Data-driven ammonia injection amount prediction method

Abstract

The application discloses an ammonia water injection amount prediction method based on data driving, and solves the problem of inaccurate ammonia water injection amount control, and comprises the following steps: collecting characteristic variable data and target characteristic data in real time, dividing by using a sliding window, and performing data preprocessing; a time series data driving model is constructed; the time series data driving model comprises a long-term encoder, a short-term encoder, a space-time fusion attention block, a flattening layer, a first BiLSTM and a prediction head; the preprocessed characteristic variable data and target characteristic data are divided into a training set, a validation set and a test set; the time series data driving model is trained by using the training set, and the Hyperband algorithm is used for hyperparameter optimization of the time series data driving model, so that the trained time series data driving model is obtained; the trained time series data driving model is verified, tested and model evaluated by using the validation set and the test set respectively. Through the above scheme, the application has the advantages of simple logic, precision and reliability.

Description

Technical Field

[0001] This invention relates to the field of environmental protection technology, and in particular to a data-driven method for predicting the amount of ammonia water injected. Background Technology

[0002] The glass manufacturing industry, as a typical high-energy-consuming and high-polluting process industry, continuously emits harmful gases such as nitrogen oxides (NOx) and sulfur dioxide during its melting and calcination processes, while consuming large amounts of fossil fuels. Studies have shown that NOx, as a key air pollutant, not only forms PM2.5 and ozone through photochemical reactions, exacerbating smog, but also triggers a chain of ecological problems such as acid deposition and eutrophication of water bodies. It should be noted that the NOx in this technology mainly consists of NO and NO2.

[0003] Among existing flue gas treatment technologies, Selective Catalytic Reduction (SCR) technology, with its high denitrification efficiency of up to 90%, has become the core method for reducing nitrogen oxide (NOx) emissions in the glass industry. This method is based on the chemical reaction between ammonia (NH3) and NOx, thus reducing NOx emissions. Injected into the flue gas duct, it reacts with NOx in the flue gas in the catalyst. A redox reaction occurs on the surface, ultimately producing harmless nitrogen (N2) and water vapor (H2O). The control effect of this process is highly dependent on the accuracy of ammonia injection prediction: insufficient ammonia injection will lead to excessive NOx emissions, while excessive ammonia injection will cause ammonia escape. Furthermore, it increases the cost of the reducing agent, and also leads to the formation of ammonium bisulfate (ABS), which clogs the catalyst pores, increases the system pressure differential, further exacerbates system instability, and increases operating costs. Therefore, achieving high-precision prediction of ammonia injection volume has become a key technical bottleneck for SCR systems to ensure economical and efficient operation while meeting environmental emission requirements.

[0004] Currently, ammonia injection control in SCR system mainly relies on two traditional methods: open-loop control with fixed NH3 / NOx molar ratio and closed-loop PID feedback control, as described in the article “Chemical and mechanistic aspects of the selective catalytic reduction of NOx by ammonia over oxide catalysts: a review”; In addition, in the article “Optimization of ammonia injection grid in hybrid selective non-catalyst reduction and selective catalyst reduction system to achieve ultra-low NOx emissions”, open-loop control calculates the theoretical ammonia injection amount by online monitoring of NOx concentration, and PID control dynamically adjusts the opening degree of ammonia injection valve according to the outlet NOx concentration deviation. The above traditional methods have obvious shortcomings in the actual operation of the glass furnace: First, the open-loop control with fixed ratio is difficult to adapt to the fluctuation of NOx generation caused by the change of fuel composition, and often needs to use the conservative over-injection strategy (NH3 / NOx ratio > 1.1), which leads to ammonia escape rate exceeding the environmental protection limit value of 3 ppm. And PID control is prone to overshoot due to the detection lag of flue gas. Second, the glass furnace changing operation will cause the flue gas temperature to change sharply (300→200°C) and the NOx concentration to fluctuate greatly (±30%), and the PID control with fixed parameters is difficult to effectively cope with this strong nonlinear process.

