An Industrial Production-Energy Coupled Forecasting Method Integrating Planning and Multidimensional State Information

Abstract

This invention discloses an industrial production-energy coupled prediction method that integrates planning and multi-dimensional state information, belonging to the field of information technology. First, a wavelet analysis-based matching method is used to extract rhythmic and waveform features from industrial energy data. Second, a contrastive learning model is constructed using a probabilistic transition contrastive learning algorithm to perform correlation and uncertainty analysis on multi-sequence feature data and encode it. Finally, the output of the contrastive learning model is concatenated with production planning data and used as input to a physical information neural network (PIN), which outputs the prediction result. The production plan is transformed into a partial differential equation and embedded in the loss function of the PSN, enabling the PSN to dynamically adjust the prediction result according to the production plan, ultimately obtaining a prediction result that satisfies the planning constraints. This invention offers high prediction accuracy, and the prediction duration meets the application requirements of industrial sites, providing decision-making guidance for subsequent energy optimization scheduling.

Description

Technical Field

[0001] This invention belongs to the field of information technology and relates to an industrial production-energy coupling prediction method that integrates planning and multi-dimensional state information. Background Technology

[0002] In process industries, production processes and energy systems are inextricably linked. Uncertainties in the production process often lead to fluctuations in energy production and consumption, disrupting the energy system's balance and resulting in energy emissions and carbon spills, impacting equipment operation and economic efficiency. Therefore, if we can anticipate changes in energy status caused by each production stage and achieve a dynamic balance between energy generation, consumption, and storage through the rational scheduling of energy buffering and conversion equipment, it will help ensure energy supply to production stages and improve energy utilization efficiency. Thus, accurate prediction of energy operating status is crucial for effective energy scheduling.

[0003] To improve the effectiveness of industrial energy forecasting, existing research has focused on constructing the intervariate dependencies of multidimensional energy data and has developed different deep neural network models, including: a deep neural network (DNN) framework based on a one-dimensional convolutional neural network and combined with WaveNet and efficiency network (Ko M, Lee K, Hur K. (2022). Feedforward Error Learning Deep Neural Networks for Multivariate Deterministic Power Forecasting[J].IEEE Transactions on Industrial Informatics, 18(9), 6214-6223); and a multi-scale adaptive graph neural network (MAGNN) based on a multi-scale pyramid network design capable of inferring intervariate dependencies at specific scales (Chen L. (2023). Multi-Scale Adaptive Graph Neural Network for Multivariate Time Series Forecasting[J].IEEE Transactions on Knowledge and Data Engineering, 35(10). 10748-10761), A method for multivariate time series spatiotemporal dynamic modeling based on neural constant differential equations and dynamic graph structures (Jin M, Zheng Y, Li Y, Chen S, Yang B, Pan S. (2023). Multivariate Time Series Forecasting With Dynamic Graph Neural ODEs[J].IEEE Transactions on Knowledge and Data Engineering, 35(9), 9168-9180), A complex neural network (ComNN) that captures the coupling relationship between multiple loads by taking aggregated multi-source information as input (Zhao P. (2024). Geometric Loss-Enabled Complex Neural Network for Multi-Energy Load Forecasting in Integrated Energy Systems[J].(IEEE Transactions on Power Systems, 39(4), 5659-5671). However, the above methods use purely data-driven deep learning models, which are not applicable to problems with prior knowledge of physical laws.

[0004] In recent years, some scholars have considered methods for fusing mechanisms and data, and have conducted research on this topic in conjunction with real-world problems. For example, physical information neural networks (PINs) incorporating self-attention mechanisms have been used to predict the remaining useful life of aircraft engines (Liao X, Chen S. (2023). Remaining useful life with self-attention assisted physics-informed neural network [J]. Advanced Engineering Informatics, 58). PINs combining sensitive features and fatigue life prediction models have been used to predict the fatigue life of materials (He G, Zhao Y, Yan C. (2023). Multiaxial fatigue life prediction using physics-informed neural networks with sensitive features. Engineering Fracture Mechanics, 289). However, industrial processes differ from general multivariate predictions. The relationships between energy equipment are mainly affected by the synchronous and asynchronous operation of the equipment. The similarity between multiple energy variables is a key part of correlation analysis. Furthermore, the mechanistic knowledge in industrial settings often appears in the form of plans, making it difficult to directly integrate into prediction models. Therefore, reasonable multivariate analysis and mechanism embedding methods may be more beneficial for industrial energy prediction. Summary of the Invention

[0005] This invention proposes an industrial production-energy coupling prediction method that integrates planning and multi-dimensional state information. This prediction method has high accuracy and can provide more reliable guidance for the optimal scheduling of industrial energy.

[0006] The technical solution of the present invention is as follows:

[0007] An industrial production-energy coupled forecasting method integrating planning and multi-dimensional state information includes: First, for continuously fluctuating industrial energy data, a wavelet analysis-based matching method is used to extract rhythmic and waveform features of the industrial energy data, thereby re-stacking the one-dimensional time series into multi-series feature data and achieving data augmentation; Second, a contrastive learning model is constructed using a probabilistic transition contrastive learning algorithm to perform correlation and uncertainty analysis on the multi-series feature data and encode it to capture the interrelationships between features of different dimensions; Finally, the output of the contrastive learning model is concatenated with the production planning data and used as the input to a physical information neural network, outputting the prediction result. The production plan is transformed into a partial differential equation and embedded in the loss function of the physical information neural network, enabling the physical information neural network to dynamically adjust the prediction result according to the production plan, ultimately obtaining a prediction result that satisfies the planning constraints. Specifically, the method includes the following steps:

[0008] Step 1, Feature Extraction and Data Augmentation:

[0009] Step 1.1, rhythm feature extraction:

[0010] Industrial energy data has rhythmic characteristics or can be segmented. The resulting energy curve is composed of energy curve segments with clear boundaries. Let the interval corresponding to the energy curve segment with a value of non-zero be the rhythm window. Perform rhythm matching on the energy curve to obtain the interval distribution position of the rhythm window, which serves as the rhythm feature.

[0011] For each rhythm window, following the wavelet transform approach, we first set the rhythm matching mother function and the transformation operations on the basis functions. To capture the distribution of the rhythm windows in the time domain and the waveforms within the rhythm windows, a convex Haar wavelet function is used as the mother function. As shown in equation (1), basis functions for energy curve segments within each rhythm window are formed by translating and scaling the generating function in the time domain. The transformation process is shown in equation (2):

[0012] (1)

[0013] (2)

[0014] in, Represents the time index of the energy curve. Represents the time-domain scaling factor. Represents the time-domain shift coefficient. The sensitivity coefficient, Indicates the upper and lower limits of the expansion / contraction amount. This indicates the upper and lower limits of the translation amount. To ensure that each basis function can independently target a curve segment within a rhythm window during rhythm matching, the stretching of the generating function does not exceed twice the average rhythm window duration.

[0015] basis functions With the energy curve Perform matching operations to obtain the matching degree matrix. The calculation process is shown in equation (3).

[0016] (3)

[0017] in, Represents the energy curve The time series length, and the elements of the matching degree matrix measure the basis function. Similarity to the energy curve.

[0018] Will Unfolding into a 3D image, the three axes of the 3D image represent the time domain dimension, the frequency domain dimension, and the matching value, respectively. The time domain dimension represents the time-shifted value of the basis function, and the frequency domain dimension represents the scaling value of the basis function. Since the energy curve is composed of curve segments within rhythm windows of different waveforms, the 3D image exhibits a clear layering phenomenon in the time domain direction. Considering that this layering may contain the rhythmic characteristics of industrial energy data, the 3D image is projected onto the plane perpendicular to the frequency domain axis to obtain the projected curve. The projected curve is a curve composed of alternating peaks and troughs. Each peak position contains an extreme point, which corresponds one-to-one with the local extrema of the 3D image, representing the local maximum matching degree between the basis function and the energy curve. At this point, the distribution positions of the basis function and the curve segments within the rhythm window coincide in the time domain (maximum similarity). Therefore, the coordinates of the extreme points in the projected curve in the 3D image are... This reflects the distribution of the rhythm window, among which Indicates the number of extreme points. Represents the time-domain coordinates of the extreme point. The frequency domain coordinates of the extreme points are represented. The coordinates of the extreme points in the three-dimensional image are searched using equations (4)-(6), which require combining the projection curve and the three-dimensional image.

