Method for improving efficiency of preparing ethylene from ethane based on refining and chemical scene

Abstract

The invention discloses a refining scene-based ethane-to-ethylene efficiency improvement method, which comprises the following steps of: in a plurality of clients, based on local operation data of each production unit, creating a mixed loss function to train a local model, and generating a local model update quantity and a group of contribution degree metadata; a central server receives local model update quantities which are uploaded by clients and are subjected to homomorphic encryption, determines aggregation weights for the update quantities according to contribution degree metadata in a plaintext form, and completes weighted aggregation in an encryption domain so as to update a global prediction model; and at the client, the global model subjected to localization fine tuning is used as a prediction simulator, the global model is embedded into an online optimization problem to solve and issue an optimal operation parameter, and continuous iteration of the model and closed-loop improvement of production efficiency are realized through performance monitoring and self-adaptive triggering steps. The reliability and performance of the model are improved, and continuous self-adaptive optimization of ethylene production efficiency is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and industrial process control technology, and in particular to a method for improving the efficiency of ethane-to-ethylene production in a refining scenario. Background Technology

[0002] Ethane cracking to ethylene is one of the core processes in modern chemical industry, and its production efficiency directly affects the economic benefits of the entire petrochemical industry chain. To improve ethylene yield, reduce unit energy consumption, and extend the operating cycle of cracking furnaces, the industry is constantly exploring more precise process modeling and optimization control technologies. Current mainstream methods mainly rely on complex models based on mechanisms such as chemical reaction kinetics, fluid mechanics, and heat transfer. While these models possess clear physical meaning, their construction process is often overly cumbersome, highly dependent on process parameters, and struggles to accurately capture dynamic nonlinear changes caused by factors such as feedstock fluctuations, catalyst deactivation, and equipment aging, thus limiting model adaptability and prediction accuracy.

[0003] With the development of artificial intelligence technology, modeling the ethane cracking process using data-driven methods such as deep neural networks has become a new research direction. However, directly applying existing data-driven technologies to complex refining scenarios typically faces three major challenges. First, data from multiple production units or branches within large petrochemical groups often form "data silos" that cannot be physically centralized due to commercial competition, data security regulations, and privacy protection requirements. Traditional machine learning models cannot train on all data in this scenario, thus severely limiting the model's generalization ability and resulting in an upper limit to model performance. Second, purely data-driven models, as a "black box," rely solely on data correlation in their training process, lacking an understanding of the inherent physicochemical laws of the process. This not only requires massive amounts of training data but also means that their predictions may violate basic chemical engineering laws such as mass conservation and energy conservation, leading to poor robustness and reliability when facing unseen or abnormal operating conditions. Third, in multi-party collaborative modeling scenarios, the process conditions, raw material batches, equipment status, and data quality of each participant (i.e., each production unit) often differ significantly, exhibiting non-independent identically distributed (Non-IID) data characteristics. If the contributions of all parties are treated equally without distinction in distributed learning, low-quality or irrelevant data will severely drag down the performance of the global prediction model, and may even cause the model to fail to converge. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a method for improving the efficiency of ethane-to-ethylene production in a refining scenario, to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method for improving the efficiency of ethane-to-ethylene production in a refining scenario. This method is based on a federated learning framework that incorporates physical mechanism constraints, and trains and iterates a global prediction model for predicting the ethane cracking process. Specific steps include:

[0007] In multiple clients associated with multiple ethane cracking production unit data sources, a local model is trained based on the local operating data of each production unit through a hybrid loss function that includes data-driven loss components and physical constraint loss components, generating a local model update quantity.

[0008] A central server receives local model updates from multiple clients and assigns an aggregation weight to each local model update based on a preset contribution evaluation strategy. The received local model updates are then weighted and aggregated to update the global prediction model.

[0009] Based on the updated global prediction model, operating parameter settings are generated to guide the ethane cracking process.

[0010] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the local operating data of each production unit is processed into feature vectors and performance tag vectors. The feature vectors include at least one of raw material fingerprint features, process operation features, and equipment status features. The performance tag vectors include values ​​related to ethylene yield, unit ethylene comprehensive energy consumption, or coking rate.

[0011] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the physical constraint loss component is calculated as follows:

[0012] Obtain the partial derivative values ​​of the predicted output of the local model with respect to one or more process input variables, and determine the magnitude of the physical constraint loss component based on the deviation between the partial derivative values ​​and the target values ​​determined by the ethane cracking chemical engineering mechanism.

