Carbon black production equipment energy management system based on internet of things

Abstract

This invention discloses an IoT-based energy management system for carbon black production equipment. The system includes: collecting energy consumption data and production condition data and unifying timestamps to form a multi-source synchronous data stream; performing standardization and resampling to form an online input dataset; constructing a time-delay structure to expand samples and form a causal discovery training sample set; initializing a directed sparse causal gating matrix and time-delay selection parameters and embedding a MambaTS selective state-space long sequence modeling model to form a parameterized association model; performing state recursive calculations to generate a state representation sequence and a model output dataset; and updating the causal gating parameters based on the model output dataset to form an updated state representation sequence. This invention achieves online consistent association modeling of energy consumption data and production condition data, improving the reliability and operational stability of online energy management associations.

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) data acquisition and industrial process energy management technology, and in particular to an IoT-based energy management system for carbon black production equipment. Background Technology

[0002] Carbon black production is a continuous industrial process, with equipment encompassing a reactor system, air and fuel supply system, quenching and collection system, fan and conveying system, and supporting power and utility systems. The production process involves simultaneous energy consumption (electricity, steam, fuel, etc.) and changes in operating conditions (temperature, pressure, flow rate, valve position, speed, load, etc.). Energy consumption data and production condition data differ significantly in time scale, sampling frequency, data accuracy, and measurement links. To address energy management needs, existing technologies generally employ IoT-based data acquisition and industrial control system data access, aggregating data from field metering devices and process instruments to an energy monitoring platform or manufacturing execution system. Based on statistical analysis and mechanistic experience models, energy consumption index systems and operating condition evaluation systems are established, supporting production operation management through energy consumption reports, energy consumption benchmark comparisons, energy consumption anomaly alarms, and operating condition traceability. Some solutions further introduce online data cleaning and resampling processing, utilizing time-series regression and multivariate correlation analysis to achieve correlation modeling between energy consumption and operating conditions, providing a data foundation for energy consumption accounting, energy consumption decomposition, and energy-saving scheduling.

[0003] In practical applications, existing technologies mostly establish online correlations between energy consumption data and production condition data by matching values ​​within the same time slice or by using sliding window-based correlation calculations. However, the carbon black production process suffers from problems such as time lag effects of operating conditions, strong coupling propagation between variables, asynchrony between instruments and metering links, inconsistent sampling periods, data gaps, and abnormal drift. These issues lead to variable alignment deviations and confusion of causal directions during the online correlation process between energy consumption variables and operating condition variables, resulting in distorted correlations and affecting the reliable input for energy consumption assessment and subsequent management strategies. Existing solutions still lack the ability to simultaneously satisfy multi-source synchronization consistency, time-delay structure expression, and acyclic causal constraints under online conditions, making it difficult to maintain the stability of the correlation model structure and parameters during continuous updates.

[0004] Therefore, how to provide an energy management system for carbon black production equipment based on the Internet of Things is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an IoT-based energy management system for carbon black production equipment. Addressing the problem of distortion in the online correlation between energy consumption data and production condition data in existing technologies, this invention proposes a technical solution that constructs an online input dataset using multi-source synchronous data streams and performs time-delay structure expansion. Based on a causal discovery training sample set, a directed sparse causal gating matrix and time-delay selection parameters are initialized. The causal gating parameter set is embedded into the intervariate interaction path of a selective state-space long sequence modeling model. A joint training objective is constructed and joint optimization is performed to obtain a parameterized correlation model. State recursion is used to generate a state representation sequence and a model output dataset. Feedback updates are performed based on the model output dataset to obtain an updated state representation sequence. This invention possesses the ability to represent the consistent correlation between energy consumption data and production condition data in online scenarios and reduces the risk of online correlation distortion.

[0006] An energy management system for carbon black production equipment based on the Internet of Things (IoT) according to an embodiment of the present invention includes the following modules: Data acquisition and synchronization module: collects energy consumption data and production status data of carbon black production equipment, performs unified timestamp processing on the collected data, and forms a multi-source synchronized data stream; Data preprocessing module: Performs standardization processing on multi-source synchronous data streams and resamples them according to a preset sampling period to form an online input dataset; Time Delay Structure Unfolding Module: Based on the online input dataset, time delay structure unfolding samples are constructed. Production condition variables are generated into time delay candidate variables within a preset candidate time delay set, and together with energy consumption variables, they form a time delay candidate parent node set, thus forming a causal discovery training sample set. Causal gating structure initialization module: Based on the causal discovery training sample set, the directed sparse causal gating matrix and time delay selection parameters are initialized, and the directed sparse causal gating matrix is ​​introduced into the cross-variable interaction path of the MambaTS selective state space long sequence modeling model to form the causal gating parameter set and the initial structure of the model with embedded causal structure. Causal Embedding Joint Training Module: Based on the initial structure of the model with embedded causal structure, a joint training objective is constructed. The joint training objective includes sequence fitting constraint terms and acyclic structure constraint terms. Joint optimization is performed to synchronously update the parameters of the MambaTS selective state space long sequence modeling model and the causal gating parameter set, forming a parameterized correlation model. State recursive association calculation module: Inputs the online input dataset into the parameterized association model to perform state recursive calculation, and generates a state representation sequence and model output dataset; Causal Feedback Update Module: Updates the causal discovery training sample set based on the model output dataset, performs feedback updates, and forms an updated state representation sequence.

