Intelligent denitration control system based on multi-model fusion and self-learning

Abstract

The invention provides an intelligent denitration control system based on multi-model fusion and self-learning, and belongs to the technical field of industrial denitration control, and the method comprises the following steps: collecting three-dimensional data of a boiler side, a combustion side and a denitration side in a DCS / SIS system in real time through an OPC protocol; s2, an intelligent data preprocessing step, wherein the step comprises the links of anomaly detection, data restoration and dynamic normalization; s3, a spatio-temporal feature engineering step, wherein the step comprises a time delay analysis link, a feature derivation link and a feature screening link; s4, a multi-model dynamic fusion step, wherein the step comprises a base model group and a fusion mechanism link; and S5, an online self-learning optimization step, the step comprises incremental learning, parameter evolution and knowledge distillation links, and aiming at the problem that algorithm optimization adjustment cannot be carried out aiming at the environmental change of the denitration process, self-adaptive evolution of control parameters is realized by constructing a'prediction-control-feedback 'complete chain.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial denitration control, and in particular to an intelligent denitration control system based on multi-model fusion and self-learning. Background Art

[0002] Traditional single-model prediction methods suffer from poor scenario adaptability and rigid parameters. With the development of deep learning technology, multi-model fusion strategies can significantly improve prediction accuracy by integrating the strengths of different algorithms. Online self-learning mechanisms can dynamically adapt to changing operating conditions, forming closed-loop optimization control. Existing technologies have not yet effectively combined dynamic model weighted fusion with real-time parameter self-correction technology, resulting in prediction lag and slow control response under complex operating conditions such as deep peak shaving.

[0003] Chinese patent publication number CN213668683U discloses an intelligent, efficient, and precise denitrification control system, which solves the problem of the ambient temperature outside the ammonia storage tank being inconsistent and having large temperature fluctuations.

[0004] However, the above-mentioned publicly disclosed solution has the following shortcomings: in actual use, the above-mentioned solution is unable to optimize and adjust the algorithm according to the environmental changes in the denitrification process.

[0005] The present invention proposes an intelligent denitrification control system based on multi-model fusion and self-learning to solve the problem. Summary of the Invention

[0006] The purpose of the present invention is to solve the hysteresis problem of static feature selection in traditional methods through joint analysis of spatiotemporal features, thereby overcoming the problems in the above-mentioned background technology.

[0007] Based on the above technical ideas, the technical solution adopted by the present invention is:

[0008] An intelligent denitrification control system based on multi-model fusion and self-learning includes the following steps:

[0009] S1 multi-source heterogeneous data acquisition step, which uses the OPC protocol to collect three-dimensional data from the boiler side, combustion side, and denitrification side of the DCS / SIS system in real time;

[0010] S2 intelligent data preprocessing step, which includes anomaly detection, data repair and dynamic normalization;

[0011] S3 spatiotemporal feature engineering step, which includes time lag analysis, feature derivation and feature screening;

[0012] S4 multi-model dynamic fusion step, which includes the base model group and fusion mechanism link;

[0013] S5 is an online self-learning optimization step, which includes incremental learning, parameter evolution, and knowledge distillation.

[0014] Further limitation of the above technical solution: the S1 multi-source heterogeneous data collection step collects the following multi-dimensional data in real time through the industrial Internet of Things gateway using the OPC UA protocol, with a sampling frequency of not less than 500ms.

[0015] Further limitation of the above technical solution is that the S1 multi-source heterogeneous data acquisition step includes the dynamic parameter analysis of the combustion process and the time series data analysis of the flue gas composition. The dynamic parameter analysis of the combustion process includes the coal powder concentration at the outlet of each pulverizer (0.5-1.2kg / m 3 ), secondary air door opening (0-100%), burner swing angle (-30°~+30°); flue gas composition time series data analysis includes SCR reactor inlet / outlet NOx concentration (50-600mg / Nm 3 ), oxygen content (3-8% vol), ammonia escape rate (0-5ppm); this step also includes equipment status parameter monitoring, equipment status parameter monitoring includes catalyst layer pressure difference (200-800Pa), ammonia injection grid valve opening (0-100%), soot blower working cycle (10-60min).

