CN105846780A - Decision tree model-based photovoltaic assembly fault diagnosis method - Google Patents

Abstract

The present invention belongs to the photovoltaic power generation technical field and provides a decision tree model-based photovoltaic assembly fault diagnosis method. The method includes the following steps that: photovoltaic assembly data are acquired, and data processing is carried out; obtained data are introduced into a decision tree-based photovoltaic assembly fault diagnosis model, the modeling steps of the model mainly comprise selection and processing of training and test sample data, tree establishment and tree pruning, decision tree model establishment and decision tree accuracy verification; and the fault type of a photovoltaic assembly is judged through a decision module. With the method of the invention adopted, manpower and resource waste caused by fault judgment errors can be avoided, the accuracy and reliability of fault judgment are improved, serious consequences to the photovoltaic assembly, caused by a fault, can be avoided, and the service life of the photovoltaic assembly can be prolonged.

Description

A kind of photovoltaic module method for diagnosing faults based on decision-tree model

Technical field

The present invention relates to photovoltaic module method for diagnosing faults, particularly relate to a kind of photovoltaic module based on decision-tree model different Chang Laohua and local shades method for diagnosing faults.

Background technology

Due to the characteristic such as inexhaustible and pollution-free of solar energy, the application of photovoltaic generation presents the state of high speed development Gesture.Monitoring and the maintenance of system running state are most important to the safe operation of photovoltaic generating system, and fault timely, reliable is pre- Alert it can be avoided that the major accident such as fire, equipment damage, and improve service life and the economic benefit of photovoltaic plant.The biggest portion Light splitting overhead utility all uses manual inspection to safeguard, block-by-block detection photovoltaic module electric parameter judges whether normally.But photovoltaic Assembly is arranged in eminence or field extreme environment, and running voltage reaches hectovolt, and manual maintenance is the most time-consuming, dangerous again.So The on-line fault diagnosis of photovoltaic module seems and becomes more and more important.

In the fault diagnosis research of photovoltaic module, between sign and the fault type of fault, there is extremely complex non-thread Property corresponding relation, this makes to set up a suitable fault diagnosis mathematical model.Wherein, the fault of photovoltaic module can be divided The most aging with abnormal for local shades.Under different duties, the output of photovoltaic module presents different changes.In order to just In the realization of method for diagnosing faults, by observing and research, it is determined that four output parameters, i.e. maximum power point electric current I_m, High-power point voltage U_m, short circuit current I_scWith open-circuit voltage U_ocFoundation as photovoltaic module fault diagnosis.

There is the method for diagnosing faults much for photovoltaic module both at home and abroad, mainly have neural network, fuzzy control, many Sensor method, infrared image analysis method, time domain reflectometry and direct-to-ground capacitance measurement method etc..Neural network and fuzzy algorithmic approach It is required for illumination meter and temperature, when neural network is broken down by study photovoltaic system between malfunction and failure cause Corresponding relation, be then saved in neural network structure and weights, the voltage that measurement is obtained or current value input train Neutral net, it is achieved fault diagnosis；Fuzzy control, by estimating output under normal circumstances, then utilizes this value Compare with real-time measurement values, it is achieved fault diagnosis；The data that FUSION WITH MULTISENSOR DETECTION method is recorded by analyte sensors judge Fault type；Infrared image analysis method is according to the photovoltaic module principle that operating temperature is different normally from fault time, by analyzing Photograph infrared image failure judgement type；Time Domain Reflectometry analytic process, by injecting a pulse to tandem photovoltaic circuit, is analyzed The signal shape returned and failure judgement type time delay；Direct-to-ground capacitance measurement method is by measuring tandem photovoltaic circuit over the ground Capacitance judges the position of open circuit.Above-mentioned BP network needs employing training sample off-line training in advance, and the adjustment of weights Being repeatedly to feed back, convergence rate is absorbed in local minimum point slowly and easily, and these shortcomings limit it and fully send out in fault diagnosis field Wave advantage；The fuzzy logic system of fuzzy control lacks self-learning capability, and this ability is in the real-time event of some high request It is again required in the case of barrier diagnosis；Infrared image analysis method and FUSION WITH MULTISENSOR DETECTION method with on-line checking, but can be advised for big For the photovoltaic system of mould, the major limitation of both approaches is to need a lot of infrared video cameras and sensor, can be further Increase the cost of electricity-generating of photovoltaic system；Direct-to-ground capacitance measurement method and Time Domain Reflectometry analytic process can only detect by off-line, the most only It is applicable to tandem photovoltaic circuit, the highest for measuring the required precision of equipment.

