Intelligent ammonia injection denitration prediction system based on time sequence neural network model

Abstract

The application is suitable for the technical field of NOx concentration prediction, and provides an intelligent ammonia injection denitration prediction system based on a time sequence neural network model, which comprises the following steps: S1, data acquisition; S2, data processing, the importance of the data is preliminarily calculated through simple multiple linear regression analysis; the application solves the problems that the current denitration ammonia injection system adopts fixed ammonia distribution, cannot adapt to the variability of denitration inlet flue gas NOx under the condition of unit flexibility peak shaving, and has problems such as excessive ammonia injection and insufficient ammonia injection amount, under the condition of meeting the unit outlet emission, excessive ammonia injection will cause the increase of consumables and the problems such as the blockage of the boiler air preheater, and too little ammonia injection amount cannot meet the NOx emission standard, the accuracy of improving the inlet NOx concentration is achieved, the average error of the predicted value and the reference value of the inlet NOx concentration is reduced, and the power plant is effectively helped to control the ammonia injection amount in real time according to the prediction result, so as to realize intelligent quantitative ammonia injection.

Description

Technical Field

[0001] This invention relates to the field of NOx concentration prediction technology, and more specifically, to an intelligent ammonia injection denitrification prediction system based on a time-series neural network model. Background Technology

[0002] NOx (nitrogen oxides), one of the air pollutants emitted from coal-fired power plant boilers, not only causes a series of cardiopulmonary and eye diseases but also generates chemicals harmful to the ozone layer. Therefore, NOx emissions have become a key indicator for evaluating the quality and environmental friendliness of coal-fired boilers. With increasing environmental awareness and environmental policies restricting NOx emissions, low-NOx combustion technologies and mature denitrification technologies are widely used in my country's coal-fired power plants to reduce NOx emissions. Selective catalytic reduction (SCR) flue gas denitrification is a technology that uses a catalyst to chemically react harmful NOx and ammonia to produce pollution-free nitrogen and water, thereby reducing NOx emissions. SCR is characterized by high denitrification efficiency and mature technology, and is currently widely used in large power plant boilers. Although SCR ammonia injection systems have achieved success in removing NOx emissions from coal-fired power plant boilers (up to 90%), reducing NOx emissions from coal combustion remains crucial in the face of increasingly stringent environmental requirements. However, current denitrification ammonia injection systems use fixed ammonia distribution, which cannot adapt to the variability of NOx in the denitrification inlet flue gas under flexible peak-shaving conditions, resulting in problems such as excessive ammonia injection and insufficient ammonia injection. Excessive ammonia injection, while meeting unit outlet emission standards, leads to increased consumables and blockage of the boiler air preheater, while insufficient ammonia injection fails to meet NOx emission standards. To effectively control NOx concentration in SCR emissions, accurate and timely prediction of the inlet NOx concentration is essential. This prediction serves as a crucial parameter for ammonia injection feedforward and participates in the closed-loop control of the total ammonia injection. The control system adjusts the ammonia injection rate in real time based on the prediction results to achieve intelligent quantitative ammonia injection. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an intelligent ammonia injection denitrification prediction system based on a time-series neural network model.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent ammonia injection denitrification prediction system based on a time-series neural network model, comprising the following steps:

[0005] S1, Data Acquisition;

[0006] S2. Data processing: Initially calculate the importance ranking of the data through simple multiple linear regression analysis;

[0007] S3. Feature selection: Use the LASSO regression model to eliminate collinearity in the model and calculate the feature importance ranking of the data.

[0008] S4. Model training and parameter optimization: using difference-based features and baseline-based models. In the non-blowback phase, the model uses a long short-term memory network as the main body, and in the blowback phase, a time-series Transformer model is used.

[0009] S5. Use ablation experiments to determine the importance of features;

[0010] S6. Conduct actual deployment during on-site forecasting.

[0011] The present invention is further configured such that: the dataset used for data collection in step S1 is the values ​​of power plant-related features within 10 days, totaling 71 statistical variables.

[0012] The present invention is further configured such that: the data processing in step S2 specifically includes multiple linear regression analysis and collinearity analysis between features;

[0013] Multiple linear regression analysis: First, determine the true NOx concentration at this moment as the regressor of the problem, and use the remaining feature values ​​as independent variables. Use the least squares method (OLS) to perform multiple linear regression analysis and calculate the regression coefficients of each feature.

