Low power factor crux diagnosis method by processing power curve data

Abstract

The invention aims to provide a low-power-factor crux diagnosis method by processing power curve data. The method comprises the following steps: constructing a low-power-factor crux Bayesian network; designing a characteristic quantity; feature quantity combination: the variables are processed and combined into variable tuples which are respectively related to calculation of transition probabilities of different nodes in the Bayesian network; the estimation method of the transition probability of each node of the transition probability neural network and the Bayesian network uses an algorithm of the neural network, the input of the neural network is an input tuple related to the transition probability of the corresponding node after a power curve of a diagnosis object is subjected to feature extraction and processing recombination, and the output of the neural network is a transition probability estimation value of the diagnosis object at the node; and outputting a crux diagnosis result. The low-power-factor diagnosis method has the beneficial effects that the low-power-factor diagnosis technical scheme provided by the invention does not need manual field equipment detection, and the manual workload and the probability of occurrence of electric shock accidents are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of low power factor problem diagnosis, and in particular to a low power factor problem diagnosis method by processing power curve data. Background Art

[0002] In the traditional power system field, low power factor problems are usually diagnosed by detecting the size and changes of power factor. Traditional power factor diagnosis methods usually use electrical parameters such as current and voltage to infer the main problems and reasons of power factor below the standard based on empirical rules and thresholds, relying on on-site equipment surveys and manual judgment. Extracting power curve features and using artificial intelligence technology of neural networks to diagnose low power factor problems is an emerging technical method in the field of low power factor diagnosis. By analyzing the power curve and combining it with neural networks, early warning of wind turbine faults can be achieved. At the same time, by performing time domain feature statistics on the output voltage curve of the DC-DC converter and using BP neural network for data processing, the classification task of multiple faults is achieved. These technologies provide strong support for the diagnosis of low power factor problems.

[0003] In general, extracting power curve features based on neural networks to diagnose low power factor problems is an effective technical method that can make full use of power curve data and combine advanced neural network algorithms to achieve accurate diagnosis and early warning of low power factor problems. However, the existing power factor diagnosis methods have the following defects:

[0004] (1) Existing power factor diagnosis methods are basically based on empirical rules and thresholds to infer the main problems and reasons why the power factor is lower than the standard, relying on on-site equipment surveys and manual judgment.

[0005] (2) On-site analysis is inefficient and requires various power outage detections, making it impossible to perform online, rapid, and batch diagnosis.

[0006] (3) Empirical rules cannot fully tap the value of data, the accuracy of analysis is low, and they rely on human empirical rules, which have the disadvantages of instability and inefficiency. Summary of the invention

[0007] The purpose of this application is to provide a method for diagnosing low power factor problems by processing power curve data, extracting the active-reactive power curve characteristic of the diagnosis object by design method, and diagnosing low power factor problems. On the one hand, there is no need to shut down the equipment on site for detection, and online rapid diagnosis and online real-time diagnosis can be achieved through the monitored power curve data; on the other hand, the value of data is fully explored, and intelligent methods are used to get rid of the dependence on people and experience rules in diagnosis, so as to achieve efficient and accurate diagnosis results.

[0008] To achieve the purpose of the present invention, the technical solution provided by the present invention is specifically as follows:

[0009] A method for diagnosing low power factor problems by processing power curve data comprises the following steps:

[0010] Step 1: Construct a Bayesian network for low power factor problems. By analyzing the Bayesian network for low power factor problems, the occurrence probability of each problem of the diagnosis object can be calculated.

[0011] Step 2: Designing feature quantities, and extracting waveform features of the collected active power curve and reactive power curve of the diagnostic object by designing feature quantities;

[0012] Step 3: Feature quantity combination, process the above variables and combine them into variable tuples that are related to calculating the transition probability of different nodes in the Bayesian network;

[0013] Step 4: Transfer probability neural network. The estimation method of the transfer probability of each node in the Bayesian network uses the neural network algorithm. The input of the neural network is the input tuple related to the corresponding node transfer probability after feature extraction and processing of the power curve of the diagnosis object, and the output is the estimated value of the transfer probability of the diagnosis object at the node.

[0014] Step 5: Output of the key diagnosis results. After extracting the characteristic quantities of the active-reactive power curve of the diagnosed object and combining them into different input tuples and inputting them into the neural network of each judgment node in the decision tree, the network will output the probabilities of the two types of diagnosis results of the corresponding nodes.

[0015] Wherein, the step 1 specifically includes the following:

[0016] Step 11: Categorize the causes of low power factor according to the different causes and solutions.

