Glass kiln waste heat resource advanced intelligent management and control system based on ai prediction model

Abstract

The application relates to the technical field of kiln temperature control, and discloses a glass kiln waste heat resource advanced intelligent control system based on an AI prediction model, which comprises a data acquisition layer, an AI prediction model layer and a control decision and execution layer; the data acquisition layer is used for collecting upstream operation parameters of a glass kiln and downstream operation parameters of a downstream waste heat boiler or a heater in real time; the AI prediction model layer is internally provided with an AI prediction model used for predicting the change trend of downstream steam temperature or heat; the control decision and execution layer comprises a PLC controller, the PLC controller comprises an advanced compensation control module and a PID feedback control module; the advanced compensation control module is used for calculating an advance adjustment amount of heater power; the PID feedback control module is used for correcting according to the real-time deviation of the actual downstream steam temperature and a set value, so as to form a composite control of AI feedforward prediction and PID feedback fine adjustment. The application can realize accurate prediction and active advanced regulation and control of waste heat resources.

Description

Technical Field

[0001] This invention relates to the field of kiln temperature control technology, specifically to an advanced intelligent management and control system for waste heat resources in glass kilns based on an AI predictive model. Background Technology

[0002] Glass kilns are the core thermal equipment in glass production lines. Their exhaust waste heat is typically recovered and utilized through waste heat boilers or heaters, generating superheated steam to drive steam turbine generators, a crucial method for energy conservation and emission reduction in the glass industry. However, the glass kiln industry currently suffers from problems such as extensive management, low levels of informatization, and severe data silos, particularly a lack of effective linkage and predictive mechanisms between the upstream glass production stage and the downstream waste heat boiler / heater stage. Existing control systems mostly employ pure feedback control (such as PID cascade control), meaning that heater power or waste heat boiler flue gas dampers are only adjusted after the downstream steam temperature has deviated from the set value. This "post-event adjustment" model leads to a series of inherent problems: (1) The regulation lag is serious. From the occurrence of fluctuations in upstream flue gas temperature or flow rate, to the transfer of heat to downstream heat exchange equipment, to the actual deviation of downstream steam temperature from the set value, and finally to the control system detecting the deviation and starting to regulate, there is a large lag time in the whole process. The steam temperature fluctuates greatly and the recovery time is long. (2) Waste heat power generation efficiency is low, and it is impossible to respond in advance to changes in operating conditions such as upstream feed adjustment, kiln pressure fluctuation, and fuel switching. Downstream steam parameters are unstable, and the turbine unit cannot operate continuously at the optimal operating point, resulting in a significant decrease in power generation efficiency. (3) The unit has high safety risks. Severe temperature fluctuations may cause safety problems such as excessive vibration of the turbine unit, water erosion of the blades, and thermal stress fatigue, which will shorten the service life of the equipment. (4) A large amount of upstream kiln operation data (such as feed amount, kiln pressure, flue gas temperature, flue gas flow rate, oxygen content, etc.) has not been used to predict downstream heat demand, forming data silos and the data value has not been fully explored.

[0003] The root cause of the above problems lies in the long time delay between changes in upstream combustion conditions and the thermal response of downstream heat exchange systems. Traditional feedback control can only start to act after a deviation occurs, and cannot make advance adjustments before the disturbance is transmitted to the downstream. Therefore, there is an urgent need for an intelligent management and control solution that can sense upstream changes in advance and proactively predict downstream heat demand.

[0004] To address the aforementioned issues, some studies have attempted to introduce predictive control concepts into the field of kiln temperature control. For example, Chinese patent CN106773682A discloses a method that uses an ARMAX (Autoregressive Moving Average) model to dynamically determine the time lag between natural gas flow and bottom temperature in a glass kiln, and combines it with an ANFIS (Adaptive Neural Fuzzy Inference System) to establish a bottom temperature prediction model. This method controls the natural gas flow using a combination of feedforward and feedback, thereby reducing bottom temperature fluctuations. However, this approach still has significant limitations: ARMAX, as a linear time series model, has limited ability to fit the complex nonlinear relationship between upstream multi-parameter coupling and downstream steam temperature, and its prediction accuracy is insufficient to meet the precise control requirements of waste heat recovery systems for steam parameters. Furthermore, the ANFIS model suffers from rule explosion under high-dimensional input conditions; as the input feature dimension increases, the number of fuzzy rules grows exponentially, resulting in low model training efficiency, insufficient generalization ability, and difficulty in stable operation in practical engineering. In addition, some solutions employ deep learning methods such as LSTM (Long Short-Term Memory) to predict the main steam temperature of boilers in thermal power plants and use the prediction difference as feedforward compensation for PID control. However, these solutions are applied to coal-fired boilers in thermal power plants, which differ significantly from glass kilns in terms of thermal characteristics, flue gas composition, and load variation patterns, making it difficult to directly apply their control strategies. Overall, existing technologies still have significant shortcomings in prediction accuracy, model generalization ability, and adaptability to waste heat recovery scenarios, and a systematic solution specifically designed for the advanced management of waste heat resources in glass kilns is lacking. Summary of the Invention

[0005] The present invention aims to provide an intelligent and proactive control system for waste heat resources in glass kilns based on an AI prediction model, which can achieve accurate prediction and proactive control of waste heat resources.

