Multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing
By combining SSAD data anomaly detection and multi-character LSTM network, the problem of difficult to capture nonlinear and complex timing characteristics in chimney NOx concentration prediction in cement production lines is solved, and higher prediction accuracy and stability are achieved.
Patent Information
- Application Number
- CN202510378513.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-28
- Publication Date
- 2025-06-27
AI Technical Summary
The prior art is difficult to effectively capture nonlinear and complex timing characteristics in chimney NOx concentration prediction in cement production lines, and the outliers in the data are not sufficiently processed, resulting in limited prediction accuracy.
The multi-featured LSTM cement chimney NOx prediction method combined with SSAD data abnormality detection processing is adopted, and abnormal data is detected and processed through sliding window slope analysis, and a multi-layer LSTM network is constructed to capture the timing change law of NOx concentration.
It improves the accuracy and stability of chimney NOx concentration prediction, can handle complex timing data and time lag effects more effectively, and reduces the negative impact of abnormal data on the prediction results.
Smart Images
Figure CN120217889A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of gas concentration prediction methods for cement production lines, and specifically to a multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection and processing. Background Art
[0002] During the cement production process, a large amount of nitrogen oxides (NOx) is generated by the decomposition furnace combustion. These pollutants pose serious hazards to the environment and human health. The emission of NOx not only exacerbates air pollution, leading to acid rain and ozone layer depletion, but also has an adverse impact on the ecological environment around the cement plant. Therefore, real-time monitoring of the chimney NOx concentration and accurate prediction are crucial for reducing environmental pollution and optimizing the production process.
[0003] Currently, the prediction methods for chimney NOx concentration mainly rely on traditional statistical models and machine learning algorithms. Traditional statistical methods (such as autoregressive integrated moving average (ARIMA) models, regression analysis) mostly focus on linear relationship modeling and usually cannot effectively capture the non-linear and complex time-series characteristics between the NOx concentration and multiple input variables. Traditional methods also fail to fully consider the time-delay effect between different variables. In the modeling process of chimney NOx concentration, there are often lag relationships between influencing factors. For example, the change in chimney NOx concentration may lag behind the changes in variables such as the SNCR first ammonia injection pump frequency and the NOx concentration in the kiln tail flue gas chamber. Traditional methods often have difficulty effectively capturing this time-delay effect, resulting in limited prediction accuracy. In addition, the NOx concentration prediction in the cement production process is affected by multiple factors, including the SNCR first ammonia injection pump frequency, the NOx concentration in the kiln tail flue gas chamber, fuel type, climate conditions, etc. The interaction between these factors is complex and non-linear. To improve the prediction accuracy, traditional methods often rely on manual selection and extraction of features, but this method is not only time-consuming but also difficult to comprehensively explore the potential laws in the data.
[0004] In practical applications, data quality is also a key factor affecting the prediction accuracy. Due to reasons such as equipment failures, equipment maintenance, environmental changes, or errors in the data collection process, the sensor data collected often contains outliers. If these abnormal data are not effectively processed, they may have a negative impact on the model training and prediction results. Therefore, how to perform effective data anomaly detection and processing has become another key issue in improving the model accuracy and stability.
[0005] To address the above challenges, in recent years, deep learning-based methods have been widely applied in chimney NOx concentration prediction. As a deep learning model with strong temporal modeling capabilities, the Long Short-Term Memory network (LSTM) can capture the complex temporal dependencies and non-linear characteristics in the changes of chimney NOx concentration. By processing multi-dimensional input data, LSTM can more accurately model the complex relationships between input variables, thereby improving the accuracy of prediction. In addition, the LSTM network can effectively process input data with time lags, further enhancing the prediction ability of the model. However, when solely relying on the LSTM neural network for prediction, the abnormal fluctuations in the data may seriously affect the training and prediction effects of the model. Therefore, it is particularly important to combine data anomaly detection methods. Summary of the Invention
[0006] The present invention provides a multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection and processing to solve the problem of poor prediction effect in the prior art that solely relies on LSTM neural network for prediction.
[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:
[0008] A multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection and processing, comprising the following steps:
[0009] Step 1: Collect the characteristic variable data of multiple consecutive time periods in the cement production line. The characteristic variable data includes the NOx concentration in the kiln tail smoke chamber, the frequency of the SNCR No. 1 ammonia injection pump, and the chimney NOx concentration data, and preprocess the characteristic variable data. The preprocessing is as follows:
[0010] Clean the collected characteristic variable data, and then use the sliding window slope analysis method to detect the cleaned characteristic variable data to eliminate backflush abnormal data; after eliminating the backflush abnormal data, use the linear interpolation method to correct the characteristic variable data;
[0011] Use the preprocessed characteristic variable data as the data sample;
[0012] Step 2: There is a correlation between the chimney NOx concentration at time k + d, the chimney NOx concentration at time k + d - 1, the SNCR No. 1 ammonia injection pump frequency at time k, and the NOx concentration in the kiln tail delay at time k. Considering that the chimney NOx concentration at time k + d - 1 cannot be obtained at time k, through time series analysis and calculation, determine the prediction model framework. In the prediction model framework, the chimney NOx concentration at time k + d is used as the output variable, and the SNCR No. 1 ammonia injection pump frequencies at times k, k - 1... k - d + 1 and the NOx concentrations in the kiln tail flue gas chamber at times k, k - 1... k - d + 1 are used as input variables; where d is the time when the chimney NOx concentration lags behind the SNCR No. 1 ammonia injection pump frequency and the NOx concentration in the kiln tail flue gas chamber.
