TWI404960B - Method for islanding phenomenon detection of photovoltaic power generating systems - Google Patents

Abstract

A method for islanding phenomenon detection of photovoltaic (PV) power generating systems is disclosed. The method is to construct an islanding detection matter-element model of the PV power generating system bases on extension theory. And an extension neural network is constructed to cooperate the islanding detection matter-element model. The extension neural network is inputted training samples to train the extension neural network. The extension neural network is employed to detect various islanding detection categories of PV power generating system after the training is accomplished.

Description

Method for detecting island phenomenon of solar photovoltaic power generation system

The invention relates to a power generation system detection method, and in particular to an extension neural network (ENN) architecture using a multi-variable detection method combined with a passive and active islanding effect detection. The method of detecting the island phenomenon is detected.

When the solar photovoltaic power generation system is connected in parallel with the mains power system (Taiwan power supply network), when the utility power fails and the solar power generation system cannot detect and cut off immediately, and the phenomenon of separate power supply is presented, it is called islanding operation (Islanding). Operation).

Once the island operation occurs, the protection device should detect and stop the operation of the island operation as soon as possible, otherwise it may adversely affect the power supply system or the power users.

The conventional island operation detection method can be divided into two categories: passive and active detection.

Passive detection technology is used to monitor the state of the load terminal, such as voltage, frequency and phase, as the basis for judging whether or not island operation occurs. Passive detection technology mainly includes voltage phase jitter detection method, frequency change rate detection method and voltage third harmonic distortion sharp increase detection method.

However, when the output of the solar photovoltaic system is close to the load consumption, the fluctuation of the voltage or frequency of the system is not enough to be detected by the power, which is called the non-detection zone of the power. The size of the area is not affected by the sensitivity of the eMule. If the sensitivity of the power is low, the area that does not feel the belt is large. If the sensitivity of the power is increased, the area that does not feel the belt can be reduced, but it is easy to have a malfunction. Under the condition of considering the stability of the electric raft, it is not appropriate to set the sensitivity of the electric shovel too high, so it is necessary to use the active detection method in combination to compensate for the lack of passive mode.

The active mode is to actively apply periodic disturbance to the voltage or frequency of the system in the parallel operation mode. Since the mains is a stable reference power supply, under normal circumstances, the active mode disturbance will not affect the system voltage or frequency too much; but when the island operation occurs, the system loses the stable reference power. The disturbance of the active mode will cause instability of the system. Even in the case of balance between power generation output and load consumption, the disturbance will destroy the power balance state, causing the voltage or frequency of the system to change significantly. At this time, the system can be detected according to the change situation. An island phenomenon has occurred.

The implementation of the active mode can achieve voltage or frequency disturbance through the control of a Power Conditioner or by using an external impedance, etc., and mainly includes an active voltage drift method, an active frequency drift method, a slip frequency shift method, and Methods such as load variation method. However, since the power system may be affected by external forces and cause power interference, the active disturbance will cause the island detection to malfunction.

The object of the present invention is to provide an islanding phenomenon detecting method for a solar photovoltaic power generation system, which is used for improving the accuracy of detecting an islanding operation of a grid-connected (mains parallel type) solar photovoltaic power generation system, improving the defects of the prior art, and reducing the non-inductive zone. (Nondetection zone) area.

Based on the above purposes, this paper proposes an islanding phenomenon detection method for solar photovoltaic power generation system. It uses the extension-based neural combination passive and active multi-variable method to detect island operation, which can quickly generate solar photovoltaic power generation during the city electrolysis. The system cuts off the load, and it can correctly distinguish the fault signal generated by the mains terminal as power quality interference or real island operation.

According to a preferred embodiment of the present invention, a method for detecting an islanding phenomenon of a solar photovoltaic power generation system selects an appropriate feature value of the islanding phenomenon detection, and transmits the solar photovoltaic system through an extension theory and an islanding characteristic parameter of the solar photovoltaic power generation system. The islanding phenomenon of the power generation system is expressed as an island detection extension element model, and the extensional nerves are used to establish the material element models corresponding to different categories, and the weights and weights of each feature are calculated and adjusted. Calculate the distance from each cluster to determine whether the system has been islanded.

