A deep learning-based reactor inlet NO x Multi-step real-time prediction method and system

Abstract

This invention discloses a reactor inlet NO based on deep learning. x A multi-step real-time prediction method and system, relating to the fields of industrial process control and environmental monitoring technology, includes the following steps: acquiring time-series data of nitrogen oxide-related operating parameters in an industrial control system; preprocessing the time-series data to obtain preprocessed time-series data; constructing a hybrid neural network model; extracting deep time-series features from the preprocessed time-series data using the pre-trained hybrid neural network model; dynamically weighting the deep time-series features to obtain a weighted feature vector; iteratively predicting based on the weighted feature vector to obtain a sequence of predicted nitrogen oxide concentration values ​​for several future time steps; inversely normalizing the sequence of predicted nitrogen oxide concentration values ​​and extracting the predicted nitrogen oxide concentration values ​​for the corresponding time steps to complete the nitrogen oxide concentration prediction. This invention solves the problem of insufficient ability to capture long-term time-series dependencies in existing technologies, resulting in low prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and environmental monitoring technology, specifically to a reactor inlet NO based on deep learning. x Multi-step real-time prediction method and system. Background Technology

[0002] In industrial processes such as thermal power generation and cement production, fuel combustion produces large amounts of nitrogen oxides (NOx). x NO is one of the main pollutants causing environmental problems such as acid rain and photochemical smog. To control NO... x In response to emissions issues, selective catalytic reduction (SCR) technology is widely used. SCR systems inject a reducing agent such as ammonia (NH3) into the flue gas, where a catalyst reduces NO2 by decreasing its concentration. x It is reduced to harmless nitrogen and water. Precise control of the ammonia injection rate is crucial to denitrification efficiency and operating costs, and optimal control of the ammonia injection rate is highly dependent on the control of NO at the SCR reactor inlet. x Accurate prediction of NO concentration. However, the SCR system is a complex nonlinear system with large delays, large inertia, and multivariable coupling. First, from boiler combustion to NO... x There is a significant delay in flue gas transport between the analyzer's detection points; secondly, NO... x The analyzer itself also exhibits a time lag in response. These delays cause the DCS (Distributed Control System) to acquire input NO. x The concentration is a historical value rather than a real-time value, which makes the ammonia injection control based on this value lag, making it difficult to achieve optimal denitrification efficiency and ammonia slip control.

[0003] Currently, Chinese patent application CN112862180A discloses an inlet NO of a denitrification system. x The concentration prediction method uses Gradient Boosting Decision Tree (GBDT) to predict the delay time, and then uses Recurrent Neural Network (RNN) to predict NO. x Concentration. Chinese patent application CN111266004B discloses an inlet NO concentration of an SCR denitrification reactor. x The delay estimation system determines the relationship between influencing factors and NO by calculating the correlation coefficient. xDelays between concentrations. These methods improve prediction accuracy to some extent, but still have the following limitations: Relatively simple model structure: Most employ traditional RNNs or their variants, limiting their ability to capture longer and more complex dependencies in time series. Limited delay handling methods: Primarily relying on correlation analysis to estimate a fixed delay time, making it difficult to adapt to dynamic delays caused by changes in operating conditions. Insufficient multi-step prediction capability: Most methods focus on single-step or short-term prediction, failing to provide the control system with sufficiently long-term future trend information for forward-looking adjustments. Insufficient feature focus: Failing to effectively distinguish the impact of various input features on NO under different operating conditions. x The influence of concentration is weighted, lacking the dynamic weighting ability similar to an attention mechanism. Summary of the Invention

[0004] The purpose of this invention is to provide a reactor inlet NO based on deep learning. x A multi-step real-time prediction method and system are proposed to address the problem of low prediction accuracy caused by the insufficient ability of existing technologies to capture long-term time-series dependencies.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a reactor inlet NO based on deep learning x A multi-step real-time prediction method includes the following steps: Acquire time-series data of nitrogen oxide-related operating parameters in an industrial control system, preprocess the time-series data, and obtain preprocessed time-series data. A hybrid neural network model is constructed, and the deep temporal features of the pre-trained hybrid neural network model are extracted. The deep temporal features are dynamically weighted to obtain a weighted feature vector. Iterative prediction is performed based on the weighted feature vector to obtain a sequence of predicted nitrogen oxide concentration values ​​for several future time steps. The predicted nitrogen oxide concentration sequence is inversely normalized, and the predicted nitrogen oxide concentration at the corresponding time step is extracted to complete the nitrogen oxide concentration prediction.

