AI-based industrial energy-saving intelligent optimization control system

Abstract

The application discloses an AI-based industrial energy-saving intelligent optimization control system, which comprises a data acquisition and processing module, a working condition semantic construction module, an energy consumption deduction module, an energy-saving evaluation module and a constraint manifold control module.The data acquisition and processing module acquires and processes multi-source operation data and constructs an industrial operation data set.The working condition semantic construction module inputs an improved DA-RNN network, filters key working condition characteristics and constructs a deformable working condition semantic structure.The energy consumption deduction module generates a candidate control action set, calculates baseline energy consumption and deduces predicted energy consumption results.The energy-saving evaluation module calculates energy-saving benefits, filters combined candidate actions and solves a minimum intervention control set.The constraint manifold control module constructs a dynamic soft constraint manifold, projects control actions and generates safe control actions.The control execution update module executes safe control actions, acquires real-time data and updates the structure and the constraint model.The application realizes industrial energy-saving optimization control by constructing a working condition semantic structure through an improved DA-RNN and combining counterfactual energy consumption deduction and dynamic constraint control.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent control and energy optimization technology, and in particular to an AI-based industrial energy-saving intelligent optimization control system. Background Technology

[0002] With the continuous expansion of industrial production scale, high-energy-consuming industries such as steel, chemicals, power, cement, and manufacturing have placed higher demands on energy efficiency. To reduce production energy consumption and improve operational efficiency, industrial sites typically use distributed control systems or programmable logic controllers (PLCs) to adjust equipment operating parameters, such as regulating valve opening, pump speed, fan frequency, and temperature setpoints to achieve energy consumption control. In recent years, with the development of data acquisition technology and the Industrial Internet, industrial systems can continuously collect large amounts of operational data. This allows for the introduction of data-driven methods for energy consumption prediction and operational optimization, making data-driven energy-saving control a crucial technical means for industrial energy management.

[0003] In existing technologies, some methods analyze the operating status of industrial systems by establishing energy consumption prediction models and optimize equipment operating parameters based on the prediction results; other methods utilize machine learning or deep learning algorithms to model industrial operating data to achieve energy consumption prediction or energy-saving control. However, industrial systems are typically characterized by frequent changes in operating conditions, complex equipment coupling relationships, and diverse energy consumption formation mechanisms. Traditional methods often struggle to simultaneously characterize the dynamic correlations between multiple variable operating conditions, resulting in poor model adaptability under different operating conditions. Existing technologies often directly make control decisions based on predicted energy consumption when optimizing energy consumption, making it difficult to distinguish between the actual energy-saving effect brought about by control actions and the energy consumption changes caused by operating condition fluctuations, thus affecting the accuracy of energy-saving strategies.

[0004] Industrial control systems are typically constrained by various factors, including equipment operating boundaries, process coupling relationships, rates of change of control variables, and safety interlocks. Existing technologies often lack a unified model of these constraints when generating control strategies, making it difficult to directly apply the optimized control actions to actual industrial systems. Particularly in scenarios involving multiple devices operating collaboratively, if the control strategy does not adequately consider system-level constraints and process coupling relationships, it can easily lead to control actions becoming unexecutable or affecting production stability.

[0005] Therefore, how to provide an AI-based intelligent optimization control system for industrial energy conservation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an AI-based intelligent optimization control system for industrial energy conservation. This invention generates a deformable semantic structure of operating conditions by constructing an improved DA-RNN network, and combines baseline energy consumption calculation under counterfactual conditions, energy consumption deduction of candidate control actions, and dynamic soft-constraint manifold projection control mechanism to uniformly model changes in industrial system operating conditions, energy consumption formation mechanisms, and control constraint relationships. This enables intelligent generation and safe execution of energy-saving control actions, thereby possessing the advantages of strong adaptability to complex operating conditions, high accuracy of energy-saving strategies, and strong executability of control actions.

[0007] According to an embodiment of the present invention, an AI-based industrial energy-saving intelligent optimization control system includes: The data acquisition and processing module collects multi-source operational data from the industrial production process, processes the multi-source operational data, and constructs an industrial operational data set. The working condition semantic construction module inputs the industrial operation data set into the improved DA-RNN network, filters key working condition features related to energy consumption changes through the input attention mechanism, models the historical working condition time series relationship through the time attention mechanism, generates multi-scale working condition semantic representation, and constructs a deformable working condition semantic structure. The energy consumption prediction module generates a set of candidate control actions based on the deformable working condition semantic structure and the set of adjustable control variables of the equipment. Under counterfactual conditions, it calculates the baseline energy consumption value of the current working condition, performs energy consumption prediction on the candidate control actions, and obtains the predicted energy consumption value corresponding to each candidate control action. The energy-saving assessment module determines the energy-saving benefits corresponding to each candidate control action based on the difference between the baseline energy consumption value and the predicted energy consumption value. Under the condition of meeting the target energy-saving requirements, it screens and combines the candidate control actions to solve the minimum set of intervention control actions required to achieve the target energy-saving level. The constrained manifold control module constructs a dynamic soft-constrained manifold based on the deformable working condition semantic structure, equipment operating boundary, process coupling relationship, upper and lower limits of control variables, control variable change rate constraints, and safety interlock constraints. The set of minimum intervention control actions is projected onto the dynamic soft-constrained manifold to obtain the safety control actions. The control execution update module sends safety control actions to the industrial control system to perform energy-saving adjustments, collects real-time operating condition data and actual energy consumption data after execution, and updates the deformable operating condition semantic structure and dynamic soft constraint manifold.

[0008] Optionally, the multi-source operating data includes process parameters, equipment status parameters, energy consumption parameters, environmental parameters, and historical control action data.

[0009] Optionally, the processing of multi-source operational data includes time alignment, outlier removal, missing value completion, and normalization of the multi-source operational data.