Furthermore, in a preferred embodiment of the present invention, an island-like phenomenon of a solar photovoltaic power generation system is detected by an extension-type neural network, and an extensional neurological theory system is combined with a passive and active islanding phenomenon. The multi-variable detection method analyzes the power quality disturbances such as voltage surge, voltage dip, power harmonic and voltage flicker at the mains terminal to identify the abnormality of the mains terminal as power quality interference or real island operation.

The method for detecting an island phenomenon according to a preferred embodiment of the present invention has the advantages of less learning data, simple learning procedure, fast learning speed, high recognition rate, and less memory to quickly detect the occurrence of an island phenomenon. .

In order to enable the reviewing committee to easily understand the technical features of the present invention and perform simulation to verify its efficacy, in this embodiment, a system architecture in which a solar photovoltaic power generation system is connected in parallel with a commercial power supply is provided, and an islanding phenomenon detecting system is installed as an example. Description.

Please refer to FIG. 1 , which is a system architecture diagram of a solar photovoltaic power generation system with an islanding detection system in parallel with the commercial power in the embodiment. The system 100 in which the solar photovoltaic power generation system is connected in parallel with the commercial power includes a power conditioner 101, an LC filter 102, an islanding effect detection controller 104, an RLC parallel load 105, and a commercial power terminal 106. In this embodiment, V _dc is the DC voltage obtained by the output of the solar photovoltaic module array via the extension maximum power tracking controller and the DC-to-DC converter.

The peak voltage of the power conditioner 101, the frequency of the power conditioner 101, and the phase difference between the voltage and current of the power conditioner 101 can be calculated via the voltage (V) and current (1) feedback signals of the responsibility demarcation point 103, and can be calculated according to the calculation. The resulting signal is used as an indicator of islanding.

The main purpose of the power conditioner 101 of the general solar photovoltaic power generation system is to convert the direct current power into an alternating current type to supply power in parallel with the mains 106 and detect island operation. Before the power conditioner 101 is connected in parallel with the mains 106, the synchronization problem needs to be considered, that is, the synchronous parallel conditions of the IEEE Std. 1547 and the general telecommunication equipment must be satisfied as follows:

(1) the voltage of the power conditioner v _{inverter range: 279V ≦ v inverter, max ≦} 342V.

(2) The voltage frequency of the power regulator f _inverter : 59.3 Hz ≦ f _inverter ≦ 60.5 Hz.

(3) The phase difference between the voltage v _{inverter of the} power conditioner and the mains voltage v _grid : ±5 degrees.

In order to make the power conditioner 101 of the solar photovoltaic power generation system closer to the actual operation, the output of the power conditioner 101 is adjusted to be synchronized with the commercial power 106 in parallel before the island operation detection. The power regulator continuously tracks the voltage and frequency of the mains 106 and the phase difference between the power conditioner 101 and the mains 106. When the control system detects the output voltage and frequency of the converter and the phase difference between the power regulator and the mains meets the above three The condition is that the power conditioner 101 is synchronized with the commercial power 106 at this time, so that the parallel switch of the power conditioner 101 is turned from OFF to ON, and starts to operate in parallel with the commercial power 106.

Please refer to FIG. 2, which is a flow chart of a method for detecting an island phenomenon of a solar photovoltaic power generation system according to a preferred embodiment of the present invention. The method for detecting the islanding phenomenon of the solar photovoltaic power generation system comprises the following steps: constructing an extension matter element model of the islanding phenomenon of the solar photovoltaic power generation system through the extension theory, wherein the extension matter element model of the islanding phenomenon comprises a plurality of sub-elemental element models, Each of the sub-element models corresponds to an island phenomenon detection category, as shown in step 210. An extension-like neural network is established to cooperate with the extrinsic matter model of the islanding phenomenon, as shown in step 220. The learning data of the known island phenomenon detection category is input, and the training procedure of the extension type neural network is performed until the set identification rate (for example, 0.001%) is met, as shown in step 230. Enter the data to be tested, and use the extended neural network architecture that has been learned to detect the islanding phenomenon, and determine the island detection category to which the data to be tested belongs, as shown in step 240.

