WO2020029327A1 - Photovoltaic array fault diagnosis method based on improved random forest algorithm - Google Patents

Abstract

Disclosed is a photovoltaic array fault diagnosis method based on an improved random forest algorithm, relating to the field of photovoltaic technology. In the method, a running state of each branch in a photovoltaic array is reflected by means of parameters of a main trunk and each branch of the photovoltaic array, and a running state of each photovoltaic cell assembly in the branch is reflected by means of a voltage difference between arrays among different branches, so as to realize fault location of the photovoltaic array; and by optimizing and improving the three parts, i.e. decision tree weighting and voting, tie processing and the importance measurement of fault features, by means of out-of-package samples, the accuracy of the fault diagnosis can be higher and the reliability of the fault diagnosis can be stronger.

Description

Fault diagnosis method of photovoltaic array based on improved random forest algorithm Technical field

The invention relates to the field of photovoltaic technology, in particular to a photovoltaic array fault diagnosis method based on an improved random forest algorithm.

Background technique

With the increase of energy demand, the increasing depletion and cost of fossil energy, and the impact of many factors such as global warming, renewable energy technologies have developed rapidly. Among them, solar energy has easy access, no noise, clean and endless. And other advantages have become an important part of renewable energy. At present, there are two main forms of solar power generation: solar thermal power and photovoltaic power. CSP is similar to thermal power generation, but the thermal energy mainly comes from large-scale mirror collection. After heating the water, it uses steam to promote the operation of traditional generators to achieve power generation. Although this type of power generation avoids the photoelectric conversion of silicon crystals, Light intensity requirements are high and power generation costs are high. Photovoltaic power generation is based on the photovoltaic effect, which directly converts solar energy into electricity. The construction period is short, and no waste residue, waste water and other pollutants will be generated during operation, especially for mountainous areas, islands and remote areas with less developed transportation. More important value.

Under normal circumstances, photovoltaic power generation uses photovoltaic cell modules in series and parallel to form a photovoltaic array to operate, and photovoltaic arrays use more photovoltaic cell modules, the probability of failure of the photovoltaic cell module itself is higher, and the photovoltaic array runs for a long time Outdoors, the operating environment is harsh, and it is prone to aging and damage. As a result, the photovoltaic cell module's power generation efficiency is reduced or even stopped. When a photovoltaic cell module in a photovoltaic array fails, it will lead to a decrease in system efficiency and adversely affect the operation and scheduling of the power system. In severe cases, it may even cause property damage and casualties. Therefore, the photovoltaic array is faulty. Diagnosis is of great significance. The currently used fault diagnosis methods mainly include two types: direct method and indirect method. Indirect method is more typical with infrared heat detection method and power generation comparison method. Direct method is more typical with ground capacitance method and time. Domain reflection method, intelligent diagnosis algorithm, and electrical characteristic detection method. Among them, fault diagnosis by combining intelligent diagnosis algorithm and electrical characteristic detection method is currently a method with great potential.

However, the current fault diagnosis methods of photovoltaic arrays rarely consider fault location. The main reason is that the current intelligent diagnostic algorithms mostly use neural networks represented by BP (Back Propagation), and the number of photovoltaic cell modules in photovoltaic arrays. There are many. When one or more photovoltaic cell modules fail, there are many forms of arrangement and combination of faults. Coupled with different fault types, sometimes there are even hundreds of fault combinations, which will require more training. The sample and training time will increase accordingly, the diagnostic accuracy cannot be guaranteed, and the diagnosis is difficult.

Summary of the invention

In view of the above problems and technical requirements, the present inventor proposes a photovoltaic array fault diagnosis method based on an improved random forest algorithm. This method determines the operating condition of a photovoltaic cell module by collecting the voltage and current of the photovoltaic cell modules between the arrays, thereby realizing a fault. Positioning.

The technical solution of the present invention is as follows:

A photovoltaic array fault diagnosis method based on an improved random forest algorithm, the method includes:

Determine the typical operating state of the photovoltaic array during operation. The photovoltaic array includes n branches, and each branch includes m photovoltaic cell modules. The typical operating state of the photovoltaic array is used to indicate the status of each photovoltaic cell module in the photovoltaic array. Running status, m and n are positive integers;

When the photovoltaic array is in each typical operating state, collect the array trunk parameters of the photovoltaic array, the corresponding array branch parameters of each branch, and the voltage difference between the arrays between different branches;

Construct fault feature vectors according to the array trunk parameters, array branch parameters, and the voltage difference between the arrays. According to the fault feature vectors, a data sample set of the photovoltaic array is constructed, and the data sample set is divided into a training sample set and a test sample set;

Based on the random forest algorithm, the photovoltaic array fault diagnosis model is trained by using the training sample set, and the photovoltaic array fault diagnosis model is tested by using the test sample set. The photovoltaic array fault diagnosis model includes N decision trees, where N is a positive integer and N ≥ 2 ;

The PV array fault diagnosis model completed by the test is used to diagnose the PV array to be diagnosed, and the voting results of N decision trees for each typical operating state are obtained;

The fault diagnosis result of the photovoltaic array to be diagnosed is obtained according to the voting results corresponding to each typical operating state, and the fault diagnosis result is used to indicate the operating state of each photovoltaic cell module in the photovoltaic array.