Summary of the invention

The defect existed for prior art, it is an object of the invention to provide a kind of photovoltaic module based on decision-tree model Method for diagnosing faults, it is possible to diagnose the operation troubles of photovoltaic module exactly, drastically increases photovoltaic module fault The accuracy of diagnosis, it is ensured that photovoltaic module reliably and securely runs.

For reaching above-mentioned purpose, the present invention uses following technical proposals:

A kind of photovoltaic module method for diagnosing faults based on decision-tree model, comprises the following steps:

Step one: collection photovoltaic module data: open-circuit voltage U_oc, short circuit current I_sc, maximum power point voltage U_mAnd electric current I_m, process through data and obtain: fill factor, curve factor FF, slope factor K and output current ratio I_m/I_sc；

Step 2: build photovoltaic module fault diagnosis model based on decision tree, respectively by training and test sample number According to the beta pruning chosen and process, contribute and set, the four processes of setting up decision-tree model and decision tree precise verification carry out structure Build photovoltaic module fault diagnosis model；

Step 3: decision-making judges, by open-circuit voltage U_oc, short circuit current I_sc, maximum power point voltage U_m, electric current I_m, and Fill factor, curve factor FF, slope factor K and output current ratio I_m/I_scImport to photovoltaic module fault diagnosis model and judge photovoltaic module Malfunction.

In described step one, four output parameters of collection photovoltaic module: open-circuit voltage U_oc, short circuit current I_sc, maximum work Rate point voltage U_mWith electric current I_m, process, through data, the fill factor, curve factor FF obtained and be shown below:

$FF=\frac{I_m{U}_m}{I_{sc}{U}_{oc}}---\left(1\right)$

Slope factor K is that in assembly I-U curve, maximum power point is to the absolute value of open-circuit voltage point straight slope, and it calculates Formula is as follows:

$K=\frac{I_m}{U_{oc}-U_m}---\left(2\right)$

In described step 2, train and obtain by testing or emulate with the data in processing links with choosing of test sample data Must, in testing or emulating, photovoltaic module is operated in normally respectively, in the case of local shades is the most aging with exception, carries out data drawing Point, a described data set part is used for simulation training, and another part is used for testing.

Achievement link in described step 2, carries out recurrence division by sample, and concrete partiting step is as follows:

First, an independent variable x is selected_i, then choose x_iOne value v_i, v_iN-dimensional space is divided into two parts, a part Institute the most all meet x_i≤v_i, the institute of another part the most all meets x_i＞ v_i, the value of property value for discontinuous variable Only two, i.e. it is equal to this value or is not equal to this value；

Then, two parts obtained above are chosen again by aforesaid operations an attribute and continues to divide, use Gini value Standard as dividing: the community set in node is A={A₁,A₂,…,A_m, wherein, the attribute in set A is that sampling obtains Photovoltaic module data: A₁=U_oc, A₂=I_oc, A₃=U_m, A₄=I_m, A₅=FF, A₆=K, A₇=I_m/I_sc, so, m=7；Right The each attribute of A, i.e. open-circuit voltage U should be gathered_oc, short circuit current I_sc, maximum power point voltage U_m, electric current I_m, fill factor, curve factor FF, tiltedly Rate factor K and output current ratio I_m/I_scObtain seven training sample set: S={s₁,s₂,…,s_n, the x in step in S correspondence_i, s_iFor the value of variable S, due to 774 groups of data of sampling altogether in experiment, so n=774；Corresponding each training sample set S, will Normally, the sampled data in the case of local shades and aging three kinds of exception is as category set: C={c₁,c₂,…,c_k, by Normal in only existing, local shades and abnormal aging three kinds of situations, so k=3；For training sample set S, correspondence is trained Value T in sample set S, as an attribute of S, S is split into incoherent subset S by it₁,S₂,…,S_w, due to use It is binary tree design, so, the v in step in T correspondence_iAnd w=2；