In the embodiment of the present invention, the method for detecting the islanding phenomenon of the solar photovoltaic power generation system is further detailed as follows:

(1) Establishing an extension matter element model of the islanding phenomenon of the solar photovoltaic power generation system (step 210)

Before establishing the extension matter model for detecting island phenomena, it is necessary to find out the characteristics of island operation detection. In the present embodiment, the following three items are selected as characteristics when the system is operated in an island:

1. Peak voltage: The maximum voltage value output by the power regulator.

2. Frequency: The frequency of the output voltage of the power regulator.

3. Phase difference: the phase difference between the output voltage and current of the power regulator.

Therefore, in the present embodiment, the voltage peak, the frequency, and the phase difference are used as the features of the matter-element model, so that the extension matter element model for detecting the island phenomenon is defined as Equation (1).

Where c ₁ and c ₂ represent the peak voltage and frequency of the power regulator output, respectively, and c ₃ represents the phase difference between the output voltage and current of the power regulator, and V _k1 ~ V _k3 are respectively the range values of the three characteristics ( The classic domain).

Because the power system may be affected by lightning strikes, salt damage, foreign objects touching or drastic load changes, the system may cause voltage surge, voltage dip, voltage flicker or power harmonics, which may cause the island detection system to malfunction. In a preferred embodiment of the present invention, an extensional neural theory is applied to detect island operations in different power quality environments to identify a solar photovoltaic system for power quality disturbance or an island operation state.

The island detection categories of the solar photovoltaic system are divided into seven categories (that is, the island detection object model is divided into seven sub-element models). The seven types of island detection categories and their symbols are respectively described as follows:

I ₁ : voltage swell

I ₂ : voltage dips

I ₃ : Injection of power harmonics

I ₄ : Normal operation

I ₅ : island operation above normal operating range

I ₆ : island operation below normal operating range

I ₇ : Voltage flashing

Among them, I ₁ : voltage swell, I ₂ : voltage dip, I ₃ : injected power harmonics and I ₇ : voltage flicker is a factor of power quality, and the following will affect the power quality in the examples. The factors are explained.

1. Voltage swell: According to IEEE Std.1159-1995, the rms voltage of the mains voltage is between 1.1p.u.~1.8p.u., and the duration is more than 0.5 cycles.

2. Voltage dip: According to the definition in IEEE Std. 1159-1995, the voltage dip refers to the effective value of the mains voltage between 0.1 p.u. and 0.9 p.u., and the duration is 0.5 to 1 minute.

3. Power Harmonics: Voltage Harmonics refers to the 3, 5 or 7th harmonic components in the mains system, and the harmonic components defined in this embodiment are 10%, 7% and 5 respectively of the fundamental frequency. %.

4. Voltage flicker: The voltage flicker in this embodiment combines a low frequency voltage source (15 Hz and 20 Hz) with a standard voltage of 60 Hz.

(2) Establish an extension class neural network (step 220)

Please refer to FIG. 3, which is a block diagram of an extension-like neural network according to a preferred embodiment of the present invention. The extension-like neural network 300 includes an input layer 310 and an output layer 320. The input layer 310 includes a plurality of input layer nodes 311 for classifying and reading input data into the extension type neural network, that is, the magnitude of the matter element features ( The connection between the input layer 310 and the output layer 320 is a weight value 330 ( W _kj ), including an upper limit of the weight value, a center of weight, and a lower limit of the weight value. Finally, the minimum value of the extension distance (ED) value of the output layer 320 belonging to each category is found, and the category of the data is determined.

The number of input layer nodes is the feature number (1~n) of each sub-element model. In this embodiment, the number of input layer nodes is 3 (that is, each sub-element model adopts three features). The number of output layer nodes is the total number of child element models (1~n _c ). In this embodiment, the number of output layer nodes is 7 (that is, seven types of island detection categories).

