Method and apparatus for determining yield prediction value, and electronic device

Abstract

This application discloses a method, apparatus, and electronic device for determining yield prediction values. The method includes: acquiring historical production data from an industrial furnace; using a feedforward neural network to capture the non-periodic global variation patterns of the historical production data, obtaining trend features in the historical production data; segmenting the historical production data to obtain local block sequences; generating a gating vector using a gating mechanism, and determining weighted features based on the gating vector and the local block sequences; using a temporal convolutional network to capture the periodic fluctuation patterns of the weighted features at different time scales, obtaining periodic features in the historical production data; and determining the predicted yield value of the chemical raw material based on the trend features and periodic features. This application solves the technical problem of low prediction accuracy for the yield of chemical raw materials in related technologies.

Description

Technical Field

[0001] This application relates to the field of industrial artificial intelligence, and more specifically, to a method, apparatus, and electronic device for determining yield prediction values. Background Technology

[0002] Yield prediction for chemical raw materials is a crucial step in optimizing production efficiency and improving economic benefits. In continuous production processes such as petrochemicals and chemicals, key performance indicators such as raw material conversion rate and product yield directly impact a company's production costs and market competitiveness. However, current yield prediction technologies face several challenges in practical applications, resulting in generally low prediction accuracy and difficulty in meeting the needs of precise control and real-time optimization:

[0003] 1. Nonlinear complexity: Chemical production processes involve complex chemical reactions and physical processes. There are nonlinear relationships between parameters such as raw material composition, reaction conditions, and equipment status. This makes it difficult for prediction models to accurately capture and express these relationships. In particular, under dynamic scenarios such as changes in raw material batches and equipment aging, the prediction error of traditional linear models will increase significantly.

[0004] 2. Time Delay and Multivariate Coupling: There is a large time delay between the feedstock entering the reactor and the final product being produced, and multiple operating parameters (such as temperature, pressure, and flow rate) are coupled with each other, jointly affecting the yield. This time delay and multivariate coupling makes real-time control and online optimization extremely difficult, and predictive models based on historical data cannot accurately reflect the immediate impact of current operating parameters on the yield.

[0005] 3. Mismatch between data characteristics and model: Time-series data for chemical production processes contain both trend changes and periodic fluctuations. Trend changes reflect long-term performance variations in equipment, while periodic fluctuations are caused by short-term factors such as production shifts, raw material batches, and environmental conditions. Existing prediction models cannot effectively distinguish and model these two types of changes, resulting in insufficient robustness and generalization ability of the prediction results.

[0006] 4. The contradiction between the generalization ability of the model and the real-time computing performance: high-precision prediction models have high computational complexity, making it difficult to achieve real-time prediction and online optimization; while lightweight models are fast, their prediction accuracy is limited.

[0007] 5. Limitations of Optimization Algorithms: Yield optimization requires not only accurate prediction models but also effective optimization algorithms. Traditional optimization methods, such as gradient descent, are prone to getting trapped in local optima when dealing with nonlinearity, multimodal functions, and complex constraints, making it difficult to find the global optimum.

[0008] In summary, the main limitations of chemical feedstock yield prediction are model complexity, understanding of data characteristics, real-time computing performance, and the effectiveness of optimization algorithms.

[0009] There is currently no effective solution to the above problems. Summary of the Invention

[0010] This application provides a method, apparatus, and electronic device for determining yield prediction values, in order to at least solve the technical problem of low accuracy in predicting the yield of chemical raw materials in related technologies.

[0011] According to one aspect of this application, a method for determining a yield prediction value is provided, comprising: acquiring historical production data of an industrial furnace used for producing chemical raw materials; capturing aperiodic global change patterns in the historical production data using a feedforward neural network to obtain trend features in the historical production data; segmenting the historical production data to obtain a local block sequence, wherein the local block sequence is a two-dimensional tensor comprising multiple local blocks; generating a gating vector using a gating mechanism, and determining weighted features based on the gating vector and the local block sequence, wherein the gating vector represents the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected to participate in subsequent calculations; capturing periodic fluctuation patterns of the weighted features at different time scales using a temporal convolutional network to obtain periodic features in the historical production data; and determining the yield prediction value of the chemical raw materials based on the trend features and the periodic features.

[0012] Optionally, the feedforward neural network is trained using the following method: Based on the multivariate coupling characteristics of chemical raw materials in the pyrolysis process, key operational variables are extracted from the training dataset, wherein the key operational variables include at least one of the following: pyrolysis depth and feed composition; time series alignment and standardization preprocessing are performed on the key operational variables to obtain standardized operational variable time series; an initial feedforward neural network is obtained, wherein the input layer dimension of the initial feedforward neural network is consistent with the variable dimension of the standardized operational variable time series; the initial feedforward neural network includes: multiple hidden layers using nonlinear activation functions, the nonlinear activation functions of the hidden layers being used to map the nonlinear characteristics of the chemical raw material pyrolysis process; the standardized operational variable time series is used as the input of the initial feedforward neural network, and historical yield data of the chemical raw material is used as the training target, the connection weights and bias parameters of the initial feedforward neural network are optimized through the backpropagation algorithm to train the initial feedforward neural network until a preset stopping condition is met, thus obtaining the feedforward neural network.

[0013] Optionally, the chemical feedstock includes at least ethylene. The method for determining the yield prediction value further includes: mapping the PONA value of the feed data in the time series of standardized operating variables to the operating variables to obtain the olefin formation potential index, wherein the operating variables are used to characterize the cracking depth, and the operating variables include at least the cracking furnace outlet temperature or hydrocarbon partial pressure; during the initial training process of the feedforward neural network, dynamic weights are assigned to the training samples based on the olefin formation potential index, wherein the olefin formation potential index is positively correlated with the weight coefficient assigned to the training samples.

[0014] Optionally, after obtaining the olefin generation potential index, the method for determining the yield prediction value further includes: generating an adaptive regularization term based on the olefin generation potential index, wherein the adaptive regularization term includes an error penalty term and an overfitting suppression term, the weight coefficient of the error penalty term is an increasing function of the olefin generation potential index, and the weight coefficient of the overfitting suppression term is a decreasing function of the olefin generation potential index; and adding the adaptive regularization term to the loss function of the backpropagation algorithm.

[0015] Optionally, a temporal convolutional network is used to capture the periodic fluctuation patterns of weighted features at different time scales to obtain periodic features in historical production data. This includes: standardizing the weighted features to obtain target weighted features; capturing multi-scale periodic features of the target weighted features using a temporal convolutional network, wherein the multi-scale periodic features include short-term fluctuation patterns and long-term periodic patterns, and the temporal convolutional network includes multiple dilated causal convolutional layers with different dilation factors; performing nonlinear transformation and integration on the multi-scale periodic features to obtain a high-level periodic feature representation; and mapping the high-level periodic feature representation to periodic features using a fully connected layer.