[0007] Optionally, the data acquisition and synchronization module specifically includes: Energy consumption data and production condition data are acquired from the carbon black production equipment to form the raw data set. Assign timestamps with a unified time base to the data records in the original collected dataset to form a timestamped dataset; The timestamped dataset is grouped according to device identifier and data type, time-aligned, and energy consumption data and production condition data within the same time window are combined into time slice data entries to generate a multi-source synchronous data stream.

[0008] Optionally, the data preprocessing module specifically includes: Based on multi-source synchronous data streams, abnormal data removal and standardization processes are performed to form a standardized dataset. The standardized dataset is resampled according to a preset sampling period to form an online input dataset.

[0009] Optionally, the time-delay structure unpacking module specifically includes: Based on the online input dataset, the energy consumption data and production condition data in the online input dataset are divided into energy consumption variable set and production condition variable set according to data type, forming a variable set dataset; A preset candidate time delay set is determined based on the variable set dataset. The preset candidate time delay set consists of multiple discrete time delay values, each of which corresponds to an integer multiple of a sampling period, forming the candidate time delay set parameters. Based on the candidate time delay set parameters, time delay candidate variables are generated for each production condition variable in the production condition variable set. The time delay candidate variables are obtained by time shifting the production condition variables according to their corresponding discrete time delay values, thus forming a time delay candidate variable set. The time-delay candidate variable set and the energy consumption variable set are concatenated according to the time slice index, and a candidate parent node set is established under each time slice index. The candidate parent node set contains all time-delay candidate variables and energy consumption variables corresponding to the same time slice index, forming a time-delay candidate parent node set dataset. The dataset of candidate parent nodes with time delay is processed by sample window segmentation. A data window of a preset length composed of continuous time slice indices is used as a training sample, and a training sample number is attached to the training sample to form a causal discovery training sample set.

[0010] Optionally, the causal gating structure initialization module specifically includes: Based on the causal discovery training sample set, the total number of variables in the energy consumption variable set and the time delay candidate variable set is determined. Based on the total number of variables, the matrix dimension of the directed sparse causal gating matrix is ​​determined, and the matrix dimension parameter is generated. The directed sparse causal gating matrix is ​​initialized based on the matrix dimension parameter. The diagonal elements of the directed sparse causal gating matrix are set to zero, and trainable gating weights are assigned to the off-diagonal elements to generate the initial directed sparse causal gating matrix. The time delay selection parameters are initialized based on the candidate time delay set parameters. The time delay selection parameters are established by variable pairs, and trainable selection weights are assigned to the discrete time delay values ​​in the candidate time delay set parameters under each variable pair to generate the initial time delay selection parameters. The initial directed sparse causal gating matrix is ​​combined with the initial time delay selection parameters to form a causal gating parameter set, and parameter indices are assigned to the causal gating parameter set to form an indexed causal gating parameter set. In the MambaTS selective state-space long sequence modeling model, an indexed causal gating parameter set is introduced into the intervariate interaction path, and a correspondence is established between the indexed causal gating parameter set and the variable interaction weights of the intervariate interaction path to form the initial structure of the model with embedded causal structure.

[0011] Optionally, the causal embedding joint training module specifically includes: The input feature order of the MambaTS selective state-space long sequence modeling model is determined based on the online input dataset, and the online input dataset is organized into a model input sequence according to the input feature order, forming a model input data sequence. Based on the causal gating parameter set and the initial model structure embedded with causal structure, a gating mapping relationship for cross-variable interaction paths is constructed. The gating mapping relationship is used to map the gating weights of the directed sparse causal gating matrix to the variable interaction weights of the cross-variable interaction paths one by one, forming a gating mapping parameter set. Based on the gating mapping parameter set, a directed sparse gating reparameterization process is performed on the cross-variable interaction path. The variable interaction weights are combined with the corresponding gating weights to generate a gating interaction weight set. The gating interaction weight set is written into the cross-variable interaction path of the MambaTS selective state space long sequence modeling model, and the state update calculation link of the MambaTS selective state space long sequence modeling model is parameterized and bound to form a parameterized association model. A joint training objective is generated based on the parametric association model and the model input data sequence. The joint training objective includes sequence fitting constraints and acyclic structure constraints. Joint optimization is performed to simultaneously update the parameters of the parametric association model and the causal gating parameter set, forming an updated parametric association model.

