A steam generator inlet pressure prediction system and method based on deep learning

Abstract

The application discloses a kind of steam generator inlet pressure prediction system and method based on deep learning, belong to steam generator operating parameter prediction technical field.The system includes distributed industrial sensor network, data acquisition module, industrial server and visual display system.The method acquires multidimensional steam data, reconstructs past 40 minutes input, future 20 minutes output time series sample after pre-processing;Pressure signal is decomposed into multiple periodic scale sub-sequences;One-dimensional time series data is folded into two-dimensional tensor, combined with dynamic delay alignment and two-dimensional convolutional neural network to extract time-period phase correlation characteristics;Through multi-scale branch prediction, attention weighted fusion and physical rule correction, the pressure prediction curve of future 20 minutes is output;Based on the prediction result, multi-objective optimization control strategy is generated and visualized.The application can predict pressure change trend in advance, with high prediction accuracy, to ensure the stable operation of steam power generation system.

Description

Technical Field

[0001] This invention belongs to the field of steam generator operating parameter prediction technology, and in particular relates to a steam generator inlet pressure prediction system and method based on deep learning. Background Technology

[0002] In the steel production process, sintering, steelmaking, and rolling processes generate a large amount of waste heat, which is converted into waste heat steam through waste heat recovery devices and then transported to a steam generator for power generation. Specifically, the RH vacuum circulating degassing furnace in the steelmaking process consumes some steam intermittently and irregularly during production. This consumption significantly affects the pressure balance of the steam pipeline network. Under normal operating conditions, the steam generator inlet pressure can be stable at 0.8-0.9 MPa; however, when the RH furnace starts up and consumes steam, the generator inlet pressure will fluctuate drastically within approximately 5-10 minutes, with a range of ±0.3 MPa. In severe cases, this can lead to generator load fluctuations, decreased efficiency, or even protective shutdowns.

[0003] Current technologies mainly rely on manual monitoring and adjustment, and traditional PID automatic control. Manual monitoring involves low data acquisition frequency, typically once per minute, with a response lag of 3-5 minutes. PID control is based solely on current pressure feedback and cannot predict future trends. Time series forecasting techniques introduced in recent years suffer from problems such as limited data dimensionality, poor model adaptability, and inability to handle nonlinear multivariate coupling, resulting in prediction accuracy typically below 75%. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies, such as slow response, inability to predict in advance, and reliance on human experience. It provides a deep learning-based steam generator inlet pressure prediction system and method, which can achieve high-precision prediction of pressure changes in the next 20 minutes, providing operators with a basis for advance adjustment and ensuring the stable and efficient operation of the steam power generation system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based steam generator inlet pressure prediction system includes: a distributed industrial sensor network 1, a data acquisition module 2, an industrial server 3, and a visualization system 4; the data acquisition module 2 is connected to the distributed industrial sensor network 1, the industrial server 3 is connected to the data acquisition module 2, and the visualization system 4 is connected to the industrial server 3.

[0006] Furthermore, the distributed industrial sensor network 1 includes: a first type of sensor group deployed at the outlet of the waste heat boiler in the sintering, steelmaking, and rolling processes to collect steam output and pressure; a second type of sensor group deployed in the steam consumption pipeline of the RH boiler to collect the real-time steam consumption and operating status of the RH boiler; a third type of sensor group deployed at the inlet of the steam generator to collect the inlet pressure and flow of the steam generator; and a fourth type of sensor group deployed on the valves of the steam pipeline network to collect the current valve opening and adjustment stroke data; wherein, the sampling frequency of all sensors is 1 second / sample, and the measurement accuracy of the pressure sensor is at least ±0.001MPa.

