TW200929566A - Method for fault diagnosis of photovoltaic power generating system - Google Patents

Abstract

A method for fault diagnosis of photovoltaic (PV) power generating system is disclosed. The method is to construct a fault matter-element model for each atmosphere conditions of the PV power generating system bases on extension theory. And an extension neural network is constructed to cooperate the fault matter-element model. The extension neural network is inputted training samples to train the extension neural network. The extension neural network is employed to diagnose various fault categories of PV power generating system after the training is accomplished.

Description

200929566 IX. Description of the Invention: [Technical Field] The present invention relates to a method for fault diagnosis of a solar power generation system, and in particular to an extension neural network (ENN) architecture A method of identifying the type of failure of a solar panel. [Prior Art] Solar stencils may be deposited on the stencil due to dust, snow, deciduous or bird droppings, resulting in "shadowing failure", or contact corrosion due to contact with different metals, deterioration of packaging materials, and moisture infiltration. The formation of internal battery rust or lightning strikes and other "template failure", both will reduce the power generation of the solar power generation system, and the power generation efficiency is also greatly attenuated, resulting in instantaneous changes in its output power, the power system Bureau 4 or the overall causes such as: instantaneous voltage drop, voltage fluctuations and harmonics affect the quality of the power, the impact on the overall power network is self-evident. Since the power output signal generated by the solar power generation system is only voltage and current', when the fault occurs, if the condition of the surface failure of the solar cell template is to be observed by visual observation, it is highly likely that the detection is misjudged. And when faced with millions of gradual solar power generation, the number of solar cell templates is huge, and it is difficult to visually confirm the area where the failure occurred. At present, the solar cell template fault diagnosis technology has been proposed, and the utility of the detection cost is high and the application cost of the diagnostic system is limited, so that the practicality is greatly reduced. In addition, the high frequency reaction measurement method 200929566 (High frequency reaction measurement) and the time domain relectometry (TDR) are used for fault diagnosis. The diagnosis method is limited by the operator having sufficient solar power generation system. Expertise and the need to ensure the accuracy of the wiring length calculation, otherwise it is prone to misjudgment of the template fault area, so the application of this fault diagnosis technology is not very helpful for the convenience of system maintenance. SUMMARY OF THE INVENTION An object of the present invention is to provide a method for fault diagnosis of a solar power generation system for improving the above-mentioned practicality and accuracy of judgment, and to provide a fault diagnosis method which is simple in structure, quick in judgment, and accurate. Based on the above object, the present invention proposes a fault diagnosis technique for a solar power generation system, and the proposed fault diagnosis technique uses a fault diagnosis method of the extension neural network theory to discriminate the type of fault occurrence. Combined with extension theory and neural network technology, the extension neural network fault diagnosis algorithm is used to identify faults. According to a preferred embodiment of the present invention, a fault diagnosis method for a solar power generation system includes expressing a fault condition of a solar power generation system as a faulty extension matter element model through a extension theory and a fault characteristic parameter of a solar power generation system. The type, wherein the faulty extension matter element model comprises a plurality of child matter element models, and each of the child matter element models corresponds to one of the fault categories. Firstly, an extension-like neural network with fault-extension matter-element model is established. 200929566: The network of the extension-type neural network has a parallel-like tube with a parallel-like tube, and the learning ability' and the extension theory are used. It is composed of the characteristics of the processing of the value of the associated function. The extension __ supervised learning rule can accept the characteristics of continuous input and discrete rotation: and extremely; In the present embodiment, the extension-type neural network includes a wheel-man layer and a layer-input layer that includes a plurality of input layer nodes, and the core layer includes a plurality of Γ: out:: matter element corresponding An output layer node, there are two connection weight values between each input layer node and == point, and each round-off layer node pair: one of the =::r barrier categories, each-transmission, point corresponding to a known fault Training of categories to train the extensional neural network for training 'If the data is mixed with the target value

