A NOx prediction method based on improved ASG filtering
By improving the ASG filtering, TCN and PASSA mechanisms, and combining them with the LSTM model, the problem of weak noise resistance and easy interference in feature extraction of NOx emission data under varying operating conditions of thermal power boilers was solved, and high-precision NOx prediction was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF TECH
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-16
Smart Images
Figure CN122224320A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of pollutant emission monitoring and artificial intelligence technology in thermal power plants, and in particular to a NOx prediction method based on an improved ASG filter. Background Technology
[0002] Nitrogen oxides are abbreviated as NOx, selective catalytic reduction (SCR) as SCR, and long short-term memory (LSTM) networks as LSTM. The energy structure is rapidly transitioning towards cleaner and lower-carbon energy, but coal-fired power units will continue to play a crucial role in stabilizing the power system for a considerable period. With the large-scale grid connection of intermittent renewable energy sources such as wind and solar power, thermal power boilers are forced to shift from traditional base-load operation to deep peak shaving and high-frequency regulation to mitigate grid fluctuations. This results in boilers operating under conditions of significant load changes and unsteady states for extended periods. These frequent changes in operating conditions exacerbate the instability of the combustion process within the furnace, causing the formation of major pollutants such as NOx to exhibit strong nonlinear, time-varying, and multivariate coupling characteristics. Currently, power plant boilers widely employ SCR technology for flue gas denitrification. However, due to the significant inertia and delay characteristics of SCR systems, relying solely on the hysteresis feedback control of hardware sensors is insufficient to meet the ammonia injection requirements under drastic changes in operating conditions, easily leading to excessive ammonia escape or instantaneous emissions exceeding standards. In these deep peak-shaving scenarios, if a high-precision NOx emission prediction model that adapts to varying operating conditions can be established, and refined feedforward control of coal-fired power units can be achieved, the response speed to pollutant generation will be greatly improved, thereby effectively avoiding environmental violations and reducing operating costs.
[0003] Although data-driven and deep learning prediction algorithms, represented by LSTM, have significantly improved the handling of nonlinear time-series data compared to traditional mechanistic models or shallow machine learning methods, existing high-performance prediction models generally suffer from feature extraction bias and weak noise resistance when dealing with industrial data accompanied by high-frequency noise under varying operating conditions. While there are attempts to introduce conventional data smoothing or standard attention mechanisms in existing technologies, these often result in the decay or loss of key dynamic features during denoising. On the one hand, using preprocessing filters with fixed parameters, such as moving averages, wavelet transforms, or traditional SG filtering, is difficult to effectively balance smoothness and feature fidelity under varying operating conditions. This can easily lead to high-frequency noise being mistaken for effective features or causing over-smoothing distortion of local abrupt changes in the actual operating conditions. On the other hand, existing attention mechanisms, when dealing with complex multivariate coupled sequences, typically assign weights to the entire time series, failing to effectively shield against a large amount of redundant information and irrelevant variable interference during the varying operating conditions. This results in a significant decrease in the model's prediction accuracy and robustness when facing strong fluctuations and high-noise environments, making it difficult to achieve an ideal balance between strong denoising and high-fidelity feature extraction. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a NOx prediction method based on improved ASG filtering. This method addresses the common problems of existing deep learning prediction models in the face of strong non-stationarity of NOx emission data and high-frequency noise interference under frequent load changes and deep peak shaving conditions of thermal power boilers. These problems include weak noise resistance, easy interference of irrelevant variables in feature extraction, and difficulty in balancing noise reduction and preservation of local abrupt features by conventional data smoothing methods.
[0005] To achieve the above objectives, the present invention provides the following solution: A NOx prediction method based on improved ASG filtering includes: Collect operating parameters of the target thermal power boiler unit; The operating parameters of the aforementioned process are input into a pre-trained hybrid prediction model for prediction to obtain NOx prediction data; the pre-training process of the hybrid prediction model includes: Construct an initial prediction network; the initial prediction network includes: an input layer, an ASG module, a TCN module, a PASSA module, and an LSTM module connected in sequence; The training set was obtained by collecting multidimensional historical operating parameters and original sequences of nitrogen oxide emission concentrations of the target thermal power boiler unit during variable load operation. The ASG module is used to extract the volatility of the training set, the sliding window size of the SG filter is dynamically adjusted according to the volatility, and the SG filter is used to clean and denoise the training set to obtain stationary multivariate time series data. The stationary multivariate time series data is mapped and feature pattern extracted using the hierarchical dilated causal convolutional structure with residual connections in the TCN module at multiple time scales to obtain high-dimensional local feature patterns. The high-dimensional local feature pattern is input into the PASSA module. In the underlying tensor operation, redundant temporal interactions that do not conform to physical causality in the high-dimensional local feature pattern are eliminated and features are screened twice according to the physical mechanism delay mask matrix to obtain the preferred feature sequence. The delay and long memory characteristics of the preferred feature sequence are modeled using the gating mechanism in the LSTM module to obtain the predicted NOx emission concentration at the target time. The error between the predicted NOx emission concentration and the original NOx emission concentration sequence is extracted to obtain the output loss. Based on the output loss, the backpropagation algorithm is used to iteratively optimize the initial prediction network until the initial prediction network converges and the model training stops, thus obtaining the hybrid prediction model.