[0007] Furthermore, the industrial server 3 integrates a data preprocessing module, which includes: a noise suppression unit, a missing value compensation unit, an outlier removal unit, a data standardization unit, and a time series reconstruction unit. The noise suppression unit uses a wavelet threshold denoising algorithm to filter out high-frequency noise in the data; the missing value compensation unit uses the KNN-Interp algorithm to complete missing values; the outlier removal unit integrates the 3σ principle and industrial rules to identify and remove outliers; the data standardization unit uses Z-Score standardization to eliminate dimensional differences; and the time series reconstruction unit reconstructs the time series data with an input window of 40 minutes, an output window of 20 minutes, and a sliding step of 1 second.

[0008] Furthermore, the industrial server 3 integrates a deep learning prediction model, which includes: a temporal feature decomposition module, a two-dimensional transformation module, a multi-scale subsequence prediction module, and a result fusion module. The temporal feature decomposition module uses an adaptive variational mode decomposition algorithm to decompose the pressure time series signal into 3-5 intrinsic mode functions, corresponding to short-term fluctuation periods, medium-term fluctuation periods, and long-term trend periods, respectively. The two-dimensional transformation module reconstructs the one-dimensional temporal subsequence of each intrinsic mode function into a two-dimensional tensor with its period length as the folding window, performs dynamic delay alignment processing, and then extracts features from the two-dimensional tensor through a 2D convolutional neural network. The multi-scale subsequence prediction module has differentiated prediction branches for sequences with different periods, including: a lightweight neural network branch for processing short-term fluctuation periods, a temporal convolutional network branch for processing medium-term fluctuation periods, and an attention-enhanced Transformer encoder branch for processing long-term trend periods. The result fusion module uses an attention-weighted fusion strategy to assign dynamic weights to the predicted subsequences of each branch and introduces a physical rule correction module based on the mass conservation law of the steam system to constrain and correct the fusion results.

[0009] Furthermore, the industrial server 3 also integrates a multi-objective optimization strategy module. The multi-objective optimization strategy module constructs the following optimization function: the main objective is to keep the steam generator inlet pressure stable within the design range and the deviation rate does not exceed ±2%; the constraint objective is to ensure normal production of the RH furnace and that the steam output of each process is not lower than the production load requirement; and the auxiliary objectives are to minimize the number of valve adjustments and minimize the steam transmission energy consumption.

[0010] Furthermore, the visualization system 4 includes: a high-frequency data monitoring module, a prediction trend module, a simulation display area, and a historical prediction comparison area; the high-frequency data monitoring module displays real-time steam data for each process updated every second in the form of dynamic curves; the prediction trend module displays the pressure prediction curve for the next 20 minutes in the form of a line graph, and marks the design range and warning threshold; the simulation display area simulates the pipeline pressure distribution and provides reference data for the pressure prediction at the next moment; the historical prediction comparison area displays a line graph comparing historical prediction results with historical actual values ​​and calculates and displays the prediction accuracy.

[0011] Furthermore, the industrial server 3 also includes a model iteration update module. The model iteration update module calculates the deviation between the predicted pressure and the actual pressure in real time. When the deviation rate exceeds 10% for several consecutive times, incremental training is triggered. Incremental training collects the latest 24 hours of data, freezes the parameters of the time series feature decomposition module and the two-dimensional transformation module, and only updates the weights of the multi-scale subsequence prediction module and the result fusion module.

[0012] Furthermore, the deep learning prediction model is trained using a composite loss function, which is the mean squared error plus a trend penalty term. The trend penalty term is determined based on whether the trend of the predicted value is consistent with the trend of the true value.

[0013] This invention also provides a deep learning-based method for predicting the inlet pressure of a steam generator, comprising the following steps: S1: Data acquisition module 2 collects multi-dimensional steam data from the steam generation end, consumption end and power generation end in real time through distributed industrial sensor network 1 at a frequency of 1 second / sample; S2: The collected multi-dimensional steam data is preprocessed by the data preprocessing module and reconstructed into a standardized time series sample with the past 40 minutes as the input window and the next 20 minutes as the output window; S3: The time-series feature decomposition module of the deep learning prediction model decomposes the standardized stress time-series signal into multiple subsequences with different periodic scales. S4: The one-dimensional time series data of each subsequence is converted into a two-dimensional tensor and features are extracted through the two-dimensional transformation module of the deep learning prediction model; S5: The multi-scale subsequence prediction module of the deep learning prediction model predicts subsequences at different periodic scales to obtain multiple predicted subsequences. S6: Through the result fusion module of the deep learning prediction model, multiple prediction sub-sequences are fused and corrected by physical rules to output the prediction curve of the steam generator inlet pressure for the next 20 minutes; S7: Through the multi-objective optimization strategy module, a control strategy including pressure stabilization, production constraints and energy consumption optimization is generated based on the pressure prediction curve; S8: The pressure prediction curve and the control strategy are displayed in real time through the visualization display system 4.