[0016] Optionally, based on trend characteristics and periodic characteristics, the predicted yield of chemical raw materials is determined, including: fusing the trend characteristics and periodic characteristics to obtain fused characteristics; performing layer normalization on the fused characteristics to obtain standardized fused characteristics; and using a fully connected output layer to map the standardized fused characteristics to the predicted yield of chemical raw materials within a preset prediction time range.

[0017] Optionally, the trend features and periodic features are fused to obtain fused features, including: obtaining the cumulative running time of the industrial furnace from a historical time to a preset time, wherein the historical time is the time corresponding to the coking and maintenance event closest to the preset time; calculating the trend term fusion weight based on the cumulative running time using a predefined weight function, wherein the trend term fusion weight is an increasing function of the cumulative running time; weighting the trend features using the trend term fusion weight, weighting the periodic features using the periodic term fusion weight, and adding the weighted trend features and the weighted periodic features to obtain the fused features, wherein the periodic term fusion weight is the difference between 1 and the trend term fusion weight.

[0018] According to another aspect of this application, an apparatus for determining the predicted yield value is also provided, comprising: an acquisition module for acquiring historical production data of an industrial furnace used for producing chemical raw materials; a first capture module for capturing non-periodic global change patterns of the historical production data using a feedforward neural network to obtain trend features in the historical production data; a processing module for segmenting the historical production data to obtain a local block sequence, wherein the local block sequence is a two-dimensional tensor comprising multiple local blocks; generating a gating vector using a gating mechanism, and determining weighted features based on the gating vector and the local block sequence, wherein the gating vector represents the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected to participate in subsequent calculations; a second capture module for capturing periodic fluctuation patterns of the weighted features at different time scales using a temporal convolutional network to obtain periodic features in the historical production data; and a determination module for determining the predicted yield value of the chemical raw materials based on the trend features and periodic features.

[0019] According to another aspect of this application, a non-volatile storage medium is also provided, the storage medium including a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-mentioned method for determining the yield prediction value.

[0020] According to another aspect of this application, an electronic device is also provided, comprising: a memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the above-described method for determining the yield prediction value.

[0021] According to another aspect of this application, a computer program is also provided, wherein when the computer program is executed by a processor, it implements the method for determining the above-mentioned yield prediction value.

[0022] According to another aspect of this application, a computer program product is also provided, the computer program product including a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned method for determining the yield prediction value.

[0023] In this application, historical production data from an industrial furnace used for producing chemical raw materials is acquired. A feedforward neural network is used to capture the non-periodic global variation patterns of the historical production data, yielding trend features. The historical production data is segmented to obtain local block sequences, each a two-dimensional tensor comprising multiple local blocks. A gating mechanism is used to generate a gating vector, and weighted features are determined based on the gating vector and the local block sequences. The gating vector represents the probability or weight of each feature channel corresponding to a process variable in the local block sequence being selected for subsequent calculations. A temporal convolutional network is used to capture the periodic fluctuation patterns of the weighted features at different time scales, yielding periodic features in the historical production data. Based on the trend and periodic features, the predicted yield of the chemical raw materials is determined. By modeling the trend and periodic features in the time-series data separately, the accuracy of predicting the yield of chemical raw materials is improved, thereby achieving the technical effect of optimizing production efficiency and solving the technical problem of low prediction accuracy for the yield of chemical raw materials in related technologies. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0025] Figure 1 This is a flowchart of a method for determining a yield prediction value according to an embodiment of this application;

[0026] Figure 2 This is a flowchart of another method for determining the yield prediction value according to an embodiment of this application;

[0027] Figure 3 This is an architecture diagram of a yield prediction model according to an embodiment of this application;

[0028] Figure 4 This is a structural diagram of a yield prediction determination device according to an embodiment of this application;

[0029] Figure 5 This is a hardware structure block diagram of a computer terminal for a method of determining a yield prediction value according to an embodiment of this application. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] According to an embodiment of this application, a method embodiment for determining a yield prediction value is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0033] Figure 1 This is a flowchart of a method for determining a yield prediction value according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0034] Step S102: Obtain historical production data of the industrial furnace, which is used to produce chemical raw materials.

[0035] Industrial furnaces can be pyrolysis furnaces, or various types of reaction furnaces, conversion furnaces, or other thermodynamic processing equipment used in petrochemical and chemical production.

[0036] Historical production data is used to record the production of chemical feedstocks in industrial furnaces under different operating conditions, including but not limited to parameters such as cracking temperature, hydrocarbon-to-water ratio, feed flow rate, and yield of key products such as ethylene. Data acquisition can cover a relatively long time span to include operational examples under various conditions.

[0037] Step S104: Use a feedforward neural network to capture the non-periodic global change patterns of historical production data to obtain trend features in the historical production data.

[0038] Global change patterns are long-term, persistent, non-periodic trends in historical production data of industrial furnaces. These patterns not only span single time periods but also reflect the systematic evolution of equipment performance, raw material properties, or external environmental factors over time, thus profoundly impacting the yield of chemical feedstocks. Identifying global change patterns helps predictive models understand more macroscopic and persistent time-series characteristics, thereby improving the accuracy of long-term yield forecasts.

[0039] Trend features refer to a set of features extracted from global change patterns using feedforward neural networks that represent the long-term development trend of historical production data. Trend features can capture long-term change patterns related to yield, such as equipment aging and catalyst activity decline.

[0040] In step S104, a feedforward neural network of the type Multilayer Perceptron (MLP) can be used to perform in-depth analysis of historical production data of industrial furnaces (e.g., cracking furnaces). An MLP is a neural network composed of alternating layers of linear layers and nonlinear activation functions. MLPs are capable of learning nonlinear and complex global change patterns in the data. By using historical production data as input, MLPs can automatically extract the long-term trend features implicit in the data during training.

[0041] Step S106: Segment the historical production data to obtain a local block sequence, wherein the local block sequence is a two-dimensional tensor comprising multiple local blocks; generate a gating vector using a gating mechanism, and determine the weighted features based on the gating vector and the local block sequence, wherein the gating vector is used to represent the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected to participate in subsequent calculations.

[0042] In step S106, historical production data is first converted into a local block sequence using a block segmentation technique. Each local block in the local block sequence includes data within a preset time range, forming a two-dimensional data structure that is easy to analyze.

[0043] Next, a gating mechanism is used to generate gate vectors. These gate vectors are generated from the local block sequence through specific calculations. Their function is to dynamically select and adjust the participation level of the feature channels corresponding to each process variable. Each element of the gate vector corresponds to a process variable, with a value between 0 and 1, representing the probability or weight of that process variable being activated in subsequent calculations.