[0013] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the contribution evaluation strategy includes generating a set of contribution metadata on the client side and uploading it along with the local model update. The contribution metadata includes at least one of the following:

[0014] The decrease in the hybrid loss function during this local model training process;

[0015] Data quality metrics used to characterize the quality or distribution characteristics of the locally running data;

[0016] And, a physical compliance score that quantifies the degree to which the local model satisfies the physical constraints after training.

[0017] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the process of determining the polymerization weight follows the following rules:

[0018] The aggregate weight value assigned to any client is positively correlated with the rate of decline, data quality, or physical compliance score contained in the contribution metadata uploaded by that client.

[0019] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the method further includes: before updating the global prediction model:

[0020] On the client side, a homomorphic encryption algorithm is used to encrypt the local model update amount;

[0021] On the central server, weighted aggregation is performed directly on encrypted local model update amounts.

[0022] As a preferred embodiment of the method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in this invention, the method further includes an iterative loop step: after updating the global prediction model, the updated global prediction model is distributed to the multiple clients as the initial model for the next round of client model training.

[0023] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the process of generating the operating parameter setpoints includes:

[0024] The updated global prediction model is used as a prediction simulator and embedded into an optimization problem with comprehensive energy consumption as the objective function. The optimization algorithm is then used to solve the problem and obtain the optimal sequence of operating parameters.

[0025] Use the first operation parameter in the sequence as the setting value of the current operation parameter.

[0026] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, a localization fine-tuning step is further included before using the global prediction model as a prediction simulator:

[0027] On the client side, the received global prediction model is trained using its associated local runtime data to generate a personalized prediction model adapted to the characteristics of the local data, and this personalized prediction model is used as the prediction simulator.

[0028] As a preferred embodiment of the ethane-to-ethylene efficiency improvement method based on a refining scenario described in this invention, the method further includes a performance monitoring and adaptive triggering step:

[0029] After deploying the global prediction model on the client side, continuously monitor the deviation between its prediction performance metrics and actual production data;

[0030] When the statistical characteristics of the deviation exceed a preset threshold, the training and iteration steps of the global prediction model of the federated learning framework are automatically re-executed.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. This invention adopts a federated learning framework, which effectively breaks down the "data silos" between different production units by training the model locally on the client and uploading only the model update. It ensures that the core operating data of each party does not leave the local machine and fully protects data privacy and business security. It also realizes the collaborative sharing of tacit knowledge across units, builds a prediction model with a global perspective, and improves the model's generalization ability and prediction accuracy.

[0033] 2. By introducing a hybrid loss function that includes physical constraint loss during client-side model training, the chemical engineering mechanism of the ethane cracking process is incorporated into the model as prior knowledge. This overcomes the drawbacks of a "black box" model driven by pure data, ensuring that the model's predictions are physically reasonable and self-consistent. This enhances the model's robustness and reliability when facing unseen or abnormal operating conditions. At the same time, the weighted aggregation strategy based on contribution metadata can intelligently evaluate and differentiate the contributions of each party, effectively suppressing low-quality data interference caused by the non-independent and identically distributed (Non-IID) characteristics of the data, accelerating model convergence, and obtaining a global model with better performance.

[0034] 3. This invention also embeds the personalized fine-tuned model into an online optimization problem as a prediction simulator, solves the problem in reverse, and issues the optimal operating parameters, directly transforming the model into actual production benefits. In addition, this invention includes performance monitoring and adaptive triggering steps, which can autonomously sense the model performance degradation and automatically start the relearning process, ensuring that the ethane cracking process can operate stably in the optimal state for a long time and achieve a continuous improvement in ethylene efficiency. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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. Wherein:

[0036] Figure 1 This is a flowchart illustrating the overall process of an ethane-to-ethylene efficiency improvement method based on a refining scenario, according to one embodiment of the present invention. Detailed Implementation

[0037] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0038] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0039] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0040] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0041] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0042] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0043] Example 1

[0044] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for improving the efficiency of ethane-to-ethylene production in a refining scenario, including:

[0045] It should be noted that the present invention adopts a federated learning framework that incorporates physical mechanism constraints to train and iterate a global prediction model for predicting the ethane cracking process. The federated learning architecture adopts a centralized architecture, namely a client-server architecture.