[0013] Step 3: Based on the prediction model framework established in Step 2, align the lag of the No. 1 pump frequency and the NOx concentration in the kiln tail flue gas chamber, which are used as input variables in the data samples obtained in Step 1, with the chimney NOx concentration used as the output variable, to obtain the aligned data samples.
[0014] Step 4: Divide the aligned data samples obtained in Step 3 into a training set and a test set, and generate an LSTM neural network; use the training set to train the LSTM neural network to obtain the LSTM neural network under the optimal model parameters;
[0015] Then, use the test set to verify the LSTM neural network under the optimal model parameters, and evaluate the accuracy of the LSTM neural network under the optimal model parameters according to the evaluation index, thereby obtaining the prediction result of the chimney NOx concentration.
[0016] Furthermore, in Step 1, when using the sliding window slope analysis method for detection, set the sliding window size to 20, and analyze 20 data points each time; set the slope change threshold to 0.1. When the slope change between adjacent windows exceeds 0.1, consider this data point as abnormal;
[0017] Use a linear regression model to calculate the slope of each window and save the slope results of each sliding window, that is, perform a linear regression on every 20 data points to calculate the slope of this window;
[0018] For two adjacent sliding windows, calculate their slope changes respectively. If the slope change is greater than the set threshold of 0.1, it is considered that there is an abnormal backflush in this data.
[0019] Furthermore, in Step 2, the determined prediction model framework is shown by the following formula:
[0020] y(k + d) = ay(k + d - 1) + bx1(k) + cx2(k)
[0021] y(k + d) = a[ay(k + d - 2) + bx1(k - 1) + cx2(k - 1)] + bx1(k) + cx2(k)
[0022] y(k + d) = a 2 [ay(k + d - 3) + bx1(k - 2) + cx2(k - 2)] + bx1(k) + abx1(k - 1) + cx2(k) + acx2(k - 1)
[0023] …………
[0024] y(k + d) = a d y(k) + bx1(k) + abx1(k - 1) +... + a d-1 bx1(k - d + 1) + cx2(k) + acx2(k - 1)... + a d-1 cx2(k - d + 1)
[0025] Where: y(k + d) is the NOx concentration of the chimney at time k + d;
[0026] y(k) is the predicted value of the NOx concentration of the chimney;
[0027] x1(k), x1(k - 1)……(k - d + 1) are the frequencies of the SNCR first ammonia injection pump at time k, k - 1……k - d + 1 respectively;
[0028] x2(k), x2(k - 1)……x2(k - d + 1) are the NOx concentrations of the NOx concentration of the kiln tail flue gas chamber at time k, k - 1……k - d + 1 respectively;
[0029] a is the influence weight of the NOx concentration of the chimney at the previous moment on the prediction result, b is the influence degree of the NOx concentration of the kiln tail flue gas chamber at the current moment on the prediction result, and c is the influence degree of the frequency of the SNCR first ammonia injection pump at the current moment on the prediction result.
[0030] Furthermore, the lag alignment process in step 3 is as follows:
[0031] Step 3.1, Determine the relationship between the input and output variables
[0032] The input variables are: The data at each time step will include the frequencies of the SNCR first ammonia injection pump from k - d + 1 to k and the NOx concentration of the kiln tail flue gas chamber, and the NOx concentration of the chimney at time k;
[0033] The output variable is: For each sample, the NOx concentration of the chimney at time k + d is used as the target output;
[0034] According to the engineering background, k + d corresponds to the future time step;
[0035] Step 3.2, Construct the time window
[0036] For each sample, the input data consists of the data of the first k + d time steps, and the output is the chimney NOx concentration at the future k + d moment. Thus, the input variables include the ammonia injection pump frequency at the first k + d time steps, the NOx concentration in the kiln tail flue gas chamber at the first k + d time steps, and the chimney NOx concentration at the k moment, and the output is the chimney NOx concentration at the future k + d moment;
[0037] Step 3.3, Determine the input format of the sample
[0038] The input of each sample includes: the ammonia injection pump frequency at the first k + d time steps, the NOx concentration in the kiln tail flue gas chamber at the first k + d time steps, and the chimney NOx concentration at the current moment k;
[0039] The output is: the chimney NOx concentration at the future k + d moment;
[0040] Step 3.4, Process the lag relationship of the data
[0041] For each time point k, the input data includes the ammonia injection pump frequency and the NOx concentration in the kiln tail flue gas chamber at the first k + d time steps, and the chimney NOx concentration at the current moment k; the output data is the chimney NOx concentration at the future k + d moment;
[0042] Thus, the input and output data are aligned through the lag relationship of time to ensure that the subsequent LSTM neural network can learn the influence of the chimney NOx concentration at the current moment on the future concentration;
[0043] Step 3.5, Generate samples using a sliding window
[0044] Use the sliding window method to gradually extract samples from the dataset; for each time point k, construct the corresponding input and output. The input includes the ammonia injection pump frequency, the NOx concentration in the kiln tail flue gas chamber at the first k + d time steps, and the chimney NOx concentration at the current moment k, and the output is the chimney NOx concentration at the future k + d moment.