Integer = the weight value of the link, and then train until it is close to the output ^ (according to the set recognition rate). In this embodiment, the training mode adopted is to input the training sample of each fault type, and the calculation of the extension distance (EXtensi() n distanee, rib) is used to estimate whether the relative weight value is correct, and only Re-input the test sample to the trained extended extension neural network according to the weighted value of the error, and determine the fault category to which the input sample belongs, so that each output layer node corresponds to the fault classification, only There will be an output layer node to indicate the fault category represented by the node. In this embodiment, according to the input test sample, the distance to the per-class cluster is calculated, and the minimum extension distance of 200929566 is the cluster category to which the test sample belongs. The fault diagnosis method according to a preferred embodiment of the present invention can be completed only when some parameters are changed when the capacity of the solar power generation system is expanded, thereby reducing the fault diagnosis update schedule and increasing the maintenance convenience of the solar power generation system. Sex and expandability. In addition, since the proposed technique is different from the learning mechanism of the traditional neural network, the calculation process and speed of the diagnosis are quite simple and rapid. [Embodiment] At present, most of the solar power generation systems in operation mainly use solar panels as the basic power generation unit. According to the power required by the load, the solar modules are connected in series to increase the output voltage and increase the output current in parallel. A solar power generation array with a large output power. In this embodiment, each piece of output is 75W' 17V, four solar panels of 4.4A are connected in series to form an array branch 'and parallel array of 10 branches to form a 3kw solar power generation array, and its rated output voltage At 68V, the rated output current is 44a, which will produce an output power of up to 3000W. BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a flow chart showing a method for fault diagnosis of a solar power generation system according to a preferred embodiment of the present invention. The method for fault diagnosis of a solar power generation system includes the following steps: The extension theory first establishes the faulty extension 7L model of the solar power generation system. The faulty (four) metamodel includes a plurality of sub-objective models, each of which corresponds to a fault category, as shown in step 210. Then, the extension-type neural network with the fault-extension matter-element model is established. The road is as shown in step 220. Re-age λ &amp; can perform #_经_(四)二故__ training data, rate (such as t program, until it meets the set identification input data to be tested, tUe 23G. After the training is completed, the fault category can be performed. The recognition/recognition of Yuancheng's extension-type neural network architecture is shown in step 2. The reading of the fault category to which the data to be tested belongs, such as the step ❹ In the inventive example, the proposed solar power system 0-break method The advance-steps are detailed as follows: 咖) 1) Establish a fault-extension matter element model of the solar power generation system (step P due to the generation of the solar voltage, its output voltage, current and; in different /Ja degrees and sunshine intensity) The following are not the same, so if you want to turn the 5 township fault category, you must establish a fault extension matter element model for each atmospheric interval. In this example, according to the average temperature and relative conditions of the Taiwan region, the atmosphere will be The conditions are divided into seven types of atmospheric intervals. Please refer to Table 1. Table 7: Classification of atmospheric intervals (W/m2) Temperature (° C ) Area A 200~400 5~10 Area B 401 -500 11 ~ 15 C area 501 ~ 600 16 ~ 18 601 ~ 700 19 to 20 Eg 701-800 21 to 25 G area 801 to 900 16-30 901-1000 31 to 40 The first section of the atmospheric section C~G is the atmospheric condition of most solar power generation systems. When the power generation system is operating in this section, fault diagnosis will be performed. However, due to insufficient strength of the a and b sections, the electrical signal output by the solar power generation system is small, which is not suitable as data for fault diagnosis, so no fault diagnosis will be performed. In the present embodiment, a solar power panel of 75 W is used to form a series of 4 kW and a 3 kW solar power generation system in the atmospheric zone G zone as an example, and a faulty extension material 兀 model is established. The fault characteristic parameters (power, money and current output by the solar power generation array) are classified into six categories (the fault element model knife is /, the individual object TG _ type) The representations of the six fault categories and their symbols are as follows: 1 Normal operation of the eight-knot system PF2: Each of the two branch paths of a template fault branch occurs in any of the ten series branches A template failure occurs in each of the ten-series spurs in the slab-fault series. Ten series two = each occurrence in the path - template failure representation = the matter element defined in (4) can be given by formula (1) with R = (N, C, V): _v rf, Cl, V γ 2 • · · = C2, V2 RP L. Fn ^ • • · k V m - , factory &quot;, Xiwei matter element, # is a feature vector C = [c,c θ' The name of ..., u 2' ···, c«] and the amount corresponding to the feature vector 200929566 Vector V 1 2, ·.·, (I, and '=(Ν 多维 物 物 兔 兔 兔 兔 兔 兔2 is the number of features of the child element. The heart, the number 'W is the sub-objective element. The above six fault categories are based on the matter-element model, and 筮) is used to establish the corresponding faulty power generation system. In this embodiment, the solar eigenvector is used. (4) and the current as the sub-element model of each sub-parameter PF! and Fl = * Fu,,, <62.8, 67.0>, L·, (40.6, 43.4) pmt, (2558.0, 2904.9) pf2 and Corpse 2 = "PF2&gt; (58.8, 63.6) Ln (38.1, 41.2) p post, <2244.5, 2621.9>, pf3 and η =, PF3, v^, (53.3, 58.5) L·, (34.4, 37.8) P 〇ut, (1838.7, 2218.2) » pf4 i?F4 PF4, v^, (46.8,51.9) ' L〇(30.3,33.6) .(1417.9,1742.9)^ » PFs ^F5 = PF5, vouP (40.2,44.7 )] W» (26.0, 28.9) (1047.5, 1295.5) &gt; 11 \pf6, (33.6, 37.4) PF6 ~= Ln (21.7,24.2) p^, (732.9,906.3) »

