A bad data identification method based on time series analysis coupled with neural network prediction

Abstract

The present invention proposes a method for identifying bad data based on time series analysis coupled with neural network prediction, and the main steps include: obtaining the original data, dimensionless processing the original data, testing the stability of the original sample parameter sequence, if it is not stable, performing differential processing on the sequence and testing again until the differential sequence is stable, establishing a differential autoregressive moving average model of the sequence, calculating the model residual, and classifying the residual value greater than the preset value into a bad sample set. A BP neural network model is established, and the bad sample set is cyclically tested through the neural network model to finally obtain a bad data set. The present invention couples time series analysis and a neural network model, uses time series analysis to preliminarily obtain bad data, and then uses a neural network model to verify and verify. It is very suitable for processing denitrification system data with the characteristics of many parameters and large fluctuations. It has high identification precision and accuracy and can be widely used for the identification of bad data in the field of flue gas denitrification.

Description

Technical Field

[0001] The present invention belongs to the field of nitrogen oxide control, and in particular relates to a bad data identification method based on time series analysis coupled with neural network prediction. Background Art

[0002] With the development of technologies such as the Internet of Things, machine learning, and big data analysis, there are more and more studies on intelligent control of denitrification systems and catalyst management based on big data analysis of SCR denitrification systems. However, due to measurement errors, equipment failures, transmission failures, and other problems in the process of big data collection of SCR denitrification systems, the original samples often contain some bad data. These bad data not only interfere with the control of the denitrification system, but also have many adverse effects on intelligent control and catalyst management based on big data analysis, limiting the promotion and application of these technologies. Therefore, how to identify bad data is of great significance to the stable and efficient operation of the SCR denitrification system.

[0003] Traditional methods for identifying bad data mainly include physical discrimination and mathematical statistics. The physical discrimination method is based on people's objective understanding of known data, and is a method of judging the deviation of measured data from normal values ​​due to external interference, human errors, etc. However, due to the large amount of data generated by the SCR denitrification system, the manual judgment workload is large and the operator needs to have rich experience, which makes this method difficult to implement. The mathematical statistics method uses mathematical statistics theory to identify bad data, but due to the late development of big data research on the SCR denitrification system, relevant mathematical models have not yet been established and lack theoretical guidance. The use of BP neural network for bad data detection is an information processing method based on machine learning that has been developed in recent years. It has the advantages of wide model adaptability, high prediction accuracy, and good model accuracy. However, the current method has the disadvantages of imperfect theory and large amount of calculation. In fact, the data acquisition system of SCR denitrification usually samples denitrification-related variables such as flue gas volume and inlet NO at a fixed sampling frequency. x The concentration and other parameters are monitored and saved, which is a typical time series. Therefore, the original data can be preliminarily simplified and analyzed by the time series analysis method to reduce the complexity of the data. Therefore, the present invention proposes a detection and identification method for bad data of SCR denitrification system based on time series analysis coupled with neural network prediction. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a bad data identification method based on time series analysis coupled with neural network prediction.

[0005] The method first preprocesses the time series collected by the SCR denitrification system, models the time series using the autoregressive moving average model, preliminarily determines the abnormal data set based on residual analysis, uses the remaining data after eliminating the abnormal data to model and learn the BP neural network, and then uses the trained neural network to test the preliminarily determined abnormal data set, identifies bad samples based on the relative error between the predicted value and the initial value of the test sample; for the identified bad samples, further optimizes the neural network to test and correct the bad samples, and finally determines the bad data.

[0006] In order to achieve the above-mentioned purpose and the above-mentioned technical effect, the present invention is implemented by the following technical scheme: a bad data identification method based on time series analysis coupled neural network prediction, characterized in that it includes the following steps:

[0007] Step 1: Obtain the original data collected by the flue gas denitrification system, supplement the sample missing data in the original data, and then perform dimensionless processing on the data to obtain sample data;

[0008] Step 2: Perform a stationary analysis on the sample data. If the sample data is stationary, proceed to the next step of analysis. If the sample data is not stationary, perform a difference process on the sample data, and then perform a stationary analysis on the difference sequence. Repeat this process until the sequence is detected to be stationary.