When dividing current attribute, Gini value is:

$Gini(S,T)=\underset{i=1}{\overset{w}{\Σ}}\frac{\left|S_i\right|}{S}(1-\underset{j=1}{\overset{k}{\Σ}}{p}^2(c_j,S_i\left)\right)---\left(3\right)$

Wherein, p (c_j,S_i) it is that sample set closes S_iIn belong to the ratio of jth class；|S_i| for subclass S_iGesture, to candidate Every kind of Gini value that may divide on this attribute of each property calculation in property set A, the division finding Gini value minimum is made For the optimum division on this attribute, the Gini value of optimum division in the most all candidate attribute, have minimum division Gini The attribute of value becomes final test attribute；

Finally, according to the value of Split Attribute, obtaining decision tree branches, data set will be divided into multiple subset, for Each subtree recalculates the Gini value of each attribute, the like, until stopping when meeting following condition contributing: the sample of node This number is 1 or sample belongs to same class.

The beta pruning of the tree in described step 2, uses the rear pruning method of cross validation, first marks off from training data One subset is used for setting up tree, then in remaining checking sub-collective estimation misclassification rate, then for different subsets repeatedly Repeat this division, then the misclassification rate obtained is averaged, obtain, about particular size tree, data were not met for new The cross validation of performance estimate, produce the tree that minimum cross validation misclassification estimates and be confirmed as the suitable big of final tree-model Little.

Described step 2 sets up decision-tree model, judges the fault type rule of photovoltaic module according to the decision tree generated As follows:

1) if FF < 0.485, then local shades fault it is judged as；

2) if 0.485 < FF < 0.615, and V_m> 31.085, then it is judged as normally working；

3) if 0.485 < FF < 0.615, and V_m< 31.085, and K < 0.135, then be judged as local shades fault；

4) if 0.485 < FF < 0.615, and V_m<31.085, and K>0.135, then be judged as abnormal degradation failure；

5) if FF>0.615, and K<0.195, then abnormal degradation failure it is judged as；

6) if FF > 0.615, and K > 0.195, then it is judged as normally working.

In described step 2, decision tree precise verification link uses classifying rules to be identified test data, utilizes and knows The effectiveness of the accuracy rate checking model of other result.

Compared with prior art, the beneficial effects of the present invention is:

1, photovoltaic method for diagnosing faults framework based on decision-tree model has been built, for analyzing the event of dissimilar photovoltaic module Barrier type is laid a good foundation；

2, avoid breakdown judge to slip up to manpower and the waste of resource, improve the accuracy of breakdown judge with reliable Property, it is to avoid the serious consequence that fault produces for photovoltaic module, extends the service life of photovoltaic module.

Accompanying drawing explanation

Fig. 1 is the flow chart of the inventive method.

Fig. 2 is described decision tree structure figure.

Fig. 3 is fault diagnosis system schematic diagram.

Detailed description of the invention

It is embodied as illustrating to the present invention the most further.

As it is shown in figure 1, a kind of photovoltaic module method for diagnosing faults based on decision-tree model, comprise the following steps:

Step one: collection photovoltaic module data: open-circuit voltage (Uoc), short circuit current (Isc), maximum power point voltage (Um) and electric current (Im), process through data and obtain: fill factor, curve factor (FF), slope factor (K) and export current ratio (I_m/I_sc)；

Step 2: build photovoltaic module fault diagnosis model based on decision tree, respectively by training and test sample number According to the beta pruning chosen and process, contribute and set, the four processes of setting up decision-tree model and decision tree precise verification carry out structure Build photovoltaic module fault diagnosis model；

Step 3: decision-making judges, by open-circuit voltage (Uoc), short circuit current (Isc), maximum power point voltage (Um), electric current (Im), and fill factor, curve factor (FF), slope factor (K) and output current ratio (I_m/I_sc) import to photovoltaic module fault diagnosis mould Type judges the malfunction of photovoltaic module.

