An industrial concentration parameter prediction method based on a PFGCNet model
By employing the frequency domain supervision strategy of the PFGCNet model and the large language model, the problem of insufficient prediction accuracy of traditional methods in industrial concentration systems is solved, achieving high-precision temperature prediction, especially the effective capture of thermodynamic coupling characteristics between adjacent devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGNAN UNIV
- Filing Date
- 2026-02-27
- Publication Date
- 2026-06-16
AI Technical Summary
In existing industrial concentration processes, traditional prediction methods struggle to capture complex nonlinear relationships and long-term time-series dependencies, and their prediction accuracy is limited, failing to meet the requirements for high-precision control. In particular, they fail to effectively guide models to focus on local thermodynamic coupling characteristics in temperature prediction of industrial concentration systems.
An industrial condensed parameter prediction method based on the PFGCNet model is adopted. The thermodynamic coupling characteristics between adjacent devices are captured by fuzzy gated convolutional flow, and the loss function of the prediction model is optimized by combining the frequency domain supervision strategy of large language model, thereby improving the temperature prediction accuracy.
It significantly improves the accuracy of temperature prediction in industrial concentration systems, accurately captures details of high-frequency disturbances, and provides a basis for refined control of the industrial concentration process.
Smart Images
Figure CN122220741A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an industrial condensed parameter prediction method based on the PFGCNet model, belonging to the fields of industrial process control and artificial intelligence technology. Background Technology
[0002] Industrial concentration is a crucial step in modern production. It involves using physical or chemical methods to concentrate certain components of a substance, increasing their concentration or extracting valuable components. The stability and precise control of its process parameters directly impact product quality and production efficiency. The industrial concentration process generates a large number of raw parameters (time-series parameters, equipment status parameters, process quality standards). Cleaning, polymerizing, compressing, or extracting features from these raw parameters to form high-value, low-redundancy, and rapidly analyzable parameters are known as industrial concentration parameters. Accurate prediction of specific industrial concentration parameters can improve industrial production efficiency, safety, and sustainability, which is of great significance in industries such as pharmaceuticals, food processing, papermaking, and chemicals.
[0003] Traditional methods for monitoring and predicting process parameters mainly fall into two categories: prediction based on empirical formulas and prediction based on simple statistical models. The former predicts by summarizing existing experimental data and establishing mathematical relationships between parameters; the latter infers future changes by fitting historical parameter trends. However, both methods have significant limitations: First, prediction accuracy is limited—empirical formulas are based on simplified assumptions and finite parameters, making it difficult to reflect real operating conditions; statistical models are highly sensitive to data quality, and noise, missing values, or outliers can significantly affect prediction results. Second, applicability is insufficient—when the system's operating mode changes or exceeds the original parameter range, both methods struggle to capture complex nonlinear relationships and long-term time-series dependencies. Therefore, traditional methods often fail to meet the high-precision requirements of modern production for condition monitoring and refined management.
[0004] Existing time series forecasting models, such as traditional machine learning methods, learn relationships between variables from historical parameters through feature engineering and simulation training to predict the future, but their ability to handle complex multivariate inputs is limited. Self-attention models, such as Transformer, capture long-distance dependencies in sequence parameters through attention mechanisms, thus performing well in time series modeling. However, industrial data often contains ambiguous transitional states, and variables exhibit strong local coupling based on physical connections (such as pipes and heat conduction). Global attention mechanisms tend to overlook the subtle interactions between these adjacent variables. Furthermore, most existing forecasting models use mean squared error (MSE) as the loss function, tending to output the average value, resulting in overly smoothed forecast curves that lose high-frequency fluctuation details reflecting the system's dynamic characteristics, failing to meet the demands of high-precision control.