[0019] (4)

[0020] (5)

[0021] (6)

[0022] in, For different time-domain translation coefficients Frequency domain coordinates of the extreme points of the matching value. For projection curves, The radius of the neighborhood of the extreme point. For Center point The frequency domain coordinates of the extreme points in the neighborhood of the radius. .

[0023] The coordinates of the extreme points obtained by the search In reality, this refers to the start time and duration of curve segments within each rhythm window. For energy curves with curve segments containing values ​​of 0, to better characterize the rhythmic characteristics of the energy curve, [the following is used:] Distribution characteristics of curve segments converted to zero numerical values As the final rhythmic feature, the transformation process is shown in equations (7) and (8).

[0024] (7)

[0025] (8)

[0026] in, This represents the midpoint of the zero-value curve segment. This indicates the duration of the zero-value curve segment.

[0027] For energy curves that do not have zero-value curve segments, calculate the value for each rhythm window. , directly As the final rhythmic feature, for the sake of a unified description, the letters are... and Replace each with a letter and It indicates that it will soon Replace with This indicates that it serves as the final rhythmic characteristic.

[0028] Step 1.2, Waveform Feature Extraction:

[0029] For energy curves that do not have zero-value curve segments, they not only have obvious rhythmic features, but also the waveform features of curve segments within each rhythm window, such as waveform amplitude, waveform area, and fluctuation trend, are equally important. They reflect semantic features related to the production process, such as energy impact and total energy accumulation. Therefore, based on the extraction of rhythmic features, it is necessary to further extract waveform features from the curve segments within each rhythm window.

[0030] Waveform matching is performed following the approach of wavelet matching. To fit various trends in the data, the beta distribution function is set as the waveform matching generating function. See equation (9).

[0031] (9)

[0032] in, and Adjustable parameters for waveform matching of the generating function, used to change the waveform shape of the generating function. It is the integral variable.

[0033] Since the waveform amplitude, waveform area, and fluctuation trend of the energy curve segment differ within each rhythm window, the transformation method of the generating function is set to vertical translation, vertical stretching, and parameter adjustment operations, thereby obtaining the waveform matching basis function. Let the waveform parameters of the waveform matching basis function be... Then the waveform matching basis function The transformation process can be described by formula (10).

[0034] (10)

[0035] in, This represents the longitudinal translation amount, serving as the translation parameter for the waveform matching basis function. This represents the longitudinal stretching amount, serving as the stretching parameter for the waveform matching basis function. Indicates the upper and lower limits of translation. Indicates the upper and lower limits of stretching. This represents a set of waveform parameters.

[0036] The waveform amplitude and area enclosed by the horizontal axis of the waveform matching basis functions obtained under different translation and stretching conditions of the waveform matching generating function also differ. Furthermore, the waveform trend of the waveform matching basis functions changes under different waveform parameters. Similar to the rhythm feature extraction process, in order to obtain the matching degree matrix... Match the waveform basis function with the energy curve segment within the rhythm window. Perform a matching operation, the process of which is shown in equation (11).

[0037] (11)

[0038] in, As a vector dot product operator, the higher the similarity between the waveform matching basis function vector and the curve segment vector, the higher the calculated matching degree.

[0039] The waveform matching basis function vector and the energy data segment vector have the maximum matching degree at the maximum point of the matching degree matrix. At this point, the waveform matching basis function vector and the energy data segment vector have the maximum matching degree. The vectors have the highest similarity. The process of searching for the coordinates of the maximum value point is shown in equation (12).

[0040] (12)

[0041] in, and Let be the longitudinal translation, longitudinal stretch, and waveform parameters of the waveform matching basis function at the maximum point, respectively. It can be used to represent the amount of energy impact. , It can be used to represent energy change trends and energy accumulation. The calculation is performed using the L2 norm of the energy curve segment.

[0042] However, during the matching process, the matrix The search space in the vertical translation dimension is too large, so the translation amount is calculated directly. This affects computational efficiency. Therefore, the translation parameters should be kept constant initially. Unchanged, for tensile parameters and waveform parameters Perform matching and obtain Assuming that the curve segment within the rhythm window is equal to the energy accumulation represented by the waveform matching basis function, then construct equation (13) using... To calculate energy impact .

[0043] (13)

[0044] in, Indicates length is And all elements are The constant vector, i.e. , The longitudinal translation is 0 and the longitudinal stretch is And the waveform parameters are The waveform matching basis function vector.

[0045] Integration With the calculated The final result of waveform matching can be obtained. The rhythm matching and waveform matching results are combined into a feature. This will enhance energy data.

[0046] For multidimensional energy data, feature extraction is performed on different dimensions to obtain its features. subscript Indicates the rhythm window index. For dimensional indexing.

[0047] Step 2, Correlation analysis of industrial energy data based on contrastive learning:

[0048] The generation and consumption of industrial energy exhibit asynchronous fluctuations in multidimensional energy data, and are also affected by upstream and downstream processes as well as human factors, resulting in uncertainty in energy fluctuations. To address this characteristic, a contrastive learning model is constructed using the Contrastive Learning with Probabilistic Transitions (CLPJ) algorithm to simultaneously perform correlation and uncertainty analysis on multidimensional energy data and encode it.

[0049] The CLPJ algorithm module includes an old encoder, a sample reconstruction module, and a new encoder, based on the enhanced multidimensional energy data features from step 1. An initial sample set is constructed, containing anchor samples, positive samples, and negative samples. The old encoder is trained based on this initial sample set, and the encoding results are provided to improve the similarity between the anchor samples and positive samples in their encoded vectors, and to minimize the similarity between these anchor samples and negative samples in their encoded vectors. Considering the uncertainty of multidimensional energy data, the sample reconstruction module redistributes the initial sample set according to the encoding results. The newly redistributed sample set is used to train the new encoder, and the final feature encoding is provided by the new encoder. Specifically:

[0050] Step 2.1, Construct the initial sample set:

[0051] To obtain positive and negative samples, the sensitivity coefficient of rhythm matching is first adjusted. Neighborhood radius And the sampling resolution of waveform parameters of waveform matching basis function Two different sets of features were obtained from all energy curve segments of data from different dimensions. And use them as augmentation samples to form an augmentation sample set. ,in For anchor samples, As a positive sample, Given the dimension of the energy data, the augmented sample set contains augmented sample pairs of data from different dimensions. For the first For a given anchor sample in dimension , its negative samples are derived from all other anchor samples except the first anchor sample. The negative sample set is composed of samples from the remaining dimensions outside the first dimension. Using anchor sample sequences Positive sample sequence and negative sample set sequence Construct the initial sample set.

[0052] Step 2.2: The Old encoder is trained based on the initial sample set and the encoding results are given.

[0053] The old encoder is a neural network encoder. and network projector Composition. Samples from the initial sample set are processed by a neural network encoder. Obtain the hidden vector Then the network projection head calculates the encoding vector. The encoding result is obtained. Since the Old encoder identifies samples from different dimensions of data by minimizing the loss function, the loss function of the Old encoder is constructed using the encoding results of anchor samples, positive samples and negative samples, as shown in Equation (14).

[0054] (14)

[0055] in and Positive sample pairs Compared with negative sample pairs The encoded vector, The set of encoding results Indicates the temperature coefficient. Indicates batch size. This is the cosine similarity operator.

