-layer meta-model. Train and predict the meta-model to obtain the final island detection model.

[0010] Furthermore, in step 1, to cover as much as possible the various operating conditions of the multi-unit parallel photovoltaic grid-connected system, a multi-unit parallel photovoltaic power generation grid-connected system is established using MATLAB / SIMULINK. The photovoltaic power plant consists of four adjustable photovoltaic (PV) arrays. Operating data of the photovoltaic grid-connected system is collected to construct a dataset S. K-fold cross-validation is used to divide the training and validation sets to provide data for training each base model. Data acquisition is performed through the following steps: controlling the number of grid-connected photovoltaic array units to vary within the range of 1 to 4, and measuring the ratio P of the active power absorbed by the load to the active power generated by the photovoltaic array. load / P PV The sampling step size is 0.001 when the load is purely resistive, with a range of 0.8 to 1.5, and a total of 100 simulations are performed. When the load contains inductive components, the load quality is 0.9 to 1.3, with a sampling step size of 0.001, and a total of 100 simulations are performed.

[0011] Furthermore, k is 5.

[0012] Furthermore, islanding detection also includes the following non-islanding operating states: grid connection fluctuation operating state, voltage sudden change operating state, power system fault operating state, and harmonic interference operating state.

[0013] Furthermore, the following steps are required to address the aforementioned grid-connected fluctuating operating state: control the number of grid-connected photovoltaic array units to vary within the range of 1 to 4, and control the photovoltaic output power fluctuation P. load Q load The value varies between 0 and 2 times the rated value, with a sampling step size of 0.001, and a total of 50 simulations are performed.

[0014] Furthermore, the following steps are required for the voltage surge operation state: control the number of grid-connected photovoltaic array units to vary within the range of 1 to 4, and under rated load, increase the rated voltage of the system by 10% to 30% with a sampling step size of 0.001; under rated load, switch in the large capacitor load, and the rated voltage drops by 15% to 30% with a sampling step size of 0.001, for a total of 50 simulations.

[0015] Furthermore, the following steps are required for the power system fault operation state: control the number of grid-connected photovoltaic array units to vary within the range of 1 to 4, and simulate power system fault conditions such as single-phase ground fault, two-phase fault, and two-phase ground fault under rated load, with a sampling step size of 0.001, for a total of 50 simulations.

[0016] Furthermore, the following steps are required to address the operation status of the aforementioned harmonic interference: control the number of grid-connected photovoltaic array units to vary within the range of 1 to 4, and keep the ratio P of the active power absorbed by the load to the active power generated by the photovoltaic system constant. load / P PV Within the range of 0.8 to 1.5, the 5th, 7th, and 11th harmonic disturbances were injected into the three phases respectively, with a sampling step size of 0.001, for a total of 50 simulations.

[0017] Furthermore, four heterogeneous base learners were selected as the four base models in the first layer, based on island detection models constructed using XGBoost, Catboost, RF, and LightGBM algorithms. A stacking strategy was employed to fuse the two-layer models, and a Bayesian classifier was chosen as the meta-model to construct the final island detection model.

[0018] A second aspect of the present invention provides an islanding detection device for a photovoltaic grid-connected system based on a stacking structure, which is used to operate the method described in the first aspect, comprising:

[0019] The data acquisition unit is used to construct a simulation model of a four-unit parallel photovoltaic grid-connected system and to collect training samples of electrical characteristic quantities such as voltage and frequency at the grid connection point.

[0020] The training unit is used to take the training samples as input vectors and perform k-fold cross-validation training on the four base models of the first layer of Stacking to obtain the island detection models built based on the training samples.

[0021] The integration unit is used to integrate the output vectors of the four base models, combine the output vectors to construct the training set of the next round meta-model, obtain the training and validation set of the second-layer meta-model, train and predict the meta-model, and obtain the final island detection model.

[0022] A third aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the island detection method described in the first aspect.

[0023] This invention relates to a method, apparatus, and computer medium for detecting islanding in a heterogeneous photovoltaic grid-connected system based on a stacking architecture. The technical advantages are as follows:

[0024] 1) This invention utilizes a stacking strategy to fuse multiple different island classification models to achieve the final island detection model. By collecting multiple mathematically processed electrical feature quantities at the grid connection point of the photovoltaic grid-connected system, and through data mining, the judgment threshold for multi-feature joint decision-making is obtained to achieve the final grid-connected system state decision.

