Method and system for optimizing energy consumption of industrial production process in whole cycle based on dynamic mechanism

Abstract

The application discloses a dynamic mechanism-based full-cycle energy consumption optimization method and system for industrial production processes, which comprises four core steps of multi-dimensional data acquisition and standardized preprocessing, dynamic mechanism-data fusion model construction, full-cycle operation stage division and parameter adaptation, data-driven intelligent optimization and closed-loop iteration, and realizes precise optimization of full-cycle energy consumption by combining a deep Q network (DQN) algorithm and an improved eagle optimization algorithm; meanwhile, an optimization system integrating multiple modules is designed to complete integrated implementation of data flow, model operation and optimization control. The application adopts the above-mentioned dynamic mechanism-based full-cycle energy consumption optimization method and system for industrial production processes, compensates for mechanism model deviation in a data-driven manner, improves prediction and optimization efficiency, is suitable for multiple industrial fields, and has remarkable energy-saving benefits.

Description

Technical Field

[0001] This invention relates to the field of energy consumption control technology in industrial production, and in particular to a method and system for optimizing energy consumption throughout the entire industrial production process based on dynamic mechanisms. Background Technology

[0002] Energy consumption in industrial production processes accounts for more than 60% of total social energy consumption. In industries such as chemical and metallurgical manufacturing, the energy consumption of core production equipment (such as cracking furnaces, reaction vessels, and smelting furnaces) throughout their entire life cycle accounts for 30%-50% of the total energy consumption of enterprises. Optimizing energy consumption throughout the entire life cycle is crucial for energy conservation, emission reduction, and improved economic efficiency in the industrial sector.

[0003] Existing industrial energy consumption optimization technologies have many core defects: insufficient model adaptability is prominent. Traditional optimization relies on static mechanism models, which can only adapt to the stable operation stage and are difficult to respond to parameter fluctuations under dynamic operating conditions such as start-up and shutdown. At the same time, they do not consider uncertain factors such as equipment aging and changes in raw material composition, resulting in large prediction deviations, usually exceeding 7%. Pure data-driven models lack physical and logical constraints, have weak generalization ability under new operating conditions, and the optimization results are prone to deviating from the actual process requirements.

[0004] Existing fusion models mostly employ fixed-weight fusion methods, failing to dynamically adjust the contributions of the mechanistic and data models based on operating conditions. Furthermore, the ambiguous model structure and unclear parameter definitions further contribute to poor model interpretability and significant challenges in engineering implementation. The efficiency of optimization algorithms also needs improvement. Most existing intelligent algorithms are not pre-trained on historical data, resulting in slow convergence speeds and iterations often exceeding 200. They also lack differentiated optimization strategies for different stages of the entire lifecycle, easily getting trapped in local optima and achieving limited energy savings, typically only 5%-8%. Traditional optimization algorithms such as genetic algorithms and particle swarm optimization also suffer from complex parameter tuning and slower convergence speeds in later stages. In addition, existing technologies lack closed-loop iteration mechanisms; optimization results are not continuously verified and corrected using real-time data, and model and algorithm parameters cannot be adaptively updated. After long-term operation, the optimization effect significantly diminishes, making it difficult to adapt to the dynamic changes in industrial production.

[0005] Therefore, there is an urgent need for a full-cycle energy consumption optimization method and system with a clear model structure, deep integration of mechanism and data, strong algorithm targeting and closed-loop iteration capability, in order to solve the pain points of poor adaptability and limited optimization effect of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for optimizing energy consumption throughout the entire industrial production process based on dynamic mechanisms. This invention solves the technical problems of poor dynamic adaptability of traditional models, insufficient integration of mechanisms and data, weak generalization ability of optimization algorithms, and lack of closed-loop iterative mechanisms. It improves optimization efficiency and global optimization capabilities, and further realizes accurate, efficient, and continuous optimization of energy consumption throughout the entire industrial production process.

[0007] To achieve the above objectives, this invention provides a method for optimizing energy consumption throughout the entire lifecycle of an industrial production process based on dynamic mechanisms, comprising the following steps: S1. Collect multi-dimensional production data throughout the entire industrial production cycle and perform preprocessing operations; S2. Construct a dynamic mechanism model, combine it with a data-driven error compensation model to establish a fusion model, and adaptively adjust the fusion weights through a sliding time window. S3. Divide the entire industrial production cycle into stages, and adjust the weights and constraints of the fusion model parameters according to the objectives of each stage. S4. The pre-trained Deep Q-Network (DQN) algorithm and the improved Skyhawk optimization algorithm are combined with the algorithm-mechanism model collaborative optimization strategy to dynamically optimize the key parameters at each stage and output the initial optimization parameter set. S5. Adjust the production conditions using the initial optimized parameter set, collect real-time data through closed-loop iteration, calculate the prediction error of the fusion model, and adjust the weights of the fusion model and the key parameters of each stage in S4 according to the prediction error range.