[0006] In some implementations, the time-series data is preprocessed, specifically including: After data cleaning, the time series data is normalized using the MinMaxScaler scaler to obtain preprocessed time series data.

[0007] In some implementations, the hybrid neural network model consists of a temporal convolutional network, an SE attention layer, and a long short-term memory network connected in sequence. The temporal convolutional network is used to extract deep temporal features from the preprocessed temporal data; The SE attention layer is used to dynamically weight the deep temporal features to obtain a weighted feature vector. The long short-term memory network is used to receive the weighted feature vector and perform iterative prediction.

[0008] In some implementations, the SE attention layer dynamically weights the deep temporal features, specifically including: After performing global average pooling on each channel of the deep temporal features, the weights of each channel are learned through at least two fully connected layers and combined with the deep temporal features to obtain a weighted feature vector.

[0009] In some implementations, the temporal convolutional network consists of a stack of residual blocks, each of which includes an extended causal convolutional layer.

[0010] In some implementations, the pre-trained hybrid neural network model is trained through the following steps: Obtain time-series data of historical nitrogen oxide-related operating parameters and corresponding actual nitrogen oxide concentrations as training samples; The mean squared error is used as the loss function, and the hybrid neural network model is trained using the Adam optimizer to obtain a pre-trained hybrid neural network model.

[0011] Secondly, a reactor inlet NO based on deep learning x Multi-step real-time prediction systems include: The data acquisition module is used to acquire time-series data of nitrogen oxide-related operating parameters in the industrial control system, and to preprocess the time-series data to obtain preprocessed time-series data. The nitrogen oxide concentration prediction module is used to construct a hybrid neural network model, extract deep time-series features from the pre-trained hybrid neural network model, dynamically weight the deep time-series features to obtain a weighted feature vector, and perform iterative prediction based on the weighted feature vector to obtain a sequence of predicted nitrogen oxide concentration values ​​for several future time steps. The result output module is used to inversely normalize the nitrogen oxide concentration prediction value sequence and extract the nitrogen oxide concentration prediction value at the corresponding time step to complete the nitrogen oxide concentration prediction.

[0012] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor, when executing the computer program, implements the aforementioned deep learning-based reactor inlet NO. x The steps of a multi-step real-time prediction method.

[0013] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the deep learning-based reactor inlet NO. x The steps of a multi-step real-time prediction method.

[0014] Fifthly, a computer program product comprising a computer program that, when executed by a processor, implements the deep learning-based reactor inlet NO. x The steps of a multi-step real-time prediction method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a reactor inlet NO based on deep learning x A multi-step real-time prediction method provides the model with raw information covering long-term operating conditions by acquiring time-series data of nitrogen oxide (NOx) related operating parameters in industrial control systems. A hybrid neural network model is constructed, and a pre-trained hybrid neural network model extracts deep temporal features from the preprocessed time-series data. This model can automatically learn and abstract key temporal evolution patterns from data spanning long periods, directly overcoming the limitations of traditional models in modeling long-range correlations. Dynamic weighting of these deep temporal features yields a weighted feature vector, further strengthening features that remain important over long periods and suppressing noise and non-critical fluctuations. Iterative prediction based on the weighted feature vector yields a sequence of predicted NOx concentration values ​​for several future time steps, ensuring temporal consistency in multi-step prediction and avoiding error accumulation due to the breakage of long-term dependencies. Through the synergistic effect of these features, the method significantly improves the prediction of nitrogen oxide (NOx) concentrations in industrial processes characterized by large delays and high inertia. x The accuracy of concentration change prediction.

[0016] Furthermore, normalization is performed using the MinMaxScaler scaler to uniformly scale the operating parameters with different dimensions and large differences in numerical range to the same interval. This effectively avoids the problems of convergence difficulties or the submergence of certain features caused by inconsistent feature scales during model training, improves the stability and speed of model training, and also makes the model more adaptable to different operating conditions, which is conducive to improving the accuracy of the final prediction results.