[0045] Furthermore, introducing the key variables at past moments as feature inputs in step 4 helps the LSTM network better capture the long-term dependencies in the time series, thereby improving the prediction accuracy of the chimney NOx concentration at future moments, that is, introducing a time lag feature construction mechanism in the LSTM neural network.
[0046] Furthermore, in step 4, the LSTM neural network is trained using the Adam optimization algorithm and the mean squared error is used as the loss function during training.
[0047] Furthermore, in step 4, multiple rounds of training are carried out, and the trained LSTM neural network is cross-validated to verify the generalization ability of the model, thereby determining the LSTM neural network under the optimal model parameters.
[0048] Furthermore, the evaluation metrics in step 4 include the goodness of fit R 2 , root mean square error RMSE, mean absolute error MAE, and maximum error MAXE between the predicted value and the true value.
[0049] The present invention performs data preprocessing on multi-dimensional input variables (such as the frequency of the SNCR No. 1 ammonia injection pump, the NOx concentration in the chimney, the NOx concentration in the kiln tail gas chamber, etc.) collected from the cement production process to ensure high-quality input of the data. The data preprocessing includes data cleaning and outlier handling.
[0050] In data cleaning, by cleaning the input data, missing values, invalid data, and inconsistent data are removed to ensure the validity of the data. When handling outliers, the SSAD (Sliding Window Slope Anomaly Detection) method is used to detect and handle abnormal data. The time series data is segmented through the sliding window technique, and the change in the slope value within each small segment is used for outlier detection. After the outliers are corrected or removed, the data is ensured to be more in line with the actual production conditions. On this basis, the main time series features are extracted and standardized to eliminate the dimensional difference and scale effect between the data.
[0051] The present invention uses the sliding window technique and slope analysis to deeply analyze the time series data. When dividing the sliding window, through the sliding window technique, the original time series data is divided into several small segments of a fixed length, and each small segment represents the data within a time window. During slope analysis, the slope of the data within each time window is calculated to reflect the trend and rate of data change. The slope value represents the change amplitude of the data within this window, providing more accurate time series features for the LSTM neural network. These slope features help the model better understand the fluctuation law of the data over time and improve the model's perception ability of time series changes. The above process can extract the dynamic characteristics in the data and provide richer input features for the LSTM neural network.
[0052] Based on the above feature extraction, the present invention constructs a multi-layer LSTM network to capture the time series change law of the NOx concentration in the chimney. The LSTM network is suitable for processing time series data with long-term and short-term dependencies and can effectively model the non-linear dynamic changes in the data.
[0053] The LSTM network structure adopts multiple stacked LSTM units. Each layer of LSTM units can learn temporal features at different levels, thus effectively capturing the long-term and short-term dependencies of the input data. By designing network structures at different levels, complex multi-feature temporal data can be processed and the prediction accuracy can be improved.
[0054] The present invention optimizes the hyperparameters (such as learning rate, number of hidden units, number of layers, etc.) of the LSTM network through methods such as cross-validation to improve the prediction performance of the model. The LSTM network can output the predicted values of the chimney NOx concentration for a period of time in the future by learning historical data.
[0055] In the prediction of the cement chimney NOx concentration, there is usually a lag effect between input variables. For example, a change in the frequency of Pump No. 1 may affect the chimney NOx concentration several minutes later. Therefore, the present invention particularly considers these lag effects to further enhance the accuracy of the prediction model.
[0056] In the LSTM chimney NOx concentration prediction model of the present invention, the lag data of the input variables (such as the lag relationship between the chimney NOx concentration and the frequency of Pump No. 1, and the NOx concentration in the kiln tail smoke chamber) is embedded into the input sequence. By introducing the lag data, the LSTM neural network can better capture the temporal correlation and dynamic changes between variables. When processing data, the present invention selects an appropriate time lag window to ensure the effectiveness of the lag data and its support for the prediction model.