[0010] Optionally, the working condition semantic construction module includes: The industrial operation data set is organized into sequential data according to time order, and a multi-dimensional feature vector containing process operation variables, equipment status variables, energy consumption measurement variables and control variables is constructed for each time step; The multidimensional feature vectors at each time step are input into the improved DA-RNN network, which includes a feature embedding layer, a group input attention module, a dual temporal coding module, a temporal attention fusion module, and a working condition semantic structure update module. The feature embedding layer performs linear mapping and normalization on process operation variables, equipment status variables and energy consumption measurement variables respectively to obtain process feature sub-vectors, equipment status feature sub-vectors and energy consumption feature sub-vectors. The three types of sub-vectors are concatenated in the feature channel dimension to form a working condition feature representation sequence. The sequence of operating condition feature representations is input into the attention module in groups. The first input attention head is used to perform weighted calculation on the energy consumption feature sub-vector to obtain the key energy consumption features. The second input attention head is used to perform weighted calculation on the process feature sub-vector and the equipment status feature sub-vector to obtain the operating condition background features. The key energy consumption features and the operating condition background features are concatenated to form the key operating condition input sequence. The key operating condition input sequence is input into the dual time-series coding module, which includes a short-term time-series coding sub-layer and a long-term time-series coding sub-layer. The short-term time-series coding sub-layer extracts the rapid operating condition fluctuation features, and the long-term time-series coding sub-layer extracts the operating condition trend features. The two types of features are input into the time attention module, and weighted fusion is performed according to the attention weights of different time scales to generate a multi-scale operating condition semantic representation. The multi-scale working condition semantic representation, the working condition semantic structure state, and the preset equipment topology relationship are input into the working condition semantic structure update module. Feature combination and nonlinear transformation are performed through multiple fully connected layers and normalization layers to generate a deformable working condition semantic structure for the current time step.

[0011] Optionally, the energy consumption prediction module includes: Based on the deformable working condition semantic structure and the set of adjustable control variables of the equipment at the current moment, a set of candidate control actions is generated. According to the equipment to which the control variable belongs, the direction of action of the control variable, and the linkage relationship of the control variable, the candidate control actions in the set of candidate control actions are organized hierarchically to form a three-level action sequence of equipment-level candidate actions, process-level candidate actions, and system-level candidate actions. Align the deformable operating condition semantic structure at the current moment with the actual control action at the previous control moment, extract the equipment action path, process transmission path and energy consumption response path corresponding to the control action at the previous control moment, solidify the equipment action path, process transmission path and energy consumption response path into the baseline action chain at the current moment, and under the counterfactual condition of keeping the baseline action chain unchanged and keeping the control action at the previous control moment unchanged, replay the energy consumption transmission process of the current operating condition at the equipment layer, process layer and system layer level by level to obtain the baseline energy consumption value of the current operating condition; For each candidate control action in the candidate control action set, a candidate intervention chain is constructed in the same order as the baseline action chain: equipment layer, process layer, and system layer. The candidate intervention chain is then embedded into the deformable operating condition semantic structure at the current moment. The local equipment state changes, local process parameter changes, and overall energy consumption transfer changes caused by the candidate control actions are analyzed in layers to obtain the layered energy consumption response results corresponding to each candidate control action. The hierarchical energy consumption response results are aggregated layer by layer according to the transmission order from the device layer to the system layer. The working condition offset caused by the candidate control action is recursively updated in combination with the deformable working condition semantic structure at the current moment to obtain the predicted energy consumption value corresponding to each candidate control action. The baseline energy consumption value is associated with the predicted energy consumption value corresponding to each candidate control action to form a one-to-one correspondence between the candidate control action and the energy consumption result.

[0012] Optionally, the energy-saving assessment module includes: Read the baseline energy consumption value and the predicted energy consumption value corresponding to each candidate control action. Compare the baseline energy consumption value with the predicted energy consumption value corresponding to each candidate control action one by one to determine the energy saving benefit corresponding to each candidate control action, and form a benefit ranking sequence of candidate control actions according to the size of the energy saving benefit. Based on the preset target energy saving requirements, the benefit ranking sequence is initially screened, and candidate control actions that achieve the target energy saving benefits are retained. The control variable change path, the set of affected equipment, and the energy consumption response level corresponding to each candidate control action are extracted to form a subset of target energy saving actions. For each candidate control action in the target energy-saving action subset, an intervention transmission chain is constructed according to the order of changes in control variables. Each control variable in the intervention transmission chain is tested back one by one. That is, while keeping the other control variables unchanged, the adjustment of a single control variable is removed in turn, and the energy-saving benefit after removing the control variable is recalculated. The core control variables necessary to maintain the target energy-saving result are determined, and the minimum necessary intervention subset of the corresponding candidate control action is formed. The minimum necessary intervention subset is hierarchically reorganized according to the equipment layer, process layer and system layer, and intra-layer merging and inter-layer trimming are performed in sequence to generate the simplest intervention chain that meets the target energy saving requirements. By comparing the number of control variables, the total adjustment range of control variables, the number of equipment linkages, and the stability of energy-saving benefits corresponding to each simplest intervention chain, the simplest intervention chain with the fewest number of control variables, the smallest total adjustment range, and the highest stability of energy-saving benefits is selected under the condition of achieving the target energy-saving requirements. The control actions corresponding to the simplest intervention chain are determined as the minimum set of intervention control actions required to achieve the target energy-saving level.

[0013] Optionally, the constrained manifold control module includes: Read the deformable working condition semantic structure and the set of minimum intervention control actions, extract the equipment status information, process association information and energy consumption response level information corresponding to the current working condition from the deformable working condition semantic structure, and combine them with the equipment operating boundary, upper and lower limits of control variables, control variable change rate limit and safety interlock limit to generate the basic constraint set under the current working condition; The basic constraint set is mapped hierarchically according to the equipment layer, process layer and system layer. Single equipment control boundary constraints are formed at the equipment layer, variable coupling constraints are formed at the process layer, and total energy consumption disturbance constraints are formed at the system layer. Based on the connection strength between each constraint object in the deformable working condition semantic structure, the constraints of each layer are associated and sorted to construct a dynamic soft constraint skeleton. The dynamic soft-constraint skeleton is input into the manifold generation module, which includes boundary contraction elements, coupled bending elements, and margin adjustment elements, wherein: The boundary contraction unit adaptively compresses the allowable adjustment range of the corresponding control variable based on the degree to which the equipment is close to the upper or lower limit of operation under the current operating conditions. The coupled bending unit continuously adjusts the connection form of the constraint boundary according to the linkage strength between process variables; The margin adjustment unit dynamically corrects the safety margin of the overall constraint space based on the risk of safety interlock triggering and the intensity of energy consumption fluctuations, generating a dynamic soft constraint manifold that adjusts in real time with changes in operating conditions. The set of minimum intervention control actions is executed in the hierarchical order of the dynamic soft constraint manifold, including equipment-level feasibility correction, process-level coupling consistency correction, and system-level safety margin correction. This ensures that each control action conforms to the dynamic soft constraint manifold layer by layer while maintaining the original energy-saving direction, thus obtaining the corresponding safety control action. Constraint compliance and stability tests are performed on each safety control action, and safety control actions that simultaneously meet the requirements of dynamic soft constraint manifold constraints and the execution requirements of the industrial control system are retained.