[0005] Furthermore, in the field of time series forecasting, existing research mostly focuses on general time series models, and there are no temperature prediction models specifically for industrial concentration systems. At the level of specific model methods, the DFCNN (Differential Convolutional Fuzzy Time Series Forecasting) model converts time series into fuzzy information and directly feeds it into a convolutional neural network; the TimesNet model converts 1D time series into 2D representations based on multiple periods and then uses a convolutional neural network; FANTF (Fuzzy Attention Network Transformer) embeds fuzzy logic into the attention matrix calculation; and FreqLLM converts time series into frequency domain information and then processes it through a large language model. While these methods have some effectiveness on general time series, they all fail to effectively guide the model to pay attention to the local thermodynamic coupling characteristics in the actual process flow, thus preventing further effective improvement in prediction accuracy. Summary of the Invention
[0006] To further improve the temperature prediction accuracy of industrial concentration systems, this invention provides an industrial concentration parameter prediction method based on the PFGCNet (Perceptual Fuzzy-Gated Convolutional Network) model. Through a unique fuzzy gated convolutional flow, it adaptively guides the convolutional kernel to focus on important local physical regions using fuzzy semantics, effectively capturing the thermodynamic coupling characteristics between adjacent devices during the concentration process, thereby significantly improving the temperature prediction accuracy of industrial concentration systems. Specifically, the industrial concentration parameter prediction method based on the PFGCNet model provided by this invention includes: Step 1: Collect historical multivariate parameters from the industrial concentration process and rearrange the variables according to the physical topology to construct an industrial concentration dataset; Step 2: Construct a prediction model based on PFGCNet; Step 3: Train the prediction model built in Step 2 using the industrial condensation dataset from Step 1. Step four: Use a frequency domain supervision strategy based on a large language model to constrain and optimize the model training process in step three; Step 5: Based on the prediction model obtained in Step 4, predict the temperature of the first effect, the temperature of the second effect, and the temperature of the third effect at future times.
[0007] Optionally, the prediction model based on PFGCNet constructed in step two includes a normalization layer, a working condition perception module, a three-branch parallel embedding module, a multi-view feature fusion module, a Transformer encoder module, an inverse normalization layer, and a large model frequency domain supervision module. The data input to the prediction model is first standardized by a normalization layer, then filtered by working condition perception, and transmitted to a three-branch parallel embedding module to extract temporal flow features, fuzzy flow features, and convolutional flow features respectively. The three features are concatenated and mapped in a multi-view feature fusion module, and then input to a Transformer encoder module for deep feature extraction of multivariate dependencies. The output of the encoder module is transmitted to an inverse normalization layer for inverse normalization. The data after inverse normalization is transmitted to a large model frequency domain supervision module to calculate the frequency domain supervision loss and update the total loss function of the prediction model. Finally, the model is optimized and trained so that the trained model can be used to predict the first-effect temperature, second-effect temperature, and third-effect temperature at future moments in the industrial concentration process.
[0008] The operating condition sensing module preprocesses data based on the status of key parameters, inputting only data that meets the conditions into the model for training. For the triple-effect concentrator, the key parameter is the steam pressure status during operation, and only industrial sensor data when the steam pressure is non-zero is selected as input into the model.
[0009] Optionally, the three-branch parallel embedding module is composed of a deep temporal stream branch, a fuzzy semantic stream branch, and a spatiotemporal convolutional stream branch arranged in parallel; The deep temporal stream branch is used to extract global dynamic trends and obtain deep features. Its expression is: , in, This represents the normalized input data. Represents timestamp characteristics. This represents the dimension reversal operation. Represents a multilayer perceptron; The fuzzy semantic flow branch uses a learnable Gaussian membership function to extract nonlinear semantics and obtain fuzzy features. Its expression is: , , in, and These are the learnable mean and standard deviation parameters, respectively. For the fuzzy membership tensor, T represents the transpose operation; The spatiotemporal convolutional flow branch utilizes a fuzzy gating mechanism to extract local physical coupling features, thereby obtaining convolutional flow features. Its expression is: , , in, Represents the Hadamard product. This represents a two-dimensional convolution operation, used to jointly extract local features across time and variable dimensions. This indicates a dimension reshaping operation. This indicates a linear projection operation.
[0010] Optionally, the multi-view feature fusion module is used to concatenate and fuse the outputs of the three branches, and its expression is: , , in, and These are learnable fusion layer parameters. This will be used as the input to the subsequent Transformer encoder; the Transformer encoder module receives... As input, an inverted attention mechanism is used, treating each variable as an independent token for computation; the output after encoder processing... It is transmitted to the denormalization layer.
[0011] Optionally, the specific implementation process of the large model frequency domain supervision module is as follows: firstly, the predicted sequence is... and the real sequence The frequency domain amplitude spectrum is extracted by dividing the data into blocks and performing a real-number fast Fourier transform. Then, a linear layer is used to transform the frequency domain amplitude spectrum. The projected features are then input into the embedding space of the pre-trained GPT-2 model, and the projected features are fed into the GPT-2 model with frozen parameters. Finally, the hidden state features of the intermediate layers of the GPT-2 model are extracted, and the mean squared error between the predicted sequence and the true sequence in the abstract semantic space is calculated as the frequency domain supervision loss function. .
