Industrial industry energy consumption amount prediction method and system based on parameter clustering

Abstract

This invention relates to the field of industrial energy management technology, specifically to a method and system for predicting industrial energy consumption based on parameter clustering. The method includes: first, collecting multi-source process parameters from the production line and constructing an enhanced dataset using pseudo-samples generated from a digital twin; determining mutual information weights by calculating nonlinear correlation entropy, recalibrating the feature space, and performing virtual-real fusion parameter clustering to extract local energy consumption centroids; using locality-sensitive hashing to map the centroids to an encrypted domain to construct a global model, and introducing a KL divergence regularization term to correct prediction bias; finally, dynamically adjusting quota weights based on an energy credit index using a blockchain smart contract, and retrieving and outputting the cluster energy consumption prediction value based on the hash space. This invention achieves precise energy efficiency control through multi-source parameter clustering and on-chain feedback.

Description

Technical Field

[0001] This invention relates to the field of industrial energy management technology, specifically to a method and system for predicting industrial energy consumption based on parameter clustering. Background Technology

[0002] Energy quota management and accurate consumption forecasting for energy-intensive manufacturing industries have become core tasks for regional energy regulation and collaborative enterprise operations.

[0003] Currently, energy consumption forecasting and assessment for industrial manufacturing clusters typically rely on fixed quota allocation methods based on historical output or statistical analysis. These methods provide a foundation for energy planning in traditional management scenarios with a uniform production environment and constant production pace. However, blade manufacturing involves extremely complex material curing kinetics and stringent environmental requirements, exhibiting significant nonlinear characteristics. In large-scale industrial management practices spanning multiple regions and factories, the interplay of evolving physical mechanisms and dynamically changing environmental factors across various production stages makes it difficult for energy management departments to obtain universally applicable and objective forecasting results using existing management tools.

[0004] The technical problem this application aims to solve is that, in the context of highly dynamic coupling of manufacturing parameters in the industry, existing methods cannot effectively establish a precise mapping between the underlying production mechanism and the upper-level management model, resulting in low accuracy in predicting energy consumption of manufacturing clusters and a lack of objective basis for energy quota evaluation.

[0005] To address this, a method and system for predicting energy consumption in industrial sectors based on parameter clustering are proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting energy consumption in industrial sectors based on parameter clustering. Through multi-source parameter clustering and on-chain feedback, precise energy efficiency control is achieved. First, multi-source process parameters from the production line are collected, and an enhanced dataset is constructed using pseudo-samples generated from digital twins. Mutual information weights are determined by calculating nonlinear correlation entropy, the feature space is recalibrated, and virtual-real fusion parameter clustering is performed to extract local energy consumption centroids. Locality-Sensitive Hashing (LSH) is used to map the centroids to an encrypted domain to construct a global model, and a KL divergence regularization term is introduced to correct prediction bias. Finally, a blockchain smart contract dynamically adjusts the quota weights based on the energy credit index, and the predicted cluster energy consumption value is retrieved and output based on the hash space.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Industrial energy consumption prediction methods based on parametric clustering include: The system acquires the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and simultaneously correlates them with the corresponding total energy consumption data per unit cycle. A pseudo-sample dataset is generated using a blade digital twin model. The mutual information weights of the blade cavity surface pressure sequence, resin flow viscosity vector, relative humidity index, and pseudo-sample dataset are calculated using a parametric clustering algorithm to extract the local energy centroid coordinates. The Locality Sensitive Hash (LSH) algorithm is used to map the local energy centroid coordinates to a high-dimensional encrypted computation domain to generate a hash encrypted feature vector; the hash encrypted feature vectors uploaded by each server are received, and a second clustering is performed in the hash space to generate a global energy efficiency model; Calculate the KL divergence between the feature distribution of the pseudo-sample dataset and the centroid coordinates of the measured local energy consumption pattern, and use the KL divergence as a deviation constraint term to correct the global energy efficiency model; upload the consistency evaluation parameters of the feature vector, calculate the energy credit index of each manufacturing plant, and automatically trigger the evaluation feedback of the regional energy quota on the chain based on the index. Based on the blade specification parameters of the production cycle to be predicted, the calibrated global energy efficiency model is invoked to output the predicted energy value.

[0008] Preferably, the process of acquiring the pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and synchronously associating them with the corresponding total energy consumption data per unit cycle, includes: using a pressure sensor array arranged inside the blade mold cavity to collect multi-point pressure changes during the resin filling process in real time, and arranging them according to the time step to form a pressure sequence inside the mold cavity; using an online ultrasonic detector or rheological monitoring terminal installed on the vacuum infusion pipeline to detect the resin flow direction and velocity in real time, and calculating the resin flow viscosity vector reflecting the rheological characteristics of the material according to the fluid dynamics model; using industrial-grade temperature and humidity transmitters deployed around the production line to continuously extract the relative humidity index of the environment during the layup process; relying on intelligent power monitoring instruments connected to the production line control system to capture electrical parameters including active power, reactive power, and current changes in real time, and calculating the total energy consumption data per unit cycle for the corresponding production cycle by integration; using a globally unified time reference, aligning the collected pressure sequence, viscosity vector, and relative humidity index with the energy consumption data of the same cycle on the time axis, and constructing a multi-source feature correlation matrix through data encapsulation.

[0009] Preferably, the construction process of the pseudo-sample dataset includes: establishing a digital twin model of blade curing dynamics based on the coupling of Fourier's thermal conductivity law and resin curing kinetic equation to describe the dynamic evolution of heat transfer and exothermic chemical reaction within the mold cavity; setting multiple layup thickness gradients and initial reaction enthalpy ranges of the glass fiber reinforced material as model input variables for Monte Carlo random sampling to simulate the temperature field distribution and energy evolution law of the mold cavity under different working conditions; extracting the total heat flux per unit time during the simulation process and fitting to generate a virtual energy consumption distribution curve; synchronously monitoring the rate of temperature rise at the center during the simulation process, and automatically associating the simulated sample with a thermal runaway working condition label when the local temperature exceeds a preset thermal stability threshold, thus constructing a pseudo-sample dataset containing mechanistic features and label attributes.

[0010] Preferably, the process of extracting the centroid coordinates of local energy consumption includes: calculating the nonlinear correlation entropy between each process feature dimension and the total energy consumption per unit cycle, and determining the mutual information weight based on the entropy value; recalibrating the feature dimensions of the pressure sequence, viscosity vector, and relative humidity index; projecting the recalibrated measured feature vectors and the mechanism pseudo-sample dataset together onto a high-dimensional energy consumption feature space; using a clustering algorithm to perform affinity search on the sample points in the space to identify multiple subspace clusters representing different process stages; calculating the spatial geometric centroid or weighted average coordinates of all sample points in each subspace cluster, and extracting the centroid coordinates of the local energy consumption pattern.

[0011] Preferably, the global energy efficiency model generation process includes: using a set of randomly generated projection vectors to map the centroid coordinates of local energy consumption patterns to binary space; generating a hash code sequence composed of 0s and 1s by calculating the dot product of the coordinate points and the projection vectors and performing binarization; utilizing the proximity preservation property of the Locality Sensitive Hash (LSH) algorithm to ensure that centroid coordinates that are close in distance in the original energy consumption feature space have the same hash code with a higher number of bits after mapping; the server receives hash code sequences from different production lines, calculates the Hamming distance between each sequence, and divides sequences with the same and / or similar bit patterns into corresponding hash buckets according to the distance; statistically analyzing the distribution density and energy efficiency weight of samples in each hash bucket, and generating a global energy efficiency model by performing reverse mapping on the frequently occurring hash bit patterns.

[0012] Preferably, the correction process of the global energy efficiency model includes: fitting the centroid coordinates of the mechanism pseudo-sample dataset and the local energy consumption mode using kernel density estimation to generate the corresponding theoretical probability distribution function P and measured probability distribution function Q; traversing the sampling points in the high-dimensional feature space, calculating the expected value of the log-likelihood difference between distribution P and distribution Q, and obtaining the KL divergence value; adding the KL divergence value as a regularization constraint term to the loss function of the global energy efficiency model; and achieving closed-loop correction of the model prediction bias through iterative optimization.