[0017] Furthermore, the temporal convolutional network captures long-term dependencies in the time series; the SE attention layer can adaptively enhance important feature channels and suppress irrelevant features; the long short-term memory network learns the sequence dynamics and performs multi-step iterative generation, forming a complementary and synergistic prediction mechanism that significantly improves the model's ability to model complex industrial time series and its prediction accuracy. Attached Figure Description

[0018] Figure 1 A reactor inlet NO based on deep learning is provided as an embodiment of the present invention. x Flowchart of a multi-step real-time prediction method; Figure 2 A reactor inlet NO based on deep learning is provided as an embodiment of the present invention. x Structure diagram of a multi-step real-time prediction system; Figure 3 A reactor inlet NO based on deep learning is provided as an embodiment of the present invention. x A flowchart illustrating the multi-step real-time prediction method. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content described herein is for explanation rather than limitation of the present invention.

[0020] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this invention are intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, systems, products, or devices.

[0021] like Figure 1 and Figure 3 As shown, this embodiment provides a reactor inlet NO based on deep learning. x A multi-step real-time prediction method includes the following steps: S1, acquire time-series data of nitrogen oxide-related operating parameters in the industrial control system, preprocess the time-series data, and obtain preprocessed time-series data; Real-time batch retrieval of data related to NO from the database of industrial control systems (such as DCS). x Time series data of multiple operating parameters related to concentration; the acquired multidimensional time series data are cleaned and aligned, and a fixed-length sliding data window is constructed, with the data within the window used as model input.

[0022] In this embodiment, the system connects to the power plant's DCS or real-time database through an API client (such as the ApiClient class). Based on a predefined point mapping table, the client periodically (e.g., once per second) batch requests multiple points related to NO. xGenerate the latest values ​​for relevant operating parameters. These parameters may include, but are not limited to: superheat, separator outlet temperature, inlet and outlet steam temperatures of each stage of superheater, main steam temperature, setpoint, and desuperheating water regulating valve opening. The API client is responsible for tasks such as request signing, sending, and failure retries to ensure the reliability of data acquisition.

[0023] The acquired real-time data points (a timestamped snapshot containing multiple parameter values) are fed into a data window manager (such as a DataWindow class). This manager maintains a fixed-length (e.g., 96 time points) first-in, first-out queue. Whenever a new data point is added, the oldest data point is removed. This sliding window design ensures that the model always makes predictions based on the most recent historical data. Before being fed into the model, the data within the window is organized into a DataFrame and normalized. Normalization uses a pre-saved MinMaxScaler scaler to scale all feature values ​​to the [0, 1] interval to eliminate dimensional differences and accelerate model convergence.

[0024] S2, construct a hybrid neural network model, use the pre-trained hybrid neural network model to extract deep time-series features of the preprocessed time-series data, dynamically weight the deep time-series features to obtain a weighted feature vector, and perform iterative prediction based on the weighted feature vector to obtain a sequence of predicted nitrogen oxide concentration values ​​for several future time steps; The preprocessed data is input into a pre-trained deep learning model. First, a Temporal Convolutional Network (TCN) layer performs convolution operations on the input multidimensional time-series data, utilizing its expanded causal convolutional structure to capture long-term dependencies in the time dimension and extract deep temporal features. The deep temporal features extracted by the TCN layer are then input into a Squeeze-and-Excitation (SE) attention layer. This SE layer processes the channel dimension of the features, first obtaining global information for each channel through global average pooling (Squeeze), then learning the weights of each channel through two fully connected layers (Excitation), and finally multiplying the learned weights with the original features to dynamically enhance important feature channels and suppress secondary feature channels. The feature vectors weighted by the SE layer are then input into a Long Short-Term Memory (LSTM) network layer. Utilizing the gating mechanism and autoregressive properties of LSTM, iterative predictions are performed at multiple future time steps in a sequence-to-sequence manner, generating NO for several future time steps. xConcentration prediction sequence. The Temporal Convolutional Network (TCN) consists of multiple stacked residual blocks, each containing two layers of dilated causal convolutions, weight normalization, an activation function (ReLU), and a dropout layer. The dilation factor of the dilated convolutions increases exponentially with the depth of the network layers to increase the receptive field and capture long-term dependencies.