[0056] Minimize the loss value using an optimization algorithm. Once the value is less than the set value, the training of the Old encoder is complete. After training, the Old encoder has the ability to analyze the similarity of multidimensional energy data. Then, the samples in the initial sample set are processed by the Old encoder after training to calculate their encoding vectors.

[0057] Step 2.3: Reconstruct the sample based on the encoding results:

[0058] The energy production and consumption process is inherently uncertain, potentially leading to difficult-to-identify negative samples. Therefore, directly treating the augmented samples of the remaining dimensions as negative samples is inappropriate. Based on the encoding vector calculated in step 2.2, negative samples that exhibit similar behavior to the anchor samples in the encoding vector should be mined into the positive samples for sample reconstruction.

[0059] To uncover potential negative samples, the sample reconstruction module redistributes the initial sample set based on the encoded vectors calculated by the old encoder, introducing uncertainty to obtain a new sample set. Specifically:

[0060] The sample reconstruction module consists of two parts: an energy evaluation module and a probability transition module. First, the energy evaluation module evaluates all negative samples. Relative to anchor sample energy value As shown in equation (15), the greater the similarity between the encoding vectors of the negative sample and the anchor sample, the higher the energy value. Then, the probability transition module calculates the probability that the negative sample will transition to a positive sample based on the energy value of the negative sample. As shown in equation (16), the transition process is simulated by random sampling.

[0061] (15)

[0062] (16)

[0063] in, An operator for calculating the cosine similarity between two encoded vectors. Energy threshold This is a hyperparameter, and can be any positive number.

[0064] The process of constructing a new sample set through the transition of negative samples in each dimension is as follows: the initial sample set and the new sample set remain unchanged in the anchor sample part, and the transition samples that successfully transition are... Add positive sample sequences In the process, the positive sample set sequence is obtained. ,Right now The new negative sample set sequence This includes the negative samples remaining after the successful transition samples, i.e. At this point, the new sample set contains anchor sample sequences. Positive sample set sequence and negative sample set sequence .

[0065] Step 2.4: The new encoder is trained based on the new sample set and the encoding results are given.

[0066] To analyze the correlation and uncertainty of multidimensional energy data, the new encoder is trained using a new sample set. The new encoder also consists of a neural network encoder. With network projector The new sample set is composed of samples processed by a neural network. Obtain the hidden vector The encoding result is then calculated by the projector. .

[0067] Unlike the initial sample set, the new sample set includes anchor samples. Positive samples Transition Samples and negative samples The purpose of model learning is to bring anchor samples closer to the model. The distance is used to exclude negative samples, and to distinguish between positive samples and transition samples, the loss function of the new encoder incorporates transition confidence weights. As the multiplier factor of the transition sample, as shown in equations (17) and (18).

[0068] (17)

[0069] (18)

[0070] Among them, the transition confidence weight The numerical value represents the transition probability of the transition sample. To calculate two vectors Similarity operators.

[0071] To accelerate the data analysis process, at fixed iteration intervals, the parameters of the New encoder corresponding to the neural network encoder and the network projector are softly updated to the Old encoder, as shown in Equation (19).

[0072] (19)

[0073] in, The parameters representing the Old encoder, The parameters representing the new encoder, This represents the soft update factor.

[0074] Minimize the loss value using an optimization algorithm. Once the value falls below a set threshold, training of the new encoder is complete. After training, only the trained neural network encoder is saved as a contrastive learning model. and its output latent vector sequence Set as a contrastive learning model The model results.

[0075] For single-dimensional energy data, its characteristics The same method was used to obtain the contrastive learning model. and output latent vectors. .

[0076] Step 3: Embed the physical information neural network of the production plan:

[0077] PINN model constructed using Physical Information Neural Network (PINN) The contrastive learning model in step 2 With PINN model The prediction model for industrial energy is constructed, and the input to the prediction model is the enhanced features from step 1. The features are then processed by the contrastive learning model. The obtained output is concatenated with the production plan data to form the PINN model. Input, The output is a predicted value for industrial energy. In addition, the production plan is transformed into a partial differential equation, which serves as the PINN model. This is part of the loss function, used to embed production planning into the PINN model. By minimizing the loss function, the PINN model learns the relationship between the production plan and its output. The production plan is a batch number plan, defined for a future period. Production batches within .

[0078] For unidimensional energy data, its enhanced features from step 1 are... Input contrastive learning model Output latent vectors In order to couple production planning information, the latent vectors are... Production planning data The splicing result is used as the PINN model. The input, including production planning data Indicates by Remaining time after the current energy state With remaining production batches , Changes in output energy data As the output of the prediction model, see equation (20). Combination The future features are calculated using the transfer equation, i.e., equation (21). After completing feature transfer, iterative prediction is performed by alternating between equations (20) and (21) to obtain the predicted sequence. Equation (22) is the formula for constraining the production plan to future energy fluctuations, where the first equation represents the initial production plan and the second equation is the dynamic form of the production plan in the iterative prediction process.

[0079] (20)

[0080] (twenty one)

[0081] (twenty two)

[0082] in, For characteristic transfer function, For the planning constraint function, For the dynamic changes of planning constraints, These are the parameters for the PINN model.

[0083] Production planning, as an overall constraint over a period of time, cannot directly affect each step of the forecasting process. To distribute constraints to each step of the forecasting process and limit the output of the forecasting model so that the forecast results meet the overall requirements of the production plan, an average characteristic is defined. With average output as follows:

[0084] Definition 1: If all energy characteristics Replaced with If the production plan is still met, then It is considered an average characteristic.

[0085] Definition 2: If Predicted features for the next iteration cycle ,but It is considered to be the average output.

[0086] Therefore, if within the remaining planned time All model outputs Then the iterative prediction result See equation (23), at this time the prediction result sequence The constraints are satisfied, as shown in equation (24).

[0087] (twenty three)

[0088] (twenty four)

[0089] Let constraint functions be defined. ,but This reflects the deviation between the model output and the production plan. The closer The more the predicted sequence satisfies the constraints, The closer to 0, the better. To embed production planning information, select variables. , For the number of selected variables, the equation Differentiation yields the partial differential equation, as shown in equation (25), where, It must include production planning data. This indicates the impact of the production plan on the output.

[0090] (25)

[0091] Embed the partial differential equation, i.e., equation (25), into the loss function of the PINN model. We learn physics knowledge from the middle, as shown in equations (26)-(28), where equation (26) is used to calculate the physical driving loss. Data-driven loss Used to measure the error between model output and fluctuations in real energy data.

[0092] (26)

[0093] (27)

[0094] (28)

[0095] in, As a balance factor, This indicates the actual amount of waveform change.

[0096] For the characteristic sequence of multidimensional energy data Input it into the prediction model, and then... The hidden vector sequence is then obtained. The concatenation result of each latent vector and the corresponding dimension of the production plan data is used as the PINN model. The input is a sequence of changes in multidimensional energy. ;Will and By combining these features, the next fluctuation state is obtained through feature transfer. The long-term prediction results of multidimensional energy are obtained through multiple iterations. For the production planning embedding problem of multidimensional energy data, since the production plans of each dimension are independent of each other, the loss value of each dimension is calculated according to equations (26)-(28), and the total loss is calculated according to equation (29). .

[0097] (29)

[0098] in For the calculated first Loss in each dimension.

[0099] Minimize loss value Or total loss value The PINN model training stops when the value falls below a set threshold, resulting in the final prediction model used to obtain the final prediction results, which meet the requirements of the production plan.

[0100] The beneficial effects of this invention are as follows: Because this invention constructs an intuitive and interpretable data structure, such as energy accumulation, energy surge, and energy fluctuation rhythm, it has advantages in judging energy fluctuation and seasonal variation patterns. In addition, this invention analyzes the synchronous and asynchronous relationships of different devices through comparative learning, that is, it analyzes and encodes samples of different dimensions, and uses PINN to embed the production plan into the loss function, which is conducive to making full use of planning information to improve the energy forecasting effect, and can provide richer information and more reliable support for the formulation of industrial energy dispatching schemes. Attached Figure Description

[0101] Figure 1 This is a flowchart of the present invention.