[0014] Assuming the NOx concentration is f(x), the above model satisfies a linear relationship: f(x) = w T Given a set of data D = {(x1, y1), (x2, y2), ..., (x...b}, where x...b is the sum of x and y1, and y2 is the sum of x and y2 ... n y n In the case of )}, the goal is to minimize the overall sum of squared errors, i.e., to find a set (w) * b * This minimizes the loss J.

[0015] The loss function for linear regression is

[0016]

[0017] The present invention is further configured as follows: collinearity analysis among features: each explanatory variable in Equation 3-1 has a variance expansion factor, which reflects the extent to which multicollinearity increases the variance of the estimated coefficients. The variance expansion factor method is used to analyze the collinearity of the variables.

[0018] The present invention is further configured such that: the feature selection in step S3 specifically includes:

[0019] LASSO ridge regression with L1 regularization has a loss function of:

[0020]

[0021] Ridge regression with L2 regularization has the following loss function:

[0022]

[0023] The present invention is further configured such that: the model training and parameter optimization in step S4 specifically includes:

[0024] Based on the difference-based feature and the baseline-based model, the algorithm takes the change of the input feature and first observes the collinearity of the features and the VIF of each feature after taking the difference. The algorithm outputs the difference between the measured value and the true value at this time according to the change of the input.

[0025] The NOx concentration prediction model based on a time-series neural network uses a Long Short-Term Memory (LSTM) network as the main body of the model in the non-backflush phase and a time-series Transformer in the backflush phase. In the non-backflush phase, the model has a data preprocessing layer at the beginning, which includes data smoothing and data normalization, followed by an LSM unit containing 400 hidden units, and finally a fully connected layer to output a specific value. The learning rate is 0.0016, the batch size during prediction is 1024, the batch size during testing is 512, and the loss function used during training is MAE. Finally, the NOx concentration prediction value is obtained by adding the difference between the model's predictions and the measured value.

[0026] Apart from the difference in input dimensions, the training process for the non-backflush stage model and the backflush stage model is the same. However, the backflush stage uses the Transformer model. Since there is no NOx concentration measurement value in the input, the NOx concentration prediction value is directly output after passing through the Transformer.

[0027] The present invention is further configured such that the ablation experiment in step S5 specifically includes:

[0028] Because time-series information is incorporated, important features may differ from those previously selected. Therefore, ablation experiments are used for feature selection.

[0029] First, train the model on the entire dataset and use the error of the model as a baseline;

[0030] Subsequently, each time one feature is deleted, the difference in accuracy of the retrained model is observed. If the feature accuracy drops too much, it indicates that the feature is very important; conversely, if the error does not drop significantly, it indicates that the feature is not important or that the feature is collinear. Therefore, a certain number of process-related additional features need to be added on this basis.

[0031] The present invention is further configured such that the actual deployment method in step S6 is as follows:

[0032] S6-1. During field operation, the DCS transmits data at intervals of 0.8s, but sometimes it exceeds 1s. If the interval exceeds 1s for a long time, frame loss will occur after a period of time. If the average transmission interval is (1+t) seconds, then frame loss will occur on average every 1 / t seconds. In this case, the value at a certain moment is missing. Therefore, the data before the algorithm input needs to be processed in a certain way. The processing method is as follows.

[0033] S6-2. The algorithm stores all the received data in memory and checks in real time whether the time is continuous. However, in order to ensure processing speed, the algorithm only maintains data for the past hour.

[0034] S6-3. If the recent data is continuous, it is directly input into the model. If it is not continuous, linear interpolation or supplementation based on the data of the previous frame is used to ensure that each timestamp in the input window has a corresponding value.

[0035] The advantages of this invention are that it significantly improves the accuracy of NOx concentration at the inlet, reduces the average error between the predicted and reference NOx concentration values, and effectively helps power plants control ammonia injection in real time based on prediction results, thereby achieving intelligent quantitative ammonia injection. Deep learning, a cutting-edge technology in time series prediction, can be used to construct a hybrid model for real-time prediction of NOx concentration at the inlet, thereby calculating the amount of ammonia injected. This model serves as a crucial parameter for ammonia injection feedforward and participates in the closed-loop control of the total ammonia injection volume. This approach offers high prediction accuracy and low delay in calculating the ammonia injection volume, reducing the amount of ammonia injected and NOx escape, thus protecting the ecological environment while improving economic efficiency. Attached Figure Description

[0036] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0037] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0038] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0039] In this invention, unless otherwise stated, the directional terms such as "up" and "down" generally refer to the directions shown in the accompanying drawings, or to the vertical, perpendicular, or gravitational direction; similarly, for ease of understanding and description, "left" and "right" generally refer to the left and right shown in the accompanying drawings; "inner" and "outer" refer to the inner and outer contours of each component itself, but the above directional terms are not intended to limit this invention.