[0046] Specifically, the present invention is executed in a star topology consisting of a central server and multiple clients. Each client corresponds to an independent ethane cracking production unit and can access the unit's local database and computing resources. The central server is responsible for task distribution, model aggregation, and global prediction model management, but it does not access any raw production data from the clients. The entire execution process is iterative, with each iteration aiming to absorb the local knowledge of each client to incrementally optimize the performance of the global prediction model.

[0047] S1. In multiple clients associated with multiple ethane cracking production unit data sources, a local model is trained based on the local operating data of each production unit through a hybrid loss function that includes data-driven loss components and physical constraint loss components, and a local model update quantity is generated.

[0048] Furthermore, this S1 step is performed on each selected participant in this round (the... (Round) Federated Learning Training Client Execution is performed independently and in parallel on each client. First, receive the latest global prediction model parameters from the central server. As the starting point for its local model training, the client then performs the following sub-steps:

[0049] S101. Local Data Preparation and Feature Processing Operations: Client The computing nodes access the historical operation database of their associated ethane cracking production units, stored locally, in real time via the OPC (OLE for Process Control) protocol or direct database connection. It is important to note that the raw time-series data stored in this database needs to undergo preprocessing operations such as data cleaning, alignment, missing value imputation, and outlier removal before it can be used. After the preprocessing operations are completed, the processed raw time-series data is used to construct structured training sample pairs. ,in, For sample index, For feature vectors, This is a performance label vector;

[0050] Specifically, this feature vector represents a digital description of the state of the ethane cracking process, and preferably, it contains information features in at least one of the following dimensions:

[0051] Raw material fingerprint characteristics: The concentrations of components such as ethane and propane, as well as the contents of key impurities such as trace sulfides, are obtained by directly using the measurement point values ​​after sensor alignment and analysis by near-infrared spectroscopy or gas chromatography. The concentrations of these components and the contents of these impurities are used as material property variables for entering the cracking furnace.

[0052] Process operating characteristics: These characteristics refer to variables that can be adjusted by the operator or the process control system (APC), such as the coil outlet temperature (COT), the steam to hydrocarbon mass ratio, and the average residence time of the material in the coil.

[0053] Equipment status characteristics: These characteristics are variables that reflect the current health status of the pyrolysis furnace, such as the tube skin temperature (TST) and the operating time of the pyrolysis furnace.

[0054] Specifically, the performance label vector is the actual performance indicator representing production efficiency at the time corresponding to the feature vector, and preferably, it includes the values ​​of one or more of the following key performance indicators:

[0055] Ethylene yield: expressed as the ratio of the mass of the target product ethylene to the mass of the feedstock ethane;

[0056] Unit ethylene comprehensive energy consumption: refers to the total energy consumed in producing one unit mass of ethylene;

[0057] It should be noted that ethylene yield and unit ethylene comprehensive energy consumption are usually not measured directly, but are calculated through material balance and energy balance mechanism models, based on multiple measured values ​​(such as raw material flow rate, product flow rate, fuel gas consumption, etc.).

[0058] Coking rate: refers to the rate at which coking occurs on the inner wall of the furnace tube, which is indirectly calculated based on factors such as furnace tube pressure drop and changes in heat transfer coefficient, or estimated through a model. It directly affects the continuous operation cycle of the pyrolysis furnace.

[0059] In addition, considering that the local model is very sensitive to the scale of the input data, after the feature vector and performance label vector are determined, they also need to be standardized.

[0060] Specifically, the present invention employs Z-score standardization to ensure that all features are within a similar numerical range, thus accelerating the model convergence process. The standardization formula for the feature vectors is as follows:

[0061]

[0062] in, Represented as the i-th dimension standardized feature vector , and are the mean and standard deviation of the i-th dimension on the model training set, respectively;

[0063] It should be noted that both the performance label vector and the feature vector are standardized using Z-score, and the standardization formula is the same, so it will not be repeated here.