[0017] Step 12: By analyzing the dependency and conditional independence between the various key points, the key points are divided into the following binary tree structure in the form of a decision tree. There are seven judgment nodes in the decision tree. Each judgment node can judge the transfer result node of the next layer according to the input, where the transfer result node is another judgment node or a key leaf node;

[0018] Step 13: Convert the dependency relationship between the nodes in the above binary tree into a probabilistic dependency relationship, and convert it into a Bayesian network in the form of a probability graph;

[0019] Step 14: By analyzing the Bayesian network of the low power factor problem, the occurrence probability of each problem of the diagnosis object can be calculated, and the formula is as follows:

[0020] P 症结一=P(A_f)×P(A_B_f)×P(B_D_t)

[0021] P 症结二 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_f)×P(E_G_f)

[0022] P 症结三 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_f)×P(E_G_t)

[0023] P 症结四 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_t)×P(E_F_f)

[0024] P 症结五 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_t)×P(E_F_t)

[0025] P 症结六 =P(A_f)×P(A_B_t)×P(B_C_t)

[0026] P 症结七 =P(A_f)×P(A_B_t)×P(B_C_f)

[0027] P 症结八 =P(A_t)

[0028] Among them, in the process of probability calculation, the transition probability values ​​of nodes A, B, C, D, E, F, and G in the Bayesian network are needed.

[0029] Among them, in step 11, the low power factor problems are classified into the following categories: (1) due to the failure of the compensation controller, the capacitor group is not put into use according to the rules, causing excessive reverse reactive power; (2) there are low power factor devices such as no-load rotating motors, and the compensation controller does not work or the compensation element fails as a whole; (3) there are many low power factor devices such as no-load rotating motors, and the compensation capacity of the compensation element is insufficient; (4) the compensation gradient of the compensation element is large, and it is not compatible with the load characteristics of the electrical equipment, resulting in over-compensation or under-compensation; (5) when the transformer is long-term no-load or light-loaded, the low-voltage side compensation control strategy cannot compensate for the reactive power loss of the transformer; (6) The compensation control strategy is not compatible with renewable energy power generation, and the power factor during power generation should be further improved; (7) The compensation controller has no effect when renewable energy is generating electricity; (8) The power factor is normal and there is no abnormality.

[0030] Wherein, the step 2 specifically includes the following:

[0031] Step 21: Define variables;

[0032] Step 22: Define the functions needed for feature extraction;

[0033] Step 23: feature extraction;

[0034] Step 24: Feature quantity combination.

[0035] Wherein, in step 21, the variables include p t , active power at time t; q t Reactive power at time t; S, rated capacity; T, the time series of the power curve, where T total Contains all moments of the power curve.

[0036] In step 22, the function includes: (1) arctan(x), cosine(x), covariance conv(X, Y) of X and Y sequences, distribution variance var(X), mean(X) of sequence x, smaller value min([x, y]) of x and y, minimum value min(X) of sequence X, number of elements count(X) of sequence X, and absolute value abs(x) of x; (2) defining an evaluation function pf_evaluation(q, p) that inputs active power and reactive power values ​​at the same time or active power and reactive power values ​​at the same time period and outputs a corresponding power factor level; (3) defining a correlation coefficient function corrcoef(x, y) that characterizes the correlation between changes in the x sequence and the y sequence; and (4) defining a threshold power function threshold(X, S, k) for calculating the active power sequence X with a rated capacity of S.

[0037] Wherein, in said step 23, said characteristic quantities include: (1) A1, the evaluation function value of the total power factor for billing; (2) B1, the evaluation function value of the total power factor during the electricity consumption period; (3) B2, the proportion of power generation in the total power consumption and power generation; (4) C1, the evaluation function value of the total power factor during the power generation period; (5) C2, the correlation coefficient between the absolute value sequence of active power during the power generation period and the evaluation function value sequence of the power factor at each moment; (6) D1, the evaluation function value of the total power factor during the power consumption period when reactive power is fed back to the power grid; (7) D2, the evaluation function value of the total power factor during the power consumption period when reactive power is consumed; (8) D3, the ratio of the total reactive power fed back to the power grid during the power consumption period to the total reactive power during the metered power consumption period :(9)D4, distribution standard deviation of reactive power fed back to the grid during the power consumption period:(10)E1, evaluation function value of total power factor during no-load period during the power consumption period:(11)E2, evaluation function value of total power factor during non-no-load period during the power consumption period:(12)E3, evaluation function value of total power factor during non-no-load period after removing active power and reactive power loss of transformer during the power consumption period:(13)E4, evaluation function value of total power factor during no-load period after removing active power and reactive power loss of transformer during the power consumption period:(14)E5, correlation coefficient between power factor evaluation function value sequence and active power sequence during no-load period during the power consumption period:(15)E6, evaluation function value sequence of power factor evaluation function value and active power sequence during non-no-load period during the power consumption period Correlation coefficient of power sequence: (16) E7, the evaluation function value of the total power factor at the moment of conversion from reactive power consumption to reactive power return or from reactive power return to reactive power consumption in the power consumption period: (17) E8, the proportion of the number of moments of conversion from reactive power consumption to reactive power return or from reactive power return to reactive power consumption in the power consumption period to the total number of moments in the power consumption period: (18) F1, the correlation coefficient of the power factor evaluation function value sequence and reactive power sequence in the no-load period in the power consumption period: (19) F2, the ratio of the total reactive power returned to the power grid after removing the transformer loss in the no-load period in the power consumption period to the total measured positive and negative reactive power: (20) F3, the total reactive power consumed after removing the transformer loss in the no-load period in the power consumption period The ratio of electricity consumption to the total amount of measured positive and negative reactive electricity: (21) F4, the distribution variance of reactive power returned to the grid during the no-load period during the power consumption period: (22) F5, the distribution variance of reactive power consumption during the no-load period during the power consumption period: (23) F6, the evaluation function value of the total power factor at the moment when reactive power consumption is converted to reactive power return or from reactive power return to reactive power consumption during the power consumption period: (24) F7, the proportion of the number of moments when reactive power consumption is converted to reactive power return or from reactive power return to reactive power consumption during the power consumption period in the total number of moments during the power consumption period: (25) G1, the evaluation function of the total power factor of the light-load period during the non-no-load period: (26) G2, the total power factor of the non-light-load period during the non-no-load period: (27) G3,Correlation coefficient between the active power sequence and the power factor function value evaluation sequence in the light load period during the non-no-load period: (28) G4, Correlation coefficient between the active power sequence and the power factor function value evaluation sequence in the non-light load period during the non-no-load period. ,