[0014] Optionally, the control execution update module includes: The safety control actions are sent to the industrial control system, where the programmable logic controller or distributed control system performs control and adjustment on the corresponding equipment. The safety control actions include the adjustment values ​​of the corresponding control variables for each equipment. After the safety control actions are executed, real-time operating data and actual energy consumption data of the industrial system are collected, the collected data are processed for time synchronization, and an updated industrial operating data set is formed. The updated industrial operation data set is input into the improved DA-RNN network. The multi-scale working condition semantic representation at the current moment is recalculated through the feature embedding layer, group input attention module, dual temporal coding module and temporal attention fusion module. The deformable working condition semantic structure is updated according to the multi-scale working condition semantic representation. The actual energy consumption data after the execution of safety control actions is compared with the predicted energy consumption value to determine the energy consumption prediction deviation. Based on the energy consumption prediction deviation, the key parameters of the improved DA-RNN network are iteratively adjusted to update the calculation results of the working condition semantic representation. The constraint boundaries and safety margins of the dynamic soft-constrained manifold are readjusted based on the updated deformable working condition semantic structure, so that the dynamic soft-constrained manifold can be adaptively updated as the current working condition changes, and the updated deformable working condition semantic structure and dynamic soft-constrained manifold are used as inputs for the next control cycle.

[0015] The beneficial effects of this invention are: This invention proposes an AI-based intelligent optimization control method for industrial energy conservation. By inputting industrial operating data into an improved DA-RNN network, and utilizing input attention and time attention mechanisms to construct a deformable semantic structure of operating conditions, a unified representation of the multivariate operating condition relationships and temporal evolution characteristics of industrial systems is achieved. Compared with traditional methods that rely on fixed features or single prediction models, this invention can continuously characterize the dynamic correlation between equipment status, process variables, and energy consumption in industrial environments with frequently changing operating conditions. This enhances the adaptability and stability of the energy conservation optimization model under complex operating conditions, enabling the energy conservation control strategy to maintain high accuracy under different production conditions.

[0016] This invention calculates the baseline energy consumption under counterfactual conditions and extrapolates the energy consumption of candidate control actions, establishing a comparison between baseline and predicted energy consumption. This allows for accurate identification of the actual energy-saving benefits brought by control actions. Compared to existing technologies that directly rely on predicted energy consumption for optimization decisions, this invention effectively distinguishes between energy consumption changes caused by operating condition fluctuations and energy-saving effects generated by control strategies. This makes energy-saving benefit assessment more objective and reliable. Furthermore, by solving for the minimum intervention control action set, it reduces the adjustment amplitude of control variables while meeting target energy-saving requirements, minimizing disturbances to industrial production processes and improving the operability of energy-saving control strategies.

[0017] This invention constructs a dynamic soft-constraint manifold to uniformly model equipment operating boundaries, process coupling relationships, upper and lower limits of control variables, rate of change limits, and safety interlock constraints. Minimum intervention control actions are projected onto this dynamic soft-constraint manifold to generate safe control actions, ensuring that the generated control strategy meets the actual execution conditions of the industrial control system. By continuously collecting real-time operating condition data and energy consumption data after control execution and updating the deformable operating condition semantic structure and the dynamic soft-constraint manifold, this invention can form a continuously adaptive energy-saving control closed loop, achieving long-term stable optimization of energy consumption in industrial systems. It boasts advantages such as high control stability, reliable energy-saving effects, and strong engineering application feasibility. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows the AI-based intelligent optimization control system for industrial energy saving proposed in this invention. Figure 2 This is a schematic diagram of the structure of the improved DA-RNN network for constructing a deformable semantic structure for the AI-based industrial energy-saving intelligent optimization control system proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 AI-based industrial energy-saving intelligent optimization control system, including: The data acquisition and processing module collects multi-source operational data from the industrial production process, processes the multi-source operational data, and constructs an industrial operational data set. The working condition semantic construction module inputs the industrial operation data set into the improved DA-RNN network, filters key working condition features related to energy consumption changes through the input attention mechanism, models the historical working condition time series relationship through the time attention mechanism, generates multi-scale working condition semantic representation, and constructs a deformable working condition semantic structure. The energy consumption prediction module generates a set of candidate control actions based on the deformable working condition semantic structure and the set of adjustable control variables of the equipment. Under counterfactual conditions, it calculates the baseline energy consumption value of the current working condition, performs energy consumption prediction on the candidate control actions, and obtains the predicted energy consumption value corresponding to each candidate control action. The energy-saving assessment module determines the energy-saving benefits corresponding to each candidate control action based on the difference between the baseline energy consumption value and the predicted energy consumption value. Under the condition of meeting the target energy-saving requirements, it screens and combines the candidate control actions to solve the minimum set of intervention control actions required to achieve the target energy-saving level. The constrained manifold control module constructs a dynamic soft-constrained manifold based on the deformable working condition semantic structure, equipment operating boundary, process coupling relationship, upper and lower limits of control variables, control variable change rate constraints, and safety interlock constraints. The set of minimum intervention control actions is projected onto the dynamic soft-constrained manifold to obtain the safety control actions. The control execution update module sends safety control actions to the industrial control system to perform energy-saving adjustments, collects real-time operating condition data and actual energy consumption data after execution, and updates the deformable operating condition semantic structure and dynamic soft constraint manifold.