[0040] Example

[0041] Please see Figure 1 The present invention provides the following technical solution: an intelligent ammonia injection denitrification prediction system based on a time-series neural network model, comprising the following steps:

[0042] S1, Data Acquisition;

[0043] S2. Data processing: Initially calculate the importance ranking of the data through simple multiple linear regression analysis;

[0044] S3. Feature selection: Use the LASSO regression model to eliminate collinearity in the model and calculate the feature importance ranking of the data.

[0045] S4. Model training and parameter optimization: using difference-based features and baseline-based models. In the non-blowback phase, the model uses a long short-term memory network as the main body, and in the blowback phase, a time-series Transformer model is used.

[0046] S5. Use ablation experiments to determine the importance of features;

[0047] S6. Conduct actual deployment during on-site forecasting.

[0048] The dataset used for data collection in step S1 consists of the values ​​of power plant-related characteristics over 10 days, totaling 71 statistical variables.

[0049] The data processing in step S2 specifically includes multiple linear regression analysis and collinearity analysis between features;

[0050] Multiple linear regression analysis: First, determine the true NOx concentration at this moment as the regressor of the problem, and use the remaining feature values ​​as independent variables. Use the least squares method (OLS) to perform multiple linear regression analysis and calculate the regression coefficients of each feature.

[0051] Assuming the NOx concentration is f(x), the above model satisfies a linear relationship: f(x) = w T Given a set of data D = {(x1, y1), (x2, y2), ..., (x...b}, where x...b is the sum of x and y1, and y2 is the sum of x and y2 ... n y n In the case of )}, the goal is to minimize the overall sum of squared errors, i.e., to find a set (w)* b * This minimizes the loss J.

[0052] The loss function for linear regression is

[0053]

[0054] Multicollinearity analysis among features: The variance expansion factor method is a method to determine whether multicollinearity exists by examining the extent to which a given explanatory variable is explained by all other explanatory variables in the equation. According to Equation 3-1, each explanatory variable has a variance expansion factor, which reflects the extent to which multicollinearity increases the variance of the estimated coefficients. The variance expansion factor method is used to analyze the multicollinearity of variables.

[0055] The feature selection in step S3 specifically includes:

[0056] LASSO ridge regression with L1 regularization has a loss function of:

[0057]

[0058] Ridge regression with L2 regularization has the following loss function:

[0059]

[0060] When using LASSO regression, the λ=50 model performed well, identifying the following important features.

[0061] The primary air volume at the inlet of the coal mill 3C (after selection), the cold secondary air flow at the outlet of the blower A (after selection), the burnout air flow at the right wall of the boiler, the hot primary air flow at the pyrolysis furnace of Unit 3 (after calculation), the nitrogen dioxide inlet concentration (6% oxygen), the feedwater flow at the outlet of the high-pressure heater (after selection), and the total boiler air volume—these characteristics play an irreplaceable role in estimating the final value. Furthermore, by observing the positive and negative values ​​of these characteristic coefficients, we can see their influence on the true value.

[0062] Step S4, model training and parameter optimization, specifically includes:

[0063] Based on the difference-based feature and the baseline-based model, the algorithm takes the change of the input feature and first observes the collinearity of the features and the VIF of each feature after taking the difference. The algorithm outputs the difference between the measured value and the true value at this time according to the change of the input.