[0014] Compared with traditional solutions, the present invention has the following advantages: (1) This invention can output the future pressure change trend in advance through high-frequency data acquisition and deep learning model, so as to give operators enough time to adjust and realize the transformation from passive response to active pre-adjustment.

[0015] (2) The present invention employs mechanisms such as time series decomposition, two-dimensional transformation, multi-scale branch prediction and physical correction to effectively capture the multi-period coupling and nonlinear dynamic delay characteristics of steam pressure, and maintain high-precision prediction in multi-variable and strongly coupled industrial scenarios.

[0016] (3) The present invention has a built-in model iteration and update mechanism. When the working conditions change, only a small amount of the latest data is needed to quickly fine-tune the model parameters without retraining. It is stable and reliable in long-term operation.

[0017] (4) The present invention can generate an optimized control strategy based on the prediction results, while ensuring stable pressure, continuous production, reducing the number of valve adjustments and reducing energy consumption, thereby improving the overall operating efficiency of the system.

[0018] (5) The visualization display system of the present invention can intuitively display real-time data, prediction curves and optimization suggestions, assisting operators to make quick and accurate decisions and reduce reliance on human experience.

[0019] (6) The core architecture of this invention can be migrated to various industrial time series prediction scenarios such as chemical, power, and new energy, supports lightweight deployment on the edge side, and has good prospects for promotion and application. Attached Figure Description

[0020] This manual includes the following figures, which illustrate the following: Figure 1 This is a schematic diagram of the hardware connection structure of the system of the present invention.

[0021] Figure 2 This is a schematic diagram of the interface layout of the visualization display system of the present invention.

[0022] Figure 3 This is a flowchart illustrating the method of the present invention.

[0023] The components include: 1. Distributed industrial sensor network; 2. Data acquisition module; 3. Industrial server; and 4. Visualization display system. Detailed Implementation

[0024] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, in order to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention, and to facilitate its implementation.

[0025] like Figure 1 As shown, this invention provides a deep learning-based steam generator inlet pressure prediction system, comprising a distributed industrial sensor network 1, a data acquisition module 2, an industrial server 3, and a visualization display system 4.

[0026] The distributed industrial sensor network 1 is deployed at key nodes of the steam system, including the steam generation end, steam consumption end, and steam transmission and power generation end. These sensors are connected to the data acquisition module 2 via an industrial fieldbus. The data acquisition module 2 uses a programmable logic controller (PLC) to receive raw electrical or digital signals from each sensor, perform analog-to-digital conversion, signal filtering, timestamp marking, and data buffering, and then uploads the packaged data to the industrial server 3 via an industrial Ethernet at a frequency of 1 second. The industrial server 3 is a computer equipped with a high-performance GPU, which integrates a data preprocessing module, a deep learning prediction model, a multi-objective optimization strategy module, and a model iteration update module. The deep learning prediction model includes a temporal feature decomposition module, a two-dimensional transformation module, a multi-scale subsequence prediction module, and a result fusion module.

[0027] like Figure 2 The diagram shows the interface layout of the visualization system. The visualization system 4 is connected to the industrial server 3 and includes: a high-frequency data monitoring module, a prediction trend module, a simulation display area, and a historical prediction comparison area. The visualization system can display the operating status of the steam system, pressure prediction results, and optimized control strategies in real time.