[0025] Once the data window is full, the preprocessed data (of shape [batch_size, seq_length, num_variables]) is fed into the core prediction model. In this embodiment, the model is exported in ONNX format (tcn_se_model.onnx) and used for efficient inference via the ONNX Runtime. The internal structure of the model is as follows: The TCN layer extracts temporal features: The input data first passes through a Temporal Convolutional Network (TCN). The TCN consists of multiple stacked Temporal Blocks, each containing two layers of dilated causal convolutions. Causal convolutions ensure that predictions do not depend on future information, while the exponentially increasing dilation rate gives the top-level network a very large receptive field, enabling it to capture long-term dependencies in the time series, which is crucial for understanding the inertia and delays in boiler combustion. Residual connections are used to prevent the gradient vanishing problem when the network is too deep. SE-Layer Channel Attention Weighting: The output feature map of the TCN (with shape [batch_size, num_channels, seq_length]) is fed into an SE attention layer (SELayer). The SE layer first performs global average pooling on each feature channel, compressing each channel into a single value, which represents the global information of that channel (Squeeze operation). Then, this global information descriptor learns the importance weights for each channel through two fully connected layers (excitation operation). The output dimension of the first fully connected layer is the number of input channels divided by the compression ratio, and the output dimension of the second fully connected layer is restored to the number of input channels. Finally, these weights are multiplied back onto the original feature channels, thereby dynamically enhancing the importance of NO. xThe prediction focuses on the representation of features that contribute significantly to the prediction (such as main steam temperature and total coal quantity) while suppressing interference from irrelevant or noisy features. The LSTM layer performs multi-step prediction: the attention-weighted feature vector is used, and its last output in the time dimension is taken as the initial state for subsequent predictions. This vector is then fed into an LSTM layer. Unlike simply using fully connected layers for single-step prediction, this invention utilizes the recurrent structure of the LSTM to generate predictions for multiple future time steps (e.g., 24 steps) in an autoregressive manner. At step t, the LSTM generates the predicted value and a new hidden state for step t based on the hidden state and output of step t-1 using a gating mechanism. This process iterates pred_length times, ultimately outputting a complete prediction sequence (with a shape of [batch_size, pred_length, output_size]). This design allows the model to better learn the temporal dynamics within the prediction sequence.

[0026] The aforementioned hybrid neural network model was obtained through training. Historical time-series data of nitrogen oxide-related operating parameters and corresponding actual nitrogen oxide concentrations were collected as training samples. A hybrid neural network model consisting of a temporal convolutional network, an SE attention layer, and a long short-term memory network was constructed. This hybrid neural network model was trained using the mean squared error loss function and the Adam optimizer. The trained hybrid neural network model was exported in ONNX format for real-time inference.

[0027] S3, the predicted nitrogen oxide concentration sequence is denormalized, and the predicted nitrogen oxide concentration at the corresponding time step is extracted to complete the nitrogen oxide concentration prediction.

[0028] After model inference is complete, the output sequence of predicted values ​​first undergoes denormalization to restore it to its original physical units. Then, depending on the configuration, the system may selectively extract predicted values ​​for specific future time points (e.g., steps 3, 6, and 12). These selected predicted values, along with the current NO... x The actual values ​​are encapsulated into API requests and written in batches to the registers (such as register_1, register_2, register_3) of the specified virtual device (such as virtual_device). In this way, downstream control systems or monitoring interfaces can easily read this forward-looking predictive information.

[0029] The entire process is encapsulated in a highly available RealTimePredictor class, which ensures smooth system operation by starting two independent threads (_data_collection_loop and _prediction_loop) for data collection and prediction. Meanwhile, all critical operations and data are recorded in detail in log files and CSV files for easy monitoring, debugging, and analysis.

[0030] like Figure 2 As shown, this embodiment provides a reactor inlet NO based on deep learning. x A multi-step real-time prediction system encapsulates functions such as data acquisition, model inference, and result write-back. It utilizes a multi-threaded mechanism to execute the data acquisition and prediction loops in parallel, ensuring the system's real-time performance and high availability. Specifically, it includes: The data acquisition module is used to acquire time-series data of nitrogen oxide-related operating parameters in the industrial control system, preprocess the time-series data to obtain preprocessed time-series data, and is equipped with an API client for connecting to the industrial control system database and periodically batch-fetching real-time operating data according to a preset list of variable points. A built-in data window manager is used to receive real-time data points and maintain a fixed-size, first-in-first-out (FIFO) data queue. When the window is full, it is formatted into the input tensors required by the model.