[0005] To solve the above problems, researchers have proposed some improved methods. Zhang et al. proposed an intelligent optimization ammonia injection control method based on multi-model state estimation in the paper "Intelligent optimization of diesel engine Selective catalytic reduction urea injection based on multi-model state estimation to reduce NH3 slip and NOx emission", which achieved the dual goals of NOx conversion efficiency and NH3 slip suppression in the SCR system. However, this method has high computational complexity and requires high hardware processing power, making it difficult to deploy in practical engineering applications. Xie et al. proposed a deep deterministic policy gradient (DDPG) based reinforcement learning method in the paper "Selective catalytic reduction system Ammonia injection control based on deep deterministic policy reinforcement learning", which trained an intelligent controller in a high-precision virtual environment to achieve dynamic adjustment of ammonia injection and effectively reduce NOx emissions and ammonia escape. However, this method relies on a large amount of simulation data and has a "reality gap" problem when migrating from the simulation environment to the real system, limiting its generalization ability and real-time deployment efficiency in practical applications. Li et al. proposed a collaborative optimization method based on a sparse auto-encoding bidirectional long short-term memory neural network (SAE-Bi-LSTM) combined with an improved particle swarm optimization (PSO) in the paper "Multi-parameter co-optimization for NOx emissions control from waste incinerators based on data-driven model and improved particle swarm optimization", which achieved collaborative optimization of waste incinerator combustion and denitration parameters. However, this method relies on iterative optimization, which may result in slow convergence speed in real-time strong nonlinear dynamic processes. Therefore, developing an ammonia injection control method with fast dynamic response capability, low hardware dependency, and adaptability to strong nonlinear processes remains a challenge.

[0006] Therefore, there is an urgent need to develop a simple and accurate ammonia injection control method based on data-driven. SUMMARY

[0007] To solve the above problems, the present application aims to provide a data-driven ammonia water injection amount prediction method, and the technical solution adopted by the present application is as follows:

[0008] The data-driven ammonia water injection amount prediction method comprises the following steps:

[0009] Real-time acquisition of characteristic variable data and target characteristic data, sliding window cutting, and data preprocessing;

[0010] A time series data-driven model is constructed; the time series data-driven model comprises a long-term encoder, a short-term encoder, a space-time fusion attention block, a flattening layer, a first BiLSTM, and a prediction head; the long-term encoder and the short-term encoder are arranged in parallel, acquire the preprocessed characteristic variable data, and use the long-term encoder to extract macro working condition features; the short-term encoder is used to extract time step transient representation data; the macro working condition features output by the long-term encoder and the time step transient representation data output by the short-term encoder are spliced in the channel dimension to form a fusion tensor, which is input into the space-time fusion attention block; the space-time fusion attention block performs fusion feature enhancement on the input fusion tensor to obtain enhanced fusion features; the flattening layer converts the enhanced fusion features output by the space-time fusion attention block to obtain one-dimensional fusion features; the first BiLSTM is connected with the flattening layer and performs deep time series integration on the one-dimensional fusion features to obtain compact time series representation; the prediction head predicts the compact time series representation output by the first BiLSTM through a fully connected layer to obtain the predicted ammonia water injection amount;

[0011] The preprocessed characteristic variable data and target characteristic data are divided into a training set, a validation set, and a test set;

[0012] The training set is used to train the time series data-driven model, and the Hyperband algorithm is used to optimize the hyperparameters of the time series data-driven model to obtain the trained time series data-driven model;

[0013] The validation set and the test set are used to validate and test the trained time series data-driven model, respectively, and the trained time series data-driven model is evaluated.

[0014] Further, the characteristic variable data comprises kiln temperature, kiln opening signal, SCR inlet flue gas temperature, SCR inlet pressure, SCR inlet flue gas flow, SCR inlet Concentration, SCR inlet Concentration and SCR inlet humidity; and the target characteristic data is ammonia water injection amount.

[0015] Further, the long-term encoder comprises a second BiLSTM and a convolution interpolation-offset module connected in sequence; the second BiLSTM acquires the preprocessed characteristic variable data to perform feature extraction, and obtains single-step steady-state features; the convolution interpolation-offset module performs cross-scale feature alignment processing on the single-step steady-state features output by the second BiLSTM, and obtains macro working condition features.

[0016] Further, the short-term encoder comprises a third BiLSTM and a fourth BiLSTM arranged in parallel, and a stacking layer connected with the third BiLSTM and the fourth BiLSTM; the third BiLSTM and the fourth BiLSTM respectively acquire characteristic variable data in different time segments intercepted in the preprocessed sequence, and perform bidirectional feature extraction on the characteristic variable data to obtain transient feature vectors in sequence bidirectional context; the stacking layer stacks and splices the transient features output by the third BiLSTM and the transient features output by the fourth BiLSTM, and obtains transient representation data containing two time steps.