In this embodiment, the initial setting value of the weight value 330 is the maximum value and the minimum value of the classical domain of the input sample feature element of the known island detecting category. E.g: versus Represents the minimum and maximum weight values connected to the jth input layer node and the kth output layer node, respectively.

(3) Supervised learning training program of extension type neural network (step 230)

The learning rule of the extension-like neural network is supervised learning. The so-called supervised learning first inputs the feature samples. If the feature samples do not match the set target values, the weight values are modified. Since the extension-type neural network belongs to supervised learning, by adjusting the weight value, the accuracy of the identification system can be effectively increased, and the error rate can be reduced.

Before you can learn the program, you must define the relevant variables first.

Set the learning sample to , where N _p is the total number of learning samples. Each sample contains the characteristics and categories of the data, and the ith sample is expressed as Where i=1, 2, 3..., N _p , n represents the total number of features of the sample, and p is the island detection category to which the ith sample belongs. To assess the accuracy of extension identifying the neural network, set N _m is the total number of the estimated overall error, and the total error ratio (E _T) can be represented as follows:

Please refer to FIG. 4, which illustrates a flow chart of performing extension-like neural network training according to FIG. 2. The supervised learning training program of the extension type neural network includes at least the following steps:

Step 421: setting an initial weight value connected between the input layer node and the output layer node, and the kth extension object model can be expressed as follows:

In equation (3), c _{j =1, 2, 3} is the jth feature of N _k , It is the classical domain value for the feature c _j , and the range of the classical domain can be determined by the training sample. V _{k 1} is the total number of clusters at the output, where the classical domain of the extensional neural theory is

among them, The weight minimum corresponding to the kth classification for the jth feature, The maximum value of the weight corresponding to the kth category for the jth feature, Learning materials for the input of the extension-like neural network.

Step 422: Calculate an initial cluster center Z _{kj of} each initial weight value.

Z _k ={ z _{k 1} , z _{k 2} ,..., z _kn } (6)

Where j =1,2,..., n ; k =1,2,..., n _c (8)

Step 423: Read the i- th learning sample of the p category

And n _c is the total number of categories.

Step 424: Calculate the learning sample by applying the extension distance (ED) The distance from the kth cluster, that is, the calculated extension distance between the training sample and each cluster, the mathematical expression is as follows:

among them, For the ith stroke learning sample, p denotes its category, and p is characterized by j : z _kj represents the center of the weight value of the jth input endpoint and the kth output endpoint.

Please refer to FIG. 5, which is a schematic diagram of the extension distance (ED) proposed in the embodiment. The extension distance (ED) can be used to represent the distance between the point x and the interval < w ^L , w ^U 〉, where z is the weight center of the cluster, w ^L is the weight cluster minimum, and w ^U is the weight cluster maximum. It can be seen from Fig. 5 that the extension distance can be calculated as a difference in distance calculation due to different numerical ranges, and thus a difference in sensitivity value is generated. In general, if the range of matter element features is large, it means that the training data is more widely blurred, so the sensitivity in distance calculation is lower; on the contrary, if the range of matter element features is small, it represents the place. The data sample is more accurate, so it can show high sensitivity in distance calculation.

Step 425: After the extension distance operation, the new cluster of learning materials is obtained as k ^* , and is available If k ^* = p at this time (that is, the input feature sample p is the same as the island detection category k ^* to which it belongs), the calculation program jumps to step 427; otherwise, step 426 is continued.

Step 426: Adjust and update the weight values of the p - th and k ^* - th clusters as follows: (a) update the weight center values of the p - th and k ^* - th clusters, that is, update the classification cluster corresponding to the input feature sample misjudgment ( The weight center value of the p - th cluster and the updated input feature sample itself should belong to the weight center value of the classification cluster ( k ^* - th cluster).

(b) updating the weight values of the p - th and k ^* - th clusters, that is, updating the weight values of the classification clusters ( p - th clusters) corresponding to the misrepresentation of the input feature samples, and updating the input feature samples themselves to belong to the classification cluster ( k ^* - th cluster) weight value.