[0044] Specifically, the gating vector is multiplied element-wise with the feature channels in the local block sequence. Essentially, the elements of the gating vector are used as weights and multiplied with the corresponding feature values ​​in the local block sequence to determine the weighted features. In other words, when an element of the gating vector is close to 1, the corresponding process variable will be considered more in subsequent calculations; conversely, if its value is close to 0, the influence of the corresponding process variable on the model will be weakened. It should be noted that this dynamic weight allocation mechanism allows the model to adaptively identify and emphasize the process variables that have the greatest impact on the prediction target (such as ethylene yield), while reducing the interference of less important variables on model training, thereby improving the performance and efficiency of the prediction model.

[0045] In short, by using gating mechanisms and gating vectors, it is possible to automatically determine which local data blocks contain process variables that are more critical to the prediction target, and then weight the data accordingly to generate weighted features. These weighted features are then used for model training, enabling the model to not only learn information from all process variables, but also to focus more on those variables that truly affect the prediction indicators, thereby achieving higher accuracy and robustness in predicting chemical feedstock yields.

[0046] Step S108: Use a temporal convolutional network to capture the periodic fluctuation patterns of weighted features at different time scales to obtain periodic features in historical production data.

[0047] Periodic fluctuation patterns refer to recurring patterns of fluctuation in historical production data that are related to specific time periods (such as a certain time period of day, a certain number of days in a week, seasonal changes, etc.). Periodic fluctuation patterns are caused by short-term factors such as changes in production shifts, raw material batch updates, and periodic changes in environmental conditions, and have a predictable periodic impact on the yield of chemical raw materials.

[0048] Periodic features refer to the set of features extracted from periodic fluctuation patterns using a Temporal Convolutional Network (TCN) that reflect short-term periodic changes in historical production data. TCNs, utilizing their unique dilated causal convolutional structure, can effectively identify and learn periodic patterns at different time scales, automatically extracting periodic features from the data. These periodic features enhance the sensitivity and ability of prediction models to short-term fluctuations in chemical feedstock yields, ensuring that the model accurately reflects the impact of periodic factors on yields. This provides a predictive basis based on current periodic features for real-time adjustment of production parameters and optimization of operating conditions. Periodic features are a key component enabling prediction models to cope with short-term periodic changes and achieve high-precision, real-time predictions.

[0049] Step S110: Determine the predicted yield of chemical raw materials based on trend and cycle characteristics.

[0050] Among them, chemical raw materials can be single compounds, such as ethylene, propylene, butadiene, etc., or they can be mixtures, such as cracked feedstock oil, natural gas, etc.

[0051] The predicted yield of chemical raw materials is the expected result of the proportion of chemical raw material output to the total raw material input.

[0052] For example, step S110 can be achieved by first fusing trend features and periodic features. Feature fusion can be achieved by directly concatenating the two features to form a new joint feature vector; or by using weighted summation, which adds the two features according to pre-set weights or weights learned by the model to emphasize the importance of different features.

[0053] The fused feature vector is then processed through one or more fully connected layers. The role of the fully connected layers is to further abstract and transform the features, mapping the input features to a new feature space through weight matrices and bias terms to capture higher-order feature combinations and ultimately output a representation suitable for predicting yield. Each fully connected layer is typically equipped with a non-linear activation function, such as ReLU, tanh, or sigmoid, to introduce non-linear relationships and enhance the model's expressive power.

[0054] Finally, the output layer transforms the processed feature vectors into predicted yield values ​​for the chemical feedstock. The output layer can also be a fully connected layer, but the number of nodes in the output layer will equal the dimension of the target dimension (e.g., if predicting the yield for the next 5 minutes, the output layer will have 5 nodes). The output layer maps the model's internal representations to the target space using learned weights and biases, directly providing the predicted yield value.

[0055] The above steps involve acquiring historical production data from an industrial furnace used to produce chemical raw materials; using a feedforward neural network to capture the non-periodic global change patterns of the historical production data to obtain trend features; segmenting the historical production data to obtain local block sequences, where each local block sequence is a two-dimensional tensor comprising multiple local blocks; generating a gating vector using a gating mechanism, and determining weighted features based on the gating vector and the local block sequences, where the gating vector represents the probability or weight of each feature channel corresponding to each process variable in the local block sequence being selected for subsequent calculations; capturing the periodic fluctuation patterns of the weighted features at different time scales using a temporal convolutional network to obtain periodic features in the historical production data; and determining the predicted yield of chemical raw materials based on the trend and periodic features. By modeling the trend and periodic features in the time series data separately, the accuracy of predicting the yield of chemical raw materials is improved, thereby achieving the technical effect of optimizing production efficiency.

[0056] The following are Figure 1 The steps shown are illustrated and explained by way of example.

[0057] According to some optional embodiments of this application, the feedforward neural network is trained by the following method: based on the multivariate coupling characteristics of chemical raw materials in the pyrolysis process, key operational variables are extracted from the training dataset, wherein the key operational variables include at least one of the following: pyrolysis depth and feed composition; the key operational variables are preprocessed by time series alignment and standardization to obtain a standardized operational variable time series; an initial feedforward neural network is obtained, wherein the input layer dimension of the initial feedforward neural network is consistent with the variable dimension of the standardized operational variable time series; the initial feedforward neural network includes: multiple hidden layers using nonlinear activation functions, the nonlinear activation functions of the hidden layers being used to map the nonlinear characteristics of the chemical raw material pyrolysis process; the standardized operational variable time series is used as the input of the initial feedforward neural network, the historical yield data of the chemical raw material is used as the training target, and the connection weights and bias parameters of the initial feedforward neural network are optimized by the backpropagation algorithm to train the initial feedforward neural network until a preset stopping condition is met, thereby obtaining the feedforward neural network.

[0058] In this embodiment, key operational variables are first extracted from the training dataset. These key operational variables include, but are not limited to, pyrolysis depth, feed composition, COT temperature, hydrocarbon-to-water ratio, feed flow rate, pressure, and feed temperature. These key operational variables have a significant impact on the efficiency and yield of the pyrolysis process.

[0059] Then, time series alignment and standardization preprocessing are performed on the key operational variables to ensure consistency among variables and stability of model training, resulting in standardized operational variable time series.

[0060] Furthermore, an initial feedforward neural network is obtained, wherein the input layer dimension of the initial feedforward neural network matches the variable dimension of the standardized operational variable time series, ensuring that the network can receive and process these key variables. The network's internal structure includes multiple nonlinear hidden layers, which employ nonlinear activation functions such as ReLU and GELU to map the nonlinear characteristics of the chemical feedstock cracking process and capture the complex interactions between variables.

[0061] The standardized operational variable time series is used as input, while historical yield data of chemical raw materials is used as the training target, i.e., the expected value of the model output. The connection weights and bias parameters of the network are continuously optimized through the backpropagation algorithm. It can be understood that the above process is essentially an iterative process of minimizing a loss function, where the loss function measures the difference between the model's predicted value and the true historical yield value.