[0006] The basic solution provided by this invention is: an advanced intelligent management and control system for waste heat resources in glass kilns based on an AI predictive model, comprising: The data acquisition layer is used to collect the upstream operating parameters of the glass kiln and the downstream operating parameters of the waste heat boiler or heater in real time. The AI ​​prediction model layer is deployed in an edge computing gateway, server or PLC. It has an AI prediction model trained with historical time series data. The AI ​​prediction model predicts the trend of downstream steam temperature or heat in the future preset time window based on the upstream operating parameter sequence in the current and past preset time windows, and outputs a sequence of predicted values ​​containing multiple future time points. The control decision and execution layer includes a PLC controller, which receives the prediction results output by the AI ​​prediction model layer. The PLC controller includes an advance compensation control module and a PID feedback control module. The advance compensation control module calculates the advance adjustment amount of the heater power based on the predicted value sequence using the cumulative heat deviation method. The cumulative heat deviation method includes: calculating the cumulative heat deviation within a future preset time window based on the predicted value sequence, and calculating the advance adjustment amount based on the conversion coefficient between the cumulative heat deviation and the heater power. The PLC controller outputs a control signal to the actuator based on the advance adjustment amount to proactively adjust the heater power before the disturbance actually affects the downstream. The PID feedback control module is used to correct the actual downstream steam temperature based on the real-time deviation from the set value, forming a composite control of AI feedforward prediction and PID feedback fine-tuning. The control decision and execution layer also includes a model adaptive update module. The model adaptive update module records the actual correction amount of the PID feedback control module and the prediction deviation of the AI ​​prediction model layer. When the actual correction amount exceeds a preset threshold and has the same sign as the prediction deviation within M consecutive sampling periods, it is determined that there is a systematic deviation in the AI ​​prediction model, and the online incremental training of the AI ​​prediction model is triggered.

[0007] The working principle and advantages of this invention are as follows: This invention relates to an AI-based predictive model-driven intelligent control system for waste heat resources in glass kilns, enabling precise prediction and proactive, advanced regulation of waste heat resources. The key features are: Firstly, a composite control architecture combining feedforward prediction and feedback fine-tuning is constructed. The control system of this invention is not a simple replacement for existing PID control, but rather, while retaining the original PID feedback control, it superimposes an AI feedforward prediction channel, forming a dual-channel parallel composite control strategy. This organically integrates the advantages of both control mechanisms: the AI ​​feedforward channel can act in advance before disturbances propagate downstream, overcoming the inherent lag defect of feedback control; the PID feedback channel serves as the final steady-state safety guarantee, continuously correcting actual deviations to ensure the system maintains steady-state accuracy under any circumstances. The two complement each other and work synergistically, improving both the system's response speed and ensuring control stability and robustness.

[0008] Secondly, the cumulative heat deviation method achieves precise advance power compensation. The advance compensation control module of this invention does not simply substitute the predicted temperature value into the PID formula for calculation. Instead, it calculates the total heat deviation within the entire prediction window based on the predicted value sequence of multiple future time points output by the AI ​​model, and then derives the power advance adjustment amount based on the conversion coefficient. This energy accumulation-based calculation method fully considers the characteristics of the large inertia and long delay of the glass furnace waste heat system. Its compensation basis is the total heat deviation that will be generated in the future, rather than the instantaneous temperature deviation at a certain moment. This makes the power adjustment smoother and more precise, avoiding frequent power fluctuations and excessive wear of actuators caused by over-responding to a single prediction point, thus helping to improve the compensation effect and system stability.

[0009] Third, a model adaptive update mechanism based on PID correction triggering is established. The model adaptive update module of this invention creatively uses the output correction of the PID feedback control module as an automatic evaluation index for the performance of the AI ​​prediction model. When the PID continuously and significantly corrects in the same direction for multiple consecutive sampling periods, and the correction direction is consistent with the direction of the prediction deviation, the system can identify that the model has developed a systematic deviation, thereby automatically triggering online incremental training. This mechanism can achieve online adaptive updates of the prediction model without shutdown, effectively addressing the impact of long-term factors such as kiln aging, heat exchanger scaling, changes in fuel characteristics, and ambient temperature drift on model accuracy. Furthermore, its triggering logic does not rely on manual judgment or offline evaluation, truly achieving automated lifecycle management of the model. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the system architecture of the advanced intelligent management and control system for waste heat resources of glass kilns based on AI prediction model according to the present invention (the human-computer interaction layer is not shown). Detailed Implementation

[0011] The following detailed explanation illustrates the specific implementation methods: Example 1 The basic implementation examples are as follows: Figure 1 As shown: The advanced intelligent management and control system for waste heat resources of glass kilns based on AI prediction models includes: a data acquisition layer, an AI prediction model layer, a control decision and execution layer, and a human-computer interaction layer.

[0012] The data acquisition layer is used to collect upstream operating parameters of the glass kiln and downstream operating parameters of the waste heat boiler or heater in real time.

[0013] Specifically, key operating parameters, i.e., upstream operating parameters, are read in real time from the existing DCS (Distributed Control System) of the glass furnace. These upstream operating parameters include at least one or more of the following: flue gas temperature, flue gas flow rate, kiln pressure, feed rate, and oxygen content. Furthermore, to further improve the accuracy of model predictions, auxiliary parameters can be collected, including but not limited to: fuel type and its calorific value, combustion air temperature, glass liquid level / discharge rate (unit: mm or t / h), ambient temperature, and ambient humidity.

[0014] Operating parameters, i.e., downstream operating parameters, are read from the field control system of the waste heat boiler / heater. These downstream operating parameters include at least one or more of the following: steam temperature, steam pressure, steam flow rate, and current heater power.

[0015] All the above-mentioned acquisition parameters are collected with a fixed sampling period, which is set to 10 seconds in this embodiment. In other embodiments, the sampling period can be dynamically adjusted according to the actual working conditions, and a value between 5 seconds and 30 seconds can achieve the technical effect of the present invention.

[0016] All raw data collected by the data acquisition layer is timestamped and then transmitted in real time to the AI ​​prediction model layer and the control decision and execution layer for online control and prediction. On the other hand, it is stored in an industrial real-time database (such as PISystem or InfluxDB) for subsequent historical data query, model training and data analysis.

[0017] The AI ​​prediction model layer is deployed in an edge computing gateway, server, or PLC. It has an AI prediction model trained on historical time series data. The AI ​​prediction model predicts the trend of downstream steam temperature or heat in the future preset time window based on the upstream operating parameter sequence in the current and past preset time windows, and outputs a sequence of predicted values ​​containing multiple future time points.

[0018] The AI ​​prediction model takes upstream operating parameter time series data from the past 10 to 30 minutes as input and outputs a sequence of downstream steam temperature predictions for the next 5 to 15 minutes. The prediction sequence contains temperature predictions for multiple time points.