[0012] Optionally, the state recursive association calculation module specifically includes: Receive online input dataset and generate state recursive input sequence according to the organization rules of model input data sequence, forming state recursive input data sequence; The state recursion input data sequence is input into the updated parameterized association model. Based on the state update calculation link of the parameterized association model, the time-slice state recursion calculation is performed to generate the state representation vector corresponding to the time slice index, forming a state representation sequence. In the updated parameterized association model, output mapping calculation is performed based on the state representation vector to generate energy consumption output vector and form energy consumption output sequence; In the updated parameterized association model, output mapping calculation is performed based on the state representation vector to generate the working condition output vector and form the working condition output sequence. The state representation sequence, energy consumption output sequence, and operating condition output sequence are aligned and concatenated according to the time slice index to form the model output dataset.

[0013] Optionally, the causal feedback update module specifically includes: Receive the model output dataset, and extract data windows composed of continuous time slice indices based on the energy consumption output sequence and operating condition output sequence in the model output dataset according to the time slice index, forming a data window sample set; Perform variable set update processing on the data window sample set, and divide the energy consumption output sequence and operating condition output sequence in the data window sample set into energy consumption variable set and production operating condition variable set according to data type to form update variable set dataset; Based on the updated variable set dataset and the preset candidate time delay set, a time delay structure expansion sample is generated. Based on the time delay structure expansion sample, an updated time delay candidate parent node set is constructed, forming an updated causal discovery training sample set. The updated causal discovery training sample set is input into the update computation link of the causal gating parameter set. Based on the updated causal discovery training sample set, the directed sparse causal gating matrix and time delay selection parameters are updated to form the updated causal gating parameter set. The updated causal gating parameter set is written into the parameterized association model and the parameter adaptive update is performed to form the updated parameterized association model; Based on the updated parameterized association model, the state recursion calculation is performed on the data window sample set to form an updated state representation sequence.

[0014] The beneficial effects of this invention are: This invention collects energy consumption data and production condition data from carbon black production equipment and performs unified timestamp processing to form a multi-source synchronous data stream. Combined with standardization and resampling to construct an online input dataset, it achieves stable input of multi-source data under the same time reference, reducing the risk of correlation bias caused by differences in sampling periods and time alignment deviations. By constructing a time-delay structure based on the online input dataset to expand samples and generate a causal discovery training sample set, it initializes a directed sparse causal gating matrix and time-delay selection parameters. The directed sparse causal gating matrix is ​​then embedded into the cross-variable interaction path of a selective state-space long sequence modeling model, realizing structured constraints on cross-variable interaction channels oriented towards time-delay effects. This improves the correlation between energy consumption variables and production conditions. This study demonstrates the ability of condition variables to express structural consistency in long-sequence modeling. By constructing a joint training objective that includes sequence fitting constraints and acyclic structure constraints and performing joint optimization, the model parameters and causal gating parameter sets of the selective state space long-sequence modeling are updated synchronously. Combined with a feedback update mechanism based on the model output dataset, this study effectively solves the problem of correlation distortion that easily occurs when energy consumption data and production condition data are correlated online in existing technologies. It overcomes the limitations of relying solely on numerical correlation matching, which leads to confusion in causal directions and difficulty in stably expressing time-delay structures. This study achieves continuous updating and consistency maintenance of the online correlation model structure and parameters, and improves the online correlation modeling capability and engineering application support level of the energy management system for carbon black production equipment. Attached Figure Description

[0015] 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 This is a schematic diagram of the process of an energy management system for carbon black production equipment based on the Internet of Things proposed in this invention.

[0016] Figure 2 This is a schematic diagram of the causal gating embedded MambaTS correlation modeling structure of an energy management system for carbon black production equipment based on the Internet of Things proposed in this invention. Detailed Implementation

[0017] Combination Figures 1-2 The present invention will be described in further detail below. These accompanying drawings are simplified schematic diagrams, illustrating only the basic structure of the invention and showing the main components relevant to the invention. Figure 1 and Figure 2 The present invention provides an energy management system for carbon black production equipment based on the Internet of Things, comprising the following modules: Data acquisition and synchronization module: collects energy consumption data and production status data of carbon black production equipment, performs unified timestamp processing on the collected data, and forms a multi-source synchronized data stream; Data preprocessing module: Performs standardization processing on multi-source synchronous data streams and resamples them according to a preset sampling period to form an online input dataset; Time Delay Structure Unfolding Module: Based on the online input dataset, time delay structure unfolding samples are constructed. Production condition variables are generated into time delay candidate variables within a preset candidate time delay set, and together with energy consumption variables, they form a time delay candidate parent node set, thus forming a causal discovery training sample set. Causal gating structure initialization module: Based on the causal discovery training sample set, the directed sparse causal gating matrix and time delay selection parameters are initialized, and the directed sparse causal gating matrix is ​​introduced into the cross-variable interaction path of the MambaTS selective state space long sequence modeling model to form the causal gating parameter set and the initial structure of the model with embedded causal structure. Causal Embedding Joint Training Module: Based on the initial structure of the model with embedded causal structure, a joint training objective is constructed. The joint training objective includes sequence fitting constraint terms and acyclic structure constraint terms. Joint optimization is performed to synchronously update the parameters of the MambaTS selective state space long sequence modeling model and the causal gating parameter set, forming a parameterized correlation model. State recursive association calculation module: Inputs the online input dataset into the parameterized association model to perform state recursive calculation, and generates a state representation sequence and model output dataset; Causal Feedback Update Module: Updates the causal discovery training sample set based on the model output dataset, performs feedback updates, and forms an updated state representation sequence.