η: learning rate of the extension-like neural network (Learning rate)

: the new weight center value after the learning feature j and the category number is p

: The old weight center value before the learning is the feature j and the category number is p

: the new weight center value after the learning feature j and the cluster number k ^*

: the old weight center value of the feature j before the learning and the cluster number k ^*

: Feature j and the category number is the lower limit of the weight of p

: Feature j and the category number is the weight of p

: feature j and the old lower limit of the weight of the category number p

: feature j and the old limit of the weight of the category number p

: Feature j and the cluster number is the lower limit of the weight of k ^*

: Feature j and the cluster number is k ^*

: feature j and the old lower limit of the weight of the cluster number k ^*

: feature j and the cluster number is the upper limit of the weight of k ^*

In this step, only the weights of the clusters p and k ^* are adjusted during the learning process, while the other weight values are not changed. Due to the above characteristics, the extension-type neural network has the advantage of faster calculation speed than other neural network architectures, and has higher adaptability in new application fields.

Please refer to FIG. 6a and FIG. 6b, which are diagrams showing the results of adjusting the two cluster weights in the learning program. The learning sample x _{ij in} Fig. 6a should belong to cluster B, but is classified as cluster A because the calculation result of the extension distance formula is ED _A < ED _B . And in Figure 6b, the new extension distance is obtained after adjusting via the weight value. ( To adjust the extension distance between the test data and cluster A, In order to adjust the extension distance of the test data to cluster B), the cluster to which the learning sample x _ij belongs is classified by cluster A (Z _A ) into cluster B (Z' _B ).

Step 427: The calculation procedure of steps 423 to 426 is repeated until all the training samples have been classified, and a learning batch (epoch) is ended.

Step 428: If the state of convergence has been reached categorizer, or total error rate (E _T) has reached the preset target value calculation program is stopped, otherwise, it returns to step 423.

In this embodiment, the 360 data obtained are divided into 235 learning materials and 125 identification data, of which 235 learning materials are calculated by the above-mentioned extensional neural learning process, and the island detection of the first table is obtained. The matter-element model, the parameters of the extension-type neural network learning process are as follows:

1. Learning rate: 0.001

2. Number of studies: 200 times

3. Number of features: 3

4. Number of categories: 7

5. Error tolerance rate: 0 (that means learning must reach 100%)

In this embodiment, the phase difference between the peak voltage and frequency of the power regulator and the voltage and current of the power regulator is used as the characteristic value of the island detection. After completing the learning through the extension-type neural network learning process, the weight upper and lower limits and weight centers of the obtained feature values are as shown in Table 2.

(4) The island state detection calculation process of the extension type neural network (step 240)

Please refer to FIG. 7 , which illustrates a flow chart of the algorithm for detecting the extremity-like neuro Island phenomenon according to FIG. 2 . When the extensional neural network completes the learning process, it can be identified or classified, and its calculation program includes:

Step 741: Read the weight value of the extension type neural network island detection after the completion of the learning, and use the weight value of the trained extension type neural network as the weight value of the extension type neural network for identification.

Step 742: Calculate the cluster weight center value of each classification cluster according to the read weight value matrix by using equation (7).

Step 743: Read the test samples for which the island state detection is to be performed. In this embodiment, there are a total of 125 test samples. The test sample is obtained by detecting the voltage and current output from the power regulator, and calculating the voltage peak, the frequency, and the phase difference between the voltage and the current, and then adding the active voltage drift method. Express it as

X _t ={ x _{t 1} , x _{t 2} ,..., x _tn } (15)

Step 744: Calculate the distance between each test sample and the classified cluster after learning by applying the definition of the extension distance (ED) proposed by (10).

Step 745: Find the minimum extension distance and find the classification cluster ( k ^* ) to which the test sample belongs. And set its corresponding output node To determine the attribution category of the test sample.

Step 746: It is judged whether all the samples are detected, and if the recognition is completed, the operation program is stopped, otherwise the process returns to step 743.