[0006] Preferably, the ASG module is used to extract the volatility of the training set, the sliding window size of the SG filter is dynamically adjusted according to the volatility, and the SG filter is used to perform cleaning and noise reduction preprocessing on the training set to obtain stationary multivariate time series data, including: Set the statistical window length, and calculate the variance of the training set according to the statistical window length to obtain local statistical features; Robust normalization coefficients are set based on the first and third quartiles of the local statistical features; the expression for the robust normalization coefficients is: ;in, The robust normalization coefficient is mentioned above; These are the local statistical features; This is the first quartile; This is the third quartile; This indicates that the minimum value is extracted. This indicates that the maximum value should be extracted. The sliding window size is obtained by performing weighted interpolation calculations on the robust normalization coefficients and the preset minimum and maximum filtering windows; the expression for the sliding window size is: ;in, The size of the sliding window; The maximum filtering window; This refers to the minimum filtering window; Based on the sliding window size, the sequences in the training set are smoothly reconstructed using the polynomial least squares method to obtain the stationary multivariate time series data.
[0007] Preferably, the stationary multivariate time-series data is mapped and feature pattern extracted using the hierarchical dilated causal convolutional structure with residual connections in the TCN module at multiple time scales to obtain high-dimensional local feature patterns, including: The stationary multivariate time series data is mapped to a high-dimensional hidden layer space of a preset dimension using a convolutional layer with a kernel size of 1. The mapped stationary multivariate time-series data is convolved using a hierarchical dilated causal convolutional network composed of multiple stacked residual blocks to obtain a dilated convolution output. The expression for a single-layer dilated convolution in the hierarchical dilated causal convolutional network is as follows: ;in, The output of the dilated convolution at the current time step; The kernel size; These are the kernel weights; For input features; For convolution kernel index; By using multiple residual blocks with different dilation factors, deep feature extraction and layer-by-layer fusion are performed on the output of each layer of dilated convolution to obtain the high-dimensional local feature pattern.
[0008] Preferably, the high-dimensional local feature pattern is input into the PASSA module. In the underlying tensor operations, redundant temporal interactions that do not conform to physical causality in the high-dimensional local feature pattern are eliminated and secondary feature filtering is performed based on the physical mechanism delay mask matrix, resulting in a preferred feature sequence, including: Linear projection is performed on the high-dimensional local feature pattern to obtain the query matrix, key matrix, and value matrix; The physical mechanism delay mask matrix is set according to the lower limit of the physical effective delay time and the upper limit of the decay time of the combustion system of the thermal power boiler. Based on the physical mechanism, the delay mask matrix is used to generate sparse attention score matrices and dense attention score matrices using an attention mechanism. The sparse attention score matrix and the dense attention score matrix are weighted and fused, and then multiplied with the value matrix to obtain the fused attention output matrix; the expression of the fused attention output matrix is: ;in, ; ; The fusion attention output matrix; , These are the first adaptive weighting weight and the second adaptive weighting weight, respectively. , These are the sparse attention score matrix and the dense attention score matrix, respectively; The value matrix; The first learning parameter; This is the second learning parameter; The fused attention output matrix is reorganized using a linear projection layer and a Dropout regularization layer in sequence to obtain the preferred feature sequence.
[0009] Preferably, the delayed and long-memory characteristics of the preferred feature sequence are modeled using the gating mechanism in the LSTM module to obtain the predicted NOx emission concentration at the target time, including: The preferred feature sequence is sequentially input into the long short-term memory network unit of the LSTM module for temporal nonlinear state update to obtain the final hidden state; The final hidden state is linearly mapped using a fully connected layer to obtain the predicted NOx emission concentration.
[0010] Preferably, the inflation factor of the hierarchical dilated causal convolutional network satisfies the following condition: ;in, This represents the number of network layers.
[0011] Preferably, the sparse attention score matrix and the dense attention score matrix are generated using an attention mechanism based on the physical mechanism delay mask matrix, including: In the sparse self-attention branch, a sparse attention score matrix is generated using a squared-corrected linear unit based on the physical mechanism delay mask matrix; the expression for the sparse attention score matrix is: ;in, The query matrix; The key matrix; This is a scaling dimension constant; The learnable bias matrix; This indicates that the maximum value should be extracted. In the dense self-attention branch, a dense attention score matrix is generated using a normalized exponential function based on the physical mechanism delay mask matrix; the expression for the dense attention score matrix is: ;in, It is a normalized exponential function.
[0012] Preferably, the physical mechanism delay mask matrix satisfies the following condition: when hour, ,otherwise, ;in, Indicates the current query time step; Indicates the time step of the history key; This is the lower limit of the effective physical delay time; This is the upper limit of the decay time; The physical mechanism delay mask matrix is in the first... Line number The elements of the column.
[0013] The present invention discloses the following technical effects: This invention provides a NOx prediction method based on an improved ASG filter. By flexibly adjusting the filter window size based on the perception of data fluctuations, it solves the problem that existing preprocessing filters with fixed parameters are unable to effectively balance smoothness and feature fidelity under varying operating conditions, which can easily lead to high-frequency noise being mistaken for effective features or causing oversmoothing distortion of local abrupt changes in the real operating conditions. It effectively eliminates high-frequency random noise while preserving the real information of abrupt changes in the operating conditions to the greatest extent. By introducing a physically prior-guided adaptive sparse self-attention mechanism (PASSA), it solves the problem that existing attention mechanisms cannot effectively shield a large amount of redundant information and irrelevant variable interference in the process of changing operating conditions when dealing with complex multivariate coupled sequences, which leads to a significant decrease in the prediction accuracy and robustness of the model when facing strong fluctuations and high noise environments. It realizes the construction of a physical mask matrix that reflects the real thermodynamic transport and combustion delay laws and the forced cut-off of interactions that do not conform to physical causal relationships. Attached Figure Description
[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 A schematic diagram of the NOx prediction process that integrates adaptive filtering and attention mechanisms provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the network hierarchy of the hybrid prediction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the feature extraction structure of the residual module of the Temporal Convolutional Network (TCN) provided in an embodiment of the present invention; Figure 4 A schematic diagram of the physical prior-guided adaptive sparse self-attention mechanism (PASSA) structure provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the temporal modeling structure of a Long Short-Term Memory (LSTM) network unit provided in an embodiment of the present invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] The purpose of this invention is to provide a NOx prediction method based on improved ASG filtering, which solves the problems of existing deep learning prediction models in the face of strong non-stationarity of NOx emission data and high-frequency noise interference under the conditions of frequent load changes and deep peak shaving of thermal power boilers. These problems include weak noise resistance, easy interference of irrelevant variables in feature extraction, and difficulty in balancing noise reduction and preservation of local abrupt features by conventional data smoothing methods.