[0017] Further, the spatio-temporal fusion attention block comprises a time global average pooling layer, a first one-dimensional convolution layer and a second one-dimensional convolution layer connected in sequence; the time global average pooling layer obtains a fusion tensor spliced in the channel dimension, and performs global average pooling processing on the fusion tensor in the corresponding entire time dimension to obtain a vector representing the global dynamic characteristics of each channel in the entire time window; the first one-dimensional convolution layer performs channel compression and dimension reduction on the global dynamic characteristic vector output by the time global average pooling layer; the second one-dimensional convolution layer restores the dimension of the data output by the first one-dimensional convolution layer and generates attention weights to obtain channel attention weights; the channel attention weights output by the second one-dimensional convolution layer are used to weight the fused feature map channel by channel to obtain an enhanced feature map.

[0018] Further, the convolution interpolation-offset module comprises an MLP and a transposed convolution layer arranged in parallel; the MLP is composed of three fully connected layers and a nonlinear activation function, and performs offset correction processing on the single-step steady-state features output by the second BiLSTM, and its expression is:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] wherein, denotes a linear transformation of the first fully connected layer; denotes a linear transformation of the second fully connected layer; denotes a linear transformation of the third fully connected layer; denotes a nonlinear activation function of the first fully connected layer; denotes a nonlinear activation function of the second fully connected layer; denotes weights of the first fully connected layer; denotes weights of the second fully connected layer; denotes weights of the third fully connected layer; denotes a long-term sequence of the input; denotes a bias of the first fully connected layer; denotes a bias of the second fully connected layer; denotes a bias of the third fully connected layer; denotes a dimension of the input feature; denotes a matrix with a dimension of ; denotes a matrix with a dimension of ; denotes a matrix with a dimension of ; denotes a one-dimensional array of 2d real numbers; denotes a number of neurons of the first fully connected layer; denotes a number of neurons of the second fully connected layer.

[0025] The transpose convolution layer performs a transpose convolution operation on a long-term sequence of a second BiLSTM output of the long-term encoder, and outputs a multi-step feature tensor with an expanded time dimension; the MLP outputs a time-specific offset after offset correction processing, and performs element accumulation with the multi-step feature tensor output by the transpose convolution layer to obtain alignment data with the same output dimension as the short-term encoder.

[0026] Further, the spatio-temporal fusion attention block comprises a time global average pooling layer, a first one-dimensional convolution layer and a second one-dimensional convolution layer connected in sequence.

[0027] Further, the prediction head comprises a regularization layer and a second fully connected layer connected in sequence.

[0028] Further, the Hyperband algorithm is used to perform hyperparameter optimization of the time series data driven model, comprising the following steps:

[0029] In any training iteration of the time series data driven model, the hyperparameters of the time series data driven model are sorted, and the hyperparameters ranked in the front are retained; the hyperparameters ranked in the front are configured with f times of additional computing resources to obtain the optimal hyperparameters.

[0030] Further, the time series data driven model adopts a coefficient of determination, a mean absolute error and a root mean square error for model evaluation.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] The long-term encoder extracts macroscopic working condition features, and the short-term encoder respectively compresses two 5-step sub-sequences through BiLSTM to obtain 2-time-step transient representations. Then, the double-time-scale features are aligned through a convolution interpolation-offset module (CILO module). The two features after alignment are spliced in the channel dimension to form a fusion tensor, and then the fusion feature is enhanced through a spatio-temporal fusion attention block (TFA module). Finally, the fusion feature is flattened and sent to the first BiLSTM, and after Dropout regularization, the output completes the end-to-end prediction of the ammonia water injection amount. The long-term encoder in the time series data driven model depicts the continuous process of the steady state working condition with a low-frequency global receptive field, which can be regarded as a "baseline" representing the macroscopic dynamics of the system; the short-term encoder captures the changes at the turning points of the working condition with a high-frequency local receptive field, forming discrete and specific "event points". The point-line coupling in the space-time dimension realizes the collaborative representation of the global continuity and local discreteness of the boiler running state.