[0025] 2) The proposed Stacking structure-based multi-source island detection model enables the model to make joint decisions on the state of the photovoltaic grid-connected system from different algorithm logics. Ultimately, it can achieve a more reliable detection effect than a single superior island detection model and provide continuous and highly accurate state detection output results for the photovoltaic grid-connected system. Attached Figure Description

[0026] Figure 1 This is a simulation diagram of the photovoltaic grid-connected system implemented in this invention.

[0027] Figure 2 This is a structural diagram of the island detection model built based on the Stacking structure of this invention.

[0028] Figure 3 This is a detection effect diagram of the island detection model built based on the Stacking structure of this invention.

[0029] Figure 4 This is a diagram showing the detection results of the comparative examples in this invention.

[0030] Figure 5This is a structural diagram of the islanding detection device for a photovoltaic grid-connected system based on a stacking structure, as described in this invention. Detailed Implementation

[0031] The technical solutions of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] Example 1 is a method for islanding detection in a photovoltaic grid-connected system based on a stacking structure, including the following steps:

[0033] Step 1: Construct a simulation model of a four-unit parallel photovoltaic grid-connected system and collect training samples of electrical characteristic quantities such as voltage and frequency at the grid connection point.

[0034] Step 2: Using the training samples as input vectors, perform k-fold cross-validation training on the four base models of the first layer of Stacking to obtain the island detection models built based on the training samples.

[0035] Step 3: Integrate the output vectors of the four base models, combine the output vectors to construct the training set of the next round meta-model, obtain the training and validation set of the second-layer meta-model, train and predict the meta-model, and obtain the final island detection model.

[0036] Figure 1 This invention utilizes MATLAB / SIMULINK to establish a multi-unit parallel photovoltaic (PV) grid-connected system. The PV power plant consists of four adjustable PV arrays. Step 1 involves constructing a simulation model of the four-unit parallel PV grid-connected system using MATLAB / SIMULINK. The PV grid-connected system consists of four adjustable PV arrays. Operating data of the PV grid-connected system is collected to construct a dataset S. K-fold cross-validation is used to divide the dataset into training and validation sets for training and validating each base model, and the output results of each model are obtained. Simultaneously, each step-up transformer is controlled by a separate Maximum Power Point Tracker (MPPT). The MPPT uses a perturbation-observation method to obtain the tracking of the maximum power point. Key parameters of the grid-connected system are shown in Table 1. Multiple adjustable active and reactive power loads are selected as loads, and the power distribution system is simplified accordingly.

[0037] Table 1 Key parameters of the simulated photovoltaic grid-connected system

[0038]

[0039] Figure 2The island detection method proposed in this invention, based on the Stacking strategy, requires the collection of 16 electrical characteristic quantities at the grid connection point of the photovoltaic grid-connected simulation system as shown in Table 2, and the construction of training and testing sets.

[0040] Table 2 Electrical Characteristic Quantities

[0041]

[0042] To cover as many conditions as possible the islanded and non-islanded operation of photovoltaic grid-connected systems, different operating states of photovoltaic grid-connected systems were simulated, as shown in Table 3.

[0043] Table 3.2 Simulation Condition Samples

[0044]

[0045]

[0046] Step 2, constructing the first-layer base model of the stacking structure, needs to consider the following three factors:

[0047] S10.1: Stacking effectively integrates multiple dissimilar classification models. This means that classification models with different algorithmic architectures can perform data mining on the photovoltaic grid-connected system's training set from different logical perspectives, approaching the task from different spatial viewpoints of the data. Therefore, model selection should not be singular; it should involve combining and selecting from classification models with different algorithmic structures.

[0048] S10.2: Considering the high safety requirements of power system operation and the processing time of actual equipment, the training cost of island detection models based on deep learning is relatively high, and each training requires a lot of time. It is difficult to adapt to the high-dimensional, large-data sample power system operation data, which will lead to long response time and is not suitable for actual engineering applications. Therefore, a machine learning model that is more suitable for actual engineering applications should be selected for task processing.