[0102] Figure 2 The graph shows the converter gas recovery rate curve, where (a) is the recovery flow rate curve and (b) is a schematic diagram of the rhythm characteristics.

[0103] Figure 3 The diagrams show the matching and projection results in rhythm feature extraction, where (a) is a diagram of the matching results and (b) is a diagram of the projection results.

[0104] Figure 4 The graph shows the consumption curve of blast furnace gas for hot blast stoves, where (a) is the consumption curve and (b) is the waveform characteristics.

[0105] Figure 5 The diagrams illustrate the changes in the waveform matching basis function, where (a) shows the longitudinal stretching and translation process, and (b) shows the waveform transformation process.

[0106] Figure 6 This is a framework diagram of a probabilistic transition contrastive learning model.

[0107] Figure 7 This is a diagram of the sample transition process.

[0108] Figure 8 A neural network framework diagram for embedding physical information of planning constraints.

[0109] Figures 9(a), 9(b), and 9(c) are the predicted results of converter gas recovery for three sets of data (1#LDG, 2#LDG, and 3#LDG) in the comparative experiment of the example, respectively.

[0110] Figure 10To compare the prediction results of blast furnace gas consumption in the experiment, (a), (b), (c), (d) and (e) show the prediction effects of five methods for total blast furnace gas consumption in this embodiment, Crossformer, MIC-RCNN-GRU, MOSMA-SVM and DMRI-Former, respectively, and (f) is the absolute error curve of a selected prediction result for each method.

[0111] Figure 11 This describes the characteristic transfer process for predicting converter gas. Detailed Implementation

[0112] The method of this invention is further illustrated and evaluated using real data from a domestic steel company as an example. Real-time flow data of converter gas LDG recovery and BFG consumption data of hot blast stove blast furnace gas, which fully reflect the characteristics of production rhythm, were selected for the experiment. The data was collected by a Supervisory Control and Data Acquisition (SCADA) system from April to June 2024, with a data sampling interval of 1 minute. The LDG data dimension was 3, and the BFG data dimension was 4. The energy data unit was km. 3 To verify the effectiveness of the method of this invention, the training set and test set were divided in a 7:3 ratio, and ablation experiments and comparative experiments were conducted. The experiments were carried out on an i9-13900K CPU and an NVIDIA RTX 4090 GPU. The model was built using the PyTorch framework, and the training platform version was Torch 1.12.1.

[0113] An industrial production-energy coupled forecasting method that integrates planning and multidimensional state information includes the following steps:

[0114] Step 1, Feature Extraction and Data Augmentation:

[0115] For continuously fluctuating energy data, a feature extraction method based on wavelet analysis is proposed to capture information such as the rhythm, energy impact, trend, and energy accumulation of energy data, transforming one-dimensional time series into feature sequences and realizing energy data enhancement that includes production process semantics.

[0116] Step 1.1, rhythm feature extraction:

[0117] In this embodiment, one dimension of the multi-dimensional data of converter gas recovery and hot blast stove blast furnace gas consumption is used as single-dimensional energy data to illustrate this step. For multi-dimensional data, different dimensions can be processed separately.

[0118] Energy data exhibits distinct rhythmic characteristics or can be segmented. Its energy curve is composed of segments with clearly defined boundaries. Let the intervals corresponding to non-zero curve segments be considered rhythm windows. Rhythm matching is performed on the energy curves to obtain the rhythm window segmentation results. Figure 2 Taking the converter gas generation curve shown in (a) as an example, the curve consists of multiple non-uniformly distributed peak segments and zero-value energy curve segments. Each peak segment represents the LDG recovery flow rate during a single steelmaking process, and the zero-value curve segment represents the period when no gas recovery is performed and production is in a break. That is, the peak segments correspond to the recovery stage, and the zero-value curve segments correspond to the break stage. Therefore, this energy curve consists of the recovery stage and the break stage. The energy curve in the recovery stage fluctuates with a sharp peak, and the value returns to zero in the break stage. For this type of energy data with zero-value curve segments, the rhythm matching result is the start time and duration of the recovery stage. And the midpoint and duration of the intermittent phases. ,like Figure 2 As shown in (b).

[0119] For each rhythm window, following the approach of wavelet transform, we first set the rhythm matching mother function and the transformation operations on the basis functions. To capture the distribution and waveform of the rhythm windows in the time domain, a convex Haar wavelet function is used as the mother function. As shown in equation (1), the basis functions of the curve segments within each rhythm window are formed by translating and scaling the generating function in the time domain. Basis functions The transformation process is shown in equation (2). Note that the stretching of the generating function does not exceed twice the average production period, which is 30 minutes for converter gas data.

[0120] basis functions With the energy curve Perform matching operations to obtain the matching degree matrix. The calculation process is shown in equation (3).

[0121] Will Unfolding into a 3D image, where the three axes represent the time domain dimension, the frequency domain dimension, and the matching value, respectively. The time domain dimension represents the time shift of the basis function, and the frequency domain dimension represents the scaling value. See [link to relevant documentation]. Figure 3 (a) Since the energy curve is composed of several curve segments, the three-dimensional image exhibits a significant layering phenomenon in the time domain. Considering that this layering may contain the rhythmic characteristics of the energy data, the three-dimensional image is projected onto the plane perpendicular to the frequency domain axis, resulting in the projected curve shown in Figure 1. Figure 3(b) The projected curve in the figure is a curve composed of alternating peaks and troughs. Each peak has an extreme point, which corresponds one-to-one with the local extrema in the 3D image. The extreme points in the figure indicate that the basis function and the energy curve have reached a local maximum matching degree. At this point, the distribution positions of the basis function and the curve segment in the recovery stage coincide in the time domain (maximum similarity). Therefore, the coordinates of the extreme points in the projected curve in the 3D image are... This reflects the distribution of the intervals in the recovery phase. In order to obtain specific coordinate values, it is necessary to search by combining the projection curve and the three-dimensional image. The extreme point search process is as shown in equations (4)-(6).

[0122] Search results In fact, these are the start times and durations of each recovery stage. To better characterize the rhythmic characteristics of the energy curve, equations (7) and (8) are used to... Characteristics of the transition to the intermittent phase As the final rhythmic feature, see Figure 2 (b)

[0123] For energy data where there are no zero-value curve segments in the energy curve, such as blast furnace gas consumption data for hot blast stoves, calculate the energy curve for each rhythm window. , directly As the final rhythmic characteristic, for the sake of a unified description, it will be and Replace with and It indicates that it will soon Replace with As its ultimate rhythmic characteristic.

[0124] Step 1.2, Waveform Feature Extraction:

[0125] For energy curves without zero-value curve segments, such as blast furnace gas consumption data, in addition to rhythmic features, waveform features such as waveform amplitude, waveform area, and fluctuation trend are equally important. They reflect semantic features related to the production process, such as energy impact and energy accumulation. Therefore, based on the extraction of rhythmic features, it is necessary to further extract waveform features from the curve segments within each rhythm window.

[0126] like Figure 4 As shown in (a), any peak and trough switching boundary of the blast furnace gas consumption curve corresponds to one furnace change, based on the rhythm characteristics obtained in step 1.1. ,like Figure 4As shown in (b), waveform matching is required for each energy segment within a rhythm window to obtain waveform features, including energy impact, energy accumulation, and trend information. The energy impact is the lowest amplitude of the energy curve segment within the interval, the energy accumulation is the area enclosed by the energy curve segment and the time axis, and the energy trend is a parameterized representation of the fluctuation trend of the energy curve segment. For energy data with zero-value curve segments, such as converter gas data, waveform feature extraction is unnecessary.

[0127] Waveform matching is performed following the approach of wavelet matching. To fit various trends in the data, the beta distribution function is set as the waveform matching generating function. See equation (9).