[0057] The present invention finally uses the trained LSTM neural network to predict the future chimney NOx concentration based on historical data and features.
[0058] After the LSTM neural network training of the present invention is completed, the optimized LSTM neural network is used to predict the chimney NOx concentration. Through multiple predictions and real-time data feedback, the prediction accuracy of the model is continuously optimized. Moreover, the present invention adopts an error feedback mechanism to adjust for the prediction error, reducing the systematic bias and variance of the model to ensure that the prediction results are closer to the actual situation.
[0059] The multi-feature LSTM cement chimney NOx prediction method of the present invention can be integrated into the production monitoring system of the cement plant to predict and monitor the chimney NOx concentration in real time. This method can not only provide data support for the optimization of the production process, but also provide more accurate emission control data for the environmental protection department.
[0060] The present invention combines the sliding window technique, slope analysis, and long short-term memory (LSTM) networks to construct a prediction model, which can accurately predict the NOx concentration in cement chimneys. It is particularly suitable for processing time series data, can effectively capture the complex temporal relationships between multiple features, and handle time lag effects and data anomalies, providing a decision-making basis for NOx emission control in cement plants. Compared with the prior art, the advantages of the present invention are as follows:
[0061] 1. Improving the ability to extract temporal features by combining sliding window slope analysis: By introducing the sliding window slope analysis technique, the present invention can accurately extract the changing trends and dynamic features in time series data. Different from traditional methods, the sliding window slope analysis can dynamically adjust the window size and the calculation method of the slope, enabling the model to capture the local changes in the data in real time and providing richer input features for the LSTM model. This technique can not only reflect the long-term trend changes of the NOx concentration but also help capture abnormal fluctuations and short-term changes, improving the prediction accuracy. Especially in complex production environments, it can effectively cope with rapidly changing working conditions.
[0062] 2. The multi-feature LSTM model captures long-term and short-term dependence relationships: Compared with existing traditional regression or simple neural network models, the present invention uses a multi-layer LSTM (long short-term memory) network to handle the prediction task of the NOx concentration in cement chimneys. The LSTM model has unique advantages in processing time series data and can effectively capture the long-term and short-term dependence relationships between input variables. Especially in a production environment with multiple features and variables, the LSTM model can comprehensively consider the interactions and influences between multiple input variables, making the prediction results more accurate and reliable. This advantage is particularly evident when dealing with complex non-linear time series data. Compared with traditional time series prediction methods, the present invention can provide more accurate predictions. BRIEF DESCRIPTION OF THE DRAWINGS
[0063] Figure 1 It is the effect diagram of SSAD sliding window slope analysis data anomaly detection and processing of the present invention.
[0064] Figure 2 It is the effect diagram of variable alignment. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0065] The present invention will be further described below in conjunction with the drawings and embodiments.
[0066] As Figure 1 shown, this embodiment discloses a multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection and processing, including the following steps:
[0067] Step 1: Based on the NOx generation mechanism in the chimney of the cement production line, collect the characteristic variable data of multiple consecutive time periods in the cement production line. The characteristic variable data includes the NOx concentration in the kiln tail flue gas chamber, the frequency of the first pump, and the NOx concentration data of the chimney. Then preprocess the characteristic variable data. The preprocessing process is as follows:
[0068] (1.1) Clean the collected characteristic variable data.
[0069] Obtain the time series file containing the NOx concentration data in the kiln tail flue gas chamber and load the data into a data frame. Since only the NOx concentration data in the kiln tail chimney is affected by backwashing, only the NOx concentration data in the kiln tail flue gas chamber is cleaned. For the non-numeric strings in the data, convert them to missing values, and then use forward filling to fill in the missing values to ensure the continuity and integrity of the data.
[0070] (2) In the collected data, due to the backwashing phenomenon caused by the periodic cleaning or maintenance of the sensors in the cement production line, the NOx concentration data in the kiln tail flue gas chamber is affected, resulting in abnormal NOx concentration data read by the sensors. Therefore, the present invention uses SSAD (Sliding Slope Anomaly Detection sliding window slope analysis method) to detect the cleaned characteristic variable data to eliminate the backwashing abnormal data.
[0071] Specifically, set the sliding window size to 20, and analyze 20 data points each time; set the slope change threshold to 0.1. When the slope change between adjacent windows exceeds 0.1, consider this data point as abnormal. Use a linear regression model to calculate the slope of each window and save the slope results of each sliding window. That is, perform a linear regression on every 20 data points to calculate the slope of this window.
[0072] For two adjacent sliding windows, calculate their slope changes. For example, if the slope of window 1 is 0.5 and the slope of window 2 is 0.45, then their slope change is |0.5 - 0.45| = 0.05. If the slope change is greater than the set threshold of 0.1, consider this data as abnormal. Detect and record all the moments when the slope change exceeds the threshold.