[0049] S10.3: The performance difference between the base models in the first layer should not be too large. The base models with excessive performance lag will restrict the effectiveness of the features of the output matrix of the first layer.

[0050] Therefore, based on the above three considerations, this invention selects four heterogeneous base learners constructed by XGBoost, Catboost, RF, and LightGBM algorithms as the first layer to improve the adaptability of the overall stacked structure to different data.

[0051] In step 2, for the base model M1 of the first layer constructed based on the XGBoost algorithm, the optimal model is optimized by minimizing the loss function (which is also the objective function) using a forward iterative approach. First, for the i-th sample of the t-th tree, the model's predicted value is:

[0052]

[0053] In the above formula, f is the predicted value for sample i after the t-th iteration; t (x i ) represents the t-th tree pair x i The predicted value; This is the predicted value for the (t-1)th tree.

[0054] As a forward iterative algorithm, XGBoost should pay more attention to the fitting of the t-th tree in the current round. Therefore, the first t-1 trees can be treated as constants for the objective function Obj. (t) The objective function is simplified by performing a second-order expansion using the Taylor formula, resulting in the final objective function:

[0055]

[0056] In step 2, for the base model M2 constructed using CatBoost in the first layer, an island detection base model M2 is constructed with a fully symmetric decision tree structure. The first stage initially employs unbiased gradient estimation, while the second stage continues to use the traditional GBDT scheme. Let F... i To construct the model after the i-th tree, the gradient value of the k-th training sample after the i-th tree is g. i (X k ,y k To ensure that the gradient estimate is unbiased for model F. i Then it is necessary to have no sample X k In the case of participation, model F i Training is then performed. This necessitates further computation of unbiased gradient estimates for all training sample data. Therefore, for each sample X... k In this invention, the M2 base model uses a single model M consisting of multiple decision trees. k Furthermore, this model does not use sample X. k Instead of iteratively updating the gradient estimate, M is used. k Estimate X k The gradient is calculated, and this estimate is used to score the tree, where the loss function is Loss(y). i ,a), where y is the label value and a is the calculated value.

[0057] In step 2, for the M3 base model of the first layer constructed based on the RF algorithm, the RF classification model, as an ensemble learning method based on decision trees, determines the final classification result by voting among each independent decision tree. This invention, based on the training set constructed from multiple feature quantities at the PCC point of a multi-machine parallel photovoltaic grid-connected system, utilizes the RF algorithm to unfold the modeling process, as follows: Figure 2 As shown, the Bootstrap method is used to randomly select multiple training sample subsets Sk (k = 1, 2, ..., n) from the training sample set. A classification model is built for each subset separately. The decision results of multiple weak classification models are combined, and the final island detection model is obtained by voting.

[0058] In step 2, for the M4 base model of the first layer constructed based on the LightGBM algorithm, the CART decision tree is used as the basic classifier. By combining multiple weak models with the forward stepwise algorithm, a final model composed of a set of regression trees is obtained. We use the negative gradient of the loss function to approximate the loss of this round instead of the residual, and then fit a CART decision tree. The negative gradient of the loss function L for the i-th sample in the t-th round is:

[0059]

[0060] Using (x i ,r ti (i = 1, 2, ..., n), fit a CART regression tree to obtain the t-th regression tree, whose corresponding leaf node region R tm Let m = 1, 2, ..., M, where M is the number of leaf nodes. Using a greedy approach, we only consider local optimization. For each leaf node, its output value is c. We find the output value c that minimizes the loss function, i.e., the best fit to the leaf node. tm :

[0061]

[0062] The decision tree fitting function for this round is obtained as follows:

[0063]

[0064] In the above formula, I(x∈R) tm Let be the training set of all nodes assigned to leaf nodes in round t, and obtain the strong classifier for this round through iterative updates:

[0065]

[0066] The final strong classifier is obtained by iteratively combining T base models:

[0067]

[0068] In the above formula, f0(x) is the initial model; f T (x) represents the final strong classification model obtained by iteratively combining T base models, which is equivalent to the final M4 island detection base model.