[0038] The method also includes the step of training the parameters of the neural network of the transfer probability of each node in the low power factor critical Bayesian network.

[0039] The steps of training the parameters of the neural network for the transfer probability of each node in the low power factor critical Bayesian network include:

[0040] The training of each node transition probability neural network requires inputting a labeled training data set and updating the network parameters by gradient descent;

[0041] Extracting each node transition probability neural network training data set from the total labeled data set, and labeling the training data set;

[0042] After extracting the label data set for training each neural network, each neural network can be trained separately.

[0043] Among them, the objective function used in training uses the cross entropy loss function, and the network parameter update optimization method uses the Mini-Batch-SGD method

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The low power factor diagnosis technical solution proposed in this application does not require manual on-site equipment detection, which reduces the manual workload and the probability of electric shock accidents. 2. The data source required for diagnosis does not require special equipment for measurement and recording, and does not require power outage tests. Only simple power measurement elements are needed to detect active-reactive power, which can realize online and real-time diagnosis tasks, provide diagnosis results at any time, and have extremely low diagnosis costs. 3. Get rid of manual experience diagnosis and simple threshold diagnosis, and use artificial intelligence methods for automatic analysis and judgment by fully extracting data feature information, which can discover hidden information and patterns, reduce the inefficiency and low accuracy of experience methods, and have strong promotion and practical value. 4. Through training optimization, the model has strong adaptability and expansion capabilities, and the model can be trained to tailor to users with regional characteristics or specific load characteristics. 5. The output is the probability of various problems in the diagnosis object. It can not only locate the main problems according to the probability size, but also discover potential and secondary problems by analyzing the probability distribution of other problems, and complete the diagnosis and early warning work. 6. The diagnosis problem of multiple problems is divided into solving the transfer probability of several nodes through the Bayesian network. Through multiple simple neural networks, the problem that requires more complex and accurate neural networks to solve is solved, reducing the demand for computing power and training data volume. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the binary tree structure of each key point in the embodiment of this application;

[0047] Figure 2 This is a schematic diagram of the conversion structure in the embodiment of the present application;

[0048] Figure 3 This is a schematic diagram of the structure of network node A in an embodiment of the present application;

[0049] Figure 4 This is a schematic diagram of the structure of a network node B in an embodiment of the present application;

[0050] Figure 5 This is a schematic diagram of the structure of the network node C in the embodiment of the present application;

[0051] Figure 6 This is a schematic diagram of the structure of a network node D in an embodiment of the present application;

[0052] Figure 7 This is a schematic diagram of the structure of a network node E in an embodiment of the present application;

[0053] Figure 8 This is a schematic diagram of the structure of a network node F in an embodiment of the present application;

[0054] Fig. 9 This is a schematic diagram of the structure of the network node G in the embodiment of the present application;

[0055] Fig.10 It is a schematic diagram of the overall diagnostic process in the embodiment of the present application;

[0056] Fig.11 A schematic diagram of the process of labeling a training data set in an embodiment of the present application;

[0057] Fig.12 Schematic diagram of training each neural network. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0059] This embodiment provides a low power factor problem diagnosis method by processing power curve data, comprising the following steps:

[0060] Step 1: Bayesian network of low power factor problems: According to the different causes and solutions of low power factor, the problems of low power factor are classified into the following categories: ① Due to the failure of the compensation controller, the capacitor bank is not put into operation according to the rules, resulting in excessive reverse reactive power; ② There are low power factor devices such as no-load rotating motors, and the compensation controller does not work or the compensation element fails as a whole; ③ There are many low power factor devices such as no-load rotating motors, and the compensation capacity of the compensation element is insufficient; ④ The compensation gradient of the compensation element is large, which is not compatible with the load characteristics of the electrical equipment, resulting in over-compensation or under-compensation; ⑤ When the transformer is unloaded or lightly loaded for a long time, the low-voltage side compensation control strategy cannot compensate for the reactive power loss of the transformer; ⑥ The compensation control strategy is not compatible with renewable energy power generation, and the power factor during power generation should be further improved; ⑦ The compensation controller has no effect during renewable energy power generation; ⑧ The power factor is normal and there is no abnormality. By analyzing the dependencies and conditional independence between the various key points, the key points are divided into the following binary tree structure in the form of a decision tree. There are seven judgment nodes in the decision tree. Each judgment node can judge the transfer result node of the next layer according to the input. The transfer result node may be another judgment node or a key leaf node. Figure 1 As shown,