Its further technical solution is that the array trunk parameters and each array branch parameter include the maximum power point voltage, the maximum power point current, the open circuit voltage and the short circuit current of the circuit, respectively.

Its further technical solution is that the photovoltaic array fault diagnosis model includes N decision trees and the corresponding weights of each decision tree, and the photovoltaic array fault diagnosis model completed by the test is used to diagnose the PV array to be diagnosed, including:

N decision trees vote for each typical running state, and each decision tree casts its own weight as the number of votes;

Statistics obtained the voting results corresponding to each typical operating state.

Its further technical solution is to obtain a photovoltaic array fault diagnosis model by training on a training sample set based on a random forest algorithm, including:

Perform N rounds of sampling on the training sample set containing p samples. For each round of sampling, use the sampling method with replacement to sample p times to obtain the sub-training sample set and out-of-package sample set corresponding to the round of sampling. The sub-training set includes For the p samples selected in this round of sampling, the out-of-package sample set includes the samples that were not sampled in this round of sampling, and p is a positive integer;

A decision tree is trained according to the sub-training sample set corresponding to each round of sampling, and the decision tree is tested by using the out-of-package sample set corresponding to the round of sampling to obtain the out-of-package accuracy rate. Weight

The training N decision trees and the weights corresponding to each decision tree are summarized to obtain a photovoltaic array fault diagnosis model.

Its further technical solution is to calculate the weight corresponding to the decision tree according to the out-of-package accuracy rate, including calculation:

Among them, w (i) is the weight corresponding to the i-th decision tree, Hoob (i) is the out-of-package accuracy rate corresponding to the i-th decision tree, i and j are parameters, 1≤i≤N and i is positive Integer.

In a further technical solution, the method further includes:

The N out-of-package sample sets are aggregated to obtain a total out-of-package sample set. For each sample in the out-of-package sample set, the samples are tested using the trained photovoltaic array fault diagnosis model to obtain the initial out-of-package accuracy rate;

The sample includes a total of K fault features. Noise is added to the k-th fault feature, and the sample is tested again using the photovoltaic array fault diagnosis model to obtain a new out-of-package accuracy rate. The initial out-of-package accuracy rate and the new out-of-package accuracy rate are calculated. The difference between the accuracy rates determines the importance metric of the kth feature, where K is a positive integer, k is a parameter, and the starting value of k is 1;

When k <K, let k = k + 1, and perform the step of adding noise to the k-th fault feature again; until k = K, obtain the importance measure values of K fault features;

For each sample in the training sample set, delete the t fault features with the smallest importance metric among the K fault features of the sample to obtain a processed training sample set;

Based on the random forest algorithm, the processed training sample set is used to construct a photovoltaic array fault diagnosis model.

Its further technical solution is to obtain the fault diagnosis result of the photovoltaic array to be diagnosed according to the voting results corresponding to each typical operating state, including:

If the typical operating state corresponding to the highest number of votes includes only one, the typical operating state corresponding to the highest number of votes is output as a fault diagnosis result;

If the typical running state corresponding to the highest number of votes includes at least two, the L decision trees with the highest weights are selected to vote for each typical running state. If there is a typical running state to obtain the selection of the most decision trees, then The typical running state is output as the fault diagnosis result; otherwise, the typical running state selected by the decision tree with the highest weight is output as the fault diagnosis result, where L is a positive integer and 2≤L <N.

The beneficial technical effects of the present invention are:

1. The present application discloses a photovoltaic array fault diagnosis method based on an improved random forest algorithm, which collects parameters of the main trunk of the photovoltaic array and each branch, and the voltage difference between the arrays between different branches. The parameters of each branch can reflect the operating status of each branch in the photovoltaic array, thereby diagnosing which branch in the photovoltaic array has what kind of failure, and the voltage difference between the arrays between different branches can reflect the branch The operating status of each photovoltaic cell module in the system, so as to diagnose which photovoltaic cell module is faulty. The combination of the two types of information effectively achieves the fault location of the photovoltaic array. At the same time, the random forest algorithm is used, which is suitable for practical applications. The characteristics of photovoltaic arrays overcome the problems of large amount of data and long training time required by traditional neural network algorithms, and can complete diagnostic tasks simply and quickly.

2. The method disclosed in this application is based on a data-driven idea, and combines random forest algorithm, decision tree weighting, voting tie processing, and variable importance measurement to realize fault diagnosis of photovoltaic arrays and effectively improve diagnostic accuracy.