[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0019] Figure 1 This is a schematic diagram of the NOx prediction process that integrates adaptive filtering and attention mechanisms provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides a NOx prediction method based on improved ASG filtering, comprising: Step 100: Collect operating parameters of the target thermal power boiler unit; Step 200: Input the operating condition parameters of the process into the pre-trained hybrid prediction model for prediction to obtain NOx prediction data; the pre-training process of the hybrid prediction model includes: Step 201: Construct the initial prediction network; the initial prediction network includes: an input layer, an ASG module, a TCN module, a PASSA module, and an LSTM module connected in sequence; Step 202: Collect the multi-dimensional historical operating parameters and the original sequence of nitrogen oxide emission concentrations of the target thermal power boiler unit during variable load operation to obtain the training set; Step 203: Use the ASG module to extract the volatility of the training set, dynamically adjust the sliding window size of the SG filter according to the volatility, and use the SG filter to perform cleaning and noise reduction preprocessing on the training set to obtain stationary multivariate time series data. Step 204: At multiple time scales, the stationary multivariate time series data is mapped and feature pattern extracted using the hierarchical dilated causal convolutional structure with residual connections in the TCN module to obtain high-dimensional local feature patterns. Step 205: Input the high-dimensional local feature pattern into the PASSA module. In the underlying tensor operation, based on the physical mechanism delay mask matrix, remove redundant temporal interactions in the high-dimensional local feature pattern that do not conform to physical causality and perform secondary feature filtering to obtain the preferred feature sequence. Step 206: Utilize the gating mechanism in the LSTM module to perform deep time-series modeling on the delay and long memory characteristics of the preferred feature sequence to obtain the predicted NOx emission concentration at the target time. Step 207: Extract the error between the predicted NOx emission concentration and the original nitrogen oxide emission concentration sequence to obtain the output loss. Based on the output loss, use the backpropagation algorithm to iteratively optimize the initial prediction network until the initial prediction network converges and the model training stops, thus obtaining the hybrid prediction model.
[0020] Preferably, this embodiment proposes a NOx prediction method based on improved ASG filtering. This method uses a deep hybrid neural network as its basic architecture. First, it employs an adaptive SG filtering (ASG) strategy based on local fluctuation statistical feature evaluation for preprocessing, fundamentally solving the problems of feature over-smoothing and distortion by dynamically adjusting the filter window size. Second, it utilizes a temporal convolutional network (TCN) to extract multi-scale local patterns from time-series data and innovatively constructs a physically prior-guided adaptive sparse self-attention mechanism (PASSA). This mechanism introduces a physical mask matrix reflecting thermodynamic delay laws, cutting off non-causal temporal interactions at the tensor operation level, and combines adaptive weighted fusion of sparse and dense branches to effectively filter out redundant interaction information and enhance the transmission of key dynamic features. Finally, it combines a long short-term memory network (LSTM) for deep modeling of long sequences. The main objective of this embodiment is to significantly improve the model's resistance to high-frequency noise interference while solving the feature extraction bias problem caused by strong coupling of multiple variables under varying operating conditions, achieving efficient and accurate prediction of NOx emission concentrations from thermal power boilers in complex industrial scenarios.
[0021] Specifically, a NOx prediction method based on improved ASG filtering includes: 1) Acquisition of multidimensional operational data and construction of prediction datasets: We obtained multidimensional historical operating parameters of thermal power boiler units under complex operating conditions such as variable load and the corresponding original sequences of nitrogen oxide (NOx) emission concentrations. We constructed these parameters into a dataset suitable for time series prediction models and divided them into training and testing sets to provide a data foundation for subsequent model training and performance verification.
[0022] 2) Adaptive filtering and data noise reduction: The original sequence was cleaned and preprocessed using an adaptive SG filtering (ASG) algorithm based on the evaluation of local fluctuation statistical features. By extracting local statistical features to perceive the intensity of data fluctuations, the sliding window size of the filter was dynamically adjusted accordingly. This effectively suppressed high-frequency random noise while preserving the true characteristics of abrupt changes in operating conditions to the greatest extent, resulting in stable multivariate time-series data.
[0023] 3) Multi-scale local feature extraction: The preprocessed stationary multivariate time-series data is input into the Temporal Convolutional Network (TCN) module. Utilizing the hierarchical dilated causal convolutional structure with residual connections in this module, the model's historical receptive field is expanded, mapping and extracting high-dimensional local feature patterns from the input data at different time scales.
[0024] 4) Physical prior-guided adaptive attention feature optimization: The extracted high-dimensional local feature patterns are fed into a physics-prior-guided adaptive sparse self-attention (PASSA) module for secondary feature filtering. By combining the upper and lower limits of the effective delay of thermal power boiler combustion to construct a physical mechanism delay mask matrix, and incorporating a parallel attention architecture with sparse and dense branches at the bottom layer, redundant background noise interactions are dynamically filtered out while preserving global dependencies, generating a preferred feature sequence that enhances key dynamic information.
[0025] 5) Long-term memory modeling and emissions prediction: The preferred feature sequence is input into a Long Short-Term Memory (LSTM) network module at time steps. Utilizing its internal gating mechanism, deep time-series modeling is performed on the large delay and long memory characteristics present in the combustion process of thermal power boilers, ultimately outputting the predicted value of nitrogen oxide emission concentration at the target time.