[0018] In this embodiment, the data acquisition and synchronization module specifically includes: Energy consumption data and production condition data are acquired from the carbon black production equipment to form the raw data set. Assign timestamps with a unified time base to the data records in the original collected dataset to form a timestamped dataset; The timestamped dataset is grouped according to device identifier and data type, time-aligned, and energy consumption data and production condition data within the same time window are combined into time slice data entries to generate a multi-source synchronous data stream.

[0019] In this embodiment, the data preprocessing module specifically includes: Based on multi-source synchronous data streams, abnormal data removal and standardization processes are performed to form a standardized dataset. The standardized dataset is resampled according to a preset sampling period to form an online input dataset.

[0020] In this embodiment, the time-delay structure expansion module specifically includes: Based on the online input dataset, the energy consumption data and production condition data in the online input dataset are divided into energy consumption variable set and production condition variable set according to data type, forming a variable set dataset; A preset candidate time delay set is determined based on the variable set dataset. The preset candidate time delay set consists of multiple discrete time delay values, each of which corresponds to an integer multiple of a sampling period, forming the candidate time delay set parameters. Based on the candidate time delay set parameters, time delay candidate variables are generated for each production condition variable in the production condition variable set. The time delay candidate variables are obtained by time shifting the production condition variables according to their corresponding discrete time delay values, thus forming a time delay candidate variable set. The time-delay candidate variable set and the energy consumption variable set are concatenated according to the time slice index, and a candidate parent node set is established under each time slice index. The candidate parent node set contains all time-delay candidate variables and energy consumption variables corresponding to the same time slice index, forming a time-delay candidate parent node set dataset. The dataset of candidate parent nodes with time delay is processed by sample window segmentation. A data window of a preset length composed of continuous time slice indices is used as a training sample, and a training sample number is attached to the training sample to form a causal discovery training sample set.

[0021] In this embodiment, the causal gating structure initialization module specifically includes: Based on the causal discovery training sample set, the total number of variables in the energy consumption variable set and the time delay candidate variable set is determined. Based on the total number of variables, the matrix dimension of the directed sparse causal gating matrix is ​​determined, and the matrix dimension parameter is generated. The directed sparse causal gating matrix is ​​initialized based on the matrix dimension parameter. The diagonal elements of the directed sparse causal gating matrix are set to zero, and trainable gating weights are assigned to the off-diagonal elements to generate the initial directed sparse causal gating matrix. The time delay selection parameters are initialized based on the candidate time delay set parameters. The time delay selection parameters are established by variable pairs, and trainable selection weights are assigned to the discrete time delay values ​​in the candidate time delay set parameters under each variable pair to generate the initial time delay selection parameters. The initial directed sparse causal gating matrix is ​​combined with the initial time delay selection parameters to form a causal gating parameter set, and parameter indices are assigned to the causal gating parameter set to form an indexed causal gating parameter set. In the MambaTS selective state-space long sequence modeling model, an indexed causal gating parameter set is introduced into the intervariate interaction path, and a correspondence is established between the indexed causal gating parameter set and the variable interaction weights of the intervariate interaction path to form the initial structure of the model with embedded causal structure.

[0022] In this embodiment, the causal embedding joint training module specifically includes: The input feature order of the MambaTS selective state-space long sequence modeling model is determined based on the online input dataset, and the online input dataset is organized into a model input sequence according to the input feature order, forming a model input data sequence. Based on the causal gating parameter set and the initial model structure embedded with causal structure, a gating mapping relationship for cross-variable interaction paths is constructed. The gating mapping relationship is used to map the gating weights of the directed sparse causal gating matrix to the variable interaction weights of the cross-variable interaction paths one by one, forming a gating mapping parameter set. Based on the gating mapping parameter set, a directed sparse gating reparameterization process is performed on the cross-variable interaction path. The variable interaction weights are combined with the corresponding gating weights to generate a gating interaction weight set. The gating interaction weight set is written into the cross-variable interaction path of the MambaTS selective state space long sequence modeling model, and the state update calculation link of the MambaTS selective state space long sequence modeling model is parameterized and bound to form a parameterized association model. A joint training objective is generated based on the parametric association model and the model input data sequence. The joint training objective includes sequence fitting constraints and acyclic structure constraints. Joint optimization is performed to simultaneously update the parameters of the parametric association model and the causal gating parameter set, forming an updated parametric association model.