3. The method disclosed in this application uses an improved random forest fault diagnosis model structure. Through the integration of multiple decision trees, a strong classifier is constructed, which uses decision tree weighting, voting tie processing, and fault feature importance measures. A series of steps were improved, and finally an improved random forest algorithm was constructed. The diagnosis results also show that although the improved algorithm has slightly longer training time, it has higher accuracy and more reliable diagnosis.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flowchart of a photovoltaic array fault diagnosis method disclosed in the present application.

Figure 2 is a schematic diagram of the construction of a 3 * 3 photovoltaic array.

Figure 3-a is the U-I output characteristic curve of a single solar cell module under different light intensities.

Figure 3-b is the U-P output characteristic curve of a single solar cell module under different light intensities.

Figure 3-c is the U-I output characteristic curve of a single solar cell module under different temperature conditions.

Figure 3-d is the U-P output characteristic curve of a single solar cell module under different temperature conditions.

FIG. 4 is an output characteristic curve of a single solar cell module at a light intensity of 600 W / m2, 800 W / m2, and 1000 W / m2.

Figure 5-a is a comparison of the output characteristic curves of a photovoltaic array when there is no fault and when a single-mode fault occurs.

Figure 5-b is a comparison diagram of the output characteristic curves of a photovoltaic array when there is no fault and when a multi-mode fault occurs.

Figure 6-a is the output characteristic curve of a photovoltaic array when a photovoltaic cell module is shorted in a branch.

Figure 6-b is the output characteristic curve of a photovoltaic array when one branch is open.

Figure 6-c is the output characteristic curve of a photovoltaic array when two photovoltaic cell modules in a branch have shadows.

Figure 6-d shows the output characteristic curve of a photovoltaic array when one photovoltaic cell module is short-circuited in one branch, and one photovoltaic cell module is shadowed in the other branch.

detailed description

The specific embodiments of the present invention are further described below with reference to the accompanying drawings.

This application discloses a photovoltaic array fault diagnosis method based on an improved random forest algorithm. Please refer to FIG. 1. The entire fault diagnosis process includes the following steps:

The first step is to build a photovoltaic array in the form of m * n. The photovoltaic array includes n branches in parallel, and each branch includes m photovoltaic cell modules in series, where m and n are positive integers. The photovoltaic cell modules in a photovoltaic array are usually solar cell modules. The specifications of each photovoltaic cell module are the same to ensure the potential balance of the photovoltaic array. Each photovoltaic cell module is connected with a bypass diode in parallel and each branch in series with an isolation diode. Protection, detection devices such as fuses, voltage and current sensors are correctly installed and working properly.

This application takes n = 3 and m = 3 as examples. For a schematic diagram of a 3 * 3 photovoltaic array, please refer to FIG. 2. It includes branch A, branch B, and branch C. Each branch includes 3 branches. For photovoltaic cells connected in series, Figure 2 does not show devices such as bypass diodes and isolation diodes. The photovoltaic cell module in the photovoltaic array of this application uses SunTech STP 270-24 / Vd solar cell module of Suntech Solar Co., Ltd. The parameter specification table of this type of solar cell module is shown in the following table:

Parameter Type	Parameter value
Open circuit voltage (Voc)	44.5V
Maximum power point voltage (Vm)	35.0V
Short-circuit current (Isc)	8.20A
Maximum power point current (Im)	7.71A
Maximum power (Pmax)	270W
Operating temperature	-40 ℃ ～ + 85 ℃
Maximum withstand voltage of the system	1000V DC
Rated current of fuse after series connection	20A
Power error	± 3%

With the temperature kept constant, the UI (voltage-current) output characteristic curve when the light intensity rises from 200 W / m ² to 1000 W / m ² (200, 400, 600, 800, and 1000 W / m ^{2 respectively} ) is as follows Figure 3-a shows the UP (voltage-power) output characteristic curve shown in Figure 3-b. Under the condition of keeping the light intensity constant, the UI output characteristic curve when the temperature drops from 35 ° C to 15 ° C (respectively 35 ° C, 30 ° C, 25 ° C, 20 ° C, and 15 ° C) is shown in Fig. The UP (voltage-power) output characteristic curve is shown in Figure 3-d. In particular, the product manual gives the output characteristic curve of this type of solar cell module under the light intensity of 600W / m ² , 800W / m ² and 1000W / m ² as shown in FIG. 4.

The second step is to determine the typical operating state of the photovoltaic array during operation. The typical operating state of the photovoltaic cell module in this application is used to indicate the operating state of each photovoltaic cell module in the photovoltaic cell module. There are many types of faults in photovoltaic arrays during actual operation. There are four main types of faults: short-circuit faults, open-circuit faults, shadow faults, and mixed faults. When one or more branches in a photovoltaic array have a photovoltaic cell component fault, the faulty branch and the output characteristic curve of the entire photovoltaic array will change, as shown in Figure 5-a. Output characteristics of faulty branches, faulty arrays, normal branches, and normal arrays when there are no faults and single-mode faults. Figure 5-b shows the faulty branches of photovoltaic arrays when there are no faults and multi-mode faults occur. , Fault array, normal branch and normal array output characteristic curve, where the normal branch refers to the branch in which each photovoltaic cell module is operating normally, the normal array refers to the photovoltaic array in which each branch is operating normally, the fault A branch is a branch in which a photovoltaic cell module fails, and a fault array refers to a photovoltaic array in which a branch fails. It can be seen from the figure that the output current of the normal branch is large, and the output of each photovoltaic cell module is also equal, while the output current capacity of the faulty branch is weak, and it can be seen that the current is significantly lower. The output characteristic curve has two regions: high voltage region and high current region. In the high voltage region, the condition of the branch can be determined by detecting the output current of each branch. In the high current region, the output voltage of the photovoltaic cell module can be detected. To achieve fault location.