[0019] The AI ​​prediction model is constructed using one or more time series prediction algorithms, such as Long Short-Term Memory (LSTM), gated recurrent units (ROUs), temporal convolutional networks (TCNNs), or Transformers. In this embodiment, the AI ​​prediction model uses a Long Short-Term Memory (LSTM) network because LSTM, as a special type of recurrent neural network, can effectively capture long-term dependencies in time series data through its unique gating mechanism (forget gate, input gate, output gate), making it particularly suitable for handling long delays (5–20 minutes) and nonlinear coupling characteristics in glass furnace waste heat systems.

[0020] Specifically, the model structure of the AI ​​prediction model layer is as follows: The input layer receives the upstream operating parameter time series data of the past 30 minutes (i.e., 180 time points, with each time point spaced 10 seconds apart). The input feature vector of each time point includes five core parameters: flue gas temperature, flue gas flow rate, kiln pressure, feed rate, and oxygen content, with a feature dimension of 5.

[0021] After passing through the input layer, the data enters two LSTM hidden layers in sequence. Each hidden layer contains 64 LSTM units. A dropout layer (with a dropout rate of 0.2) is set between the two LSTM layers to prevent overfitting.

[0022] The output of the LSTM layer is fed into an attention layer, which automatically learns the importance weights of different historical moments for the current prediction task, enhancing the model's ability to focus on key historical information. The output of the attention layer passes through a fully connected layer (containing 32 neurons with ReLU activation function) and finally reaches the output layer. The output layer contains 10 nodes, each corresponding to a predicted steam temperature value for the next 10 minutes (one prediction point per minute, for a total of 10 prediction points).

[0023] The model is trained using a custom composite loss function, which consists of two parts: the first part is the mean squared error (MSE), which measures the deviation between the predicted temperature value and the actual temperature value; the second part is the temperature change rate error, which constrains the consistency between the trend of the predicted curve and the actual trend, and prevents unreasonable and violent oscillations in the predicted curve.

[0024] In other embodiments, the AI ​​prediction model may also employ one or more combinations of Gated Recurrent Units (GRUs), Temporal Convolutional Networks (TCNs), or Transformer architectures. GRUs, as a simplified variant of LSTMs, have the advantages of fewer parameters and faster training speeds; TCNs, through dilated convolution operations, can process long sequence data in parallel, resulting in high inference efficiency; and Transformers, through their self-attention mechanism, can capture the dependencies between any two positions in a sequence, offering potential advantages for systems with extremely long latency. Each of these three models has its own focus, and those skilled in the art can flexibly choose the appropriate model based on the computational resource constraints and accuracy requirements of the actual application scenario.

[0025] Furthermore, the AI ​​prediction model layer also includes a multi-model fusion module, which is used to deploy multiple different types of prediction models simultaneously, evaluate the prediction accuracy of each model in the most recent time window in real time, and dynamically select the output of the current best model as the prediction result based on the evaluation results, or perform weighted fusion of the prediction results of each model.

[0026] Specifically, in this embodiment, in addition to the LSTM model described above, the system also deploys a GRU model as a backup prediction model. The GRU model adopts the same input and output structure as the LSTM model, but the number of hidden layer units is adjusted to 48 to balance the computational overhead.

[0027] The multi-model fusion module runs in real time, and its operation is as follows: Within each control cycle (30 seconds), the LSTM and GRU models run inference independently, each outputting a steam temperature prediction sequence for the next 10 minutes. The multi-model fusion module records the prediction accuracy of both models over the past hour (i.e., 120 sampling points), using the root mean square error (RMSE) between the predicted and measured values ​​within the most recent hour as the accuracy evaluation metric. Then, the multi-model fusion module compares the RMSE values ​​of the LSTM and GRU models, selecting the output of the model with the smaller current RMSE as the final prediction result for the current cycle, which is then used for subsequent lead compensation control.

[0028] Through this dynamic selection mechanism, the system can automatically adapt to the performance differences of different models under different working conditions, effectively improving the robustness of predictions. In other embodiments, the multi-model fusion module can also adopt a weighted fusion method, that is, assign different fusion weights according to the accuracy evaluation results of each model (the higher the accuracy, the greater the weight), and perform a weighted average of the predicted values ​​output by each model to obtain the fused prediction result.

[0029] In addition, the AI ​​prediction model is trained offline in advance using the following methods: The first step was sample data collection. Operational data for three consecutive months was collected from the DCS system at the glass kiln production site, with a fixed sampling interval of 10 seconds. This three-month period covered different seasons (summer high-temperature season and winter low-temperature season), different load conditions (60%, 80%, and 100% of design load), and different fuel types (switching periods between natural gas and producer gas), ensuring the diversity and representativeness of the training sample. The collected data included five core upstream parameters (exhaust gas temperature, flue gas flow rate, kiln pressure, feed rate, and oxygen content), auxiliary parameters (fuel calorific value, combustion air temperature, and glass liquid level), and downstream steam temperature (as a label value). The total data volume was approximately (90 days × 24 hours × 360 sampling points / hour) = 777,600 sample points.

[0030] The second step is data preprocessing. The collected raw data is cleaned, specifically including: 1) removing obvious outliers caused by sensor malfunctions (such as data points with a temperature of 0 or exceeding the sensor's upper limit), and replacing them with the median value of all historical data for that parameter; 2) performing linear interpolation to fill in missing data; 3) normalizing all parameters using the Z-score standardization method to convert each parameter into standard normal distribution data with a mean of 0 and a standard deviation of 1, in order to eliminate the influence of different units on model training. The mean and standard deviation parameters used during normalization are saved separately and used for the same normalization process on newly collected data during online inference.

[0031] The third step is sample construction. A sliding window method is used to construct training sample pairs. The input X for each sample is the time-series data of the five upstream core parameters over the past 30 minutes (180 time points), with a shape of (180, 5); the output Y for each sample is the steam temperature data for the next 10 minutes (10 time points, one prediction point per minute), with a shape of (10, 1). The sliding step size is set to 1 minute (i.e., 6 sampling points), with a 29-minute overlap between adjacent samples. Using this method, approximately (777,600 - 180 - 10 + 1) ≈ 777,411 training samples are constructed from 777,600 sampling points. The constructed samples are then randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0032] The fourth step is model training. The LSTM model is trained using samples from the training set. Training employs mini-batch gradient descent, with each batch containing 64 samples. The Adam optimizer is used, with an initial learning rate of 0.001, which decays to 0.9 times after every 10 complete iterations. Early stopping is employed during training to prevent overfitting: training is terminated early when the loss function value on the validation set shows no improvement for 20 consecutive iterations, and the optimal model parameters from the validation set are saved.