[0021] In this embodiment, the multi-source operating data includes process parameters, equipment status parameters, energy consumption parameters, environmental parameters, and historical control action data.

[0022] In this embodiment, the processing of multi-source operational data includes time alignment, outlier removal, missing value completion, and normalization of the multi-source operational data.

[0023] In this embodiment, the working condition semantic construction module includes: The industrial operation data set is organized into sequential data according to time order, and a multi-dimensional feature vector containing process operation variables, equipment status variables, energy consumption measurement variables and control variables is constructed for each time step; The multidimensional feature vectors at each time step are input into the improved DA-RNN network, which includes a feature embedding layer, a group input attention module, a dual temporal coding module, a temporal attention fusion module, and a working condition semantic structure update module. The feature embedding layer performs linear mapping and normalization on process operation variables, equipment status variables and energy consumption measurement variables respectively to obtain process feature sub-vectors, equipment status feature sub-vectors and energy consumption feature sub-vectors. The three types of sub-vectors are concatenated in the feature channel dimension to form a working condition feature representation sequence. The sequence of operating condition feature representations is input into the attention module in groups. The first input attention head performs weighted calculations on the energy consumption feature sub-vectors to obtain key energy consumption features. The second input attention head performs weighted calculations on the process feature sub-vectors and equipment status feature sub-vectors to obtain background operating condition features. The key energy consumption features and the background operating condition features are concatenated to form the key operating condition input sequence, where: The energy consumption feature sub-vectors are weighted using the first input attention head, specifically as follows: The energy consumption feature sub-vectors are input into the energy consumption attention calculation unit. The corresponding query vector, key vector and value vector are generated through linear mapping. The relevance score between the query vector and the key vector is calculated. The relevance score is normalized to obtain the energy consumption attention weight. The energy consumption feature sub-vectors are weighted and summed according to the energy consumption attention weight to obtain the key energy consumption features. The second input attention head is used to perform weighted calculations on the process feature sub-vector and the equipment state feature sub-vector, specifically as follows: The process feature vector and the equipment status feature vector are combined along the feature dimension to form the operating condition feature vector. The operating condition feature vector is input into the operating condition attention calculation unit, and a query vector, key vector and value vector are generated through linear mapping. The correlation score between the query vector and the key vector is calculated, and the correlation score is normalized to obtain the operating condition attention weight. The operating condition feature vector is weighted and summed according to the operating condition attention weight to obtain the operating condition background feature. The key operating condition input sequence is fed into a dual-time-series coding module, which includes a short-term time-series coding sub-layer and a long-term time-series coding sub-layer. The short-term time-series coding sub-layer extracts rapid operating condition fluctuation features, while the long-term time-series coding sub-layer extracts operating condition trend features. These two types of features are then input into a time-attention module, where they are weighted and fused according to the attention weights at different time scales to generate a multi-scale operating condition semantic representation. The short-term temporal coding sublayer extracts rapid operating condition fluctuation features, specifically: The key operating condition input sequence is divided into sliding segments according to a preset short time window. Within each time window, the operating condition features of adjacent time steps are continuously encoded. The features within the window are recursively updated through a gated loop unit to obtain short-term time series features that reflect the magnitude and speed of changes in operating conditions within a short time scale, thus forming rapid operating condition fluctuation features. The long-term time-series coding sublayer extracts operating condition trend features, specifically: The key operating condition input sequence is aggregated according to multiple continuous time windows, the feature vectors in each time window are sequence encoded, and the correlation between different time windows is progressively updated through long-term dependency modeling units to obtain long-term time series features that reflect the changing trend and evolution direction of operating conditions over a long time scale, thus forming operating condition trend features. The multi-scale working condition semantic representation, the working condition semantic structure state, and the preset equipment topology relationships are input into the working condition semantic structure update module. Feature combination and nonlinear transformation are performed through multiple fully connected layers and normalization layers to generate a deformable working condition semantic structure for the current time step, wherein: The preset equipment topology is a pre-constructed equipment association structure based on the process flow connection relationship, energy transfer relationship and control signal transmission relationship between each piece of equipment. It is used to represent the correspondence between upstream and downstream equipment, the direction of energy flow and the path of control variable action between different equipment. Generate a deformable semantic structure for the current time step, specifically: The multi-scale working condition semantic representation is concatenated with the working condition semantic structure state of the previous time step in the feature dimension, and the feature vectors corresponding to each device node are associated and mapped in combination with the preset device topology relationship to obtain a structural feature representation that includes device node features and device association relationship. The structural feature representation is input into a multi-layer fully connected layer for feature combination and nonlinear transformation. The feature reconstruction is performed on the operating condition influence relationship and energy consumption coupling relationship between different device nodes to obtain the updated semantic features of the device nodes. The updated semantic features of the device nodes are scaled through a normalization layer, and the connection structure between the device nodes is reorganized according to the preset device topology to generate a deformable semantic structure of the current time step.

[0024] In this embodiment, the energy consumption prediction module includes: Based on the deformable working condition semantic structure and the set of adjustable control variables of the equipment at the current moment, a set of candidate control actions is generated. According to the equipment to which the control variable belongs, the direction of action of the control variable, and the linkage relationship of the control variable, the candidate control actions in the set of candidate control actions are organized hierarchically to form a three-level action sequence of equipment-level candidate actions, process-level candidate actions, and system-level candidate actions. Align the deformable operating condition semantic structure at the current moment with the actual control action at the previous control moment, extract the equipment action path, process transmission path and energy consumption response path corresponding to the control action at the previous control moment, solidify the equipment action path, process transmission path and energy consumption response path into the baseline action chain at the current moment, and under the counterfactual condition of keeping the baseline action chain unchanged and keeping the control action at the previous control moment unchanged, replay the energy consumption transmission process of the current operating condition at the equipment layer, process layer and system layer level by level to obtain the baseline energy consumption value of the current operating condition; For each candidate control action in the candidate control action set, a candidate intervention chain is constructed in the same order as the baseline action chain: equipment layer, process layer, and system layer. The candidate intervention chain is then embedded into the deformable operating condition semantic structure at the current moment. The local equipment state changes, local process parameter changes, and overall energy consumption transfer changes caused by the candidate control actions are analyzed in layers to obtain the layered energy consumption response results corresponding to each candidate control action. The hierarchical energy consumption response results are aggregated layer by layer according to the transmission order from the device layer to the system layer. The working condition offset caused by the candidate control action is recursively updated in combination with the deformable working condition semantic structure at the current moment to obtain the predicted energy consumption value corresponding to each candidate control action. The baseline energy consumption value is associated with the predicted energy consumption value corresponding to each candidate control action to form a one-to-one correspondence between the candidate control action and the energy consumption result.