The above four types of faults may occur during the operation of photovoltaic arrays. In addition to the normal operating conditions, the operating states of photovoltaic arrays mainly include five categories: normal operating states, short circuit fault states, open circuit fault states, shadow fault states, and mixed fault states :

(1) Normal operating state, that is, all branches are operating normally.

(2) Short circuit fault state, that is, any one branch or multiple branches are short-circuited. It should be noted that it is impossible for all photovoltaic cell modules to be short-circuited in a branch. If all photovoltaic cell modules are short-circuited, the fuse device is blown and the branch will be opened.

(3) Open circuit fault state, that is, any branch or multiple branches are open.

(4) The shadow fault state, that is, any branch or multiple branches have shadows. Unlike the short-circuit fault state, there can be shadows of all photovoltaic cell modules in a branch, so the number of this state is large. .

(5) Mixed fault states, which mainly include the case where a short-circuit fault and a shadow fault occur simultaneously in the photovoltaic array.

For these five types of operating states, two-level state classification can be adopted. First, rough classification is performed to locate the branches in the photovoltaic array. This step does not involve fault location. You can choose some of the most common operations based on actual conditions and experience. status. It should be noted that in photovoltaic systems, the major faults of the system have evolved on the basis of minor faults. The probability of direct major faults is very small, and the types and combinations of major faults are too many, so this application mainly Consider light and moderate faults, that is, consider that at least one branch in the photovoltaic array is in a normal operating state. For example, in the photovoltaic array shown in FIG. 2, assuming that the branch C in the photovoltaic array is always in a normal operating state, the following 20 operating states can be selected as the result of the first-level state classification, as shown in the following table:

Taking the operating states corresponding to the status labels F2, F7, F10, and F18 as examples, their output characteristic curves are shown in order in Figures 6-a, 6-b, 6-c, and 6-d. On the basis of the first-level status classification, the second-level status classification can be performed to locate the photovoltaic cell module in the branch, thereby achieving fault location. Taking the status label F2 as an example, this situation indicates that a photovoltaic cell module in a branch is short-circuited. In the 3 * 3 photovoltaic array shown in FIG. 2, assuming that the branch C in the photovoltaic array is in a normal operating state, the The condition indicates that there is a short-circuit of a photovoltaic cell module on branch A or branch B. There are 6 cases in total, and the results of the second-level state classification based on this are shown in the following table:

Similarly, the results of the second-level status classification based on the first-level status classification corresponding to the status label F3 can be shown in the following table:

The rest can be deduced by analogy. At this point, the typical operating states of the photovoltaic array are obtained through two-level state classification.

The third step is to collect the circuit parameters of the photovoltaic array. When the photovoltaic array is in each typical operating state, collect the array trunk parameters of the photovoltaic array, the corresponding array branch parameters of each branch, and the voltage difference between the arrays between different branches. The array trunk parameters of the photovoltaic array are parameters on the main trunk of the photovoltaic array. The array trunk parameters in this application include the maximum power point voltage V _m , the maximum power point current I _m , and the open circuit voltage V _{oc of the} main trunk. And short-circuit current I _sc , which is because these four parameters can well characterize the UI output characteristic curve of the photovoltaic cell module. Similarly, for each branch in the photovoltaic array, the maximum power point voltage V _m , the maximum power point current I _m , the open circuit voltage V _oc, and the short-circuit current I _sc of the branch are collected as the array branch parameters. As shown in Figure 2, in actual operation, the circuit parameters of a photovoltaic array can be measured by configuring a voltage sensor and a current sensor in the photovoltaic array. For an m × n scale photovoltaic array, it is necessary to configure [n × (m-1) / 2] voltage sensors and n + 1 current sensors, where the symbol [] means rounding up n × (m-1) / 2. The current sensor is set in each branch and the main trunk to detect the current of the branch and the current of the main trunk. The voltage sensor is set between different branches to detect the voltage difference between the arrays between different branches. As shown in FIG. 2, three voltage sensors U _a , U _b and U _{c are provided} . The voltage sensor U _{a is} disposed between the branch A and the branch B, and the voltage sensor U _{b is} disposed between the branch B and the branch C. The voltage sensor U _{c is} arranged between the branch C and the branch A. The detection of the branch current and the detection of the voltage difference between the branches are essentially to detect the voltage and current of each photovoltaic cell module. The changes in the parameter values such as voltage and current reflect the operation of the photovoltaic cell module, thereby achieving line faults. The principle of positioning is as follows:

In a photovoltaic array, the voltage matrix U _{pv array} of the photovoltaic _array is:

Among them, U ₁₁ represents the output voltage of the first photovoltaic cell module of the first branch in the photovoltaic array, that is, the output voltage of PV1 in FIG. 2, and U _mn represents the m-th branch of the n-th branch in the photovoltaic array The output voltage of the photovoltaic cell module, that is, the output voltage of PV9 in FIG. 2, and so on. Then the photovoltaic array equation holds as follows:

Where U _array is the output voltage of the photovoltaic array, then

That is, the output voltage of each photovoltaic cell module should be 1 / m of the output voltage of the entire photovoltaic array.

Define the PV matrix weight matrix A _{pv array} , that is, the potential of the voltage sensor placement point between the branches is:

Among them, u ₁₁ = U ₂₁ -U ₁₁ , u ₂₁ = U ₃₁ -U ₂₁ , u _{(m-1) 1} = U _m1 -U _{(m-1) 1} , and so on. When the photovoltaic array is fault-free, In the normal running state, the values of the elements in each column in the weight matrix A _{pv array} are equal difference sequences, and the values of the elements in each row are equal. If a photovoltaic cell module in one branch fails, the voltage of other photovoltaic cell modules will also change. Use the following formula to calculate the output voltage of the changed photovoltaic cell module as

U × m = U _bad × (mx)

Among them, U is the output voltage of the photovoltaic cell module in normal operation, m is the total number of photovoltaic cell modules included in the branch, x is the number of photovoltaic cell modules that have failed in the faulty branch, and U _bad is the faulty branch. The output voltage of a photovoltaic cell module operating normally in the circuit. For example, in Figure 2, the output voltages of photovoltaic cell modules PV1, PV2 and PV3 are

When the PV cell module PV1 fails, it will be short-circuited by the bypass diode, then the output of the PV cell module PV1 is 0, and x = 1 in the above formula, then

Then the output voltage across the photovoltaic cell modules PV2 and PV3 is

That is, the output voltage of the corresponding photovoltaic cell modules PV2 and PV3 will increase. When the photovoltaic array is in a fault state (including short circuit, open circuit, shadow, and mixed faults), the elements of the rows of the non-faulty branch in the weight matrix A _{pv array} are equal, and the elements of the rows of the faulty branch and the non-faulty branch The elements of each row of the road are not equal; and the elements of each column of the branch with a fault are not equal difference sequences. The following table shows the characteristic values of the voltage difference between the arrays collected by each voltage sensor in the case of a partial short-circuit fault:

Status label	Ua	Ub	Uc	Status label	Ua	Ub	Uc
F1	Uarray / 3	Uarray / 3	Uarray / 3	F3-46	2Uarray / 3	2Uarray / 3	Uarray / 3
F2-1	Uarray / 6	Uarray / 3	2Uarray / 3	F3-56	-Uarray / 3	2Uarray / 3	Uarray / 3
F2-2	Uarray / 6	Uarray / 3	Uarray / 6	F4-14	Uarray / 2	Uarray / 6	2Uarray / 3
F2-3	2Uarray / 3	Uarray / 3	Uarray / 6	F4-15	0	Uarray / 6	2Uarray / 3
F2-4	2Uarray / 3	Uarray / 6	Uarray / 3	F4-16	0	2Uarray / 3	2Uarray / 3
F2-5	Uarray / 6	Uarray / 6	Uarray / 3	F4-24	Uarray / 2	Uarray / 6	Uarray / 6
F2-6	Uarray / 6	2Uarray / 3	Uarray / 3	F4-25	0	Uarray / 6	Uarray / 6
F3-12	-Uarray / 3	Uarray / 3	2Uarray / 3	F4-26	0	2Uarray / 3	Uarray / 6
F3-13	2Uarray / 3	Uarray / 3	2Uarray / 3	F4-34	Uarray	Uarray / 6	Uarray / 6
F3-23	2Uarray / 3	Uarray / 3	-Uarray / 3	F4-35	Uarray / 2	Uarray / 6	Uarray / 6
F3-45	2Uarray / 3	-Uarray / 3	Uarray / 3	F4-36	Uarray / 2	2Uarray / 3	Uarray / 6

For the meaning of the above status label, please refer to the content in the second step above. Those skilled in the art understand that the situation indicated by status label F4-14 is that PV1 on branch A and PV4 on branch B are short-circuited. By analogy, this application does not repeat them one by one.