[0033] Step 5: Model Testing and Deployment. The trained, optimal model is tested using samples from the test set. Evaluation metrics include root mean square error (RMSE) and mean absolute percentage error (MAPE). A model is considered to meet deployment requirements when the RMSE is below 2% and the MAPE is below 1%. The LSTM and GRU models that meet the accuracy requirements are converted to TensorFlow Lite format (FP16 accuracy) and deployed to the TensorFlow Lite runtime environment of an industrial edge computing server for online inference.

[0034] In this embodiment, the prediction function of the AI ​​prediction model layer is specifically manifested as follows: Based on the current and past 30-minute upstream operating parameter sequence (180 time points), the model outputs a sequence of predicted downstream steam temperatures for the next 10 minutes (10 values, corresponding to the predicted temperatures for the 1st to 10th minute). This predicted sequence is sent to the PLC controller in real time via Socket communication as the basis for proactive compensation control.

[0035] The inference frequency of the AI ​​prediction model layer is set to run once every 30 seconds (i.e., output a new sequence of prediction values ​​every 30 seconds).

[0036] It should be noted that the predictive inference frequency can be dynamically adjusted according to changes in operating conditions—a lower frequency is maintained during steady-state operation to save computing resources, while the frequency can be increased to once every 10 seconds when operating conditions fluctuate drastically.

[0037] The AI ​​prediction model layer supports transfer learning. It uses the AI ​​prediction model trained on historical data from a first glass furnace (e.g., a float glass furnace on a production line) as a pre-trained model. This pre-trained model is then fine-tuned using small sample data (e.g., 3-7 days of operational data) from a second glass furnace (e.g., a bottle / jar glass furnace on another production line) to obtain an AI prediction model adapted to the second glass furnace. This approach significantly reduces the model training time and data accumulation requirements when deploying the system in each new furnace, improving the system's portability and deployment efficiency.

[0038] The AI ​​prediction model layer also supports incremental training of the model on a regular basis (e.g., once a month) using newly collected operating data to adapt to long-term operating condition drift caused by kiln aging, fuel replacement and seasonal changes.

[0039] The control decision and execution layer includes a PLC controller, which receives the prediction results (a sequence of predicted steam temperature values ​​for the next 10 minutes) output by the AI ​​prediction model layer; at the same time, it collects analog signals such as downstream actual steam temperature, steam pressure, and steam flow through an analog input module (model SM 331, 8×12 bits); and collects the operating status and fault signals of each actuator through a digital input module.

[0040] Specifically, the PLC controller communicates with the glass furnace DCS system via OPC or Modbus protocol and with the AI ​​prediction model layer via industrial Ethernet. When communication is interrupted, the system automatically switches to pure PID control mode as a degradation strategy. Specifically, if the network communication interruption between the AI ​​prediction model and the PLC lasts for more than three scan cycles, the system automatically identifies it as a communication fault and switches to pure PID control mode, where the PID feedback control module independently undertakes all control tasks, ensuring system safety and preventing loss of control.

[0041] The PLC controller contains two parallel control logic units: an AI feedforward prediction module and a PID feedback control module. The two work together to form a composite control system that combines AI feedforward prediction with PID feedback fine-tuning.

[0042] The PID feedback control module is used to correct the actual downstream steam temperature based on the real-time deviation from the setpoint, forming a composite control of AI feedforward prediction and PID feedback fine-tuning. Its specific implementation mechanism is as follows: The PID feedback control module uses the real-time deviation between the steam temperature setpoint (200℃) and the actual downstream steam temperature as input, and calculates the control output using a standard PID positional algorithm. In this embodiment, the PID parameters are determined through on-site tuning as follows: proportional gain... Integral time Second, differential time The control cycle of the PID feedback control module is consistent with the data acquisition cycle, which is once every 10 seconds. The output of the PID feedback control module serves as the basic control variable, used to maintain the steam temperature near the set value under steady-state conditions, ensuring the steady-state control accuracy of the system.

[0043] The parameters of the PID feedback control module are adaptively adjusted based on the prediction results of the AI ​​prediction model layer or changes in operating conditions. When the future steam temperature fluctuation predicted by the AI ​​prediction model exceeds the preset threshold, the PID proportional gain is temporarily increased or a gain scheduling strategy is adopted.

[0044] Specifically, when the fluctuation range of the future steam temperature predicted by the AI ​​prediction model (i.e., the difference between the maximum and minimum values ​​in the predicted value sequence) is less than 5℃, the operating condition is determined to be stable, and the PID parameters remain at normal values. , Second, (seconds). When the fluctuation range predicted by the AI ​​is greater than or equal to 5℃ and less than 10℃, it is determined that there is a certain fluctuation in the operating condition, and the PID response speed is appropriately enhanced: the proportional gain is temporarily increased to The points accumulation time has been shortened to Seconds. When the AI ​​predicts a fluctuation range greater than or equal to 10℃, it is determined that the operating condition is fluctuating drastically, and the PID response is significantly enhanced: the proportional gain is temporarily increased to [value missing]. The points accumulation time has been shortened to Seconds, the differential time is extended to This is to enhance the damping effect of proactive regulation. After the fluctuation amplitude returns to the normal range, the PID parameters smoothly transition back to their normal values ​​within 3 minutes to avoid system oscillations caused by sudden parameter changes.

[0045] The advance compensation control module calculates the advance adjustment amount of the heater power based on the predicted value sequence using the cumulative heat deviation method. The cumulative heat deviation method includes: calculating the cumulative heat deviation within a future preset time window (e.g., 10 minutes) based on the predicted value sequence, and calculating the advance adjustment amount based on the conversion coefficient between the cumulative heat deviation and the heater power. The PLC controller outputs a control signal to the actuator based on the advance adjustment amount to proactively adjust the heater power before the disturbance actually affects the downstream.