[0023] In this embodiment, the state recursive association calculation module specifically includes: Receive online input dataset and generate state recursive input sequence according to the organization rules of model input data sequence, forming state recursive input data sequence; The state recursion input data sequence is input into the updated parameterized association model. Based on the state update calculation link of the parameterized association model, the time-slice state recursion calculation is performed to generate the state representation vector corresponding to the time slice index, forming a state representation sequence. In the updated parameterized association model, output mapping calculation is performed based on the state representation vector to generate energy consumption output vector and form energy consumption output sequence; In the updated parameterized association model, output mapping calculation is performed based on the state representation vector to generate the working condition output vector and form the working condition output sequence. The state representation sequence, energy consumption output sequence, and operating condition output sequence are aligned and concatenated according to the time slice index to form the model output dataset.

[0024] In this embodiment, the causal feedback update module specifically includes: Receive the model output dataset, and extract data windows composed of continuous time slice indices based on the energy consumption output sequence and operating condition output sequence in the model output dataset according to the time slice index, forming a data window sample set; Perform variable set update processing on the data window sample set, and divide the energy consumption output sequence and operating condition output sequence in the data window sample set into energy consumption variable set and production operating condition variable set according to data type to form update variable set dataset; Based on the updated variable set dataset and the preset candidate time delay set, a time delay structure expansion sample is generated. Based on the time delay structure expansion sample, an updated time delay candidate parent node set is constructed, forming an updated causal discovery training sample set. The updated causal discovery training sample set is input into the update computation link of the causal gating parameter set. Based on the updated causal discovery training sample set, the directed sparse causal gating matrix and time delay selection parameters are updated to form the updated causal gating parameter set. The updated causal gating parameter set is written into the parameterized association model and the parameter adaptive update is performed to form the updated parameterized association model; Based on the updated parameterized association model, the state recursion calculation is performed on the data window sample set to form an updated state representation sequence.

[0025] Example 1: To verify the feasibility of this invention in practice, a carbon black production equipment that has undergone IoT access upgrades was selected as the verification object. The verification scenarios covered three typical processes: stable load operation, load step adjustment, and operating condition drift caused by raw material fluctuations. The data collected included equipment energy consumption data and production condition data. The energy consumption data covered the power of the main fan, the power of the induced draft fan, the power of the feeding system, and comprehensive power metering information. The production condition data covered the reactor temperature, feed rate, flue gas oxygen content, and key valve opening information. All data entered the data acquisition and synchronization module through the equipment-side acquisition link and formed multi-source synchronization. The data stream then enters the data preprocessing module to form an online input dataset. Based on this, the time-delay structure expansion module constructs a causal discovery training sample set. Further, the causal gating structure initialization module constructs a causal gating parameter set and forms an initial model structure embedding the causal structure. Then, the causal embedding joint training module completes the parameterized association model training and update. Finally, the state recursive association calculation module generates a state representation sequence and a model output dataset, and the causal feedback update module completes the online feedback update, enabling the system to continuously output the updated state representation sequence for state characterization and consistency association assessment on the energy management side.

[0026] In the aforementioned scenarios, existing methods often employ time-aligned correlation calculations of energy consumption data and production condition data, moving regression, and segmented modeling of operating conditions based on empirical thresholds to achieve online correlation. These methods are prone to significant fluctuations in correlation relationships as the time window changes, especially in situations with strong multivariate coupling, variable time delays, and frequent operating condition drift. This manifests as energy consumption changes being incorrectly attributed to non-dominant operating condition variables, or misalignment between energy consumption and operating condition responses during load step transitions, resulting in distorted correlations. This invention specifically addresses the structural distortion in the online correlation process between energy consumption data and production condition data. By embedding a directed sparse gating structure driven by a causal gating parameter set into the cross-variable interaction path of the MambaTS selective state-space long-sequence modeling model, the model is modulated by both acyclic structure constraints and time delay selection constraints during long-sequence recursion, reducing the interference of non-causal paths on variable interaction weights. Furthermore, in feedback updates, the model output dataset is used to reconstruct the causal discovery training sample set to track operating condition drift.

[0027] During implementation, the data acquisition and synchronization module timestamps the original acquired dataset using a unified time base and aligns it by device identifier and data type. Energy consumption data and production condition data within the same time window are combined into time-slice data entries to form a multi-source synchronous data stream. The data preprocessing module performs anomaly removal and standardization on the multi-source synchronous data stream to form a standardized dataset. It then resamples the dataset according to a preset sampling period of 5 seconds to form an online input dataset. Anomaly removal employs a combined rule of mutation threshold and missing segment imputation based on time-slice indices. When the missing segment length does not exceed 6 time-slice indices, neighborhood linear imputation is used; when it exceeds 6 time-slice indices, the corresponding time-slice data entry is directly removed, and a removal mask is recorded for subsequent sample window segmentation to avoid invalid segments.

[0028] The time-delay structure expansion module receives the online input dataset and divides it into a set of energy consumption variables and a set of production condition variables according to data type, forming a variable set dataset. The candidate time-delay set parameters are set as discrete time-delay values ​​consisting of 0, 1, 2, 3, 4, 6, and 8 sampling periods, covering a time-delay range of 0 to 40 seconds. For each production condition variable in the production condition variable set, a time-delay candidate variable set is generated according to the candidate time-delay set parameters, and concatenated with the energy consumption variable set according to time slice index. A candidate parent node set is established under each time slice index to form a time-delay candidate parent node set dataset. The sample window segmentation process uses a preset length data window of 120 consecutive time slice indices as a training sample, corresponding to a 10-minute span, and adds a training sample number to the training sample to form a causal discovery training sample set. During segmentation, windows containing more than 5% of the data with rejection masks are skipped to avoid abnormal fragments contaminating the training samples.