[0016] Further limitation of the above technical solution: the anomaly detection link in the S2 intelligent data preprocessing step includes using an improved isolation forest algorithm, setting the number of subtrees N = 200, the anomaly score threshold θ = 0.65, and performing joint outlier detection on the multidimensional data stream; the data repair link in the S2 intelligent data preprocessing step includes dynamically interpolating and constructing a two-layer LSTM-Attention network, wherein: the encoder layer: contains 128 LSTM units, the time step T = 60; the attention layer: uses the scaled dot product attention mechanism to calculate the feature weights; the decoder layer: outputs the reconstruction value of the missing data, and the reconstruction error is <2.5%. This step also includes an adaptive normalization link, which includes establishing a working condition perception module. When the load change rate is >2% / min, sliding window normalization (window width W = 300s) is enabled.

[0017] Further limiting the above technical solution, the S3 spatiotemporal feature engineering step, in which the time lag analysis step includes using the maximum information coefficient (MIC) and improved dynamic time warping (DTW) joint analysis method to determine the optimal time lag τ∈[30s,300s] between each variable and NOx concentration; the feature derivation step includes generating a combustion efficiency factor, Where ΔO2 is the difference in oxygen content between the inlet and outlet of the preheater, Q coal is the amount of coal fed into the furnace, T furnaceis the furnace temperature; it also includes the dynamic pulverized coal residence time,

[0018] Further limiting the above technical solution, the S4 multi-model dynamic fusion step, in which the base model group includes a spatiotemporal convolutional network, a physics-guided neural network and online sequence extreme learning, the spatiotemporal convolutional network includes a time branch: 4 layers of dilated causal convolution, with a dilation coefficient d = [1, 2, 4, 8]; a spatial branch: the number of graph convolution layers K = 3, and the adjacency matrix is ​​constructed based on the device topology relationship; the physics-guided neural network includes embedding the Arrhenius equation: And constraint layer: the output value must satisfy The online sequence extreme learning includes the number of hidden layer nodes H = 150 and the regularization parameter λ = 1e-4.

[0019] Further limitation of the above technical solution is that the S4 multi-model dynamic fusion step includes a fusion mechanism link including weight calculation, Where β = 0.8, RMSE i The prediction error of each model in the past 30 minutes, including the fusion output, Where γ = 0.2 is the physical model compensation coefficient.

[0020] Further limitation of the above technical solution is that the S5 online self-learning optimization step, the incremental learning link in this step includes establishing a circular data buffer (capacity C = 10,000 samples), triggering the model update when any of the following conditions is detected: operating condition fluctuation index > 0.7, and prediction error of 10 consecutive samples > 5%.

[0021] Further limitations on the above technical solution include the S1 multi-source heterogeneous data collection step, which includes establishing a three-dimensional data cube. The time axis stores the second-level data of the last 72 hours, using time slice block storage (30 minutes per block). The spatial axis includes the temperature field distribution of 32 sections of the boiler (resolution 0.5m×0.5m). The equipment axis records equipment health status parameters such as the pulverizer vibration spectrum (0-2000Hz) and the fan bearing temperature (0-150°C).

[0022] To further limit the above technical solution, the S2 intelligent data preprocessing step includes an improved isolation forest. The improved isolation forest includes a dynamic subsampling strategy: when the data dimension D>20, feature group sampling (5-8 relevant features per group) is adopted, and an incremental update mechanism: the tree structure is updated every time 1000 new samples are received, and the update time is <500ms.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. Dynamic feature engine, which solves the lag problem of static feature selection in traditional methods through joint analysis of spatiotemporal features;

[0025] 2. Intelligent fusion mechanism, using online weight adjustment based on prediction error backpropagation to improve adaptability to working conditions;

[0026] 3. A closed-loop learning system builds a complete chain of "prediction-control-feedback" to achieve adaptive evolution of control parameters;

[0027] 4. Fusion of virtual and real, combining data-driven models with physical reaction equations to enhance the interpretability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 The figure is a flow chart of an intelligent denitrification control system based on multi-model fusion and self-learning according to the present invention. DETAILED DESCRIPTION

[0030] The following is combined with Figure 1 The present invention is described in further detail.