Stupid 200929566 The second table of things (~) represents the sub-objects of the six fault forms of the solar power generation array, and, for the fault set ^ respectively represent the "fault category, each sub-element uses three feature elements. f is the electric dust (v." output from the solar power generation array, the output turbulence ^", and the power (匕) of the output. The magnitude of each sub-element model is the classical domain of its corresponding feature. The classical domain is based on the maximum and minimum values of the intensity and temperature of the g-zone in the atmospheric interval, and the six voltages, currents, and powers generated by the solar power generation array when six types of faults occur. The maximum and minimum ranges are based on. (2) Establishing an extension-like neural network (step): According to FIG. 2, a diagram of a pullable neural network according to a preferred embodiment of the present invention is illustrated. The architecture diagram. The extension class neural network _ includes a layer, and the _output layer 42. The input layer 41 〇 includes a plurality of layers, ie, 4 U, for receiving the input feature samples, that is, the value of the matter element features. (smart to private), and set and levy the sample童#山思一A集 produces input specials-rounds=multiple matter elements, each element-to-object pair is out of layer (pointing ~〇421. in the turn-out layer 42〇一土—round-out node 421 The change occurs, and the identification result of the borrower is obtained. (4) “The most ^^ value of the classical domain of the characteristic sample 12 200929566 meta-characteristics-the weight value (the minimum value of the classical domain of the X feature. The number of input layer nodes is the sub-primitive element) In the embodiment, the number of features of the wheel A s ~ (1~m), the number of nodes in the input layer is 3 (three features:). ^ + s ~ P is the sub-element model using ^ output layer nodes For the established work-a-(1~η), the total number of models of the sub-objects of Benshini, such as mat, is 6 (that is, six faults are shown in the second table). Therefore, the network architecture A total of 9 extension-type neural crests in 9 embodiments are required, ie, 9 points and 36 connection numbers. In this embodiment, the weight value 43 is the fault category 耠a-like + &amp; 耵 initial sigh value is known Affiliation (such as the maximum and minimum input layer sections of the classic domain of the 2nd element represent the minimum and maximum weights connected to the jth: point and the output layer nodes respectively 230) The supervised learning training program of the neural network (step sm of the extensional neural network) represents the adjustment of the input value of the learning habits, so that the law is supervised learning, which learns or reaches the target (4) The extension type neural network has a better recognition rate, and the output characteristic a with the same target value must first define the relevant variables. For the learning S program, the training sample must be set to | the total number. ,..., where 'the characteristics of the training sample sample her Weng ^, Ν, ~, ..., ~} ' m represents the evaluation of the fault category of the extension of the two samples. For the number of differences between the 1 stone core extension neural network' The total error ratio p N Ά γ = - gas can be expressed as the following difference, ° accuracy, ^ is the overall estimation error (2) 13 200929566 Please refer to Figure 3, which shows the flow chart of the network. The training in Figure 1 can include the following types of steps: _ network supervised learning training program to the initial Sr, 1 and = connected between the input layer node and the output layer node. The form of the extension meta-model is as follows: "乂, A, 匕.