[0033] The long-term encoder of the present application models macroscopic steady-state features, and the short-term encoder captures transient disturbances, and the feature alignment is realized through the convolution interpolation-offset module (CILO module), which effectively aligns the long and short time series, thereby balancing stability and sensitivity, and solving the problem of traditional single-scale models that are difficult to deal with the melting furnace fire change and temperature sudden change.

[0034] The spatio-temporal fusion attention block (TFA module) of the present application adopts a temporal global average pooling (Temporal GAP, TGAP). It performs pooling in the time step (T) dimension to obtain a vector representing the global dynamic characteristics of each channel in the entire time window. The transformation is based on the characteristics of the time series data mode, so that the attention mechanism plays a key role in time series prediction.

[0035] The convolution interpolation-offset module (CILO module) of the present application realizes the up-sampling of long-term features in the time dimension through transposed convolution, and combines a learnable offset correction mechanism, so that the expanded multi-step features have differentiated expression while maintaining the global trend, thereby enhancing the alignment and fusion effect of cross-scale features.

[0036] In summary, the present application has the advantages of simple logic, precision and reliability, and has high practical value and popularization value in the field of environmental protection technology. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as limiting the scope of protection. For those skilled in the art, other related drawings can also be obtained from these drawings without creative labor.

[0038] Figure 1 The structural schematic diagram of the time sequence data driven model of the present application.

[0039] Figure 2 The framework diagram of the space-time fusion attention block of the present application.

[0040] Figure 3 The framework diagram of the convolution interpolation-offset module of the present application.

[0041] Figure 4 The hyperparameter tuning flowchart of the present application.

[0042] Figure 5 The data set distribution diagram of the present application. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present application more clear, the following will further illustrate the present application by combining with the drawings and embodiments. The embodiments of the present application include but are not limited to the following embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0044] In the present embodiment, the term "and / or" is only used to describe the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B can represent the following three cases: A exists alone, A and B exist together, and B exists alone.

[0045] In the present embodiment, the terms "first" and "second" in the specification and claims are used to distinguish different objects, and are not used to describe the specific order of the objects. For example, the first target object and the second target object are used to distinguish different target objects, and are not used to describe the specific order of the target objects.

[0046] In the present embodiment, the words such as "exemplary" or "for example" are used to represent as an example, illustration or description. Any embodiment or design scheme described as "exemplary" or "for example" in the present embodiment should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. On the contrary, the words such as "exemplary" or "for example" are intended to present the related concept in a specific way.

[0047] In the description of the embodiments in this application, unless otherwise stated, "multiple" means two or more. For example, multiple processing units means two or more processing units; multiple systems means two or more systems.

[0048] like Figures 1 to 5 As shown, this embodiment, based on the characteristics of the SCR process in a glass melting furnace, provides a data-driven method for predicting ammonia injection rate to effectively predict the ammonia injection volume while meeting low NOx emission requirements. First, a baseline model is established, and an ammonia injection rate prediction model is developed based on this model. Then, an alignment module and a feature fusion enhancement module are designed based on the dual-timescale characteristics. Finally, the Hyperband algorithm is used to optimize the model parameters, and the optimization results are applied to the model to achieve efficient and stable prediction.

[0049] The first step involves real-time acquisition of feature variable data and target feature data, segmentation using a sliding window, and data preprocessing. In this embodiment, a sliding window Pearson analysis is employed, selecting kiln temperature, kiln opening signal, flue gas temperature at the SCR inlet, pressure at the SCR inlet, flue gas flow rate at the SCR inlet, and other parameters. Concentration and SCR inlet Seven core parameters were used for concentration, with the target characteristic data being the ammonia injection rate. A multi-level data quality control strategy was employed during data preprocessing. First, linear interpolation was used to initially handle missing values. Subsequently, a multi-level outlier handling mechanism was established: the first stage employed hard screening based on equipment physical limits, followed by conformal piecewise cubic Hermite interpolation for data repair; the second stage used dynamic statistical detection with a 56-hour sliding window to identify and correct outliers deviating from the window mean ±3σ. Finally, normalization was used to eliminate dimensional differences between parameters.