[0009] Step 3: Construct the difference autoregressive moving average model ARIMA;

[0010] Step 4: Calculate the residual of the differential autoregressive moving average model, add samples with residuals greater than the preset value to the abnormal data set, and classify the remaining samples into the normal sample set;

[0011] Step 5: Construct BP neural network;

[0012] Step 6: Take the abnormal data set and 10% of the random normal samples as test samples, and the remaining 90% of the normal samples as training samples. Use the learned neural network model to predict the test samples, and reclassify the test samples into the bad sample set or the training sample set according to the preset conditions. Perform a cyclic test until the preset number of times is reached.

[0013] Step 7: Check and correct the bad sample set to determine the final bad data.

[0014] Preferably, the above-mentioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the original data in step 1 is data collected at equal time intervals, each data contains at least two or more parameters; and at least contains parameter export NO x At least one of concentration, denitrification efficiency or catalyst activity is used as its label parameter.

[0015] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the dimensionless method in step 1 is calculated using any one of the following formulas 1 to 3:

[0016]

[0017]

[0018]

[0019] In the formula, Z ni is the normalized value of parameter Z, i is the original value of parameter Z, Z max is the maximum value of parameter Z, Z min is the minimum value of parameter Z, σ is the standard deviation of parameter Z, and μ is the mean value of parameter Z.

[0020] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the stationarity analysis in step 2 adopts at least one of the two methods of ADF test or KPSS test.

[0021] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the differential processing in step 2 refers to obtaining the differential sequence by calculation using the following formula:

[0022]

[0023] In the formula, is the difference sequence we want, Z k is the parameter Z at time k, Z k-1 is the parameter Z at time k-1.

[0024] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the ARIMA model steps in step 3 are as follows:

[0025] 1) Calculate and plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the difference sequence, and preliminarily determine the autoregressive order p and moving average order q of the model based on the tailing or truncation properties of the ACF and PACF diagrams;

[0026] 2) Determine the order p and q of the model using either the Akaike Information Criterion (AIC Criterion) or the Bayesian Information Criterion (BIC Criterion);

[0027] 3) Based on the model order determined above, the model parameters are estimated using the maximum likelihood function method or the least squares method to preliminarily determine the model:

[0028]

[0029] In the formula, y t is the difference sequence to be found, ε t is the white noise residual, a j and θ i is the model coefficient to be estimated;

[0030] 4) Calculate the model residual and check whether the residual is white noise. If so, the model is successfully established. If not, repeat steps 2) to 4) until the residual is verified to be white noise.

[0031] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the preset value in step 4 is twice the standard deviation of the residual sequence.

[0032] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the preset condition of step 6 means that if the relative error is less than 10%, it is classified as a training sample, otherwise it is classified as a bad sample, and the preset number of times is at least 3 times.

[0033] Preferably, the aforementioned bad data identification method based on time series analysis coupled with neural network prediction is characterized in that the inspection and correction steps in step 7 are as follows:

[0034] 1) Delete bad samples, use optimized samples to build and train BP neural network, use the trained neural network to test bad samples, and calculate relative error;

[0035] 2) Screen out samples with a relative error of no less than 10% and confirm them as bad data; screen out samples with a relative error of no more than 5% and correct them as good data;

[0036] 3) Repeat step 1) and step 2) for the remaining samples after screening in step 2) until any of the following conditions is met:

[0037] Condition I: no sample remains after executing step 2);

[0038] Condition II: Repeat the execution for no less than 3 times;

[0039] 4) All remaining samples are confirmed as bad data.

[0040] The beneficial effects of the present invention are:

[0041] 1. SCR denitrification big data is a typical time series, which has the characteristics of large data volume, frequent short-term fluctuations, and severe noise interference that most time series have, making the analysis and processing of time series more complicated and difficult. The differential autoregressive moving average model is a traditional and effective time series processing method with the advantages of mature theory, small amount of calculation, and fast processing speed. Through the preprocessing of time series analysis, data error values ​​and outliers can be effectively screened, the distribution of bad data can be preliminarily understood, and the complexity of original data can be reduced.