In order to prove the feasibility of an islanding phenomenon detection method for a solar photovoltaic power generation system according to a preferred embodiment of the present invention, a Kyocera KC40T solar photovoltaic module is used to form a 516W solar photovoltaic power generation system for island detection simulation. . Please refer to Figure 8 for the island operation detection simulation diagram when the 98Ω resistive load is used for 4 seconds. The waveforms of the converter output voltage 801 and the load voltage 802 are respectively displayed from top to bottom, followed by the waveform of the frequency 803 of the output voltage of the converter, and the phase difference between the output voltage and the current of the converter is 804. The waveform diagram is finally a schematic diagram of the city electrolysis signal 805 and the load excision signal 806. After the output of the power conditioner is connected in parallel with the mains, the system is connected to the city at 4 seconds. The island detection system detects the island operation (about 8.3 milliseconds) in 0.5 cycles and cuts off the load.

Please refer to Figure 9 for the simulation of the island operation detection in the case of a sudden voltage rise. The waveforms of the voltage swell 901, the converter output voltage 902 and the load voltage 903 are respectively displayed from top to bottom, and finally the schematic diagram of the city electrolysis signal 904 and the load disconnection signal 905. When the system first occurs after 3 cycles of voltage swell, the island terminal runs again after about 4 seconds. It can be seen from the figure that the islanding phenomenon detection method of the solar photovoltaic power generation system in this embodiment can discriminate the voltage swell. The situation is signal interference, not island operation, so the load is not cut off, but the load is cut off after 0.5 cycles after the island operation occurs.

Please refer to Figure 10 for the simulation of the island operation detection in the case of a voltage dip. The waveform diagrams of the voltage dip 1001, the converter output voltage 1002 and the load voltage 1003 are respectively displayed from top to bottom, and finally the schematic diagram of the municipal electrolysis signal 1004 and the load disconnection signal 1005. In the case of a voltage dip in the system, the island terminal operates again in about 4 seconds. It can also be seen from the figure. The islanding phenomenon detection method of the solar photovoltaic power generation system in this embodiment can also clearly determine the voltage dip. It is a power interference, so the load is cut off within 0.5 cycles after the actual island operation occurs.

Figure 11 is a simulation diagram of island operation detection in the case of power harmonics. The waveforms of the added power harmonic 1101, the converter output voltage 1102 and the load voltage 1103 are respectively displayed from top to bottom, and finally the schematic diagram of the city electrolysis signal 1104 and the load disconnection signal 1105. Simulate the occurrence of harmonic component interference before the island operation, to observe whether it affects the islanding detection. The added harmonic components include the 3rd, 5th and 7th harmonics, and the components of each harmonic are 10% of the fundamental frequency. , 7% and 5%. As shown in the figure, the islanding operation occurs only about 4 seconds, and the islanding phenomenon detecting method of the solar photovoltaic power generation system in this embodiment is not misjudged due to harmonic interference of the system before the islanding operation.

Figure 12 is a simulation diagram of island operation detection in the case of voltage flicker. The waveforms of the input voltage flicker 1201, the converter output voltage 1202 and the load voltage 1203 are respectively displayed from top to bottom, and finally the schematic diagram of the city electrolysis signal 1204 and the load excision signal 1205. The island operation detection under voltage flicker is synthesized with a standard voltage of 60 Hz at 15 Hz and 20 Hz, and an island operation occurs in about 4 seconds, and the controller cuts off the load in 0.5 cycles.

The above simulation results show that the islanding phenomenon detecting method of the solar photovoltaic power generation system in a preferred embodiment of the present invention can correctly detect the islanding phenomenon and quickly cut off the load from the solar photovoltaic power generation system within a standardized time.

Although the present invention has been described above in terms of a preferred embodiment, it is not intended to limit the invention, and it is obvious to those skilled in the art that various changes and modifications can be made without departing from the spirit and scope of the invention. The scope of the invention is defined by the scope of the appended claims.