[0073] (3) After eliminating the backwashing abnormal data, use the linear interpolation method to correct the characteristic variable data to fill in the missing values, and smooth the NOx data in the kiln tail flue gas chamber to ensure the continuity and effectiveness of the data.
[0074] Specifically, use linear interpolation to correct the detected abnormal data. If a large slope change is detected in a certain sliding window, first calculate the normal change rate of the NOx concentration in the kiln tail flue gas chamber, as shown in the following formula:
[0075]
[0076] Among them: normal_slope is the slope between two points in the smoothed time series data, used to detect the rate of change of the data over time; smoothed_values is the time series data after smoothing, used to reduce noise and obtain a smoother trend; window_size is the size of the sliding window, which is the number of consecutive data points selected for calculating the slope; time_diff is the time difference between adjacent data points within the sliding window; i is the position in the smoothed time series data.
[0077] Next, linear interpolation is used to correct the NOx concentration value in the kiln tail flue gas chamber, and the abnormal data is corrected to data close to the normal slope. Suppose the concentration value of the data point in sliding window 5 is 60 at time t5 = 50, and the concentration value of the data in window 6 is 65 at time t6 = 70. Through interpolation calculation, these data points are linearly transitioned to a suitable trend value. The corrected data will replace the original abnormal data and be merged with the original data to form a new smoothed NOx concentration time series of the kiln tail flue gas chamber. Figure 1 This is the effect diagram of the SSAD sliding window slope analysis data anomaly detection and processing of the present invention.
[0078] Thus, by preprocessing the feature variable data, the preprocessed feature variable data is used as the data sample for the subsequent LSTM neural network.
[0079] Step 2: Determine the prediction model framework.
[0080] According to the process analysis of the cement production line, there is a strong correlation between the NOx concentration in the chimney at time k + d and the NOx concentration in the chimney at time k + d - 1, the frequency of the SNCR first ammonia injection pump at time k, and the NOx concentration in the kiln tail at time k. Among them, d is the time when the NOx concentration in the chimney lags behind the frequency of the SNCR first ammonia injection pump and the NOx concentration in the kiln tail flue gas chamber. Therefore, the following prediction model framework can be constructed:
[0081] y(k + d) = ay(k + d - 1) + bx1(k) + cx2(k)
[0082] Among them: y(k + d) is the NOx concentration in the chimney at time k + d; x1(k) is the frequency of the first pump at time k; x2(k) is the NOx concentration in the kiln tail flue gas chamber at time k; y(k + d - 1) is the NOx concentration in the chimney at time k + d - 1;
[0083] a is the influence weight of the NOx concentration in the chimney at the previous moment on the prediction result.
[0084] b is the influence degree of the NOx concentration in the kiln tail flue gas chamber at the current moment on the prediction result.
[0085] c is the influence degree of the frequency of the SNCR No. 1 ammonia injection pump at the current moment on the prediction result.
[0086] However, in actual engineering, the chimney NOx concentration at the moment of k + d - 1 cannot be obtained at the moment of k. Therefore, through time series analysis and calculation in this embodiment, the final prediction model can be obtained as follows:
[0087] y(k + d) = ay(k + d - 1) + bx1(k) + cx2(k)
[0088] y(k + d) = a[ay(k + d - 2) + bx1(k - 1) + cx2(k - 1)] + bx1(k) + cx2(k)
[0089] y(k + d) = a 2 [ay(k + d - 3) + bx1(k - 2) + cx2(k - 2)] + bx1(k) + abx1(k - 1)
[0090] + cx2(k) + acx2(k - 1)
[0091] And so on...
[0092] y(k + d) = a d y(k) + bx1(k) + abx1(k - 1) +... + a d-1 bx1(k - d + 1) + cx2(k)
[0093] + acx2(k - 1)... + a d-1 cx2(k - d + 1)
[0094] Among them: y(k + d) is the chimney NOx concentration at the moment of k + d;
[0095] y(k) is the predicted value of the chimney NOx concentration;
[0096] x1(k), x1(k - 1) …… (k - d + 1) are the frequencies of the SNCR No. 1 ammonia injection pump at the moments of k, k - 1 …… k - d + 1 respectively;
[0097] x2(k), x2(k - 1), ……, x2(k - d + 1) are the chimney NOx concentrations in the tail gas chamber of the kiln at the moments of k, k - 1 …… k - d + 1 respectively;
[0098] a is the influence weight of the chimney NOx concentration at the previous moment on the prediction result, b is the influence degree of the chimney NOx concentration in the tail gas chamber of the kiln at the current moment on the prediction result, and c is the influence degree of the frequency of the SNCR No. 1 ammonia injection pump at the current moment on the prediction result.