[0042] 2. The neural network model has strong nonlinear fitting and generalization capabilities, and has shown excellent results in pattern recognition, regression prediction, data identification, etc. However, the neural network theory is immature, the amount of calculation is large, and the prediction accuracy is affected to a certain extent by the quality of the training samples. Pre-selecting and filtering out bad sample sets through time series analysis can reduce the impact of bad samples in the neural network modeling process. Using neural network prediction to predict and test abnormal data can prevent misjudgment of simple time series analysis and improve the accuracy of judgment.

[0043] 3. The neural network is used to perform cyclic testing on the initially obtained bad sample set and 10% of the randomly selected normal samples to avoid the possibility of missing bad samples due to relying solely on time series analysis; finally, the neural network is used to test and correct the bad samples identified by the coupling of the two methods to avoid misjudgment of bad samples and improve recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The present invention is an operational flow chart of the method. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0046] This embodiment takes the normal operation data of the SCR denitration system of a coal-fired power plant for one year as an example, and detects the bad data in these samples using the method of the present invention by artificially adding bad data.

[0047] Example 1

[0048] A bad data identification method based on time series analysis coupled with neural network prediction, such as Figure 1 As shown, the specific implementation steps are as follows:

[0049] Step 1: Collect 1000 sets of accurate raw data U0 from the SCR denitrification system. Each set of samples includes inlet pressure, inlet flue gas volume, inlet NO x concentration, ammonia injection amount, flue gas temperature, outlet pressure, outlet NO xA total of 10 parameters are included, namely, concentration, denitration efficiency, ammonia escape concentration, and SCR denitration catalyst activity. 50 groups are randomly selected from the original data U0 as sample data, and an error of 10% is artificially added as a bad sample. The processed samples are dimensionless processed according to the following formula (1), which is recorded as sample set U1:

[0050]

[0051] In the formula, Z ni is the normalized value of parameter Z, i is the original value of parameter Z, Z max is the maximum value of parameter Z, Z min is the minimum value of parameter Z.

[0052] Step 2: Use the ADF test to perform a stationary analysis on the above samples. If the sample is detected to be stationary, proceed to the next step of analysis. If the sample is not stationary, perform a first-order difference on the sample, perform a stationary analysis on the differential sequence DU, and repeat the process until the detected differential sequence DU is stationary. The obtained differential sequence DU is used as the sample for subsequent analysis.

[0053] Step 3: Use the following steps to build the autoregressive moving average model of the difference series:

[0054] 1) Calculate and plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the difference sequence, and preliminarily determine the autoregressive order p and moving average order q of the model based on the tailing or truncation position of the ACF and PACF diagrams;

[0055] 2) The Bayesian Information Criterion (BIC criterion) is used to determine the order p and q of the model;

[0056] 3) Based on the model order determined above, the maximum likelihood function method is used to estimate the model parameters and preliminarily determine the model.

[0057]

[0058] Where: y t is the difference sequence to be determined, ε t is the white noise residual, a j and θ i are the model coefficients to be estimated.

[0059] 4) Calculate the model residuals and use the LBQ test (Ljung-Box test) to check whether the residuals are white noise. If so, the model is successfully established. If not, repeat steps 2) to 4) until the residuals are white noise.

[0060] Step 4: Calculate the residual R of the above model, and classify the samples whose residual values ​​are greater than 2 times the standard deviation of the residual sequence into the bad sample set BU, and the rest of the samples into the normal sample set GU.

[0061] Step 5: Take the first 9 parameters of each group of samples as the neural network input value, and the last parameter as the neural network output value, and construct a BP neural network with a single hidden layer. The hidden layer and the output layer use the hyperbolic tangent function.

[0062] Step 6: Take the abnormal data set and 10% of the random normal sample set data as test samples TU, and the remaining samples as training samples. Use the learned neural network model to predict the test samples, and classify the samples with prediction errors exceeding 10% into the bad sample set BU again, and the rest into the training sample set TU. Repeat this step 3 times.

