NOx concentration prediction method, system and equipment based on denitration system and storage medium

Abstract

The invention relates to the technical field of flue gas denitration, and particularly provides a NOx concentration prediction method, system and device based on a denitration system and a storage medium, and the method comprises the steps: obtaining historical working condition data, carrying out the standardization processing of the historical working condition data, and obtaining standard time series data; vectorizing the standard time sequence data to obtain a vector data set, and segmenting the vector data set by using a sliding window with a preset length to obtain a model training set; based on the model training set, training a pre-constructed initial NOx concentration prediction model to obtain a trained enhanced NOx concentration prediction model; based on denitration system operation data collected in real time, standardization processing is carried out, the trained enhanced NOx concentration prediction model is input, and a NOx concentration prediction value under the target working condition is obtained. According to the invention, local features and global features extracted by different encoder layers are effectively fused.

Description

NOx concentration prediction methods, systems, equipment, and storage media based on denitrification systems Technical Field

[0001] This invention belongs to the field of flue gas denitrification technology, specifically relating to a method, system, equipment, and storage medium for predicting NOx concentration based on a denitrification system. Background Technology

[0002] In the denitrification system of thermal power units, accurately predicting the NOx concentration at the outlet of the SCR reactor is crucial for ensuring emission compliance, optimizing ammonia injection, and reducing operating costs. NOx formation is influenced by various operating parameters, including load, primary and secondary air ratio, combustion temperature, coal feed rate, and ammonia injection rate. These parameters exhibit strong nonlinear coupling and fluctuate continuously with changes in unit load, damper operation, and coal quality, resulting in NOx concentration sequences characterized by frequent local disturbances and significant long-term dependencies. Against this backdrop, with the continuous accumulation of unit operating data, using deep learning models to predict NOx concentrations has become an industry trend. Existing research often uses CNNs to extract short-term local variation features, LSTM and GRU models to handle temporal dependencies within a certain range, while Transformers, due to their self-attention mechanism, can capture dependencies over longer time spans and are gradually being introduced into industrial sequence prediction tasks.

[0003] However, in practical engineering applications, the above models still have significant shortcomings. First, the encoder structure of the standard Transformer is an independent stacking method, with no feature interaction between the encoder layers. That is, each layer only processes the attention results of its own layer and cannot fuse the local features captured by lower layers with the global features captured by higher layers. Therefore, when faced with multivariate coupling relationships such as "airflow changes affect coal feed, coal feed affects ammonia injection response, and ammonia injection changes determine NOx trends," traditional Transformers are prone to insufficient feature understanding and inadequate extraction of correlations between variables, resulting in low NOx prediction accuracy. Second, RNN / LSTM is prone to memory decay under long sequence conditions and cannot stably capture cross-variable dependencies over tens of seconds or even longer periods; while CNN alone can capture local perturbations, it cannot express the overall trend. Summary of the Invention

[0004] In view of the above-mentioned shortcomings of the prior art, the present invention provides a method, system, device and storage medium for predicting NOx concentration based on a denitrification system to solve the above-mentioned technical problems.

[0005] In a first aspect, the present invention provides a method for predicting NOx concentration based on a denitrification system, comprising: acquiring historical operating condition data; standardizing the historical operating condition data to obtain standard time series data; vectorizing the standard time series data to obtain a vector dataset; segmenting the vector dataset using a sliding window of a preset length to obtain a model training set; training a pre-constructed initial NOx concentration prediction model based on the model training set to obtain a trained enhanced NOx concentration prediction model, wherein the enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers for feature interaction and fusion between different encoder layers; and obtaining the predicted NOx concentration value under the target operating condition based on real-time collected denitrification system operating data, after standardization processing and input into the trained enhanced NOx concentration prediction model.

[0006] In one optional implementation, historical operating condition data is standardized to obtain standard time series data, including: sampling historical multivariate operating condition data of the denitrification system operation process at preset time intervals in an industrial time series database to obtain a historical raw data sequence composed of multiple measuring points, including primary air volume, secondary air volume, total air volume, damper opening, coal feeder coal quantity, ammonia injection valve opening, oxygen content, and NOx concentration; based on the historical raw data sequence, the quartiles of each measuring point are calculated.

[0007] in, It is the first quartile. The third quartile is given by IQR, which is the interquartile range. or If x is an outlier, then x is determined to be an outlier and is deleted, where x is a data point in the historical original data sequence; the historical original data sequence after deleting outliers is interpolated and normalized using the Lagrange interpolation method.

[0008] In one optional implementation, standard time series data is vectorized to obtain a vector dataset. The vector dataset is then segmented using a sliding window of a preset length to obtain a model training set. This includes: combining the standard time series data into a single vector based on the values ​​of each measurement point at the same time scale, resulting in a vector dataset; indexing the vector dataset according to time order to obtain an indexed vector dataset; sequentially extracting continuous vectors from the indexed vector dataset using a sliding window of a preset length, resulting in a set of sliding window segments, where each sliding window segment consists of continuous vectors corresponding to the window length; and using the first N continuous vectors from the sliding window segment set as input features and the (N+1)th vector from the sliding window segment set as the corresponding label to obtain the model training set.