[0128] Because the waveform amplitude, waveform area, and fluctuation trend of the energy curve segment differ within each rhythm window, the transformation methods of the generating function include vertical translation, vertical stretching, and parameter adjustment operations, thereby obtaining the waveform matching basis function. Let the waveform parameters be... Then the waveform matching basis function The transformation process can be described by formula (10).

[0129] The waveform amplitude and the area enclosed by the horizontal axis of the waveform matching basis functions obtained under different translation and stretching conditions of the waveform matching generating function are different, such as... Figure 5 As shown in (a), the waveform trend of the basis function changes under different waveform parameters. Figure 5 (b) Similar to the rhythm feature extraction process, in order to obtain the matching degree matrix... Match the waveform basis function with the energy curve segment within the rhythm window. Perform a matching operation, the process of which is shown in equation (11).

[0130] The waveform matching basis function vector and the energy data segment vector have the maximum matching degree at the maximum point of the matching degree matrix. At this point, the waveform matching basis function vector and the energy data segment vector have the maximum matching degree. The vectors have the highest similarity. The process of searching for the coordinates of the maximum point is shown in equation (12), then the equation... It can be used to represent the amount of energy impact. , It can be used to represent energy change trends and energy accumulation. The calculation is performed using the L2 norm of the energy curve segment.

[0131] However, during the matching process, The search space in the vertical translation dimension is too large, for example, the gas consumption of a hot blast stove can reach 200 km. 3 / h, which means the longitudinal translation range is 0-250, and the translation amount can be calculated directly. This affects computational efficiency. Therefore, the translation parameters should be kept constant initially. The value remains unchanged at 0. Matching and searching are performed on the stretching parameters and waveform parameters to obtain... Assuming that the curve segment within the rhythm window represents the same amount of accumulated energy as the waveform matching basis function, equation (13) can be constructed, and then... To calculate the impact amount .

[0132] Integration With the calculated The final result of waveform matching can be obtained. The rhythm matching and waveform matching results are combined into a feature. .

[0133] For multidimensional data, feature extraction is performed on different dimensions separately to obtain its features. subscript Indicates the rhythm window index. This is a dimensional index for multidimensional data. Based on 3D converter gas data, the features obtained are: , Where 3 indicates that the data dimension is 3. The number of features or rhythm windows obtained from extracting converter gas data; based on 4D hot blast stove blast furnace gas data, the features obtained are: , Where 4 indicates that the data dimension is 4. To extract the number of features or the number of rhythm windows from the hot blast stove blast furnace gas data, all features waveform parameters .

[0134] Step 2, Correlation analysis of industrial energy data based on contrastive learning:

[0135] To address the correlations and uncertainties among industrial equipment, a contrastive learning approach is used to extract asynchronous patterns from energy data samples across different dimensions and analyze the correlations between multidimensional data. Based on this, a sample reallocation module with probabilistic transitions is proposed to simulate potential synchronization issues in equipment operating conditions.

[0136] Industrial energy generation and consumption exhibit asynchronous fluctuations across different dimensions of energy data. For example, the overlap rate of recovery stages in multiple converters is low, reducing energy shocks caused by simultaneous recovery. Energy fluctuations are uncertain due to the influence of upstream and downstream processes and human factors; for instance, the furnace switching time and duration of hot blast stoves typically fluctuate within a certain range. To address this characteristic, a contrastive learning algorithm based on probabilistic transitions is used to construct a contrastive learning model, simultaneously performing correlation and uncertainty analysis on multidimensional energy data and encoding it. Figure 6 As shown.

[0137] The CLPJ algorithm module includes an old encoder, a sample reconstruction module, and a new encoder, based on the enhanced multidimensional energy data features from step 1. An initial sample set is constructed, containing anchor samples, positive samples, and negative samples. The old encoder is trained based on this initial sample set, and the encoding results are provided to improve the similarity between the anchor samples and positive samples in their encoded vectors, and to minimize the similarity between these anchor samples and negative samples in their encoded vectors. Considering the uncertainty of multidimensional energy data, the sample reconstruction module redistributes the initial sample set according to the encoding results. The newly redistributed sample set is used to train the new encoder, and the final feature encoding is provided by the new encoder. Specifically:

[0138] Step 2.1, Construct the initial sample set:

[0139] To obtain positive and negative samples, the sensitivity coefficient of rhythm matching is first adjusted. Neighborhood radius And the sampling resolution of waveform parameters of waveform matching basis function Two different sets of features were obtained from all energy curve segments of data from different dimensions. And use them as augmentation samples to form an augmentation sample set. ,in For anchor samples, As a positive sample, Given the total number of dimensions in the energy data, the augmented sample set contains augmented sample pairs of data from different dimensions. For the first For a given anchor sample in dimension , its negative samples are derived from all other anchor samples except the first anchor sample. The negative sample set is composed of samples from the remaining dimensions outside the first dimension. Using anchor sample sequences Positive sample sequence and negative sample set sequence Construct the initial sample set.

[0140] Step 2.2: The Old encoder trains the model based on the initial sample set and provides the encoding results:

[0141] The old encoder is a neural network encoder. and network projector Composition. Samples from the initial sample set are processed by a neural network encoder. Obtain the hidden vector Then the network projection head calculates the encoding vector. The encoding results are obtained. Since the Old encoder guides the model to recognize samples from different dimensions of data by minimizing the loss function, the loss function of the Old encoder is constructed using the encoding results of anchor samples, positive samples, and negative samples, as shown in equation (14), where .

[0142] Minimize the loss value using an optimization algorithm. Once the value is less than the set value, the training of the Old encoder is complete. After training, the Old encoder has the ability to analyze the similarity of multidimensional energy data. Then, the samples in the initial sample set are processed by the Old encoder after training to calculate their encoding vectors.

[0143] Step 2.3: Reconstruct the sample based on the encoding results:

[0144] The energy production and consumption process is inherently uncertain, potentially leading to difficult-to-identify negative samples. Therefore, directly treating the augmented samples of the remaining dimensions as negative samples is inappropriate. Based on the encoding vector calculated in step 2.2, negative samples that exhibit similar behavior to the anchor samples in the encoding vector should be mined into the positive samples for sample reconstruction.

[0145] To uncover potential negative samples, the sample reconstruction module redistributes the initial sample set based on the encoded vectors calculated by the old encoder, introducing uncertainty to obtain a new sample set. Specifically:

[0146] The sample reconstruction module consists of two parts: an energy evaluation module and a probability transition module. First, the energy evaluation module evaluates all negative samples. Relative to anchor sample energy value As shown in equation (15), where the encoding vectors of negative samples and anchor samples are... The greater the similarity, the higher the energy value. Then, the probability transition module calculates the probability that a negative sample will transition to a positive sample based on its energy value. As shown in equation (16), the transition process is simulated by random sampling.

[0147] Figure 7 This describes the process of constructing a new sample set through transitions in each dimension. In the diagram, the initial and new sample sets remain unchanged in the anchor sample portion, while the transition samples represent successful transitions. Add positive sample sequences In the process, the positive sample set sequence is obtained. ,Right now The new negative sample set sequence This includes the negative samples remaining after the successful transition samples, i.e. At this point, the new sample set contains anchor sample sequences. Positive sample set sequence and negative sample set sequence .

[0148] Step 2.4: The new encoder trains the model based on the new sample set and provides the encoding results.

[0149] To analyze the correlation and uncertainty of multidimensional energy data, the new encoder is trained using a new sample set. The new encoder also consists of a neural network encoder. With network projector The new sample set is composed of samples processed by a neural network. Obtain the hidden vector The encoding result is then calculated by the projector. .