[0099] In the prediction model framework, the NOx concentration of the chimney at time k + d is used as the output variable, and the frequencies of the SNCR No. 1 ammonia injection pump at times k, k - 1, …, k - d + 1 and the NOx concentrations in the NOx chamber at the kiln tail at times k, k - 1, …, k - d + 1 are used as input variables. As shown by the formula of the prediction model framework, the NOx concentration of the chimney at time k + d will be predicted by the LSTM neural network with three variables as the subsequent input of the LSTM neural network: the frequencies of the SNCR No. 1 ammonia injection pump and the NOx concentrations in the NOx chamber at the kiln tail from time k - d + 1 to time k, and the NOx concentration of the chimney at time k.
[0100] Step 3. Based on the prediction model framework established in Step 2, align the lag of the No. 1 pump frequency and the NOx concentration in the NOx chamber at the kiln tail, which are used as input variables, and the NOx concentration of the chimney, which is used as the output variable, in the data samples obtained in Step 1 to obtain the aligned data samples for LSTM neural network training. Figure 2 It is the variable alignment effect diagram. The specific process is as follows:
[0101] Step 3.1. Determine the relationship between the input and output variables
[0102] The input variables are: the data at each time step will include the frequencies of the SNCR No. 1 ammonia injection pump from time k - d + 1 to time k and the NOx concentration in the NOx chamber at the kiln tail, and the NOx concentration of the chimney at time k (the current time k).
[0103] The output variable is: for each sample, the NOx concentration of the chimney at time k + d is used as the target output.
[0104] According to the engineering background, k + d corresponds to the future time step, and k + d is set to 180 seconds.
[0105] Step 3.2. Construct the time window
[0106] For each sample, the input data consists of the data of the first k + d (180) time steps, and the output is the NOx concentration of the chimney at the future time k + d (the future 180 seconds). Therefore, the input variables include the ammonia injection pump frequencies in the first k + d (first 180 seconds), the NOx concentrations in the NOx chamber at the kiln tail in the first k + d (first 180 seconds), and the NOx concentration of the chimney at time k, and the output is the NOx concentration of the chimney at the future time k + d (the future 180 seconds).
[0107] Step 3.3. Determine the input format of the sample
[0108] The input of each sample includes: the ammonia injection pump frequencies in the first k + d (first 180) time steps, the NOx concentrations in the NOx chamber at the kiln tail in the first k + d (first 180) time steps, and the NOx concentration of the chimney at the current time k;
[0109] The output is the chimney NOx concentration at the future k + d moment (the future 180 seconds).
[0110] Step 3.4, Process the lag relationship of the data
[0111] For each time point k, the input data includes the ammonia injection pump frequency and the NOx concentration in the kiln tail flue gas chamber at the previous k + d (the previous 180) time steps, as well as the chimney NOx concentration at the current moment k; the output data is the chimney NOx concentration at the future k + d moment (the future 180 seconds).
[0112] Thus, the input and output data are aligned through the lag relationship of time to ensure that the subsequent LSTM neural network can learn the influence of the chimney NOx concentration at the current moment on the future concentration.
[0113] Step 3.5, Generate samples using a sliding window
[0114] Use the sliding window method to gradually extract samples from the dataset; for each time point k, construct the corresponding input and output. The input includes the ammonia injection pump frequency, the NOx concentration in the kiln tail flue gas chamber at the previous k + d (the previous 180) time steps, and the chimney NOx concentration at the current moment k, and the output is the chimney NOx concentration at the future k + d moment (the future 180 seconds). Specifically, select the ammonia injection pump frequency and the NOx concentration in the kiln tail flue gas chamber from the k - d + 1 moment to the k moment in the data, use the chimney NOx concentration at the k moment, and take the chimney NOx concentration at the k + d moment as the output target.
[0115] Step 4, Divide the aligned data samples obtained in Step 3 into a training set and a test set, generate an LSTM neural network, and use the training set to train the LSTM neural network to obtain the LSTM neural network under the optimal model parameters.
[0116] In this embodiment, the LSTM neural network includes multiple LSTM layers, and a time lag feature introduction mechanism is introduced into the LSTM to fully learn the temporal relationship between different features. The Adam optimization algorithm is used for training when the LSTM neural network is trained, and the loss function during training is the mean squared error (MSE).
[0117] The specific process of constructing the lag feature is as follows:
[0118] In the LSTM model, the core idea of the lag feature is to help the model capture the temporal dependence relationship in the time series data by introducing the data of past time steps. It includes input feature construction and target output construction.
[0119] The construction of input features ensures that the input for each sample includes the SNCR ammonia injection pump frequency at the past lag moments and the NOx concentration data in the kiln tail flue gas chamber, as well as the NOx concentration in the chimney at the current moment. The construction of the target output ensures that the target output is the NOx concentration in the chimney at the future lag moments. Here, lag represents the time when the NOx concentration value in the chimney lags behind the ammonia injection pump frequency of the first unit and the NOx concentration in the kiln tail flue gas chamber. According to the analysis of the actual engineering process, lag is considered to be 180 seconds.