[0062] The training process continues until the network performance meets preset stopping conditions, such as reaching a predetermined minimum loss function value, loss function convergence, or a preset number of training epochs, thus ensuring that the model can make stable and effective predictions after sufficient training. The resulting feedforward neural network can not only efficiently handle key operational variables in the chemical feedstock cracking process, but also predict the cracking yield based on these variables, providing data-driven decision-making basis for real-time optimization of the cracking furnace, thereby improving production efficiency and economic benefits.

[0063] According to some alternative embodiments of this application, the chemical raw materials include at least: ethylene.

[0064] The yield prediction method may also include the following steps: mapping the PONA values ​​of the feed data in the time series of standardized operating variables to the operating variables to obtain the olefin formation potential index, wherein the operating variables are used to characterize the cracking depth, and the operating variables include at least: cracking furnace outlet temperature or hydrocarbon partial pressure; during the initial training process of the feedforward neural network, dynamic weights are assigned to the training samples based on the olefin formation potential index, wherein the olefin formation potential index is positively correlated with the weight coefficient assigned to the training samples.

[0065] In this embodiment, a dynamic weight allocation strategy is used to enhance the model's predictive performance. Specifically, the dynamic weight allocation strategy involves mapping the PONA values ​​(Paraffins, Olefins, Naphthenes, Aromatics values, i.e., the composition ratios of saturated hydrocarbons, olefins, cycloalkanes, and aromatics) of the feed data in the time series of the standardized operational variables to the operational variables to obtain the olefin formation potential index. The operational variables play a crucial role in the cracking process, characterizing the cracking depth, and include at least the outlet temperature or hydrocarbon partial pressure of the cracking furnace. Furthermore, the olefin formation potential index is used to assign dynamic weights to each training sample during the initial training of the feedforward neural network. It should be noted that during the training of the feedforward neural network, weights are dynamically assigned to each training sample based on the olefin formation potential index. The magnitude of the weight is positively correlated with the olefin formation potential reflected in the sample; that is, samples showing higher olefin formation potential under given cracking conditions will be given greater weights and thus receive more attention during training. The aforementioned dynamic weighting mechanism ensures that the model can pay more attention to and learn from changes in operating conditions and feedstock composition that significantly contribute to olefin production, thereby improving the model's accuracy and robustness in predicting high-yield conditions.

[0066] In the steps described above, by mapping the feed composition (PONA value) in the time series of standardized operating variables to operating variables (such as cracker outlet temperature or hydrocarbon partial pressure), the probability of heavy hydrocarbons being converted into ethylene under different cracking conditions can be quantitatively described. The olefin formation potential index can reveal the potential for olefin yield under specific operating conditions. Furthermore, by introducing the olefin formation potential index and a dynamic weighting mechanism, the model can better capture and utilize data points reflecting key characteristics of the cracking process, especially under complex and variable operating conditions, where samples that significantly impact feedstock conversion rates are more fully learned and simulated. This not only accelerates model training convergence but also improves the adaptability and prediction accuracy of the predictive model in the face of actual production and feedstock fluctuations, thereby enabling more effective production parameter control and optimization.

[0067] In some optional embodiments of this application, after obtaining the olefin generation potential index, the yield prediction method further includes the following steps: generating an adaptive regularization term based on the olefin generation potential index, wherein the adaptive regularization term includes an error penalty term and an overfitting suppression term, the weight coefficient of the error penalty term is an increasing function of the olefin generation potential index, and the weight coefficient of the overfitting suppression term is a decreasing function of the olefin generation potential index; and adding the adaptive regularization term to the loss function of the backpropagation algorithm.

[0068] In this embodiment, the weighting coefficient of the error penalty term is set as an increasing function of the olefin formation potential index. This means that for samples exhibiting high olefin formation potential under specific pyrolysis conditions, the penalty for model prediction errors will be stronger. Since high potential index conditions represent higher production efficiency and economic value, the model's prediction accuracy on these samples is particularly important. Strict error control can guide the model to focus more on learning and simulating these high-value conditions.

[0069] The weighting coefficients of the overfitting suppression term are set as a decreasing function of the olefin formation potential index. During training, the weighting coefficients for suppressing overfitting decrease as the potential index increases, and vice versa. This is because for samples with low olefin formation potential and high complexity, the model is prone to overfitting, i.e., overlearning the detailed features of these samples while neglecting the model's performance and generalization ability under simplified conditions. By increasing the weight of the overfitting suppression term, the model can be made more conservative on complex samples, avoiding overcomplication and thus maintaining robustness in predictions under different operating conditions.

[0070] Furthermore, this embodiment adds an adaptive regularization term as defined above to the loss function used in the BP algorithm, forming a composite loss function. This composite loss function considers not only the difference between the model's predicted results and the actual yield (i.e., the error penalty term), but also the balance between model complexity and generalization ability (i.e., the overfitting suppression term). This ensures that the network pursues both accurate predictions during training and avoids the risk of decreased model generalization performance due to overfitting the training data.

[0071] For example, the composite loss function can be expressed as:

[0072]

[0073] in, It is the original loss function, used to measure the model output. Compared with actual yield The error; and These represent the olefin formation potential index, respectively. Error penalty term and overfit suppression term, and These are the dynamic weighting coefficients of these two terms, which are associated with the olefin formation potential index according to increasing and decreasing functions, respectively, to achieve an adaptive adjustment effect.

[0074] It should be noted that by dynamically adjusting the weight of the adaptive regularization term in the loss function in each training iteration, the performance of the feedforward neural network in the task of predicting the yield of chemical feedstock cracking can be significantly improved.

[0075] As some optional embodiments of this application, capturing the periodic fluctuation patterns of weighted features at different time scales using a temporal convolutional network to obtain periodic features in historical production data can be achieved through the following steps: standardizing the weighted features to obtain target weighted features; capturing multi-scale periodic features of the target weighted features using a temporal convolutional network, wherein the multi-scale periodic features include short-term fluctuation patterns and long-term periodic patterns, and the temporal convolutional network includes multiple dilated causal convolutional layers with different dilation factors; performing nonlinear transformation and integration on the multi-scale periodic features to obtain a high-level periodic feature representation; and mapping the high-level periodic feature representation to periodic features using a fully connected layer.

[0076] In this embodiment, the weighted features are first standardized to ensure effective learning of the features within the network. The standardized target weighted features are then used as input to the TCN to capture its inherent periodic variation patterns.

[0077] Furthermore, dilated causal convolutional layers in temporal convolutional networks are utilized to capture multi-scale periodic features of the target weighted features. A TCN with dilated causal convolution is employed to efficiently capture long-term dependencies in periodic terms. Causality ensures that predictions do not use future information, and the dilation mechanism allows neurons to capture longer-term patterns with exponentially expanded receptive fields.

[0078] For a one-dimensional input X and a convolution kernel f: .

[0079] in: : Dilated convolution operation with dilation factor d; K: Size of convolution kernel.