[0046] In this embodiment, the cumulative heat deviation method specifically involves: obtaining each temperature prediction value in the predicted value sequence, calculating the product of the cumulative sum of the differences between each temperature prediction value and the target temperature and the sampling interval, and then multiplying it by the specific heat capacity of steam to obtain the cumulative heat deviation; dividing the cumulative heat deviation by the total duration of the prediction time window and then multiplying it by the conversion coefficient to obtain the advance adjustment amount of the heater power.

[0047] Right now: ; ; in, This refers to the specific heat capacity of steam at constant pressure. This is the steam mass flow rate (measured in real time by a downstream flow sensor). The time interval between each prediction time point (60 seconds in this example). If A positive value indicates that the predicted steam temperature is too high, and the heater power needs to be reduced; if... A negative value indicates that the predicted steam temperature is too low, and the heater power needs to be increased.

[0048] K is the conversion coefficient between heater power and heat (unit: kW / kJ, determined by the heater's electro-thermal conversion efficiency, and taken as 0.95 in this embodiment). The calculated... This refers to the power value (i.e., advance adjustment amount) that the heater needs to adjust in advance so that the steam temperature will approach the target temperature within the next 10 minutes.

[0049] Furthermore, the calculated Limiting is applied. The upper limit of the limit is 80% of the heater's rated power, and the lower limit is 0 (i.e., it will not be reduced to negative power) to prevent heater overload or excessive power fluctuations caused by extreme values ​​predicted by the model.

[0050] After limiting The analog output module of the PLC controller outputs the signal to the analog control terminal of the actuator (such as the power regulator). The actuator continuously adjusts the output power of the electric heater according to the signal, thereby actively compensating for heat before the flue gas temperature disturbance actually affects the downstream steam temperature.

[0051] In this embodiment, the advance compensation amount is dynamically determined based on the actual measured delay time of the system. The physical delay time for the glass furnace flue gas to flow from the upstream measuring point (furnace exhaust port) to the downstream heater / waste heat boiler inlet depends on the flue gas pipe length, flue gas velocity, and flue gas temperature. In this embodiment, the actual measured delay time is 6 minutes. Therefore, the advance compensation control module prioritizes the predicted value near the 6th minute in the AI ​​prediction sequence, using the prediction deviation at that moment as the main basis for calculating the primary weight of the cumulative heat deviation. In other embodiments, if the delay time is other values, the key time points in the prediction value sequence can be adjusted accordingly.

[0052] The control decision and execution layer also includes a condition change detection module, a model adaptive update module, and an early warning and interlocking module.

[0053] The operating condition change detection module is used to monitor the rate of change of upstream operating parameters in real time. When the rate of change of the feed rate, kiln pressure or flue gas temperature exceeds the preset threshold, it is determined to be an operating condition change event, triggering the AI ​​prediction model layer to perform incremental prediction or increase the prediction frequency, and dynamically adjusting the advance amount of advance compensation after the operating condition change event occurs.

[0054] In this embodiment, the operating condition change detection module is implemented as a software function block in a PLC to monitor the rate of change of three key parameters: feed rate, kiln pressure, and flue gas temperature in real time.

[0055] Specifically, the operating condition change detection module calculates the rate of change of each parameter within the last 30 seconds using a sliding window method (i.e., the current value minus the value 30 seconds ago, divided by 30 seconds). An operating condition change event is determined to have occurred when any one of the following conditions is met: the rate of change in feed rate exceeds 0.5 t / h / min (i.e., a change of more than 0.5 tons per hour per minute), the rate of change in kiln pressure exceeds 5 Pa / min, or the rate of change in flue gas temperature exceeds 5℃ / min. After an operating condition change event occurs, the PLC controller immediately sends an event trigger signal to the AI ​​server of the AI ​​prediction model. Upon receiving the trigger signal, the AI ​​server immediately performs model inference (not limited by the conventional 30-second inference cycle, but performing inference ahead of time) and sends the new prediction result to the PLC controller. Simultaneously, after an operating condition change event occurs, the advance compensation control module dynamically adjusts the advance compensation lead: shifting the focus from the conventional 6 minutes to 4 minutes to respond more quickly to sudden changes in operating conditions. Within 3 minutes of the triggering of the operating condition change event, the system maintains a high response state (i.e., the AI ​​inference frequency is increased to once every 10 seconds, and the proportional gain Kp in the PID parameter is temporarily increased to 3.2), and automatically returns to normal parameters after 3 minutes.

[0056] Through the operating condition change detection module, when an operating condition change event occurs, the AI ​​prediction model can be immediately triggered to perform incremental prediction or increase the prediction frequency, and dynamically adjust the lead amount of the advance compensation, while temporarily enhancing the response strength of the PID feedback control module. This event-triggered mechanism complements the conventional periodic prediction—periodic prediction ensures the quality of conventional control under steady-state conditions, while the event-triggered mechanism ensures that the system can respond as quickly as possible when operating conditions change abruptly, enabling the system to effectively cope with non-stationary operating conditions commonly encountered in glass furnace production, such as adjustments to the feed rate and fuel switching.

[0057] The model adaptive update module is used to record the actual correction amount of the PID feedback control module and the prediction deviation of the AI ​​prediction model layer. When the actual correction amount exceeds the preset threshold (10% of the PID output range in this embodiment) within M consecutive sampling periods (M is 60, i.e., 10 consecutive minutes) and has the same sign as the prediction deviation, it is determined that the AI ​​prediction model has a systematic deviation (i.e. the model continuously overestimates or underestimates, causing the PID to continuously correct in the same direction), and the online incremental training of the AI ​​prediction model is triggered.

[0058] When the model adaptive update module triggers online incremental training, it uses real-time data from the most recent period (e.g., 7 days before the trigger time) as new samples. The data undergoes the same cleaning and normalization processes as offline training (the normalization parameters are the same as those saved during offline training). The data from the most recent 7 days is used as the incremental training set. The same loss function and optimizer as offline training are used. The AI ​​prediction model is fine-tuned using a sliding window approach, and the model parameters are updated. After the update is completed, the old model is replaced with the new model.