[0013] Preferably, the process of automatically triggering the evaluation feedback of regional energy quotas on the blockchain based on the index includes: the smart contract calls the hash-encrypted feature vectors uploaded by each factory, calculates the Hamming distance between the hash-encrypted feature vectors and the high-energy-efficiency cluster centers in the global energy efficiency model, and obtains the consistency evaluation parameters; the consistency evaluation parameters are weighted and summed with the total energy consumption per unit period to generate the energy credit index of each manufacturing factory; the smart contract compares the energy credit index with the quota adjustment threshold stored in the blockchain ledger in real time, and automatically adjusts the energy quota weight factor of the factory on the blockchain according to the comparison deviation; the smart contract synchronizes the adjusted energy quota weight factor to the distributed nodes of the regional energy regulatory platform to complete the closed-loop evaluation feedback of energy quotas.

[0014] Preferably, the output process of the energy prediction value includes: extracting the geometric dimension parameters and structural ply parameters of the blades within the production cycle to be predicted; using the same projection matrix as that used to generate the hash encryption feature vector, mapping the extracted parameters to a high-dimensional encryption computation domain to generate a hash feature vector to be predicted; inputting the hash feature vector to be predicted into a calibrated global energy efficiency model, and searching for the target energy consumption pattern cluster with the smallest Hamming distance in the hash space; retrieving the energy efficiency centroid coordinates associated with the target energy consumption pattern cluster, and performing a product correction in conjunction with the preset rated power of the equipment and the production task of each production line to obtain the energy consumption prediction score of a single production line; summing the energy consumption prediction scores of all production lines in the cluster, and outputting the predicted energy consumption value of the entire blade manufacturing cluster within the prediction cycle.

[0015] An industrial energy consumption prediction system based on parametric clustering includes: Data acquisition module: acquires the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and synchronously associates it with the corresponding total energy consumption data per unit cycle. Twin Simulation Module: Generates a pseudo-sample dataset through a digital twin model of the blade; uses a parametric clustering algorithm to calculate the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector, the relative humidity index, and the mutual information weights of the pseudo-sample dataset, and extracts the coordinates of the local energy-consuming centroid. Encryption module: The Locality Sensitive Hash (LSH) algorithm is used to map the local energy centroid coordinates to a high-dimensional encryption computation domain to generate a hash encryption feature vector; the hash encryption feature vectors uploaded by each server are received and secondary clustering is performed in the hash space to generate a global energy efficiency model; Prediction module: Calculates the KL divergence between the feature distribution of the pseudo-sample dataset and the centroid coordinates of the measured local energy consumption pattern, and uses the KL divergence as a deviation constraint term to correct the global energy efficiency model; uploads the consistency evaluation parameters of the feature vectors, calculates the energy credit index of each manufacturing plant, and automatically triggers the evaluation feedback of regional energy quotas on the chain based on the index; calls the calibrated global energy efficiency model based on the blade specification parameters of the production cycle to be predicted, and outputs the predicted energy value.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By generating mechanism pseudo-samples through the built-in blade solidification dynamics digital twin model and using KL divergence for bias correction, this "virtual-real fusion" method enables the prediction model to no longer rely solely on historical statistical patterns, but to have logical support based on physical energy balance. Even in complex manufacturing environments where production parameters fluctuate frequently, the model can still accurately capture the impact of underlying process evolution on energy consumption, providing more scientific and cutting-edge energy consumption prediction values.

[0017] 2. By utilizing Locality Sensitive Hash (LSH) to map sensitive process parameters to a high-dimensional encrypted computation domain, core production data such as pressure sequences and resin formulations from each factory are ensured to participate in global clustering based on hash space similarity features, without leaving the factory or disclosing original values. This not only protects the company's core trade secrets but also enables the construction of an industry-wide common energy efficiency model. This allows regional monitoring platforms to objectively and fairly locate and analyze the energy consumption levels of individual factories based on industry-wide big data benchmarks.

[0018] 3. By linking blockchain smart contracts with the energy credit index, an automated closed loop from "data prediction" to "management feedback" is achieved. Based on consistent evaluation parameters of production characteristics, it can automatically generate transparent and tamper-proof energy credit profiles for each manufacturing plant and directly trigger energy quota adjustment commands on the blockchain. This decentralized management model eliminates subjective biases that may arise from human intervention. Simultaneously, the tiered feedback mechanism can incentivize enterprises to optimize process configurations in real time, thereby driving the entire blade manufacturing cluster towards a low-carbon, efficient, and green manufacturing model. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the industrial energy consumption prediction method based on parameter clustering of the present invention. Figure 2This is a schematic diagram of the industrial energy consumption prediction system based on parameter clustering according to the present invention. Figure 3 This is a schematic diagram of the process for extracting the coordinates of the local energy centroid in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1 to 3 This invention provides a method and system for predicting industrial energy consumption based on parameter clustering. The technical solution is as follows:

[0022] Example 1 Industrial energy consumption prediction methods based on parametric clustering include: The system acquires the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and simultaneously correlates them with the corresponding total energy consumption data per unit cycle. A pseudo-sample dataset is generated using a blade digital twin model. The mutual information weights of the blade cavity surface pressure sequence, resin flow viscosity vector, relative humidity index, and pseudo-sample dataset are calculated using a parametric clustering algorithm to extract the local energy centroid coordinates. The Locality Sensitive Hash (LSH) algorithm is used to map the local energy centroid coordinates to a high-dimensional encrypted computation domain to generate a hash encrypted feature vector; the hash encrypted feature vectors uploaded by each server are received, and a second clustering is performed in the hash space to generate a global energy efficiency model; Calculate the KL divergence between the feature distribution of the pseudo-sample dataset and the centroid coordinates of the measured local energy consumption pattern, and use the KL divergence as a deviation constraint term to correct the global energy efficiency model; upload the consistency evaluation parameters of the feature vector, calculate the energy credit index of each manufacturing plant, and automatically trigger the evaluation feedback of the regional energy quota on the chain based on the index. Based on the blade specification parameters of the production cycle to be predicted, the calibrated global energy efficiency model is invoked to output the predicted energy value.

[0023] Furthermore, the data acquisition process of the local production line includes: using a pressure sensor array arranged inside the cavity of the blade mold to collect multi-point pressure changes during the resin filling process in real time, and arranging them according to the time step to form a surface pressure sequence inside the mold cavity; using an online ultrasonic detector or rheological monitoring terminal installed on the vacuum filling pipeline to detect the resin flow direction and velocity in real time, and calculating the resin flow viscosity vector reflecting the rheological properties of the material according to the fluid dynamics model; using industrial-grade temperature and humidity transmitters deployed around the production line to continuously extract the environmental relative humidity index during the layup process; relying on intelligent power monitoring instruments connected to the production line control system to capture electrical parameters including active power, reactive power and current changes in real time, and integrally calculating the total energy consumption data per unit cycle for the corresponding production cycle; using a globally unified time reference, aligning the collected pressure sequence, viscosity vector and relative humidity index with the energy consumption data of the same cycle on the time axis, and constructing a multi-source feature correlation matrix through data encapsulation.

[0024] On the inner surface of the blade mold cavity, 16 piezoresistive pressure sensors are pre-embedded at equal intervals according to the pressure gradient distribution direction from the blade root to the blade tip. At the beginning of the pouring stage, these sensors monitor the pressure pulses generated by the resin flow in real time at a frequency of 10 times per second. The system extracts the readings of all sensors at each sampling moment and arranges them horizontally according to the order of collection time, thereby forming a pressure sequence on the inner surface of the mold cavity that reflects the spatiotemporal evolution of pressure.