[0064] After using feature interpolation, collinearity is significantly reduced, making many features beneficial for model prediction. Therefore, when using the LASSO model, a large penalty term λ reduces the number of features used and affects model accuracy; λ=1 yields better results. In this regression model, 42 feature coefficients are 0, thus the model selects the following 30 features:

[0065] Coal mill 3C outlet air-coal mixture temperature (after selection), coal mill 3C inlet primary air volume (after selection), coal mill D inlet primary air volume (after selection), coal mill 3F outlet air-coal mixture temperature (after selection), coal mill 3G inlet primary air volume (after selection), blower A outlet cold secondary air flow rate (after selection), blower B outlet cold secondary air flow rate (after selection), boiler right side hot secondary air header temperature, boiler left wall burnout air flow rate, boiler right wall burnout air flow rate, third-stage superheater outlet steam temperature C (after selection), third-stage superheater outlet steam temperature B (after selection), third-stage superheater outlet steam temperature D (after selection), superheater first-stage desuperheater inlet... Steam flow rate A (after selection), superheater first-stage desuperheater inlet steam flow rate B (after selection), superheater first-stage desuperheater inlet steam flow rate D (after selection), total coal quantity before correction, pulverizer 3C current (B phase), LDC generator actual power, No. 3 unit pyrolysis furnace hot primary air flow rate {after calculation}, No. 3 furnace reactor inlet flue gas temperature (after selection), nitrogen dioxide inlet concentration (6% oxygen content), No. 3 unit reactor L inlet and outlet flue gas differential pressure, No. 3 unit reactor R inlet and outlet flue gas differential pressure, economizer inlet feedwater temperature, furnace outlet flue gas oxygen content average value, boiler total air volume, coal feeder C instantaneous coal quantity feedback, coal feeder G instantaneous coal quantity feedback, unit AGC target value.

[0066] Treating this problem as a time series problem helps the model's prediction. A time window of 60 seconds was chosen, which achieves a balance between algorithm accuracy and prediction complexity, and is greater than the response time of all features.

[0067] The NOx concentration prediction model based on a time-series neural network uses a Long Short-Term Memory (LSTM) network as the main body of the model in the non-backflush phase and a time-series Transformer in the backflush phase. In the non-backflush phase, the model has a data preprocessing layer at the beginning, which includes data smoothing and data normalization, followed by an LSM unit containing 400 hidden units, and finally a fully connected layer to output a specific value. The learning rate is 0.0016, the batch size during prediction is 1024, the batch size during testing is 512, and the loss function used during training is MAE. Finally, the NOx concentration prediction value is obtained by adding the difference between the model's predictions and the measured value.

[0068] Apart from the difference in input dimensions, the training process for the non-backflush stage model and the backflush stage model is the same. However, the backflush stage uses the Transformer model. Since there is no NOx concentration measurement value in the input, it directly outputs the predicted NOx concentration value after passing through the Transformer.

[0069] The ablation experiment in step S5 specifically includes:

[0070] Because time-series information is incorporated, important features may differ from those previously selected. Therefore, ablation experiments are used for feature selection.

[0071] First, train the model on the entire dataset and use the error of the model as a baseline;

[0072] Then, each time one feature is deleted, the difference in accuracy of the retrained model is observed. If the feature accuracy drops too much, it means that the feature is very important; conversely, if the error does not drop significantly, it means that the feature is not important or that the feature is collinear. Therefore, some process-important features need to be added on this basis.