[0026] 6) Model optimization and iteration: The error loss between the predicted value and the actual emission sequence is calculated, and the network parameters of the hybrid prediction model are iteratively optimized using the backpropagation algorithm until the model converges, resulting in a lightweight and high-precision NOx prediction model that has been trained.
[0027] refer to Figure 2 This is a schematic diagram of the network hierarchy of the hybrid prediction model provided in this embodiment. In this network architecture, data flows sequentially through the input layer, the adaptive SG filter layer, the temporal convolutional module (TCN layer) containing multiple cascaded layers, the physically prior-guided adaptive sparse self-attention layer (PASSA attention layer), and the LSTM temporal module containing multiple stacked layers. Based on this network architecture, the specific implementation steps of the NOx prediction method based on improved ASG filtering provided in this embodiment are as follows: Optionally, obtain runtime data and build a time-series dataset: The original sequence of multidimensional historical operating parameters and nitrogen oxide (NOx) emission concentrations of thermal power boiler units during variable load operation was obtained, and a dataset was constructed and divided into training and testing sets.
[0028] In one embodiment, for complex operating conditions such as frequent peak shaving and frequency regulation of coal-fired power units, relevant combustion condition variables are collected as multi-dimensional input features through the distributed control system of the industrial site. To adapt to the time-series dependency extraction requirements of deep learning models, the model input time step is set to 50. A sliding window technique is used to extract and construct sample pairs along the time axis for the multi-dimensional operating condition variables and the corresponding original NOx emission concentration sequences. Finally, the constructed dataset is divided into a training set and a test set; the training set is used for subsequent iterative updates of the weight parameters within the hybrid network, and the test set is used to evaluate the generalization ability of the final model when facing unknown variable operating condition data.
[0029] Preferably, adaptive filtering preprocessing based on the ASG algorithm: refer to Figure 2 The input data then enters the adaptive SG filter layer. In this stage, the original sequence is preprocessed by cleaning and denoising using an adaptive SG filter algorithm based on local fluctuation statistical features.
[0030] Given the strong non-stationarity and high-frequency noise characteristics of NOx emission data collected from industrial sites, directly using conventional filters with fixed window lengths can easily lead to over-smoothing distortion of local abrupt changes in real-world operating conditions. Therefore, this embodiment assesses the severity of data fluctuations by extracting local statistical features and dynamically adjusts the sliding window size of the SG filter, specifically including: 1) Quantification of local fluctuation characteristics: Set the statistics window length to To accurately capture the fluctuation level of non-stationary signals, this embodiment uses local variance instead of mean or information entropy as the evaluation criterion. Calculate the current time... Local variance As a local statistical feature, its calculation formula is: .
[0031] in, For the sequence data within the statistics window, This is to calculate the mean of the data within the statistical window. The larger the value, the more drastic the fluctuation of the operating data within that local range, and the richer the dynamic characteristics it contains. Theoretically, a smaller filtering window should be used to preserve these characteristics.
[0032] 2) High robustness normalization treatment: If we directly address the local variance Conventional linear mappings (such as Min-Max normalization) are highly susceptible to interference from outliers in high-frequency noise, leading to distortion of the normalization coefficients. Therefore, this embodiment introduces a robust statistical method with a higher collapse point. The first quartile of the entire local variance sequence is extracted. and the third and fourth quartiles Calculate the robust normalization coefficient : .
[0033] This truncation and normalization mechanism ensures that when the local variance is within the normal fluctuation range... Smooth mapping in It effectively shields against interference from extreme outliers.
[0034] 3) Adaptive weighting and odd-number reconstruction of the filter window: Based on the calculated robust normalization coefficient Within the preset minimum filtering window and maximum filter window Weighted interpolation is performed between them. In a specific embodiment, to match the data sampling frequency and combustion delay characteristics of the thermal power boiler, it is preferable to set... , .
[0035] Calculate the current time Theoretical sliding window size : .
[0036] Because the window of a standard SG filter must satisfy the centrosymmetry property, meaning the actual window length must be a positive odd number, in practical implementation, the calculated continuous values need to be... Perform rounding and odd-number conversion to generate an odd-numbered window size for practical applications. : .
[0037] In the formula, This is the Gaussian rounding symbol, indicating that the value within the parentheses is rounded down. This mechanism is used to handle drastic fluctuations in local operating conditions (…). When the window quickly converges to... To track sudden changes in operating conditions; when the operating condition data tends to stabilize ( When the window automatically expands to [size], To filter out steady-state high-frequency noise to the maximum extent.
[0038] 4) Local polynomial least squares reconstruction: Based on the dynamically determined odd-numbered sliding window size mentioned above The original sequence is smoothly reconstructed using the polynomial least squares method. The half-window width of the current dynamic window is defined as... The order of the polynomial fitting is set to . (In this embodiment, it is preferably set to) ), then at the current time Smooth reconstruction of numerical values The calculation formula is: .
[0039] in, The original data points within the sliding window. These are the polynomial convolution fitting coefficients, which are determined by the fitting order. It is determined together with the current window size.
[0040] Through this step, the ASG filter layer generates high-quality, stationary, multivariate time-series data at the output, fundamentally eliminating the problem of local abrupt feature distortion caused by fixed window filtering, and providing high-fidelity feature input for subsequent temporal convolution modules.