[0025] Online ultrasonic detection terminals are installed at key nodes in the vacuum filling pipeline (such as the injection port and return port). These terminals utilize the propagation speed and energy attenuation characteristics of ultrasonic waves in heterogeneous fluids to detect the resin state. The specific calculation logic is as follows: the system uses the real-time acquired acoustic wave attenuation rate, compares it with a pre-set resin characteristic database, and converts the acoustic parameters into dynamic viscosity values ​​of the resin. To characterize the evolution of viscosity with the filling progress, the system extracts viscosity values ​​at multiple key characteristic moments from the start of filling to the gelation point, and encapsulates these discrete values ​​into a resin flow viscosity vector reflecting the material's rheological properties.

[0026] Industrial-grade temperature and humidity transmitters are deployed on supports above the layup station to continuously monitor and record the relative humidity percentage in the workshop. Simultaneously, intelligent power monitoring instruments with communication capabilities are installed in the main control cabinet of the production line. These instruments monitor and provide real-time feedback on the active power, reactive power, and instantaneous current fluctuations during the production process. By accumulating the power value multiplied by time within each sampling micro-element (i.e., the integral accumulation of power over time), the total energy consumption data for the blade throughout the entire production cycle from layup to curing is obtained.

[0027] To address the issue of inconsistent sampling step sizes among different sensors, the system connects to a globally unified time reference server (such as an NTP time server) within the factory. The acquisition process is as follows: The system assigns a high-precision timestamp to each collected pressure data, viscosity vector, and humidity index in real time. Subsequently, using the start time of the production work order issued by the production execution system as the zero point, the energy consumption data is linked to process parameters one by one on the time axis. Finally, the system uses a data encapsulation protocol to merge these multi-dimensional parameters at the same time step, constructing a multi-source feature correlation matrix with time-series correlation, which serves as the input benchmark for subsequent energy efficiency prediction.

[0028] By synchronously collecting and temporally correlating multidimensional physical parameters and energy efficiency data from the production line, a highly observable foundational dataset was provided for subsequent modeling. By aligning implicit fluid characteristics, environmental fluctuations, and explicit energy consumption on a unified time axis, the dynamic coupling logic of process actions and energy conversion during blade manufacturing was effectively reconstructed. This structured fusion of multi-source data ensures the authenticity of the mapping relationship between input features and energy efficiency target values, providing objective data support for improving the accuracy of energy consumption prediction under complex operating conditions.

[0029] Furthermore, the construction process of the pseudo-sample dataset includes: establishing a digital twin model of blade curing dynamics based on the coupling of Fourier's thermal conductivity law and resin curing kinetic equation to describe the dynamic evolution of heat transfer and exothermic chemical reaction within the mold cavity; setting multiple layup thickness gradients and initial reaction enthalpy ranges of glass fiber reinforced material as input variables for Monte Carlo random sampling to simulate the temperature field distribution and energy evolution law of the mold cavity under different working conditions; extracting the total heat flux per unit time during the simulation process and fitting to generate a virtual energy consumption distribution curve; synchronously monitoring the rate of temperature rise at the center during the simulation process, and automatically associating the simulated sample with a thermal runaway working condition label when the local temperature exceeds the preset thermal stability threshold, thus constructing a pseudo-sample dataset containing mechanistic features and label attributes.

[0030] When constructing a digital twin model of blade curing dynamics, a three-dimensional multilayer composite material model of a typical thickness region of the blade is first established in finite element analysis software. Basic thermophysical parameters such as density, thermal conductivity, and specific heat capacity of glass fiber reinforced material are set. The core logic of the model is to use Fourier's law of thermal conductivity as the basic criterion for spatial heat conduction and embed the Arrhenius-type curing kinetic rate equation as an internal heat generation term. When the resin begins to crosslink, each calculation unit calculates the heat release in real time according to the current temperature and degree of curing and feeds it back into the heat conduction equation to form a dynamic thermo-chemical coupling field simulation environment.

[0031] To cover complex on-site production conditions, this embodiment employs Monte Carlo random sampling technology for large-scale sample expansion, pre-setting the probability distribution of variables: the glass fiber layup thickness is set within the range of 10mm to 150mm, and sampling is performed according to a Gaussian distribution to simulate the structural characteristics of blades of different specifications; at the same time, considering the fluctuation of resin activity under different storage temperatures or batches, the initial total enthalpy of the reaction is set between 300J / g and 500J / g for uniform random sampling. In this way, more than 5,000 sets of virtual production condition combinations are automatically generated, each set representing a specific "thickness-material property" production background.

[0032] During the simulation calculation, the entire process from the end of the pouring to complete curing (usually 10 to 24 hours) is continuously simulated in 10-second simulation steps. The heat flux density change on the inner surface of the model is monitored in real time. The total energy exchange across the mold boundary per unit time is calculated and converted into active power consumption under laboratory benchmark. Then, the cubic spline interpolation algorithm is used to connect these discrete energy points into a smooth trajectory, thereby generating a corresponding virtual energy consumption distribution curve for each set of virtual working conditions. This curve can accurately reflect the time when the heat release peak occurs and its demand for external heating compensation.

[0033] For automated labeling of samples, the temperature rise slope at the geometric center of the blades (usually the area with the most severe heat accumulation) is monitored synchronously during simulation. In this example, the critical heating rate is set to 5℃ / min, and the upper limit of the center peak temperature is 180℃. Once the simulation results show that any indicator exceeds the above threshold, the operating condition is immediately determined to be in a "thermal runaway" risk state, and a numerical label "1" is automatically added to the data entry; otherwise, it is determined to be a process safety state, and the label is recorded as "0". In this way, not only is energy consumption data generated, but also clear physical boundary labels are given to the data.

[0034] Finally, the generated process parameter combinations (thickness, enthalpy), temperature response sequences, virtual energy consumption curves, and thermal runaway labels are standardized and packaged to form a structured mechanism pseudo-sample dataset. This dataset is stored in the form of a database file and serves as a prior knowledge base for subsequent parameter clustering algorithms. Compared with simple field measurement data, this dataset greatly expands the model's ability to identify high-risk, low-frequency operating conditions, ensuring that the prediction model still has mechanism-level robustness when facing special production cycles.

[0035] By constructing a digital twin model based on physical mechanisms and generating a pseudo-sample dataset, the problem of scarce data on extreme industrial conditions is effectively addressed. Monte Carlo sampling is used to simulate energy evolution under different thicknesses and reactivity levels, enabling the model to learn in advance the energy consumption patterns under physical boundary constraints, particularly the characteristic distribution under high-risk states such as thermal runaway. This mechanism-guided data augmentation method ensures that the predictive model not only possesses data fitting capabilities but also physical interpretability consistent with the laws of thermodynamics.

[0036] Furthermore, the process of extracting the centroid coordinates of local energy consumption includes: calculating the nonlinear correlation entropy between each process feature dimension and the total energy consumption per unit cycle, and determining the mutual information weight based on the entropy value; recalibrating the feature dimensions of the pressure sequence, viscosity vector, and relative humidity index; projecting the recalibrated measured feature vectors and the mechanism pseudo-sample dataset together onto the high-dimensional energy consumption feature space; using a clustering algorithm to perform affinity search on the sample points in the space to identify multiple subspace clusters representing different process stages; calculating the spatial geometric centroid or weighted average coordinates of all sample points in each subspace cluster, and extracting the centroid coordinates of the local energy consumption pattern.

[0037] Specifically, the measured intramold pressure sequence, resin flow viscosity vector, and real-time relative humidity index are first preprocessed. Since the raw data collected by the sensors are continuously fluctuating values, the system uses an equal-frequency partitioning method to divide the numerical range of each dimension into multiple intervals with equal probability. For example, by statistically analyzing the full distribution of pressure data, the values ​​are arranged from smallest to largest and divided into ten intervals, each representing a pressure state level. Through this mapping, the original continuous curve is transformed into a discrete sequence composed of interval numbers, providing a statistical benchmark for subsequent calculations of the correlation between characteristics and energy efficiency.