[0073] Thirty key features were selected from all known features, and 12 additional process features were added to form the 42 features used in the model, as follows: "Average oxygen content in flue gas at furnace outlet", "Differential pressure of flue gas at inlet and outlet of reactor L of Unit 3", "Differential pressure of flue gas at inlet and outlet of reactor R of Unit 3", "O2 content at the inlet of reactor 3", "Flue gas flow rate at the inlet of reactor 3 (Pitto tube)", "Actual power output of LDC generator", "Nitrogen dioxide inlet concentration (6% oxygen)", "Feedwater flow rate at high pressure heater outlet (after selection)", "Heat on the right side of boiler". "Secondary air header temperature", "Boiler right side hot secondary air header pressure (after selection)", "Boiler right wall burnout air flow rate", "Boiler total air volume", "Boiler left side hot secondary air header temperature", "Boiler left side hot secondary air header pressure (after selection)", "Boiler left wall burnout air flow rate", "Superheater first stage desuperheater inlet steam flow rate A (after selection)", "Superheater first stage desuperheater inlet steam flow rate B (after selection)", "Superheater first stage desuperheater inlet steam flow rate C (after selection)", "Superheater first stage desuperheater inlet steam flow rate D (after selection)", "Unit A" "GC target value", "Primary air volume at the inlet of coal mill 3A (after selection)", "Primary air volume at the inlet of coal mill 3B (after selection)", "Primary air volume at the inlet of coal mill 3C (after selection)", "Primary air volume at the inlet of coal mill 3E (after selection)", "Primary air volume at the inlet of coal mill 3F (after selection)", "Primary air volume at the inlet of coal mill 3G (after selection)", "Primary air pressure at the inlet of coal mill A", "Primary air pressure at the inlet of coal mill B", "Primary air pressure at the inlet of coal mill C", "Primary air volume at the inlet of coal mill D (after selection)", "Coal mill "D Inlet Primary Air Pressure", "E Inlet Primary Air Pressure of Coal Mill", "F Inlet Primary Air Pressure of Coal Mill", "G Inlet Primary Air Pressure of Coal Mill", "Steam Temperature at the Outlet of the Third-Stage Superheater A (after selection)", "Steam Temperature at the Outlet of the Third-Stage Superheater B (after selection)", "Steam Temperature at the Outlet of the Third-Stage Superheater C (after selection)", "Steam Temperature at the Outlet of the Third-Stage Superheater D (after selection)", "Economizer Inlet Feedwater Temperature", "Cold Secondary Air Flow Rate at the Outlet of Blower A (after selection)", "Cold Secondary Air Flow Rate at the Outlet of Blower B (after selection)", "Total Coal Quantity Before Correction".

[0074] In on-site prediction, it is necessary to combine the prediction with data transmitted from DCS. The arrival time of DCS data is closely related to the algorithm's prediction. Furthermore, since on-site operation may be unstable, if the transmission time interval is not fixed, unexpected situations such as frame loss may occur. This section explains how to handle the uncertainty of data arrival time and how to deal with data frame loss.

[0075] The specific deployment method in step S6 is as follows:

[0076] S6-1. During field operation, the DCS transmits data at intervals of 0.8s, but sometimes this interval exceeds 1s. If the interval exceeds 1s for an extended period, frame drops will occur after a while. When the transmission interval is less than 1s, the algorithm checks the timestamp changes in the data at a frequency of 1ms during program idle time. If a change occurs, the following program is executed: Assuming a message at time t_with_ms (in milliseconds) is received, and a message at time t_with_ms+800ms is received in the next moment, since the algorithm requires data at the second level, the first message t_with_ms is first converted to t_with_s. The processing method is to directly discard the following milliseconds and treat them as the data of this second. Then, when the next message t_with_ms+800ms arrives, the difference between t_with_s and t_with_ms+800ms is compared. If it is greater than 0.7s, then t_with_ms+800ms is rounded down to the new t_with_s. If it is less than 0.7s, the current t_with_s is continued, and the algorithm waits. This processing method ensures that the algorithm can capture timestamp changes at a very high frequency, and for data senders with uncertain sending times, it can robustly control the long-term timestamp error to within 1 second.

[0077] The handling method when the sending interval is greater than 1 second is as follows: If the sending interval is greater than 1 second, then frame loss will inevitably occur. If the average sending interval is (1+t) seconds, then frame loss will occur on average every 1 / t seconds. In this case, the value at a certain moment is missing. If it is input into the algorithm without processing, the algorithm will output an infinite value or even report an error due to invalid input. Therefore, it is necessary to process the data before inputting it into the algorithm. The processing method is as follows:

[0078] S6-2. The algorithm stores all the received data in memory and checks in real time whether the time is continuous. However, in order to ensure processing speed, the algorithm only maintains data for the past hour.

[0079] S6-3. If the recent data is continuous, it is directly input into the model. If it is not continuous, linear interpolation or supplementation based on the data of the previous frame is used to ensure that each timestamp in the input window has a corresponding value.

[0080] Therefore, although a large amount of missing data may affect the model's prediction accuracy to some extent, it will not cause errors during operation.