[0009] In an optional implementation, the method further includes: a step of constructing an initial NOx concentration prediction model, comprising: constructing a feature extraction layer of a neural network for extracting local change features, wherein the feature extraction layer includes two one-dimensional convolutional layers, each layer having several convolutional kernels and employing the ReLU activation function, for extracting local features from standard time series data to obtain a feature map containing local perturbation information; based on the feature map containing local perturbation information, mapping the feature map containing local perturbation information to the feature dimension of a Transformer model through a linear mapping layer, and adding positional encoding based on sine and cosine functions to give the input sequence temporal positional information, resulting in a mapped feature sequence with added positional encoding; based on the mapped feature sequence with added positional encoding, constructing a Transformer sequence modeling network containing multiple encoder layers and multiple decoder layers, wherein each encoder layer includes a multi-head self-attention module and a forward fully connected network, wherein the multi-head self-attention module includes multiple attention heads for modeling the internal correlation of the mapped feature sequence with added positional encoding, resulting in the encoding result of the Transformer sequence modeling network; based on the Transformer... The encoding results of the sequence modeling network are processed by a cross-layer multi-head attention module between adjacent encoder layers. This allows the encoding results of the Transformer sequence modeling network to undergo cross-layer attention calculation, resulting in cross-layer feature interaction results. Based on these cross-layer feature interaction results, a deep neural network output layer for regression output is constructed. This output layer includes two hidden layers, each employing the ReLU activation function and incorporating a dropout structure to improve the model's generalization ability. The cross-layer feature interaction results are then input into a single-neuron output layer to generate NOx concentration prediction values, resulting in a deep regression output structure. Based on this deep regression output structure, it is combined with a convolutional neural network and an enhanced Transformer network to obtain an initial NOx concentration prediction model.

[0010] In one optional implementation, a pre-built initial NOx concentration prediction model is trained based on a model training set to obtain a trained enhanced NOx concentration prediction model. The enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers for feature interaction and fusion between different encoder layers. This includes: inputting the model training set into the pre-built initial NOx concentration prediction model, obtaining the forward propagation result of the initial NOx concentration prediction model through a forward propagation step; inputting the forward propagation result of the initial NOx concentration prediction model into a feature extraction layer containing two one-dimensional convolutional layers, obtaining a convolutional feature extraction result through a convolutional feature extraction step; mapping the convolutional feature extraction result to the feature dimension of a Transformer through a linear mapping layer and adding positional encoding, obtaining a mapping sequence with added positional encoding through a feature mapping step; and inputting the mapping sequence with added positional encoding into a Transformer containing multiple encoder layers. The sequence modeling network is processed through an encoding computation step to obtain the encoding result of the Transformer sequence modeling network. Based on the encoding result of the Transformer sequence modeling network, a cross-layer multi-head attention module is set between adjacent encoder layers, and a cross-layer feature interaction result is obtained through a cross-layer attention computation step. Based on the cross-layer feature interaction result, the cross-layer feature interaction result is input to the output layer of the deep neural network, and a NOx concentration prediction output is obtained through a regression output step. Based on the NOx concentration prediction output, the NOx concentration prediction output is compared with the true NOx concentration label, and a loss calculation result is obtained through a loss calculation step. Based on the loss calculation result, a backpropagation gradient is obtained through a backpropagation step. Based on the backpropagation gradient, an updated initial NOx concentration prediction model is obtained through a parameter update step. Based on the updated initial NOx concentration prediction model, an enhanced NOx concentration prediction model is obtained through an iterative training step.

[0011] Secondly, this invention provides a NOx concentration prediction system based on a denitrification system, comprising: a data standardization module for acquiring historical operating condition data and standardizing the historical operating condition data to obtain standard time series data; a model training set construction module for vectorizing the standard time series data to obtain a vector dataset, and dividing the vector dataset using a sliding window of a preset length to obtain a model training set; a model training module for training a pre-constructed initial NOx concentration prediction model based on the model training set to obtain a trained enhanced NOx concentration prediction model, wherein the enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers for feature interaction and fusion between different encoder layers; and a result output module for obtaining the predicted NOx concentration value under the target operating condition based on real-time collected denitrification system operating data, after standardization processing and input into the trained enhanced NOx concentration prediction model.

[0012] In one optional implementation, the model training set construction module includes: a vector transformation unit, used to combine standard time series data into the same vector according to the values ​​of each measurement point at the same time scale, based on standard time series data, to obtain a vector dataset; a vector indexing unit, used to index and arrange the vector dataset according to time order, based on the vector dataset, to obtain an indexed vector dataset; a vector truncation unit, used to truncate continuous vectors in the indexed vector dataset sequentially using a sliding window of a preset length, based on the indexed vector dataset, to obtain a set of sliding window segments, wherein each sliding window segment consists of continuous vectors corresponding to the window length; and a model training set construction unit, used to take the first N continuous vectors in the sliding window segment set as input features and the N+1th vector in the sliding window segment set as the corresponding label, to obtain the model training set.