[0025] In this embodiment, the energy-saving assessment module includes: Read the baseline energy consumption value and the predicted energy consumption value corresponding to each candidate control action. Compare the baseline energy consumption value with the predicted energy consumption value corresponding to each candidate control action one by one to determine the energy saving benefit corresponding to each candidate control action, and form a benefit ranking sequence of candidate control actions according to the size of the energy saving benefit. The benefit ranking sequence is initially screened according to the preset target energy saving requirements. Candidate control actions that achieve the target energy saving benefits are retained. The control variable change path, the set of affected equipment, and the energy consumption response level corresponding to each candidate control action are extracted to form a subset of target energy saving actions. The preset target energy saving requirements are the minimum energy saving benefits set by the industrial production system under the current production task. They are used to limit the energy saving benefits achieved by the candidate control actions after implementation to not be lower than the preset energy saving threshold. The preset energy saving threshold is determined in advance based on the target energy consumption per unit product, the current production load, the equipment operating status, and the production stability requirements. For each candidate control action in the target energy-saving action subset, an intervention transmission chain is constructed according to the order of changes in control variables. Each control variable in the intervention transmission chain is tested back one by one. That is, while keeping the other control variables unchanged, the adjustment of a single control variable is removed in turn, and the energy-saving benefit after removing the control variable is recalculated. The core control variables necessary to maintain the target energy-saving result are determined, and the minimum necessary intervention subset of the corresponding candidate control action is formed. The minimum necessary intervention subset is hierarchically reorganized according to the equipment layer, process layer and system layer, and intra-layer merging and inter-layer trimming are performed in sequence. Intra-layer merging is used to merge control variables that act on the same equipment and have the same adjustment direction, and inter-layer trimming is used to delete duplicate intervention quantities that can be achieved by passing them from the control variables of the previous level, so as to generate the simplest intervention chain that meets the target energy saving requirements. By comparing the number of control variables, the total adjustment range of control variables, the number of equipment linkages, and the stability of energy-saving benefits corresponding to each simplest intervention chain, the simplest intervention chain with the fewest number of control variables, the smallest total adjustment range, and the highest stability of energy-saving benefits is selected under the condition of achieving the target energy-saving requirements. The control actions corresponding to the simplest intervention chain are determined as the minimum set of intervention control actions required to achieve the target energy-saving level.

[0026] In this embodiment, the constrained manifold control module includes: Read the deformable working condition semantic structure and the set of minimum intervention control actions, extract the equipment status information, process association information and energy consumption response level information corresponding to the current working condition from the deformable working condition semantic structure, and combine them with the equipment operating boundary, upper and lower limits of control variables, control variable change rate limit and safety interlock limit to generate the basic constraint set under the current working condition; The basic constraint set is mapped hierarchically according to the equipment layer, process layer, and system layer. Single-equipment control boundary constraints are formed at the equipment layer, variable coupling constraints at the process layer, and total energy consumption disturbance constraints at the system layer. Based on the connection strength between constraint objects in the deformable operating condition semantic structure, the constraints at each layer are correlated and sorted to construct a dynamic soft constraint skeleton, wherein: A single device control boundary constraint is formed at the device layer, specifically as follows: Based on the allowable value range, upper limit of operation, lower limit of operation, upper limit of adjustment rate, and safety interlock boundary of the control variables corresponding to each device, an independent constraint interval is established for the control variables of each device, and the current state of the control variables of each device is matched with the independent constraint interval to form the single device control boundary constraint of the corresponding device. Variable coupling constraints are formed at the process level, specifically as follows: Based on the linkage, transmission and mutual constraint relationships among various process variables in industrial production, the temperature, pressure and flow rate changes and load changes caused by different equipment control variables are correlated and mapped to establish linkage limit ranges among multiple process variables, forming variable coupling constraints. A total energy consumption disturbance constraint is formed at the system level, specifically as follows: Based on the allowable fluctuation range of the overall energy consumption of the industrial system under the current operating conditions, the target energy consumption requirements per unit product, and the production stability requirements, a unified constraint is imposed on the overall energy consumption change range that may be caused by each candidate control action, limiting the total energy consumption disturbance of the system after the control action is implemented to not exceed the preset allowable range, thus forming a total energy consumption disturbance constraint. The dynamic soft-constraint skeleton is input into the manifold generation module, which includes boundary contraction elements, coupled bending elements, and margin adjustment elements, wherein: The boundary contraction unit adaptively compresses the allowable adjustment range of the corresponding control variable based on the degree to which the equipment is close to the upper or lower limit of operation under the current operating conditions. The coupled bending unit continuously adjusts the connection form of the constraint boundary according to the linkage strength between process variables; The margin adjustment unit dynamically corrects the safety margin of the overall constraint space based on the risk of safety interlock triggering and the intensity of energy consumption fluctuations, generating a dynamic soft constraint manifold that adjusts in real time with changes in operating conditions. The safety margin refers to the allowable range of change reserved for the adjustment of control variables under the premise of meeting the equipment operating boundary constraints, process coupling constraints, and safety interlock constraints. It is used to ensure that even if there are fluctuations in operating conditions or measurement errors during the implementation of control actions, the equipment protection or safety interlock will not be triggered, thus maintaining the stable operation of the industrial system. The set of minimum intervention control actions is executed in the hierarchical order of the dynamic soft constraint manifold, including equipment-level feasibility correction, process-level coupling consistency correction, and system-level safety margin correction. This ensures that each control action conforms to the dynamic soft constraint manifold layer by layer while maintaining the original energy-saving direction, thus obtaining the corresponding safety control action. Constraint compliance and stability tests are performed on each safety control action, and safety control actions that simultaneously meet the requirements of dynamic soft constraint manifold constraints and the execution requirements of the industrial control system are retained.