[0038] For each discretized process feature, the degree of coupling between it and the total energy consumption per unit cycle is measured by calculating the correlation entropy. Specifically, the logic is as follows: When a certain process feature (such as viscosity) is within a specific range, the determinism of the energy consumption data distribution is statistically analyzed. If changes in the process feature significantly reduce the uncertainty of the energy consumption data, a strong correlation is determined. By calculating the gain of this information, the system derives a mutual information value characterizing the correlation strength. During this process, the system can automatically identify nonlinear influence characteristics. For example, when fluctuations in the viscosity vector within a specific range cause drastic changes in energy efficiency, the correlation strength in that dimension is determined to be extremely high.

[0039] In this embodiment, the nonlinear correlation entropy is implemented through a mutual information quantification mechanism based on Shannon information theory to accurately measure the contribution of process parameters to energy consumption fluctuations. First, the discretized process characteristics and the total energy consumption per unit cycle are set as two independent discrete random variables. By traversing historical datasets, the frequency of combinations of each process characteristic value and energy consumption value is statistically analyzed to estimate their joint probability distribution and their respective marginal probability distributions. Subsequently, based on the information gain principle, the degree of reduction in energy consumption uncertainty under a known process characteristic is calculated. This calculation process compares the information difference between the joint and independent distributions to obtain a mutual information value characterizing the nonlinear coupling strength between the process characteristic and energy consumption. The obtained mutual information value is normalized and used as the core weight coefficient for this feature dimension. The larger the mutual information value, the stronger the explanatory power and the higher the correlation of the process characteristic to energy consumption changes. In the subsequent high-dimensional space recalibration step, the system proportionally stretches the corresponding coordinate axes according to this weight, so that in cluster analysis, small changes in these key process parameters can produce more significant spatial displacements, dominating the energy efficiency pattern identification process.

[0040] Based on the calculated mutual information values, weight coefficients are assigned to each feature dimension. Features with higher correlation strength (larger mutual information values) receive larger weight coefficients. Subsequently, these weight coefficients are used to recalibrate the high-dimensional feature space, i.e., to "stretch" or "compress" the coordinate axes of each dimension. For the pressure and viscosity dimensions with higher weights, the system stretches their coordinate scale in mathematical space; for the humidity dimension with lower weights, its coordinate scale is compressed. The result of this recalibration is that in subsequent calculations, small changes in core process parameters will produce larger spatial displacements, thus dominating the identification of the entire energy efficiency mode.

[0041] The recalibrated measured feature vectors and the pre-constructed mechanism pseudo-sample dataset are projected together into the aforementioned weighted feature space. Within this space, a clustering search algorithm is run to find the region with the densest sample distribution based on the weighted distance between points. Due to the introduction of pseudo-samples representing physical mechanisms, dense regions supported by both "theoretical simulation points" and "on-site measured points" will appear in the space. The algorithm identifies these dense regions as multiple subspace clusters representing different production processes or energy efficiency states. Each cluster is essentially an "energy efficiency label" under a specific process combination.

[0042] For each identified subspace cluster, the coordinate information of all sample points within that cluster is extracted. Finally, the arithmetic mean of these sample points across all dimensions is calculated to determine the center coordinates of the cluster. Specifically, the system sums the weighted values ​​of pressure, viscosity, humidity, etc., for all points within the cluster and divides them by the total number of samples. The resulting average coordinates are the centroid coordinates of the local energy consumption mode. These centroid coordinates, as representative features of this production cycle, not only condense massive amounts of raw sampling data but also completely preserve the core mapping relationship between process and energy consumption.

[0043] In this embodiment, the clustering algorithm is preferably a distance-based K-means clustering algorithm. Specifically, the recalibrated measured feature vectors and the mechanism pseudo-sample feature vectors are jointly input into the K-means clustering model, and weighted Euclidean distance is used as the affinity function between samples. The weights of each feature dimension are consistent with the aforementioned mutual information weights. The number of clusters K can be set according to the preset number of process stages and energy efficiency segments. For example, when the blade manufacturing process is divided into four stages: the initial glue injection stage, the exothermic peak stage, the heat preservation and curing stage, and the cooling and demolding stage, K is set to 4. The cluster centers are updated iteratively through K-means, so that each cluster center is simultaneously supported by both measured sample points and mechanism pseudo-sample points, thereby forming local energy consumption pattern centroids representing different process stages.

[0044] By automatically weighting and recalibrating the dimensions of process features using mutual information, and combining pseudo-samples to perform affinity clustering that integrates virtual and real data, efficient feature extraction and compression of high-dimensional manufacturing data are achieved. The mutual information mechanism can quantitatively identify key process variables that significantly impact energy efficiency, enhance the expressive power of core features through spatial stretching, and effectively suppress the interference of workshop environmental noise on model input. Simultaneously, the introduction of physically constrained pseudo-samples fills the distribution gaps of measured data under extreme or low-frequency operating conditions, reducing the computational complexity of subsequent hash encryption while ensuring the objectivity and physical interpretability of the mapping relationship between input features and energy consumption.

[0045] Furthermore, the generation process of the global energy efficiency model includes: using a set of randomly generated projection vectors to map the centroid coordinates of local energy consumption patterns to binary space; generating a hash code sequence composed of 0s and 1s by calculating the dot product of the coordinate points and the projection vectors and performing binarization; utilizing the proximity preservation property of the Locality Sensitive Hash (LSH) algorithm to ensure that centroid coordinates that are close in distance in the original energy consumption feature space have the same hash code with a higher number of bits after mapping; the server receives hash code sequences from different production lines, calculates the Hamming distance between each sequence, and divides sequences with the same and / or similar bit patterns into corresponding hash buckets according to the distance; statistically analyzes the distribution density and energy efficiency weight of samples in each hash bucket, and calculates and generates the global energy efficiency model by performing reverse mapping on the frequently occurring hash bit patterns.

[0046] Specifically, a set of randomly generated projection vector arrays is first pre-set on the local server. These vectors represent a set of randomly pointing "reference directions" in mathematical space. After obtaining the centroid coordinates of the local energy consumption mode, the coordinate point is multiplied by each random vector in the array. The physical meaning of this operation is to project multidimensional process feature points onto these random directions, resulting in a series of scalar values ​​of different positive and negative values. Then, binarization is performed: if the projected value is greater than or equal to zero, it is recorded as "1" in that position; if it is less than zero, it is recorded as "0". By judging the projection of all random vectors, the system transforms the originally complex physical quantity coordinates into a hash code sequence composed of "0" and "1".

[0047] In this embodiment, by controlling the number and distribution of random projection vectors, the spatial characteristics of the LSH (Local Sensitive Hash) algorithm are used to ensure data similarity. The logic is that if the distance between the coordinates of two centroids in the original process space is very close (meaning that their process parameters such as pressure and viscosity are highly similar), then when projected by random vectors, they are highly likely to fall on the same side of the projection plane. As a result, there will be more "0"s or "1"s at the same position in the generated hash code. This mapping method preserves the "operating condition similarity" in the form of "encoding similarity" without exposing specific physical values ​​(such as precise pressure values).

[0048] The regional central server receives hash-encoded sequences from different geographical locations and production lines. Because these sequences consist only of binary numbers, the server cannot deduce the original process formula of the factory, thus achieving data privacy protection. The server performs Hamming distance calculations on all received encoded sequences, comparing the number of different values ​​at the same bit in two encoded sequences. The smaller the Hamming distance, the more similar the production conditions behind the two codes are. Through this calculation, the server can quickly identify which factories are operating at the same frequency from massive amounts of random data.

[0049] Based on the calculated Hamming distance, the server divides the encoded sequences with the same or highly similar bit patterns into the corresponding hash buckets. Each hash bucket essentially represents a typical operating condition category in the industry. The server counts the distribution density of samples in each hash bucket in real time (i.e., the frequency of occurrence of this type of operating condition) and the associated energy efficiency weight. By analyzing which hash bit patterns (i.e., bucket numbers) occur most frequently, the server identifies the most universal energy consumption pattern prototype in the manufacturing cluster.