[0041] Specifically, multi-scale local feature extraction based on the TCN module: like Figure 2 The data flow shown is the same as Figure 3 The module's internal structure, as shown, inputs the stationary multivariate time-series data output from the aforementioned steps into the Temporal Convolutional Network (TCN) module. Addressing the dynamic characteristics of thermal power boiler combustion systems, such as large delays and strong inertia, this embodiment utilizes a hierarchical dilated causal convolutional structure with residual connections within the TCN. This achieves an exponential expansion of the historical receptive field with a logarithmic number of network layers, thereby mapping and extracting high-dimensional local feature patterns from the input data at different time scales. The specific steps are as follows: 1) High-dimensional hidden space mapping: First, a normalized convolutional layer with a kernel size of 1 is used ( Conv performs a cross-channel linear combination of the input low-dimensional stationary multivariate time-series data, mapping it to a high-dimensional hidden layer space of a preset dimension. In this embodiment, it is preferable to uniformly map the 4-dimensional input sequence to a 128-dimensional feature space to provide sufficient representation dimensions for subsequent deep convolution feature extraction.
[0042] 2) Hierarchical dilated causal convolution extraction: The mapped high-dimensional feature sequence is fed into a hierarchical dilated causal convolutional network composed of multiple stacked residual blocks. Causal convolution strictly adheres to time constraints, ensuring that feature computation at the current time step depends only on data from the current and historical time steps, avoiding future information leakage. Simultaneously, a dilated convolution strategy is introduced, significantly expanding the effective receptive field without increasing model computational complexity or parameter count by inserting intervals between sampling points of the convolution kernel. A single-layer dilated convolution at time... The formula for calculating the output is: .
[0043] In the formula, It is the high-dimensional feature tensor input from the previous layer; Output the dilated convolutional features at the current time step; These are the learnable weight coefficients corresponding to the convolution kernel; The kernel size (preferably set to [size] in this embodiment) ), This is the kernel index.
[0044] To cover the entire long reaction cycle of the boiler from fuel intake to NOx production, the expansion factor within the network... With the number of hidden layers in the network It increases exponentially, satisfying: .
[0045] In this embodiment, three stacked dilated causal convolutions were constructed, with the corresponding dilation factors set sequentially as follows: .
[0046] 3) Residual connectivity and feature regularization: As the number of layers in a TCN network increases, it is prone to gradient vanishing or gradient exploding, making model training difficult to converge. Therefore, as... Figure 4 As shown, residual connections are introduced in each network layer of this module, allowing gradients to be directly propagated across layers during backpropagation. The final feature output calculation formula for residual blocks with residual connections is as follows: .
[0047] In the formula, These are the original input features of the residual block; This is the residual transformation function.
[0048] To ensure the stability of feature extraction and suppress overfitting, the residual transformation function Internally, it includes: dilated causal convolution operation, weight normalization operation for normalizing the internal covariate shift, ReLU activation function for providing nonlinear mapping capability, and Dropout regularization operation for randomly discarding neurons at a certain ratio (preferably set to 0.15) after each activation.
[0049] After deep feature extraction and layer-by-layer fusion of multiple residual blocks with different inflation factors, the TCN module outputs a high-dimensional local feature pattern with multi-scale temporal dependencies and long-term memory without losing time series resolution, and passes it to the downstream attention module for secondary optimization.
[0050] 4) Physics prior fusion and adaptive attention feature optimization based on the PASSA module: refer to Figure 2 While the high-dimensional local feature patterns extracted by the TCN module possess multi-scale time dependence, the complex multivariate coupling and physical delays in thermal power boiler systems inevitably introduce redundant irrelevant variable interactions and residual background noise into this high-dimensional feature space. If these are fed directly into the time-series prediction network without filtering, the model is highly susceptible to learning spurious correlations that violate thermodynamic causality (i.e., mistaking occasional noise within the non-delay window for features). Therefore, this embodiment constructs a... Figure 4 The illustrated Physical Prior-Guided Adaptive Sparse Self-Attention (PASSA) module performs secondary optimization and recombination of the feature sequence through underlying tensor fusion of industrial physical priors. Specifically: 1) Construction of characteristic linear projection and physical delay mask matrix: First, the high-dimensional local feature patterns output by the TCN module are input into the PASSA module. Through linear projection, these patterns are mapped onto the feature subspace, generating query matrices. (Query), Key Matrix (Key) and value matrix (Value).
[0051] In specific industrial implementation scenarios, this embodiment constructs a physical mechanism delay mask matrix. The specific logic is as follows: Assume the data sampling period of the distributed control system (DCS) is 10 seconds, and the time step of the model input is set to 50. Based on cross-correlation analysis of historical operating data from boiler combustion kinetics and the denitrification system, it is known that the significant impact of ammonia injection on NOx concentration at the flue gas outlet typically has a physical delay of 30 to 120 seconds. This is mapped to the model's time step, i.e., setting a lower limit for the effective physical delay time. Upper limit of decay time .
[0052] Based on the hyperparameters set above, a matrix with the same dimension as the attention score matrix is automatically initialized. For any element in the matrix The program assigns values using conditional statements: when the time step difference value meets the condition... When the historical operation is determined to be within the actual physical delay response period, a value is assigned. If the difference is less than 3 or greater than 12, it is determined to be redundant noise interference without causal physical correlation, and a value is assigned. .
[0053] 2) Two-branch attention scores are calculated independently: refer to Figure 4 The PASSA module employs a dual-branch parallel architecture. When calculating the attention score, it forcibly injects the aforementioned physical delay mask matrix through underlying matrix addition, and independently calculates the dense and sparse attention scores respectively. In the Dense Self-Attention (DSA) branch, the dense attention score matrix is calculated using the standard normalized exponential function (Softmax). This branch aims to preserve the global dependencies and long-term trends of the time series, preventing the loss of information about the series' continuity. Its calculation formula is as follows: .
[0054] in, This is a scaling dimension constant used to prevent the gradient from vanishing due to an excessively large dot product result; This is a learnable bias matrix. Because the matrix... Elements that exceed the physical delay range are set to After Softmax activation, the weights of these historical time steps that do not conform to thermodynamic laws are forced to zero, completely eliminating noise spurious correlations.