[0028] like Figure 3 The diagram shows a flowchart of a deep learning-based method for predicting the inlet pressure of a steam generator, which includes the following steps.

[0029] S1: Multi-dimensional steam data from the steam generation, consumption, and power generation ends are collected in real time via a distributed industrial sensor network 1 at a frequency of 1 second / sample. Specifically, a first type of sensor is installed on the outlet pipes of the waste heat boilers in the sintering, steelmaking, and rolling processes to collect real-time steam output and pressure. A second type of sensor is installed on the steam consumption pipes of the RH boiler to collect the real-time steam consumption and operating status of the RH boiler. The operating status is represented by a switch signal, with 0 indicating shutdown and 1 indicating operation. A third type of sensor is installed on the steam pipe before the steam generator inlet to collect the generator inlet pressure and flow rate. A fourth type of sensor is installed on the valves in the steam network to collect the current valve opening percentage and adjustment stroke data. All sensors are industrial-grade sensors, and the sampling frequency is uniformly set to 1 second / sample, i.e., data is collected once per second to ensure sufficient time resolution. Among them, the steam production measurement accuracy of the first type of sensor is ±0.1t / h, the consumption measurement accuracy of the second type of sensor is ±0.05t / h, the pressure measurement accuracy of the third type of sensor is at least ±0.001MPa, the flow measurement accuracy is ±0.1t / h, and the opening accuracy of the fourth type of sensor is ±0.5%.

[0030] S2: The data preprocessing module in industrial server 3 receives steam data from data acquisition module 2 in real time and performs three levels of preprocessing on the data. The first level is noise suppression. In this embodiment, a wavelet threshold denoising algorithm is used to process the original signal. Specifically, the Db4 wavelet basis is selected, and the number of decomposition layers is set to 3. Soft threshold denoising is performed on high-frequency coefficients, and then the signal is reconstructed, thereby effectively filtering out high-frequency interference and retaining the true trend of pressure changes. The second level is intelligent missing value completion. In this embodiment, the KNN-Interp algorithm is used, with the nearest neighbor number K=10. The 10 most similar complete operating conditions to the current operating condition are searched from the historical database. The pressure values ​​at the corresponding time of these 10 operating conditions are weighted and averaged as the result of missing value completion. If the missing value is continuously missing for more than 10 seconds, it is further completed by combining the steam pipeline flow conservation equation. The third level is dynamic outlier removal. The system simultaneously applies the 3σ principle and industrial rules to identify outliers. For example, when the RH boiler operating status is 0 but the steam consumption is >0.5t / h, it is determined to be an outlier caused by sensor failure or data packet loss. For identified outliers, the system replaces them with a weighted moving average of data from adjacent 10-second intervals. The weights decay over time with a decay coefficient of 0.9, meaning data closer to the current moment has a higher weight, thus ensuring data continuity and reasonableness. After these three levels of processing, the data preprocessing module performs standardization and temporal reconstruction. Standardization uses the Z-Score method to calculate the historical mean and standard deviation for each feature dimension, converting each data point into a dimensionless value with a mean of 0 and a standard deviation of 1 to eliminate differences between different physical dimensions. Temporal reconstruction uses a sliding window model of "input window - output window," using 2400 one-second samples from the past 40 minutes as the input window and 1200 one-second samples from the next 20 minutes as the output window, with a window sliding step of 1 second to ensure real-time prediction.

[0031] S3: The standardized time-series samples, after preprocessing and reconstruction, are fed into the deep learning prediction model. First, the time-series feature decomposition module uses an adaptive variational mode decomposition algorithm to decompose the original pressure time-series signal. In this embodiment, based on the frequency characteristics of the input data, the pressure signal can be automatically decomposed into 3 to 5 intrinsic mode functions (IMFs). Specifically, the first IMF corresponds to the short-term fluctuation period, typically 5-10 minutes; the second IMF corresponds to the medium-term fluctuation period, typically 30-60 minutes; and the third IMF corresponds to the long-term trend period, typically 2-4 hours. For each subsequence obtained after decomposition, the model also performs the same decomposition on the corresponding input variables, establishing a cross-variable correlation mapping between the "input variable subsequence - pressure subsequence".