Below above three step is further elaborated.

The photovoltaic module needing to gather in described step one is at four output parameters respectively: open-circuit voltage (U_oc), short circuit Electric current (I_sc), maximum power point voltage (U_m) and electric current (I_m).The fill factor, curve factor obtained is processed through dataTiltedly The rate factorAbsolute value for maximum power point in assembly I-U curve to open-circuit voltage point straight slope.

Described step 2 builds photovoltaic module fault diagnosis model based on decision tree, is divided into four processes, and it specifically walks Rapid as follows:

The first step: training being chosen with test sample data and processing

Training in this method is chosen with the data in processing links by testing acquisition, in experiment with test sample data Photovoltaic module is operated in normally respectively, in the case of local shades and abnormal the most aging three kinds.Gather four outputs of photovoltaic module Parameter: maximum power point electric current (I_m), maximum power point voltage (U_m), short circuit current (I_sc), open-circuit voltage (U_oc).Owing to providing The attribute carrying out selecting to decision tree only has maximum power point voltage (U_m), maximum power point electric current (I_m), open-circuit voltage (U_oc) With short circuit current (Isc), decision tree only can consider to carry out dividing internal node according to the value of wherein certain attribute, be difficult to find it In rule of combination, it is therefore necessary to optimize and create new attribute.It is calculated fill factor, curve factorSlope factorWith output current ratio (I_m/I_sc), by simulating, verifying, can be together with the four of photovoltaic module output parameter It is shown in Table 1 as new community set.

Table 1 community set

Gathering data during 774 groups of photovoltaic module work, and divide data, described data set 90% is used for instructing Practicing, 10% is used for testing, and part training data is as shown in table 2, and partial test data are as shown in table 3.

Table 2 part training data

Table 3 partial test data

Second step: the beta pruning contribute and set

First, an independent variable x is selected_i, then choose x_iOne value v_i, v_iN-dimensional space is divided into two parts, a part Institute the most all meet x_i￡ v_i, the institute of another part the most all meets x_i>v_i, the value of property value for discontinuous variable Only two, i.e. it is equal to this value or is not equal to this value.

Then, two parts obtained above are chosen again by aforesaid operations an attribute and continues to divide, use Gini value to make Standard for dividing: the community set in node is A={A₁,A₂,…,A_m, training sample set is combined into S={s₁,s₂,…,s_n, class Ji not be combined into C={c₁,c₂,…,c_k}.For the attribute that training sample set S, T are S, S is split into incoherent son by it Collection S₁,S₂,…,S_w.When dividing current attribute, Gini value is:Its In, p (c_j,S_i) sample set conjunction S_iIn belong to the ratio of jth class；|S_i| subclass S_iGesture.Candidate attribute is concentrated Every kind of Gini value that may divide on this attribute of each property calculation, finds minimum the dividing as on this attribute of Gini value Optimum division, the Gini value of optimum division in the most all candidate attribute, having the minimum attribute dividing Gini value becomes Final test attribute.

Finally, according to the value of Split Attribute, can obtain decision tree branches, data set will be divided into multiple subset, Each subtree is recalculated to the Gini value of each attribute, the like, until stopping contributing a period of time meeting following condition: The sample number of node is 1 or sample belongs to same class, decision tree and highly arrives the threshold value that user sets.

Due in 774 groups of data being gathered, carry out internal node each time when dividing, FF, K and V_mThese 3 attributes Gini value is less than the Gini value of other 4 attributes, so decision tree generates attribute and only have selected in the attribute of provided 7 3, it is respectively as follows: FF, K and V_m。

For the beta pruning of tree, use the rear pruning method of cross validation, from training data, first mark off a subset use In setting up tree, then in remaining checking sub-collective estimation misclassification rate.Then it is repeated several times this stroke for different subsets Point, then the misclassification rate obtained is averaged, obtains about particular size tree for the friendship of the new performance not meeting data Fork checking is estimated.The tree producing minimum cross validation misclassification estimation is confirmed as the suitable size of final tree-model.