[0055] In the Sparse Self-Attention (SSA) branch, to address the drawback of traditional Softmax which disperses attention across all time steps and introduces noise, this embodiment uses SquaredReLU instead of Softmax to compute the sparse attention score matrix. This mechanism can concentrate the majority of attention resources on a few key time steps of sudden changes in operating conditions, effectively shielding low-relevance background noise. Its calculation formula is as follows: .
[0056] 3) Adaptive gated weighting and feature fusion: To dynamically balance noise reduction requirements and information integrity under different operating conditions, an adaptive gating mechanism is implemented within the module. Two independent, learnable parameters are set. and These correspond to the channel weights of the dense and sparse branches, respectively. The adaptive weighted weights of the two branches are calculated using exponential normalization. and : .
[0057] .
[0058] The attention score matrices of the two branches are fused using the aforementioned adaptive weighting, and then combined with the value matrix. Multiplying these matrices yields a fusion attention output matrix that enhances key dynamic information. : .
[0059] This gating strategy automatically increases network bandwidth when data noise is severe. Enhance sparse filtering; when operating conditions are stable, the network will increase... To preserve the overall trend.
[0060] 4) Feature Recombination and Output: The fusion attention output matrix Features are recombined sequentially through a linear projection layer and a Dropout regularization layer. While maintaining the global receptive field, the risk of network overfitting is further reduced, and a refined and optimized high-quality preferred feature sequence is output, which is then fed into the downstream LSTM module for final temporal modeling.
[0061] Furthermore, long-term memory modeling and emission prediction based on LSTM modules: like Figure 2 The data flow shown is the same as Figure 5 The unit-based temporal modeling structure shown inputs the generated preferred feature sequence into the Long Short-Term Memory (LSTM) network unit step by step. Since NOx generation in a thermal power boiler combustion system is a dynamic physical process with large delays and strong inertia, the emission concentration at the current moment is often influenced by the cumulative effects of operating conditions over a relatively long period. Therefore, this embodiment utilizes the synergistic effect of the forget gate, input gate, and output gate within the LSTM network to perform nonlinear updates to the temporal state, thereby addressing the gradient decay problem in long sequence training. The specific modeling and processing steps are as follows: 1) Multi-gated state nonlinear update: Let the preferred input feature vector at the current time step be... The hidden layer state at the previous time step was The cell state at the previous time step was The core state update within the unit is as follows: First, through the Gate of Oblivion The decision was made to discard outdated historical status information that is irrelevant to the current operating conditions: .
[0062] Secondly, through the input gate The decision is made to incorporate key information such as the current time-step mutation load, and to generate candidate cell states for the current time step. : .
[0063] .
[0064] Then, by combining the outputs of the forget gate and the input gate, the updated cell state at the current time step is calculated. : .
[0065] Finally, through the output gate Calculate the updated hidden state at the current time step. : .
[0066] .
[0067] In the formula, It is the Sigmoid activation function. The hyperbolic tangent activation function is used. It represents the Hadamardi (or Hadama) stack; , , , These are the weight matrices corresponding to each gating structure; , , , These are the corresponding bias vectors.
[0068] 2) Deep network stacking and regularization to prevent overfitting: To ensure the model has sufficient nonlinear fitting capability to cope with extremely complex varying operating conditions, this embodiment constructs a deep LSTM network with multiple hidden layers. Preferably, three LSTM hidden layers are stacked, with each hidden layer configured with 128 neurons. Simultaneously, to prevent the model from overfitting to high-dimensional input features during training and to enhance its generalization ability under unknown conditions, a Dropout regularization mechanism is introduced between each LSTM hidden layer, with the Dropout ratio set to 0.15. This configuration forces the network to randomly detach some neuron connections during training, thereby learning more robust physical mapping features.
[0069] 3) Regression output of prediction results: Get the final hidden state after all time steps and time-series state updates. The data is then fed into a fully connected layer for linear mapping and dimensionality compression, ultimately outputting the predicted regression results of nitrogen oxide (NOx) emission concentrations from thermal power boilers at the target time.
[0070] Optionally, model iterative optimization and performance evaluation verification: The predicted NOx emission concentration at the target time, output from the aforementioned steps, is compared with the actual monitoring sequence. The error loss between the predicted and actual sequences is calculated. The model parameters are iteratively optimized using the backpropagation algorithm until convergence, resulting in the final prediction model after training. To fully verify the effectiveness and advancement of the NOx prediction method integrating adaptive filtering and attention mechanisms proposed in this embodiment, the training process and evaluation metrics of the model are explained below: 1) Network hyperparameter settings and model training: During the model training phase, to ensure the stability and convergence speed of parameter updates within the hybrid network, this embodiment uses the Adam optimizer for gradient descent optimization. The initial learning rate is uniformly set to 0.0001, the training batch size is set to 128, and the maximum number of epochs is set to 50. Simultaneously, an early stopping mechanism is introduced during training. Training is automatically terminated when the loss on the validation set no longer decreases within a certain number of consecutive epochs to prevent the network from overfitting to high-frequency industrial noise.
[0071] 2) Quantification of model evaluation indicators: After training, the performance of the final lightweight, high-precision NOx prediction model is evaluated using the initially partitioned independent test set. This embodiment selects three core evaluation metrics commonly used in the field of time series forecasting to objectively quantify the model: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The specific calculation formulas for each metric are defined as follows: The root mean square error (RMSE) is highly sensitive to extreme prediction biases and can effectively measure the overall dispersion of prediction results. Its calculation formula is as follows: .
[0072] The mean absolute error (MAE) reflects the average level of the absolute error between the predicted and actual values, exhibiting good robustness. Its calculation formula is as follows: .