[0008] Preferably, in S1, the multi-dimensional production data includes raw material characteristic data, operating status data, energy consumption and product data, and equipment status data; Raw material characteristic data includes feed composition Material density Viscosity Sulfur content Hydrogen-to-carbon ratio ; Operating status data includes reaction temperature Furnace pressure Smoke components Feed flow rate Steam ratio ; Energy consumption and product data include fuel consumption. Electricity consumption Product yield Product component distribution ; Equipment status data includes coking thickness on furnace tubes. Equipment runtime Burner operating status Valve opening .

[0009] Preferably, in S1, the preprocessing operation specifically includes: Missing data are filled using linear interpolation, and outliers are removed using the isolated forest algorithm. Pearson correlation analysis was used to screen features strongly correlated with energy consumption, and correlation coefficients were retained. Characteristic parameters ≥ 0.7; The data is mapped to the [0,1] interval using minimum-maximum normalization, as shown in the formula: ; in, This is the original data. The minimum value in the dataset. The maximum value in the dataset. This is standardized data after normalization.

[0010] Preferably, in S2, the dynamic mechanism model includes a reaction kinetics sub-model, a heat and proton transfer model, and a device loss sub-model. The reaction kinetics sub-model, used to describe the material conversion law, is expressed as follows: ; in, Components The rate of change of concentration over time, In order to generate components Components Total quantity Components concentration, Components Genetic components The reaction rate constant; Indicates the composition of components The generated components Total quantity Components Genetic components The reaction rate constant, For material flow rate, This represents the reaction path length. The heat and proton transfer model is used to describe the laws governing heat and mass transfer, and its expression is as follows: ; in, For material density, For isobaric specific heat capacity, The reaction temperature. For time, Thermal conductivity, The heat of reaction, The convective heat transfer coefficient, Ambient temperature; The equipment loss sub-model, used to describe the impact of equipment aging on energy consumption, is expressed as follows: ; in, Additional energy consumption due to equipment aging. The aging factor is... This represents the percentage of time the device is in use. This refers to the rated power of the equipment.

[0011] Preferably, in S2, the error compensation model adopts a two-layer LSTM neural network model, with the input being the input parameters and predicted values ​​of the dynamic mechanism model, and the output being the prediction bias; The fusion model employs a weighted fusion strategy, and the fusion formula is as follows: ; in, The energy consumption prediction value of the fusion model. For the weights of the dynamic mechanism model, For the weights of the error compensation model, This represents the energy consumption prediction value from the dynamic mechanism model. This represents the energy consumption prediction value of the error compensation model; The weights are adaptively updated based on the prediction error within the sliding time window, and the update formula is: ; in, The standard deviation of the prediction error for the dynamic mechanism model. The standard deviation of the prediction error for the error compensation model.

[0012] Preferably, in S3, the entire industrial production cycle is divided into a startup and adaptation phase, a stable operation phase, and a load reduction and shutdown phase. For each phase's objectives, the key parameter weights and optimization constraints of each sub-model in the fusion model are adjusted, with the following adjustment strategies: Increase the weights of the heat transfer and proton transfer model during the initial adaptation phase. Optimize heating rate and fuel consumption ; During the stable operation phase, the weight coefficients of each sub-model are balanced. Duration of stay And steam ratio Increase the weight of the equipment loss sub-model during the load reduction and shutdown phase. Optimize the load reduction rate and cooling rate .

[0013] Preferably, in S4, the reward function in the Deep Q-Network (DQN) algorithm... The expression is as follows: ; in, For energy consumption reduction rate, For product yield, The reaction temperature. To set the temperature, , , These are the weighting coefficients; Action value function The expression is as follows: ; in, This is the current state. For the current action, For instant rewards, As a discount factor, For the next state, For the next action, Value of network actions for the target; The improved Tianying optimization algorithm aims to minimize energy consumption per unit product and maximize product yield, using production process parameters as constraints, and constructs a fitness function: ; in, Energy consumption per unit of product , These are positive and negative adjustment coefficients, used to balance the optimization weights of reducing energy consumption and stabilizing product yield. For product yield Preferably, in S4, the algorithm-mechanism model co-optimization strategy includes the following steps: Step 1: Adjust the algorithm parameters based on the prediction error of the fusion model: When 3% < When the mutation rate is ≤5%, increase the mutation probability or explorer ratio of the improved Skyhawk optimization algorithm; when When the percentage is greater than 5%, the majority of individuals in the population are moved into the process constraint boundary, the global optimization weight is increased, and the fusion model weight is simultaneously triggered for emergency update. Step 2: Embed the output of the dynamic mechanism model into the fitness function. The modified fitness function is as follows: ; in, , The penalty coefficient is the mechanism constraint coefficient. Energy consumption per unit of product For product yield, Additional energy consumption due to equipment aging. Components The rate of change of concentration over time; increased during the start-up adaptation phase and the load reduction shutdown phase. Increase during stable operation phase ; Establish Weights of the heat transfer and proton transfer model Equipment loss sub-model weights Linear relationship: ; in, for The linkage coefficient, for The linkage coefficient, Based on the offset.