100. . . system

101. . . Power conditioner

102. . . LC filter

103. . . Demarcation point of responsibility

104. . . Island phenomenon detection controller

105. . . RLC parallel load

106. . . Mains terminal

210~240. . . step

300. . . Extensional neural network

310. . . Input layer

311. . . Input layer node

320. . . Output layer

330. . . Weights

421~428. . . step

741~746. . . step

801. . . Converter output voltage

802. . . Load voltage

803. . . Converter output voltage frequency

804. . . Phase difference between voltage and current

805. . . City Electrolysis Signal

806. . . Load cutoff signal

901. . . Voltage swell

902. . . Converter output voltage

903. . . Load voltage

904. . . City Electrolysis Signal

905. . . Load cutoff signal

1001. . . Voltage dip

1002. . . Converter output voltage

1003. . . Load voltage

1004. . . City Electrolysis Signal

1005. . . Load cutoff signal

1101. . . Power harmonic

1102. . . Converter output voltage

1103. . . Load voltage

1104. . . City Electrolysis Signal

1105. . . Load cutoff signal

1201. . . Voltage flashing

1202. . . Converter output voltage

1203. . . Load voltage

1204. . . City Electrolysis Signal

1205. . . Load cutoff signal

The above and other objects, features, advantages and embodiments of the present invention will become more apparent and understood.

FIG. 1 is a system architecture diagram of a solar photovoltaic power generation system having a phenomenon detection system in parallel with a commercial power supply in the embodiment.

2 is a flow chart showing a method for detecting an island phenomenon of a solar photovoltaic power generation system according to a preferred embodiment of the present invention.

FIG. 3 is a block diagram showing an extension type neural network according to a preferred embodiment of the present invention.

Figure 4 is a flow chart showing the extension of the neural network training in accordance with Figure 2.

Figure 5 is a schematic diagram showing the extension distance (ED) proposed in the present embodiment.

Figures 6a and 6b are diagrams showing the results of adjustments of the two cluster weights in the learning program.

Fig. 7 is a flow chart showing the calculation of the extensional nerve island detection according to Fig. 2.

Fig. 8 is a simulation diagram showing the island operation detection when a 98 Ω resistive load is generated in 4 seconds.

Figure 9 is a simulation diagram of the islanding operation detection in the case of a sudden voltage rise.

Figure 10 is a simulation diagram showing the island operation detection in the event of a voltage dip.

Figure 11 is a simulation diagram showing the island operation detection in the case of power harmonics.

Figure 12 is a simulation diagram showing the island operation detection in the case of voltage flicker.

210~240. . . step

Claims (9)