Rk 尧 = 1, 2, K2 c3, 匕

In equation (3), c ^ XT J , the domain value and the range of the classical domain can be determined by the training sample. Where (4) (5) w&gt;9' (i = l-, Np) /, medium, % is the minimum weight of the y'th feature corresponding to the & and π is the first feature corresponding to the third The weight of each category is the maximum. In the present embodiment, the magnitude value corresponding to each feature in the child matter model of each category in the second table is used as the initial weight value connected between each input layer node and each output layer node. Step 222. Calculate an initial cluster center of each initial weight value. Z* ={~,~,.··,~} (6) +^)/2 (7) k = l 2,..., „ ; j = l&gt;2,.··,m (8) 14 200929566 Step 223: Read the z-th training sample and its classification number. ~ {xn» , pen (9) Step 224. Apply the extension distance (ED) to calculate the distance between the training sample γ and the first classification cluster. , that is, calculate the extension distance between the training sample and each classification cluster. The mathematical expression is as follows:

EDik = Y (10) Μ A: = 1, 2, .... Please refer to Fig. 4' for a schematic diagram of the extension distance (ED) proposed in the present embodiment. The extension distance (ED) can be used to represent the distance between the point and the range <V, />, where the center of the weight of the cluster is the weight cluster minimum and β is the weight cluster maximum. It can be seen from Fig. 4 that the extension distance can be different from the numerical value difference, and thus the difference in the sensitivity value is generated. In general, if the range of matter elements is large, it means that the training materials are more widely blurred, so they are expressed in the distance.

The sensitivity of the calculation is relatively low; conversely, if the range of the matter element is small, the representation of the required data sample is more accurate, so the sensitivity of the distance δ can be expressed. +Step 225: After calculating and comparing, find 々•, so that young;=mjn{£^}, if 々·=;; (ie, the input feature sample p is the same as the fault category 所属 to which it belongs), the calculation program jumps to step 227 Otherwise, proceed to step 226. Step 226: Adjust and update the weight values of the slave and the cluster as follows: U) Update the weight of the center of the p and the transfer, that is, update the input feature sample itself should belong to the classification cluster (corpse - The weight center value of the cluster is recognized and the weight center value of the classification cluster (recognition cluster) corresponding to the 200929566 new input feature sample misjudgment. (11) (12)

(b) Update the weight value of the household-money set, and (4) update the input characteristics. The sample book S should be the weight value of the classification cluster (Μ cluster) and the weighted value of the classified cluster cluster corresponding to the error of the updated input feature sample. &lt;(new) = &lt;(oW)+7(^-2^&gt;) (13)