[0061] The dependency relationship between the nodes in the above binary tree is converted into a probabilistic dependency relationship, and then converted into a Bayesian network in the form of a probability graph. Its structure is shown as follows: Figure 2 As shown: By analyzing the Bayesian network of low power factor problems, the probability of occurrence of each problem of the diagnosis object can be calculated, and the formula is as follows:

[0062] P 症结一 =P(A_f)×P(A_B_f)×P(B_D_t)

[0063] P 症结二 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_f)×P(E_G_f)

[0064] P 症结三 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_f)×P(E_G_t)

[0065] P 症结四 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_t)×P(E_F_f)

[0066] P 症结五 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_t)×P(E_F_t)

[0067] P 症结六 =P(A_f)×P(A_B_t)×P(B_C_t)

[0068] P 症结七 =P(A_f)×P(A_B_t)×P(B_C_f)

[0069] P 症结八 =P(A_t)

[0070] In the process of probability calculation, the transfer probability values ​​of nodes A, B, C, D, E, F, and G in the Bayesian network are needed. The following is a series of schemes to calculate the transfer probability by extracting the characteristic quantities of the active-reactive power curve data of the diagnosed object.

[0071] Step 2: Design characteristic quantities; by designing characteristic quantities, the waveform characteristics of the collected active power curve and reactive power curve of the diagnostic object are extracted. The numerical values ​​of these characteristic quantities can reflect different low power factor problems, which are equivalent to various indicators of medical diagnosis.

[0072] 2.1 Variable meaning:

[0073]

[0074] 2.2 Define the functions needed for feature extraction:

[0075] (1) Inverse tangent function arctan(x), cosine function cos(x), covariance conv(X, Y) of sequence X and sequence Y, distribution variance var(X), mean value mean(X) of sequence x, smaller value min([x, y]) of x and y, minimum value min(X) of sequence X, number of elements in sequence X count(X), and absolute value abs(x) of x;

[0076] (2) Define the evaluation function pf_evaluation(q,p) that inputs the active power and reactive power values ​​at the same time or the active power and reactive power values ​​at the same time period, and outputs the corresponding power factor level:

[0077]

[0078] In the process of calculating electricity charges, when calculating the total power factor used for settlement, the calculation rules are more stringent for the power factor standard during the renewable energy power generation period. The evaluation function designed above takes this aspect into account, making the evaluation function value used to quantify the power factor more sensitive to the changes in the actual power factor of the diagnosed object during the power generation period.

[0079] (3) Define the correlation coefficient function corrcoef(x,y) that represents the correlation between the changes in the x and y sequences:

[0080]

[0081] (4) Define the threshold power function threshold(X,S,k) for the active power sequence X of the equipment rated capacity S, where k is a proportional parameter between 0 and 1. Input the power curve X and output the threshold power of the power curve. The moment when the power is greater than this power is a non-no-load moment or a non-light-load moment, and the moment when the power is less than this power is a no-load moment or a light-load moment. By controlling the value of k, it is possible to control whether the threshold power is a light-load threshold power or a no-load threshold power. 0.1 can be set as the no-load threshold and 0.3 as the light-load threshold. The higher k is, the higher the threshold is.

[0082]

[0083] 2.3 Feature extraction:

[0084] (1) A1, evaluation function value of billing total power factor:

[0085]

[0086] The generation time sequence T is defined as follows generate , T generate =(t|p t <0,t∈T total) and the electricity consumption time series T consume , T consume =(t|p t ≥0,t∈T total )

[0087] (2) B1, the evaluation function value of the total power factor during the power consumption period (active power is positive):

[0088]

[0089] (3) B2, the proportion of electricity generation in the total electricity consumption and electricity generation:

[0090]

[0091] (4) C1, evaluation function value of total power factor during power generation period (active power is negative):

[0092]

[0093] (5) C2, the correlation coefficient between the absolute value sequence of active power during the power generation period and the power factor evaluation function value sequence at each moment: C1 = corrcoef((abs(p t )|t∈T generate ),(pf_evaluation(q t ,p t )|t∈T generate )) The following defines the time sequence T of the period of reactive power feeding back to the power grid during the power consumption period: reverse , T reverse =(t|q t <0,t∈T consume ) and the time sequence T of the reactive power consumption period in the power consumption period positive , T positive =(t|q t ≥0,t∈T consume )

[0094] (6) D1, the evaluation function value of the total power factor during the period of reactive power feeding back to the power grid during the power consumption period:

[0095]

[0096] (7) D2, the evaluation function value of the total power factor during the period of reactive power consumption in the power consumption period:

[0097]

[0098] (8) D3, the ratio of total reactive power sent back to the grid during the power consumption period to the total reactive power during the power consumption period:

[0099] (9) D4, distribution standard deviation of reactive power sent back to the grid during the power consumption period:

[0100]

[0101] The no-load threshold power P during the power consumption period is defined as follows: load-threshold , P load-threshold =threshold((P t |t∈T consume ),S,0.1). Taking this as the threshold, the period not less than this power is the non-no-load period, and its time sequence is T load =(t|P t ≥P load-threshold ,t∈T consume ), the period less than this power is the no-load period and its time sequence is T no-load =(t|P t <P load-threshold ,t∈T consume ).