[0069] In step 3, the output of the first-layer base model is stacked and used to train the second-layer meta-model. By using the stacking strategy, the two-layer model is fused to achieve the construction of the final island detection model.

[0070] In step 3, considering that the first-layer model established in step 2 has high overall accuracy and performance, and to prevent the overfitting problem of the final output result of the stacked model, the present invention selects the Naive Bayes classifier, which is based on classical mathematical theory and has a simple structure, as the meta-model.

[0071] In step 3, the second layer meta-model based on a Bayesian classifier under the Stacking structure is constructed. The specific steps are as follows:

[0072] For island detection tasks, there are only two classification cases: "0" and "1", and a given category y is associated with its corresponding features x1 to x2. n According to Bayes' theorem, we can obtain:

[0073]

[0074] For Naive Bayes, it is necessary to assume that each feature is independent of the others.

[0075] P(x i |y,x1,...,x i-1 ,x i+1 ,...,x n )=P(x i |y)

[0076] If all features i hold true, the formula can be further simplified to:

[0077]

[0078] Therefore, for each input sample data P(y|x1,...,x) n If we calculate the classification result, we can obtain the following:

[0079]

[0080] To better analyze the output stability of the final islanding detection model, an islanding operation state of a photovoltaic grid-connected system with perfectly matched power was selected for detection. The islanding detection method of this invention was compared with a single islanding detection model based on the LightGBM algorithm, which showed better performance. The output of the islanding detection method was converted into raw probability values; that is, when the probability value was greater than 0.5, it was determined to be an islanding event; otherwise, it was a non-islanding operation event. When the model detected a non-islanding operation state, the output result was 0; when the model detected an islanding operation state, it output a signal with an amplitude of 1, thus allowing the simulation system to obtain the islanding pulse signal. Figure 3 The image shows the waveform diagram of the fully matched islanded operation state. Figure 3 It can be seen that when the grid-connected system enters the islanding operation state in 0.5s, the islanding detection scheme proposed in this chapter, which uses a stacking structure to fuse multiple models, can quickly provide an islanding signal. Moreover, it has a relatively certain output probability value both during non-islanded grid-connected operation and after entering the islanding operation state. The system can obtain a relatively certain islanding decision result that continuously conforms to the actual operating conditions.

[0081] Figure 4 The diagram shows an intelligent islanding detection method based on the LightGBM algorithm. Although the grid-connected system can quickly receive the islanding signal after entering islanding operation, the low confidence level and low model output probability value in the short period after islanding occurs lead to some instances being output as non-islanding. For situations requiring continuous output of status signals, there is still a possibility of misjudgment, making it unsuitable for applications requiring continuous output of grid-connected system status determination results. By comparing the output probability diagrams of the two methods, it can be seen that the method proposed in this chapter can produce more stable and confident output values, resulting in a higher confidence level for the final output status determination result.

[0082] Example 2 is an islanding detection device for a photovoltaic grid-connected system based on a stacking structure, which is used to operate the method described in Example 1, including:

[0083] The data acquisition unit is used to construct a simulation model of a four-unit parallel photovoltaic grid-connected system and to collect training samples of electrical characteristic quantities such as voltage and frequency at the grid connection point.

[0084] The training unit is used to take the training samples as input vectors and perform k-fold cross-validation training on the four base models of the first layer of Stacking to obtain the island detection models built based on the training samples.

[0085] The integration unit is used to integrate the output vectors of the four base models, combine the output vectors to construct the training set of the next round meta-model, obtain the training and validation set of the second-layer meta-model, train and predict the meta-model, and obtain the final island detection model.

[0086] Embodiment 3 is a computer-readable storage medium storing computer-executable instructions for causing a computer to execute the island detection method of Embodiment 1.