[0014] Preferably, S5 includes the following steps: S51. Collect optimized real-time industrial production data and calculate the actual prediction error of the fusion model using the following formula. : in, The energy consumption prediction value of the fusion model. This represents the actual energy consumption value collected in real time. S52, Based on prediction error Adjust the size based on feedback: If the actual prediction error 2.8%, keep the existing parameters unchanged, and continue to perform dynamic parameter optimization of S4; If the actual prediction error is 2.8% < 3%, keep the existing parameters unchanged, and continue the conventional optimization through S4 only; If the actual prediction error Based on the algorithm-mechanism model collaborative optimization strategy Adjust the core parameters of the improved Skyhawk optimization algorithm in S4 and update the adaptive weights of the fusion model synchronously. , ; The target network parameters of the Deep Q-Network (DQN) algorithm are iteratively adjusted, and the weight coefficients of the reward function are also adjusted. ; Correcting and improving the weight coefficients of the Skyhawk optimization algorithm Adjustment coefficient of fitness function , ; S53. Repeat the optimization steps S2-S4 until the prediction error is reached. If the value is ≤2.8%, stop the iteration and use the updated parameters as the new base parameters to continue the full-cycle energy consumption optimization.

[0015] This invention also provides a full-cycle energy consumption optimization system for industrial production processes based on dynamic mechanisms, comprising: The data acquisition module consists of distributed sensors, online analyzers, and metering equipment, used to collect multi-dimensional production data; the data preprocessing module is used for data cleaning, feature extraction and normalization, and outputs a standardized dataset. The fusion model building module integrates the dynamic mechanism model unit, LSTM error compensation unit, and weight adaptive unit to output fusion prediction results; the stage adaptation module is used to divide the production stage and adjust the model parameters and constraints accordingly. The intelligent optimization module integrates the Deep Q-Network (DQN) algorithm and the improved Skyhawk optimization algorithm, and outputs optimization operation parameters; the closed-loop control module is used to verify the optimization effect based on real-time data and trigger iterative updates of model and algorithm parameters. The output module is used to transmit optimized parameters to the production control system (DCS) to achieve automated control.

[0016] Therefore, the present invention employs the above-mentioned method and system for optimizing energy consumption throughout the entire industrial production process based on dynamic mechanisms, and the beneficial effects are as follows: (1) This invention integrates dynamic mechanism and data-driven model, and through closed-loop iteration, the energy consumption prediction error is reduced to ≤2.8%, which greatly improves the prediction accuracy and provides reliable support for precise optimization.

[0017] (2) The method and system of this invention improve the energy consumption of the whole cycle by 3-5 percentage points compared with the existing methods, while taking into account the product yield, and the energy saving and economic benefits are outstanding.

[0018] (3) This invention is adaptable to all stages of production and can be extended to multiple fields such as chemical and metallurgical industries. It has strong engineering feasibility and can cope with dynamic working conditions and equipment aging effects.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is an overall flowchart of an embodiment of the energy consumption optimization method for the entire life cycle of industrial production processes based on dynamic mechanisms of the present invention; Figure 2 This is an overall block diagram of an embodiment of the energy consumption optimization system for the entire life cycle of industrial production processes based on dynamic mechanisms of the present invention; Figure 3This is a schematic diagram illustrating the process of constructing a fusion model for an embodiment of the energy consumption optimization method for the entire life cycle of industrial production based on dynamic mechanisms of the present invention. Figure 4 This is a schematic diagram comparing the cumulative energy consumption throughout the entire lifecycle of the industrial production process based on dynamic mechanisms, as shown in Implementation Example 1 of the present invention. Figure 5 This is a schematic diagram of the full-cycle energy consumption prediction error of the experimental group in the first embodiment of the present invention, which is based on the dynamic mechanism of the full-cycle energy consumption optimization method and system for industrial production processes. Detailed Implementation

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] like Figure 1 As shown, the method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms includes the following steps: S1. Collect multi-dimensional production data throughout the entire industrial production cycle and perform preprocessing operations. The collected multi-dimensional production data includes raw material characteristic data, operating status data, energy consumption and product data, and equipment status data throughout the entire industrial production cycle.

[0024] Raw material characteristic data includes feed composition Material density Viscosity Sulfur content Hydrogen-to-carbon ratio .

[0025] Operating status data includes reaction temperature Furnace pressure Smoke components Feed flow rate Steam ratio .

[0026] Energy consumption and product data include fuel consumption. Electricity consumption Product yield Product component distribution .

[0027] Equipment status data includes coking thickness on furnace tubes. Equipment runtime Burner operating status Valve opening .

[0028] The collected multi-dimensional data undergoes preprocessing operations, specifically: Missing data were filled using linear interpolation, and outliers were removed using the Isolation Forest algorithm. Pearson correlation analysis was used to screen features strongly correlated with energy consumption, and correlation coefficients were retained. Characteristic parameters ≥ 0.7.

[0029] The minimum-maximum normalization method is used to map the data to the [0,1] interval to eliminate dimensional differences. The formula is as follows: ; in, This is the original data. The minimum value in the dataset. For the maximum value of the dataset, This is standardized data after normalization.