[0031] Example 1: This example provides an intelligent denitrification control system based on multi-model fusion and self-learning, such as Figure 1 As shown, the following steps are included:

[0032] S1 multi-source heterogeneous data acquisition step, which uses the OPC protocol to collect three-dimensional data from the boiler side, combustion side, and denitrification side of the DCS / SIS system in real time;

[0033] S2 intelligent data preprocessing step, which includes anomaly detection, data repair and dynamic normalization;

[0034] S3 spatiotemporal feature engineering step, which includes time lag analysis, feature derivation and feature screening;

[0035] S4 multi-model dynamic fusion step, which includes the base model group and fusion mechanism link;

[0036] S5 is an online self-learning optimization step, which includes incremental learning, parameter evolution, and knowledge distillation.

[0037] The S1 multi-source heterogeneous data collection step collects the following multi-dimensional data in real time through the industrial Internet of Things gateway using the OPC UA protocol, with a sampling frequency of no less than 500ms.

[0038] The S1 multi-source heterogeneous data acquisition step includes the dynamic parameter analysis of the combustion process and the time series data analysis of the flue gas composition. The dynamic parameter analysis of the combustion process includes the coal powder concentration at the outlet of each pulverizer (0.5-1.2 kg / m 3 ), secondary air door opening (0-100%), burner swing angle (-30°~+30°); flue gas composition time series data analysis includes SCR reactor inlet / outlet NOx concentration (50-600mg / Nm 3 ), oxygen content (3-8% vol), ammonia escape rate (0-5ppm); this step also includes equipment status parameter monitoring, equipment status parameter monitoring includes catalyst layer pressure difference (200-800Pa), ammonia injection grid valve opening (0-100%), soot blower working cycle (10-60min).

[0039] The anomaly detection link in the S2 intelligent data preprocessing step includes using an improved isolation forest algorithm, setting the number of subtrees N=200, the anomaly score threshold θ=0.65, and performing joint outlier detection on the multidimensional data stream. The data repair link in the S2 intelligent data preprocessing step includes dynamically interpolating and constructing a two-layer LSTM-Attention network, wherein: the encoder layer: contains 128 LSTM units, the time step T=60; the attention layer: uses the scaled dot product attention mechanism to calculate the feature weights; the decoder layer: outputs the reconstruction value of the missing data, and the reconstruction error is <2.5%. This step also includes an adaptive normalization link, which includes establishing a working condition perception module. When the load change rate is >2% / min, sliding window normalization (window width W=300s) is enabled.

[0040] S3 spatiotemporal feature engineering step, in which the time lag analysis step includes the use of the maximum information coefficient (MIC) and improved dynamic time warping (DTW) joint analysis method to determine the optimal time lag τ∈[30s,300s] between each variable and NOx concentration; the feature derivation step includes generating the combustion efficiency factor, Where ΔO2 is the difference in oxygen content between the inlet and outlet of the preheater, Q coal is the amount of coal fed into the furnace, T furnace is the furnace temperature; it also includes the dynamic pulverized coal residence time,

[0041] The S4 multi-model dynamic fusion step includes a base model group including a spatiotemporal convolutional network, a physics-guided neural network, and online sequence extreme learning. The spatiotemporal convolutional network includes a time branch: 4 layers of dilated causal convolution, with a dilation coefficient d = [1, 2, 4, 8]; a spatial branch: the number of graph convolution layers K = 3, and the adjacency matrix is ​​constructed based on the device topology relationship; the physics-guided neural network includes an embedded Arrhenius equation: And constraint layer: the output value must satisfy The online sequence extreme learning includes the number of hidden layer nodes H = 150 and the regularization parameter λ = 1e-4.

[0042] The S4 multi-model dynamic fusion step includes the fusion mechanism link including weight calculation, Where β = 0.8, RMSE i The prediction error of each model in the past 30 minutes, including the fusion output, Where γ = 0.2 is the physical model compensation coefficient.

[0043] The S5 online self-learning optimization step includes an incremental learning link that establishes a circular data buffer (capacity C = 10,000 samples). When any of the following conditions is detected, a model update is triggered: the operating condition fluctuation index is greater than 0.7, and the prediction error of 10 consecutive samples is greater than 5%.