wL.(Kew)= kj WUh.{new)= k J ~^χυ yL(old) \ zU[old), (14) 7 of the formulas (11) to (14) represent the learning rate (Learning rate) . In this step, 'only the clusters and the weights of *· are adjusted during the learning process' and their (iv) weights do not change. Due to the above characteristics, the extension-like neural network has advantages over other neural network architectures and has high adaptability in new application fields. ^: Refer to Figure 5a and Figure 5, which are diagrams showing the results of adjustments made in the two clusters. From the state test data and the group 隼... (4) The original f material cut Γ (take) and the original state measurement B's extension distance (%) produces a significant change. And for the adjusted test data and the cluster A's extension distance' is the post-adjustment test (4) Taixus ^Gongko and cluster B extension distance), so the training sample (10) belongs to the cluster changed from cluster Α(ζ·α) to Cluster Μ%). 16 200929566 Step 227: Repeat steps 223 to 226 to calculate the course, all the training samples have been classified, and end a learning batch (Epoch). Step 228: If the classification procedure has reached convergence, or If the total error rate (out) has reached the preset target value, the calculation program is stopped, otherwise it returns to step 223. (4) Fault diagnosis calculation process of extension type neural network (step 24〇)

Please refer to Figure 6, which shows the flow chart for troubleshooting according to the second time. After the extension type neural network completes the learning process, it can be identified or classified, and its calculation program includes: Step 241 · Read the weight value matrix of the pullable neural network after the training element is completed, and the training is completed. The weight value of the extension-like neural network is used as the weight value of the extension-type neural network for identification. Step 242: Using formula (7), calculate the cluster weight center value of each classification cluster according to the read interest_step 243: Read the test sample xtm} to be diagnosed (15)~ Step 244: Apply the definition of the extension distance (ed) proposed by (1G), and calculate the distance between the test sample and each classification cluster. Step 245. Find the classification cluster to which the test sample belongs ((1) make the thief =}, and set its corresponding output section = ι, by which the sample belongs to the cluster _ 1 all test samples are classified and the operation program is stopped, otherwise Going back to step 243. In the present embodiment, 'the data of the atmospheric interval G area 36 is used as the training 17 200929566 sample, and 6000 data is used as the test sample. The output can be based on the electric «, current and output power diagnosis method: identification category Six kinds, each only has a -identification knot; = word respectively represents six fault categories, and then compared with the target value. In order to verify the feasibility of the extension-type neurological fault diagnosis method of the present invention, the atmospheric interval G area is to be treated. Test data as a test. Table 3

8 test data arbitrarily selected in the district, the table "for the G 丫 &lt; sunshine intensity, temperature and = branch number is the sun-power system of the health-like g parameter, and the output value of the electric dust, the positive array, and Also as the result of the fault diagnosis of the object to be tested, the meter is compared with the fault categories listed in the test chart of the third table. The diagnosis method can accurately identify the fault category. The field light (10) will test the data of the atmospheric interval C~G as shown in the table. The identification results shown in the sixth table can also prove that the fault diagnosis method is in the 5 atmosphere interval, and the six fault categories can be individually identified. The table 7 shows the comparison between the proposed extension neural network and other obstacle diagnosis methods. It can be seen from the seventh table that the structure of the extension type god=road fault reading method is simpler than the inverse transmission type neural algorithm. The neural network architecture requires only 9 nodes and 36 links. In addition, the extension-based neural network only trains to adjust the core and lower weight values of the existing links in the learning phase, so when the amount of training data is large or the data is added When you can show the extension god The adaptive characteristics of the network. The comparison between the number of iterations and the accuracy of the identification can be seen that the extension time of the extension-type neural network 18 200929566 is shorter, and the rate of test data is 6 On the other hand, only the gentleman &lt; Gu Yunjiao 5 pen error, so the identification accuracy is higher than other fault diagnosis broad methods. In addition, the optimized extension neural network optimization weight value upper and lower limits range It can be determined by the expert experience value before training, so it is easy to highlight the fault category to which the identification output belongs after training.