[0030] S2, such as Figure 3 As shown, a dynamic mechanism model is constructed that includes a reaction kinetics sub-model, a heat transfer and proton transfer sub-model, and a device loss sub-model; a fusion model is established by combining a data-driven error compensation model, and the fusion weights are adaptively adjusted through a sliding time window.

[0031] The reaction kinetics sub-model, used to describe the material conversion law, is expressed as follows: ; in, Components The rate of change of concentration over time In order to generate components Components Total quantity Components concentration, Components Genetic components The reaction rate constant; Indicates the composition of components The generated components Total quantity Components Genetic components The reaction rate constant, For material flow rate, This represents the reaction path length.

[0032] The heat and proton transfer model is used to describe the laws governing heat and mass transfer, and its expression is as follows: ; in, For material density, For isobaric specific heat capacity, The reaction temperature. For time, Thermal conductivity, The heat of reaction, The convective heat transfer coefficient, The ambient temperature.

[0033] The equipment loss sub-model, used to describe the impact of equipment aging on energy consumption, is expressed as follows: ; in, Additional energy consumption due to equipment aging. The aging factor is... The percentage of device runtime. This refers to the rated power of the equipment.

[0034] The error compensation model in this invention adopts a two-layer LSTM neural network model. The input is the input parameters and predicted values ​​of the dynamic mechanism model, and the output is the prediction deviation.

[0035] The fusion model employs a weighted fusion strategy, and the fusion formula is as follows: ; in, The energy consumption prediction value of the fusion model. For the weights of the dynamic mechanism model, For the weights of the error compensation model, This represents the energy consumption prediction value from the dynamic mechanism model. Let be the energy consumption prediction value of the error compensation model, and + =1.

[0036] The weights are adaptively updated based on the prediction error within the sliding time window, and the update formula is: ; in, The standard deviation of the prediction error for the dynamic mechanism model. The standard deviation of the prediction error for the error compensation model.

[0037] S3. The entire industrial production cycle is divided into three phases: startup and adaptation phase, stable operation phase, and load reduction and shutdown phase.

[0038] To achieve the goals at each stage, the parameter weights and optimization constraints of the fusion model are adjusted using the following strategies: The adaptation phase lasts for 0- ,in, From equipment ignition to reaching the set threshold operating parameters, the goal during the 6-10 hour period is rapid temperature rise, avoiding fuel waste, and reducing energy consumption during startup. Therefore, it is necessary to increase the weights of the heat transfer and proton transfer model. The value should be between 0.6 and 0.7; optimize the heating rate. Values ​​range from 4-8℃ / min; fuel consumption The value ranges from 50 to 200 kg / h.

[0039] The duration of the stable operation phase is - ,in, For the 6-8 hours prior to shutdown, equipment operating parameters, including temperature, flow rate, and pressure, are stabilized. The objective during this phase is to balance energy consumption reduction with product yield improvement; therefore, the weights of the equilibrium dynamic mechanism model are determined. =0.5; Optimize reaction temperature The temperature range is 800-900℃; residence time The value ranges from 0.2 to 0.6 s; steam ratio The value ranges from 0.3 to 0.5.

[0040] The duration of the load reduction and shutdown phase is - ,in, During shutdown, the feed flow rate and operating temperature are gradually reduced. The goal of this stage is to minimize energy loss during the cooling and load reduction process and protect the equipment; therefore, the weight of the equipment loss sub-model is increased. The value should be between 0.65 and 0.75 to optimize the load reduction rate. The value is between 5% and 12% / h, and the cooling rate is... The value is between 3-6℃ / min.

[0041] S4. The pre-trained Deep Q-Network (DQN) algorithm and the improved Skyhawk optimization algorithm are combined with the algorithm-mechanism model collaborative optimization strategy to dynamically optimize the key parameters at each stage and output the initial optimization parameter set.

[0042] The initialization parameter set is set based on the industrial production process constraint range and historical best parameter statistics. The core parameter initialization expressions and rules are as follows: Reward function in Deep Q-Network (DQN) algorithm The expression is as follows: ; in, For energy consumption reduction rate, Here, represents the product yield, and represents the reaction temperature. To set the temperature, , , Let be the weighting coefficient, satisfying =1.

[0043] Action value function The expression is as follows: ; in, This is the current state. For the current action, For instant rewards, As a discount factor, For the next state, For the next action, Value of network actions for the target.

[0044] This invention improves the Tianying optimization algorithm to find the optimal value within the optimization interval of key parameters output in real-time production status. Specifically: Set population size The iteration factor is 30-60. Levy flight step length factor 1.5, Number of iterations It is 80-150.

[0045] With the optimization objectives of minimizing unit product energy consumption and maximizing product yield, and with industrial production process constraints as the boundary conditions, the following constraints are set: furnace tube temperature ≤ 900℃, feed flow rate ≥ 5t / h. A fitness function is then constructed: ; in, Energy consumption per unit of product , These are positive and negative adjustment coefficients, used to balance the optimization weights of reducing energy consumption and stabilizing product yield; The yield is the product yield.

[0046] A dynamic Lévy flight strategy is introduced to optimize the position update formula. In the early stage of iteration, a small step size is used to focus on the refinement of local parameters. In the later stage of iteration, a random step size is used to enhance the global optimization capability and avoid local optima.