[0080] TCN (Transient Causal Convolutional Network) efficiently processes information at different time scales through multiple dilated causal convolutional layers with varying dilation factors, thereby extracting short-term fluctuation patterns and long-term periodic patterns. The core advantage of dilated causal convolution lies in its ability to significantly expand the receptive field of each neuron without increasing network depth, thus capturing dependencies in longer sequences. Furthermore, through the internal structure of TCN, such as residual connections and batch normalization, these multi-scale periodic features are further nonlinearly transformed and integrated to obtain high-level periodic feature representations.

[0081] Batch normalization standardizes each feature channel, stabilizing the training process of deep networks and accelerating convergence. The specific formula is:

[0082]

[0083] in This represents the mean of the current small batch of data. This represents the variance of the current small batch of data.

[0084] Finally, a fully connected layer is used to map the high-level periodic feature representation to periodic features, completing the transformation from raw data to a periodic feature representation. The fully connected layer uses linear mapping and non-linear activation functions (such as ReLU and GELU) to convert the high-level representation of periodic features into a format suitable for direct use by the prediction model, thereby improving the model's sensitivity to periodic fluctuations and its prediction accuracy.

[0085] In some optional embodiments of this application, the predicted yield of chemical raw materials can be determined based on trend features and periodic features through the following steps: fusing trend features and periodic features to obtain fused features; performing layer normalization on the fused features to obtain standardized fused features; and using a fully connected output layer to map the standardized fused features to the predicted yield of chemical raw materials within a preset prediction time range.

[0086] In this embodiment, trend features and periodic features are fused to form a fused feature that comprehensively reflects the dynamic characteristics of the chemical feedstock cracking process. Layer normalization is then applied to the fused feature. Layer normalization involves calculating the mean and variance of each column of the feature matrix (corresponding to each feature channel) individually, followed by standardization to eliminate the dimensional influence between features and ensure the stability of the model input. The fully connected output layer transforms the fused feature into an output within the prediction domain—that is, the expected change in the chemical feedstock cracking yield over a future period—through a series of linear transformations and nonlinear activation functions. The fully connected output layer is the core component of the prediction model; it not only directly determines the accuracy of the prediction results but also flexibly adapts to different prediction time ranges, providing strong support for real-time optimization and decision-making in chemical production processes.

[0087] As some alternative embodiments of this application, fusing trend features and periodic features to obtain fused features can be achieved through the following steps: obtaining the cumulative running time of the industrial furnace from a historical time to a preset time, wherein the historical time is the time corresponding to the most recent coking and maintenance event before the preset time; calculating the trend term fusion weight based on the cumulative running time using a predefined weighting function, wherein the trend term fusion weight is an increasing function of the cumulative running time; weighting the trend features using the trend term fusion weight, weighting the periodic features using the periodic term fusion weight, and adding the weighted trend features and the weighted periodic features to obtain the fused features, wherein the periodic term fusion weight is the difference between 1 and the trend term fusion weight.

[0088] Among them, the decoking maintenance event refers to the mandatory process operation performed to remove carbon deposits inside the furnace tubes and convection section through oxidation reaction in order to restore the design thermal efficiency and fluid flow capacity of the pyrolysis furnace.

[0089] In this embodiment, the trend term fusion weight is calculated based on the cumulative operating time of the industrial furnace since the last coking maintenance event, and the trend term fusion weight is a monotonically increasing function of the cumulative operating time. This is because, over time, factors such as equipment wear and tear, and decreased catalyst activity gradually accumulate, increasing their trend-based impact on the production process. Therefore, a longer operating time implies a larger trend term weight, indicating that the model should pay more attention to the impact of equipment aging and long-term operation on production yield. The periodic term fusion weight is set to 1 minus the trend term fusion weight, ensuring that the sum of the two is always 1. This allows the model to consider the impact of equipment aging while also not ignoring periodic fluctuations caused by changes in operating cycles and raw material batches. The automatic adjustment mechanism of the periodic term fusion weight automatically balances the relative importance of long-term trends and short-term cycles in the model's predictions based on the cumulative operating time.

[0090] By weighting trend features with trend-term fusion weights and periodic features with periodic-term fusion weights, and then summing the two to obtain the fused features, the flexibility and accuracy of model predictions can be ensured. Because the impact of equipment aging and periodic operations on production varies at different operating stages, dynamically adjusting the fusion weights allows the model to focus more on the impact of periodic operations in the early stages of equipment aging and consider more of the trend changes brought about by equipment aging as the equipment approaches its maintenance cycle, thus improving the model's adaptability and accuracy in predicting yields.

[0091] The above scheme takes into account both long-term trends and short-term cyclical factors in the operation of the cracking furnace. Through adaptive regression analysis, the prediction model can better adapt to changes in the actual production process.

[0092] For example, a weighting function can be defined as a piecewise function.

[0093] When the cumulative running time is less than or equal to the typical desiccation cycle × β: the trend term fusion weight = α × (cumulative running time / typical desiccation cycle) ^ δ; when the cumulative running time is greater than or equal to the typical desiccation cycle × β: the trend term fusion weight = 1 - (1 - α) × exp(-η × (cumulative running time / typical desiccation cycle - β)).

[0094] Where: α is the benchmark weight coefficient, ranging from 0.3 to 0.7; β is the inflection point ratio, ranging from 0.4 to 0.8; δ is the power law exponent, ranging from 1.2 to 2.5; and η is the decay rate coefficient, ranging from 2.0 to 5.0.

[0095] It's worth explaining that the aging status of the pyrolysis furnace can be quantified by calculating the ratio of its cumulative operating time from the most recent decoking maintenance to the current moment to a typical decoking cycle. In the early stages of equipment operation, when the cumulative operating time is less than β times the inflection point ratio of a typical decoking cycle, the impact of equipment aging on production is relatively small, the trend term fusion weight is low, and the model focuses more on periodic changes and current operating conditions. However, when the cumulative operating time exceeds the inflection point ratio β, the impact of equipment aging becomes significant, the trend term fusion weight gradually increases, and the model begins to consider the trend changes brought about by equipment aging more.

[0096] Piecewise functions allow the trend term's weight to change differently at different stages of equipment operation. In the early stages of operation, there is a slow and continuous increase in weight, reflecting the gradual accumulation of the effects of equipment aging. In the later stages of operation, the trend term weight rapidly approaches 1, indicating that the impact of equipment aging becomes significant, and the weight quickly adjusts to primarily consider the trend term, i.e., the impact of equipment aging on production yield.

[0097] By dynamically adjusting the fusion weights of the trend and periodic terms, the yield prediction model can better adapt to changes in the aging state of the cracking furnace equipment. In the early stages of equipment operation, the model focuses more on the impact of periodic operating conditions; while in the later stages of equipment operation, the model focuses on long-term trend changes caused by equipment aging. This adaptive weight adjustment mechanism improves the accuracy and practicality of the model's predictions, providing more precise data support for parameter optimization in chemical production processes.