[0050] In the second step, a time series data-driven model is constructed. The time series data-driven model comprises a long-term encoder, a short-term encoder, a space-time fusion attention block (TFA module), a flattening layer, a first BiLSTM, and a prediction head. The long-term encoder and the short-term encoder are arranged in parallel, obtain the preprocessed feature variable data, extract macro working condition features by using the long-term encoder, and extract transient representation data of a time step by using the short-term encoder. The macro working condition features output by the long-term encoder and the transient representation data of the time step output by the short-term encoder are spliced in a channel dimension to form a fusion tensor, which is input into the space-time fusion attention block. The space-time fusion attention block performs fusion feature enhancement on the input fusion tensor to obtain enhanced fusion features. The flattening layer converts the enhanced fusion features output by the space-time fusion attention block to obtain one-dimensional fusion features. The first BiLSTM is connected with the flattening layer and performs deep time series integration on the one-dimensional fusion features to obtain compact time series representation. The prediction head performs prediction on the compact time series representation output by the first BiLSTM through a fully connected layer to obtain a predicted ammonia water injection amount.

[0051] In the present embodiment, the long-term encoder comprises a second BiLSTM and a convolution interpolation-offset module (CILO module) connected in sequence. The second BiLSTM obtains the preprocessed feature variable data to extract features and obtain single-step steady-state features. The convolution interpolation-offset module performs cross-scale feature alignment processing on the single-step steady-state features output by the second BiLSTM to obtain macro working condition features.

[0052] In addition, the short-term encoder comprises a third BiLSTM and a fourth BiLSTM arranged in parallel, and a stacking layer connected with the third BiLSTM and the fourth BiLSTM. The third BiLSTM and the fourth BiLSTM respectively obtain feature variable data of different time segments intercepted in the preprocessed sequence, and perform bidirectional feature extraction on the feature variable data to obtain transient feature vectors of sequence bidirectional context. The stacking layer stacks and splices the transient features output by the third BiLSTM and the transient features output by the fourth BiLSTM to obtain transient representation data containing two time steps.

[0053] As Figure 2As shown, the spatio-temporal fusion attention block (TFA module) includes a time global average pooling layer, a first one-dimensional convolution layer and a second one-dimensional convolution layer connected in sequence; the time global average pooling layer obtains a fused tensor spliced in the channel dimension, and performs global average pooling processing on the fused tensor in the corresponding entire time dimension to obtain a vector representing the global dynamic characteristics of each channel in the entire time window; the first one-dimensional convolution layer performs channel compression and dimension reduction on the vector output by the time global average pooling layer; the second one-dimensional convolution layer restores the dimension of the data output by the first one-dimensional convolution layer and generates attention weights to obtain channel attention weights; and the channel attention weights output by the second one-dimensional convolution layer are used for channel-by-channel weighting on the fused feature map to obtain an enhanced feature map. In this embodiment, the TFA module adopts a temporal global average pooling (TGAP). It performs pooling in the time step (T) dimension to obtain a vector representing the global dynamic characteristics of each channel in the entire time window.

[0054] In the TFA module, two 1D convolution layers (i.e., the first Conv1D and the second Conv1D) are used to generate attention weights. Since the 1D convolution has the characteristics of local connection and weight sharing, it can better capture the local patterns in the time dimension, while the FC layer used in the SE Net will destroy such patterns; at the same time, the parameter amount of the 1D convolution is much less than that of the FC layer, and the calculation is more efficient, which is more suitable for deployment in industrial control systems that require real-time performance.

[0055] Here, the convolution interpolation-offset module (CILO module) includes an MLP and a transposed convolution layer arranged in parallel; the MLP is composed of three fully connected layers and a nonlinear activation function, and performs offset correction processing on the single-step steady-state feature output by the second BiLSTM.

[0056] In the double-time-scale structure (i.e., the long-term encoder and the short-term encoder), the long-term encoder outputs a single-step steady-state feature of (1x194) after being extracted by the second BiLSTM, while the short-term encoder outputs a transient feature of (2x194) after being stacked by the third BiLSTM and the fourth BiLSTM. There is a step difference in the time dimension between the two, and if they are directly spliced, the dimensions will be inconsistent. The CILO module is used for processing in this embodiment, which realizes the up-sampling of the long-term feature in the time dimension through the transposed convolution, and combines the learnable offset correction mechanism, so that the expanded multi-step feature has differentiated expression while maintaining the global trend, thereby enhancing the alignment and fusion effect of the cross-scale features.