3rd step: set up decision-tree model, generates decision tree as shown in Figure 2, and the decision tree according to generating judges photovoltaic The fault type rule of assembly is as follows:

1) if FF < 0.485, it can be determined that for local shades fault；

2) if 0.485 < FF < 0.615, and V_m> 31.085, it can be determined that for normal work；

3) if 0.485 < FF < 0.615, and V_m< 31.085, and K < 0.135, it can be determined that for local shades fault；

4) if 0.485 < FF < 0.615, and V_m<31.085, and K>0.135, it can be determined that for abnormal degradation failure；

5) if FF>0.615, and K<0.195, it can be determined that for abnormal degradation failure；

6) if FF > 0.615, and K > 0.195, it can be determined that for normal work.

4th step: decision tree precise verification

Use classifying rules that test data are identified, utilize the effectiveness of the accuracy rate checking model of recognition result. Table 4 is shown that model accuracy situation

Table 4 model accuracy situation

Test result display decision-tree model test data accuracy reaches 98.32%, shows that this model is for detecting light Photovoltaic assembly fault type is the most easy.

The decision tree generated is write as if-then form by described step 3, in write single-chip microcomputer.

Use fault detect experimental provision as shown in Figure 3, in the case of intensity of illumination is higher, test 205 groups of experiments Gather the data of photovoltaic module；In the case of intensity of illumination is more weak, test 222 groups of data, judged by the decision tree generated The malfunction of photovoltaic module.Test result is the most as shown in table 5 and table 6.

Experimental result when table 5 intensity of illumination is higher

Experimental result when table 6 intensity of illumination is more weak

Test result indicate that, intensity of illumination higher and more weak time, the mould of photovoltaic fault diagnosis based on decision-tree model The judging nicety rate of type all more than 95%, shows that photovoltaic method for diagnosing faults based on decision-tree model is feasible and real , and decision-making judging nicety rate is high, rate of false alarm is relatively low.

Claims (7)

1. a photovoltaic module method for diagnosing faults based on decision-tree model, it is characterised in that comprise the following steps:

Step one: collection photovoltaic module data: open-circuit voltage U_oc, short circuit current I_sc, maximum power point voltage U_mWith electric current I_m, warp Cross data process to obtain: fill factor, curve factor FF, slope factor K and output current ratio I_m/I_sc；

Step 2: build photovoltaic module fault diagnosis model based on decision tree, respectively by training and test sample data The beta pruning choosing and process, contribute and set, the four processes setting up decision-tree model and decision tree precise verification are to build light Photovoltaic assembly fault diagnosis model；

Step 3: decision-making judges, by open-circuit voltage U_oc, short circuit current I_sc, maximum power point voltage U_m, electric current I_m, and fill Factor FF, slope factor K and output current ratio I_m/I_scImport to photovoltaic module fault diagnosis model and judge the fault of photovoltaic module State.

Photovoltaic module method for diagnosing faults based on decision-tree model the most according to claim 1, it is characterised in that: described In step one, four output parameters of collection photovoltaic module: open-circuit voltage U_oc, short circuit current I_sc, maximum power point voltage U_mWith Electric current I_m, process, through data, the fill factor, curve factor FF obtained and be shown below:

$FF=\frac{I_m{U}_m}{I_{sc}{U}_{oc}}---\left(1\right)$

Slope factor K be in assembly I-U curve maximum power point to the absolute value of open-circuit voltage point straight slope, its computing formula As follows:

$K=\frac{I_m}{U_{oc}-U_m}---\left(2\right)$

Photovoltaic module method for diagnosing faults based on decision-tree model the most according to claim 1, it is characterised in that: described In step 2, training is chosen with the data in processing links by testing or emulating acquisition with test sample data, tests or imitative Middle photovoltaic module is operated in normally respectively, in the case of local shades is the most aging with exception, divides data, described data A collection part is for simulation training, and another part is used for testing.