[0012] Optionally, the loss function of the prediction model is a hybrid time-frequency sensing loss function. Including the temporal reconstruction loss function frequency domain supervised loss function Furthermore, it is dynamically adjusted at training time step t, and its expression is: ,in, The weights represent the frequency domain supervised loss function.
[0013] Optionally, the weights of the frequency domain supervision loss function Adjustments are made using a piecewise linear preheating strategy, expressed as follows: , in, and These represent the start and end time steps of the preheating process, respectively. This represents the target weight value.
[0014] Optionally, when the industrial concentration process is a triple-effect concentration process, the industrial concentration parameters include the inlet flow rate, steam pressure, first-effect pressure, second-effect pressure, third-effect pressure, large-circulation outlet temperature, outlet flow rate, outlet density, cumulative outlet, and the first-effect temperature, second-effect temperature, and third-effect temperature as prediction targets during the time period T before prediction.
[0015] Optionally, the variable rearrangement based on physical topology refers to forcibly rearranging physically adjacent or directly interactive sensor variables into adjacent positions in the dimension of the data matrix. The rearranged variables are in the following order: liquid flow rate, steam pressure, first-effect pressure, first-effect temperature, second-effect pressure, second-effect temperature, third-effect pressure, third-effect temperature, large-circulation outlet temperature, outlet flow rate, outlet density, and outlet accumulation.
[0016] The present invention also provides a method for predicting industrial concentration parameters of honeysuckle extract, which is based on the above method.
[0017] The beneficial effects of this invention are: By employing a unique fuzzy-gated convolutional flow, the method adaptively guides the convolutional kernel to focus on important local physical regions using fuzzy semantics, effectively capturing the thermodynamic coupling characteristics between adjacent devices during the concentration process. A pre-trained large model is introduced, utilizing its powerful general representation capabilities for supervised training in the frequency domain, solving the problem of "over-smoothing" in traditional model predictions and accurately reproducing the details of rapid temperature parameter fluctuations. Experiments demonstrate that this method accurately captures high-frequency perturbations in the prediction of first-, second-, and third-effect temperatures, and significantly reduces errors compared to existing mainstream models, providing a reliable basis for the refined control of industrial concentration processes. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart of an industrial concentration parameter prediction method based on the PFGCNet model provided in an embodiment of the present invention.
[0020] Figure 2 This is a schematic diagram of a triple-effect concentration system for honeysuckle extract.
[0021] Figure 3 This is a schematic diagram of the variable physical topology rearrangement in the triple-effect concentration process of honeysuckle extract in one embodiment of the present invention.
[0022] Figure 4 This is a schematic diagram of the overall architecture of a prediction model based on PFGCNet provided in one embodiment of the present invention.
[0023] Figure 5 This is a schematic diagram illustrating the calculation principle of the frequency domain supervision loss function for a large model provided in one embodiment of the present invention.
[0024] Figure 6 This is a schematic diagram comparing the prediction results of the first effect temperature of the triple-effect evaporator A by various models in Embodiment 2 of the present invention; Figure 7 This is a schematic diagram comparing the prediction results of the two-effect temperature of the triple-effect evaporator A by various models in Embodiment 2 of the present invention; Figure 8 This is a schematic diagram comparing the predicted triple-effect temperature of the triple-effect evaporator A by various models in Embodiment 2 of the present invention; Figure 9 This is a schematic diagram comparing the prediction results of the first effect temperature of the triple-effect evaporator B by various models in Embodiment 2 of the present invention; Figure 10 This is a schematic diagram comparing the prediction results of the two-effect temperature of the triple-effect evaporator B by various models in Embodiment 2 of the present invention; Figure 11 This is a schematic diagram comparing the predicted triple-effect temperature of the triple-effect evaporator B using various models in Embodiment 2 of the present invention. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0026] Example 1: This embodiment provides a method for predicting industrial condensation parameters based on the PFGCNet model (Perceptual Fuzzy-Gated Convolutional Network), such as... Figure 1 As shown, the method includes: Step 1: For a specific concentration process, collect historical multivariate parameters from the industrial concentration process, rearrange the variables according to the physical topology, and construct an industrial concentration dataset. The specific concentration process includes the commonly used triple-effect concentration process. This embodiment takes the triple-effect concentration process of honeysuckle extract as an example. Figure 2A complete schematic diagram of a triple-effect concentration system is shown. Its core principle is to utilize the inverse relationship between vacuum degree and boiling point to achieve the cascade utilization of steam by connecting three effects in series, ultimately obtaining honeysuckle extract concentrate. The industrial concentration parameters include the inlet flow rate, steam pressure, first-effect pressure, second-effect pressure, and third-effect pressure during the predicted time period T, the outlet temperature of the large circulation system, the outlet flow rate, the outlet density, the cumulative outlet, and the first-effect temperature, second-effect temperature, and third-effect temperature as the prediction targets.