[0013] In an optional implementation, the system further includes: an initial NOx concentration prediction model construction module comprising: a feature extraction layer construction unit, used to construct a feature extraction layer of a neural network for extracting local change features, wherein the feature extraction layer includes two one-dimensional convolutional layers, each layer having several convolutional kernels and employing the ReLU activation function, used to extract local features from standard time series data to obtain a feature map containing local perturbation information; a mapping feature sequence construction unit, used to map the feature map containing local perturbation information to the feature dimension of a Transformer model through a linear mapping layer, and add positional encoding based on sine and cosine functions to give the input sequence temporal positional information, resulting in a mapped feature sequence with added positional encoding; and an encoding result construction unit, used to construct a Transformer sequence modeling network containing multiple encoder layers and multiple decoder layers based on the mapped feature sequence with added positional encoding, wherein each encoder layer includes a multi-head self-attention module and a forward fully connected network, wherein the multi-head self-attention module includes multiple attention heads, used to model the internal correlation of the mapped feature sequence with added positional encoding, resulting in a Transformer model. The system comprises the following components: a sequence modeling network encoding result unit; a cross-layer feature interaction result construction unit, which, based on the encoding result of the Transformer sequence modeling network, sets up a cross-layer multi-head attention module between adjacent encoder layers, so that the encoding result of the Transformer sequence modeling network undergoes cross-layer attention calculation to obtain the cross-layer feature interaction result; a deep regression output unit, which, based on the cross-layer feature interaction result, constructs a deep neural network output layer for regression output, the output layer including two hidden layers, each using the ReLU activation function and incorporating a dropout structure to improve the model's generalization ability, and inputs the cross-layer feature interaction result into a single-neuron output layer to generate NOx concentration prediction values, thus obtaining the deep regression output structure; and an initial prediction model construction unit, which, based on the deep regression output structure, combines the deep regression output structure with a convolutional neural network and an enhanced Transformer network to obtain an initial NOx concentration prediction model.

[0014] Thirdly, an apparatus is provided, comprising: a memory for storing a NOx concentration prediction program based on a denitrification system; and a processor for executing the NOx concentration prediction program based on a denitrification system to implement the steps of the NOx concentration prediction method based on a denitrification system as provided in the first aspect.

[0015] Fourthly, a computer-readable storage medium is provided, on which a NOx concentration prediction program based on a denitrification system is stored, wherein when the NOx concentration prediction program based on a denitrification system is executed by a processor, the steps of the NOx concentration prediction method based on a denitrification system provided in the first aspect are implemented.

[0016] The beneficial effects of this invention are as follows: The NOx concentration prediction method, system, device, and storage medium based on denitrification systems provided by this invention improve data consistency and usability by standardizing historical operating data, reducing interference from outliers, sampling fluctuations, and missing measurement points on model training. By constructing training samples through vectorization and sliding windows, multi-dimensional operating parameters can be input into the model on a unified time scale, ensuring that the model can simultaneously learn local perturbations and overall trends, which is beneficial for improving the completeness of time-series modeling. By introducing an enhanced Transformer structure containing multiple encoder layers and setting cross-layer multi-head attention modules, local and global features extracted by different encoder layers are effectively fused, improving the problem of insufficient extraction of multivariate coupling relationships in traditional models, thereby improving the accuracy and stability of NOx concentration prediction under complex and variable operating conditions.

[0017] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 is a schematic flowchart of a method according to an embodiment of the present invention.

[0020] Figure 2 is a schematic block diagram of a system according to an embodiment of the present invention.

[0021] Figure 3 is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0024] The NOx concentration prediction method based on the denitrification system provided in this embodiment of the invention is executed by a computer device, and correspondingly, the NOx concentration prediction system based on the denitrification system runs in the computer device.

[0025] Figure 1 is a schematic flowchart of a method according to an embodiment of the present invention. The entity executing the process in Figure 1 can be a NOx concentration prediction system based on a denitrification system. Depending on different requirements, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0026] As shown in Figure 1, the method includes: S1, acquiring historical operating condition data, standardizing the historical operating condition data, and obtaining standard time series data.

[0027] S2. Vectorize the standard time series data to obtain a vector dataset. Use a sliding window of preset length to divide the vector dataset to obtain the model training set.

[0028] S3. Based on the model training set, train the pre-built initial NOx concentration prediction model to obtain the trained enhanced NOx concentration prediction model. The enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers to perform feature interaction and fusion between different encoder layers.

[0029] S4. Based on the real-time collected operating data of the denitrification system, the data is standardized and input into the trained enhanced NOx concentration prediction model to obtain the predicted NOx concentration under the target operating conditions.

[0030] In one embodiment of the present invention, based on step S1, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0031] Historical operating condition data is acquired and standardized to obtain standard time series data. This includes sampling historical multivariate operating condition data of the denitrification system at preset time intervals in an industrial time series database, resulting in a historical raw data sequence composed of multiple measuring points. These measuring points include primary air volume, secondary air volume, total air volume, damper opening, coal feeder coal quantity, ammonia injection valve opening, oxygen content, and NOx concentration. Based on the historical raw data sequence, the quartiles of each measuring point are calculated.

[0032] in, It is the first quartile. The third quartile is given, and the IQR is the interquartile range.

[0033] like or If x is an outlier, it is determined to be an outlier and is deleted, where x is a data point in the historical original data sequence. The historical original data sequence after outlier deletion is then interpolated using Lagrange interpolation and normalized.

[0034] Specifically, firstly, historical multi-factor operating data of the denitrification system during operation is retrieved from the industrial time-series database according to a sampling period consistent with the unit's operation (e.g., 5 seconds). This historical multi-factor operating data includes at least the following measurement points: primary air volume, secondary air volume, total air volume, damper opening, coal feeder coal quantity, ammonia injection valve opening, oxygen content, and NOx concentration. These measurement points are aligned along the time dimension and arranged chronologically to form a historical raw data sequence. Secondly, based on this historical raw data sequence, the first quartile Q1, the third quartile Q3, and the interquartile range IQR = Q3 are calculated for each measurement point using a box plot method. Q1, when a data point is less than Q1 When the value is 1.5 × IQR or greater than Q3 + 1.5 × IQR, the data point is identified as an outlier and deleted from the historical raw data sequence. Then, missing value processing is performed on the historical raw data sequence after outlier removal. For short-span gaps, Lagrange interpolation is used to fill them in chronological order. For long-term gaps caused by unit shutdowns or communication interruptions, they can be filled according to agreed-upon rules, ensuring that each measuring point forms a continuous time series across the entire time axis. Finally, normalization is performed on the data of each measuring point after interpolation and filling, preferably using maximum-minimum normalization. The original value x of each measuring point is normalized according to x' = (x... x_min) / (x_max The data is linearly scaled (x_min) and (x_max) are the minimum and maximum values ​​of the measurement point within the selected historical interval, respectively. After normalization, standard time series data with uniform dimensions, time alignment, and continuity are obtained, which serve as the input basis for subsequent vectorization and model training.