[0150] Unlike the initial sample set, the new sample set includes anchor samples. Positive samples Transition Samples and negative samples The purpose of model learning is to bring anchor samples closer to the model. The distance is used to exclude negative samples, and to distinguish between positive samples and transition samples, the loss function of the new encoder incorporates transition confidence weights. The loss function of the new encoder is shown in equations (17) and (18), which serve as the multiplier factor for the transition sample.

[0151] To accelerate the data analysis process, every fixed iteration cycle, such as 20 iterations, the parameters of the New encoder corresponding to the neural network encoder and the network projector are softly updated to the Old encoder, as shown in Equation (19).

[0152] Minimize the loss value using an optimization algorithm. Once the value falls below a set threshold, training of the new encoder is complete. After training, only the trained neural network encoder is saved as a contrastive learning model. and its output latent vector sequence Set as a contrastive learning model The model results are used as input for downstream models.

[0153] Saved contrastive learning model This is used to encode feature samples of converter gas and blast furnace gas, with the model's output dimension set to 10. Therefore, for converter gas features... , Their outputs are respectively For the characteristics of blast furnace gas data in hot blast stoves , Then its output .

[0154] Step 3: Embed the physical information neural network of the production plan:

[0155] PINN model constructed using Physical Information Neural Network (PINN) The contrastive learning model in step 2 With PINN model The prediction model for industrial energy is constructed, and the input to the prediction model is the enhanced features from step 1. The features are then processed by the contrastive learning model. The obtained output is concatenated with the production plan data to form the PINN model. Input, The output is a predicted value for industrial energy. In addition, the production plan is transformed into a partial differential equation, which serves as the PINN model. Part of the loss function is used to embed production planning into the PINN model. By minimizing the loss function, the PINN model learns the relationship between the production plan and the PINN model's output. Figure 8 As shown. The production plan is a batch production plan, which defines a specific period of time in the future. Production batches within .

[0156] For unidimensional energy data, its features are enhanced in step 1. Input contrastive learning model Output latent vectors In order to couple production planning information, the latent vectors are... Production planning data The splicing result is used as the PINN model. The input, including production planning data Indicates by Remaining time after the current energy state With remaining production batches , Changes in output energy data As the output of the prediction model, see equation (20). Combination The future features are calculated using the transfer equation, i.e., equation (21). After completing feature transfer, iterative prediction is performed by alternating between equations (20) and (21) to obtain the predicted sequence. Equation (22) is the formula for constraining the production plan to future energy fluctuations, where the first equation represents the initial production plan and the second equation is the dynamic form of the production plan in the iterative prediction process.

[0157] Production planning, as an overall constraint over a period of time, cannot directly affect each step of the forecasting process. To distribute constraints to each step of the forecasting process and limit the output of the forecasting model so that the forecast results meet the overall requirements of the production plan, an average characteristic is defined. With average output as follows:

[0158] Definition 1: If all energy characteristics Replaced with If the production plan is still met, then It is considered an average characteristic.

[0159] Definition 2: If Predicted features for the next iteration cycle ,but It is considered to be the average output.

[0160] Therefore, if within the remaining planned time All model outputs Then the iterative prediction result See equation (23), at this time the prediction result sequence The constraints are satisfied, as shown in equation (24).

[0161] Let constraint functions be defined. ,but This reflects the deviation between the model output and the production plan. The closer The more the predicted sequence satisfies the constraints, The closer to 0, the better. To embed production planning information, select variables. , For the number of selected variables, the equation Differentiation yields the partial differential equation, as shown in equation (25), where, It must include production planning data. This indicates the impact of the production plan on the output.

[0162] Embed the partial differential equation, i.e., equation (25), into the loss function of the PINN model. We learn physics knowledge from the middle, as shown in equations (26)-(28), where equation (26) is used to calculate the physical driving loss. Data-driven loss Used to measure the error between model output and fluctuations in real energy data.

[0163] For the characteristic sequence of multidimensional energy data Input it into the prediction model, and then... The hidden vector sequence is then obtained. To couple production planning information, each latent vector is associated with the planning data of its corresponding dimension. By splicing As the PINN model The PINN model outputs a multidimensional energy change sequence as input. ;Will and By combining these features, the next fluctuation state is obtained through feature transfer. The long-term prediction results of multidimensional energy are obtained through multiple iterations. For the production planning embedding problem of multidimensional energy data, since the production plans of each dimension are independent of each other, the loss value of each dimension is calculated according to equations (26)-(28), and the total loss is calculated according to equation (29). .

[0164] Minimize loss value Or total loss value Training of the PINN model stops once the value falls below a set threshold, yielding the final prediction model used to obtain the final prediction result. The training loss of the PINN module is calculated using automatic differentiation. By embedding the production plan into the prediction model's training process, the stability of the prediction model during iterative prediction is improved, ensuring that the final prediction result meets the requirements of the production plan.

[0165] Regarding the problem of converter recovery rate prediction Figure 11 This is a schematic diagram of the LDG recycling process. The production plan in the diagram specifies the planned time. Production batches within This is used to limit the distribution of recovery and intermittent phases within the planned timeframe. Since the recovery phases of LDG are relatively stable, the interval characteristics of the intermittent phases are adopted. As a rhythmic characteristic of the converter gas energy curve, the output of the PINN model in this prediction task is the time interval between the midpoint of the current intermittent phase and the midpoint of the next intermittent phase.

[0166] Figure 11 The diagram shows the relationship between production planning and feature transfer, due to the duration of the recovery phase. Generally, the fluctuations are not significant, so a fixed value is used. replace The production plan formula and transfer equation are shown in equations (30) and (31).

[0167] (30)

[0168] (31)

[0169] in The value is 15 minutes.

[0170] Based on the definitions of average characteristics and average output, the average duration of the zero-value curve segment in the average characteristics is... It can be scheduled by time and planned batches Calculate the average output Calculated from the average characteristics, see equation (32):

[0171] (32)

[0172] According to the constraint function Defined, then To reflect the impact of historical rhythms and production plans on gas recovery, the following was selected: As a variable ,but Given 3, the partial differential equation is obtained from equation (25), as shown in equation (33):

[0173] (33)

[0174] Considering the multidimensional energy data, with converter gas data having a dimension of 3, the PINN model loss function is further obtained, as shown in equations (34) and (35):

[0175] (34)

[0176] (35)

[0177] The sum of equations (34) and (35) is the total loss of the PINN model. By minimizing the total loss, when the model is updated 1000 times, the total loss value is less than the set value, and the training is completed and the PINN model is saved.

[0178] Regarding the prediction of blast furnace gas consumption for hot blast stoves, the production plan specifies the planning time. Production batches within This limits the number of rhythm windows within the planned timeframe. (Using features) As a feature of energy data, the output of the PINN model in this prediction task... for Since the output only contains the duration of the rhythm window If it is affected by the production plan and there are no intermittent periods, then the constraint function... .choose As a variable, the derivative yields the partial differential equation shown in equation (36).

[0179] (36)

[0180] The data-driven loss value is calculated according to equation (34), and the physical-driven loss value is calculated by combining equations (35) and (36). The two loss values ​​are summed and then optimized. When the model is updated 1000 times, the total loss value is less than the set value. The training is completed and the PINN model is saved.

[0181] To verify the effectiveness of the proposed method, the model prediction results were reconstructed into real-time data, and experimental metrics were statistically analyzed. The metrics used included root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and cumulative time bias (CTB), where CTB represents the prediction deviation at the time node of the last production batch within the planned time, reflecting the error accumulation during the iterative prediction process.

[0182] To evaluate the predictive capability of this method in industrial energy scenarios, comparative experiments were conducted using the kernel learning method MOSMA-SVM, multivariate prediction methods such as Cossformer, data pattern-related interpretable transformer networks (DMRI-Fomer), and the hybrid prediction method MIC-RCNN-GRU. All methods make predictions based on the extracted data features, and the inputs all include historical data and production plan data.