[0044] S1 multi-source heterogeneous data collection step, which includes the establishment of a three-dimensional data cube. The time axis stores the second-level data of the last 72 hours, using time slices for block storage (30 minutes per block). The spatial axis includes the temperature field distribution of 32 sections of the boiler (resolution 0.5m×0.5m). The equipment axis records equipment health status parameters such as the pulverizer vibration spectrum (0-2000Hz) and the fan bearing temperature (0-150°C).

[0045] S2 intelligent data preprocessing step, which includes an improved isolation forest. The improved isolation forest includes a dynamic subsampling strategy: when the data dimension D>20, feature group sampling (5-8 relevant features per group) is adopted, and an incremental update mechanism: the tree structure is updated every time 1000 new samples are received, and the update time is <500ms.

[0046] Example 2: This example provides an intelligent denitrification control system based on multi-model fusion and self-learning, such as Figure 1 As shown, typical working cases are also included:

[0047] Case 1: 75% load rapid lifting test

[0048] Control effect:

[0049] NOx concentration fluctuation range: 112-128 mg / Nm3 (Set value 120)

[0050] Ammonia injection valve operation times: 6 times / hour (28 times for traditional system)

[0051] Ammonia consumption: 78kg / h (92kg / h for traditional system)

[0052] Case 2: Coal calorific value mutation test

[0053] Event: Received low calorific value coal (calorific value decreased by 15%)

[0054] System Response:

[0055] Detects a decrease in combustion efficiency within 8 seconds

[0056] Trigger feature library update (5 new derivative features)

[0057] The model weights are automatically adjusted to: STCN (0.38) → PGNN (0.41)

[0058] Control stability is maintained: NOx fluctuation <±5%.

[0059] Example 3: This example provides an intelligent denitrification control system based on multi-model fusion and self-learning, such as Figure 1 As shown, this system is also deployed in a 1000MW ultra-supercritical coal-fired unit. The hardware configuration is as follows:

[0060] Data collection layer:

[0061] Edge computing gateway: Huawei Atlas 500 intelligent edge server (quad-core ARM CPU + 8GB memory);

[0062] Sensor network: 96 high-precision sensors (K-type thermocouples, Rosemount oxygen analyzers, Endress+Hauser pressure transmitters);

[0063] Communication protocol: OPC UA over TSN, transmission cycle 500ms, jitter <5ms.

[0064] Computing Platform:

[0065] Training server: NVIDIA DGX A100 (8×A100 GPUs, 320GB video memory);

[0066] Inference engine: TensorRT 8.4 optimized model, deployed on an industrial PC (Intel i9-12900K + RTX 4090).

[0067] Control execution layer:

[0068] Ammonia injection actuator: Siemens SIPART PS2 intelligent positioner (positioning accuracy 0.1%);

[0069] Emergency interlock: Triple redundant PLC system (Schneider Modicon M580).

[0070] The above contents are further detailed descriptions of the present invention in conjunction with specific preferred embodiments, so as to facilitate those skilled in the art to understand and apply the present invention. It should not be considered that the specific implementation of the present invention is limited to these descriptions.

Claims

1. An intelligent denitrification control system based on multi-model fusion and self-learning, characterized in that: The following steps are involved: S1 multi-source heterogeneous data acquisition step, which uses the OPC protocol to collect three-dimensional data from the boiler side, combustion side, and denitrification side of the DCS / SIS system in real time; S2 intelligent data preprocessing step, which includes anomaly detection, data repair and dynamic normalization; S3 spatiotemporal feature engineering step, which includes time lag analysis, feature derivation, and feature screening; S4 multi-model dynamic fusion step, which includes the base model group and fusion mechanism link; S5 is an online self-learning optimization step, which includes incremental learning, parameter evolution, and knowledge distillation.

2. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 1 is characterized in that: The S1 multi-source heterogeneous data collection step collects the following multi-dimensional data in real time through the industrial Internet of Things gateway using the OPC UA protocol, with a sampling frequency of no less than 500ms.

3. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 2 is characterized in that: The S1 multi-source heterogeneous data acquisition step includes the dynamic parameter analysis of the combustion process and the time series data analysis of the flue gas composition. The dynamic parameter analysis of the combustion process includes the coal powder concentration at the outlet of each pulverizer (0.5-1.2 kg / m 3 ), secondary air door opening (0-100%), burner swing angle (-30°~+30°); flue gas composition time series data analysis includes SCR reactor inlet / outlet NOx concentration (50-600mg / Nm 3 ), oxygen content (3-8% vol), ammonia escape rate (0-5ppm); this step also includes equipment status parameter monitoring, equipment status parameter monitoring includes catalyst layer pressure difference (200-800Pa), ammonia injection grid valve opening (0-100%), soot blower working cycle (10-60min).

4. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 3 is characterized in that: The anomaly detection link in the S2 intelligent data preprocessing step includes using an improved isolation forest algorithm, setting the number of subtrees N=200, the anomaly score threshold θ=0.65, and performing joint outlier detection on the multidimensional data stream. The data repair link in the S2 intelligent data preprocessing step includes dynamically interpolating and constructing a two-layer LSTM-Attention network, wherein: the encoder layer: contains 128 LSTM units, the time step T=60; the attention layer: uses the scaled dot product attention mechanism to calculate the feature weights; the decoder layer: outputs the reconstruction value of the missing data, and the reconstruction error is <2.5%. This step also includes an adaptive normalization link, which includes establishing a working condition perception module. When the load change rate is >2% / min, sliding window normalization (window width W=300s) is enabled.

5. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 4 is characterized in that: S3 spatiotemporal feature engineering step, in which the time lag analysis step includes the use of the maximum information coefficient (MIC) and improved dynamic time warping (DTW) joint analysis method to determine the optimal time lag τ∈[30s,300s] between each variable and NOx concentration; the feature derivation step includes generating the combustion efficiency factor, Where ΔO2 is the difference in oxygen content between the inlet and outlet of the preheater, Q coal is the amount of coal fed into the furnace, T furnace is the furnace temperature; it also includes the dynamic pulverized coal residence time, 6. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 5 is characterized in that: The S4 multi-model dynamic fusion step, in which the base model group includes a spatiotemporal convolutional network, a physical guided neural network and an online sequence extreme learning, and the spatiotemporal convolutional network includes a time branch: 4 layers of dilated causal convolution, with a dilation coefficient d = [1, 2, 4, 8]; a spatial Branch: The number of graph convolution layers is K = 3, and the adjacency matrix is ​​constructed based on the device topology; the physical guidance neural network includes embedding the Arrhenius equation: And constraint layer: the output value must satisfy The online sequence extreme learning includes the number of hidden layer nodes H = 150 and the regularization parameter λ = 1e-4.

7. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 6 is characterized in that: The S4 multi-model dynamic fusion step includes the fusion mechanism link including weight calculation, Where β = 0.8, RMSE i The prediction error of each model in the past 30 minutes, including the fusion output, Where γ = 0.2 is the physical model compensation coefficient.

8. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 7 is characterized in that: The S5 online self-learning optimization step includes an incremental learning link that establishes a circular data buffer (capacity C = 10,000 samples). When any of the following conditions is detected, a model update is triggered: the operating condition fluctuation index is greater than 0.7, and the prediction error of 10 consecutive samples is greater than 5%.

9. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 8 is characterized in that: S1 multi-source heterogeneous data collection step, which includes the establishment of a three-dimensional data cube. The time axis stores the second-level data of the last 72 hours, using time slices for block storage (30 minutes per block). The spatial axis includes the temperature field distribution of 32 sections of the boiler (resolution 0.5m×0.5m). The equipment axis records equipment health status parameters such as the pulverizer vibration spectrum (0-2000Hz) and the fan bearing temperature (0-150°C).

10. The intelligent denitrification control system based on multi-model fusion and self-learning according to claim 9 is characterized in that: S2 intelligent data preprocessing step, which includes an improved isolation forest. The improved isolation forest includes a dynamic subsampling strategy: when the data dimension D>20, feature group sampling (5-8 relevant features per group) is adopted, and an incremental update mechanism: the tree structure is updated every time 1000 new samples are received, and the update time is <500ms.

Images