Photovoltaic module method for diagnosing faults based on decision-tree model the most according to claim 1, it is characterised in that: described Achievement link in step 2, carries out recurrence division by sample, and concrete partiting step is as follows:

First, an independent variable x is selected_i, then choose x_iOne value v_i, v_iN-dimensional space is divided into two parts, owning of a part Point all meets x_i≤v_i, the institute of another part the most all meets x_i＞ v_i, for discontinuous variable, the value of property value only has two Individual, i.e. it is equal to this value or is not equal to this value；

Then, two parts obtained above are chosen again by aforesaid operations an attribute and continues to divide, use Gini value conduct The standard divided: the community set in node is A={A₁,A₂,…,A_m, wherein, the attribute in set A is the light that sampling obtains Photovoltaic assembly data: A₁=U_oc, A₂=I_oc, A₃=U_m, A₄=I_m, A₅=FF, A₆=K, A₇=I_m/I_sc, so, m=7；Corresponding collection Close each attribute of A, i.e. open-circuit voltage U_oc, short circuit current I_sc, maximum power point voltage U_m, electric current I_m, fill factor, curve factor FF, slope because of Sub-K and output current ratio I_m/I_scObtain seven training sample set: S={s₁,s₂,…,s_n, the x in step in S correspondence_i, s_iFor The value of variable S, due to 774 groups of data of sampling altogether in experiment, so n=774；Corresponding each training sample set S, will be just Often, the sampled data in the case of local shades and abnormal aging three kinds is as category set: C={c₁,c₂,…,c_k, due to only Exist normal, local shades and abnormal aging three kinds of situations, so k=3；For training sample set S, by correspondence training sample Value T in set S, as an attribute of S, S is split into incoherent subset S by it₁,S₂,…,S_w, owing to using two Fork tree design, so, the v in step in T correspondence_iAnd w=2；

When dividing current attribute, Gini value is:

$Gini(S,T)=\underset{i=1}{\overset{w}{\&Sigma;}}\frac{\left|S_i\right|}{S}(1-\underset{j=1}{\overset{k}{\&Sigma;}}{p}^2(c_j,S_i\left)\right)---\left(3\right)$

Wherein, p (c_j,S_i) it is that sample set closes S_iIn belong to the ratio of jth class；|S_i| for subclass S_iGesture, to candidate attribute Every kind of Gini value that may divide on this attribute of each property calculation in collection A, finds the division of Gini value minimum as this Optimum division on attribute, the Gini value of optimum division in the most all candidate attribute, have minimum division Gini value Attribute becomes final test attribute；

Finally, according to the value of Split Attribute, obtaining decision tree branches, data set will be divided into multiple subset, for each Subtree recalculates the Gini value of each attribute, the like, until stopping when meeting following condition contributing: the sample number of node It is 1 or sample belongs to same class.

Photovoltaic module method for diagnosing faults based on decision-tree model the most according to claim 1, it is characterised in that: described The beta pruning of the tree in step 2, uses the rear pruning method of cross validation, first mark off from training data a subset for Set up tree, then in remaining checking sub-collective estimation misclassification rate, then this division be repeated several times for different subsets, Again the misclassification rate obtained is averaged, obtains about particular size tree, the intersection of the new performance not meeting data being tested Card is estimated, the tree producing minimum cross validation misclassification estimation is confirmed as the suitable size of final tree-model.

Photovoltaic module method for diagnosing faults based on decision-tree model the most according to claim 1, it is characterised in that: described Step 2 is set up decision-tree model, judges that according to the decision tree generated the fault type rule of photovoltaic module is as follows:

1) if FF < 0.485, then local shades fault it is judged as；

2) if 0.485 < FF < 0.615, and V_m> 31.085, then it is judged as normally working；

3) if 0.485 < FF < 0.615, and V_m< 31.085, and K < 0.135, then be judged as local shades fault；

4) if 0.485 < FF < 0.615, and V_m<31.085, and K>0.135, then be judged as abnormal degradation failure；

5) if FF>0.615, and K<0.195, then abnormal degradation failure it is judged as；

6) if FF > 0.615, and K > 0.195, then it is judged as normally working.

Photovoltaic module method for diagnosing faults based on decision-tree model the most according to claim 1, it is characterised in that: described In step 2, decision tree precise verification link uses classifying rules to be identified test data, utilizes the accurate of recognition result The effectiveness of rate checking model.