[0102] (10) E1, the evaluation function value of the total power factor during the no-load period during the power consumption period:

[0103]

[0104] (11) E2, the evaluation function value of the total power factor during the non-no-load period during the power consumption period:

[0105]

[0106] The transformer active power loss p is defined as follows: loss , p loss =min((p t |t∈T no-load )); and transformer reactive power loss q loss ,q loss =min((q t |t∈(t|t∈T no-load ,q t ≥0)))

[0107] (12) E3, the evaluation function value of the total power factor in the non-no-load period after removing the active power and reactive power loss of the transformer during the power consumption period:

[0108]

[0109] (13) E4, the total power factor evaluation function value of the no-load period after removing the transformer active power and reactive power losses during the power consumption period:

[0110]

[0111] (14) E5, the correlation coefficient between the power factor evaluation function value sequence and the active power sequence during the no-load period in the power consumption period:

[0112] E5 = corrcoef((p t |t∈T no-load ),(pf_evaluation(q t ,p t )|t∈T no-load ))

[0113] (15) E6, the correlation coefficient between the power factor evaluation function value sequence and the active power sequence in the non-no-load period during the power consumption period:

[0114] E6 = corrcoef((p t |t∈T load ),(pf_evaluation(q t ,p t )|t∈T load ))

[0115] The following defines the reactive power crossing time sequence T at which the reactive power is converted from consuming reactive power to sending reactive power or from sending reactive power to consuming reactive power during the power consumption period: tran =(t|q t ×q t+1 <0,t∈T consume )

[0116] (16) E7, the evaluation function value of the total power factor at the moment of conversion from reactive power consumption to reactive power return or from reactive power return to reactive power consumption during the power consumption period:

[0117]

[0118] (17) E8, the proportion of the number of moments in the power consumption period when the reactive power is converted to the reactive power return or the reactive power return to the reactive power consumption moment in the total number of moments in the power consumption period:

[0119]

[0120] (18) F1, the correlation coefficient between the power factor evaluation function value sequence and the reactive power sequence during the no-load period in the power consumption period: F1 = corrcoef((abs(q t )|t∈T no-load ),(pf_evaluation(q t ,p t )|t∈T no-load ))

[0121] (19) F2, the ratio of the total reactive power sent back to the grid during the no-load period after removing the transformer loss to the total measured positive and negative reactive power:

[0122]

[0123] (20) F3, the ratio of total reactive power consumed during the no-load period after removing transformer loss to total measured positive and negative reactive power:

[0124]

[0125] (21) F4, distribution variance of reactive power fed back to the grid during the no-load period during the power consumption period:

[0126]

[0127] (22) F5, distribution variance of reactive power consumption during the no-load period in the power consumption period:

[0128]

[0129] (23) F6, switching from consuming reactive power to sending reactive power or from sending reactive power to sending reactive power during the power consumption period

[0130] The evaluation function value of the total power factor from the moment of reactive power consumption to the moment of reactive power consumption:

[0131]

[0132] (24) F7, switching from consuming reactive power to sending reactive power or from sending reactive power to sending reactive power during the power consumption period

[0133] The proportion of the time when the reactive power is consumed in the total number of time during the power consumption period is:

[0134]

[0135] The light load threshold power P during the non-no-load period is defined as follows: light-threshold , P light-threshold =threshold((P t |t∈T load ),S,0.3). Taking this as the threshold, the period with power not less than this in the non-no-load period is the non-light-load period, and its time sequence is T no-light =(t|P t ≥P light-threshold ,t∈T load ), on the contrary, the period less than this power is the light load period and its time sequence is T light =(t|P t <P light-threshold ,t∈T load)

[0136] (25) G1, evaluation function of total power factor in light load period in non-no-load period:

[0137]

[0138] (26) G2, total power factor during non-light load period during non-no-load period:

[0139]

[0140] (27) G3, active power sequence and power factor function value evaluation sequence of light load period in non-no-load period

[0141] Column correlation coefficient:

[0142] G3=corrcoef((p t |t∈T light ),(pf_evaluation(q t ,p t )|t∈T light ))

[0143] (28) G4, Evaluation of active power sequence and power factor function value in non-light load period during non-no-load period

[0144] Serial Correlation Coefficient

[0145] G4=corrcoef((p t |t∈T no-light ),(pf_evaluation(q t ,p t )|t∈T no-light ))

[0146] Step 3: Feature combination

[0147] The above variables are processed and combined into variable tuples related to the calculation of the transfer probability of different nodes in the Bayesian network. The variable processing method is mainly to calculate the ratio of the power factor evaluation function values ​​of different time periods according to the transfer probability representation content to enhance the representation ability of the input elements. These variables are combined into input tuples for calculating different transfer probabilities, and after processing based on machine learning methods, the estimated value of the transfer probability can be output.