[0098] Figure 2 This is a flowchart of another method for determining the yield prediction value according to an embodiment of this application. Figure 3 This is an architecture diagram of a yield prediction model according to an embodiment of this application. The following is in conjunction with... Figure 2 and Figure 3 This section details the specific steps involved in determining the yield forecast.

[0099] Step S201, data reading.

[0100] During the production data reading training phase, production data from the production line was read. This data primarily consisted of time-series data, collected from historical production data of the cracking furnace, including COT temperature, hydrocarbon-to-water ratio, feed flow rate, and ethylene yield. For the time-series data, the data was averaged every 10 seconds, and a sliding window method was used for data segmentation. Data was then combined and input every 64 data points.

[0101] Step S202, model learning.

[0102] The input data is decomposed using the STL method, a typical approach in time series decomposition.

[0103] First, bilateral filtering is used to denoise the time series data. The main idea of ​​bilateral filtering is to use neighboring values ​​with similar values ​​to smooth the time series:

[0104]

[0105] Where J represents a filter window of length 2H+1, and the filter weights are given by two Gaussian functions:

[0106]

[0107] Where 1 / z is the normalization factor, and These are two parameters that control smoothness.

[0108] After noise removal, the trend term is extracted using the mean sliding window method:

[0109]

[0110] Where m = 2k + 1 is the length of the sliding window.

[0111] Extract the remaining terms as periodic terms: .

[0112] Step S203, extracting periodic feature blocks ( Figure 3 The seasonal-block extraction method extracts periodic features.

[0113] 1. Patching: The input time series is divided into overlapping or non-overlapping local patches, which transforms the dimension of time steps into the dimension of the number of patches, thereby reducing the computational complexity of subsequent processing and giving the model the ability to directly model local patterns.

[0114] For input features (L is the sequence length, C is the feature dimension), reshape it into , where P is the patch length and N is the number of patches.

[0115] 2. SwitchU (Reversible Switching Unit): This is an innovative feature control unit that dynamically and reversibly selects which feature channels or paths to activate through a learnable gating mechanism (such as Gumbel-Softmax), enhancing the model's expressive power without significantly increasing computational burden.

[0116]

[0117] in, For the input features, F(X) employs a transformation consisting of two linear layers. The gate is a gating vector calculated from the input, with values ​​between 0 and 1. This represents element-wise multiplication (Hadamard product).

[0118] 3. Batch normalization: Standardize each feature channel to stabilize the training process of deep networks and accelerate convergence.

[0119]

[0120] in, This represents the mean of the current small batch of data. This represents the variance of the current small batch of data.

[0121] 4. Temporal Convolutional Networks: TCNs with dilated causal convolutions are used to efficiently capture long-term dependencies in periodic terms. Causality ensures that predictions do not use future information, and the dilation mechanism allows neurons to capture longer-term patterns with exponentially expanding receptive fields.

[0122] For a one-dimensional input X and a convolution kernel f:

[0123]

[0124] in: : Dilated convolution operation with dilation factor d; K: Size of convolution kernel.

[0125] 5. MLP Fusion: Located at the end of the Seasonal Block, it fuses the multi-scale periodic features extracted from all TCN layers in the block. It consists of one or more multilayer perceptron layers, integrating features into a high-level representation through nonlinear transformation. Formula:

[0126]

[0127] in, It is a non-linear activation function GELU.

[0128] Step S204, extract blocks through trend features ( Figure 3 Trend features are extracted using the Trand-Block algorithm.

[0129] 1. Layer Normalization: Standardizes all feature channels of a single sample. It is particularly suitable for sequence models, insensitive to mini-batch size, and allows for stable training. Formula:

[0130]

[0131] in, and These are the mean and variance calculated along the feature dimension.

[0132] 2. Fully Connected Layer: Function: Effectively learns global and long-term patterns in trend terms through simple linear transformations and non-linear activations. Its concise structure avoids overfitting smooth trend signals.

[0133] Step S205, Output & Fusion.

[0134] 1. Layer normalization: Standardize the features before fusion to ensure stable data distribution.

[0135] 2. Layer Connection: This connects the periodic features output by the Seasonal-Block. Trend characteristics of Trand-Block output Perform concatenation or additive fusion.

[0136]

[0137] 3. Fully-connected Output Layer: Maps the fused high-level features to the final prediction space, outputting the predicted ethylene yield for a future time period (e.g., the next 5 minutes). .

[0138]

[0139] Where h is the prediction horizon.

[0140] The above steps abandon the traditional approach of treating time-series data as a whole for modeling, and innovatively adopt time-series decomposition technology. First, the sequences of key performance indicators such as ethylene yield are adaptively decomposed into trend features and periodic features. For the slowly changing nature of trend features, a multilayer perceptron (MLP) is used for fitting to capture the underlying long-term evolution patterns; for the complex fluctuations of periodic features, a temporal convolutional network is used for modeling, leveraging its advantages of dilated causal convolution and residual connections to efficiently capture multi-scale periodic patterns. Finally, the two prediction results are fused to achieve high-precision and highly robust yield prediction.

[0141] The above steps embed a high-precision yield prediction model as a surrogate model within the framework of a heuristic optimization algorithm (such as a genetic algorithm or particle swarm optimization). This prediction model acts as the "eyes" of the optimization algorithm, quickly assessing the potential yield results of any set of candidate operating parameters (COT, hydrocarbon-to-water ratio) over a future period. The heuristic algorithm then uses this reliable prediction as a foundation to conduct a global search within a complex constraint space, seeking the optimal combination of operating parameters that maximizes the predicted yield. This method effectively combines the accuracy of the prediction model with the global search capability of the optimization algorithm.

[0142] The above steps fully consider the actual conditions of industrial sites. The design of the temporal decomposition module and TCN module significantly reduces model redundancy and the number of parameters, improves computational efficiency, and meets real-time requirements. The algorithm as a whole is trained on unlabeled historical data, reducing the dependence on data labeling.

[0143] Figure 4 This is a structural diagram of a yield prediction determination device according to an embodiment of this application, as shown below. Figure 4 As shown, the device includes:

[0144] The acquisition module 40 is used to acquire historical production data of the industrial furnace, which is used to produce chemical raw materials.

[0145] The first capture module 42 is used to capture the non-periodic global change patterns of historical production data using a feedforward neural network, and obtain the trend features in the historical production data.

[0146] The processing module 44 is used to segment historical production data to obtain a local block sequence, wherein the local block sequence is a two-dimensional tensor including multiple local blocks; a gating mechanism is used to generate a gating vector, and the weighted features are determined based on the gating vector and the local block sequence, wherein the gating vector is used to represent the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected to participate in subsequent calculations.

[0147] The second capture module 46 is used to capture the periodic fluctuation patterns of weighted features at different time scales using a temporal convolutional network, thereby obtaining periodic features in historical production data.

[0148] Module 48 is used to determine the predicted yield of chemical raw materials based on trend and cycle characteristics.