Table 3 Test data of atmospheric interval G Test item Sunshine intensity (W/m2) Temperature (°C) Fault _ Switching voltage (V) Current (A) Power (W) Fault category 1 917 37 4 41.0 26.5 1086.5 pf5 2 983 31 0 66.6 43.1 2878.4 PFi ___3__ 950 33 4 49.3 31.9 1572.6 pf5 4 928 38 5 34.7 22.4 777.3 pf6 5 931 32 2 54.9 35.5 1948.9 pf3 6 998 36 3 51.7 33.5 1731.9 pf4 7 946 34 1 61.3 39.7 2433.6 pf2 8 905 35 3 47.1 30.4 1431.8 pf4 19 200929566

The fault diagnosis result of the atmospheric interval G of the fourth table - test the output value of the first sample corresponding to each fault category. The atmospheric identification result item S Oil 〇 〇 〇α 〇i4 〇i5 〇i6 Interval category 1 0 0 0 0 1 0 G pf5 2 1 0 0 0 0 0 G PFi 3 0 0 0 0 1 0 G pf5 4 0 0 0 0 0 1 G pf6 5 0 0 1 0 0 0 G pf3 6 0 0 0 1 0 0 G pf4 7 0 1 0 0 0 0 G pf2 8 0 0 0 1 0 0 G pf4

Table 5 Atmospheric interval C~G Test data Test item reference intensity (W/m2) Temperature (°C) Fault switch number voltage (V) Current (A) Power (W) Atmospheric interval fault category 1 1000 37 3 51.9 33.6 1742.3 G pf4 2 850 26 0 61.8 40.0 2476.7 F PFi 3 680 19 4 30.3 19.6 594.9 D pf5 4 510 17 2 30.1 19.5 588.6 C pf3 5 770 23 1 51.1 33.1 1692.8 E pf2 6 570 16 5 21.3 13.8 295.0 C pf6 6 Table Atmospheric Interval C~G Fault Diagnosis Results 2009 200929566 昙The third sample corresponds to the output value of each fault category Test 〇,7 〇i2 〇u 〇(4 Oil 〇ie 1 0 0 0 1 0 0 2 1 0 0 0 0 0 3 0 0 0 0 1 o' 4 0 0 1 0 0 0 5 0 1 0 0 0 0 6 0 0 0 0 0 1 Atmospheric interval identification result category c DEFG 0 ----— 0 0 0 1 Pf4 0 0 0 1 0 PF, 1 0 1 0 0 0 pf5 1 0 0 0 0 pf3 0 -— 0 1 0 0 pf2 1 0 丄丄0 pf6

Table 7: Identification and comparison results under different fault diagnosis algorithms (G-zone in the atmospheric interval) Number of generations of algorithm generation, number of required data Cs/Ns Identification accuracy (%) 叮拓神__ via the network 10 36 5995/6000 99.917 Inverted-transfer-like neural network 1500 6000 5844/6000 97.401 K-Means classification grouping method ί£Ί ~ 500 6000 4511/6000 75.183 fs. Identifying the correct number of samples 丄 Total test 檨 too eye-catching reference 7 FIG. 2 is a learning curve diagram of an extension type neural network for identification according to a preferred embodiment of the present invention. Figure 7 is a learning curve for identifying the extensional neural pathways with a learning rate of 0.2. The total error rate shown in the figure is known as 'because the initial weight of the extensional neural network 21 200929566 reset has been considered The maximum and minimum characteristic values of each fault identification category, so the initial value of the total error rate is much smaller than that of the general neural network, so its convergence is faster and more accurate than other methods. Although the present invention has been described above in terms of a preferred embodiment, it is not intended to limit the invention, and various modifications and refinements can be made without departing from the spirit and scope of the invention. Therefore, the scope of the invention is defined by the scope of the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS In order to make the above and other objects, features, advantages and embodiments of the present invention more comprehensible, the detailed description of the drawings is as follows: Figure 1 shows a A flow chart of a method for fault diagnosis of a solar power generation system of a preferred embodiment. 2 is a block diagram showing an extension type neural network in accordance with a preferred embodiment of the present invention. Figure 3 is a flow chart showing the training of the extension-like neural network in the figure. Figure 4 is a schematic illustration of an extendable distance (ED) in accordance with a preferred embodiment of the present invention. Figure 5a is a schematic diagram of the two cluster weights before adjustment in the learning program. Figure 5b is a schematic diagram of the adjustment of the two cluster weights in the learning program. Figure 6 is drawn according to the first! The process of troubleshooting in the figure 22 200929566 Figure. FIG. 7 is a learning curve diagram 221 to 228 for identifying an extension-like neural network according to a preferred embodiment of the present invention: Step 400: Extension-type neural network 411: Input layer node 421: Output layer node