[0032] S4: The two-dimensional transformation module in the deep learning prediction model converts the one-dimensional time-series data of each subsequence into a two-dimensional tensor, and performs dynamic delay alignment and feature extraction. First, the principal period length is automatically estimated; for example, the principal period of the short-term mode is 10 minutes, or 600 1-second samples. Then, using this period length as a folding window, the one-dimensional time-series subsequence is reconstructed into a two-dimensional tensor. Assuming the input window contains 2400 samples and the principal period of the short-term mode is 600 samples, it can be reconstructed into a 4×600 two-dimensional tensor. For medium- or long-term modes, if the input window length is insufficient to contain a complete period, a periodic extension method is used to supplement the data. After completing the two-dimensional tensor reconstruction, the model performs dynamic delay alignment. Since there is a 5-10 minute delay between the steam consumption of the RH boiler and the generator inlet pressure response, and this delay is not a fixed value but varies non-linearly with factors such as steam flow rate, pipe length, and current pressure in the pipeline network, this embodiment pre-calculates the delay time under the current operating condition based on the steam pipeline network pressure loss calculation formula and the real-time collected pipeline flow data. Then, a tensor translation operation is performed on the subsequence corresponding to the RH furnace consumption in the two-dimensional tensor to align it with the pressure subsequence on the time axis. The aligned two-dimensional tensor is then input into a three-layer 2D convolutional neural network, with each convolution kernel being 3×3 and a stride of 1×1. The 2D convolutional neural network can extract features simultaneously along the time direction and the periodic phase direction.

[0033] S5: After feature extraction, the multi-scale subsequence prediction module in the deep learning prediction model receives feature maps from the 2D convolutional neural network and designs differentiated prediction branches for sub-branch periodic sequences of different periods. The short-term fluctuation branch uses a lightweight neural network with a number of hidden units of [number missing]. This branch has few parameters and fast computation, making it suitable for quickly capturing short-term dynamic changes. The medium-term fluctuation branch uses a temporal convolutional network, which expands the receptive field through dilated convolution. In this embodiment, the dilation coefficients are set to 2^0, 2^1, and 2^2, and the receptive field can cover approximately 120 time steps, effectively capturing the medium-term impact of multi-process collaborative changes. The long-term trend branch uses an attention-enhanced Transformer encoder, retaining only the encoder part, with 2 attention heads and 128 hidden dimensions. This branch can focus on key turning points in long-term trends, such as pressure changes caused by production plan switching.

[0034] S6: The deep learning prediction model's result fusion module fuses and corrects the predicted sub-sequences output from each branch. The fusion method employs an attention-weighted strategy, assigning dynamic weights to the predicted sub-sequences of each sub-branch cycle: through the training process, it learns the contribution of different sub-branch cycle sequences to the final pressure. For example, during short-term fluctuations caused by RH boiler consumption, the weight of IMF1 is increased to 0.6-0.7; under stable operating conditions, the weight of IMF3 is increased to 0.4-0.5, with the sum of the weights being 1. The sum of the three weights is always 1. After weighted summation, a preliminary fused prediction sequence is obtained. To ensure that the prediction results conform to the basic physical laws of the steam system, the model further introduces a physical rule correction module. This module is based on the mass conservation law of the steam system, i.e., total steam production = steam consumption + rate of change of pipeline storage. Based on the measured data at the current moment, the mass conservation deviation is calculated. If the flow conversion value corresponding to the fused prediction sequence deviates from the conservation constraint by more than a preset threshold, the weights of each branch are automatically adjusted and recalculated until the physical constraint is met or the iteration limit is reached. The corrected final prediction sequence is the steam generator inlet pressure change curve for the next 20 minutes.