[0183] Table 1 summarizes the RMSE and MAE metrics of five methods on LDG#1, LDG#2, and LDG#3 recovery data. The table shows that this embodiment performs best on multiple datasets, exhibiting strong generalization ability and stability. While other methods perform well on LDG#2 and LDG#3 data, their accuracy differs significantly from that of LDG#1. The performance of multivariate prediction methods DMRI-Fomer, Cossformer, and MIC-RCNN-GRU is attributed to their lack of multidimensional correlation analysis for asynchronous patterns, insufficient analysis of asynchronous patterns in multiple LDG recovery processes, and failure to consider uncertainties during the recovery process. The predictive performance of MOSMA-SVM is inferior to other methods because it tends to overlook the impact of planning information on the next prediction output when selecting input factors through correlation analysis. Figures 9(a), 9(b), and 9(c) show the predicted results of converter gas recovery for the three sets of data (1#LDG, 2#LDG, and 3#LDG) for each method. The predicted results of LDG recovery rate in this embodiment are more in line with the actual situation.

[0184] Table 2 further summarizes the CTB of different methods. It can be seen that the cumulative prediction bias in this embodiment is significantly lower than other methods, indicating that embedding production planning using PINN is superior to the traditional method of simply using planning data as input. Unlike typical black-box models, this approach effectively captures the dynamic relationship between output variables and production planning variables, thereby reducing iterative errors caused by long-term forecasting.

[0185] Table 3 summarizes the RMSE, MAE, and MAPE indices of five methods on four sets of hot blast stove data. The table shows that this embodiment demonstrates the best prediction performance across multiple data sets, while other methods show significant differences in their indices across different data sets. This may be because these methods fail to identify the correlation between hot blast stoves and the uncertainty of energy fluctuations among multiple hot blast stove devices. Considering the impact of hot blast stoves on the BFG energy balance, the real-time consumption of multiple hot blast stoves is further superimposed, and the total consumption flow prediction curves of different methods are compared, as shown below. Figure 10 As shown. Figure 10Figures (a)-(e) show the total flow prediction results of different methods, where the actual BFG flow curve is divided into different time periods by dashed lines. The energy impact and waveform trend of the BFG differ significantly within each time period. As can be seen from the figures, the prediction result of this embodiment best matches the actual energy curve. Among other methods, DMRIFormer and MOSMA-SVM perform well in predicting the overall waveform trend, but their estimation of the magnitude of the energy impact is not as good as this method (as shown in t3, t4, t5). Conversely, Cossformer exhibits a large error in estimating the energy waveform trend, resulting in spikes in the prediction results (as shown in t1, t2, t4). In the prediction results of the MIC-RCNN-GRU method, the peaks of the energy waveform are misaligned with the actual curve (as shown in t1, t2, t4), possibly due to inaccurate estimation of the hot blast stove's switching rhythm. Figure 10 (f) The absolute error curve for the total predicted amount is further given. In the figure, the error of this embodiment can be maintained at 40 km for most of the time. 3 Within the / h range, and maintained a low level across all methods.

[0186] In addition, Table 4 summarizes the CTB indexes for different methods. Since the hot blast stove changing rhythm is stable, the time deviations of the five methods are not significant, but this embodiment still maintains the most stable prediction performance.

[0187] Based on the above experimental results, the prediction accuracy and CTB of this embodiment are superior to the comparative methods in different industrial energy data, demonstrating the superiority of this embodiment in the correlation analysis of multidimensional data and its embedding into production planning.

[0188] Table 1: Statistics of LDG recovery prediction indicators in comparative experiments

[0189]

[0190] Table 2: CTB statistics of LDG recovery in comparative experiments

[0191]

[0192] Table 3: Statistics of BFG Consumption Prediction Indicators in Comparative Experiments

[0193]

[0194] Table 4: CTB Statistical Results of BFG Consumption in the Comparative Experiment

[0195]

Claims

1. A method for industrial production-energy coupled forecasting that integrates planning and multidimensional state information, characterized in that, The process includes: First, for continuously fluctuating industrial energy data, a wavelet analysis-based matching method is used to extract rhythmic and waveform features from the industrial energy data, thereby re-stacking the one-dimensional time series into multi-series feature data; Second, a contrastive learning model is constructed using a probabilistic transition contrastive learning algorithm to perform correlation and uncertainty analysis on the multi-series feature data and encode it to capture the interrelationships between features of different dimensions; Finally, the output of the contrastive learning model is concatenated with production plan data and used as the input to a physical information neural network, which then outputs the prediction results. The production plan is transformed into a partial differential equation and embedded in the loss function of the physical information neural network, enabling the physical information neural network to dynamically adjust the prediction results according to the production plan, ultimately obtaining prediction results that meet the plan constraints. Rhythm feature extraction specifically involves: The energy curve of industrial energy data consists of energy curve segments with clear boundaries. Let the interval corresponding to the energy curve segment with a value of non-zero be the rhythm window. A convex Haar wavelet function is used as the generating function. The basis functions for energy curve segments within each rhythm window are formed by translating and scaling the generating function in the time domain. : (1) (2) in, Represents the time index of the energy curve. Represents the time-domain scaling factor. Represents the time-domain shift coefficient. The sensitivity coefficient, Indicates the upper and lower limits of the expansion / contraction amount. Indicates the upper and lower limits of the translation amount; basis functions With the energy curve Perform matching operations to obtain the matching degree matrix. : (3) in, Represents the energy curve The timing length; Will Unfolding the image into a 3D image, where the three axes are the time domain dimension, the frequency domain dimension, and the matching value, respectively, the 3D image is projected onto the plane perpendicular to the frequency domain axis to obtain the projection curve. The coordinates of the extreme points of each peak position of the projection curve in the 3D image are then searched. , Represents the time-domain coordinates of the extreme point. Represents the frequency domain coordinates of the extreme points. Indicates the number of extreme points; For energy curves with curve segments containing values ​​of 0, Distribution characteristics of curve segments converted to zero numerical values As the final rhythmic feature, among which This represents the midpoint of the zero-value curve segment. Indicates the duration of a zero-value curve segment; For energy curves that do not have a zero-value curve segment, As the final rhythmic characteristic, for the sake of a unified description, it will be Replace with express.