[0035] In one embodiment of the present invention, based on step S2, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0036] Standard time series data is vectorized to obtain a vector dataset. The vector dataset is then segmented using a sliding window of a preset length to obtain the model training set. This process includes: combining the standard time series data into vectors based on the values ​​of each measurement point at the same time scale; indexing the vector dataset according to time order to obtain an indexed vector dataset; sequentially extracting continuous vectors from the indexed vector dataset using a sliding window of a preset length to obtain a set of sliding window segments, where each sliding window segment consists of continuous vectors corresponding to the window length; and using the first N continuous vectors from the sliding window segment set as input features and the (N+1)th vector as the corresponding label to obtain the model training set.

[0037] Specifically, firstly, the model training set generated by the sliding window is divided into multiple training batches according to the batch size. The input feature sequence of each training batch is then fed into a pre-constructed initial NOx concentration prediction model for forward propagation. During forward propagation, the input sequence first passes through a feature extraction network consisting of two one-dimensional convolutional layers to extract the change features of each operating condition at the local time scale, obtaining a local feature map. Subsequently, the local feature map is mapped to the feature dimension of the Transformer model through a linear mapping layer, and a sine-cosine positional encoding is added to give the input sequence temporal positional information.

[0038] Next, the position-encoded sequence is fed into a Transformer sequence modeling network containing multiple encoder layers. Each encoder layer includes a multi-head self-attention module and a feedforward fully connected network. The multi-head self-attention module uses eight attention heads to model the correlations within the input sequence. To enhance the model's ability to extract complex nonlinear relationships between multivariate operating parameters, a cross-layer multi-head attention module is set between adjacent encoder layers. This allows the features extracted by the upper encoder layer to interact with the features of the lower encoder layer through a multi-head attention operation using the query matrix Q, key matrix K, and value matrix V. This enables the encoder layers to share features and perform deep fusion, resulting in a cross-layer enhanced global temporal feature representation.

[0039] Subsequently, the sequence features output from the final encoder layer are input into the decoder structure, and further processed through a multi-head attention mechanism and a feedforward fully connected network to obtain fused deep temporal features. These features are then fed into a deep neural network containing two hidden layers (with 256 and 128 hidden nodes respectively, and ReLU activation function and Dropout layer) to complete regression prediction, and output the model's predicted NOx concentration for the current batch of samples.

[0040] For each training batch, the Huber loss function is calculated by comparing the predicted values ​​with the actual NOx concentrations to obtain the error term. The gradient is then calculated using the backpropagation algorithm, and the model parameters are updated using the Adam optimizer with an initial learning rate of 0.001 and a learning rate decay strategy (decreasing by a factor of 0.9 every 10 epochs). An early stopping mechanism is employed during training: training is terminated when the validation set loss does not significantly decrease for 15 consecutive epochs to prevent overfitting.

[0041] Through the above iterative training process, the model can adaptively learn local perturbation features and long-range dependency information, and finally obtain a trained enhanced NOx concentration prediction model with convergent parameters and cross-layer multi-head attention enhancement capabilities. This model can more accurately extract the correlation between multivariate operating condition parameters and is suitable for high-precision NOx concentration prediction under complex and unsteady operating conditions.

[0042] In one possible implementation, the method further includes a step of constructing an initial NOx concentration prediction model, comprising: constructing a feature extraction layer of a neural network for extracting local change features, wherein the feature extraction layer includes two one-dimensional convolutional layers, each layer having several convolutional kernels and employing the ReLU activation function, for extracting local features from standard time series data to obtain a feature map containing local perturbation information; based on the feature map containing local perturbation information, mapping the feature map containing local perturbation information to the feature dimension of a Transformer model through a linear mapping layer, and adding positional encoding based on sine and cosine functions to give the input sequence temporal positional information, resulting in a mapped feature sequence with added positional encoding; based on the mapped feature sequence with added positional encoding, constructing a Transformer sequence modeling network containing multiple encoder layers and multiple decoder layers, wherein each encoder layer includes a multi-head self-attention module and a forward fully connected network, wherein the multi-head self-attention module includes multiple attention heads for modeling the internal correlation of the mapped feature sequence with added positional encoding, resulting in a Transformer model. The encoding results of the sequence modeling network are used as follows: Based on the encoding results of the Transformer sequence modeling network, a cross-layer multi-head attention module is set between adjacent encoder layers, so that the encoding results of the Transformer sequence modeling network are processed by cross-layer attention to obtain cross-layer feature interaction results; Based on the cross-layer feature interaction results, a deep neural network output layer for regression output is constructed. The output layer includes two hidden layers, which respectively adopt the ReLU activation function and add a dropout structure to improve the generalization ability of the model. The cross-layer feature interaction results are input to a single neuron output layer to generate NOx concentration prediction values, resulting in a deep regression output structure; Based on the deep regression output structure, the deep regression output structure is combined with a convolutional neural network and an enhanced Transformer network to obtain an initial NOx concentration prediction model.