[Description of Main Component Symbols] 210 to 240: Steps 241 to 245: Step 410: Input Layer 420: Output Layer 430: Weight Value

twenty three

Claims (1)

200929566 X. Patent application scope: 1. A land-based I e &amp; barrier: breaking method for diagnosing the fault condition of a solar photovoltaic system, the fault diagnosis The method comprises: (= according to the matter-element model defined in the extension theory and a plurality of fault characteristic parameters for the too-first power generation system--fault extension matter element model, wherein the fault-liftable matter element model is used Indicates the type of failure of the solar power generation system; (1) Establishing-cooperating with the fault may include an extension-like neural network of a matter-element model, wherein the extension-type neural network includes an input layer and an output layer, the input layer includes a plurality of round-trip nodes, the output layer Including a plurality of objects 每 'every material element corresponding _ output layer node, there are two connection weight values between each input layer node and each output layer node, and each output layer node corresponds to solar power generation (four) _ Type-the input layer node corresponds to the parent-fault characteristic parameter; (〇 training the extension-type neural network, inputting a training data of a known fault type to train the extension-type neural network And () using the trained extension-type neural network to diagnose the failure condition of the solar power generation system. For example, the solar power generation system described in the first paragraph of the patent, the fault diagnosis method of the system' a) includes: establishing a plurality of atmospheric intervals according to the average temperature and relative sunshine conditions of the area in which the solar power generation system is located, and distinguishing the atmospheric conditions into the atmospheric intervals; and 24 200929566 A fault extension matter element model corresponding to each atmospheric interval is established, and a corresponding fault extension matter element model is established according to the fault condition of the solar power generation system in each of the atmospheric intervals. The failure method of the solar power generation system according to the above, wherein each of the faulty extension matter elements of the atmospheric interval comprises a plurality of sub-element models, each sub-element model corresponding to a classification cluster, and each of the classification clusters is The fault diagnosis method of the solar power generation system in the faulty condition of the solar power generation system. 4. The fault diagnosis method of the solar power generation system according to the item or the third item of the whistle patent scope, wherein the (4) The matter-element model defined in the theory is ^ = (N, C, V) = h Rf2 ,, ~v C2, V2 *·· ··. - ~, K :中为-Multidimensional matter element '# is the name of the thing (R), ", has the feature vector C=[C&quot; C2,...,C«] and the feature number corresponding to the feature vector, the child 7L," is The total number of sub-objects, m is the per-sub-element 25 200929566, the voltage output first, the current output by the solar power generation system, and the power output by the solar power generation system. 6. The method for fault diagnosis of a solar power generation system as described in the patent scope of the invention in June is characterized in that the faulty extension matter element model includes a plurality of features, each of which corresponds to one of the fault characteristic parameters. .鬌7· The β-break method of the solar power generation system described in claim 6 wherein the number of the round-trip nodes is the characteristic number of the fault-extension matter element model, and the number of the output layer nodes is The failure-prolongable matter element model of the sub-Ning Yin search scope of the third method of the fault diagnosis method, wherein the step (1) comprises: 9 (Cl) identifies all the input layers connected to the extension-type neural network The weight value of the layer node 'where the initial value of the weight value is the characteristic element === yuan Γ 'per-class cluster, the maximum value and the minimum value of the field 1; knot η: count per-weight The weighted center value of the value is the average value of the weighted value of the λ 卩 point and the output layer node. For the ==::: and its classification number 'the training samples are all distances; ((4)〇 The training sample and each of the classification clusters - extension 26 200929566 (c5) find the minimum value of the extension distance, so that the output value of the output layer node corresponding to the classification cluster is 丨, whereby