[0057] Thirdly, the pre-processed feature variable data and target feature data are divided into a training set, a validation set and a test set. In this embodiment, the data set is divided into a training set, a validation set and a test set in a ratio of 7:1.5:1.5 to better prevent overfitting and ensure model reliability. The distribution of the training set, the validation set and the test set is shown in Table 2. Figure 5 In the training process, the loss function adopts Mean Squared Error (MSE) to minimize the mean square error between the prediction and the actual ammonia water spraying amount.

[0058] Fourthly, the training set is used to train the time series data driven model, and the Hyperband algorithm is used to optimize the hyperparameters of the time series data driven model to obtain the trained time series data driven model. In this case, the hyperparameters of the time series data driven model are sorted in each training iteration of the time series data driven model, and the top 1 / f hyperparameters are retained; the top 1 / f hyperparameters are configured with f times of additional computing resources to obtain the optimal hyperparameters.

[0059] Fifthly, the validation set and the test set are used to validate and test the trained time series data driven model respectively, and the trained time series data driven model is evaluated. In this case, the time series data driven model is evaluated by using the coefficient of determination, the mean absolute error and the root mean square error.

[0060] In this embodiment, the trained time series data driven model, the baseline model BiLSTM and four mainstream models (RF, KNN, XGBoost and CatBoost) are used for data testing. The real data and the prediction results are compared and visualized, and the highest R² value or the lowest MAE and RMSE are used as evaluation criteria to determine the best model performance. The R² values of the six models decrease in the following order: the trained time series data driven model (0.986), BiLSTM (0.978), XGBoost (0.962), CatBoost (0.961), KNN (0.920) and RF (0.909). In summary, DSAT-Net shows the best prediction performance among the six comparison models, with an R² value as high as 0.986, which is significantly better than the traditional machine learning methods (RF, KNN) and the gradient boosting models (XGBoost, CatBoost), and is 0.8 percentage points higher than the time series baseline model BiLSTM.

[0061] The above embodiment is only a preferred embodiment of the present application, and does not limit the protection scope of the present application. Any design principle of the present application and any changes made on the basis of non-creative labor shall fall within the protection scope of the present application.

Claims

1. A data-driven based ammonia injection amount prediction method, characterized by, The method comprises the following steps: Real-time acquisition of characteristic variable data and target characteristic data, sliding window segmentation, and data preprocessing; A time series data-driven model is constructed; the time series data-driven model comprises a long-term encoder, a short-term encoder, a space-time fusion attention block, a flattening layer, a first BiLSTM, and a prediction head; the long-term encoder and the short-term encoder are arranged in parallel, acquire the preprocessed characteristic variable data, extract macro working condition features by using the long-term encoder; extract transient representation data of the time step by using the short-term encoder; the macro working condition features output by the long-term encoder and the transient representation data of the time step output by the short-term encoder are spliced in the channel dimension to form a fusion tensor, which is input into the space-time fusion attention block; the space-time fusion attention block performs fusion feature enhancement on the input fusion tensor to obtain enhanced fusion features; The flattening layer converts the enhanced fusion features output by the space-time fusion attention block to obtain one-dimensional fusion features; The first BiLSTM is connected with the flattening layer and performs deep time series integration on the one-dimensional fusion features to obtain compact time series representations; the prediction head predicts the compact time series representations output by the first BiLSTM through a fully connected layer to obtain a predicted ammonia water injection amount; The preprocessed characteristic variable data and target characteristic data are divided into a training set, a validation set, and a test set; The training set is used to train the time series data-driven model, and the Hyperband algorithm is used to optimize the hyperparameters of the time series data-driven model to obtain a trained time series data-driven model; The validation set and the test set are used to validate and test the trained time series data-driven model, respectively, and the trained time series data-driven model is evaluated.

2. The data-driven based ammonia injection amount prediction method according to claim 1, characterized by, The characteristic variable data include kiln temperature, kiln opening signal, flue gas temperature at the SCR inlet, pressure at the SCR inlet, flue gas flow rate at the SCR inlet, and other parameters. Concentration, SCR inlet Concentration and humidity at the SCR inlet; the target characteristic data is the amount of ammonia water injected.