[0029] The causal gating structure initialization module determines the matrix dimension of the directed sparse causal gating matrix based on the total number of variables in the energy consumption variable set and the time-delay candidate variable set in the causal discovery training sample set. The diagonal elements of the matrix are fixed at zero, while the off-diagonal elements are initialized as trainable gating weights with a mean of 0 and a standard deviation of 0.02. Time-delay selection parameters are established on a variable-pair basis, and trainable selection weights are assigned to each discrete time-delay value, with the selection weights initialized to a uniform distribution. The initial directed sparse causal gating matrix and the initial time-delay selection parameters are combined to form a causal gating parameter set, and parameter indices are assigned to form an indexed causal gating parameter set. This indexed causal gating parameter set is then introduced into the intervariate interaction path of the MambaTS selective state-space long sequence modeling model, and a one-to-one correspondence is established between it and the variable interaction weights of the intervariate interaction path to form the initial structure of the model embedding the causal structure.

[0030] The causal embedding joint training module determines the input feature order of the model input data sequence on the online input dataset. The input feature order is organized according to the priority of the energy consumption variable set and the order of the production condition variable set, forming the model input data sequence. The MambaTS selective state space long sequence modeling model adopts a 6-layer selective state space block, with the state representation vector dimension set to 128. The variable interaction weights of the cross-variable interaction path adopt a trainable matrix and perform directed sparse gating reparameterization under the constraints of the gating mapping parameter set. The joint training objective includes a sequence fitting constraint term and an acyclic structure constraint term. The sequence fitting constraint term consists of the mean square error of aligning the energy consumption output sequence and the production condition output sequence on the online input dataset. The acyclic structure constraint term consists of a differentiable penalty term of the directed sparse causal gating matrix satisfying the acyclic constraint. The weight coefficients are set as follows: sequence fitting constraint term weight 1, acyclic structure constraint term weight 0.3, and gating sparse regularization weight 0.05. The joint optimization uses the AdamW optimizer with a learning rate of 0.0008, a weight decay of 0.01, a batch size of 32 training samples, and a total of 3000 training iterations. In each iteration, the parameterized association model parameters and the causal gating parameter set are updated synchronously, and a sparse projection is performed on the gating weights after every 200 iterations to maintain the directed sparse structure.

[0031] The state recursion correlation computation module receives the online input dataset after training and generates a state recursion input data sequence. This sequence is then input into the updated parameterized correlation model. Time-slice state recursion computation is performed according to the time-slice index to generate a state representation sequence. Two output mapping computation links are executed on the same state representation vector to generate an energy consumption output sequence and a working condition output sequence. These three sequences are then aligned and concatenated according to their time-slice indices to form the model output dataset. The causal feedback update module takes the model output dataset as input and extracts data windows composed of consecutive time-slice indices to form a data window sample set. The data window length is set to 60 time-slice indices, corresponding to a 5-minute span. The variable set is then updated on the data window sample set, and an updated causal discovery training sample set is constructed based on a preset candidate time delay set. The update computation link performs a gating parameter update once when each data window arrives, with 80 iterations and a learning rate of 0.0005. After the update, the updated causal gating parameter set is written into the parameterized correlation model, and adaptive parameter updates are performed. Finally, state recursion computation is performed on the data window sample set to form the updated state representation sequence.

[0032] To quantify the usability of the consistent correlation results output by the system, 25 consecutive data windows were selected as evaluation segments within the validation period. Evaluation metrics included the mean absolute error, root mean square error, and mean absolute percentage error of the energy consumption output sequence against the online input dataset's energy consumption observations, as well as the root mean square error and coefficient of determination of the operating condition output sequence against the online input dataset's operating condition observations. After 3000 training iterations, the parameterized correlation model entered a stable phase. The mean absolute error of the energy consumption output sequence stabilized at the level of 0.18 kWh per window, the root mean square error at the level of 0.24 kWh per window, and the mean absolute percentage error at the level of 1.9%. The root mean square error of the reactor temperature in the operating condition output sequence stabilized at the level of 2.6 degrees Celsius, and the coefficient of determination stabilized at the level of 0.96. The comparison before and after the causal feedback update module was triggered shows that during the load step and raw material fluctuation periods, the updated causal gating parameter set significantly converged the gating interaction weights of the cross-variable interaction paths. The gating weights corresponding to non-dominant operating condition variables were compressed to a sparse interval close to zero, and the time delay selection weights of dominant operating condition variables were concentrated around 2 and 3 sampling periods, which significantly reduced the misalignment between energy consumption response and operating condition response in the time slice index dimension.