[0043] Specifically, firstly, a convolutional neural network (CNN) feature extraction layer is constructed to extract local variation features. The feature extraction layer includes two one-dimensional convolutional layers. Each one-dimensional convolutional layer has 32 convolutional kernels. The kernel size and stride are set according to the length of the input sequence. ReLU is used as the activation function after the convolution operation to extract local perturbation features from the multi-dimensional working condition time series and obtain a one-dimensional feature map containing multi-dimensional local features.

[0044] Secondly, a Transformer sequence modeling network is constructed to model long-range dependencies. The one-dimensional feature map output by the CNN is mapped to the input feature dimension d_{model} of the Transformer through a linear mapping layer. In order to preserve temporal information, fixed position encoding based on sine and cosine functions is added to the mapped sequence.

[0045] Furthermore, a Transformer main structure comprising multiple encoder layers and multiple decoder layers is constructed. Each encoder layer includes a multi-head self-attention module and a forward fully connected network. The multi-head attention module contains eight attention heads, each of which performs self-attention calculations on the query matrix Q, key matrix K, and value matrix V of the input sequence. A cross-layer multi-head attention module is set between any two adjacent encoder layers, enabling cross-layer attention calculations between the output features of the preceding encoder layer and the input features of the following encoder layer, thereby achieving deep interaction between multi-layer features. Each decoder layer includes a multi-head attention module for receiving encoder output, a local self-attention module, and a forward fully connected network for further fusing global features from different encoder layers.

[0046] Subsequently, a deep neural network (DNN) output layer is constructed to perform the regression output. The output layer includes two hidden layers: the first hidden layer contains 256 neurons, and the second hidden layer contains 128 neurons. Each hidden layer uses the ReLU activation function, and a dropout layer is added after each hidden layer with dropout ratios of 0.4 and 0.3, respectively, to enhance the generalization ability of the model. Finally, the output layer is set to one neuron to output the predicted NOx concentration value.

[0047] The CNN–Transformer–DNN hybrid network obtained by combining the above structures constitutes the initial NOx concentration prediction model, which is used to receive the model training set and train the parameters in subsequent steps.

[0048] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0049] Based on the model training set, a pre-built initial NOx concentration prediction model is trained to obtain a trained enhanced NOx concentration prediction model. This enhanced model includes multiple encoder layers, with cross-layer multi-head attention modules between adjacent encoder layers for feature interaction and fusion between different encoder layers. The process includes: inputting the model training set into the pre-built initial NOx concentration prediction model, obtaining the forward propagation result of the initial NOx concentration prediction model through a forward propagation step; inputting the forward propagation result of the initial NOx concentration prediction model into a feature extraction layer containing two one-dimensional convolutional layers, obtaining the convolutional feature extraction result through a convolutional feature extraction step; mapping the convolutional feature extraction result to the feature dimension of a Transformer through a linear mapping layer and adding positional encoding, obtaining a mapping sequence with added positional encoding through a feature mapping step; and inputting the mapping sequence with added positional encoding into a Transformer sequence modeling network containing multiple encoder layers, obtaining the Transformer sequence modeling result through an encoding calculation step. The sequence modeling network's encoding results are used as follows: Based on the Transformer sequence modeling network's encoding results, a cross-layer multi-head attention module is set between adjacent encoder layers, and a cross-layer feature interaction result is obtained through a cross-layer attention calculation step; Based on the cross-layer feature interaction result, the cross-layer feature interaction result is input into the deep neural network's output layer, and a NOx concentration prediction output is obtained through a regression output step; Based on the NOx concentration prediction output, the NOx concentration prediction output is compared with the true NOx concentration label, and a loss calculation result is obtained through a loss calculation step; Based on the loss calculation result, a backpropagation gradient is obtained through a backpropagation step; Based on the backpropagation gradient, an updated initial NOx concentration prediction model is obtained through a parameter update step; Based on the updated initial NOx concentration prediction model, an enhanced NOx concentration prediction model is obtained through an iterative training step.

[0050] Specifically, based on the model training set, the pre-constructed initial NOx concentration prediction model is trained to obtain the trained enhanced NOx concentration prediction model. This includes: dividing the model training set constructed by the sliding window into a training set, a validation set, and a test set in a 7:2:1 ratio; during the training phase, dividing the input feature sequences in the training set into multiple batches according to the batch size (e.g., a batch size of 64); and feeding the input feature sequences of each batch into the initial NOx concentration prediction model for forward propagation; during the forward propagation process, the input feature sequences first pass through two layers of one-dimensional convolutional feature extraction layers, each layer having several convolutional kernels and using the ReLU activation function to extract local variation features of the multivariate operating condition time series, obtaining a feature map containing local perturbation information; subsequently, mapping the feature map to a preset Transformer feature dimension through a linear mapping layer, and superimposing a positional encoding constructed based on sine and cosine functions, so that the features at each time step carry temporal positional information, forming an input sequence for sequence modeling.