[0031] The data acquisition module includes the following functions: a TTP request signature mechanism, using the HMAC SHA256 algorithm to sign and authenticate API requests; a failure retry mechanism, automatically retrying when an API request fails, with configurable retries and latency; batch data reading functionality, allowing a single request to retrieve the latest values ​​of multiple operating parameters; a thread-safe data window manager, using a locking mechanism to ensure data consistency in a multi-threaded environment; a data integrity check function, ensuring that the window contains all necessary operating parameters; and a data normalization function, using a pre-saved scaler to standardize the data.

[0032] The nitrogen oxide concentration prediction module is used to construct a hybrid neural network model. This model extracts deep temporal features from the pre-processed time-series data using a pre-trained hybrid neural network model. These deep temporal features are dynamically weighted to obtain a weighted feature vector. Iterative prediction is then performed based on this weighted feature vector to obtain a sequence of predicted nitrogen oxide concentration values ​​for several future time steps. A pre-trained deep learning model stored in ONNX (Open Neural Network Exchange) format is loaded. The core architecture of this model is a combination of TCN, SE attention layer, and LSTM. This module receives tensors output from the data processing module, performs model inference, and outputs multi-step prediction results.

[0033] The nitrogen oxide concentration prediction module includes the following functions: an ONNX runtime inference engine that supports efficient model inference; a dimensionality conversion function for model input and output to ensure that the data format matches the model requirements; and a denormalization function for prediction results to restore the predicted values ​​to actual physical units.

[0034] The result output module is used to inversely normalize the predicted nitrogen oxide concentration sequence and extract the predicted nitrogen oxide concentration values ​​at the corresponding time steps to complete the nitrogen oxide concentration prediction. The prediction sequence output by the nitrogen oxide concentration prediction module is parsed, and the predicted values ​​for specific prediction step sizes (such as the 3rd, 6th, and 12th future steps) are selected according to the configuration. These values ​​are then written to the specified virtual device or control system register via the API client.

[0035] The results output module includes the following functions: prediction step size selection function, which can be configured to output the predicted value at a specific time step; API write-back function, which supports writing the prediction results to multiple registers of the virtual device; and error handling mechanism, which records errors and retryes when API write fails.

[0036] The entire reactor inlet nitrogen oxide concentration prediction system also includes the following functions: a multi-threaded execution framework, in which data acquisition loops and prediction loops are executed in parallel in independent threads; a logging system, including a main log, data recording logs, and prediction result logs; and a CSV file output function, which saves real-time data and prediction results to CSV files for subsequent analysis.

[0037] This embodiment has the following advantages: First, it has high prediction accuracy. It combines the powerful long-term dependency capture capability of TCN with the dynamic focusing capability of SE-Layer on key features, enabling a deeper understanding of complex working conditions and NO. x First, the nonlinear relationship between concentrations is identified, thus improving prediction accuracy. Second, deep features extracted through TCN-SE are combined with LSTM for sequence-to-sequence generation, achieving end-to-end multi-step prediction while maintaining high accuracy over long periods, providing ample response time for the control system. Third, the overall method is encapsulated in a highly available integrated script, utilizing multi-threaded parallel processing for data acquisition and model prediction, and employing the efficient ONNX runtime for model inference, ensuring low latency throughout the prediction process and meeting the needs of industrial real-time control. Fourth, the model structure is independent of specific latency assumptions and can adaptively learn inherent time dependencies from the data. This method is easy to deploy and migrate to different industrial scenarios, guiding ammonia injection control in SCR systems.

[0038] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0039] This embodiment also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program (in this embodiment, the computer program includes a computing component and an iterative component, capable of model calculation and model updating). The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to realize the corresponding method flow or corresponding function. The processor described in this embodiment can be used in a deep learning-based reactor inlet NO. x Operation of multi-step real-time prediction methods.