[0064] Specifically, the present invention uses a feedforward neural network (i.e., a multilayer perceptron (MLP)) as the backbone of the local model. In one feasible implementation of the present invention, the structure of the local model is as follows: an input layer (the number of neurons equals the feature vector). The system consists of 3-5 hidden layers (each containing 64-256 neurons, using ReLU as the activation function) and an output layer (the number of neurons equals the performance label vector). (dimensions of the pyrolysis process), used to fit the complex nonlinear relationships of the pyrolysis process;

[0065] S102. Training by creating a local model with a mixed loss function: Client Use its local dataset For its local model (with) The model is trained multiple times (using the initial representation of local model parameters) with the goal of finding a set of updated model parameters. The core of this training lies in minimizing the mixed loss function;

[0066] It should be noted that the creation of the hybrid loss function aims to require that the local model not only be able to learn, but that the learning process conforms to chemical engineering mechanisms;

[0067] Furthermore, the hybrid loss function consists of two parts: a data-driven loss component and a physical constraint loss component.

[0068] Furthermore, for N samples in the training set, calculate the local model's predicted value. and performance tag vector The mean squared error between the two is used to allow the local model to learn while fully fitting historical data, with the aim of making the model's predictions as close as possible to real production data. This is the data-driven loss component. Represented as:

[0069]

[0070] in, Represents local dataset The total number of samples;

[0071] Furthermore, the purpose of the physical constraint loss component is to embed known chemical engineering mechanisms as prior knowledge into the local model, ensuring that the predictive behavior of the local model conforms to basic physical laws. This process does not use any performance label vectors. Instead, it creates some input sample points out of thin air and checks whether the local model behaves reasonably on these input sample points.

[0072] Specifically, in one implementable process of the present invention, the physical constraints of the chemical engineering mechanism include, but are not limited to, the following three forms:

[0073] Monotonicity constraint: Assuming that within the normal operating range, the ethylene yield increases with furnace tube outlet temperature. The value increases monotonically with the increase of the input sample points; by randomly generating a batch of virtual input sample points. The physical constraint loss components under this monotonic constraint are obtained. for:

[0074]

[0075] Where M represents the virtual input sample points. The total number, It is expressed as the partial derivative of the ethylene yield output by the local model with respect to the input furnace tube outlet temperature, obtained through automatic differentiation; This is expressed as a predicted ethylene yield. It is a modified linear unit function. When the partial derivative is positive (in accordance with the chemical engineering mechanism), the output of ReLU is 0 and no loss is generated. When the partial derivative is negative (violating the chemical engineering mechanism), the output of ReLU is positive, which generates a quadratic penalty term (positive loss), thereby driving the model parameters to adjust in the direction that satisfies the monotonicity constraint.

[0076] Mass conservation constraint: Ideally, the mass of ethane input should equal the total mass of unreacted ethane, ethylene, hydrogen, methane, and other byproducts output. Therefore, this can be simplified to a single constraint: the sum of the mass fractions of all products (including unreacted substances) must be 1. (Assuming the model...) The output is a vector containing the mass fraction of each product. Let P be the total number of vectors representing the mass fractions of each product. Then, under this mass conservation constraint, the physical constraint loss component... It can be represented as:

[0077]

[0078] Boundary constraints: Since the yield of any product cannot be negative, the theoretical yield of ethylene should have an upper limit (approximately 55% yield from 80% conversion). The physical constraint loss component under this boundary constraint is then obtained. :

[0079]

[0080] in, This represents the theoretical upper limit of ethylene yield;

[0081] Furthermore, based on the physical constraints defined above and the data-driven loss component mentioned earlier, the created hybrid loss function can be expressed mathematically as follows:

[0082]

[0083] in, For a mixed loss function, , , These are represented as hyperparameters for each physical constraint term, primarily used to balance the importance of different physical constraint terms and satisfy... It can be set based on expert experience or through methods such as grid search, which will not be elaborated here;

[0084] S103. Generating Local Model Updates: The client uses its local dataset and the hybrid loss function created above to perform multiple rounds of gradient descent training using the Adam optimizer. In each round of training, updates are generated based on... The total gradient is calculated to fine-tune the local model parameters;

[0085] Furthermore, after training concludes, to protect data privacy, the client... Instead of directly uploading its pre-trained local model parameters, it calculates and generates update values ​​for the local model parameters. :

[0086]

[0087] It should be noted that this update volume Essentially, it represents the change or gradient direction of the local model relative to the initial global prediction model after absorbing local data knowledge (learning) and physical constraint loss.

[0088] S2. A central server receives local model updates from multiple clients and assigns an aggregation weight to each local model update based on a preset contribution evaluation strategy. The received local model updates are then weighted and aggregated to update the global prediction model.