[0051] The input sequence with superimposed positional encoding is fed into a Transformer sequence modeling network containing multiple encoder layers for temporal dependency modeling. Each encoder layer includes a multi-head self-attention module and a forward fully connected network. The multi-head self-attention module uses multiple attention heads to model the correlation within the input sequence, and the forward fully connected network performs a nonlinear transformation on the attention-weighted features. A cross-layer multi-head attention module is set between any two adjacent encoder layers, using the features output by the previous encoder layer as the key matrix and value matrix, and the current features of the next encoder layer as the query matrix. Multi-head attention calculation is performed to obtain the cross-layer feature interaction result, which is then fused with the original features of the next encoder layer. This achieves deep feature interaction and fusion between different encoder layers, enhancing the model's ability to express complex relationships between multiple operating parameters. The sequence features processed by multiple encoder layers and cross-layer multi-head attention modules can be optionally further processed by a decoding structure or directly flattened and fed into the output layer of a deep neural network containing two hidden layers. The hidden layers use the ReLU activation function and add a dropout structure. Finally, the NOx concentration prediction value of the batch of samples is obtained through a single neuron output layer.

[0052] A loss function is calculated for each batch of predicted NOx concentrations and their corresponding real NOx concentration labels. The preferred loss function is the Huber loss function. Based on the loss function results, the gradients of the parameters of each network layer are calculated using the backpropagation algorithm. The Adam optimizer is used to update the weight and bias parameters of the convolutional layers, encoder layers, multi-head attention modules, and the output layer of the deep neural network with an initial learning rate (e.g., 0.001). During training, a learning rate decay strategy and an early stopping strategy are combined, and the loss change is monitored on the validation set. Training is terminated when the validation set loss no longer decreases significantly within several training epochs. Through the above iterative updates based on the model training set, the parameters of the initial NOx concentration prediction model converge, resulting in an enhanced NOx concentration prediction model after training, which includes multiple encoder layers and multi-head attention modules between adjacent encoder layers.

[0053] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0054] Specifically, based on real-time collected denitrification system operation data, after standardization processing, the data is input into the trained enhanced NOx concentration prediction model to obtain the predicted NOx concentration under the target operating conditions. This includes: during unit operation, the industrial time-series database acquires multiple operating condition parameters in real time, such as primary air volume, secondary air volume, total air volume, damper opening, coal feeder coal quantity, ammonia injection valve opening, oxygen content, furnace temperature, and SCR inlet flue gas temperature, according to a fixed sampling period (e.g., 5 seconds) consistent with the model training, forming a real-time operating condition data sequence; the real-time operating condition data sequence is preprocessed according to the data processing rules adopted in the training phase, including: using the Lagrange interpolation method to interpolate and complete short-time-space value points, and performing maximum-minimum normalization on the interpolated data according to the maximum and minimum values ​​of the corresponding measurement points in the training phase, so that the real-time operating condition features maintain the same scale and distribution as the training data, resulting in a real-time standardized feature vector.

[0055] The real-time standardized feature vectors are converted into a vector form consistent with the training phase according to the model input structure, and combined into the input sequence required for the current prediction time. This input sequence is then fed into the trained enhanced NOx concentration prediction model for inference. During inference, the input sequence first passes through a one-dimensional convolutional feature extraction layer to extract local perturbation features in the real-time operating conditions, resulting in a real-time convolutional feature map. The real-time convolutional feature map is mapped to the input dimension of the Transformer model, and a positional encoding composed of sine and cosine is added, enabling the model to recognize the temporal location information of the real-time data. Subsequently, the sequence with the positional encoding is input into an enhanced Transformer network containing multiple encoder layers. The cross-layer multi-head attention module between adjacent encoder layers performs cross-layer attention calculation on the feature representations of different encoder layers to capture the deep correlation between operating parameters and obtain real-time global temporal features.

[0056] Global time-series features are input into the regression layer of a deep neural network, and nonlinear feature transformation is performed sequentially through two hidden layers (using ReLU activation function and dropout). Finally, the predicted NOx concentration value at the current moment is generated through a single neuron output layer. This predicted NOx concentration value is used as the NOx concentration prediction result under the target operating condition, and can be used for ammonia injection feedforward control, emission early warning, and air distribution strategy optimization.

[0057] In some embodiments, the NOx concentration prediction system based on the denitrification system may include multiple functional modules composed of computer program segments. The computer programs of each program segment in the NOx concentration prediction system based on the denitrification system may be stored in the memory of a computer device and executed by at least one processor to perform the function of predicting NOx concentration based on the denitrification system (see Figure 1 for details).

[0058] In this embodiment, the NOx concentration prediction system based on the denitrification system can be divided into multiple functional modules according to its functions, as shown in Figure 2. A module, as referred to in this invention, is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0059] The data standardization module acquires historical operating condition data and standardizes it to obtain standard time series data. The model training set construction module vectorizes the standard time series data to obtain a vector dataset, and then segments the vector dataset using a sliding window of preset length to obtain the model training set. The model training module trains the pre-built initial NOx concentration prediction model based on the model training set to obtain the trained enhanced NOx concentration prediction model. The enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers for feature interaction and fusion between different encoder layers. The result output module standardizes the real-time collected denitrification system operating data and inputs it into the trained enhanced NOx concentration prediction model to obtain the predicted NOx concentration value under the target operating condition.

[0060] Figure 3 illustrates that the NOx concentration prediction method based on a denitrification system provided in this embodiment can be applied to equipment. Those skilled in the art will understand that the equipment structure involved in the embodiments of this invention does not constitute a limitation on the equipment. The equipment may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. In the embodiments of this invention, the equipment includes, but is not limited to, laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The equipment may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of this application described and / or claimed herein.