[0033] Table 1 is a comparison table of the measured and predicted values ​​of multiple indicators for five data window samples extracted in the validation section. The measured energy consumption value is taken from the cumulative difference of the comprehensive power metering of the corresponding window in the online input dataset, the predicted energy consumption value is taken from the cumulative value of the energy consumption output sequence window, the measured value of the operating condition is taken from the mean value of the online input dataset window, and the predicted value of the operating condition is taken from the mean value of the operating condition output sequence window.

[0034] Table 1. Comparison of Measured and Predicted Values ​​of Multiple Indicators in the Data Window Sample

[0035] As shown in Table 1, the maximum deviation between predicted and measured energy consumption values ​​across the five samples was 0.25 kWh; the maximum deviation between predicted and measured reactor temperature was 2.7 degrees Celsius; the maximum deviation between predicted and measured feed rate was 1.9 kg / min; the maximum deviation between predicted and measured main fan power was 6.5 kW; and the maximum deviation between predicted and measured flue gas oxygen content was 0.06 percent. Based on the overall statistical results of the validation section, the parameterized association model can simultaneously maintain the alignment consistency between the energy consumption output sequence and the operating condition output sequence under long-sequence state recursion. After operating condition drift occurs, the causal feedback update module reconstructs and updates the causal discovery training sample set through the model output dataset and updates the causal gating parameter set. This allows the gating interaction weights and time delay selection parameters of the intervariate interaction path to adaptively adjust with the operating state, thereby continuously forming an updated state representation sequence during online operation and maintaining the online consistency between energy consumption data and production operating condition data. This completes the entire operational process of causal relationship modeling, state recursive calculation, and feedback update for the energy management system of carbon black production equipment based on the Internet of Things.

Claims

1. A carbon black production plant energy management system based on Internet of Things, characterized in that, Includes the following modules: Data acquisition and synchronization module: collects energy consumption data and production status data of carbon black production equipment, performs unified timestamp processing on the collected data, and forms a multi-source synchronized data stream; Data preprocessing module: Performs standardization processing on multi-source synchronous data streams and resamples them according to a preset sampling period to form an online input dataset; Time Delay Structure Unfolding Module: Based on the online input dataset, time delay structure unfolding samples are constructed. Production condition variables are generated into time delay candidate variables within a preset candidate time delay set, and together with energy consumption variables, they form a time delay candidate parent node set, thus forming a causal discovery training sample set. Causal gating structure initialization module: Based on the causal discovery training sample set, the directed sparse causal gating matrix and time delay selection parameters are initialized, and the directed sparse causal gating matrix is ​​introduced into the cross-variable interaction path of the MambaTS selective state space long sequence modeling model to form the causal gating parameter set and the initial structure of the model with embedded causal structure. Causal Embedding Joint Training Module: Based on the initial structure of the model with embedded causal structure, a joint training objective is constructed. The joint training objective includes sequence fitting constraint terms and acyclic structure constraint terms. Joint optimization is performed to synchronously update the parameters of the MambaTS selective state space long sequence modeling model and the causal gating parameter set, forming a parameterized correlation model. State recursive association calculation module: Inputs the online input dataset into the parameterized association model to perform state recursive calculation, and generates a state representation sequence and model output dataset; Causal Feedback Update Module: Updates the causal discovery training sample set based on the model output dataset, performs feedback updates, and forms an updated state representation sequence.

2. The carbon black production plant energy management system based on the Internet of Things according to claim 1, characterized in that, The data acquisition and synchronization module specifically includes: Energy consumption data and production condition data are acquired from the carbon black production equipment to form the raw data set. Assign timestamps with a unified time base to the data records in the original collected dataset to form a timestamped dataset; The timestamped dataset is grouped according to device identifier and data type, time-aligned, and energy consumption data and production condition data within the same time window are combined into time slice data entries to generate a multi-source synchronous data stream.

3. The carbon black production plant energy management system based on the Internet of Things according to claim 1, characterized in that, The data preprocessing module specifically includes: Based on multi-source synchronous data streams, abnormal data removal and standardization processes are performed to form a standardized dataset. The standardized dataset is resampled according to a preset sampling period to form an online input dataset.

4. The energy management system for carbon black production equipment based on the Internet of Things according to claim 1, characterized in that, The time-delay structure expansion module specifically includes: Based on the online input dataset, the energy consumption data and production condition data in the online input dataset are divided into energy consumption variable set and production condition variable set according to data type, forming a variable set dataset; A preset candidate time delay set is determined based on the variable set dataset. The preset candidate time delay set consists of multiple discrete time delay values, each of which corresponds to an integer multiple of a sampling period, forming the candidate time delay set parameters. Based on the candidate time delay set parameters, time delay candidate variables are generated for each production condition variable in the production condition variable set. The time delay candidate variables are obtained by time shifting the production condition variables according to their corresponding discrete time delay values, thus forming a time delay candidate variable set. The time-delay candidate variable set and the energy consumption variable set are concatenated according to the time slice index, and a candidate parent node set is established under each time slice index. The candidate parent node set contains all time-delay candidate variables and energy consumption variables corresponding to the same time slice index, forming a time-delay candidate parent node set dataset. The dataset of candidate parent nodes with time delay is processed by sample window segmentation. A data window of a preset length composed of continuous time slice indices is used as a training sample, and a training sample number is attached to the training sample to form a causal discovery training sample set.