Claims

1. A method for islanding detection in a photovoltaic grid-connected system based on a stacking structure, characterized in that, Includes the following steps: Step 1: Construct a simulation model of a four-unit parallel photovoltaic grid-connected system and collect voltage and frequency training samples at the grid connection point; Step 2: Select the island detection models built based on XGBoost, Catboost, RF and LightGBM algorithms as the four heterogeneous base learners of the four base models in the first layer. Use the training samples as input vectors to perform k-fold cross-validation training on the four base models in the first layer of Stacking to obtain the island detection models built based on the training samples. In step 2, for the base model M4 of the first layer constructed based on the LightGBM algorithm, the CART decision tree is used as the basic classifier. By combining multiple weak models with the forward stepwise algorithm, a final model composed of a set of regression trees is obtained. The negative gradient of the loss function is used to replace the residuals to fit an approximation of the loss in this round, and then a CART decision tree is fitted; The first round Loss function for each sample The negative gradient is: in, For the first The true value of each sample; Indicates sample The corresponding feature vector; For the first The model prediction value for each sample; For the first Trees for input feature vectors The model's predicted values; use , Fit a CART regression tree to obtain the first... A regression tree, and its corresponding leaf node region , ,in The number of leaf nodes; Using a greedy approach, considering only local optimization, for each leaf node, the output value is... Find the output value that minimizes the loss function, i.e., the value that best fits the leaf nodes: in, For the m-th leaf node The best output value of the regression tree; The decision tree fitting function for this round is obtained as follows: in, For the first The training set of all nodes assigned to leaf nodes in each round is used to obtain the strong classifier for this round through iterative updates: pass The final strong classifier is obtained by iteratively combining the base models: in, This is the initial model; for The final strong classification model obtained by iteratively combining the base models is equivalent to the final M4 island detection base model. ; Step 3: Integrate the output vectors of the four base models and select a Bayesian classifier as the meta-model. The output vectors are combined to form the training set for the next round of meta-model, thus obtaining the training and validation set for the second-layer meta-model. The meta-model is then trained and used for prediction to obtain the final island detection model.

2. The island detection method as described in claim 1, characterized in that, The simulation model for constructing a four-unit parallel photovoltaic grid-connected system in step 1 includes establishing a multi-unit parallel photovoltaic power generation grid-connected system using MATLAB / SIMULINK. The photovoltaic power generation grid-connected system consists of four adjustable photovoltaic arrays.

3. The island detection method as described in claim 2, characterized in that, The simulation model of the four-unit parallel photovoltaic grid-connected system in step 1 also includes collecting the operating data of the photovoltaic grid-connected system to construct a data set S, using k-fold cross-validation to divide the training set and the validation set, so as to perform data training and validation of each base model, and obtain the output results of each model.

4. The island detection method as described in claim 2, characterized in that, It also includes the following non-islanded operating states: grid connection fluctuation operating state, voltage sudden change operating state, power system fault operating state, and harmonic interference operating state.

5. The island detection method as described in claim 4, characterized in that, To address the aforementioned grid-connected fluctuating operating state, the following steps are required: control the number of grid-connected photovoltaic array units to vary within the range of 1 to 4, and control the photovoltaic output power fluctuation P. load Q load The value varies between 0 and 2 times the rated value, with a sampling step size of 0.001, and a total of 50 simulations are performed.

6. The island detection method as described in claim 4, characterized in that, The following steps are required to address the aforementioned voltage fluctuation operation state: The number of grid-connected photovoltaic array units was controlled to vary from 1 to 4. Under rated load, the rated voltage of the system was suddenly increased by 10% to 30%, with a sampling step size of 0.

001. Under rated load, a large capacitor load was switched in, and the rated voltage suddenly dropped by 15% to 30%, with a sampling step size of 0.

001. A total of 50 simulations were performed.

7. The island detection method as described in claim 4, characterized in that, The following steps are required to address the aforementioned power system fault operation state: The number of grid-connected photovoltaic array units was controlled to vary from 1 to 4. Under rated load, the model simulated single-phase ground fault, two-phase fault, and two-phase ground fault conditions with a sampling step size of 0.001, for a total of 50 simulations.

8. The island detection method as described in claim 4, characterized in that, The following steps are required to address the aforementioned harmonic interference during operation: The number of grid-connected photovoltaic array units is controlled to vary within the range of 1 to 4, and the ratio P of the active power absorbed by the load to the active power generated by the photovoltaic system is controlled. load / P PV Within the range of 0.8 to 1.5, the 5th, 7th, and 11th harmonic disturbances were injected into the three phases respectively, with a sampling step size of 0.001, for a total of 50 simulations.