[0148] Next, we construct the input tuple according to the relationship between the variable values ​​and the transition probabilities.

[0149] (1) Determine whether the power factor is normal. The transfer probability of node A: A_in = [A1];

[0150] (2) Determine whether the problem is related to renewable energy generation. The transfer probability of node B: B_in = [A1 / B1, B2];

[0151] (3) Determine whether the compensator has an effect when renewable energy is generating electricity. The transfer probability of node C is C_in = [C1, C2];

[0152] (4) Determine whether the reactive power compensator fails and the transfer probability of the over-generated reactive power node D

[0153] D_in=[D1 / B1,D2 / B1,D3,D4];

[0154] (5) Determine the transfer probability of whether there is a problem node E during the light load period

[0155] E_in=[E1 / B1,E2 / B1,E3 / B1,E4 / B1,E5,E6,E7,E8];

[0156] (6) Determine whether the control strategy has the transition probability of the problem node F

[0157] F_in=[F1,F2,F3,F4,F5,F6,F7];

[0158] (7) Determine whether the compensation capacity is insufficient and the transfer probability of node G

[0159] G_in=[G1 / E2,G2 / E2,G3,G4];

[0160] Step 4: Transfer Probability Neural Network

[0161] The estimation method of the transition probability of each node in the Bayesian network uses the algorithm of the neural network. The input of the neural network is the input tuple related to the corresponding node transition probability after feature extraction and processing of the power curve of the diagnosis object, and the output is the estimated value of the transition probability of the diagnosis object at the node. Each transition probability neural network is fully connected, and its structure, input tuple and output transition probability are shown as follows: Node A is as follows Figure 3 As shown; Node B is as Figure 4 As shown; Node C is Figure 5 As shown; Node D is Figure 6 As shown; judgment node E is as Figure 7 As shown; judgment node F is as Figure 8 As shown; the judgment node G is as follows Fig. 9 shown.

[0162] Step 5: Output of problem diagnosis results:

[0163] (1) After extracting the characteristic quantities of the active-reactive power curve of the diagnosed object and combining them into different input tuples and inputting them into the neural network of each judgment node in the decision tree, the network will output the probabilities of the two types of diagnosis results of the corresponding nodes.

[0164] (2) According to the jump connection and conditional probability relationship between the output results of each judgment node in the diagnostic Bayesian network, the probability of each low power factor problem occurring in the diagnosed object is:

[0165] P 功率因数正常，无异常情况 =A_out1

[0166] P 补偿控制策略与新能源发电不适配，应进一步提升发电期间功率因数 =A_out2×B_out1×C_out1

[0167] P 补偿控制器在新能源发电时无作用 =A_out2×B_out1×C_out2

[0168] P 因为补偿控制器失灵，导致电容器组不按照规则投入，引起反向无功过多 =A_out2×B_out2×D_out1

[0169] P 补偿元件补偿的梯度较大，和用电设备负荷特性不适配，导致过补偿或欠补偿 =A_out2×B_out2×D_out2×E_out1×F_out1

[0170] P 在变压器长期空载或轻载的情况下，低压侧补偿控制策略无法补偿变压器无功损耗 =A_out2×B_out2×D_out2×E_out1×F_out2P 有空载旋转电机等低功率因数设备，且补偿控制器不工作或补偿元件整体失灵 =A_out2×B_out2×D_out2×E_out2×G_out1

[0171] P 空载旋转电机等低功率因数设备较多，补偿元件补偿容量不足 =A_out2×B_out2×D_out2×E_out2×G_out2

[0172] (3) The problem with the highest probability among the eight types of problems is the main problem, and the probability values ​​of other problems can also be used as a reference, because the low power factor problem is not caused by a single problem in some cases, but there are many potential problems in parallel. This diagnostic method can not only determine the main problem, but also discover potential problems.

[0173] The above 1-5 diagnostic process can be summarized as follows: Fig.10 shown.

[0174] Step 6: Train the transfer probability neural network parameters of each node in the low power factor Bayesian network:

[0175] The training of each node transition probability neural network requires the input of a labeled training data set, and the network parameters are updated by gradient descent to reduce the error between the network output and the label. The process of extracting each node transition probability neural network training data set from the total labeled data set and labeling the training data set is as follows Fig.11 As shown in the figure: After extracting the label data set for training each neural network, each neural network can be trained separately. The objective function used in training uses the cross entropy loss function, and the network parameter update optimization method uses the Mini-Batch-SGD method. The training process is as follows Fig.12 .