[0027] In this embodiment, the control execution update module includes: The safety control actions are sent to the industrial control system, where the programmable logic controller or distributed control system performs control and adjustment on the corresponding equipment. The safety control actions include the adjustment values ​​of the corresponding control variables for each equipment. After the safety control actions are executed, real-time operating data and actual energy consumption data of the industrial system are collected, the collected data are processed for time synchronization, and an updated industrial operating data set is formed. The updated industrial operation data set is input into the improved DA-RNN network. The multi-scale working condition semantic representation at the current moment is recalculated through the feature embedding layer, group input attention module, dual temporal coding module and temporal attention fusion module. The deformable working condition semantic structure is updated according to the multi-scale working condition semantic representation. The actual energy consumption data after the execution of safety control actions is compared with the predicted energy consumption value to determine the energy consumption prediction deviation. Based on the energy consumption prediction deviation, the key parameters of the improved DA-RNN network are iteratively adjusted to update the calculation results of the working condition semantic representation. The constraint boundaries and safety margins of the dynamic soft-constrained manifold are readjusted based on the updated deformable working condition semantic structure, so that the dynamic soft-constrained manifold can be adaptively updated as the current working condition changes, and the updated deformable working condition semantic structure and dynamic soft-constrained manifold are used as inputs for the next control cycle.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a large cement production enterprise with a new dry-process cement production line producing 5,000 tons of clinker per day. This production line includes a raw material mill system, a preheater system, a rotary kiln system, and a grate cooler system. The rotary kiln and grate cooler systems account for over 70% of the total energy consumption of the entire production line. During actual production, the enterprise found that system energy consumption changed significantly when the raw material moisture content changed, the kiln tail temperature fluctuated, or the production load was adjusted. Traditional control methods mainly rely on empirical parameters for adjustment, such as manually controlling parameters like fan frequency, fuel injection rate, and grate cooler airflow. This method can maintain stable production when operating conditions change slightly, but when the production load changes significantly, the control strategy is difficult to adjust in time, easily leading to decreased energy efficiency. There is a clear coupling relationship between multiple devices; for example, changes in the kiln tail fan frequency affect the kiln temperature, which in turn affects fuel consumption and clinker quality. Therefore, optimizing a single device often fails to achieve overall energy savings.

[0029] The AI-based intelligent optimization control method for industrial energy conservation proposed in this invention is deployed on this production line. First, multi-source operational data is collected through the existing distributed control system of the production line. The collected data includes parameters such as rotary kiln temperature, kiln tail fan frequency, grate cooler airflow, fuel injection rate, motor current, production load, and energy consumption per unit clinker. The data sampling period is set to 5 seconds, and the collected data undergoes time alignment, outlier filtering, and normalization to construct an industrial operation data set. Subsequently, the industrial operation data set is input into an improved DA-RNN network. Through an input attention mechanism, key operating condition features with significant impact on energy consumption are automatically filtered, such as kiln tail fan frequency, fuel injection rate, rotary kiln temperature, and production load. Simultaneously, a time attention mechanism is used to model the temporal relationships of historical operating conditions, generating a multi-scale semantic representation of the current operating condition and constructing a deformable semantic structure for the operating condition. This structure can describe the operating relationships between equipment and the energy consumption trends under different operating conditions.

[0030] After obtaining the semantic structure of the deformable operating condition, the system generates a set of candidate control actions based on the adjustable control variables of the equipment. These candidate control actions include adjusting the kiln tail fan frequency, fine-tuning the fuel injection rate, and adjusting the grate cooler airflow. Under the counterfactual condition of keeping the current control actions unchanged, the system first calculates the baseline energy consumption value for the current operating condition, then extrapolates the energy consumption of each candidate control action to obtain the predicted energy consumption value after executing different control actions. By comparing the difference between the baseline energy consumption value and the predicted energy consumption value, the system can determine the energy-saving benefits corresponding to different control actions and select candidate control actions while meeting the energy-saving target. Subsequently, the system solves for the set of minimum intervention control actions according to the principle of minimum intervention, keeping the variation range of control variables within a small range to avoid significant disturbances to the production process. After obtaining the minimum intervention control actions, a dynamic soft-constrained manifold is constructed based on the equipment operating boundary, process coupling relationship, rate of change of control variables, and safety interlocking conditions. The minimum intervention control actions are then projected onto this dynamic soft-constrained manifold to obtain control actions that can be safely executed in the actual industrial system.

[0031] Ultimately, the system sends safety control actions to the rotary kiln system and grate cooler system through a distributed control system, and continuously collects real-time operating data and energy consumption data to update the deformable operating condition semantic structure and dynamic soft constraint manifold, thereby forming an energy-saving optimization control closed loop.

[0032] The system underwent three consecutive months of field testing. During the testing period, the production line produced an average of approximately 5,100 tons of clinker per day, with production load fluctuating between 85% and 100%. By comparing energy consumption data before and after system implementation, it can be found that, while ensuring clinker quality and production stability, the method of this invention can significantly reduce the energy consumption of the production line.

[0033] Table 1. Comparison of energy consumption before and after implementing the method of this invention in a cement clinker production line. Operational phase Unit clinker power consumption (kWh / t) Fuel consumption (kg standard coal / t) Average power of kiln tail blower (kW) Rotary kiln main motor power (kW) Daily clinker production (t) Average value before implementation 92.5 109 1850 2650 5080 First month after implementation 89.7 106 1785 2580 5120 Second month after implementation 88.8 105 1770 2555 5095 3 months after implementation 88.2 104 1765 2540 5105 Average value after implementation 88.9 105 1773 2558 5107 As shown in Table 1, before implementing the method of this invention, the average power consumption per ton of clinker in this cement clinker production line was 92.5 kWh / ton, fuel consumption was approximately 109 kg of standard coal per ton, the average power of the kiln tail blower was approximately 1850 kW, the power of the rotary kiln main motor was approximately 2650 kW, and the daily clinker production was approximately 5080 tons. This data indicates that the production line can maintain stable operation under the traditional experience-based control mode, but the equipment operating parameters are usually adjusted according to fixed rules, failing to fully adapt to changes in actual operating conditions. This results in a certain degree of energy redundancy in the blower system and kiln system, leading to a relatively high overall energy consumption level.