9. The island detection method as described in claim 2, characterized in that, The simulation model also includes data acquisition of the island's operational status: The number of grid-connected photovoltaic array units is varied within the range of 1 to 4, and the ratio P of the active power absorbed by the load to the active power generated by the photovoltaic system is... load / P PV The sampling step size is 0.001 when the load is purely resistive, with a range of 0.8 to 1.5, and a total of 100 simulations are performed. When the load contains inductive components, the load quality factor is between 0.9 and 1.3, with a sampling step size of 0.001, and a total of 100 simulations are performed.

10. The island detection method as described in claim 1, characterized in that, For the base model M1 of the first layer built based on the XGBoost algorithm, the optimal model is optimized by minimizing the loss function through forward iteration, where the loss function is the objective function. Firstly, regarding the first... The first tree For each sample, the model's predicted value is: in, For the first After round of iterations, the samples The predicted value; For the first tree pair The predicted value; For the first The predicted value of each tree; In the forward iteration of the XGBoost algorithm, more attention should be paid to the fitting of the t-th tree in the current round. Therefore, the previous iteration should be... Treating the trees as constants for the objective function The objective function is simplified by performing a second-order expansion using the Taylor formula, resulting in the final objective function: in, leaf node The sum of the first derivatives of the included samples; It is the sum of the second derivatives.

11. The island detection method as described in claim 1, characterized in that, For the first-layer base model M2 constructed based on the CatBoost algorithm, an island detection base model M2 is constructed using a fully symmetric decision tree structure. In the first stage, unbiased gradient estimation is adopted, while in the second stage, the traditional GBDT scheme is continued. Let... To construct the first The model after the first tree, the first one constructed The tree after the first The gradient value of each training sample is For each sample The base model M2 uses a single model consisting of multiple decision trees. Furthermore, this model does not use samples. Instead of iteratively updating the gradient estimate, we use... estimate The gradient is calculated, and the estimated value is used to score the tree, where the loss function is... , For the first The true value of each sample The value is calculated using the formula.

12. The island detection method as described in claim 1, characterized in that, For the base model M3 of the first layer constructed based on the RF algorithm, the training set constructed by obtaining multiple feature quantities at the PCC point of the multi-machine parallel photovoltaic grid-connected system is expanded using the RF algorithm; the Bootstrap method is used to randomly extract multiple training sample subsets from the training sample set. For each subset, a classification model is trained and constructed separately. The decision results of multiple weak classification models are combined, and the final island detection model is obtained through voting.

13. The island detection method as described in claim 1, characterized in that, Step 3 includes using a stacking strategy to fuse two-layer models and construct the final island detection model.

14. The island detection method as described in claim 1, characterized in that, The specific steps for constructing the second layer meta-model based on a Bayesian classifier within the Stacking structure are as follows: For island detection tasks, there are only two classification cases: "0" and "1", and the category is given. Its corresponding features arrive According to Bayes' theorem, we can obtain: in, Let y be the probability of category y across all categories; In order to the corresponding features arrive The posterior probability of class y appearing in the data; Input features for sample data The total probability; for Features appearing in the class arrive Class conditional probability; For Naive Bayes, it is necessary to assume that each feature is independent of the others, and all features... If all of these conditions are met, the formula can be further simplified to: in, Each is an independent feature Distinguish as The probability of the category; Therefore, for each input sample data The final classification result can then be obtained as follows: in, This represents the classification result of the final model.

15. The island detection method as described in claim 1, characterized in that, The value of k is 5.

16. An islanding detection device for a photovoltaic grid-connected system based on a stacking structure, used to operate the method as described in any one of claims 1-15, characterized in that, include: The data acquisition unit is used to construct a simulation model of a four-unit parallel photovoltaic grid-connected system and to collect training samples of electrical characteristic quantities such as voltage and frequency at the grid connection point. The training unit is used to take the training samples as input vectors and perform k-fold cross-validation training on the four base models of the first layer of Stacking to obtain the island detection models built based on the training samples. The integration unit is used to integrate the output vectors of the four base models, combine the output vectors to construct the training set of the next round meta-model, obtain the training and validation set of the second-layer meta-model, train and predict the meta-model, and obtain the final island detection model.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the island detection method as described in any one of claims 1-15.

Images