[0063] Step 7: Remove all marked bad samples from the original samples, use the remaining samples to reconstruct and train the BP neural network, and use the trained neural network to test and correct the bad samples. The specific steps are as follows:

[0064] 1) Delete bad samples, use optimized samples to build and train BP neural network, use the trained neural network to test bad samples, and calculate relative error;

[0065] 2) Screen samples with a relative error greater than 20% and determine them as bad data; screen samples with a relative error less than 5% and correct them as non-bad data;

[0066] 3) Repeat step 1) and step 2) for the remaining samples after screening in step 2) until any of the following conditions is met: condition 1, no remaining samples after executing step 2); condition 2, the number of repetitions is not less than 3 times;

[0067] 4) All remaining samples are confirmed as bad data BD.

[0068] After the above steps, the method identified 47 groups of bad data, and all the identified samples were set bad samples. The recognition rate of bad data obtained by this method was 94%, and the accuracy rate was 100%.

[0069] Example 2

[0070] A bad data identification method based on time series analysis coupled with neural network prediction, such as Figure 1 As shown, the specific implementation steps are as follows:

[0071] Step 1: Collect 1000 sets of accurate raw data U0 from the SCR denitrification system. Each set of samples includes inlet pressure, inlet flue gas volume, inlet NO xconcentration, ammonia injection amount, flue gas temperature, outlet pressure, outlet NO x A total of 10 parameters are included, namely, concentration, denitration efficiency, ammonia escape concentration, and SCR denitration catalyst activity. 50 groups are randomly selected from the original data U0 as samples, and an error of 10% is artificially added as bad samples. The processed samples are dimensionless processed according to the following formula (1), which is recorded as sample set U1:

[0072]

[0073] In the formula, Z ni is the normalized value of parameter Z, i is the original value of parameter Z, Z max is the maximum value of parameter Z, Z min is the minimum value of parameter Z.

[0074] Step 2: Use KPSS test to perform stationarity analysis on the above samples. If the sample is detected to be stationary, proceed to the next step of analysis. If the sample is not stationary, perform first-order difference on the sample, perform stationarity analysis on the difference sequence DU, and repeat the process until the detected difference sequence DU is stationary. The obtained stationary difference sequence DU is used as the sample for subsequent analysis.

[0075] Step 3: Use the following steps to build the autoregressive moving average model of the difference series:

[0076] 1) Calculate and plot the autocorrelation function (ACF) and partial autocorrelation function (PACF) of the difference sequence, and preliminarily determine the autoregressive order p and moving average order q of the model based on the tailing or truncation position of the ACF and PACF diagrams;

[0077] 2) Akaike information criterion (AIC criterion) was used to determine the order p and q of the model;

[0078] 3) Based on the model order determined above, the least squares method is used to estimate the model parameters and preliminarily determine the model.

[0079]

[0080] Where: y t is the difference sequence to be found, ε t is the white noise residual, a j and θ i are the model coefficients to be estimated.

[0081] 4) Calculate the model residuals and use the LBQ test (Ljung-Box test) to check whether the residuals are white noise. If so, the model is successfully established. If not, repeat steps 2) to 4) until the residuals are white noise.

[0082] Step 4: Calculate the residual R of the above model, and classify the samples whose residual values ​​are greater than 2 times the standard deviation of the residual sequence into the bad sample set BU, and the rest of the samples into the normal sample set GU.

[0083] Step 5: Take the first 9 parameters of each group of samples as the neural network input value, and the last parameter as the neural network output value, and construct a BP neural network with a single hidden layer. The hidden layer and the output layer use the Sigmoid function.

[0084] Step 6: Take the abnormal data set and 10% of the random normal sample set data as test samples TU, and the remaining samples as training samples. Use the learned neural network model to predict the test samples, and classify the samples with prediction errors exceeding 10% into the bad sample set BU again, and the rest into the training sample set TU. Repeat this step 3 times.

[0085] Step 7: Remove all marked bad samples from the original samples, use the remaining samples to reconstruct and train the BP neural network, and use the trained neural network to test and correct the bad samples. The specific steps are as follows:

[0086] 1) Delete bad samples, use optimized samples to build and train BP neural network, use the trained neural network to test bad samples, and calculate relative error;

[0087] 2) Screen samples with a relative error greater than 20% and determine them as bad data; screen samples with a relative error less than 5% and correct them as non-bad data;

[0088] 3) Repeat step 1) and step 2) for the remaining samples after screening in step 2) until any of the following conditions is met: condition 1, no remaining samples after executing step 2); condition 2, the number of repetitions is not less than 3 times;

[0089] 4) All remaining samples are confirmed as bad data BD.