[0027] The dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1:2.
[0028] To enhance the model's ability to perceive physical coupling, this embodiment innovatively introduces a physical topology variable rearrangement mechanism. The motivation behind this design lies in the fact that traditional deep learning models often treat the input sensor variables as independent channels or random arrangements, ignoring the objectively existing "spatial neighborhood" relationships in industrial concentration processes. However, in triple-effect concentration processes, such as... Figure 2 As shown, heat and materials are transferred unidirectionally along the physical link of "first effect → second effect → third effect", and there is a strong instantaneous coupling between temperature and pressure within the same effect.
[0029] Therefore, based on the actual process piping connections and thermodynamic conduction chains, this step forcibly rearranges physically adjacent or directly interacting sensor variables into adjacent positions along the data matrix dimension. The rearranged variable physical topology is as follows: Figure 3 As shown. The physical meaning of this rearrangement is to map the "locality of physical space" to the "locality of feature space", so that subsequent convolution operations (Conv2D) can use their local receptive fields to directly capture the material and energy transfer features between upstream and downstream devices, rather than just extracting statistical correlations.
[0030] The rearranged data is then normalized by subtracting the mean from each variable in each sample and dividing by the standard deviation. The expression is as follows: ,in, Represents input data, Represents the mean of the sample. The standard deviation is represented by the following expression: Where n is the number of data points in the training set. It is the i-th data point in the training set.
[0031] Step 2: Construct a prediction model based on PFGCNet; like Figure 4As shown, the constructed PFGCNet-based industrial condensed parameter prediction model includes a condition perception module, a three-branch parallel embedding module, a multi-view feature fusion module, a Transformer encoder module, an inverse normalization layer, and a large model frequency domain supervision module; detailed descriptions are as follows: 1. Operating Condition Sensing Module: In actual industrial production, sensors operate on a continuous data acquisition basis, inevitably resulting in a mixture of production and non-production states. Inputting all sensor data into a model for training would introduce a large amount of irrelevant noise, hindering the model's effective learning of industrial processes. To address this issue, this invention implements a "condition-aware filtering" strategy.
[0032] This module preprocesses data based on the state of key parameters, inputting only qualified data into the model for training. For the triple-effect concentrator, the key parameter is the steam pressure state during operation. Only industrial sensor data when the steam pressure is non-zero is selected and input into the model for training via a sliding window sampling method. Once a boundary is detected, the window immediately jumps to the next valid operating condition segment for sampling, avoiding the introduction of spurious dependencies between actual and non-production states in the model, thus promoting rapid model convergence and improved accuracy.
[0033] 2. Three-branch parallel embedded module; Traditional time series prediction models typically employ a single linear embedding layer to map input data into a high-dimensional space. While effective for standard datasets, this approach suffers from significant limitations when dealing with complex industrial data condensation processes. Industrial data exhibits multimodal physical characteristics, and a single perspective cannot fully characterize the system state. To address this, this invention designs a "three-branch parallel embedding architecture," whose core objective is to perform "feature decoupling and multi-perspective reconstruction" of industrial data. These three components are used for macro-trend following, data uncertainty and nonlinearity extraction, and local physical coupling analysis in complex environments. Working in parallel, they holographically encode the original data from the temporal, semantic, and spatial dimensions, complementing each other to provide a rich feature representation containing physical mechanisms for the subsequent Transformer encoder.