An islanding phenomenon detection method for a solar photovoltaic power generation system is used for diagnosing a commercial parallel solar photovoltaic power generation system, comprising: (a) establishing an islanding phenomenon extension matter element model, which is based on a matter element defined in the extension theory The model and a plurality of features of the island operation detection of the solar photovoltaic power generation system, wherein the island phenomenon extension material element model is used to indicate an island operation state of the solar photovoltaic power generation system, wherein the island photovoltaic operation detection of the solar photovoltaic power generation system The characteristic includes: a peak voltage, which is a maximum voltage value of a power regulator output; a frequency, which is a frequency of the power regulator output voltage; and a phase difference, which is the power regulator output voltage a phase difference from the current; (b) establishing an extension-like neural network that cooperates with the islanding extinction matter element model, wherein the extension-type neural network includes an input layer and an output layer, the input layer and the The connection between the output layers is a weight value; (c) training the extension-type neural network, inputting a learning data of a known island detection category The neural network is trained; and (d) using the trained extension-type neural network to determine whether the islanding phenomenon of the solar photovoltaic system occurs; wherein the island detection category of the solar photovoltaic system includes a voltage swell, Voltage dips, injected power harmonics, normal operation, island operation above normal operating range, island operation below normal operating range, and voltage flicker. For example, the method for detecting an islanding phenomenon of a solar photovoltaic power generation system according to claim 1, wherein the islanding phenomenon extension element model is Where c ₁ and c ₂ represent the peak voltage and frequency of the power regulator output, respectively, and c ₃ represents the phase difference between the output voltage and current of the power regulator, and V _k1 ~ V _k3 are respectively c ₁ , c ₂ and The range of characteristic values of c ₃ . The method for detecting an island phenomenon of a solar photovoltaic power generation system according to claim 2, wherein the weight value includes an upper limit of the weight value, a weight center, and a lower limit of the weight value. The method for detecting an island phenomenon of a solar photovoltaic power generation system according to claim 3, wherein the step (c) comprises: (c1) setting an initial weight value connected between the input layer node and the output layer node; (c2) Calculating a weighted center value of each weight value, the weighted center value is an average value of two weight values connecting the input layer and the output layer; (c3) inputting a learning sample and a category number thereof, wherein the learning sample is the a type of the extension metamodel; (c4) calculating a reachable distance between the learning sample and each of the clusters; (c5) finding the minimum value of the extension distance, so that the output value of the output layer node corresponding to the classification cluster is 1 , whereby the island detection category of the solar photovoltaic system is an island corresponding to the output layer node. a detection category; (c6) when the island detection category represented by the output layer node is different from the island detection category to which the learning sample belongs, updating the weight value and the weight center value; and (c7) calculating one The total error rate, which is the ratio of the number of errors to the total number of learning samples. When the total error rate is less than a predetermined value, the training ends; otherwise, the process returns to step (c3). The method for detecting an islanding phenomenon of a solar photovoltaic power generation system according to claim 4, wherein the mathematical expression of the extension distance is k =1,2,..., n _c ; where For the i-th study sample, p denotes its category, and p is characterized by j : Z _kj represents the center of the weight value of the jth input endpoint and the kth output endpoint. The method for detecting an island phenomenon of a solar photovoltaic power generation system according to claim 5, wherein in the step (c6), the method for updating the weight center value is only for the weight center value of the classification cluster to which the learning sample belongs and The weight center value of the error output classification cluster is adjusted, wherein the adjustment of the weight center value of the classification cluster to which the learning sample belongs may be expressed as follows: among them, : the weighted central value of the classification cluster to which the learning sample belongs after an update, : is the weight center value of the classification cluster to which the learning sample belongs before updating, and η: is a set learning rate, : for the learning sample; the adjustment of the weight center value of the error output classification cluster can be expressed as follows: among them, : The value of the weighted center of the classification cluster is output for the error after an update, : for a pre-update error output classification cluster weight center value, η: is the set learning rate, : For the study sample. The method for detecting an island phenomenon of a solar photovoltaic power generation system according to claim 6, wherein in the step (c6), the method for updating the weight value is only for the weight value of the classification cluster to which the learning sample belongs and an error output. The weight value of the classification cluster is adjusted, and the weighting value of the classification cluster to which the learning sample belongs may be expressed as follows: among them, : the minimum weight of the classification cluster to which the learning sample belongs after an update, The minimum weight of the classification cluster to which the learning sample belongs before updating. : the maximum weight of the classification cluster to which the learning sample belongs after an update, The maximum weight of the classification cluster to which the learning sample belongs before updating, η: is a set learning rate, : for the learning sample; the error output cluster weight value can be adjusted as follows: among them, : the minimum weight of the classification cluster for an updated error output, The minimum weight of the classification cluster for a pre-update error output, : The maximum weight of the classification cluster for an updated error output, The maximum weight of the classification cluster is output for a pre-update error, η: is the set learning rate, : For the study sample. The method for detecting an islanding phenomenon of a solar photovoltaic power generation system according to claim 7, wherein the learning sample is Where N _{p is} the total number of learning samples, each sample contains the characteristics and categories of the data, and the i- th sample is expressed as Where i =1, 2, 3..., N _p , n represents the total number of features of the sample, and p is the island detection category to which the ith sample belongs. The method for detecting an island phenomenon of a solar photovoltaic power generation system according to claim 1, wherein the step (d) comprises: (d1) obtaining a weight value of the extension-type neural network of the training completion, and completing the training The weight value of the extension-type neural network is used as the weight value of the extension-type neural network for identification; (d2) the weight center value of each classification cluster is calculated according to the obtained weight value; (d3) reading the test sample for which island state detection is to be performed; (d4) calculating the extension distance of the test sample from each of the clusters; and (d5) finding a minimum set of extension distances to classify the cluster The output value of the corresponding output layer node is 1 to display the category of the cluster to which the test sample belongs.