[0090] Through the above steps, the method identified 45 groups of bad data, and all the identified samples were set bad samples. The recognition rate of bad data obtained by this method was 90%, and the accuracy rate was 100%.

Claims

1. A bad data identification method based on time series analysis coupled with neural network prediction, characterized in that: The following steps are involved: Step 1: Obtain the original data collected by the flue gas denitrification system, supplement the sample missing data in the original data, and then perform dimensionless processing on the data to obtain sample data; Step 2: Perform a stationary analysis on the sample data. If the sample data is stationary, proceed to the next step of analysis. If the sample data is not stationary, perform a difference process on the sample data, and then perform a stationary analysis on the obtained difference sequence. Repeat this process until the difference sequence is stationary. Step 3: Construct the difference autoregressive moving average model ARIMA; Step 4: Calculate the residual of the differential autoregressive moving average model, add samples with residuals greater than the preset value to the abnormal data set, and classify the remaining samples into the normal sample set; Step 5: Construct BP neural network; Step 6: Take the abnormal data set and 10% of the random normal samples as test samples, and the remaining 90% of the normal samples as training samples, establish a neural network model, and then predict the test samples. According to the preset conditions, the test samples are reclassified as bad sample sets or training sample sets, and the cycle test is performed until the preset number of times is reached; Step 7: Check and correct the bad sample set to determine the final bad data.

2. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The original data in step 1 are data collected at equal time intervals, each of which contains at least two parameters; and at least contains parameter export NO x At least one of concentration, denitrification efficiency or catalyst activity is used as its label parameter.

3. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The dimensionless method in step 1 is calculated by any one of the following formulas 1 to 3: In the formula, Z ni is the normalized value of parameter Z, i is the original value of parameter Z, Z max is the maximum value of parameter Z, Z min is the minimum value of parameter Z, σ is the standard deviation of parameter Z, and μ is the mean value of parameter Z.

4. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The stationarity analysis in step 2 adopts at least one of the two methods of ADF test or KPSS test.

5. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The difference processing in step 2 refers to calculating the difference sequence using the following formula: In the formula, is the difference sequence we want, Z k is the parameter Z at time k, Z k-1 is the parameter Z at time k-1.

6. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The ARIMA model steps of step 3 are as follows: 1) Calculate and draw the ACF graph and the PACF graph of the difference sequence. According to the tailing or truncation properties of the ACF graph and the PACF graph, preliminarily determine the autoregressive order p and the moving average order q of the model; 2) using either the Akaike Information Criterion (AIC Criterion) or the Bayesian Information Criterion (BIC Criterion) to determine the order p and q of the model; 3) Based on the model order determined above, the model parameters are estimated using the maximum likelihood function method or the least squares method to preliminarily determine the model: In the formula, y t is the difference sequence to be determined, ε t is the white noise residual, a j and θ i is the model coefficient to be estimated; 4) Calculate the model residual and check whether the residual is white noise. If so, the model is successfully established. If not, repeat steps 2) to 4) until the residual is verified to be white noise.

7. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The preset value in step 4 is twice the standard deviation of the residual sequence.

8. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The preset condition of step 6 is that samples with a relative error less than 10% are classified as training samples, otherwise they are classified as bad samples, and the preset number of times is at least 3 times.

9. The bad data identification method based on time series analysis coupled neural network prediction according to claim 1 is characterized in that: The inspection and correction steps in step 7 are as follows: 1) Delete bad samples, use optimized samples to build and train BP neural network, use the trained neural network to test bad samples, and calculate relative error; 2) Screen out samples with a relative error of no less than 10% and confirm them as bad data; screen out samples with a relative error of no more than 5% and correct them as good data; 3) Repeat step 1) and step 2) for the remaining samples after screening in step 2) until any of the following conditions is met: Condition I: no sample remains after executing step 2); Condition II: Repeat the execution for no less than 3 times; 4) All remaining samples are confirmed as bad data.

Images