When the photovoltaic array has a shadow fault, the voltage difference between the arrays collected by each voltage sensor will also be within a range. The characteristic values of some shadow fault conditions are shown in the following table:

Status label	Ua	Ub	Uc
F9-1	Uarray / 6 <Ua <Uarray / 3	Uarray / 3	Uarray / 3 <Uc <2Uarray / 3
F9-2	Uarray / 6 <Ua <Uarray / 3	Uarray / 3	Uarray / 6 <Uc <Uarray / 3
F9-3	Uarray / 3 <Ua <2Uarray / 3	Uarray / 3	Uarray / 6 <Uc <Uarray / 3
F9-4	Uarray / 3 <Ua <2Uarray / 3	Uarray / 6 <Ub <Uarray / 3	Uarray / 3
F9-5	Uarray / 6 <Ua <Uarray / 3	Uarray / 6 <Ub <Uarray / 3	Uarray / 3

F9-6

Uarray / 6 <Ua <Uarray / 3

Uarray / 3 <Ub <2Uarray / 3

Uarray / 3

The situation indicated by the status label F9-1 is: there is a shadow on PV1 on the branch A. Similarly, those skilled in the art can understand the meanings of the remaining status labels, which are not described in detail in this application.

It can be known from the above discussion that when a photovoltaic array fails, it will be reflected in the weight matrix A _{pv array} , which will cause the voltage difference between the arrays collected by the voltage sensor to change. Therefore, this application analyzes the voltage difference between the collection arrays. Can realize the fault location of photovoltaic array.

In addition, as can be seen from the above, the voltage difference between the arrays collected by the voltage sensor is proportional to the output voltage U _{array of the} photovoltaic array, so it can be further simplified. Instead of directly using the collected voltage difference between the arrays, the array The voltage difference is divided by the output voltage U _array of the photovoltaic _{array to} obtain a voltage coefficient, and the voltage difference is used to represent the voltage difference between the arrays.

The fourth step is the construction of fault feature vectors and data sample sets. Construct fault feature vectors based on the collected array trunk parameters, array branch parameters, and the voltage difference between the arrays. After the fault feature vectors are obtained, they can be summarized and constructed into a data sample set of photovoltaic arrays. Data samples are collected from the fault feature vectors. The collection method is a method known to those skilled in the art, which is not described in this application. The data sample set is divided to obtain a training sample set and a test sample set, which can usually be divided into a training sample set and a test sample set according to a ratio of 2: 1.

The fifth step is the construction of a photovoltaic array fault diagnosis model based on an improved random forest algorithm. Suppose the training sample set is X, and the dimension of the fault feature vector is K, that is, the number of fault features of the sample is K, K is a positive integer; the number of decision trees to be established in fault diagnosis is N, N≥2, and N is Positive integer, the steps of constructing a photovoltaic array fault diagnosis model based on an improved random forest algorithm using training sample sets can be briefly described as:

1. Perform N rounds of sampling on the training sample set X containing p samples. For the i-th round of sampling, use sampling with replacement sampling p times. Because the sampling is performed with replacement, some samples will be repeated. In the case where some samples are not drawn, the p samples extracted in this round of sampling can constitute the sub-training sample set X _i corresponding to this round of sampling, and the training samples are concentrated in the unsampled samples in this round of sampling. the samples corresponding to the outer wheel sets of samples packet _{oob i (out-of-bag} ), p is a positive integer, i is the parameter, 1≤i≤N and i is a positive integer.

2. A decision tree is obtained by training the sub-training sample set corresponding to each round of sampling, that is, the sub-training sample set X _{i is used to} train the decision tree Q _i . When the nodes of the decision tree are split, q fault features are randomly selected from K fault features, where q is a positive integer and q≤K, and the Gini coefficient corresponding to each fault feature in the q fault features is calculated. The fault feature corresponding to the coefficient is used as the best split feature for node splitting. The calculation method of the Gini coefficient can use the existing calculation formula, which is not described in detail in this application.

3. For the i-th round of sampling, since the samples in the corresponding out-of-package sample set oob _i do not participate in the training of the decision tree Q _i , it can be used as a test set, and the out-of-package sample set oob _{i is used} for the decision tree Q. _i is tested to obtain the out-of-packet accuracy H ^oob (i) of the decision tree Q _i .

In ordinary random forest algorithms, the weights of decision trees are equal by default. Therefore, some decision trees with excellent performance may not show their advantages and may even be overwhelmed by a majority vote. The tree is weighted, and the out-of-package accuracy rate of the decision tree is used to evaluate the performance of the decision tree. After all the decision tree training is completed, the out-of-package accuracy rate of each decision tree is calculated, and the decision is obtained for the i-th sampling training. Tree Q _i , using its corresponding out-of-package accuracy rate H ^oob (i) to calculate the weight of the decision tree Q _i is as follows:

Among them, w (i) is the weight corresponding to the i-th decision tree, and j is a parameter.

4. Add all decision trees and the corresponding weights of each decision tree to the decision tree set, so as to form a photovoltaic array fault diagnosis model.