[0149] Optionally, the feedforward neural network is trained using the following method: Based on the multivariate coupling characteristics of chemical raw materials in the pyrolysis process, key operational variables are extracted from the training dataset, wherein the key operational variables include at least one of the following: pyrolysis depth and feed composition; time series alignment and standardization preprocessing are performed on the key operational variables to obtain standardized operational variable time series; an initial feedforward neural network is obtained, wherein the input layer dimension of the initial feedforward neural network is consistent with the variable dimension of the standardized operational variable time series; the initial feedforward neural network includes: multiple hidden layers using nonlinear activation functions, the nonlinear activation functions of the hidden layers being used to map the nonlinear characteristics of the chemical raw material pyrolysis process; the standardized operational variable time series is used as the input of the initial feedforward neural network, and historical yield data of the chemical raw material is used as the training target, the connection weights and bias parameters of the initial feedforward neural network are optimized through the backpropagation algorithm to train the initial feedforward neural network until a preset stopping condition is met, thus obtaining the feedforward neural network.

[0150] Optionally, the chemical feedstock includes at least ethylene. The device for determining the yield prediction also has the following functions: mapping the PONA values ​​of the feed data in the time series of standardized operating variables to the operating variables to obtain the olefin formation potential index, wherein the operating variables are used to characterize the cracking depth and include at least the cracking furnace outlet temperature or hydrocarbon partial pressure; during the initial training process of the feedforward neural network, dynamic weights are assigned to the training samples based on the olefin formation potential index, wherein the olefin formation potential index is positively correlated with the weight coefficient assigned to the training samples.

[0151] Optionally, after obtaining the olefin generation potential index, the yield prediction determination device also has the following functions: generating an adaptive regularization term based on the olefin generation potential index, wherein the adaptive regularization term includes an error penalty term and an overfitting suppression term, the weight coefficient of the error penalty term is an increasing function of the olefin generation potential index, and the weight coefficient of the overfitting suppression term is a decreasing function of the olefin generation potential index; and adding the adaptive regularization term to the loss function of the backpropagation algorithm.

[0152] Optionally, a temporal convolutional network is used to capture the periodic fluctuation patterns of weighted features at different time scales to obtain periodic features in historical production data. This specifically includes the following steps: standardizing the weighted features to obtain target weighted features; capturing multi-scale periodic features of the target weighted features using a temporal convolutional network, where the multi-scale periodic features include short-term fluctuation patterns and long-term periodic patterns, and the temporal convolutional network includes multiple dilated causal convolutional layers with different dilation factors; performing nonlinear transformation and integration on the multi-scale periodic features to obtain a high-level periodic feature representation; and mapping the high-level periodic feature representation to periodic features using a fully connected layer.

[0153] Optionally, the yield prediction value of chemical raw materials is determined based on trend characteristics and periodic characteristics, specifically including the following steps: fusing trend characteristics and periodic characteristics to obtain fused characteristics; performing layer normalization on the fused characteristics to obtain standardized fused characteristics; and using a fully connected output layer to map the standardized fused characteristics into the yield prediction value of chemical raw materials within a preset prediction time range.

[0154] Optionally, the trend features and periodic features are fused to obtain fused features, specifically including the following steps: obtaining the cumulative running time of the industrial furnace from a historical time to a preset time, wherein the historical time is the time corresponding to the coking and maintenance event closest to the preset time; calculating the trend term fusion weight based on the cumulative running time using a predefined weight function, wherein the trend term fusion weight is an increasing function of the cumulative running time; weighting the trend features using the trend term fusion weight, weighting the periodic features using the periodic term fusion weight, and adding the weighted trend features and the weighted periodic features to obtain the fused features, wherein the periodic term fusion weight is the difference between 1 and the trend term fusion weight.

[0155] It should be noted that the above Figure 4 The modules in can be program modules (e.g., a set of program instructions that implements a specific function) or hardware modules. For the latter, they can be represented in the following forms, but are not limited to these: each of the above modules is represented by a processor, or the functions of each of the above modules are implemented by a processor.

[0156] It should be noted that, Figure 4 Preferred embodiments of the shown examples can be found in [reference needed]. Figure 1 The relevant descriptions of the embodiments shown will not be repeated here.

[0157] Figure 5 A hardware block diagram of a computer terminal for implementing a method for determining yield prediction values ​​is shown. Figure 5 As shown, the computer terminal 50 may include one or more processors 502 (shown as 502a, 502b, ..., 502n in the figure) 502 (processor 502 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 504 for storing data, and a transmission module 506 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 5 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 50 may also include... Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown.

[0158] It should be noted that the aforementioned one or more processors 502 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 50. As involved in the embodiments of this application, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0159] The memory 504 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the yield prediction value determination method in the embodiments of this application. The processor 502 executes various functional applications and data processing by running the software programs and modules stored in the memory 504, thereby realizing the above-mentioned yield prediction value determination method. The memory 504 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 504 may further include memory remotely located relative to the processor 502, and these remote memories can be connected to the computer terminal 50 via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0160] The transmission module 506 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 50. In one example, the transmission module 506 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 506 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0161] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 50.

[0162] It should be noted here that, in some optional embodiments, the above... Figure 5 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 5 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0163] It should be noted that, Figure 5 The computer terminal shown is used to execute Figure 1 The method for determining the yield prediction value shown above also applies to this electronic device, and will not be repeated here.

[0164] This application also provides a non-volatile storage medium, which includes a stored program, wherein the program, when running, controls the device where the storage medium is located to execute the above-mentioned method for determining the yield prediction value.

[0165] A non-volatile storage medium performs the following functions: acquires historical production data of an industrial furnace used to produce chemical raw materials; captures aperiodic global change patterns in the historical production data using a feedforward neural network to obtain trend features in the historical production data; segments the historical production data to obtain local block sequences, where each local block sequence is a two-dimensional tensor comprising multiple local blocks; generates a gating vector using a gating mechanism, and determines weighted features based on the gating vector and the local block sequences, where the gating vector represents the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected for subsequent calculations; captures periodic fluctuation patterns of the weighted features at different time scales using a temporal convolutional network to obtain periodic features in the historical production data; and determines the predicted yield of the chemical raw materials based on the trend features and periodic features.

[0166] This application also provides an electronic device, including: a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the above-described method for determining the yield prediction value.

[0167] The processor runs a program that performs the following functions: acquires historical production data from an industrial furnace used to produce chemical raw materials; captures aperiodic global change patterns in the historical production data using a feedforward neural network to obtain trend features; segments the historical production data to obtain local block sequences, where each local block sequence is a two-dimensional tensor comprising multiple local blocks; generates a gating vector using a gating mechanism and determines weighted features based on the gating vector and the local block sequences, where the gating vector represents the probability or weight of each feature channel corresponding to a process variable in the local block sequence being selected for subsequent calculations; captures periodic fluctuation patterns of the weighted features at different time scales using a temporal convolutional network to obtain periodic features in the historical production data; and determines the predicted yield of the chemical raw materials based on the trend features and periodic features.