[0073] Mean Absolute Percentage Error (MAPE) is used to intuitively measure the relative proportion of prediction error to the true value. Its calculation formula is as follows: .
[0074] In the above formula, The total number of samples in the test set. For the first The actual NOx emission monitoring observation value at each moment. These are the predicted NOx concentration values for the corresponding time points output by the model in this embodiment. The smaller the values of each indicator, the higher the prediction accuracy of the model.
[0075] 3) Analysis of experimental results and verification of beneficial effects: To further highlight the advantages of the hybrid prediction model (ASG+TCN-PASSA-LSTM architecture) constructed in this embodiment when processing industrial data with strong non-stationarity and high-frequency noise, this embodiment conducted a comparative experiment with mainstream time series benchmark prediction models in the prior art under the same experimental environment and dataset conditions. The comparison models included: long-sequence linear benchmark model (DLinear), deep convolutional network (ResNet), temporal convolutional network (TCN), basic recurrent neural network (RNN) and its variant (GRU), and a model based on self-attention mechanism (Transformer). The quantitative evaluation results of the prediction performance of each model on independent test sets are shown in Table 1.
[0076] Table 1
[0077] As shown in Table 1, the fusion model proposed in this embodiment exhibits significant performance advantages across all evaluation metrics. Specifically, the root mean square error (RMSE) of this embodiment is as low as 2.2744, a substantial reduction of approximately 79.8% compared to the second-best performing GRU model (RMSE = 11.2688). Furthermore, traditional attention mechanism models such as Transformer are prone to overfitting (RMSE as high as 20.8274) when faced with industrial data accompanied by high-frequency noise. This embodiment, however, actively filters out low-correlation noise interactions by introducing a physically prior-guided adaptive sparse self-attention (PASSA) mechanism, successfully overcoming the technical deficiency of being sensitive to noise.
[0078] 4) Ablation experiments to validate the model's network architecture: To verify the indispensability of each key module in the TCN-PASSA-LSTM hybrid network architecture of this embodiment and the synergistic gain effect between modules, an ablation study was designed for the system. By progressively stacking the Temporal Convolutional Network (TCN), the Physically Prior-Guided Adaptive Sparse Self-Attention Mechanism (PASSA), and the Adaptive Filtering (ASG) module, the contribution of each component to the overall performance was quantitatively analyzed. The experimental results are shown in Table 2.
[0079] Table 2
[0080] As shown in Table 2, the progressive integration evolution path of the model clearly demonstrates the synergistic gain effect of "1+1>2" generated by the superposition of each module. The basic LSTM model, constrained by long sequence gradients and strong noise interference, has an RMSE as high as 40.2388. After introducing the TCN module, its multi-scale dilated convolution captures local features, reducing the error to 33.7673. Further introducing the PASSA module constructed in this embodiment, redundant noise information is filtered out through low-level physical mask constraints and dynamic feature sparsity filtering mechanisms, significantly reducing the error to 24.4846. Finally, after integrating the ASG adaptive filtering module to form a complete model, the non-stationary fluctuation problem of the data is solved from the source, causing a cliff-like drop in error, ultimately reducing it to 2.2744. Preferably, the above ablation experiment data fully demonstrates that the hybrid prediction architecture proposed in this embodiment organically integrates and achieves end-to-end synergistic optimization from data cleaning and key feature extraction to deep temporal modeling, producing superior technical effects that cannot be predicted by existing single prediction models or conventional combinations.
[0081] The beneficial effects of this invention are as follows: (1) The adaptive SG filtering (ASG) algorithm based on the evaluation of local fluctuation statistical features constructed in this invention breaks through the limitations of traditional fixed window filtering. It can dynamically sense data fluctuations and adjust the window size, thus achieving an adaptive balance between suppressing high-frequency noise and preserving local mutation features from the source.
[0082] (2) The present invention designs a feature extraction architecture that integrates Temporal Convolutional Network (TCN) and Physical Prior Guided Adaptive Sparse Self-Attention Mechanism (PASSA). This architecture innovatively injects the combustion delay prior into the attention computation layer in the form of a mask matrix, and through the dynamic fusion of sparse and dense dual branches, it accurately removes redundant noise interaction information, effectively enhancing the key feature expression and model noise robustness under complex working conditions.
[0083] (3) This invention combines Long Short-Term Memory (LSTM) networks to perform deep time-series modeling of the large delay characteristics of the combustion process, which greatly improves the prediction accuracy. This invention significantly reduces the prediction error under varying operating conditions, providing reliable technical support for the environmentally friendly refined feedforward control and ultra-low emissions of coal-fired power units.
[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0085] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A NOx prediction method based on improved ASG filtering, characterized in that, include: Collect operating parameters of the target thermal power boiler unit; The operating parameters of the process are input into a pre-trained hybrid prediction model for prediction to obtain NOx prediction data; The pre-training process of the hybrid prediction model includes: Construct an initial prediction network; the initial prediction network includes: an input layer, an ASG module, a TCN module, a PASSA module, and an LSTM module connected in sequence; The training set was obtained by collecting multidimensional historical operating parameters and original sequences of nitrogen oxide emission concentrations of the target thermal power boiler unit during variable load operation. The ASG module is used to extract the volatility of the training set, the sliding window size of the SG filter is dynamically adjusted according to the volatility, and the SG filter is used to clean and denoise the training set to obtain stationary multivariate time series data. The stationary multivariate time series data is mapped and feature pattern extracted using the hierarchical dilated causal convolutional structure with residual connections in the TCN module at multiple time scales to obtain high-dimensional local feature patterns. The high-dimensional local feature pattern is input into the PASSA module. In the underlying tensor operation, redundant temporal interactions that do not conform to physical causality in the high-dimensional local feature pattern are eliminated and features are screened twice according to the physical mechanism delay mask matrix to obtain the preferred feature sequence. The delay and long memory characteristics of the preferred feature sequence are modeled using the gating mechanism in the LSTM module to obtain the predicted NOx emission concentration at the target time. The error between the predicted NOx emission concentration and the original NOx emission concentration sequence is extracted to obtain the output loss. Based on the output loss, the backpropagation algorithm is used to iteratively optimize the initial prediction network until the initial prediction network converges and the model training stops, thus obtaining the hybrid prediction model.