[0059] In addition, the updated model and the original model are compared using validation data from the most recent 24 hours. If the updated model's RMSE on the validation set is more than 0.5°C lower than the original model, the updated model is used to replace the original model for subsequent online predictions. If the updated model's accuracy does not improve significantly, the update is abandoned, and the original model is retained to avoid reducing the model's generalization ability due to overfitting to short-term noise.

[0060] The warning and interlocking module is equipped with a graded warning threshold. When the predicted temperature value output by the AI ​​prediction model layer exceeds the first-level warning threshold (5°C higher or lower than the target temperature value; for example, when the target temperature value is 200°C, if the predicted temperature reaches 195°C or 205°C within the next 10 minutes, it is determined to exceed the first-level warning threshold), the system issues a warning. When the predicted temperature exceeds the Level 2 warning threshold (8°C higher or lower than the target temperature; for example, if the target temperature is 200°C and the predicted temperature reaches 192°C or 208°C within the next 10 minutes, it is considered to have exceeded the Level 2 warning threshold), the system automatically executes protective actions. These protective actions include, but are not limited to, limiting the heater power change rate (limiting it to no more than 5% / second of the rated power to prevent sudden power changes due to extreme deviations in model predictions), switching to the backup heater, and reducing the operating load (reducing the power generation load to 80% of the rated value to reduce the impact of steam temperature fluctuations on the turbine). When the predicted temperature returns to the normal range and remains stable for more than 3 minutes, the above protective actions are automatically deactivated and the system returns to normal operation mode.

[0061] The human-machine interaction layer includes a touch screen, which is used to display in real time upstream operating parameters (exhaust gas temperature, flue gas flow rate, kiln pressure, feed rate, oxygen content), downstream operating parameters (steam temperature, steam pressure, steam flow rate, current heater power), superimposed comparison of the temperature change curve predicted by the AI ​​prediction model (drawn as a dashed line) and the actual temperature curve (drawn as a solid line) in the same coordinate system, model confidence level indication (the 95% confidence interval is expressed as the predicted temperature value ± 1.96 times the standard deviation of the prediction error), early warning information, system status (normal / early warning / fault / degradation mode), etc.

[0062] The human-machine interface layer allows operators to set prediction lead time (adjustable range 3–15 minutes, default 6 minutes), intervention thresholds (i.e., the offset of each warning threshold, adjustable range ±2℃–±10℃), PID parameters (proportional gain Kp, integral time Ti, derivative time Td, only available to authorized operators), and model selection parameters (manual selection of LSTM or GRU as the primary model, or maintaining automatic selection mode). It also supports manual takeover control in emergency situations, switching to pure PID control mode with a single click (completely controlled independently by the PID feedback control module, with AI feedforward signals isolated). Manual takeover operation records are stored as log data for subsequent model optimization.

[0063] Based on the configuration and functions of the above modules, the complete control flow of the system in online operation mode in this embodiment is as follows: Data acquisition steps: The data acquisition layer acquires upstream operating parameters from the glass furnace DCS system via the OPC protocol at a fixed sampling period (10 seconds in this embodiment), and simultaneously acquires downstream operating parameters via the analog input module. When a communication failure occurs, the system automatically switches to degraded mode (pure PID control) and displays warning information in the human-machine interface layer.

[0064] Operating condition change detection steps: In each sampling period, the operating condition change detection module in the control decision and execution layer calculates the rate of change of feed rate, kiln pressure, and flue gas temperature within the most recent preset time window. When the rate of change of any parameter exceeds a preset threshold, it is determined to be an operating condition change event. After an operating condition change event occurs, the operating condition change detection module immediately sends an event trigger signal to the AI ​​prediction model layer. Upon receiving the trigger signal, the AI ​​prediction model layer immediately executes model inference and sends the new predicted value sequence back to the control decision and execution layer. Simultaneously, the operating condition change detection module triggers the lead compensation control module to dynamically adjust the lead amount of the lead compensation and temporarily adjusts the control parameters of the PID feedback control module. This high-response state continues for a preset time before automatically recovering.

[0065] AI Prediction and Multi-Model Fusion Steps: Under normal conditions (no operational change events), the AI ​​prediction model layer automatically performs model inference at a preset inference frequency. During inference, the AI ​​prediction model layer reads the upstream operating parameter time series data within the most recent preset time window from the local cache, and after standardization, inputs it into multiple different types of prediction models (LSTM and GRU models in this embodiment). Each model completes its own inference calculation and outputs a sequence of predicted downstream steam temperature values ​​within the future preset time window. The multi-model fusion module in the AI ​​prediction model layer uses the prediction accuracy of each model within the most recent preset time window as the evaluation index, selecting the output of the model with the highest accuracy as the final prediction result, or weighting and fusing the prediction results of each model as the final prediction result. The final prediction result is sent to the control decision and execution layer via communication. When an anomaly occurs in the AI ​​prediction model layer, a fault status code is automatically sent to the control decision and execution layer, which automatically switches to pure PID control mode upon receiving the code.

[0066] Advance compensation decision-making steps: After receiving the predicted value sequence sent by the AI ​​prediction model layer, the advance compensation control module in the control decision and execution layer obtains the predicted temperature values ​​at multiple future time points. Taking the predicted point corresponding to the actual system delay time as the main focus, the cumulative heat deviation method is used to calculate the power advance adjustment amount. After amplitude limiting processing, it is output to the execution mechanism through the analog output module.

[0067] Execution and Feedback Correction Steps: The actuator (in this embodiment, a power regulator and an electric heater) adjusts the heater's output power according to the control signals output from the control decision and execution layer, thereby changing the flue gas temperature entering the waste heat boiler and affecting the steam temperature generated by the waste heat boiler. Simultaneously, the PID feedback control module in the control decision and execution layer continuously performs PID calculations based on the real-time deviation between the actual downstream steam temperature and the target temperature in each control cycle. Its output, combined with the output of the lead compensation control module, acts on the actuator, forming a composite control of AI feedforward prediction and PID feedback fine-tuning to ensure steady-state control accuracy. The parameters of the PID feedback control module are adaptively adjusted based on the prediction results of the AI ​​prediction model layer or changes in operating conditions.