[0034] This module processes the normalized input data in parallel. It contains three branches: Deep temporal stream: Used to extract global dynamic trends. Input data is concatenated with embedded timestamps, then dimensionally inverted, and mapped to hidden dimensions using a multilayer perceptron (MLP). Its expression is: ,in, This means transforming the dimension from (B, L, C) to (B, C, L).
[0035] Fuzzy Semantic Flow: Recovering the nonlinear fuzzy semantics of data using learnable Gaussian membership functions.
[0036] First, calculate the fuzzy membership tensor. Its expression is: , in, and These are the learnable center and width parameters, respectively; then, fuzzy features are obtained through linear layer mapping. .
[0037] Fuzzy gated convolutional flow: This is the core of capturing local physical coupling. In industrial field data, due to sensor drift or micro-fluctuations in operating conditions, the system state is often in a "either / or" fuzzy zone. Traditional hard threshold activation functions (such as ReLU) are prone to causing the loss of effective weak signals or the false amplification of noise.
[0038] Therefore, this invention utilizes the calculated fuzzy membership tensor As a "soft-gated signal," it adaptively modulates the input data. Its physical meaning is similar to a dynamic filter: when the data point falls within the central region of the Gaussian function (i.e., high membership, representing typical operating conditions or strong features), the gate opens, allowing the feature to pass through completely; when the data point deviates from the center (i.e., low membership, representing noise or irrelevant disturbances), the gate automatically suppresses the signal. Its expression is: ,in, This represents the Hadamard product. This modulation operation based on the Hadamard product essentially uses semantic-level judgments (whether it conforms to a specific fuzzy concept) to guide the extraction of physical features, thereby accurately locking the local coupling patterns that truly contribute to temperature prediction in complex, noisy environments.
[0039] Then Expanded to a pseudo-image format, it utilizes a two-dimensional convolutional layer (Conv2D) to slide jointly along the time and variable axes, with the following expression: The convolutional output undergoes dimensionality reshaping and linear projection to obtain convolutional flow features. Its expression is ; This indicates a dimension reshaping operation. This indicates a linear projection operation.
[0040] 3. Multi-view feature fusion module; The output features of the three branches mentioned above are concatenated and fused to construct a unified variable embedding, the expression of which is: , , in, and These are learnable fusion layer parameters. It will be used as input for the subsequent Transformer encoder.
[0041] 4. Transformer encoder module; This module is the core backbone of the prediction model, used for deep interaction and integration of fused multi-view features. This invention adopts an inverted Transformer architecture (iTransformer), which, unlike traditional time-step attention, maps the complete time series of each variable to an independent semantic token.
[0042] Specifically, the feature tensor input to this module is ,in For batch size, The number of variables (corresponding to process parameters such as inlet flow rate, pressure of each effect, and temperature). This represents the dimension of the model's hidden layers. This module consists of... The encoder layers are stacked with the same layer structure. Each encoder layer contains two core sub-layers: Multi-head Self-Attention (MSA) sub-layer and Position-by-Point Feedforward Network (FFN) sub-layer.
[0043] (1) Multi-head self-attention sublayer (MSA); This sublayer is designed to capture the dynamic correlations between different process variables (e.g., how steam pressure fluctuations are transmitted and affect the first-effect temperature).
[0044] First, input features Through three learnable linear projection matrices Mapped as a matrix of queries, keys, and values: , To extract features from different subspaces, a multi-head mechanism is employed. It is split into h heads along the channel dimension. For the i-th head ( ), calculate Scaled Dot-ProductAttention: , in, For each head dimension, This is a scaling factor used to prevent the gradient from vanishing due to excessively large dot product values. In this calculation, the attention weight matrix... A dependency graph among N variables was explicitly modeled.
[0045] Finally, the outputs of all the heads are concatenated and passed through a linear layer. Fusion: , This sublayer uses residual connections and layer normalization for processing: .
[0046] (2) Location-based feedforward network sublayer (FFN); This sublayer is used to perform nonlinear transformation and feature extraction on the feature representation of each variable. FFN consists of two linear transformation layers and a nonlinear activation function. In this embodiment, GELU (Gaussian Error Linear Unit) is chosen as the activation function to provide smoother gradient propagation than ReLU.
[0047] Its calculation expression is: ; in, , , Usually set to To expand the feature space.
[0048] Similarly, this sublayer is followed by residual connections and layer normalization: .