[0061] The device 300 may include a processor 310, a memory 320, and a communication unit 330. These components communicate via one or more buses. Those skilled in the art will understand that the server structure shown in the figure does not constitute a limitation of the present invention. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0062] The memory 320 can be used to store execution instructions of the processor 310. The memory 320 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. When the execution instructions in the memory 320 are executed by the processor 310, the device 300 is able to perform some or all of the steps in the above method embodiments.

[0063] The processor 310 serves as the control center of the storage device, connecting various parts of the electronic device via various interfaces and lines. It executes software programs and / or modules stored in the memory 320, and calls data stored in the memory to perform various functions of the electronic device and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 310 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.

[0064] The communication unit 330 is used to establish a communication channel, enabling the storage device to communicate with other devices. It can receive user data sent by other devices or send user data to other devices.

[0065] The present invention also provides a computer storage medium, wherein the computer storage medium may store a program, which, when executed, may include some or all of the steps provided in the embodiments of the present invention. The storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0066] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other medium capable of storing program code. It includes several instructions to cause a computer device (which may be a personal computer, a server, or a second device, network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present invention.

[0067] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

[0068] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or modules may be electrical, mechanical, or other forms.

[0069] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0070] In addition, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0071] Although the present invention has been described in detail with reference to the accompanying drawings and preferred embodiments, the present invention is not limited thereto. Various equivalent modifications or substitutions can be made to the embodiments of the present invention by those skilled in the art without departing from the spirit and essence of the invention, and such modifications or substitutions should all be within the scope of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should also be covered within the protection scope of the present invention.

Claims

1. A method for predicting NOx concentration based on a denitrification system, characterized in that, include: Acquire historical operating condition data, standardize the historical operating condition data, and obtain standard time series data; Standard time series data is vectorized to obtain a vector dataset. The vector dataset is then segmented using a sliding window of a preset length to obtain a model training set. Based on the model training set, a pre-built initial NOx concentration prediction model is trained to obtain a trained enhanced NOx concentration prediction model. The enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers for feature interaction and fusion between different encoder layers. Based on real-time collected denitrification system operation data, after standardization processing, the data is input into the trained enhanced NOx concentration prediction model to obtain the predicted NOx concentration value under the target operating condition.

2. The method according to claim 1, characterized in that, Historical operating condition data is acquired and standardized to obtain standard time series data. This includes sampling historical multivariate operating condition data of the denitrification system at preset time intervals in an industrial time series database, resulting in a historical raw data sequence composed of multiple measuring points. These measuring points include primary air volume, secondary air volume, total air volume, damper opening, coal feeder coal quantity, ammonia injection valve opening, oxygen content, and NOx concentration. Based on the historical raw data sequence, the quartiles of each measuring point are calculated. in, It is the first quartile. The third quartile is given by IQR, which is the interquartile range. or If x is an outlier, then x is determined to be an outlier and is deleted, where x is a data point in the historical original data sequence; the historical original data sequence after deleting outliers is interpolated and normalized using the Lagrange interpolation method.

3. The method according to claim 1, characterized in that, Standard time series data is vectorized to obtain a vector dataset. The vector dataset is then segmented using a sliding window of a preset length to obtain the model training set. This process includes: combining the standard time series data into vectors based on the values ​​of each measurement point at the same time scale; indexing the vector dataset according to time order to obtain an indexed vector dataset; sequentially extracting continuous vectors from the indexed vector dataset using a sliding window of a preset length to obtain a set of sliding window segments, where each sliding window segment consists of continuous vectors corresponding to the window length; and using the first N continuous vectors from the sliding window segment set as input features and the (N+1)th vector as the corresponding label to obtain the model training set.

4. The method according to claim 1, characterized in that, Also includes: The steps for constructing an initial NOx concentration prediction model include: constructing a feature extraction layer of a neural network for extracting local change features. This feature extraction layer comprises two one-dimensional convolutional layers, each with several convolutional kernels and a ReLU activation function, used to extract local features from standard time-series data, obtaining a feature map containing local perturbation information; based on this feature map, mapping it to the feature dimension of a Transformer model through a linear mapping layer, and adding positional encoding based on sine and cosine functions to give the input sequence temporal positional information, resulting in a mapped feature sequence with added positional encoding; based on this mapped feature sequence with added positional encoding, constructing a Transformer sequence modeling network containing multiple encoder layers and multiple decoder layers. Each encoder layer includes a multi-head self-attention module and a forward fully connected network. The multi-head self-attention module includes multiple attention heads used to model the internal correlation of the mapped feature sequence with added positional encoding, obtaining the encoding result of the Transformer sequence modeling network; based on the encoding result of the Transformer sequence modeling network, setting cross-layer multi-head attention modules between adjacent encoder layers, enabling the Transformer... The encoding results of the sequence modeling network are processed through cross-layer attention calculation to obtain cross-layer feature interaction results. Based on the cross-layer feature interaction results, a deep neural network output layer for regression output is constructed. The output layer includes two hidden layers, which respectively adopt the ReLU activation function and add a dropout structure to improve the generalization ability of the model. The cross-layer feature interaction results are input into the single neuron output layer to generate NOx concentration prediction values, resulting in a deep regression output structure. Based on the deep regression output structure, the deep regression output structure is combined with a convolutional neural network and an enhanced Transformer network to obtain an initial NOx concentration prediction model.