[0035] The deep learning prediction model was trained using an optimization strategy specifically designed for this industrial scenario. The loss function was a composite form: mean squared error plus a trend penalty term. The mean squared error is calculated as the average of the squared errors between the predicted and true values; the trend penalty term calculates whether the trend of the predicted value matches the trend of the true value. If the trends are opposite, an additional penalty is applied, with the penalty coefficient α determined to be 0.1 based on the validation set. The model used historical production data from the past 6 months, totaling approximately 15 million 1-second samples, divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The AdamW optimizer was used, with an initial learning rate of 1e-4, weight decay of 1e-5, and a cosine annealing learning rate scheduling algorithm. The training consisted of 66 epochs with a batch size of 512. An early stopping mechanism was also enabled, terminating training early when the validation set loss no longer decreased after 6 consecutive epochs to avoid overfitting. After training, the mean absolute percentage error on the test set was less than 5%, and the prediction accuracy exceeded 90%.

[0036] Considering that steel production conditions change over time, the model's prediction accuracy may decrease. The Industrial Server 3 also integrates a model iteration and update module. This module calculates the deviation rate between the current predicted pressure and the actual feedback pressure in real time. When the deviation rate exceeds 10% for five consecutive times, incremental training is automatically triggered. Incremental training collects the latest 24 hours of 1-second-level data, totaling 86,400 samples. It freezes the parameters of the underlying temporal feature decomposition and two-dimensional transformation modules of the deep learning prediction model, updating only the fully connected layer weights of the multi-scale subsequence prediction module and the result fusion module. Incremental training takes ≤1.5 hours, ensuring the model quickly adapts to changes in operating conditions. Furthermore, a full training run is performed monthly, comprehensively optimizing model parameters using the latest accumulated historical data to ensure long-term stability.

[0037] The multi-objective optimization strategy module in S7 Industrial Server 3 automatically generates a control strategy to stabilize the inlet pressure of the steam generator based on the 20-minute pressure prediction curve output by a deep learning model. The optimization function contains three levels of objectives: the primary objective is to stabilize the generator inlet pressure within the design range, with a pressure deviation rate not exceeding ±2%; the constraint objective is to ensure normal operation of the RH boiler, while ensuring that the steam output of each process does not fall below its minimum production load requirement, thus avoiding impact on the main process; the auxiliary objective is to minimize the number of adjustments to the steam pipeline regulating valves and minimize energy consumption during steam transport. This multi-objective optimization function is solved using a weighted summation combined with constraint penalties, outputting specific control commands. These commands can be directly sent to the automated execution system or displayed to operators via a large visualization screen for decision-making reference.

[0038] S8: The visualization system 4 displays real-time monitoring data, pressure prediction curves, and optimized control strategies. The large screen uses an industrial-grade touchscreen, retrieving the latest data from the industrial server 3 and refreshing it every 10 seconds. The high-frequency data monitoring area displays real-time data for each process, including steam flow, pressure, RH furnace consumption, and generator inlet pressure, updated every second, in the form of a dynamic line graph. Operators can drag and drop to view historical data from the past 24 hours. The prediction trend area uses a solid blue line to plot the historical pressure of the past 40 minutes and a red dashed line to plot the predicted pressure curve for the next 20 minutes. Design boundaries of 0.6MPa and 0.9MPa are marked with green horizontal lines; a yellow warning or red alarm signal is triggered when the prediction curve reaches a boundary. The simulation display area uses fluid simulation animation; pipe colors change from green to red to indicate pressure from low to high, and arrow direction and size indicate steam flow direction and velocity. A prominent text prompt, "Next moment generator inlet predicted pressure: X.XX MPa," provides a clear reference for on-site personnel to adjust valves. The historical prediction comparison area uses a line graph to show the comparison between the prediction results of the most recent predictions and the actual values ​​at the corresponding times, and calculates and displays the current prediction accuracy in real time.