[0176] Application use cases, first: cooperate with terminal metering and collection equipment, after the collection equipment transmits back and stores the power curve metering data, use the diagnostic tool of this patented technical solution, through the user's active initiation or the software's automatic initiation, to obtain the active-reactive power curve data within a certain period of time from the storage space. Statistical feature extraction is performed on the power curve data, and then the feature is input into the trained diagnostic neural network, the probability of each low power factor problem is output, and the probability of the main problem and other problems are presented to the user, completing an online diagnosis, guiding the user to pay attention to the problem in time and troubleshoot and repair it, improve energy efficiency, and reduce energy costs. After solving the low power factor problem, add the diagnostic object to the training set for subsequent optimization of the diagnostic neural network and improving diagnostic capabilities.

[0177] Second: As an ancillary product of the electricity bill service, when the interest rate adjustment electricity bill item in the electricity bill is high, the characteristic values ​​of the monthly power curve data of the household's electricity bill are extracted through analysis and input into the diagnostic neural network. The output information such as the main and secondary problems will be presented to provide a reference for users to improve quality and reduce fees in the next step.

[0178] Experimental verification: Use Python to write a program for training and diagnosis using the patented method to verify the diagnostic accuracy of the patented method. 138 labeled samples were selected as the training data set, and the results were good.

[0179] The optional implementation modes of the embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above implementation modes. Within the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical scheme of the embodiments of the present invention, and these simple modifications all belong to the protection scope of the embodiments of the present invention.

Claims

1. A method for diagnosing low power factor problems by processing power curve data, characterized in that: The steps include: Step 1: Construct a Bayesian network for low power factor problems. By analyzing the Bayesian network for low power factor problems, the occurrence probability of each problem of the diagnosis object can be calculated. Step 2: Designing feature quantities, and extracting waveform features of the collected active power curve and reactive power curve of the diagnostic object by designing feature quantities; Step 3: Feature quantity combination, process the above variables and combine them into variable tuples that are related to calculating the transition probability of different nodes in the Bayesian network; Step 4: Transfer probability neural network. The estimation method of the transfer probability of each node in the Bayesian network uses the neural network algorithm. The input of the neural network is the input tuple related to the corresponding node transfer probability after feature extraction and processing of the power curve of the diagnosis object, and the output is the estimated value of the transfer probability of the diagnosis object at the node. Step 5: Output of the key diagnosis results. After extracting the characteristic quantities of the active-reactive power curve of the diagnosed object and combining them into different input tuples and inputting them into the neural network of each judgment node in the decision tree, the network will output the probabilities of the two types of diagnosis results of the corresponding nodes.

2. The method for diagnosing low power factor problems by processing power curve data according to claim 1, characterized in that: The step 1 specifically includes the following: Step 11: Categorize the causes of low power factor according to the different causes and solutions. Step 12: By analyzing the dependency and conditional independence between the various key points, the key points are divided into the following binary tree structure in the form of a decision tree. There are seven judgment nodes in the decision tree. Each judgment node can judge the transfer result node of the next layer according to the input, where the transfer result node is another judgment node or a key leaf node; Step 13: Convert the dependency relationship between the nodes in the above binary tree into a probabilistic dependency relationship, and convert it into a Bayesian network in the form of a probability graph; Step 14: By analyzing the Bayesian network of the low power factor problem, the occurrence probability of each problem of the diagnosis object can be calculated, and the formula is as follows: P 症结一 =P(A_f)×P(A_B_f)×P(B_D_t) P 症结二 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_f)×P(E_G_f) P 症结三 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_f)×P(E_G_t) P 症结四 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_t)×P(E_F_f) P 症结五 =P(A_f)×P(A_B_f)×P(B_D_f)×P(D_E_t)×P(E_F_t) P 症结六 =P(A_f)×P(A_B_t)×P(B_C_t) P 症结七 =P(A_f)×P(A_B_t)×P(B_C_f) P 症结八 =P(A_t) Among them, in the process of probability calculation, the transition probability values ​​of nodes A, B, C, D, E, F, and G in the Bayesian network are needed.

3. The method for diagnosing low power factor problems by processing power curve data according to claim 2, characterized in that: In step 11, the causes of low power factor are classified into the following categories: (1) due to the failure of the compensation controller, the capacitor bank is not put into operation according to the rules, resulting in excessive reverse reactive power; (2) there are low power factor devices such as no-load rotating motors, and the compensation controller does not work or the compensation element fails as a whole; (3) there are many low power factor devices such as no-load rotating motors, and the compensation capacity of the compensation element is insufficient; (4) the compensation gradient of the compensation element is large, which is not compatible with the load characteristics of the electrical equipment, resulting in over-compensation or under-compensation; (5) when the transformer is long-term no-load or light-loaded, the low-voltage side compensation control strategy cannot compensate for the reactive power loss of the transformer; (6) the compensation control strategy is not compatible with renewable energy power generation, and the power factor during power generation should be further improved; (7) the compensation controller has no effect when renewable energy is generating electricity; (8) the power factor is normal and there is no abnormality.

4. The method for diagnosing low power factor problems by processing power curve data according to claim 1, characterized in that: The step 2 specifically includes the following: Step 21: Define variables; Step 22: Define the functions needed for feature extraction; Step 23: feature extraction; Step 24: Feature quantity combination.