[0050] After obtaining the frequently occurring hash bit patterns, the server uses reverse mapping logic to associate these abstract encoded features with the synchronously uploaded de-identified energy consumption distribution patterns. Based on the sample distribution patterns within the hash bucket, it calculates the energy efficiency baseline corresponding to each type of operating condition, thereby generating a global energy efficiency model. This model can not only reflect the average energy efficiency level of the entire industry, but also identify the optimal energy consumption range under different process combinations, providing a unified and scientific "standard" for subsequent energy consumption prediction and evaluation feedback for individual factories.

[0051] After constructing the projection matrix containing a set of random vectors using the LSH algorithm, the process further includes an adaptive orthogonal correction step for the projection matrix. Specifically, the cloud server monitors the variance of the measured energy consumption data associated with each hash bucket in real time. When the variance of a specific hash bucket exceeds a preset splitting threshold, it is determined that a feature collision has occurred in that hash bucket. The cloud server calculates the maximum difference direction of all original feature vectors in the hash bucket based on principal component analysis and generates an orthogonal rotation correction matrix. The orthogonal rotation correction matrix is ​​then sent to the edge, where the edge uses this matrix to rotate and update the random vectors of the corresponding dimension in the original projection matrix, so that the feature vectors that caused the collision split into different hash coding bit patterns in the next mapping.

[0052] Specifically, the cloud server continuously monitors the data stability within each hash bucket (i.e., each type of energy consumption mode) in the global energy efficiency model. For a specific hash bucket, the server calculates the actual energy consumption values ​​corresponding to all historical samples accumulated in that bucket in real time, and calculates the variance of these values. For example, if the energy consumption of samples in a certain hash bucket is generally concentrated around 3200 with a small variance, then the mode is considered stable. However, if the variance of a certain hash bucket suddenly exceeds a preset splitting threshold (e.g., the variance exceeds 2500), it means that two significantly different sets of data have been mixed into the bucket: one set may be a low-energy consumption condition (e.g., 3000), and the other set may be a high-energy consumption condition (e.g., 3500). Physically, this indicates that a feature collision has occurred, that is, two fundamentally different process conditions have been incorrectly compressed into the same hash code due to poor projection angle.

[0053] Once a collision is detected, the cloud server immediately initiates a difference analysis program to extract the original feature vectors corresponding to all samples in the hash bucket of the problem. Principal component analysis is then used for calculation. Specifically, in the multidimensional feature space, the geometric direction in which these sample points are most dispersed and have the most obvious differences is found, i.e., the direction of maximum difference. This direction represents the axis with the greatest distinguishing power between the two sets of confused data (condition A and condition B). The reason why the current random projection matrix cannot separate them is that the existing projection vector is exactly perpendicular to this direction of maximum difference, causing the projected values ​​to overlap.

[0054] To address the aforementioned issues, an orthogonal rotation correction matrix is ​​generated. This matrix is ​​used to rotate and adjust the spatial coordinate axes. The computational objective is to make the projection vectors in the original projection matrix as parallel as possible to the calculated direction of maximum difference. By utilizing algorithms such as Schmidt orthogonalization, the orthogonality and independence of the vectors in the projection matrix are maintained while adjusting the projection angle. This generates a new set of more targeted projection parameters, which is equivalent to adjusting the angle from which the algorithm observes the data, specifically to distinguish the two sets of originally confused data.

[0055] The cloud server encapsulates the generated orthogonal rotation correction matrix into an update instruction and sends it to the edge manufacturing plant. Upon receiving the instruction, the edge plant uses this matrix to update its locally stored random projection matrix online. When the plant generates data for the two types of operating conditions again, the updated matrix is ​​used for projection calculation, producing drastically different results: the low-energy-consumption data that caused the collision may still have its projection value in the negative range, generating a hash bit "0"; while the high-energy-consumption data, due to the rotation of the projection plane, has its projection value flipped to the positive range, generating a hash bit "1". Thus, the mixed data causing prediction bias is successfully split into two different hash buckets, achieving adaptive improvement in model accuracy. By monitoring variance to trigger orthogonal rotation, the system can actively correct the mapping angle to adapt to the direction of maximum difference, achieving effective separation of mixed operating conditions and overcoming the feature overlap defect caused by mapping failure in traditional hash algorithms under dynamic environments.

[0056] By employing random projection and binarization mapping using the Locality Sensitive Hashing (LSH) algorithm, feature aggregation and collaborative modeling of sensitive process data within an encrypted domain were achieved. This approach, while masking the original physical values, preserves the similarity correlations between operating conditions using the proximity-preserving property, effectively solving the privacy and security challenges in cross-plant data integration. The hash bucket classification mechanism based on Hamming distance significantly reduces the computational complexity of cloud servers, enabling efficient aggregation of energy efficiency patterns across production lines. This provides a technical means to build a universally applicable industry energy efficiency benchmark model that balances privacy protection and computational efficiency.

[0057] Furthermore, the correction process of the global energy efficiency model includes: fitting the centroid coordinates of the mechanism pseudo-sample dataset and the local energy consumption mode using kernel density estimation to generate the corresponding theoretical probability distribution function P and measured probability distribution function Q; traversing the sampling points in the high-dimensional feature space to calculate the expected value of the log-likelihood difference between distribution P and distribution Q, and obtaining the KL divergence value; adding the KL divergence value as a regularization constraint term to the loss function of the global energy efficiency model; and achieving closed-loop correction of the model prediction bias through iterative optimization.

[0058] Specifically, the pseudo-sample dataset (representing the theoretical boundary) and the centroid coordinates of the local energy consumption mode (representing the measured state) generated in the previous steps are extracted first. Since these data points are discretely distributed in the feature space, a kernel density estimation method is used for smoothing. The specific operation is as follows: a smooth kernel function (such as a Gaussian kernel) is placed at the center of each data point. By superimposing the contributions of all kernel functions, continuous theoretical probability distribution function P and measured probability distribution function Q are generated respectively. The physical meaning of this step is to transform the discrete sampling points into a probability density cloud map that reflects the energy efficiency evolution law, where P represents the ideal distribution under the constraints of physical laws, and Q represents the characteristic distribution in the actual operation of the factory.

[0059] After obtaining two distribution functions, the differences between the two are compared by traversing the sampling region in the high-dimensional feature space. Under the same process parameters, the log-likelihood difference between the measured distribution Q and the theoretical distribution P is calculated. Then, the expected values ​​of these differences in the entire feature space are accumulated to obtain the KL divergence value, which reflects the information distance between the two. If the KL divergence value is small, it indicates that the measured working condition is highly consistent with the physical mechanism. If the value increases, it indicates that there may be prediction deviations caused by sensor offset, equipment aging, or external environmental interference in the production site. This value quantitatively indicates the extent to which the global energy efficiency model needs to be corrected.

[0060] The calculated KL divergence value is used as a regularization constraint term and added to the total loss function of the global energy efficiency model. During model training, the loss function must not only minimize the residual between the predicted and actual values, but also minimize this KL divergence term. This sets a physical guideline for the model's self-optimization, forcing the model's prediction logic to fit the measured data without deviating from the probability distribution defined by the digital twin mechanism. This constraint mechanism effectively prevents the model from overfitting or causing prediction distortion due to on-site noise interference.

[0061] Finally, the iterative optimization program is launched. By continuously adjusting the internal weight parameters of the global energy efficiency model, the optimal solution that minimizes the total loss function (including the residual term and the KL divergence term) is sought. In each iteration, the model is fine-tuned according to the offset direction of the KL divergence feedback until the predicted distribution and the theoretical distribution reach a statistical dynamic equilibrium. The final corrected model has both the ability to perceive the on-site working conditions in real time and the robustness in accordance with the laws of thermodynamics, realizing closed-loop correction of energy consumption prediction deviations across regions and multiple scenarios.