In the sixth step, the photovoltaic array fault diagnosis model is further optimized. As the size of a photovoltaic array expands, the number of branches and voltage sensors running at the same time will increase, and the dimension K of the fault feature vector will increase, which will increase the difficulty of model training. The fault characteristics of the sample are measured for importance to delete some fault characteristics. The specific steps are as follows:

1. Aggregate N out-of-package sample sets oob _i to obtain a total out-of-package sample set. For each sample in the out-of-package sample set, use the trained photovoltaic array fault diagnosis model to test the sample to obtain the corresponding sample. The initial out-of-package accuracy e, specifically: The PV array fault diagnosis model includes a total of N decision trees. Using each decision tree that is not trained with the sample to test the sample, you can determine the correct test and the wrong decision The number of trees to obtain the initial out-of-package accuracy e.

2. The sample includes a total of K fault characteristics. Noise is added to the k-th fault characteristic, and the sample is tested again using the photovoltaic array fault diagnosis model to obtain a new out-of-package accuracy rate e1, and the initial out-of-package accuracy rate is calculated. And the new out-of-package accuracy rate, e-e1, to determine the importance metric of the k-th feature, where K is a positive integer, k is a parameter, and the starting value of k is 1.

3. When k <K, let k = k + 1, and perform step 2 again, that is, the step of adding noise to the k-th fault feature. Until k = K, importance metric values of all K fault features of the sample are obtained.

4. The importance metrics of the K fault features calculated using the samples in the outsourced sample set are generally consistent, but there are errors in actual operation. Therefore, the After the importance metric values of the K fault features are calculated for each sample, they are summed to determine the t fault features with the smallest importance metric value among the K fault features. Then for each sample in the training sample set, t fault features with the smallest importance metric among the K fault features of the sample are deleted to obtain a processed training sample set, 1 ≦ t <K.

5. Based on the random forest algorithm, the processed training sample set is used to construct a photovoltaic array fault diagnosis model.

The seventh step is to test the fault diagnosis model of the photovoltaic array. By using the test sample set to test the PV array fault diagnosis model, indicators such as diagnostic accuracy, average training time, and test time can be obtained. According to the F1-F20 listed above, the first-level state type of the photovoltaic array is diagnosed without fault feature deletion. The improved random forest algorithm and ordinary random forest algorithm are used for fault diagnosis. The training sample set and the test sample set. Both algorithms use the same decision tree model. The number of decision trees is 40. The experimental results are compared as follows:

diagnosis method	Training time / s	Test time / s	Diagnostic accuracy /%
Ordinary random forest	0.976120	0.38079	89
Improve Random Forest	1.472387	0.38711	91

As can be seen from the above table, although the algorithm based on the improved random forest provided in this application slightly increases the training time, the diagnosis accuracy is higher.

The second-level status classification can be regarded as a sub-class of the first-level status classification. When the size of a photovoltaic array is large, there are many permutations and combinations of the states. Taking the short-circuit fault status as an example, the short-circuit fault status can be obtained as the second-level status. The number after classification is 49, and the comparison of the experimental results is as follows: Similarly, the number of decision trees is set to 40, the average diagnosis accuracy of the decision tree is 64.13%, and the dimension of the fault feature vector is K = 19. After the 4-dimensional fault feature, the average diagnostic accuracy of the decision tree rose to 72.95%, and the diagnostic accuracy of the entire photovoltaic array fault diagnostic model also improved.

diagnosis method	Training time / s	Test time / s	Diagnostic accuracy /%
Ordinary random forest	2.0646	0.49109	94.49
Improve Random Forest	2.1171	0.50935	96.12

The eighth step is to use the completed photovoltaic array fault diagnosis model to diagnose the PV array to be diagnosed, that is, to use N decision trees to vote for each typical operating state. The voting method is used to vote, that is, each decision tree The weight corresponding to itself is cast as the number of votes.

After the voting is completed, the voting results corresponding to each typical operating state are statistically obtained, and then the fault diagnosis results of the photovoltaic array to be diagnosed are obtained according to the voting results corresponding to each typical operating state. The obtained fault diagnosis results are used to indicate each photovoltaic in the photovoltaic array. The operating status of the battery assembly, specifically:

1) The number of typical operating states that detect the highest number of votes.

2) If the typical running state corresponding to the highest number of votes includes only one, the typical running state corresponding to the highest number of votes is directly output as the fault diagnosis result, otherwise, the following step 3) is performed.

3) When the typical running state corresponding to the highest number of votes includes at least two, this application creatively adds a draw processing strategy, that is, the L decision trees with the highest weights are selected to vote again for each typical running state, L It is a positive integer and 2≤L <N, and it is selected according to actual needs. If there is a typical operating state to obtain the selection of the most decision trees, the typical operating state is output as the fault diagnosis result; otherwise, the typical operating state selected by the decision tree with the highest weight is output as the fault diagnosis result. For example, taking L = 3 as an example, when re-voting, if the voting results of two decision trees are the same, the result is output, otherwise the voting result of the decision tree with the highest weight is output.