[0081] Obviously, the embodiments described above are merely some, not all, embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0082] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0083] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0085] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An intelligent ammonia injection denitrification prediction system based on a time-series neural network model, characterized in that, It includes the following steps: S1, Data Acquisition; S2. Data processing: Initially calculate the importance ranking of the data through simple multiple linear regression analysis; S3. Feature selection: Use the LASSO regression model to eliminate collinearity in the model and calculate the feature importance ranking of the data. S4. Model training and parameter optimization: using difference-based features and baseline-based models. In the non-blowback phase, the model uses a long short-term memory network as the main body, and in the blowback phase, a time-series Transformer model is used. S5. Use ablation experiments to determine the importance of features; S6. Conduct actual deployment during on-site forecasting; Step S4, model training and parameter optimization, specifically includes: Based on the difference-based feature and the baseline-based model, the algorithm takes the change of the input feature and first observes the collinearity of the features and the VIF of each feature after taking the difference. The algorithm outputs the difference between the measured value and the true value at this time according to the change of the input. The NOx concentration prediction model based on a time-series neural network uses a Long Short-Term Memory (LSTM) network as the main body of the model in the non-backflush phase and a time-series Transformer in the backflush phase. In the non-backflush phase, the model has a data preprocessing layer at the beginning, which includes data smoothing and data normalization, followed by an LSM unit containing 400 hidden units, and finally a fully connected layer to output a specific value. The learning rate is 0.0016, the batch size during prediction is 1024, the batch size during testing is 512, and the loss function used during training is MAE. Finally, the NOx concentration prediction value is obtained by adding the difference between the model's predictions and the measured value. Apart from the difference in input dimensions, the training process for the non-backflush stage model and the backflush stage model is the same. However, the backflush stage uses the Transformer model. Since there is no NOx concentration measurement value in the input, the NOx concentration prediction value is directly output after passing through the Transformer.

2. The intelligent ammonia injection denitrification prediction system based on a time-series neural network model according to claim 1, characterized in that: The dataset used for data collection in step S1 consists of the values ​​of power plant-related characteristics over 10 days, totaling 71 statistical variables.

3. The intelligent ammonia injection denitrification prediction system based on a time-series neural network model according to claim 1, characterized in that: The data processing in step S2 specifically includes multiple linear regression analysis and collinearity analysis between features; Multiple linear regression analysis: First, determine the true NOx concentration at this moment as the regression coefficient of the problem, and use the remaining feature values ​​as independent variables. Use the least squares method to perform multiple linear regression analysis and calculate the regression coefficients of each feature. Assuming NOx concentration is They satisfy a linear relationship: In a set of data In the case of minimizing the overall sum of squared errors, the objective is to find a set of... This resulted in losses. Minimum; The loss function for linear regression is （3-1）。 4. The intelligent ammonia injection denitrification prediction system based on a time-series neural network model according to claim 3, characterized in that: Multicollinearity analysis among features: Each explanatory variable in Equation 3-1 has a variance expansion factor, which reflects the extent to which multicollinearity increases the variance of the estimated coefficients. The variance expansion factor method is used to analyze the multicollinearity of variables.

5. The intelligent ammonia injection denitrification prediction system based on a time-series neural network model according to claim 4, characterized in that: The feature selection in step S3 specifically includes: LASSO ridge regression with L1 regularization has a loss function of: （5-1） Ridge regression with L2 regularization has the following loss function: （5-2）。 6. The intelligent ammonia injection denitrification prediction system based on a time-series neural network model according to claim 1, characterized in that: The ablation experiment in step S5 specifically includes: First, train the model on the entire dataset and use the error of the model as a baseline; Then, each time one feature is deleted, the difference in accuracy of the retrained model is observed. If the accuracy of a feature drops too much, it indicates that the feature is very important. Conversely, if the error does not decrease significantly, it indicates that the feature is unimportant or that the feature is collinear. Therefore, a certain number of additional process features need to be added on this basis.

7. The intelligent ammonia injection denitrification prediction system based on a time-series neural network model according to claim 1, characterized in that: The specific deployment method in step S6 is as follows: S6-1. During field operation, the DCS transmits data at intervals of 0.8s, but sometimes it exceeds 1s. If the interval exceeds 1s for a long time, frame loss will occur after a period of time. If the average transmission interval is (1+t) seconds, then frame loss will occur on average every 1 / t seconds. In this case, the value at a certain moment is missing. Therefore, the data before the algorithm input needs to be processed in a certain way. The processing method is as follows. S6-2. The algorithm stores all the received data in memory and checks the continuity of timestamps in real time. However, in order to ensure processing speed, the algorithm only maintains data for the past hour. S6-3. If the recent data is continuous, input the data directly into the model. If it is not continuous, use linear interpolation or supplement the data based on the previous frame to ensure that each timestamp in the input window has a corresponding value.

Images