[0049] After L layers of stacking, the encoder outputs the final high-dimensional feature representation. This feature deeply integrates temporal trends, fuzzy semantics, local physical coupling, and global multivariate dependencies, providing a complete information foundation for subsequent accurate prediction of the first, second, and third effect temperatures.
[0050] 5. Anti-normalization layer Features of encoder output Project the prediction onto the predicted length F to obtain the normalized prediction result. Finally, the mean and standard deviation of the original data are used for inverse standardization to obtain the final predicted temperatures for the first, second, and third effects. .
[0051] Step 3: Train the prediction model built in Step 2 using the industrial condensation dataset from Step 1. The internal parameters and weights of the prediction model are optimized by comparing the training results with the true parameters, and an early stopping method is used to avoid overfitting. The training of the prediction model ends when the validation set loss of the prediction model no longer decreases significantly as the training process continues.
[0052] Step four: Use a large model frequency domain supervision strategy to constrain and optimize the model training process in step three; Specifically, this is achieved by constructing a hybrid time-frequency sensing loss function. To guide model training.
[0053] The loss function is derived from the temporal reconstruction loss. and frequency domain supervision loss Composition, its expression is: , in, To ensure training stability, a piecewise linear warmup strategy is employed for the dynamic weights that vary with the number of training steps t. The weights are initially zero and then linearly increased to the target value. This loss function not only constrains numerical accuracy but also forces the model to align with the energy distribution of the real data in the frequency domain. This compels the prediction results to retain the "high-frequency physical details" crucial for industrial control (such as subtle oscillations before equipment malfunctions), preventing the prediction curve from degenerating into a smooth mean line.
[0054] in The MSE loss function is defined by the following formula:
[0055] Where N represents the number of samples. Represents the true value of industrial parameters. This represents the predicted value output by the model.
[0056] The core motivation for introducing the large model frequency domain supervision strategy here is to solve the "oversmoothing" problem commonly found in regression tasks. The traditional mean squared error (MSE) loss function tends to fit the average trend of the data, thus mathematically "smoothing out" the high-frequency components that represent the dynamic characteristics of the system (such as rapid temperature fluctuations or instantaneous pressure shocks).
[0057] Large model frequency domain supervision loss The specific calculation process is as follows: 1. For the predicted sequence and the real sequence The process involves patching the data and performing a real-number Fast Fourier Transform (RFFT) on each patch to extract the frequency domain amplitude spectrum. : ; 2. Using linear layers to transform the amplitude spectrum Projected into the input embedding space of the pre-trained large language model (GPT-2); 3. Input the projected features into the GPT-2 model with frozen parameters, and extract the abstract feature representations of its intermediate hidden layers (such as layers 3, 7, and 11). ; 4. Calculate the mean squared error between the predicted sequence and the true sequence in the LLM abstract feature space: .
[0058] This step leverages the powerful general representation capabilities of large models to force the prediction model to learn the details of high-frequency physical fluctuations in temperature parameters, thus avoiding overly smoothed prediction results.
[0059] The ingenious design here leverages the general pattern recognition capabilities of the pre-trained large language model (GPT-2). Although GPT-2 is trained for text, its feature space learned from massive amounts of data contains general "sequence textures" and "frequency distribution patterns." By transforming the time series to the frequency domain and projecting it onto the semantic space of GPT-2, the question of "whether the predicted curve is realistic" is essentially transformed into a judgment of "whether the spectral characteristics of the predicted curve conform to natural laws."
[0060] Step 5: Based on the prediction model obtained in Step 4, predict the temperature of the first effect, the temperature of the second effect, and the temperature of the third effect at future times; The relevant parameters of the time period to be predicted are input into the trained and optimized PFGCNet-based prediction model, and the model outputs the predicted values of the first-effect temperature, second-effect temperature and third-effect temperature at future times.
[0061] Example 2 This embodiment provides a method for predicting industrial concentration parameters. The method is based on the prediction model described in Embodiment 1. The method includes: inputting the parameters to be predicted into the prediction model based on PFGCNet to predict the industrial concentration parameters.