[0120] The original data includes the ammonia injection pump frequency u1(t) of the first unit, the NOx concentration u2(t) in the kiln tail flue gas chamber, and the NOx concentration u3(t) in the chimney, where t is the time index.
[0121] The input features are constructed as follows:
[0122] X t = [u1(t - lag), u1(t - lag + 1),..., u1(t - 1), u2(t - lag), u2(t - lag + 1),..., u2(t - 1), u3(t)]
[0123] Among them, lag = 180 seconds, indicating that the ammonia injection pump frequency of the first unit, the NOx concentration in the kiln tail flue gas chamber, and the NOx concentration in the chimney at the past 180 time steps (i.e., the past 180 seconds) are used as the input.
[0124] The target output Y t represents the NOx concentration in the chimney at the future 180 seconds.
[0125] Y t = u3(t + lag)
[0126] Among them, u3(t + lag) represents the NOx concentration in the chimney at t + lag, that is, at the future 180 seconds.
[0127] For a series of samples, the target output matrix Y is:
[0128] Y = [u3(1 + lag), u3(2 + lag),..., u3(N)]
[0129] In this embodiment, the LSTM neural network is trained for multiple rounds, and the trained LSTM neural network model is cross-validated to verify the generalization ability of the model and calculate the accuracy of the model, thereby obtaining the LSTM neural network under the optimal model parameters.
[0130] Then, the LSTM neural network under the optimal model parameters is verified using the test set, and the accuracy of the LSTM neural network under the optimal model parameters is evaluated according to the evaluation index, thereby obtaining the prediction result of the NOx concentration in the chimney. The evaluation index includes the goodness of fit R 2, Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Maximum Error (MAXE) between the predicted value and the true value.
[0131] The preferred embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Among the various specific technical features described in the above specific embodiments, they can be combined in any suitable manner without contradiction. As long as such a combination does not violate the idea of the present invention, it should also be regarded as the content disclosed in this disclosure. To avoid unnecessary repetition, the present invention will not separately describe various possible combination methods.
[0132] The present invention is not limited to the specific details in the above embodiments. Within the technical concept scope of the present invention and without departing from the design idea of the present invention, various variations and improvements made by those skilled in the art to the technical solution of the present invention should all fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.
Claims
1. A multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing, characterized in that: The following steps are involved: Step 1: Collect characteristic variable data of multiple continuous time periods in the cement production line, including NOx concentration in the kiln tail smoke chamber, frequency of SNCR No. 1 ammonia injection pump and chimney NOx concentration data, and pre-process the characteristic variable data, the pre-processing is as follows: The collected characteristic variable data are cleaned, and then the sliding window slope analysis method is used to detect the cleaned characteristic variable data to eliminate the backwash abnormal data; after eliminating the backwash abnormal data, the linear interpolation method is used to correct the characteristic variable data; The preprocessed feature variable data is used as data sample; Step 2: Based on the correlation between the chimney NOx concentration at time k+d and the chimney NOx concentration at time k+d-1, the frequency of the No. 1 ammonia injection pump of SNCR at time k, and the delayed NOx concentration at the kiln tail at time k, and considering that the chimney NOx concentration at time k+d-1 cannot be obtained at time k, the prediction model framework is determined through time series analysis and calculation. In the prediction model framework, the chimney NOx concentration at time k+d is used as the output variable, and the frequency of the No. 1 ammonia injection pump of SNCR at time k, time k-1... time k-d+1 and the NOx concentration in the kiln tail smoke chamber at time k, time k-1... time k-d+1 are used as input variables; wherein d is the time that the chimney NOx concentration lags behind the frequency of the No. 1 ammonia injection pump of SNCR and the NOx concentration in the kiln tail smoke chamber; Step 3: Based on the prediction model framework established in step 2, the frequency of pump No. 1 and the NOx concentration in the kiln tail smoke chamber as input variables in the data sample obtained in step 1 are lagged and aligned with the chimney NOx concentration as the output variable to obtain the aligned data sample; Step 4: Divide the aligned data samples obtained in step 3 into a training set and a test set, and generate an LSTM neural network; use the training set to train the LSTM neural network to obtain the LSTM neural network under the optimal model parameters; Then, the test set was used to verify the LSTM neural network under the optimal model parameters, and the accuracy of the LSTM neural network under the optimal model parameters was evaluated according to the evaluation indicators, thereby obtaining the prediction results of chimney NOx concentration.