3. The data-driven-based ammonia injection amount prediction method according to claim 1 or 2, characterized by, The long-term encoder comprises a second BiLSTM and a convolution interpolation-offset module connected in sequence; the second BiLSTM acquires the preprocessed characteristic variable data to extract features to obtain single-step steady-state features; the convolution interpolation-offset module performs cross-scale feature alignment processing on the single-step steady-state features output by the second BiLSTM to obtain macro working condition features.

4. The data-driven based ammonia injection amount prediction method according to claim 3, characterized by, The short-term encoder comprises a third BiLSTM and a fourth BiLSTM arranged in parallel, and a stacking layer connected with the third BiLSTM and the fourth BiLSTM; the third BiLSTM and the fourth BiLSTM acquire characteristic variable data of different time segments in the preprocessed sequence, respectively, and perform bidirectional feature extraction on the characteristic variable data to obtain transient feature vectors of sequence bidirectional context; the stacking layer stacks and splices the transient features output by the third BiLSTM and the transient features output by the fourth BiLSTM to obtain transient representation data containing two time steps.

5. The data-driven-based ammonia injection amount prediction method according to claim 4, characterized by, The spatio-temporal fusion attention block comprises a time global average pooling layer, a first one-dimensional convolution layer and a second one-dimensional convolution layer connected in sequence; the time global average pooling layer obtains a fusion tensor spliced in a channel dimension, and performs global average pooling processing on the fusion tensor in a corresponding entire time dimension to obtain a vector representing global dynamic characteristics of each channel in an entire time window; the first one-dimensional convolution layer performs channel compression and dimension reduction on the global dynamic characteristic vector output by the time global average pooling layer; the second one-dimensional convolution layer restores the dimension of the data output by the first one-dimensional convolution layer and generates attention weights to obtain channel attention weights; The channel attention weights output by the second one-dimensional convolution layer are used to perform channel-by-channel weighting on the fused feature map to obtain an enhanced feature map.

6. The data-driven based ammonia injection amount prediction method according to claim 4, wherein The convolution interpolation-offset module comprises an MLP and a transpose convolution layer arranged in parallel; the MLP is composed of three fully connected layers and a nonlinear activation function, and performs offset correction processing on the single-step steady-state features output by the second BiLSTM, and its expression is as follows: wherein, represents the linear transformation of the first fully connected layer; represents the linear transformation of the second fully connected layer; represents the linear transformation of the third fully connected layer; represents the nonlinear activation function of the first fully connected layer; represents the nonlinear activation function of the second fully connected layer; represents the weight of the first fully connected layer; represents the weight of the second fully connected layer; represents the weight of the third fully connected layer; represents the input long-term sequence; represents the bias of the first fully connected layer; represents the bias of the second fully connected layer; represents the bias of the third fully connected layer; represents the dimension of the input feature; represents a matrix with a dimension of ; represents a matrix with a dimension of ; represents a matrix with a dimension of ; represents a one-dimensional array of 2d real numbers; represents the number of neurons of the first fully connected layer; represents the number of neurons of the second fully connected layer; the transpose convolution layer performs transpose convolution operation on the long-term sequence output by the second BiLSTM of the long-term encoder, and outputs a multi-step feature tensor with expanded time dimension; the MLP outputs a time series specific offset after offset correction processing, and performs element accumulation with the multi-step feature tensor output by the transpose convolution layer, to obtain alignment data with the same output dimension as the short-term encoder.

7. The data-driven based ammonia injection amount prediction method according to claim 6, wherein The spatio-temporal fusion attention block comprises a time global average pooling layer, a first one-dimensional convolution layer and a second one-dimensional convolution layer connected in sequence.

8. The data-driven based ammonia injection amount prediction method according to claim 7, wherein, The prediction head comprises a regularization layer and a second fully connected layer connected in sequence.

9. The data-driven based ammonia injection amount prediction method according to claim 8, wherein, The Hyperband algorithm is used to perform hyperparameter optimization of the time series data-driven model, comprising the following steps: In each training iteration of the time series data-driven model, the hyperparameters of the time series data-driven model are sorted, and the 1 / f hyperparameters in the front of the sorting are retained; the 1 / f hyperparameters in the front of the sorting are configured with f times of additional computing resources to obtain optimal hyperparameters.

10. The data-driven based ammonia injection amount prediction method according to claim 9, wherein The time series data-driven model uses the coefficient of determination, the mean absolute error and the root mean square error for model evaluation.

Images