[0068] Model Adaptive Update Steps: The model adaptive update module in the control decision and execution layer periodically calculates the actual correction amount of the PID feedback control module and the prediction deviation of the AI ​​prediction model layer. When the actual correction amount exceeds a preset threshold and has the same sign as the prediction deviation within M consecutive sampling periods, it is determined that the AI ​​prediction model has a systematic bias. At this time, the model adaptive update module automatically triggers online incremental training, using the real-time collected data within the most recent preset time window as the new sample set, and performs fine-tuning training based on the deployed model parameters. After the fine-tuning training is completed, the updated model and the original model are compared and tested for accuracy using the most recent validation data. If the accuracy improvement of the updated model exceeds a preset value, the updated new model replaces the original model for subsequent online predictions; if the accuracy does not improve significantly, the update is abandoned, and the original model is retained.

[0069] Early warning and interlocking procedures: The early warning and interlocking module in the control decision and execution layer checks whether the predicted value sequence reaches the tiered early warning threshold within each AI prediction cycle. When the predicted value exceeds the first-level early warning threshold, the early warning and interlocking module issues a warning through the human-machine interface layer. When the predicted value exceeds the second-level early warning threshold, the early warning and interlocking module automatically executes protective actions, including but not limited to limiting the heater power change rate, switching to the standby heater, and sending a load reduction request to the turbine control system—one or more of these actions. Once the predicted temperature returns to the normal range and remains stable for a preset time, the early warning and interlocking module automatically deactivates the aforementioned protective actions and returns to normal operation mode.

[0070] Human-machine interaction and data upload steps: The human-machine interaction and data upload layer displays upstream operating parameters, downstream operating parameters, a comparison between AI-predicted temperature change curves and actual temperature curves, model confidence indicators, early warning information, and system status in real time. At the same time, it uploads operating data, prediction results, and control commands to the factory-level DCS system or cloud platform through standard communication protocols. Operators can set prediction lead time, intervention threshold, PID parameters, and model selection parameters through the human-machine interaction layer. It also supports manual takeover control by operators in emergency situations. The manual takeover operation records are stored as log data for subsequent model optimization.

[0071] This embodiment provides an AI-based predictive model-based intelligent control system for waste heat resources in glass kilns, which enables accurate prediction and proactive control of waste heat resources.

[0072] Example 2 An advanced intelligent management and control system for waste heat resources in glass kilns based on an AI predictive model is proposed. Building upon Example 1, this system introduces a multi-timescale dynamic feature extraction mechanism into the AI ​​predictive model layer. The core understanding of this mechanism is that the waste heat system of glass kilns is affected by disturbances at different time scales, and the rate of change of different operating parameters varies significantly.

[0073] Specifically, upstream operating parameters are categorized into the following three types based on their rate of change: Slow-varying characteristics include kiln pressure setpoints, oxygen content setpoints, and the thermal state of kiln refractory materials. These parameters typically change very slowly, with time constants on the order of hours or even days. Slow-varying characteristics reflect the long-term drift of the overall thermal state of the kiln, and their impact on downstream steam temperature exhibits a slow, trend-like change. Slow-varying characteristics are extracted using a longer time window (e.g., the past 2 hours) and a lower sampling frequency (e.g., sampling once every 60 seconds) to capture their long-term trends.

[0074] Medium-speed characteristics include flue gas temperature, feed rate, and fuel calorific value. These parameters are affected by factors such as production rhythm adjustments and raw material batch variations, with time constants on the order of minutes (approximately 5–30 minutes). Medium-speed characteristics are the main factors influencing downstream steam temperature, and their changes have a clear causal relationship with steam temperature fluctuations. Medium-speed characteristics are extracted using a medium time window (e.g., the past 30 minutes) and a medium sampling frequency (e.g., sampling once every 10 seconds), consistent with the conventional LSTM input window in Example 1.

[0075] Rapidly changing characteristics include instantaneous fluctuations in flue gas flow rate and combustion air pressure. These parameters are affected by factors such as burner pulsation and induced draft fan speed fluctuations, with time constants on the order of seconds (approximately 1–10 seconds). Although the fluctuation amplitude of rapidly changing characteristics may be small, their high-frequency components may excite the dynamic response of the heat exchange system under specific conditions (such as when the fan frequency is adjusted), and cannot be ignored. Rapidly changing characteristics are extracted using a relatively short time window (e.g., the past 2 minutes) and a high sampling frequency (e.g., sampling once every 2 seconds).

[0076] The three types of features are processed by independent feature encoding networks. Slow-varying features are input into a feature encoder consisting of fully connected layers, and the output is a slow-varying feature vector; medium-varying features are input into an LSTM encoder, and the output is a medium-varying feature vector; fast-varying features are input into another LSTM encoder (or a one-dimensional convolutional network, because it is more efficient at extracting features from high-frequency signals), and the output is a fast-varying feature vector.

[0077] The three feature vectors are integrated through a cross-scale attention fusion module. This module first calculates the correlation matrix among the three feature vectors, identifying the interrelationships between features at different time scales—for example, a step change in feed rate (medium speed) may lead to a slow recovery in kiln pressure (slow change), while simultaneously causing instantaneous oscillations in flue gas flow (fast change), indicating a cross-scale coupling relationship among the three. Then, the cross-scale attention fusion module calculates the attention weights of each feature vector based on the aforementioned correlation matrix, performing weighted fusion of the three feature vectors to generate a multi-scale fused feature vector. Finally, the fused feature vector is input into the prediction output layer to obtain a sequence of predicted future steam temperatures.

[0078] The advantages of the multi-timescale dynamic feature extraction mechanism are: by dividing parameters with different rates of change into frequency bands, it avoids the feature aliasing and information loss problems caused by mixing all parameters into the same time window, enabling the model to simultaneously capture the long-term trend, medium-term changes, and short-term fluctuations of the system, resulting in higher prediction accuracy. The cross-scale attention fusion module further explores the coupling relationship between features at different time scales, making the model's prediction results more consistent with the inherent laws of actual physical processes, thus improving the model's interpretability compared to Example 1.