[0062] To verify the performance of the PFGCNet-based prediction model proposed in this invention, this invention compares its prediction performance with that of the iTransformer prediction model, PatchTST prediction model, TimeXer prediction model, DLinear prediction model, and FreqLLM prediction model at 5 minutes, 10 minutes, and 15 minutes in the enrichment process. For iTransformer, please refer to Liu Y, Hu T, Zhang H, et al. itransformer: Inverted transformers are effective for time series forecasting[J]. arXivpreprint arXiv:2310.06625, 2023. The PatchTST prediction model can be found in Nie Y. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers[J]. arXiv preprint arXiv:2211.14730, 2022. For the TimeXer prediction model, please refer to Wang Y, Wu H, Dong J, et al. Timexer: Empoweringtransformers for time series forecasting with exogenous variables[J]. Advances in Neural Information Processing Systems, 2024, 37: 469-498. For the DLinear prediction model, please refer to Zeng A, Chen M, Zhang L, et al. Are transformers effective for time series forecasting?[C] / / Proceedings of the AAAI conference on artificial intelligence. 2023, 37(9): 11121-11128. For details on the FreqLLM forecasting model, please refer to Wang S, Gao M, Wang Z, et al. FreqLLM: frequency-aware large language models for time series forecasting[C] / / Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence. 2025: 3389-3397. Mean squared error (MSE), mean absolute error (MAE), and weighted average absolute percentage error (WMAPE) are used as evaluation criteria; Mean Squared Error (MSE): This measures prediction error by calculating the average of the squared differences between the predicted and actual values. It is more sensitive to larger errors. Its expression is: .
[0063] Mean Absolute Error (MAE): This measure of prediction accuracy is calculated by averaging the absolute differences between the predicted and actual values. It provides a direct reflection of the average error level, and its expression is as follows: .
[0064] Weighted Average Absolute Percentage Error (WMAPE): This measure of relative error is calculated by dividing the sum of absolute errors by the sum of true values. It is more robust when dealing with industrial data containing zero values or small values. Its expression is: .
[0065] in, Represents the true value. This represents the predicted value, and N represents the number of samples.
[0066] Table 1 summarizes the evaluation index values of the PFGCNet-based prediction model constructed using the method of this application, as well as five other models, on triple-effect evaporator A. Table 2 summarizes the evaluation index values of the PFGCNet-based prediction model constructed using the method of this application, as well as five other models, on triple-effect evaporator B. Triple-effect evaporators A and B are two independent honeysuckle concentration section devices. Table 1: Summary of Results for Triple-Effect Evaporator A
[0067] Table 2: Summary of Results for Triple-Effect Evaporator B
[0068] As shown in Tables 1 and 2, the PFGCNet index constructed in this application is optimal under 5-minute, 10-minute, and 15-minute prediction lead times. This indicates that its average deviation from the actual value is the smallest, its accuracy is the highest, and its explanatory power for data variation is the strongest. This proves that the PFGCNet-based prediction model proposed in this invention has a more accurate prediction effect compared to other prediction models.
[0069] from Figures 6 to 11 As shown in the six comparison charts of predicted and actual values, the prediction method of this application exhibits excellent prediction performance under different devices and batches. This is mainly due to the fuzzy gated convolution strategy used, which focuses on the thermodynamic conduction chain. Whether it is a stationary sequence or a sequence with high-frequency perturbations, the model can accurately track the complex temperature change trend of the three-effect concentrator system. Moreover, the model effectively avoids the mean regression phenomenon generated by traditional prediction models in the high-frequency perturbation part, and accurately tracks the fluctuating trend of temperature changes. This is thanks to the frequency domain supervised loss module based on the GPT-2 model.
[0070] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.
[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting industrial enrichment parameters based on the PFGCNet model, characterized in that, The method includes: Step 1: Collect historical multivariate parameters from the industrial concentration process and rearrange the variables according to the physical topology to construct an industrial concentration dataset; Step 2: Construct a prediction model based on PFGCNet; Step 3: Train the prediction model built in Step 2 using the industrial condensation dataset from Step 1. Step four: Use a frequency domain supervision strategy based on a large language model to constrain and optimize the model training process in step three; Step 5: Based on the prediction model obtained in Step 4, predict the temperature of the first effect, the temperature of the second effect, and the temperature of the third effect at future times.