[0047] Adjusting weighting coefficients based on production stage characteristics Start the adaptation phase The value is 0.7, indicating a stable operating phase. The value is 0.5, during the load reduction and shutdown phase. The value is 0.6. The algorithm parameters are adaptively adjusted by subdividing the prediction error interval of the fusion model, and the algorithm-mechanism model co-optimization strategy is executed simultaneously to dynamically optimize key parameters at each stage, outputting an initial optimized parameter set. The algorithm-mechanism model co-optimization strategy includes the following steps: Step 1: Adjust the algorithm parameters based on the prediction error of the fusion model: When 3% < When the mutation rate is ≤5%, increase the mutation probability of the improved Skyhawk optimization algorithm to 1.2-1.5 times the original or the proportion of explorers to 60%-70% to improve population diversity.

[0048] when When the population reaches >5%, reset 30%-40% of the individuals in the population to within the process constraint boundary, increase the global optimization weight to 0.6-0.7, and simultaneously trigger an emergency update of the fusion model weights.

[0049] Step 2: Embed the output of the dynamic mechanism model into the fitness function. The modified fitness function is as follows: ; in, , The penalty coefficient is the mechanism constraint coefficient. Energy consumption per unit of product For product yield, Additional energy consumption due to equipment aging. Components The rate of change of concentration over time; increased during startup and load reduction phases. Increase adjustment during the stable phase .

[0050] Establish Weights of the heat transfer and proton transfer model Equipment loss sub-model weights Linear relationship: ; in, for The linkage coefficient, for The linkage coefficient, Based on the offset.

[0051] S5. Adjust production conditions using the initial optimized parameter set, collect real-time data through closed-loop iteration, calculate the prediction error of the fusion model, and adjust the weights of the fusion model and the key parameters of each stage in S4 according to the prediction error range. Specifically: Each interval Every hour, a closed-loop iteration verification is triggered. The iteration period is adapted to the sliding time window of the fusion model in S2, and includes the following steps: S51. Collect optimized real-time industrial production data and calculate the actual prediction error of the fusion model. The calculation formula is: in, The energy consumption prediction value of the fusion model. This represents the actual energy consumption value collected in real time.

[0052] S53, If the actual prediction error The 2.8% result indicates that the current fusion model and optimization algorithm parameters are adapted to the actual working conditions. The existing parameters of the fusion model weights, the deep Q-network DQN algorithm, and the improved Tianying optimization algorithm remain unchanged, and the dynamic parameter optimization of S4 continues.

[0053] If the actual prediction error is 2.8% < 3%, parameters are basically compatible, keep the existing parameters unchanged, and let S4 continue to seek optimization in a routine manner to avoid frequent adjustments that may cause fluctuations in operating conditions.

[0054] If the actual prediction error Insufficient parameter adaptability; optimization strategy based on algorithm-mechanism model. The interval subdivision rule first adjusts the core parameters of the improved Skyhawk optimization algorithm in S4 according to the collaborative strategy, and then simultaneously updates the adaptive weights of the fusion model. , .

[0055] Iteratively adjust the target network parameters of the Deep Q-Network (DQN) algorithm and adjust the weight coefficients of the reward function. .

[0056] Correcting and improving the weight coefficients of the Skyhawk optimization algorithm Adjustment coefficient of fitness function , .

[0057] S54. After completing the above parameter updates, repeat the optimization steps S2-S4 until the prediction error of the fusion model is within acceptable limits. If the value is ≤2.8%, stop the iteration and use the updated parameters as the new base parameters to continue the full-cycle energy consumption optimization.

[0058] like Figure 2 As shown, the present invention provides a dynamic mechanism-based full-cycle energy consumption optimization system for industrial production processes, used to implement the above method, comprising: The data acquisition module consists of distributed sensors, online analyzers, and metering equipment, used to collect multi-dimensional production data; the data preprocessing module is used for data cleaning, feature extraction and normalization, and outputs a standardized dataset.

[0059] The fusion model building module integrates the dynamic mechanism model unit, the LSTM error compensation unit, and the weight adaptive unit to output the fusion prediction results; the stage adaptation module is used to divide the production stage and adjust the model parameters and constraints accordingly.

[0060] The intelligent optimization module integrates the Deep Q-Network (DQN) algorithm and the improved Skyhawk optimization algorithm, and outputs optimized operation parameters; the closed-loop control module is used to verify the optimization effect based on real-time data and trigger iterative updates of model and algorithm parameters.

[0061] The output module is used to transmit optimized parameters to the production control system (DCS) to achieve automated control.

[0062] Example 1: A 100,000-ton / year ethylene unit of a petrochemical enterprise was selected, using an SRT-III type cracking furnace with naphtha as feedstock to produce ethylene and propylene. The entire cycle operation lasted 72 hours, covering the three stages of start-up, stabilization, and load reduction shutdown. Before the experiment, traditional static mechanism optimization was used, which had problems such as fuel waste and large energy consumption fluctuations. The energy consumption reduction rate for the entire cycle was only 6.8%, and the prediction error was 7.5%.