[0089] Furthermore, within a preset communication window, the central server listens for and receives data from all clients participating in this round of training (assuming a total of...). The client uploads data packets (individual data packets), and to ensure absolute security and privacy during the communication process, the client uploads encrypted ciphertext instead of plaintext model parameter update data.

[0090] Specifically, when the client generates local model parameter update volume Then, it is immediately encrypted using a homomorphic encryption algorithm (such as Paillier or CKKS) to generate the encrypted local model parameter update. ;

[0091] It should be noted that the characteristic of homomorphic encryption is that it allows specific mathematical operations (such as addition and scalar multiplication) to be performed directly on the ciphertext. The decrypted result is the same as the result of performing the same operation on the plaintext. Therefore, the data packet received by the central server from the client contains a tuple:

[0092] in, It is plaintext contribution metadata associated with the update amount. This contribution metadata does not contain any sensitive production data and is only used by the server to evaluate the value of the client's contribution.

[0093] Furthermore, upon receiving all After receiving the data packets from each client, the central server updates the parameters for each encrypted local model. Determine a fair and efficient aggregation weight, denoted as . This process abandons the traditional federated learning approach of weighting by data volume, and instead relies on contribution metadata uploaded by the client. Conduct an assessment;

[0094] Specifically, the strategy employed in this assessment is as follows:

[0095] S201, Parsing Contribution Metadata The central server retrieves contribution metadata. Extract at least one of the following three dimensions of information:

[0096] The decrease in the mixed loss function ( This value reflects the client's... The "learning progress" during local model training is typically the difference in loss before and after training. ,in, and These represent the loss values ​​before and after local model training, respectively.

[0097] Data quality indicators ( This metric quantifies the client The quality and uniqueness of the local dataset, in one feasible process of the present invention, is a comprehensive score. This comprehensive score combines the size of the local dataset, the entropy of the feature distribution (reflecting the diversity of the local dataset), and the KL divergence calculated by comparing the distribution of the local dataset with the global average distribution after aggregating all client data. The larger the divergence value, the more unique the client data is.

[0098] Furthermore, since the core of federated learning is that the original data does not leave the local machine, while the global average distribution directly aggregates all the data, the present invention adopts a global distribution approximation method to replace the global average distribution, so as to achieve the global average distribution without leaking the client's original data.

[0099] Specifically, each client Locally, each dimension of the data features is divided into equal intervals, and the number of samples falling into each interval is calculated to form a local histogram. Subsequently, each client, through Secure Multi-Party Computation (SMPC), jointly calculates the sum of all local histograms. This process does not require any party to know the specific histograms of the others; ultimately, the central server uses the aggregated global histogram. This can approximate the average distribution of global data;

[0100] Physical conformity score ( This rating quantifies the client's performance. The degree to which the trained local model adheres to the preset physical mechanisms, in one feasible process of the present invention, is defined as follows: ,in, It is the value of the physical constraint loss component after training. The higher the value, the more the local model's behavior conforms to physical laws;

[0101] S202, Calculate contribution metadata The extracted scores for each dimension: To eliminate the influence of dimensions, the central server first normalizes the raw values ​​of each dimension across all clients, obtaining a normalized score. Therefore, regarding the aforementioned decrease... Its normalization formula can be expressed as:

[0102]

[0103] in, This represents the decrease after normalization. Similarly, the normalized data quality score can be obtained using the same normalization formula. Physical conformity score This will not be elaborated upon here;

[0104] S203. Calculate the aggregate weight: The central server, based on the above strategy, weights and merges the scores of the three dimensions into the final aggregate weight. :

[0105]

[0106] in, Similar to the hyperparameters in the aforementioned hybrid loss function, they represent the importance of "learning efficiency," "data value," and "model reliability" in the aggregation decision, respectively, and are non-negative. In the early stages of model training, we can set them to relatively large values. Values ​​can be increased later in training to encourage dataset diversity. and The value of is used to focus on the convergence and physical consistency of the model;

[0107] Furthermore, for all clients Their respective aggregation weights were determined. Afterwards, the central server performs an aggregation operation to update the global prediction model. It is important to note that this process must be performed in an encrypted domain to ensure that even the central server cannot snoop on the model update details of any individual client.

[0108] Furthermore, the update rules for the global prediction model parameters are as follows:

[0109] Furthermore, because homomorphic encryption algorithms support scalars of ciphertext and plaintext (i.e., aggregated weights) The central server can directly calculate the encrypted global update amount by performing multiplication operations on the ciphertext and addition operations on the ciphertext.