[0034] After deploying the industrial energy-saving intelligent optimization control method proposed in this invention, a deformable semantic structure of operating conditions is constructed by improving the DA-RNN network. Based on counterfactual conditions, the baseline energy consumption and the predicted energy consumption of candidate control actions are calculated. The system can automatically identify energy-saving control strategies under different operating conditions. In the first month after implementation, the unit clinker power consumption decreased to 89.7 kWh per ton, fuel consumption decreased to 106 kg of standard coal per ton, the average power of the kiln tail fan decreased to 1785 kW, and the power of the rotary kiln main motor decreased to 2580 kW. At the same time, the daily clinker production reached 5120 tons, indicating that production efficiency remained stable while reducing energy consumption.

[0035] As the system continues to operate and updates the deformable working condition semantic structure and dynamic soft-constraint manifold through real-time data, the energy-saving effect becomes even more stable. In the second and third months after implementation, the unit clinker power consumption decreased to 88.8 kWh / ton and 88.2 kWh / ton, respectively, and fuel consumption decreased to 105 kg / ton and 104 kg / ton of standard coal, respectively. The three-month average unit clinker power consumption was 88.9 kWh / ton, a decrease of approximately 3.9% compared to before implementation, while daily clinker production remained at around 5100 tons. This indicates that the method of this invention effectively reduces industrial energy consumption without affecting production stability and has good practical application results.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-based intelligent optimization control system for industrial energy conservation, characterized in that, include: The data acquisition and processing module collects multi-source operational data from the industrial production process, processes the multi-source operational data, and constructs an industrial operational data set. The working condition semantic construction module inputs the industrial operation data set into the improved DA-RNN network, filters key working condition features related to energy consumption changes through the input attention mechanism, models the historical working condition time series relationship through the time attention mechanism, generates multi-scale working condition semantic representation, and constructs a deformable working condition semantic structure. The energy consumption prediction module generates a set of candidate control actions based on the deformable working condition semantic structure and the set of adjustable control variables of the equipment. Under counterfactual conditions, it calculates the baseline energy consumption value of the current working condition, performs energy consumption prediction on the candidate control actions, and obtains the predicted energy consumption value corresponding to each candidate control action. The energy-saving assessment module determines the energy-saving benefits corresponding to each candidate control action based on the difference between the baseline energy consumption value and the predicted energy consumption value. Under the condition of meeting the target energy-saving requirements, it screens and combines the candidate control actions to solve the minimum set of intervention control actions required to achieve the target energy-saving level. The constrained manifold control module constructs a dynamic soft-constrained manifold based on the deformable working condition semantic structure, equipment operating boundary, process coupling relationship, upper and lower limits of control variables, control variable change rate constraints, and safety interlock constraints. The set of minimum intervention control actions is projected onto the dynamic soft-constrained manifold to obtain the safety control actions. The control execution update module sends safety control actions to the industrial control system to perform energy-saving adjustments, collects real-time operating condition data and actual energy consumption data after execution, and updates the deformable operating condition semantic structure and dynamic soft constraint manifold.

2. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The multi-source operating data includes process parameters, equipment status parameters, energy consumption parameters, environmental parameters, and historical control action data.

3. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The processing of multi-source operational data includes time alignment, outlier removal, missing value completion, and normalization.

4. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The working condition semantic construction module includes: The industrial operation data set is organized into sequential data according to time order, and a multi-dimensional feature vector containing process operation variables, equipment status variables, energy consumption measurement variables and control variables is constructed for each time step; The multidimensional feature vectors at each time step are input into the improved DA-RNN network, which includes a feature embedding layer, a group input attention module, a dual temporal coding module, a temporal attention fusion module, and a working condition semantic structure update module. The feature embedding layer performs linear mapping and normalization on process operation variables, equipment status variables and energy consumption measurement variables respectively to obtain process feature sub-vectors, equipment status feature sub-vectors and energy consumption feature sub-vectors. The three types of sub-vectors are concatenated in the feature channel dimension to form a working condition feature representation sequence. The sequence of operating condition feature representations is input into the attention module in groups. The first input attention head is used to perform weighted calculation on the energy consumption feature sub-vector to obtain the key energy consumption features. The second input attention head is used to perform weighted calculation on the process feature sub-vector and the equipment status feature sub-vector to obtain the operating condition background features. The key energy consumption features and the operating condition background features are concatenated to form the key operating condition input sequence. The key operating condition input sequence is input into the dual time-series coding module, which includes a short-term time-series coding sub-layer and a long-term time-series coding sub-layer. The short-term time-series coding sub-layer extracts the rapid operating condition fluctuation features, and the long-term time-series coding sub-layer extracts the operating condition trend features. The two types of features are input into the time attention module, and weighted fusion is performed according to the attention weights of different time scales to generate a multi-scale operating condition semantic representation. The multi-scale working condition semantic representation, the working condition semantic structure state, and the preset equipment topology relationship are input into the working condition semantic structure update module. Feature combination and nonlinear transformation are performed through multiple fully connected layers and normalization layers to generate a deformable working condition semantic structure for the current time step.

5. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The energy consumption prediction module includes: Based on the deformable working condition semantic structure and the set of adjustable control variables of the equipment at the current moment, a set of candidate control actions is generated. According to the equipment to which the control variable belongs, the direction of action of the control variable, and the linkage relationship of the control variable, the candidate control actions in the set of candidate control actions are organized hierarchically to form a three-level action sequence of equipment-level candidate actions, process-level candidate actions, and system-level candidate actions. Align the deformable operating condition semantic structure at the current moment with the actual control action at the previous control moment, extract the equipment action path, process transmission path and energy consumption response path corresponding to the control action at the previous control moment, solidify the equipment action path, process transmission path and energy consumption response path into the baseline action chain at the current moment, and under the counterfactual condition of keeping the baseline action chain unchanged and keeping the control action at the previous control moment unchanged, replay the energy consumption transmission process of the current operating condition at the equipment layer, process layer and system layer level by level to obtain the baseline energy consumption value of the current operating condition; For each candidate control action in the candidate control action set, a candidate intervention chain is constructed in the same order as the baseline action chain: equipment layer, process layer, and system layer. The candidate intervention chain is then embedded into the deformable operating condition semantic structure at the current moment. The local equipment state changes, local process parameter changes, and overall energy consumption transfer changes caused by the candidate control actions are analyzed in layers to obtain the layered energy consumption response results corresponding to each candidate control action. The hierarchical energy consumption response results are aggregated layer by layer according to the transmission order from the device layer to the system layer. The working condition offset caused by the candidate control action is recursively updated in combination with the deformable working condition semantic structure at the current moment to obtain the predicted energy consumption value corresponding to each candidate control action. The baseline energy consumption value is associated with the predicted energy consumption value corresponding to each candidate control action to form a one-to-one correspondence between the candidate control action and the energy consumption result.

6. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The energy-saving assessment module includes: Read the baseline energy consumption value and the predicted energy consumption value corresponding to each candidate control action. Compare the baseline energy consumption value with the predicted energy consumption value corresponding to each candidate control action one by one to determine the energy saving benefit corresponding to each candidate control action, and form a benefit ranking sequence of candidate control actions according to the size of the energy saving benefit. Based on the preset target energy saving requirements, the benefit ranking sequence is initially screened, and candidate control actions that achieve the target energy saving benefits are retained. The control variable change path, the set of affected equipment, and the energy consumption response level corresponding to each candidate control action are extracted to form a subset of target energy saving actions. For each candidate control action in the target energy-saving action subset, an intervention transmission chain is constructed according to the order of changes in control variables. Each control variable in the intervention transmission chain is tested back one by one. That is, while keeping the other control variables unchanged, the adjustment of a single control variable is removed in turn, and the energy-saving benefit after removing the control variable is recalculated. The core control variables necessary to maintain the target energy-saving result are determined, and the minimum necessary intervention subset of the corresponding candidate control action is formed. The minimum necessary intervention subset is hierarchically reorganized according to the equipment layer, process layer and system layer, and intra-layer merging and inter-layer trimming are performed in sequence to generate the simplest intervention chain that meets the target energy saving requirements. By comparing the number of control variables, the total adjustment range of control variables, the number of equipment linkages, and the stability of energy-saving benefits corresponding to each simplest intervention chain, the simplest intervention chain with the fewest number of control variables, the smallest total adjustment range, and the highest stability of energy-saving benefits is selected under the condition of achieving the target energy-saving requirements. The control actions corresponding to the simplest intervention chain are determined as the minimum set of intervention control actions required to achieve the target energy-saving level.

7. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The constrained manifold control module includes: Read the deformable working condition semantic structure and the set of minimum intervention control actions, extract the equipment status information, process association information and energy consumption response level information corresponding to the current working condition from the deformable working condition semantic structure, and combine them with the equipment operating boundary, upper and lower limits of control variables, control variable change rate limit and safety interlock limit to generate the basic constraint set under the current working condition; The basic constraint set is mapped hierarchically according to the equipment layer, process layer and system layer. Single equipment control boundary constraints are formed at the equipment layer, variable coupling constraints are formed at the process layer, and total energy consumption disturbance constraints are formed at the system layer. Based on the connection strength between each constraint object in the deformable working condition semantic structure, the constraints of each layer are associated and sorted to construct a dynamic soft constraint skeleton. The dynamic soft-constraint skeleton is input into the manifold generation module, which includes boundary contraction elements, coupled bending elements, and margin adjustment elements, wherein: The boundary contraction unit adaptively compresses the allowable adjustment range of the corresponding control variable based on how close the equipment is to the upper or lower limit of operation under the current operating conditions. The coupled bending unit continuously adjusts the connection form of the constraint boundary according to the linkage strength between process variables; The margin adjustment unit dynamically corrects the safety margin of the overall constraint space based on the risk of safety interlock triggering and the intensity of energy consumption fluctuations, generating a dynamic soft constraint manifold that adjusts in real time with changes in operating conditions. The set of minimum intervention control actions is executed in the hierarchical order of the dynamic soft constraint manifold, including equipment-level feasibility correction, process-level coupling consistency correction, and system-level safety margin correction. This ensures that each control action conforms to the dynamic soft constraint manifold layer by layer while maintaining the original energy-saving direction, thus obtaining the corresponding safety control action. Constraint compliance and stability tests are performed on each safety control action, and safety control actions that simultaneously meet the requirements of dynamic soft constraint manifold constraints and the execution requirements of the industrial control system are retained.

8. The AI-based intelligent optimization control system for industrial energy saving according to claim 1, characterized in that, The control execution update module includes: The safety control actions are sent to the industrial control system, where the programmable logic controller or distributed control system performs control and adjustment on the corresponding equipment. The safety control actions include the adjustment values ​​of the corresponding control variables for each equipment. After the safety control actions are executed, real-time operating data and actual energy consumption data of the industrial system are collected, the collected data are processed for time synchronization, and an updated industrial operating data set is formed. The updated industrial operation data set is input into the improved DA-RNN network. The multi-scale working condition semantic representation at the current moment is recalculated through the feature embedding layer, group input attention module, dual temporal coding module and temporal attention fusion module. The deformable working condition semantic structure is updated according to the multi-scale working condition semantic representation. The actual energy consumption data after the execution of safety control actions is compared with the predicted energy consumption value to determine the energy consumption prediction deviation. Based on the energy consumption prediction deviation, the key parameters of the improved DA-RNN network are iteratively adjusted to update the calculation results of the working condition semantic representation. The constraint boundaries and safety margins of the dynamic soft-constrained manifold are readjusted based on the updated deformable working condition semantic structure, so that the dynamic soft-constrained manifold can be adaptively updated as the current working condition changes, and the updated deformable working condition semantic structure and dynamic soft-constrained manifold are used as inputs for the next control cycle.

Images