2. According to claim 1, a multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing is characterized in that: In step 1, when the sliding window slope analysis method is used for detection, the sliding window size is set to 20, and 20 data points are analyzed each time; the slope change threshold is set to 0.1, and when the slope change of adjacent windows exceeds 0.1, the data point is considered abnormal; Use the linear regression model to calculate the slope of each window and save the slope result of each sliding window, that is, perform a linear regression on every 20 data points to calculate the slope of the window; For two adjacent sliding windows, their slope changes are calculated respectively. If the slope change is greater than the set threshold value of 0.1, it is considered that the data has a backwash anomaly.
3. According to claim 1, a multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing is characterized in that: In step 2, the prediction model framework is determined as shown in the following formula: y(k+d)=ay(k+d-1)+bx1(k)+cx2(k) y(k+d)=a[ay(k+d-2)+bx1(k-1)+cx2(k-1)]+bx1(k)+cx2(k) y(k+d)=a 2 [ay(k+d-3)+bx1(k-2)+cx2(k-2)]+bx1(k)+abx1(k-1) +cx2(k)+acx2(k-1) ………… y(k+d)=a d y(k)+bx1(k)+abx1(k-1)+...+a d-1 bx1(k-d+1)+cx2(k) +acx2(k-1)...+a d-1 cx2(k-d+1) Where: y(k+d) is the chimney NOx concentration at time k+d; y(k) is the predicted value of chimney NOx concentration; x1(k), x1(k-1)…(k-d+1) are the frequencies of SNCR No. 1 ammonia injection pump at time k, time k-1…k-d+1 respectively; x2(k), x2(k-1)…x2(k-d+1) are the NOx concentrations in the kiln tail smoke chamber at time k, time k-1…time k-d+1 respectively; a is the influence weight of the chimney NOx concentration at the previous moment on the prediction result, b is the influence degree of the NOx concentration in the kiln tail smoke chamber at the current moment on the prediction result, and c is the influence degree of the SNCR No. 1 ammonia injection pump frequency at the current moment on the prediction result.
4. The multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing according to claim 1 is characterized in that: The lag alignment process in step 3 is as follows: Step 3.1: Determine the relationship between input and output variables The input variables are: the data of each time step will include the frequency of SNCR No. 1 ammonia injection pump and the NOx concentration in the kiln tail smoke chamber from k-d+1 to k, and the NOx concentration in the chimney at k; The output variables are: for each sample, the chimney NOx concentration at time k+d is used as the target output; According to the engineering context, k + d corresponds to the future time step; Step 3.2: Build time window For each sample, the input data consists of the data of the previous k+d time steps, and the output is the chimney NOx concentration at the future k+d time step. Therefore, the input variables include the ammonia injection pump frequency at the previous k+d time step, the NOx concentration in the kiln tail smoke chamber at the previous k+d time step, and the chimney NOx concentration at time k. The output is the chimney NOx concentration at the future k+d time step. Step 3.3: Determine the sample input format The input of each sample includes: the frequency of the ammonia injection pump in the previous k+d time steps, the NOx concentration in the kiln tail smoke chamber in the previous k+d time steps, and the NOx concentration in the chimney at the current time k; The output is: chimney NOx concentration at time k+d in the future; Step 3.4: Processing the lag relationship of data For each time point k, the input data includes the ammonia injection pump frequency and the NOx concentration in the kiln tail smoke chamber in the previous k+d time steps, as well as the chimney NOx concentration at the current time k; the output data is the chimney NOx concentration at the future k+d time steps; Thus, the input and output data are aligned through the time lag relationship, ensuring that the subsequent LSTM neural network can learn the impact of the current chimney NOx concentration on the future concentration; Step 3.5: Sliding window generates samples Using the sliding window method, samples are gradually extracted from the data set. For each time point k, the corresponding input and output are constructed. The input includes the ammonia injection pump frequency, the NOx concentration in the kiln tail smoke chamber and the chimney NOx concentration at the current time k in the previous k+d time steps. The output is the chimney NOx concentration at the future k+d time steps.
5. The multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing according to claim 1 is characterized in that: Introducing key variables from the past as feature inputs as described in step 4 helps the LSTM network better capture long-term dependencies in time series, thereby improving the prediction accuracy of chimney NOx concentrations at future moments, that is, introducing a time lag feature construction mechanism into the LSTM neural network.
6. The multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing according to claim 1 is characterized in that: In step 4, the LSTM neural network is trained using the Adam optimization algorithm, and the mean square error is used as the loss function during training.
7. The multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing according to claim 1 is characterized in that: In step 4, multiple rounds of training are performed, and the trained LSTM neural network is cross-validated to verify the generalization ability of the model, thereby determining the LSTM neural network under the optimal model parameters.
8. The multi-feature LSTM cement chimney NOx prediction method combined with SSAD data anomaly detection processing according to claim 1 is characterized in that: The evaluation indicators described in step 4 include the goodness of fit R 2 , root mean square error RMSE, mean absolute error MAE, and the maximum error MAXE between the predicted value and the true value.