[0040] This embodiment also provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement a deep learning-based reactor inlet NO in the above embodiment. x The corresponding steps of a multi-step real-time prediction method.

[0041] This embodiment also provides a computer program product, which includes a computer program that, when executed by a processor, implements a deep learning-based reactor inlet NO as described in the above embodiment. x The corresponding steps of a multi-step real-time prediction method.

[0042] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0043] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A deep learning based reactor inlet NO x Multi-step real-time prediction method characterized in that, The method comprises the following steps: obtaining time series data of nitrogen oxide related working condition parameters in an industrial control system, preprocessing the time series data to obtain preprocessed time series data; constructing a hybrid neural network model, extracting deep time series features of the preprocessed time series data using the pre-trained hybrid neural network model, dynamically weighting the deep time series features to obtain a weighted feature vector, iteratively predicting according to the weighted feature vector, and obtaining a nitrogen oxide concentration prediction value sequence of future time steps; de-normalizing the nitrogen oxide concentration prediction value sequence and extracting the nitrogen oxide concentration prediction value of the corresponding time step to complete the nitrogen oxide concentration prediction.

2. The reactor inlet NO based on deep learning according to claim 1 x The multi-step real-time prediction method is characterized by, The preprocessing of the time series data specifically comprises: after data cleaning of the time series data, performing normalization processing by using a MinMaxScaler scaler to obtain preprocessed time series data.

3. A reactor inlet NO based on deep learning according to claim 1 x The multi-step real-time prediction method is characterized by, The hybrid neural network model is composed of a time convolution network, an SE attention layer and a long short-term memory network connected in sequence; The time convolution network is used to extract deep time series features of the preprocessed time series data. The SE attention layer is used to dynamically weight the deep time series features to obtain a weighted feature vector. The long short-term memory network is used to receive the weighted feature vector for iterative prediction.

4. A deep learning based reactor inlet NOx concentration prediction system according to claim 3, wherein the neural network is a convolutional neural network. x A multi-step real-time prediction method, characterized in that, The SE attention layer dynamically weights the deep time series features, specifically comprising: after global average pooling of each channel of the deep time series features, learning the weight of each channel through at least two fully connected layers and combining the deep time series features to obtain a weighted feature vector.

5. A deep learning based reactor inlet NOx concentration prediction system according to claim 3, wherein the neural network is a convolutional neural network. x A multi-step real-time prediction method, characterized in that, The time convolution network is composed of a plurality of residual blocks stacked together, and each residual block includes a dilated causal convolution layer.

6. A deep learning based reactor inlet NOx concentration prediction system according to claim 3, wherein the neural network is a convolutional neural network. x A multi-step real-time prediction method, characterized in that, The pre-trained hybrid neural network model is trained by the following steps: obtaining time series data of historical nitrogen oxide related working condition parameters and corresponding actual nitrogen oxide concentration values as training samples; using mean square error as a loss function, and training the hybrid neural network model through an Adam optimizer to obtain a pre-trained hybrid neural network model.

7. A deep learning based reactor inlet NOx x A multi-step real-time prediction system characterized by, It comprises: a data acquisition module for obtaining time series data of nitrogen oxide related working condition parameters in an industrial control system, preprocessing the time series data to obtain preprocessed time series data; a nitrogen oxide concentration prediction module for constructing a hybrid neural network model, extracting deep time series features of the preprocessed time series data using the pre-trained hybrid neural network model, dynamically weighting the deep time series features to obtain a weighted feature vector, iteratively predicting according to the weighted feature vector, and obtaining a nitrogen oxide concentration prediction value sequence of future time steps; a result output module for de-normalizing the nitrogen oxide concentration prediction value sequence and extracting the nitrogen oxide concentration prediction value of the corresponding time step to complete the nitrogen oxide concentration prediction.

8. An electronic device, comprising: A computer program product comprising a memory, a processor, and a computer program stored in the memory and loadable into the processor, the processor implementing the method according to any one of claims 1 to 6 when executing the computer program. x Steps of the multi-step real-time prediction method.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, the computer program is executed by the processor to realize the reactor inlet NOx concentration prediction method based on deep learning in any one of claims 1-6 x Steps of the multi-step real-time prediction method.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements a deep learning based reactor inlet NOx x Steps of the multi-step real-time prediction method.

Images