2. The industrial production-energy coupling prediction method according to claim 1, characterized in that, Specifically, the steps include the following: Step 1, Feature Extraction and Data Augmentation: Step 1.1, rhythm feature extraction; Step 1.2, Waveform Feature Extraction: For energy curves that do not have zero-value curve segments, following the wavelet matching approach, the beta distribution function is set as the waveform matching mother function: （9） in, and Adjustable parameters for waveform matching generating function. For integration variables; Let the waveform parameters of the waveform matching basis function be... Then the waveform matching basis function The transformation process is as follows: （10） in, Indicates the longitudinal translation amount. Indicates the longitudinal stretching amount. Indicates the upper and lower limits of translation. Indicates the upper and lower limits of stretching. Represents a set of waveform parameters; Match the waveform basis function with the energy curve segment within the rhythm window. Perform matching operations to obtain the matching degree matrix. : （11） in, It is a vector dot product operator; from Search for the coordinates of the maximum value point ,in, and Let be the longitudinal translation, longitudinal stretch, and waveform parameters of the waveform matching basis function at the maximum point, respectively. Used to indicate energy impact , Used to represent trends in energy change and cumulative energy levels. The final waveform matching result is obtained by using the L2 norm of the energy curve segment. Merge rhythm matching and waveform matching results into features. To enhance energy data; For multidimensional energy data, feature extraction is performed on different dimensions to obtain its features. subscript Indicates the rhythm window index. For dimension indexing; Step 2, Correlation analysis of industrial energy data based on contrastive learning: A contrastive learning model is constructed using the probabilistic transition contrastive learning (CLPJ) algorithm to simultaneously perform correlation and uncertainty analysis and encode multi-dimensional energy data. The CLPJ algorithm module includes an old encoder, a sample reconstruction module, and a new encoder. Step 2.1, Construct the initial sample set: First, adjust the sensitivity coefficient for rhythm matching. Neighborhood radius And the sampling resolution of waveform parameters of waveform matching basis function Two different sets of features were obtained from all energy curve segments of data from different dimensions. , And use them as augmentation samples to form an augmentation sample set. ,in For anchor samples, As a positive sample, The dimension of the energy data; for the first For a given anchor sample in dimension , its negative samples are derived from all other anchor samples except the first anchor sample. The negative sample set is composed of samples from the remaining dimensions outside the first dimension. Using anchor sample sequences Positive sample sequence and negative sample set sequence Construct the initial sample set; Step 2.2: The Old encoder is trained based on the initial sample set and the encoding results are given. The old encoder is a neural network encoder. and network projector Composition; the samples in the initial sample set are processed by a neural network encoder. Obtain the hidden vector Then the network projection head calculates the encoding vector. The encoding results are obtained; the loss function of the Old encoder is constructed using the encoding results of anchor samples, positive samples, and negative samples: （14） in and Positive sample pairs Compared with negative sample pairs The encoded vector, for The set of encoding results Indicates the temperature coefficient. Indicates batch size. For cosine similarity operator; Minimize the loss value using an optimization algorithm. Once the value is less than the set value, the training of the Old encoder is complete. The samples in the initial sample set are processed by the Old encoder after training to calculate their encoding vectors. Step 2.3: Reconstruct the sample based on the encoding results: The sample reconstruction module includes an energy evaluation module and a probability transition module; firstly, the energy evaluation module evaluates all negative samples. Relative to anchor sample energy value Then, the probability transition module calculates the probability that a negative sample will transition to a positive sample based on the energy value of the negative sample. And the transition process is simulated through random sampling; The process of constructing a new sample set through the transition of negative samples in each dimension is as follows: the initial sample set and the new sample set remain unchanged in the anchor sample part, and the transition samples that successfully transition are... Add positive sample sequences In the process, the positive sample set sequence is obtained. The new negative sample set sequence This includes the remaining negative samples excluding those that successfully transitioned; the new sample set contains the anchor sample sequence. Positive sample set sequence and negative sample set sequence ; Step 2.4: The new encoder is trained based on the new sample set and the encoding results are given. The new encoder is also a neural network encoder. With network projector The new sample set is composed of samples processed by a neural network. Obtain the hidden vector The encoding result is then calculated by the projector. ; The loss function of the new encoder incorporates transition confidence weights. As a multiplier factor for the transition sample: （17） Among them, the transition confidence weight The numerical value represents the transition probability of the transition sample. Operators for calculating similarity; Minimize the loss value using an optimization algorithm. Once the value falls below a set threshold, training of the new encoder is complete. After the new encoder is trained, only its trained neural network encoder is saved as a contrastive learning model. and its output latent vector sequence Set as a contrastive learning model The model results; For single-dimensional energy data, its characteristics The same method was used to obtain the contrastive learning model. and output the latent vector. ; Step 3: Embed the physical information neural network of the production plan: PINN model constructed using the PINN physical information neural network. The contrastive learning model in step 2 With PINN model Predictive models constituting industrial energy; For unidimensional energy data, its enhanced features from step 1 are... Input contrastive learning model Output latent vectors , will the latent vector Production planning data The splicing result is used as the PINN model. The input, including production planning data Indicates by Remaining time after the current energy state With remaining production batches , Changes in output energy data As the output of the prediction model, see equation (20). Combination The future features are calculated using the transfer equation, i.e., equation (21). After completing feature transfer, iterative prediction is performed by alternating between equations (20) and (21) to obtain the predicted sequence. Equation (22) is the constraint formula of the production plan on future energy fluctuations, where the first equation represents the initial production plan and the second equation is the dynamic form of the production plan in the iterative prediction process. （20） （21） （22） in, For characteristic transfer function, For the planning constraint function, For the dynamic changes of planning constraints, For PINN model parameters; Define average characteristics With average output as follows: Definition 1: If all energy characteristics Replaced with If the production plan is still met, then It is considered an average characteristic; Definition 2: If Predicted features for the next iteration cycle ,but It is considered to be the average output; Therefore, if within the remaining planned time All model outputs , Then the iterative prediction result See equation (23), at this time the prediction result sequence The constraints are satisfied, as shown in equation (24); （23） （24） Let constraint functions be defined. ,but This reflects the deviation between the model output and the production plan; variables are selected to embed production plan information. , For the number of selected variables, the equation Taking the derivative, we obtain the partial differential equation: （25） Embedding equation (25) into the loss function of the PINN model middle: （26） （27） （28） in, For physical driving loss, For data-driven loss, As a balance factor, This indicates the actual amount of waveform change; Feature sequences of multidimensional energy data Input it into the prediction model, and then... The hidden vector sequence is then obtained. The concatenation result of each latent vector and the corresponding dimension of the production plan data is used as the PINN model. The input is a sequence of changes in multidimensional energy. ;Will and By combining these features, the next fluctuation state is obtained through feature transfer. The long-term prediction results of multidimensional energy are obtained through multiple iterations. For the production planning embedding problem of multidimensional energy data, the loss value of each dimension is calculated according to equations (26)-(28), and the total loss is calculated according to equation (29). ; （29） in For the calculated first Loss in each dimension; Minimize loss value Or total loss value The PINN model training stops when the value falls below a set threshold, resulting in the final prediction model used to obtain the final prediction results, which meet the requirements of the production plan.

3. The industrial production-energy coupling prediction method according to claim 2, characterized in that, In step 1.1, in order to ensure that each basis function in the rhythm matching process can independently target a curve segment within a rhythm window, the stretching of the parent function does not exceed twice the average rhythm window duration.

4. The industrial production-energy coupled prediction method according to claim 2, characterized in that, In step 1.1, the coordinates of the extreme points in the 3D image are searched using equations (4)-(6): （30） （31） （32） in, For different time-domain translation coefficients Frequency domain coordinates of the extreme points of the matching value. For projection curves, The radius of the neighborhood of the extreme point. For Center point The frequency domain coordinates of the extreme points in the neighborhood of the radius. .

5. The industrial production-energy coupling prediction method according to claim 2, characterized in that, In step 1.1, equations (7) and (8) are used to... Distribution characteristics of curve segments converted to zero numerical values : （33） （34）。 6. The industrial production-energy coupling prediction method according to claim 2, characterized in that, In step 1.2, from The process of searching for the coordinates of the maximum point is shown in equation (12): （35）。 7. The industrial production-energy coupling prediction method according to claim 2, characterized in that, In step 1.2, because The search space in the vertical translation dimension is too large, so the translation amount is calculated directly. This affects computational efficiency; therefore, we should first maintain... Unchanged, targeting and Perform matching and obtain Assuming that the curve segment within the rhythm window is equal to the energy accumulation represented by the waveform matching basis function, then construct equation (13) using... Calculate energy impact ; （36） in, Indicates length is And all elements are a constant vector, The longitudinal translation is 0 and the longitudinal stretch is And the waveform parameters are The waveform matching basis function vector.

8. The industrial production-energy coupling prediction method according to claim 2, characterized in that, In step 2.3, the energy value and the probability of a negative sample transitioning to a positive sample are mentioned. The calculation formula is as follows: （37） （38） in, An operator for calculating the cosine similarity between two encoded vectors. Energy threshold This is a hyperparameter, and can be any positive number.

9. The industrial production-energy coupled prediction method according to claim 2, characterized in that, In step 2.4, to accelerate the data analysis process, at fixed iteration intervals, the parameters of the Newencoder corresponding to the neural network encoder and network projector are softly updated to the Old encoder: （39） in, The parameters representing the Old encoder, The parameters representing the new encoder, This represents the soft update factor.

10. The industrial production-energy coupled prediction method according to claim 2, characterized in that, In step 2.4, the two vectors are calculated. Similarity operators The calculation formula is: （40）。

Images