[0039] The embodiments verify the main technical effects of the present invention. In terms of prediction accuracy, the model's average absolute percentage error is less than 5%, and the prediction accuracy exceeds 90%, significantly outperforming existing prediction methods based on ARIMA or a single LSTM. Regarding real-time response, due to the use of 1-second high-frequency sampling and sliding window reconstruction, the system can output prediction results for the next 20 minutes every second, far superior to the 1-minute sampling interval and 3-5 minute response lag of traditional manual monitoring. In terms of operational adaptability, when the RH furnace is restarted after maintenance or the sintering machine's trolley is replaced, causing changes in operating conditions, the model iteration update module automatically triggers incremental training. Only the latest 24-hour data is needed to complete parameter fine-tuning within 1.5 hours, and the prediction accuracy quickly recovers to over 90%. In terms of control performance, the multi-objective optimization strategy, by adjusting valve opening and steam distribution in advance, reduces the pressure fluctuation caused by RH furnace consumption from ±0.3MPa to ±0.1MPa, reduces generator load fluctuation by approximately 70%, and does not trigger low-pressure alarms or protective shutdowns throughout the process. In terms of visualization and human decision-making, the display system enables operators to obtain early warning information and optimization suggestions 5-15 minutes before stress changes occur, which greatly reduces reliance on personal experience and reduces the risk of misoperation.

[0040] Furthermore, the embodiments also verify the versatility of the present invention. Its core "time series decomposition-two-dimensional transformation-multi-scale prediction-fusion correction" architecture can be transferred to other industrial time series scenarios such as chemical reactor temperature prediction. After lightweight quantization pruning, the single inference time can be achieved on edge devices with a time consumption of no more than 0.8 seconds.

[0041] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A deep learning-based steam generator inlet pressure prediction system, characterized in that, include: The system comprises a distributed industrial sensor network (1), a data acquisition module (2), an industrial server (3), and a visualization display system (4); the data acquisition module (2) is connected to the distributed industrial sensor network (1), the industrial server (3) is connected to the data acquisition module (2), and the visualization display system (4) is connected to the industrial server (3).

2. The deep learning-based steam generator inlet pressure prediction system according to claim 1, characterized in that, The distributed industrial sensor network (1) includes: a first type of sensor group deployed at the outlet of the waste heat boiler in the sintering process, steelmaking process, and rolling process to collect steam output and pressure; a second type of sensor group deployed in the steam consumption pipeline of the RH furnace to collect the real-time steam consumption and operating status of the RH furnace; a third type of sensor group deployed at the inlet of the steam generator to collect the inlet pressure and flow of the steam generator; and a fourth type of sensor group deployed on the valves of the steam pipeline network to collect the current valve opening and adjustment stroke data; wherein, the sampling frequency of all sensors is 1 second / sample, and the measurement accuracy of the pressure sensor is at least ±0.001MPa.

3. The deep learning-based steam generator inlet pressure prediction system according to claim 1, characterized in that, The industrial server (3) integrates a data preprocessing module, which includes: a noise suppression unit, a missing value compensation unit, an outlier removal unit, a data standardization unit, and a time series reconstruction unit. The noise suppression unit uses a wavelet threshold denoising algorithm to filter out high-frequency noise in the data. The missing value completion unit completes missing values ​​based on the KNN-Interp algorithm. The outlier removal unit integrates the 3σ principle with industrial rules to identify and remove outliers. The data standardization unit uses Z-Score standardization to eliminate dimensional differences. The time series reconstruction unit reconstructs time series data in a mode with an input window of 40 minutes, an output window of 20 minutes, and a sliding step of 1 second.