[0062] In this embodiment, the global energy efficiency model achieves closed-loop correction of prediction bias by constructing a composite loss function. The specific training and optimization process is as follows: First, a composite optimization objective that balances prediction accuracy and mechanistic constraints is set for the model. This objective is composed of a weighted combination of a prediction residual term and a physical regularization term. The first part of the loss term measures the difference between the energy consumption value predicted by the model and the energy consumption value monitored in actual production. By minimizing this bias, the model can be ensured to highly fit the measured data on site, guaranteeing the basic accuracy of the prediction. The second part of the loss term is the KL divergence regularization term, whose core logic is to calculate the information distance between the theoretical probability distribution generated by the digital twin model and the local energy centroid distribution extracted from on-site measurements. This term serves as the physical guideline for the model, constraining the parameter update process to ensure it does not deviate from the reasonable range defined by the laws of thermodynamics. A preset regularization weight coefficient is introduced into the total loss function. The role of this coefficient is to dynamically balance the relationship between fitting the measured data and adhering to the physical mechanism. By adjusting this coefficient, the intensity of physical constraints in correcting the model can be controlled, preventing the model from developing physical-level logical distortions due to overfitting to on-site noise. In each iteration of model parameter updates, the optimization algorithm simultaneously drives the prediction residuals and distribution differences to converge toward their minimum values. This synchronous optimization mechanism enables the final model to not only perceive complex changes in operating conditions in real time but also possesses strong robustness and physical interpretability, achieving automated correction of energy consumption prediction biases across regions and multiple scenarios.

[0063] By utilizing theoretical and measured distribution functions generated through kernel density estimation, and incorporating the KL divergence between the two into the loss function as a regularization constraint, closed-loop correction of the global energy efficiency prediction model was achieved. This mechanism, by quantifying the information distance between the physical mechanism distribution and the measured distribution in the field, forces the model to simultaneously consider data fit and physical logic consistency during optimization, effectively suppressing prediction fluctuations caused by field noise or abnormal operating conditions. This correction method not only improves the robustness of energy consumption prediction in complex environments but also provides more objective and manufacturing-mechanism-compliant data for subsequent energy quota allocation at the management decision-making level.

[0064] Furthermore, the process of automatically triggering the evaluation feedback of regional energy quotas on the blockchain based on the index includes: the smart contract calls the hash-encrypted feature vectors uploaded by each factory, calculates the Hamming distance between the hash-encrypted feature vectors and the high-efficiency cluster centers in the global energy efficiency model, and obtains the consistency evaluation parameters; the consistency evaluation parameters are weighted and summed with the total energy consumption per unit period to generate the energy credit index of each manufacturing factory; the smart contract compares the energy credit index with the quota adjustment threshold stored in the blockchain ledger in real time, and automatically adjusts the energy quota weight factor of the factory on the blockchain according to the comparison deviation; the smart contract synchronizes the adjusted energy quota weight factor to the distributed nodes of the regional energy regulatory platform to complete the closed-loop evaluation feedback of energy quotas.

[0065] In the blockchain network of the regional energy regulatory platform, the pre-set energy efficiency evaluation smart contract is automatically triggered. The contract first extracts the hash encrypted feature vector uploaded by each factory from the distributed ledger. Then, the contract retrieves the hash code of the high-efficiency cluster center representing the optimal energy use state in the global energy efficiency model. By comparing the difference between the two bits, the Hamming distance is calculated. The smaller the distance, the closer the factory's actual production process is to the industry-recognized green and high-efficiency model. This comparison result is defined as the consistency evaluation parameter, which is used to quantitatively measure the degree of matching between the factory's process execution and low-carbon goals.

[0066] The smart contract obtains the measured value of the total energy consumption per unit cycle of the factory within the same production cycle. The contract executes the weighted summation logic: the consistency evaluation parameters (representing "how well it is doing") and the total energy consumption data (representing "how little it is using") are merged and calculated according to the preset weight ratio, and finally the energy credit index of the factory is generated. This index takes into account both technological leadership and resource conservation. For example, even if a factory has low total energy consumption, if its technological characteristics are too far from the Hamming model of high energy efficiency (indicating that there may be technological hidden dangers or low-end production capacity), its credit index will be correspondingly reduced.

[0067] The blockchain ledger securely stores a matrix of quota adjustment thresholds set by the regional energy regulatory authority. The smart contract compares the generated energy credit index with this threshold in real time at millisecond levels. The comparison logic is that the contract calculates the direction and magnitude of the deviation between the index and the threshold. If the credit index is significantly higher than the excellent threshold, the factory is judged to have a "green manufacturing" demonstration effect; if it is lower than the warning threshold, its energy efficiency is judged to be low, and administrative intervention procedures are required.

[0068] After completing the comparison, the smart contract first calculates the relative deviation rate between the energy credit index and a preset threshold benchmark. Specifically, it subtracts the threshold benchmark from the credit index and then divides the result by the threshold benchmark to obtain a percentage value reflecting the degree of creditworthiness. For example, if a factory's credit index is 85 and the threshold is 80, the factory's relative deviation rate is positive 6.25%; if the credit index is 60, the relative deviation rate is negative 25%.

[0069] The smart contract pre-configures a tiered adjustment mapping table to convert the aforementioned percentages into specific weighting factor correction values. In this embodiment, the adjustment steps are set as follows: Positive incentive range: When the relative deviation rate increases by 5%, the smart contract automatically increases the "energy quota weight factor" of the factory in the next cycle by 0.05 units based on the previous benchmark (the upper limit is set to 1.5). Negative penalty interval: When the relative deviation rate decreases by more than 10%, the weight factor is directly reduced by 0.1 units; if the deviation rate further expands to more than 30%, the weight factor triggers the "rapid reduction" mechanism, reducing by 0.3 units at a time (the lower limit is set at 0.5). Through this clear ladder mapping logic, the abstract credit assessment is directly transformed into a coefficient variable that can be used for resource allocation.

[0070] To ensure the stability of the region's total energy supply, the smart contract performs a global normalization operation after adjusting the weight factors of individual factories. The specific process is as follows: the contract extracts the latest weight factors of all participating nodes in the region, calculates the arithmetic sum of all weight factors, and then divides the independent weight factor of each factory by the sum to obtain the final allocation ratio of each factory in the current region's total energy supply. This step ensures that no matter how individual factories are rewarded or punished, the total energy quota in the region remains under control.

[0071] Once the weighting factor and final allocation ratio are calculated, the smart contract immediately calls the underlying ledger's state update interface to bind and store the factory's unique identifier with its corresponding latest weighting factor. This adjustment instruction is packaged into a new block as a transaction. After verification by the consensus algorithm, this data is synchronized to all power supply nodes, energy regulatory terminals, and distributed nodes at the factory level within the region. Each node automatically executes the physical power allocation limit by reading the latest weighting factor on the chain, thus completing a closed loop from energy efficiency assessment to administrative resource allocation without human intervention.

[0072] By constructing a complete technical architecture from physical mechanism perception to blockchain-based automated management, an intelligent transformation of energy efficiency supervision in industrial manufacturing clusters has been achieved. Digital twins and mechanistic pseudo-samples enhance the model's predictive robustness in complex dynamic environments, ensuring the scientific rigor of the evaluation criteria. The application of locality-sensitive hashing (LSH) technology overcomes obstacles to cross-plant data collaboration while protecting the privacy of core enterprise processes. Finally, by automatically linking energy efficiency performance with energy quotas through blockchain smart contracts, a transparent, fair, and real-time closed-loop allocation mechanism is constructed. This not only significantly improves the accuracy of regional energy utilization efficiency supervision but also provides reliable technical support for the precise quantification and scientific control of carbon emissions in the industrial sector.