5. The energy management system for carbon black production equipment based on the Internet of Things according to claim 1, characterized in that, The causal gating structure initialization module specifically includes: Based on the causal discovery training sample set, the total number of variables in the energy consumption variable set and the time delay candidate variable set is determined. Based on the total number of variables, the matrix dimension of the directed sparse causal gating matrix is ​​determined, and the matrix dimension parameter is generated. The directed sparse causal gating matrix is ​​initialized based on the matrix dimension parameter. The diagonal elements of the directed sparse causal gating matrix are set to zero, and trainable gating weights are assigned to the off-diagonal elements to generate the initial directed sparse causal gating matrix. The time delay selection parameters are initialized based on the candidate time delay set parameters. The time delay selection parameters are established by variable pairs, and trainable selection weights are assigned to the discrete time delay values ​​in the candidate time delay set parameters under each variable pair to generate the initial time delay selection parameters. The initial directed sparse causal gating matrix is ​​combined with the initial time delay selection parameters to form a causal gating parameter set, and parameter indices are assigned to the causal gating parameter set to form an indexed causal gating parameter set. In the MambaTS selective state-space long sequence modeling model, an indexed causal gating parameter set is introduced into the intervariate interaction path, and a correspondence is established between the indexed causal gating parameter set and the variable interaction weights of the intervariate interaction path to form the initial structure of the model with embedded causal structure.

6. The energy management system for carbon black production equipment based on the Internet of Things according to claim 1, characterized in that, The causal embedding joint training module specifically includes: The input feature order of the MambaTS selective state-space long sequence modeling model is determined based on the online input dataset, and the online input dataset is organized into a model input sequence according to the input feature order, forming a model input data sequence. Based on the causal gating parameter set and the initial model structure embedded with causal structure, a gating mapping relationship for cross-variable interaction paths is constructed. The gating mapping relationship is used to map the gating weights of the directed sparse causal gating matrix to the variable interaction weights of the cross-variable interaction paths one by one, forming a gating mapping parameter set. Based on the gating mapping parameter set, a directed sparse gating reparameterization process is performed on the cross-variable interaction path. The variable interaction weights are combined with the corresponding gating weights to generate a gating interaction weight set. The gating interaction weight set is written into the cross-variable interaction path of the MambaTS selective state space long sequence modeling model, and the state update calculation link of the MambaTS selective state space long sequence modeling model is parameterized and bound to form a parameterized association model. A joint training objective is generated based on the parametric association model and the model input data sequence. The joint training objective includes sequence fitting constraints and acyclic structure constraints. Joint optimization is performed to simultaneously update the parameters of the parametric association model and the causal gating parameter set, forming an updated parametric association model.

7. The energy management system for carbon black production equipment based on the Internet of Things according to claim 1, characterized in that, The state recursive correlation calculation module specifically includes: Receive the online input dataset and generate a state recursive input sequence according to the organization rules of the model input data sequence, thus forming a state recursive input data sequence; The state recursion input data sequence is input into the updated parameterized association model. Based on the state update calculation link of the parameterized association model, the time-slice state recursion calculation is performed to generate the state representation vector corresponding to the time slice index, forming a state representation sequence. In the updated parameterized association model, output mapping calculation is performed based on the state representation vector to generate energy consumption output vector and form energy consumption output sequence; In the updated parameterized association model, output mapping calculation is performed based on the state representation vector to generate the working condition output vector and form the working condition output sequence. The state representation sequence, energy consumption output sequence, and operating condition output sequence are aligned and concatenated according to the time slice index to form the model output dataset.

8. The energy management system for carbon black production equipment based on the Internet of Things according to claim 1, characterized in that, The causal feedback update module specifically includes: Receive the model output dataset, and extract data windows composed of continuous time slice indices based on the energy consumption output sequence and operating condition output sequence in the model output dataset according to the time slice index, forming a data window sample set; Perform variable set update processing on the data window sample set, and divide the energy consumption output sequence and operating condition output sequence in the data window sample set into energy consumption variable set and production operating condition variable set according to data type to form update variable set dataset; Based on the updated variable set dataset and the preset candidate time delay set, a time delay structure expansion sample is generated. Based on the time delay structure expansion sample, an updated time delay candidate parent node set is constructed, forming an updated causal discovery training sample set. The updated causal discovery training sample set is input into the update computation link of the causal gating parameter set. Based on the updated causal discovery training sample set, the directed sparse causal gating matrix and time delay selection parameters are updated to form the updated causal gating parameter set. The updated causal gating parameter set is written into the parameterized association model and the parameter adaptive update is performed to form the updated parameterized association model; Based on the updated parameterized association model, the state recursion calculation is performed on the data window sample set to form an updated state representation sequence.

Images