[0079] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. An advanced intelligent management and control system for waste heat resources in glass kilns based on an AI predictive model, characterized in that, include: The data acquisition layer is used to collect the upstream operating parameters of the glass kiln and the downstream operating parameters of the waste heat boiler or heater in real time. The AI ​​prediction model layer is deployed in an edge computing gateway, server or PLC. It has an AI prediction model trained with historical time series data. The AI ​​prediction model predicts the trend of downstream steam temperature or heat in the future preset time window based on the upstream operating parameter sequence in the current and past preset time windows, and outputs a sequence of predicted values ​​containing multiple future time points. The control decision and execution layer includes a PLC controller, which receives the prediction results output by the AI ​​prediction model layer. The PLC controller includes an advance compensation control module and a PID feedback control module. The advance compensation control module calculates the advance adjustment amount of the heater power based on the predicted value sequence using the cumulative heat deviation method. The cumulative heat deviation method includes: calculating the cumulative heat deviation within a future preset time window based on the predicted value sequence, and calculating the advance adjustment amount based on the conversion coefficient between the cumulative heat deviation and the heater power. The PLC controller outputs a control signal to the actuator according to the advance adjustment amount, so as to proactively adjust the heater power before the disturbance actually affects the downstream. The PID feedback control module is used to correct the actual downstream steam temperature based on the real-time deviation from the set value, forming a composite control of AI feedforward prediction and PID feedback fine-tuning. The control decision and execution layer also includes a model adaptive update module. The model adaptive update module records the actual correction amount of the PID feedback control module and the prediction deviation of the AI ​​prediction model layer. When the actual correction amount exceeds a preset threshold and has the same sign as the prediction deviation within M consecutive sampling periods, it is determined that there is a systematic deviation in the AI ​​prediction model, and the online incremental training of the AI ​​prediction model is triggered.

2. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The upstream operating parameters include at least one or more of the following: exhaust gas temperature, flue gas flow rate, kiln pressure, feed rate, and oxygen content; the downstream operating parameters include at least one or more of the following: steam temperature, steam pressure, steam flow rate, and current heater power.

3. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The AI ​​prediction model is constructed using one or more time series prediction algorithms, such as Long Short-Term Memory Network, Gated Recurrent Unit, Temporal Convolutional Network, or Transformer. The AI ​​prediction model layer also includes a multi-model fusion module, which is used to deploy multiple different types of prediction models at the same time, evaluate the prediction accuracy of each model in the most recent time window in real time, and dynamically select the output of the current best model as the prediction result based on the evaluation results, or perform weighted fusion of the prediction results of each model.

4. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The AI ​​prediction model takes upstream operating parameter time series data from the past 10 to 30 minutes as input and outputs a sequence of downstream steam temperature predictions for the next 5 to 15 minutes, the prediction sequence containing temperature predictions for multiple time points. The cumulative heat deviation method specifically involves: obtaining each predicted temperature value in the predicted value sequence, calculating the product of the cumulative sum of the differences between each predicted temperature value and the target temperature and the sampling interval, and then multiplying it by the specific heat capacity of steam to obtain the cumulative heat deviation; dividing the cumulative heat deviation by the total duration of the prediction time window and then multiplying it by the conversion coefficient to obtain the advance adjustment amount of the heater power.

5. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The data acquisition layer acquires data at a fixed sampling period of 5 to 30 seconds; The PLC controller communicates with the glass furnace DCS system via OPC or Modbus protocol and with the AI ​​prediction model layer via industrial Ethernet. When communication is interrupted, the system automatically switches to pure PID control mode as a degradation strategy.

6. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The control decision and execution layer also includes a working condition change detection module; The operating condition change detection module monitors the rate of change of upstream operating parameters in real time. When the rate of change of the feed rate, kiln pressure or flue gas temperature exceeds the preset threshold, it is determined to be an operating condition change event, triggering the AI ​​prediction model layer to perform incremental prediction or increase the prediction frequency, and dynamically adjusting the advance amount of advance compensation after the operating condition change event occurs.

7. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The parameters of the PID feedback control module are adaptively adjusted according to the prediction results of the AI ​​prediction model layer or changes in operating conditions; when the future steam temperature fluctuation predicted by the AI ​​prediction model exceeds the preset threshold, the PID proportional gain is temporarily increased or a gain scheduling strategy is adopted. When the model adaptive update module triggers online incremental training, it uses real-time data from the most recent period as new samples, fine-tunes the AI ​​prediction model using a sliding window approach, updates the model parameters, and replaces the old model with the new model after the update is complete.

8. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, The control decision and execution layer also includes an early warning and interlocking module; The early warning and interlocking module is equipped with graded early warning thresholds. When the predicted temperature value output by the AI ​​prediction model layer exceeds the first-level early warning threshold, the system issues an early warning. When the predicted temperature value exceeds the level 2 warning threshold, the system automatically performs protective actions, including but not limited to limiting the rate of change of heater power, switching to the standby heater, and reducing the operating load, or one or more of these actions.

9. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model as described in claim 1, characterized in that, It also includes a human-computer interaction layer, which includes a touch screen for real-time display of upstream operating parameters, downstream operating parameters, a comparison between the temperature change curve predicted by the AI ​​prediction model and the actual temperature curve, model confidence level indication, early warning information and system status; The human-machine interaction layer allows operators to set prediction lead time, intervention threshold, PID parameters and model selection parameters, and supports manual takeover control by operators in emergency situations. The manual takeover operation record is stored as log data for subsequent model optimization.

10. The advanced intelligent management and control system for waste heat resources of glass kilns based on an AI prediction model according to claim 1, characterized in that, The AI ​​prediction model layer supports transfer learning. The AI ​​prediction model trained with historical data from the first glass furnace is used as a pre-trained model. The pre-trained model is then fine-tuned using small sample data from the second glass furnace to obtain an AI prediction model that is adapted to the second glass furnace. The AI ​​prediction model layer also supports periodic incremental training of the model using newly collected operational data to adapt to long-term operating condition drift caused by kiln aging, fuel replacement, and seasonal changes.

Images