[0073] Furthermore, the output process of the energy prediction value includes: extracting the geometric dimension parameters and structural ply parameters of the blades within the production cycle to be predicted; using the same projection matrix as that used to generate the hash encryption feature vector, mapping the extracted parameters to a high-dimensional encryption computation domain to generate the hash feature vector to be predicted; inputting the hash feature vector to be predicted into the calibrated global energy efficiency model, and searching for the target energy consumption pattern cluster with the smallest Hamming distance in the hash space; retrieving the energy efficiency centroid coordinates associated with the target energy consumption pattern cluster, and performing a product correction in combination with the preset rated power of the equipment and the production task of each production line to obtain the energy consumption prediction score of a single production line; summing the energy consumption prediction scores of all production lines in the cluster, and outputting the predicted energy consumption value of the entire blade manufacturing cluster within the prediction cycle.

[0074] Before entering the prediction stage, the management platform first extracts the production plan data for the prediction period (such as the next week), locks the geometric dimension parameters (such as the total blade length and maximum chord length) and structural ply parameters (such as the number of main beam layers and axial ply density) of each blade to be manufactured, and then calls the locally stored random projection matrix that is completely consistent with the one used to generate the historical feature vector. The above physical parameters are input into the matrix to perform dot product operation, and binarization is completed through zero threshold judgment to generate the prediction hash feature vector of the prediction condition. This step ensures that the prediction input and the model base library communicate within the same encrypted dimension.

[0075] The generated hash feature vector to be predicted is sent to the global energy efficiency model in the cloud. The cloud server performs similarity retrieval in the hash space (hash bucket). The specific logic is as follows: calculate the Hamming distance between the vector to be predicted and the center of each energy consumption pattern cluster in the model, and automatically match the target energy consumption pattern cluster with the smallest distance (i.e. the highest bit overlap). Since this cluster contains the fusion features of a large number of historical operating conditions and mechanism samples, this process is essentially finding a digital template that is closest in terms of energy consumption pattern for the blade to be produced.

[0076] Once the target cluster is identified, the coordinates of the local energy efficiency centroid associated with that cluster are immediately retrieved (these coordinates represent the energy consumption intensity characteristics under a unit benchmark). To ensure that the prediction results are more realistic, a single production line operating condition correction mechanism is introduced: Equipment efficiency calibration: Extract the rated power and energy efficiency attenuation coefficient of the equipment in the production line as the first-level correction factor; Production load superposition: Read the current production task volume (such as the number of planned continuous injections) as the second-level correction factor, and perform continuous multiplication operation between the baseline energy consumption represented by the centroid coordinate and the above correction factor. This method ensures that the predicted value includes both general mechanism laws and takes into account the hardware differences and task load of specific production lines, and finally outputs the energy consumption prediction score of the single production line.

[0077] Finally, the management platform executes the above prediction process for all production lines involved in the prediction period, summing up the prediction scores of all production lines in operation within the cluster, and finally outputting the predicted energy consumption value of the manufacturing cluster in the prediction period. This value is visualized through the dashboard of the monitoring platform and serves as a direct reference for regional power load scheduling, energy procurement and storage, and carbon emission index calculation.

[0078] Privacy-preserving energy consumption prediction is achieved through fast hash space retrieval and multidimensional product correction. The use of a random projection matrix consistent with the training matrix ensures the closed-loop consistency of the prediction logic. Furthermore, the fast Hamming distance comparison mechanism, by combining the abstract energy efficiency centroid with real-world constraints such as equipment rated power and production scheduling, successfully transforms the "mechanism" into "execution data," effectively solving the computational bias caused by traditional prediction methods neglecting individual equipment differences.

[0079] By integrating physical mechanism modeling, multi-source process feature extraction, and privacy-preserving collaborative computing, an energy forecasting and quota management system covering the entire manufacturing process was constructed. Utilizing digital twin pseudo-samples and mutual information weighting mechanisms, the model's feature recognition capabilities and physical interpretability under complex dynamic conditions were effectively enhanced. Through locality-sensitive hashing and KL divergence calibration technologies, privacy shielding and closed-loop correction of prediction biases were achieved in cross-factory data collaboration. Finally, energy efficiency evaluation and quota allocation were automatically linked through blockchain smart contracts, providing transparent and objective decision-making basis for regional energy regulation and promoting the scientific and digital transformation of energy efficiency control in the industry.

[0080] Example 2

[0081] On production line 1 of factory A in the park, the system uses 18 piezoresistive pressure sensors embedded in the inner surface of the mold to capture pressure fluctuations during resin infusion at a frequency of 2Hz, forming a pressure sequence P. The online rheometer measures the initial viscosity of the resin as 280mPa·s and generates a 24-dimensional viscosity vector V as the curing progress progresses.

[0082] Because the factory is located near the sea, humidity sensors provide real-time feedback that the relative humidity in the workshop is 78%. Power meters simultaneously record the total power consumption of this blade from layup to demolding as 3250 kWh. The system uses an NTP server to align the above process characteristics with energy efficiency data, encapsulating them into a measured characteristic matrix for this cycle.

[0083] For this 110-meter-class blade, the digital twin model preset the thickness gradient (40mm to 160mm) in the main beam region and the resin reactivity fluctuation (enthalpy 380-460J / g). The system ran 3000 Monte Carlo simulations in the background to simulate the energy consumption trajectory of "premature gelation" that may occur under extreme high temperature and humidity conditions. When the central peak temperature shown in the simulation exceeded 175°C, the system automatically labeled the simulated trajectory with a "thermal runaway risk" tag. These pseudo-samples, together with the measured data from Plant A, constitute the mixed dataset to be analyzed.

[0084] The system discretizes the data from Factory A and calculates that under the current high humidity environment, the mutual information value between the viscosity vector V and the total energy consumption E reaches the highest value (0.72), while the mutual information value between the ambient humidity H is only 0.08.

[0085] Based on this, the system automatically stretches the "viscosity axis" within the space by a factor of 9. Within this weighted space, the measured data points and pseudo-sample points with physical mechanisms are rapidly clustered, identifying four local energy consumption subspaces, including the "initial injection molding section" and the "exothermic peak section." The system calculates the geometric centroid coordinates of these four clusters, which are the four sets of local energy consumption centroid coordinates for Factory A in this cycle.

[0086] Factory A performs a dot product operation on these four sets of centroid coordinates using a pre-set 64-bit random projection matrix. Bits with values ​​greater than zero are mapped to "1", and those with values ​​less than zero are mapped to "0", generating a hash code sequence of the form 10110...01. This code is uploaded to the park's monitoring server. The server calculates and finds that its Hamming distance to the "standard high-efficiency mode bucket" in the global model is only 3 (the bit patterns are highly similar), indicating that although Factory A is in a high-humidity environment, its process control logic still meets the industry's high-efficiency benchmark.

[0087] The monitoring server uses kernel density estimation to compare the measured distribution Q of factory A with the theoretical distribution P of the mechanism, and calculates the KL divergence value to be 0.12, which shows a small statistical bias. The system uses 0.12 as a penalty term and substitutes it into the loss function to regularize the global energy efficiency model. The corrected model eliminates the prediction disturbance caused by the high humidity environment, making the energy efficiency prediction benchmark for the next cycle more in line with the actual physical boundary of factory A.

[0088] The blockchain smart contract reads the credit performance of Factory A: because of its small Hamming distance and energy efficiency, the generated energy credit index is 92. The contract automatically compares the threshold (85) and judges it as excellent, and then triggers the quota correction: the energy quota weight factor of Factory A in the next cycle is increased from 1.0 to 1.05. Finally, the system retrieves the corresponding centroid coordinates according to the production plan of 12 blades of Factory A next week, multiplies them by the weight factor of 1.05 and the rated power of the equipment, and outputs the predicted value: the energy consumption of Factory A next week is expected to be 39,200kWh. This value is synchronized to the grid dispatch department to complete the accurate energy supply and demand matching.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting energy consumption in industrial sectors based on parametric clustering, characterized in that, include: The system acquires the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and simultaneously correlates them with the corresponding total energy consumption data per unit cycle. A pseudo-sample dataset is generated using a blade digital twin model. The mutual information weights of the blade cavity surface pressure sequence, resin flow viscosity vector, relative humidity index, and pseudo-sample dataset are calculated using a parametric clustering algorithm to extract the local energy centroid coordinates. The Locality Sensitive Hash (LSH) algorithm is used to map the local energy centroid coordinates to a high-dimensional encrypted computation domain to generate a hash encrypted feature vector; the hash encrypted feature vectors uploaded by each server are received, and a second clustering is performed in the hash space to generate a global energy efficiency model; Calculate the KL divergence between the feature distribution of the pseudo-sample dataset and the centroid coordinates of the measured local energy consumption pattern, and use the KL divergence as a bias constraint term to correct the global energy efficiency model; Upload the consistency evaluation parameters of the feature vector, calculate the energy credit index of each manufacturing plant, and automatically trigger the evaluation feedback of the regional energy quota on the chain based on the index. Based on the blade specification parameters of the production cycle to be predicted, the calibrated global energy efficiency model is invoked to output the predicted energy value.

2. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The process of acquiring the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and synchronously associating them with the corresponding total energy consumption data per unit cycle, includes: using a pressure sensor array arranged inside the blade mold cavity to collect multi-point pressure changes during the resin filling process in real time, and arranging them according to the time step to form the surface pressure sequence inside the mold cavity; using an online ultrasonic detector or rheological monitoring terminal installed on the vacuum infusion pipeline to detect the resin flow direction and velocity in real time, and calculating the resin flow viscosity vector reflecting the rheological characteristics of the material according to the fluid dynamics model; using industrial-grade temperature and humidity transmitters deployed around the production line to continuously extract the relative humidity index of the environment during the layup process; relying on intelligent power monitoring instruments connected to the production line control system to capture electrical parameters including active power, reactive power, and current changes in real time, and calculating the total energy consumption data per unit cycle for the corresponding production cycle by integration; using a globally unified time reference, aligning the collected pressure sequence, viscosity vector, and relative humidity index with the energy consumption data of the same cycle on the time axis, and constructing a multi-source feature correlation matrix through data encapsulation.

3. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The construction process of the pseudo-sample dataset includes: establishing a digital twin model of blade curing dynamics based on the coupling of Fourier's thermal conductivity law and resin curing kinetic equation to describe the dynamic evolution of heat transfer and exothermic chemical reaction within the mold cavity; setting multiple layup thickness gradients and initial reaction enthalpy ranges of glass fiber reinforced material as input variables for Monte Carlo random sampling to simulate the temperature field distribution and energy evolution of the mold cavity under different operating conditions; extracting the total heat flux per unit time during the simulation process and fitting a virtual energy consumption distribution curve; synchronously monitoring the rate of temperature rise at the center during the simulation process, and automatically associating the simulated samples with thermal runaway operating condition labels when the local temperature exceeds the preset thermal stability threshold, thus constructing a pseudo-sample dataset containing mechanistic features and label attributes.

4. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The process of extracting the centroid coordinates of local energy consumption includes: calculating the nonlinear correlation entropy between each process feature dimension and the total energy consumption per unit cycle, and determining the mutual information weight based on the entropy value; recalibrating the feature dimensions of the pressure sequence, viscosity vector, and relative humidity index; projecting the recalibrated measured feature vectors and the mechanism pseudo-sample dataset together onto a high-dimensional energy consumption feature space; using a clustering algorithm to perform affinity search on the sample points in the space to identify multiple subspace clusters representing different process stages; calculating the spatial geometric centroid or weighted average coordinates of all sample points in each subspace cluster, and extracting the centroid coordinates of the local energy consumption pattern.

5. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The generation process of the global energy efficiency model includes: using a set of randomly generated projection vectors to map the centroid coordinates of local energy consumption patterns to binary space; generating a hash code sequence composed of 0s and 1s by calculating the dot product of the coordinate points and the projection vectors and performing binarization; using the proximity preservation property of the Locality Sensitive Hash (LSH) algorithm to ensure that centroid coordinates that are close in distance in the original energy consumption feature space have the same hash code with a higher number of bits after mapping; the server receives hash code sequences from different production lines, calculates the Hamming distance between each sequence, and divides sequences with the same and / or similar bit patterns into corresponding hash buckets according to the distance; statistically analyzes the distribution density and energy efficiency weight of samples in each hash bucket, and calculates and generates the global energy efficiency model by performing reverse mapping on the frequently occurring hash bit patterns.

6. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The correction process of the global energy efficiency model includes: fitting the centroid coordinates of the mechanism pseudo-sample dataset and the local energy consumption mode using kernel density estimation to generate the corresponding theoretical probability distribution function P and measured probability distribution function Q; traversing the sampling points in the high-dimensional feature space to calculate the expected value of the log-likelihood difference between distribution P and distribution Q, and obtaining the KL divergence value; adding the KL divergence value as a regularization constraint term to the loss function of the global energy efficiency model; and iteratively optimizing the closed-loop correction of the model prediction bias.

7. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The process of automatically triggering regional energy quota evaluation feedback on the blockchain based on the index includes: the smart contract calls the hash-encrypted feature vector uploaded by each factory, calculates the Hamming distance between the hash-encrypted feature vector and the high-energy-efficiency cluster center in the global energy efficiency model, and obtains the consistency evaluation parameters; the consistency evaluation parameters are weighted and summed with the total energy consumption per unit period to generate the energy credit index of each manufacturing factory; the smart contract compares the energy credit index with the quota adjustment threshold stored in the blockchain ledger in real time, and automatically adjusts the energy quota weight factor of the factory on the blockchain according to the comparison deviation; the smart contract synchronizes the adjusted energy quota weight factor to the distributed nodes of the regional energy regulatory platform to complete the closed-loop evaluation feedback of energy quotas.

8. The industrial energy consumption prediction method based on parametric clustering according to claim 1, characterized in that, The process of outputting the energy prediction value includes: extracting the geometric dimension parameters and structural ply parameters of the blades within the production cycle to be predicted; using the same projection matrix as that used to generate the hash encryption feature vector, mapping the extracted parameters to a high-dimensional encryption computation domain to generate the hash feature vector to be predicted; inputting the hash feature vector to be predicted into a calibrated global energy efficiency model, and searching for the target energy consumption pattern cluster with the smallest Hamming distance in the hash space; retrieving the energy efficiency centroid coordinates associated with the target energy consumption pattern cluster, and performing a product correction in conjunction with the preset rated power of the equipment and the production task of each production line to obtain the energy consumption prediction score of a single production line; summing the energy consumption prediction scores of all production lines in the cluster, and outputting the predicted energy consumption value of the entire blade manufacturing cluster within the prediction cycle.

9. An industrial energy consumption prediction system based on parameter clustering, characterized in that, include: Data acquisition module: acquires the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector during the vacuum infusion stage, and the relative humidity index of the layup environment, and synchronously associates it with the corresponding total energy consumption data per unit cycle. Twin Simulation Module: Generates a pseudo-sample dataset through a digital twin model of the blade; uses a parametric clustering algorithm to calculate the surface pressure sequence inside the blade mold cavity, the resin flow viscosity vector, the relative humidity index, and the mutual information weights of the pseudo-sample dataset, and extracts the coordinates of the local energy-consuming centroid. Encryption module: The Locality Sensitive Hash (LSH) algorithm is used to map the local energy centroid coordinates to a high-dimensional encryption computation domain to generate a hash encryption feature vector; the hash encryption feature vectors uploaded by each server are received and secondary clustering is performed in the hash space to generate a global energy efficiency model; Prediction module: Calculates the KL divergence between the feature distribution of the pseudo-sample dataset and the centroid coordinates of the measured local energy consumption pattern, and uses the KL divergence as a bias constraint term to correct the global energy efficiency model; The system uploads the consistency evaluation parameters of the feature vectors, calculates the energy credit index of each manufacturing plant, and automatically triggers the evaluation feedback of regional energy quotas on the blockchain based on the index. Based on the blade specification parameters of the production cycle to be predicted, the system calls the calibrated global energy efficiency model and outputs the predicted energy value.

Images