[0063] This embodiment employs a before-and-after comparison method, operating the same pyrolysis furnace under both the conventional method (control group) and the method of this invention (experimental group), while maintaining consistent interference factors. Data is collected using sensors and analyzers, and the effectiveness is verified by comparing indicators such as energy consumption, prediction accuracy, and product yield. Experimental tools include distributed sensors, an online gas chromatograph, an industrial control computer, and MATLAB R2023b.

[0064] 72 hours of multi-dimensional data were collected, with a sampling frequency of 5 minutes per sampling period, covering four core parameters: raw material characteristics, operating status, energy consumption, and product and equipment status. Linear interpolation was used to complete missing data, and the isolated forest algorithm was used to remove outliers. Pearson correlation analysis was used to select 8 features that were strongly correlated with energy consumption (correlation coefficient ≥ 0.7), and then min-max normalization was used to map them to the [0,1] interval.

[0065] The 72-hour period is divided into three stages, and the model weights and optimization parameters are adjusted accordingly, as follows: Table 1. Division of Operational Stages and Parameter Adjustment for the Pyrolysis Furnace Throughout its Life Cycle

[0066] A closed-loop iteration is triggered every 2 hours to calculate the prediction error and adjust the parameters according to the error range: when When ≤2.8%, maintain the parameter; when 2.8% < When ≤3%, conventional optimization is used; when If the error is greater than 3%, adjust the algorithm and model weights, and repeat steps S2-S4 until the error is less than or equal to 2.8%.

[0067] Table 2 Comparison of Core Performance Indicators

[0068] As shown in Table 2, the experimental group achieved an energy consumption reduction rate of 11.5% over the entire cycle, which is 4.7 percentage points higher than the 6.8% reduction rate of the traditional method (control group). This demonstrates outstanding energy-saving benefits and aligns with the core needs of energy conservation and consumption reduction in industrial production.

[0069] The energy consumption prediction error of the experimental group was only 2.3%, which was 5.2 percentage points lower than the 7.5% of the control group and far below the target threshold of 2.8%. This shows that the dynamic mechanism-data fusion model and closed-loop iterative mechanism constructed in this invention effectively improved the prediction accuracy.

[0070] After optimization, the average ethylene yield increased from 28.6% to 30.2%, achieving a dual benefit of reduced energy consumption and increased product yield, thus avoiding the negative impact of energy saving on production efficiency.

[0071] The experimental group algorithm converged in 112 iterations, which is 98 fewer than the 210 iterations in the control group, improving the convergence speed by 46.7%, thus meeting the efficiency requirements of real-time control in industrial production.

[0072] Tables 3 and 4 further corroborate the stability of the optimization method of the present invention, wherein Table 3 and Figure 4 The extent of energy consumption optimization at each stage can be observed intuitively, as shown in Table 4 and Figure 5 It can verify the closed-loop convergence effect of the prediction error.

[0073] Table 3. Comparison of Cumulative Energy Consumption Over the Entire Life Cycle (Unit: 10) 3 kg standard oil)

[0074] As shown in Table 3, the method of this invention has a significant advantage in energy consumption throughout the entire cycle. The cumulative energy consumption of the experimental group at each time point is lower than that of the control group. The cumulative energy consumption of the experimental group over the 72-hour cycle is 86.7 × 10⁻⁶. 3 kg of standard oil, control group: 98.7 × 10 3 The energy consumption per kg of standard oil over the entire cycle is significantly lower than the static result, indicating that the energy-saving effect is achieved throughout the entire production process.

[0075] Specifically, during the startup and adaptation phase (0-8h), the experimental group's cumulative energy consumption over 8 hours was 7.6 × 10⁻⁶. 3 kg of standard oil, compared to 8.5 × 10 kg of control group. 3 The 10.6% reduction in kg standard oil indicates the effectiveness of the weight optimization of the heat and mass transfer model during the start-up adaptation phase, which effectively reduces fuel waste during the start-up heating process.

[0076] During the stable operation phase (8-64h), the energy consumption gap gradually widened, with the experimental group consuming a cumulative energy of 78.8×10⁻⁶ at 64h. 3 kg of standard oil, compared to the control group 90.5 × 10 3The 12.9% reduction in standard fuel consumption highlights the continuous energy-saving effect achieved through precise parameter optimization during stable operation.

[0077] During the load reduction and shutdown phase (64-72h), the energy consumption growth rate of the experimental group was lower than that of the control group, further verifying the rationality of the weight adjustment strategy of the equipment loss model, which can reduce energy loss during the shutdown phase while protecting the equipment.

[0078] It is evident that the energy consumption of the experimental group was consistently lower than that of the control group throughout the entire cycle, with the difference remaining manageable and without any abnormal fluctuations. This demonstrates the synergistic effect of the dynamic mechanism fusion model, stage adaptation, and closed-loop iteration mechanism, which enables stable adaptation to different operating conditions throughout the entire cycle.

[0079] Table 4 and Figure 5 This can further verify the closed-loop convergence effect of the prediction error, and complement the data in Table 3 to prove the comprehensive effectiveness of the optimization method.