What has been described above are only preferred embodiments of the present application, and the present invention is not limited to the above embodiments. It can be understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims (7)

1. A photovoltaic array fault diagnosis method based on an improved random forest algorithm, characterized in that the method includes:

   Determine a typical operating state of a photovoltaic array during operation, the photovoltaic array includes n branches, and each branch includes m photovoltaic cell components, and the typical operating state of the photovoltaic array is used to indicate The operating state of each photovoltaic cell module, m and n are positive integers;

   When the photovoltaic array is in each of the typical operating states, collecting an array trunk parameter of the photovoltaic array, an array branch parameter corresponding to each branch, and a voltage difference between the arrays between different branches;

   A fault feature vector is constructed according to the array trunk parameters, array branch parameters, and voltage difference between the arrays. A data sample set of the photovoltaic array is constructed based on the fault feature vectors, and the data sample set is divided into training sample sets. And test sample sets;

   Use the training sample set to train a photovoltaic array fault diagnosis model based on a random forest algorithm, and use the test sample set to test the photovoltaic array fault diagnosis model. The photovoltaic array fault diagnosis model includes N decision trees. N is a positive integer and N≥2;

   Diagnose the photovoltaic array to be diagnosed by using the photovoltaic array fault diagnosis model completed by the test, and obtain voting results of the N decision trees for each of the typical operating states;

   A fault diagnosis result of the photovoltaic array to be diagnosed is obtained according to a voting result corresponding to each of the typical operating states, and the fault diagnosis result is used to indicate an operating state of each photovoltaic cell module in the photovoltaic array.
2. The method according to claim 1, wherein:

   The array trunk parameters and each of the array branch parameters include a maximum power point voltage, a maximum power point current, an open circuit voltage, and a short circuit current of a circuit, respectively.
3. The method according to claim 1 or 2, wherein the photovoltaic array fault diagnosis model includes N decision trees and weights corresponding to each of the decision trees, then the photovoltaics completed by the test are used. The array fault diagnosis model diagnoses the PV array to be diagnosed, including:

   The N decision trees vote for each of the typical operating states, and each of the decision trees casts its corresponding weight as the number of votes;

   The voting results corresponding to each of the typical operating states are obtained by statistics.
4. The method according to claim 3, wherein the random forest algorithm based on the training sample set is used to obtain a photovoltaic array fault diagnosis model, comprising:

   Perform N rounds of sampling on the training sample set containing p samples. For each round of sampling, use the sampling method with replacement to sample p times to obtain the sub-training sample set and out-of-package sample set corresponding to the round of sampling. The set includes p samples selected in the round of sampling, and the out-of-package sample set includes samples that are not drawn in the round of sampling in the training sample set, and p is a positive integer;

   A decision tree is trained according to the sub-training sample set corresponding to each round of sampling, and the decision tree is tested using the out-of-package sample set corresponding to the round of sampling to obtain an out-of-package accuracy rate, which is calculated according to the out-of-package accuracy rate A weight corresponding to the decision tree;

   The training N decision trees and the weights corresponding to each decision tree are summarized to obtain the photovoltaic array fault diagnosis model.
5. The method according to claim 4, wherein the calculating the weight value corresponding to the decision tree according to the out-of-package accuracy rate comprises calculating:

   

   Among them, w (i) is the weight corresponding to the i-th decision tree, Hoob (i) is the out-of-package accuracy rate corresponding to the i-th decision tree, i and j are parameters, 1≤i≤N and i is positive Integer.
6. The method according to claim 4, further comprising:

   The N out-of-package sample sets are aggregated to obtain a total out-of-package sample set. For each sample in the out-of-package sample set, the samples are tested by using the trained photovoltaic array fault diagnosis model to obtain an initial Out-of-package accuracy

   The sample includes a total of K fault characteristics, noise is added to the k-th fault characteristic, and the sample is retested using the photovoltaic array fault diagnosis model to obtain a new out-of-package accuracy rate, and the initial package is calculated. The difference between the external accuracy rate and the new out-of-package accuracy rate determines the importance metric of the k-th feature, K is a positive integer, k is a parameter, and the starting value of k is 1;

   When k <K, let k = k + 1, and perform the step of adding noise to the k-th fault feature again; until k = K, obtain the importance metric values of the K fault features;

   For each sample in the training sample set, delete t fault features with the smallest importance metric among the K fault features of the sample to obtain a processed training sample set;

   A photovoltaic array fault diagnosis model is constructed by using the processed training sample set based on a random forest algorithm.
7. The method according to claim 3, wherein the obtaining a fault diagnosis result of the photovoltaic array to be diagnosed according to a voting result corresponding to each of the typical operating states comprises:

   If the typical operating state corresponding to the highest number of votes includes only one, outputting the typical operating state corresponding to the highest number of votes as the fault diagnosis result;

   If the typical running state corresponding to the highest number of votes includes at least two, the L decision trees with the highest weights are selected to vote again for each of the typical running states. If there is a typical running state to obtain the selection of the most decision trees, Then output the typical operating state as the fault diagnosis result; otherwise, output the typical operating state selected by the decision tree with the highest weight as the fault diagnosis result, where L is a positive integer and 2 ≦ L <N.

Images