5. The method for diagnosing low power factor problems by processing power curve data according to claim 4, characterized in that: In step 21, the variables include p t , active power at time t; q t Reactive power at time t; S, rated capacity; T, the time series of the power curve, where T total Contains all moments of the power curve.

6. The method for diagnosing low power factor problems by processing power curve data according to claim 4, characterized in that: In step 22, the function includes: (1) arctan(x), cosine(x), covariance conv(X, Y) of X sequence and Y sequence, distribution variance var(X) of X sequence, mean(X) of sequence x, smaller value min([x, y]) of x and y, minimum value min(X) of sequence X, number of elements count(X) of sequence X, and absolute value abs(x) of x; (2) defining an evaluation function pf_evaluation(q, p) that inputs active power and reactive power values ​​at the same time or active power and reactive power values ​​at the same time period and outputs a corresponding power factor level; (3) defining a correlation coefficient function corrcoef(x, y) that characterizes the correlation degree of changes in x sequence and y sequence; (4) defining a threshold power function threshold(X, S, k) for calculating the active power sequence X with a rated capacity of S.

7. The method for diagnosing low power factor problems by processing power curve data according to claim 4, characterized in that: In step 23, the characteristic quantities include: (1) A1, the evaluation function value of the total power factor for billing; (2) B1, the evaluation function value of the total power factor during the electricity consumption period; (3) B2, the proportion of power generation in the total power consumption and power generation; (4) C1, the evaluation function value of the total power factor during the power generation period; (5) C2, the correlation coefficient between the absolute value sequence of active power in the power generation period and the evaluation function value sequence of the power factor at each moment; (6) D1, the evaluation function value of the total power factor during the power consumption period when reactive power is fed back to the power grid; (7) D2, the evaluation function value of the total power factor during the power consumption period when reactive power is consumed; (8) D3, the ratio of the total reactive power fed back to the power grid during the power consumption period to the total reactive power of the metered power consumption period; (9) 9) D4, distribution standard deviation of reactive power fed back to the grid during the power consumption period: (10) E1, evaluation function value of total power factor during no-load period during the power consumption period: (11) E2, evaluation function value of total power factor during non-no-load period during the power consumption period: (12) E3, evaluation function value of total power factor during non-no-load period after removing transformer active power and reactive power loss during the power consumption period: (13) E4, evaluation function value of total power factor during no-load period after removing transformer active power and reactive power loss during the power consumption period: (14) E5, correlation coefficient between power factor evaluation function value sequence and active power sequence during no-load period during the power consumption period: (15) E6, power factor evaluation function value sequence and active power sequence during non-no-load period during the power consumption period =Correlation coefficient of the sequence: (16) E7, the evaluation function value of the total power factor at the moment when reactive power consumption is converted to reactive power return or from reactive power return to reactive power consumption in the power consumption period: (17) E8, the proportion of the number of moments when reactive power consumption is converted to reactive power return or from reactive power return to reactive power consumption in the power consumption period to the total number of moments in the power consumption period: (18) F1, the correlation coefficient of the power factor evaluation function value sequence and the reactive power sequence in the no-load period in the power consumption period: (19) F2, the ratio of the total reactive power returned to the power grid after removing the transformer loss in the no-load period in the power consumption period to the total measured positive and negative reactive power: (20) F3, the total reactive power consumed after removing the transformer loss in the no-load period in the power consumption period The ratio of the amount to the total amount of measured positive and negative reactive power: (21) F4, the distribution variance of reactive power returned to the grid during the no-load period during the power consumption period: (22) F5, the distribution variance of reactive power consumed during the no-load period during the power consumption period: (23) F6, the evaluation function value of the total power factor at the moment when reactive power consumption is converted to reactive power return or from reactive power return to reactive power consumption during the power consumption period: (24) F7, the proportion of the number of moments when reactive power consumption is converted to reactive power return or from reactive power return to reactive power consumption during the power consumption period in the total number of moments during the power consumption period: (25) G1, the evaluation function of the total power factor of the light-load period during the non-no-load period: (26) G2, the total power factor of the non-light-load period during the non-no-load period: (27) G3,Correlation coefficient between the active power sequence and the power factor function value evaluation sequence in the light load period during the non-no-load period: (28) G4, Correlation coefficient between the active power sequence and the power factor function value evaluation sequence in the non-light load period during the non-no-load period. , 8. The method for diagnosing low power factor problems by processing power curve data according to claim 1, characterized in that: The method also includes the step of training the parameters of the neural network for the transfer probability of each node in the low power factor critical Bayesian network.

9. The method for diagnosing low power factor problems by processing power curve data according to claim 8, characterized in that: The steps of training the parameters of the transfer probability neural network of each node in the low power factor crux Bayesian network include: The training of each node transition probability neural network requires inputting a labeled training data set and updating the network parameters by gradient descent; Extracting each node transition probability neural network training data set from the total labeled data set, and labeling the training data set; After extracting the label data set for training each neural network, each neural network can be trained separately.

10. The method for diagnosing low power factor problems by processing power curve data according to claim 9, characterized in that: The objective function used in training is the cross entropy loss function, and the network parameter update optimization method uses the Mini-Batch-SGD method.

Images