2. The method according to claim 1, characterized in that, The prediction model based on PFGCNet constructed in step two includes a normalization layer, a working condition perception module, a three-branch parallel embedding module, a multi-view feature fusion module, a Transformer encoder module, an inverse normalization layer, and a large model frequency domain supervision module. The data input to the prediction model is first standardized by a normalization layer, then filtered by working condition perception, and then transmitted to a three-branch parallel embedding module to extract temporal flow features, fuzzy flow features, and convolutional flow features respectively. The three features are concatenated and mapped in a multi-view feature fusion module, and then input to a Transformer encoder module for deep feature extraction of multivariate dependencies. The output of the encoder module is transmitted to an inverse normalization layer for inverse normalization. The data after inverse normalization is transmitted to a large model frequency domain supervision module to calculate the frequency domain supervision loss and update the total loss function of the prediction model. Finally, the model is optimized and trained so that the trained model can be used to predict the first-effect temperature, second-effect temperature, and third-effect temperature at future moments in the industrial concentration process.
3. The method according to claim 2, characterized in that, The three-branch parallel embedding module consists of a deep temporal stream branch, a fuzzy semantic stream branch, and a spatiotemporal convolutional stream branch arranged in parallel. The deep temporal stream branch is used to extract global dynamic trends and obtain deep features. Its expression is: ， in, This represents the normalized input data. Represents timestamp characteristics. This represents the dimension reversal operation. Represents a multilayer perceptron; The fuzzy semantic flow branch uses a learnable Gaussian membership function to extract nonlinear semantics and obtain fuzzy features. Its expression is: ， ， in, and These are the learnable mean and standard deviation parameters, respectively. For the fuzzy membership tensor, T represents the transpose operation; The spatiotemporal convolutional flow branch utilizes a fuzzy gating mechanism to extract local physical coupling features, thereby obtaining convolutional flow features. Its expression is: ， ， in, Represents the Hadamard product. This represents a two-dimensional convolution operation, used to jointly extract local features across time and variable dimensions. This indicates a dimension reshaping operation. This indicates a linear projection operation.
4. The method according to claim 3, characterized in that, The multi-view feature fusion module is used to concatenate and fuse the outputs of the three branches, and its expression is: ， ， in, and These are learnable fusion layer parameters. This will be used as the input to the subsequent Transformer encoder; the Transformer encoder module receives... As input, an inverted attention mechanism is used, treating each variable as an independent token for computation; the output after encoder processing... It is transmitted to the denormalization layer.
5. The method according to claim 4, characterized in that, The specific implementation process of the large model frequency domain supervision module is as follows: First, the predicted sequence is... and the real sequence The frequency domain amplitude spectrum is extracted by dividing the data into blocks and performing a real-number fast Fourier transform. Then, a linear layer is used to transform the frequency domain amplitude spectrum. The projected features are then input into the embedding space of the pre-trained GPT-2 model, and the projected features are fed into the GPT-2 model with frozen parameters. Finally, the hidden state features of the intermediate layers of the GPT-2 model are extracted, and the mean squared error between the predicted sequence and the true sequence in the abstract semantic space is calculated as the frequency domain supervision loss function. .
6. The method according to claim 5, characterized in that, The loss function of the prediction model is a hybrid time-frequency sensing loss function. Including the temporal reconstruction loss function frequency domain supervised loss function Furthermore, it is dynamically adjusted at training time step t, and its expression is: ,in, The weights represent the frequency domain supervised loss function.
7. The method according to claim 6, characterized in that, The weights of the frequency domain supervision loss function Adjustments are made using a piecewise linear preheating strategy, expressed as follows: ， in, and These represent the start and end time steps of the preheating process, respectively. This represents the target weight value.
8. The method according to claim 7, characterized in that, When the industrial concentration process is a triple-effect concentration process, the industrial concentration parameters include the inlet flow rate, steam pressure, first-effect pressure, second-effect pressure, and third-effect pressure, large-circulation outlet temperature, outlet flow rate, outlet density, cumulative outlet, and the first-effect temperature, second-effect temperature, and third-effect temperature as prediction targets during the time period T before prediction.
9. The method according to claim 8, characterized in that, The variable rearrangement based on physical topology refers to forcibly rearranging physically adjacent or directly interactive sensor variables into adjacent positions in the dimension of the data matrix. The rearranged variables are in the following order: liquid inlet flow rate, steam pressure, first-effect pressure, first-effect temperature, second-effect pressure, second-effect temperature, third-effect pressure, third-effect temperature, large-circulation outlet temperature, outlet flow rate, outlet density, and outlet accumulation.
10. A method for predicting industrial concentration parameters of honeysuckle extract, characterized in that, The method is implemented based on the method described in any one of claims 1-9.