the solar power generation The fault category of the system is the fault corresponding to the output layer node. (c6) when the fault category represented by the output layer node is different from the fault category to which the training sample belongs, updating the weight value and the weight center value; and (c7) calculating a total error rate, the total The error rate is the ratio of the number of errors to the total number of training samples. When the oil is rented off, the training error ends when the error rate is less than a preset value; otherwise, the process returns to step (c3). The method for diagnosing solar radiation according to item 8 of the patent scope, wherein the mathematical expression of the extension distance is Μ k = l, 2, .. ζ, middle and young* • for the training sample 舆久八s 隹 隹 隹 is the extension distance of the _ training sample and its classification number, &lt; the center value of the weight of the classification, 'the characteristic corresponds to the minimum moxibustion, π is the weight of the second classification. The maximum weight of the two-phase corresponding phase classification ^Beibei, "for the total number of classifications. 27 200929566 The weight center value of an error output classification cluster is adjusted, 纟t The adjustment of the weight center value of the classification cluster to which the training sample belongs can be expressed As follows·· Z ^ =Zp/+V(x^ Z o Ld' PJ · 甘φ , -new v. , 'A &amp; is the weighted center value of the classification cluster to which the training sample belongs after an update, 5, · is the weight center value of the classification cluster to which the training sample belongs before updating; For the training rate set; 4 is the training sample; the adjustment of the weight center value of the error output classification cluster can be expressed as follows: old /, medium ' is the weighted center value of the error output classification cluster after an update; zg The weighted center value of the classification cluster is output for a pre-update error; ^ is the learning rate of the set =; X# is the training sample. 11. The method for fault diagnosis of a solar power generation system according to claim 8 or claim 1, wherein in the step (c6), the method of updating the weight value is only for the classification cluster to which the training sample belongs. The weight value and the weight value of the error output classification cluster are adjusted. The adjustment method of the weight value of the classification cluster to which the training sample belongs may be expressed as follows: +η(4 •Ζ U{old), Pj. The minimum weight value of the classification cluster to which the training sample belongs is the weight minimum value of the classification cluster to which the training sample belongs before updating, and is the weight maximum value of the classification cluster to which the training sample belongs after being updated; The weight of the cluster to which the sample belongs 28 200929566 The maximum value; 7 is the set learning rate; 4 is the training sample; the error output cluster weight value can be adjusted as follows: L(new) -4Γ)) jL{ Ne\\f) ', k'j is an update after the error output classification cluster has a very small weight · ynL^old) 'The heart is—the minimum value of the weight of the error output classification cluster before the update; h is the weight maximum of the error output classification cluster after an update; you ^. The maximum value of the weight of the cluster is just updated for the error; ^ is the learning rate set for the test; &lt; is the training sample. 12. The method of diagnosis according to the solar photovoltaic power generation system of claim n, wherein in the step (e3), the training sample comprises a voltage output by the solar power generation system, and the solar power generation system The output current and the power output by the solar power generation system. 13. The method of fault diagnosis of a solar power generation system according to the invention of claim 8 or 8, wherein the step (d) comprises: (d1) obtaining a weight of the extension-type neural network of the training completion Value, the weight value of the trained extension type neural network is used as the weight value of the extension type neural network for identification; (d2) calculating the weight center value of each classification cluster according to the obtained weight value; (d3) reading the test sample to be diagnosed; 29 200929566 (d4) calculating the extension distance of the test sample and each classification cluster; and (d5) finding a minimum set of extension distances to classify The output layer node corresponding to the cluster has an output value of 1, which displays the category of the cluster to which the test sample belongs. XI. Schema: Secondary page 30