4. The deep learning-based steam generator inlet pressure prediction system according to claim 3, characterized in that, The industrial server (3) also integrates a deep learning prediction model, which includes: a time-series feature decomposition module, a two-dimensional transformation module, a multi-scale subsequence prediction module, and a result fusion module. The time-series feature decomposition module decomposes the pressure time-series signal into 3-5 intrinsic mode functions, corresponding to short-term fluctuation cycles, medium-term fluctuation cycles, and long-term trend cycles, respectively. The two-dimensional transformation module reconstructs the one-dimensional time-series subsequence of each intrinsic mode function into a two-dimensional tensor with its period length as the folding window, performs dynamic delay alignment processing, and then extracts the features in the two-dimensional tensor through a 2D convolutional neural network. The multi-scale subsequence prediction module has differentiated prediction branches for sequences with different cycles, including: a lightweight neural network branch for processing short-term fluctuation cycles, a time-series convolutional network branch for processing medium-term fluctuation cycles, and an attention-enhanced Transformer encoder branch for processing long-term trend cycles. The result fusion module uses an attention-weighted fusion strategy to assign dynamic weights to the predicted subsequences of each branch and introduces a physical rule correction module based on the mass conservation law of steam systems to constrain and correct the fusion results.

5. The deep learning-based steam generator inlet pressure prediction system according to claim 4, characterized in that, The industrial server (3) also integrates a multi-objective optimization strategy module. The multi-objective optimization strategy module constructs the following optimization function: the main objective is to keep the steam generator inlet pressure stable within the design range and the deviation rate does not exceed ±2%. The constraint objective is to ensure normal production of the RH furnace and that the steam output of each process is not lower than the production load requirement. The auxiliary objectives are to minimize the number of valve adjustments and minimize the steam transmission energy consumption.

6. The deep learning-based steam generator inlet pressure prediction system according to claim 1, characterized in that, The visualization display system (4) includes: a high-frequency data monitoring module, a prediction trend module, a simulation display area, and a historical prediction comparison area; the high-frequency data monitoring module displays the steam data of each process updated in real time at 1 second in the form of dynamic curves; the prediction trend module displays the pressure prediction curve for the next 20 minutes in the form of a line graph, and marks the design range and warning threshold; the simulation display area simulates the pipeline pressure distribution and provides the pressure prediction reference data for the next moment; the historical prediction comparison area displays a line graph comparing the historical prediction results with the historical actual values ​​and calculates and displays the prediction accuracy.

7. The deep learning-based steam generator inlet pressure prediction system according to claim 4, characterized in that, The industrial server (3) also includes a model iteration update module. The model iteration update module calculates the deviation between the predicted pressure and the actual pressure in real time. When the deviation rate exceeds 10% for multiple consecutive times, incremental training is triggered. The incremental training collects the latest 24 hours of data, freezes the parameters of the time series feature decomposition module and the two-dimensional transformation module, and only updates the weights of the multi-scale subsequence prediction module and the result fusion module.

8. The deep learning-based steam generator inlet pressure prediction system according to claim 4, characterized in that, The deep learning prediction model is trained using a composite loss function, which is the mean squared error plus a trend penalty term. The trend penalty term is determined based on whether the trend of the predicted value is consistent with the trend of the true value.

9. The prediction method for a deep learning-based steam generator inlet pressure prediction system according to any one of claims 1-8, characterized in that, Includes the following steps: S1: The distributed industrial sensor network (1) collects multi-dimensional steam data from the steam generation end, consumption end and power generation end in real time at a frequency of 1 second / sample; S2: The collected multi-dimensional steam data is preprocessed by the data preprocessing module and reconstructed into a standardized time series sample with the past 40 minutes as the input window and the next 20 minutes as the output window; S3: The time-series feature decomposition module of the deep learning prediction model decomposes the standardized stress time-series signal into multiple subsequences with different periodic scales. S4: The one-dimensional time series data of each subsequence is converted into a two-dimensional tensor and feature extraction is performed through the two-dimensional transformation module of the deep learning prediction model; S5: The multi-scale subsequence prediction module of the deep learning prediction model is used to predict subsequences at different periodic scales to obtain multiple predicted subsequences. S6: The result fusion module of the deep learning prediction model merges multiple prediction sub-sequences and corrects them with physical rules to output the prediction curve of the steam generator inlet pressure for the next 20 minutes. S7: Through the multi-objective optimization strategy module, a control strategy including pressure stabilization, production constraints and energy consumption optimization is generated based on the pressure prediction curve; S8: The pressure prediction curve and the control strategy are displayed in real time through the visualization display system (4).

Images