[0080] As shown in Table 4, the reliability of the dynamic mechanism fusion model and closed-loop control mechanism of this invention is demonstrated. The prediction error shows a rapid convergence and stable maintenance trend. The error was 4.8% at 2 hours, and dropped to 2.7% in just 8 hours, successfully falling below the preset error threshold of 2.8%, achieving rapid target attainment. Throughout the subsequent cycle, the prediction error remained stable in the range of 2.2%-2.4%, without significant rebound fluctuations, and remained below the error threshold throughout, exhibiting extremely strong stability.

[0081] Table 4. Prediction Error of Energy Consumption for the Experimental Group Throughout the Cycle

[0082] This change process demonstrates that the closed-loop iterative mechanism can play an effective role. Real-time data feedback and parameter adjustment every 2 hours can continuously correct the weights and algorithm parameters of the fusion model, keeping the prediction accuracy within the target range. The dynamic mechanism-data fusion model can accurately cope with parameter fluctuations under different operating conditions throughout the entire cycle, providing reliable data support for precise energy consumption optimization.

[0083] Therefore, the present invention adopts the above-mentioned energy consumption optimization method and system for the entire cycle of industrial production process based on dynamic mechanism. Through deep integration of mechanism and data, phased adaptation and closed-loop iteration, it solves the pain points of poor adaptability and limited optimization effect of existing technologies, and takes into account prediction accuracy, energy saving efficiency and generalization ability. It has significant energy saving benefits and wide engineering application value.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms, characterized in that, Includes the following steps: S1. Collect multi-dimensional production data throughout the entire industrial production cycle and perform preprocessing operations; S2. Construct a dynamic mechanism model, combine it with a data-driven error compensation model to establish a fusion model, and adaptively adjust the fusion weights through a sliding time window. S3. Divide the entire industrial production cycle into stages, and adjust the weights and constraints of the fusion model parameters according to the objectives of each stage. S4. The pre-trained Deep Q-Network (DQN) algorithm and the improved Skyhawk optimization algorithm are combined with the algorithm-mechanism model collaborative optimization strategy to dynamically optimize the key parameters at each stage and output the initial optimization parameter set. S5. Adjust the production conditions using the initial optimized parameter set, collect real-time data through closed-loop iteration, calculate the prediction error of the fusion model, and adjust the weights of the fusion model and the key parameters of each stage in S4 according to the prediction error range.

2. The method for optimizing energy consumption throughout the entire industrial production process based on dynamic mechanisms according to claim 1, characterized in that, In S1, multi-dimensional production data includes raw material characteristic data, operating status data, energy consumption and product data, and equipment status data; Raw material characteristic data includes feed composition Material density Viscosity Sulfur content Hydrogen-to-carbon ratio ; Operating status data includes reaction temperature Furnace pressure Smoke components Feed flow rate Steam ratio ; Energy consumption and product data include fuel consumption. Electricity consumption Product yield Product component distribution ; Equipment status data includes coking thickness on furnace tubes. Equipment runtime Burner operating status Valve opening .

3. The method for optimizing energy consumption throughout the entire industrial production process based on dynamic mechanisms according to claim 2, characterized in that, In S1, the preprocessing operation is as follows: Missing data are filled using linear interpolation, and outliers are removed using the isolated forest algorithm. Pearson correlation analysis was used to screen features strongly correlated with energy consumption, and correlation coefficients were retained. Characteristic parameters ≥ 0.7; The data is mapped to the [0,1] interval using minimum-maximum normalization, as shown in the formula: ； in, This is the original data. The minimum value in the dataset. The maximum value in the dataset. This is standardized data after normalization.

4. The method for optimizing energy consumption throughout the entire industrial production process based on dynamic mechanisms according to claim 3, characterized in that, In S2, the dynamic mechanism model includes a reaction kinetics sub-model, a heat and proton transfer model, and a device loss sub-model. The reaction kinetics sub-model, used to describe the material conversion law, is expressed as follows: ； in, Components The rate of change of concentration over time, In order to generate components Components Total quantity Components concentration, Components Genetic components The reaction rate constant; Indicates the composition of components The generated components Total quantity Components Genetic components The reaction rate constant, For material flow rate, This represents the reaction path length. The heat and proton transfer model is used to describe the laws governing heat and mass transfer, and its expression is as follows: ； in, For material density, For isobaric specific heat capacity, The reaction temperature. For time, Thermal conductivity, The heat of reaction, The convective heat transfer coefficient, The ambient temperature; The equipment loss sub-model, used to describe the impact of equipment aging on energy consumption, is expressed as follows: ； in, Additional energy consumption due to equipment aging. The aging factor is... This represents the percentage of time the device is in use. This refers to the rated power of the equipment.

5. The method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms according to claim 4, characterized in that, In S2, the error compensation model adopts a two-layer LSTM neural network model. The input is the input parameters and predicted values ​​of the dynamic mechanism model, and the output is the prediction deviation. The fusion model employs a weighted fusion strategy, and the fusion formula is as follows: ； in, The energy consumption prediction value of the fusion model. For the weights of the dynamic mechanism model, For the weights of the error compensation model, This represents the energy consumption prediction value from the dynamic mechanism model. This represents the energy consumption prediction value of the error compensation model; The weights are adaptively updated based on the prediction error within the sliding time window, and the update formula is: ； in, The standard deviation of the prediction error for the dynamic mechanism model. The standard deviation of the prediction error for the error compensation model.