[0168] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0169] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0170] In the above embodiments of this application, the information collected is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data all comply with relevant laws, regulations and standards, take necessary protective measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0171] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0172] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0173] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0174] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0175] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for determining a yield prediction value, characterized in that, include: Acquire historical production data of an industrial furnace used for producing chemical raw materials; By using a feedforward neural network to capture the non-periodic global change patterns of the historical production data, trend characteristics in the historical production data can be obtained. The historical production data is segmented to obtain a local block sequence, wherein the local block sequence is a two-dimensional tensor comprising multiple local blocks; a gating mechanism is used to generate a gating vector, and a weighted feature is determined based on the gating vector and the local block sequence, wherein the gating vector is used to represent the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected to participate in subsequent calculations; By using a temporal convolutional network to capture the periodic fluctuation patterns of the weighted features at different time scales, the periodic features in the historical production data are obtained. Based on the trend characteristics and the periodic characteristics, the predicted yield of the chemical raw material is determined; Based on the trend characteristics and the periodic characteristics, the predicted yield of the chemical raw material is determined, including: The trend feature and the periodic feature are fused to obtain the fused feature; The fused features are subjected to layer normalization to obtain standardized fused features; The standardized fusion features are mapped to the predicted yield values ​​of the chemical raw materials within a preset prediction time range using a fully connected output layer. The trend feature and the periodic feature are fused to obtain the fused feature, including: The cumulative running time of the industrial furnace from a historical time to a preset time is obtained, wherein the historical time is the time corresponding to the most recent coke removal maintenance event before the preset time; Based on the cumulative running time, the trend term fusion weight is calculated using a predefined weight function, wherein the trend term fusion weight is a monotonically increasing function of the cumulative running time. The trend feature is weighted using the trend term fusion weight, and the periodic feature is weighted using the periodic term fusion weight. The weighted trend feature and the weighted periodic feature are then added together to obtain the fusion feature, wherein the periodic term fusion weight is the difference between 1 and the trend term fusion weight.

2. The method according to claim 1, characterized in that, The feedforward neural network was trained using the following method: Based on the multivariate coupling characteristics of the chemical raw material in the pyrolysis process, key operational variables are extracted from the training dataset, wherein the key operational variables include at least one of the following: pyrolysis depth and feed composition; The key operational variables are subjected to time series alignment and standardization preprocessing to obtain standardized operational variable time series; An initial feedforward neural network is obtained, wherein the dimension of the input layer of the initial feedforward neural network is consistent with the dimension of the variable in the time series of the standardized operational variable; the initial feedforward neural network includes: multiple hidden layers using nonlinear activation functions, wherein the nonlinear activation functions of the hidden layers are used to map the nonlinear characteristics of the chemical raw material cracking process; The standardized operational variable time series is used as the input of the initial feedforward neural network, and the historical yield data of the chemical raw material is used as the training target. The connection weights and bias parameters of the initial feedforward neural network are optimized by the backpropagation algorithm to train the initial feedforward neural network until a preset stopping condition is met, thus obtaining the feedforward neural network.

3. The method according to claim 2, characterized in that, The chemical raw material includes at least ethylene, and the method further includes: The PONA values ​​of the feed data in the time series of the standardized operating variables are mapped to the operating variables to obtain the olefin generation potential index. The operating variables are used to characterize the cracking depth and include at least the cracking furnace outlet temperature or hydrocarbon partial pressure. During the training process of the initial feedforward neural network, dynamic weights are assigned to training samples based on the olefin generation potential index, wherein the olefin generation potential index is positively correlated with the weight coefficient assigned to the training samples.

4. The method according to claim 3, characterized in that, After obtaining the olefin formation potential index, the method further includes: An adaptive regularization term is generated based on the olefin generation potential index, wherein the adaptive regularization term includes an error penalty term and an overfitting suppression term, the weight coefficient of the error penalty term is an increasing function of the olefin generation potential index, and the weight coefficient of the overfitting suppression term is a decreasing function of the olefin generation potential index. Add the adaptive regularization term to the loss function of the backpropagation algorithm.

5. The method according to claim 1, characterized in that, By using a temporal convolutional network to capture the periodic fluctuation patterns of the weighted features at different time scales, the periodic features in the historical production data are obtained, including: The weighted features are standardized to obtain the target weighted features; The temporal convolutional network is used to capture the multi-scale periodic features of the target weighted features, wherein the multi-scale periodic features include: short-term fluctuation patterns and long-term periodic patterns, and the temporal convolutional network includes: multiple dilated causal convolutional layers with different dilation factors. The multi-scale periodic features are nonlinearly transformed and integrated to obtain a high-level periodic feature representation. The high-level periodic feature representation is mapped to the periodic feature using a fully connected layer.

6. A device for determining a yield prediction value, characterized in that, include: An acquisition module is used to acquire historical production data of an industrial furnace, wherein the industrial furnace is used to produce chemical raw materials; The first capture module is used to capture the non-periodic global change patterns of the historical production data using a feedforward neural network, and to obtain the trend features in the historical production data. The processing module is used to segment the historical production data to obtain a local block sequence, wherein the local block sequence is a two-dimensional tensor including multiple local blocks; a gating mechanism is used to generate a gating vector, and a weighted feature is determined based on the gating vector and the local block sequence, wherein the gating vector is used to represent the probability or weight of the feature channel corresponding to each process variable in the local block sequence being selected to participate in subsequent calculations; The second capture module is used to capture the periodic fluctuation patterns of the weighted features at different time scales using a temporal convolutional network, thereby obtaining the periodic features in the historical production data. A determination module is used to determine the predicted yield value of the chemical raw material based on the trend feature and the periodic feature: fusing the trend feature and the periodic feature to obtain a fused feature; performing layer normalization on the fused feature to obtain a standardized fused feature; mapping the standardized fused feature to the predicted yield value of the chemical raw material within a preset prediction time range using a fully connected output layer; fusing the trend feature and the periodic feature to obtain the fused feature includes: obtaining the cumulative running time of the industrial furnace from a historical time to a preset time, wherein the historical time is the time corresponding to the nearest coking maintenance event to the preset time; calculating the trend term fusion weight based on the cumulative running time using a predefined weighting function, wherein the trend term fusion weight is a monotonically increasing function of the cumulative running time; weighting the trend feature using the trend term fusion weight, weighting the periodic feature using the periodic term fusion weight, and adding the weighted trend feature and the weighted periodic feature to obtain the fused feature, wherein the periodic term fusion weight is the difference between 1 and the trend term fusion weight.

7. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the non-volatile storage medium to perform the method for determining the yield prediction value as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, include: A memory and a processor, the processor being configured to run a program stored in the memory, wherein the program, when running, executes the method for determining the yield prediction value as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for determining the yield prediction value as described in any one of claims 1 to 5.

Images