5. The method according to claim 1, characterized in that, Based on the model training set, a pre-built initial NOx concentration prediction model is trained to obtain a trained enhanced NOx concentration prediction model. This enhanced model includes multiple encoder layers, with cross-layer multi-head attention modules between adjacent encoder layers for feature interaction and fusion between different encoder layers. The process includes: inputting the model training set into the pre-built initial NOx concentration prediction model, obtaining the forward propagation result of the initial NOx concentration prediction model through a forward propagation step; inputting the forward propagation result of the initial NOx concentration prediction model into a feature extraction layer containing two one-dimensional convolutional layers, obtaining the convolutional feature extraction result through a convolutional feature extraction step; mapping the convolutional feature extraction result to the feature dimension of a Transformer through a linear mapping layer and adding positional encoding, obtaining a mapped sequence with added positional encoding through a feature mapping step; and inputting the mapped sequence with added positional encoding into a Transformer sequence modeling network containing multiple encoder layers, obtaining the Transformer sequence modeling result through an encoding calculation step. The sequence modeling network's encoding results are used as follows: Based on the Transformer sequence modeling network's encoding results, a cross-layer multi-head attention module is set between adjacent encoder layers, and a cross-layer feature interaction result is obtained through a cross-layer attention calculation step; Based on the cross-layer feature interaction result, the cross-layer feature interaction result is input into the deep neural network's output layer, and a NOx concentration prediction output is obtained through a regression output step; Based on the NOx concentration prediction output, the NOx concentration prediction output is compared with the true NOx concentration label, and a loss calculation result is obtained through a loss calculation step; Based on the loss calculation result, a backpropagation gradient is obtained through a backpropagation step; Based on the backpropagation gradient, an updated initial NOx concentration prediction model is obtained through a parameter update step; Based on the updated initial NOx concentration prediction model, an enhanced NOx concentration prediction model is obtained through an iterative training step.

6. A NOx concentration prediction system based on a denitrification system, characterized in that, include: The data standardization module is used to acquire historical operating condition data, standardize the historical operating condition data, and obtain standard time series data. The model training set construction module is used to vectorize standard time series data to obtain a vector dataset, and then use a sliding window of preset length to split the vector dataset to obtain the model training set. The model training module is used to train the pre-built initial NOx concentration prediction model based on the model training set to obtain the trained enhanced NOx concentration prediction model. The enhanced NOx concentration prediction model includes multiple encoder layers, and a cross-layer multi-head attention module is set between adjacent encoder layers to perform feature interaction and fusion between different encoder layers. The results output module is used to obtain the predicted NOx concentration under the target operating conditions by taking real-time collected denitrification system operation data, standardizing it, and inputting it into the trained enhanced NOx concentration prediction model.

7. The system according to claim 1, characterized in that, The model training set construction module includes: a vector transformation unit, used to combine standard time series data into a single vector based on the values ​​of each measurement point at the same time scale, to obtain a vector dataset; a vector indexing unit, used to index and arrange the vector dataset in chronological order, to obtain an indexed vector dataset; a vector truncation unit, used to truncate continuous vectors in the indexed vector dataset sequentially using a sliding window of a preset length, to obtain a set of sliding window segments, where each sliding window segment consists of continuous vectors corresponding to the window length; and a model training set construction unit, used to take the first N continuous vectors in the sliding window segment set as input features and the (N+1)th vector in the sliding window segment set as the corresponding label, to obtain the model training set.

8. The system according to claim 1, characterized in that, Also includes: The initial NOx concentration prediction model construction module includes: a feature extraction layer construction unit, used to construct a feature extraction layer of a neural network for extracting local change features. The feature extraction layer includes two one-dimensional convolutional layers, each with several convolutional kernels and a ReLU activation function, used to extract local features from standard time series data, obtaining a feature map containing local perturbation information; a mapping feature sequence construction unit, used to map the feature map containing local perturbation information to the feature dimension of a Transformer model through a linear mapping layer, and add positional encoding based on sine and cosine functions to give the input sequence temporal positional information, obtaining a mapped feature sequence with added positional encoding; an encoding result construction unit, based on the mapped feature sequence with added positional encoding, constructing a Transformer sequence modeling network containing multiple encoder layers and multiple decoder layers. Each encoder layer includes a multi-head self-attention module and a forward fully connected network. The multi-head self-attention module includes multiple attention heads, used to model the internal correlation of the mapped feature sequence with added positional encoding, obtaining the encoding result of the Transformer sequence modeling network; and a cross-layer feature interaction result construction unit, used to construct a feature sequence modeling network based on the Transformer model. The sequence modeling network's encoding results are processed by a cross-layer multi-head attention module between adjacent encoder layers. This allows the Transformer sequence modeling network's encoding results to undergo cross-layer attention calculations, resulting in cross-layer feature interaction results. A deep regression output unit is used to construct a deep neural network output layer for regression output based on the cross-layer feature interaction results. This output layer includes two hidden layers, each employing the ReLU activation function and incorporating dropout to improve model generalization. The cross-layer feature interaction results are then input into a single-neuron output layer to generate NOx concentration prediction values, resulting in a deep regression output structure. An initial prediction model construction unit is used to combine the deep regression output structure with a convolutional neural network and an enhanced Transformer network to obtain an initial NOx concentration prediction model.

9. A device, characterized in that, include: A memory for storing NOx concentration prediction programs based on the denitrification system; A processor for implementing the steps of the NOx concentration prediction method based on any one of claims 1-5 when executing a NOx concentration prediction program based on a denitrification system.

10. A computer-readable storage medium storing a computer program, characterized in that, The readable storage medium stores a NOx concentration prediction program based on a denitrification system, which, when executed by a processor, implements the steps of the NOx concentration prediction method based on a denitrification system as claimed in any one of claims 1-5.

Images