2. The NOx prediction method based on improved ASG filtering according to claim 1, characterized in that, The ASG module is used to extract the volatility of the training set. The sliding window size of the SG filter is dynamically adjusted based on the volatility, and the SG filter is used to perform cleaning and noise reduction preprocessing on the training set to obtain stationary multivariate time series data, including: Set the statistical window length, and calculate the variance of the training set according to the statistical window length to obtain local statistical features; Robust normalization coefficients are set based on the first and third quartiles of the local statistical features; the expression for the robust normalization coefficients is: ;in, The robust normalization coefficient is mentioned above; These are the local statistical features; This is the first quartile; This is the third quartile; This indicates that the minimum value is extracted. This indicates that the maximum value should be extracted. The sliding window size is obtained by performing weighted interpolation calculations on the robust normalization coefficients and the preset minimum and maximum filtering windows; the expression for the sliding window size is: ;in, The size of the sliding window; The maximum filtering window; This refers to the minimum filtering window; Based on the sliding window size, the sequences in the training set are smoothly reconstructed using the polynomial least squares method to obtain the stationary multivariate time series data.
3. The NOx prediction method based on improved ASG filtering according to claim 1, characterized in that, The stationary multivariate time-series data is mapped and feature pattern extracted using a hierarchical dilated causal convolutional structure with residual connections in the TCN module across multiple time scales, resulting in high-dimensional local feature patterns, including: The stationary multivariate time series data is mapped to a high-dimensional hidden layer space of a preset dimension using a convolutional layer with a kernel size of 1. The mapped stationary multivariate time-series data is convolved using a hierarchical dilated causal convolutional network composed of multiple stacked residual blocks to obtain a dilated convolution output. The expression for a single-layer dilated convolution in the hierarchical dilated causal convolutional network is as follows: ;in, The output of the dilated convolution at the current time step; The kernel size; These are the kernel weights; Input features; For convolution kernel index; By using multiple residual blocks with different dilation factors, deep feature extraction and layer-by-layer fusion are performed on the output of each layer of dilated convolution to obtain the high-dimensional local feature pattern.
4. The NOx prediction method based on improved ASG filtering according to claim 1, characterized in that, The high-dimensional local feature pattern is input into the PASSA module. In the underlying tensor operations, redundant temporal interactions that do not conform to physical causality in the high-dimensional local feature pattern are eliminated and secondary feature selection is performed based on the physical mechanism delay mask matrix, resulting in a preferred feature sequence, including: Linear projection is performed on the high-dimensional local feature pattern to obtain the query matrix, key matrix, and value matrix; The physical mechanism delay mask matrix is set according to the lower limit of the physical effective delay time and the upper limit of the decay time of the combustion system of the thermal power boiler. Based on the physical mechanism, the delay mask matrix is used to generate sparse attention score matrices and dense attention score matrices using an attention mechanism. The sparse attention score matrix and the dense attention score matrix are weighted and fused, and then multiplied with the value matrix to obtain the fused attention output matrix; the expression of the fused attention output matrix is: ;in, ; ; The fusion attention output matrix; , These are the first adaptive weighting weight and the second adaptive weighting weight, respectively. , These are the sparse attention score matrix and the dense attention score matrix, respectively; The value matrix; The first learning parameter; This is the second learning parameter; The fused attention output matrix is reorganized using a linear projection layer and a Dropout regularization layer in sequence to obtain the preferred feature sequence.
5. The NOx prediction method based on improved ASG filtering according to claim 1, characterized in that, The LSTM module utilizes a gating mechanism to perform deep time-series modeling on the delay and long memory characteristics of the preferred feature sequence, yielding a predicted NOx emission concentration at the target time, including: The preferred feature sequence is sequentially input into the long short-term memory network unit of the LSTM module for temporal nonlinear state update to obtain the final hidden state; The final hidden state is linearly mapped using a fully connected layer to obtain the predicted NOx emission concentration.
6. The NOx prediction method based on improved ASG filtering according to claim 3, characterized in that, The inflation factor of the hierarchical dilated causal convolutional network satisfies the following condition: ;in, This represents the number of network layers.
7. The NOx prediction method based on improved ASG filtering according to claim 4, characterized in that, Based on the physical mechanism, the delay mask matrix is used to generate sparse attention score matrices and dense attention score matrices using an attention mechanism, including: In the sparse self-attention branch, a sparse attention score matrix is generated using a squared-corrected linear unit based on the physical mechanism delay mask matrix; the expression for the sparse attention score matrix is: ;in, The query matrix; The key matrix; This is a scaling dimension constant; The learnable bias matrix; This indicates that the maximum value should be extracted. In the dense self-attention branch, a dense attention score matrix is generated using a normalized exponential function based on the physical mechanism delay mask matrix; the expression for the dense attention score matrix is: ;in, It is a normalized exponential function.
8. A NOx prediction method based on improved ASG filtering according to claim 4, characterized in that, The physical mechanism delay mask matrix satisfies the following condition: when hour, ,otherwise, ;in, Indicates the current query time step; Indicates the time step of the history key; This is the lower limit of the effective physical delay time; This is the upper limit of the decay time; The physical mechanism delay mask matrix is in the first... Line number The elements of the column.