6. The method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms according to claim 5, characterized in that, In S3, the entire industrial production cycle is divided into three phases: startup and adaptation, stable operation, and load reduction and shutdown. To address the objectives of each phase, the key parameter weights and optimization constraints of each sub-model in the fusion model are adjusted using the following strategies: Increase the weights of the heat transfer and proton transfer model during the initial adaptation phase. Optimize heating rate and fuel consumption ; During the stable operation phase, the weight coefficients of each sub-model are balanced, and the reaction temperature is optimized. Duration of stay Compared to steam; During the load reduction and shutdown phase, increase the weight of the equipment loss sub-model. Optimize the load reduction rate and cooling rate .

7. The method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms according to claim 6, characterized in that, In S4, the reward function in the Deep Q-Network (DQN) algorithm The expression is as follows: ； in, For energy consumption reduction rate, For product yield, The reaction temperature. To set the temperature, , , These are the weighting coefficients; Action value function The expression is as follows: ； in, This is the current state. For the current action, For instant rewards, As a discount factor, For the next state, For the next action, Value of network actions for the target; The improved Tianying optimization algorithm aims to minimize energy consumption per unit product and maximize product yield, using production process parameters as constraints, and constructs a fitness function: ； in, Energy consumption per unit of product , These are positive and negative adjustment coefficients, used to balance the optimization weights of reducing energy consumption and stabilizing product yield. The yield is the product yield.

8. The method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms according to claim 7, characterized in that, In S4, the algorithm-mechanism model co-optimization strategy includes the following steps: Step 1: Adjust the algorithm parameters based on the prediction error of the fusion model: When 3% < When the mutation rate is ≤5%, increase the mutation probability or explorer ratio of the improved Skyhawk optimization algorithm; when When the percentage is greater than 5%, the majority of individuals in the population are moved into the process constraint boundary, the global optimization weight is increased, and the fusion model weight is simultaneously triggered for emergency update. Step 2: Embed the output of the dynamic mechanism model into the fitness function. The modified fitness function is as follows: ； in, , The penalty coefficient is the mechanism constraint coefficient. Energy consumption per unit of product For product yield, Additional energy consumption due to equipment aging. Components The rate of change of concentration over time; increased during the start-up adaptation phase and the load reduction shutdown phase. Increase during stable operation phase ; Establish Weights of the heat transfer and proton transfer model Equipment loss sub-model weights Linear relationship: ； in, for The linkage coefficient, for The linkage coefficient, Based on the offset.

9. The method for optimizing energy consumption throughout the entire lifecycle of industrial production processes based on dynamic mechanisms as described in claim 8, characterized in that, S5 includes the following steps: S51. Collect optimized real-time industrial production data and calculate the actual prediction error of the fusion model using the following formula. : in, The energy consumption prediction value of the fusion model. This represents the actual energy consumption value collected in real time. S52, Based on prediction error Adjust the size based on feedback: If the actual prediction error 2.8%, keep the existing parameters unchanged, and continue to perform dynamic parameter optimization of S4; If the actual prediction error is 2.8% < 3%, keep the existing parameters unchanged, and continue the conventional optimization through S4 only; If the actual prediction error Based on the algorithm-mechanism model collaborative optimization strategy, the core parameters of the improved Tianying optimization algorithm in S4 are adjusted, and the adaptive weights of the fusion model are updated synchronously. , ; The target network parameters of the Deep Q-Network (DQN) algorithm are iteratively adjusted, and the weight coefficients of the reward function are also adjusted. ; Correcting and improving the weight coefficients of the Skyhawk optimization algorithm Adjustment coefficient of fitness function , ; S53. Repeat the optimization steps S2-S4 until the prediction error is reached. If the value is ≤2.8%, stop the iteration and use the updated parameters as the new base parameters to continue the full-cycle energy consumption optimization.

10. A dynamic mechanism-based full-cycle energy consumption optimization system for industrial production processes, used to implement the method described in any one of claims 1-9, characterized in that, include: The data acquisition module consists of distributed sensors, online analyzers, and metering equipment, and is used to collect multi-dimensional production data. The data preprocessing module is used for data cleaning, feature extraction and normalization, and outputs a standardized dataset. The fusion model building module integrates the dynamic mechanism model unit, LSTM error compensation unit, and weight adaptive unit to output fusion prediction results; the stage adaptation module is used to divide the production stage and adjust the model parameters and constraints accordingly. The intelligent optimization module integrates the Deep Q-Network (DQN) algorithm and the improved Skyhawk optimization algorithm, and outputs optimization operation parameters; the closed-loop control module is used to verify the optimization effect based on real-time data and trigger iterative updates of model and algorithm parameters. The output module is used to transmit optimized parameters to the production control system (DCS) to achieve automated control.

Images