[0110]

[0111] in, Represented as an encrypted global update value;

[0112] In addition, after calculating the encrypted global update amount, the central server updates the global prediction model in two ways. :

[0113] The first method is server-side updates: a central server encrypts the global update volume. Send it to a trusted third party or hardware security module (HSM) with the private key for decryption to obtain the plaintext global update. Subsequently, the central server completed the update locally:

[0114]

[0115] The second method is client-collaborative decryption: This method employs a threshold encryption approach based on Secure Multi-Party Computation (SMPC). Multiple clients collaboratively generate a public key through a secure interactive protocol and distribute their private key shares among themselves. It's important to note that since each client only holds a share of the private key, it cannot reconstruct the complete private key independently. The central server broadcasts the encrypted global update to all clients. Each client receiving the request uses its unique private key share to perform a decryption operation on the encrypted global update. This decryption operation does not yield the final plaintext but generates a partial decryption result. The client then sends this partial decryption result to the central server. To prevent malicious clients from submitting forged partial decryptions, this process typically includes a zero-knowledge proof to demonstrate that the partial decryption result was indeed generated from the correct private key share and the received encrypted global update, without exposing the private key share itself. Once the central server has collected at least F (where F is a preset decryption threshold value, an integer greater than or equal to 2 and less than or equal to the total number of clients K) valid partial decryption results, it can combine these results using a publicly available aggregation algorithm to recover the final plaintext.

[0116] It should be noted that by updating the global prediction model, attackers are unable to steal model information by compromising the central server. To crack the encryption, attackers must simultaneously compromise and control at least F different clients within the same decryption window, thus increasing the attacker's cost and enhancing the model's security.

[0117] S3. Based on the updated global prediction model, generate operating parameter settings to guide the ethane cracking process;

[0118] It should be noted that after the global prediction model is updated, it is necessary to use the model to solve the production operation parameters of the ethane cracking process under the optimal state, so as to realize the online optimization control of the ethane cracking process and improve the ethylene production efficiency.

[0119] Furthermore, this step is executed periodically in each client's local control system (e.g., every 5 to 15 minutes) to adapt to real-time changes in production conditions.

[0120] In addition, in order to better adapt the global prediction model to the unique equipment characteristics and operating condition drift of each production unit, a quick local fine-tuning step can be performed before it is used for optimization.

[0121] Specifically, the localization fine-tuning steps are as follows:

[0122] The client receives the latest global model parameters Instead of using it directly, it is used as an initial parameter. A small batch of high-quality local running data that has been accumulated recently (e.g., the past 24 hours) is used to fine-tune the training for several rounds (e.g., 1 to 5 rounds). This process still uses the hybrid loss function created above for fine-tuning training, but the number of training rounds is relatively small. The purpose is to allow the global prediction model to quickly adapt to the latest local dynamics without "forgetting" the global knowledge.

[0123] After fine-tuning and training, the client can obtain a personalized prediction model. This personalized prediction model has both a global perspective and takes into account local characteristics, so it can be used as a prediction simulator to provide more accurate predictions.

[0124] Furthermore, the actual production optimization objective is formalized into a mathematical optimization problem, namely, the preferred solution of this invention aims to minimize the overall energy consumption of ethylene production and improve ethylene production efficiency. To this end, this invention constructs an objective function. ,in, These are the production operation parameters to be optimized;

[0125] Furthermore, the process operation characteristics defined above are used as decision variables of the objective function. Correspondingly, the raw material fingerprint characteristics and equipment status characteristics defined above are used as state variables of the objective function. It should be noted that at the current moment of optimization, both decision variables and state variables are considered as known and fixed inputs, and their current values ​​are obtained from the database.

[0126] It should be noted that since the objective function's unit is the comprehensive energy consumption of ethylene, its calculation requires the use of the prediction results from the personalized prediction model (or directly using the global prediction model). Based on this, we define the output of the personalized prediction model as a predicted performance label vector. This predicted performance label vector includes the performance indicators output by the prediction model, and its mathematical form is: ,in, For transpose operation, This is expressed as the predicted ethylene yield. Expressed as the predicted total energy consumption per unit of ethylene, This is expressed as the predicted coking rate. This invention provides a personalized prediction model for client k, and the preferred objective of this solution is to minimize the overall energy consumption of ethylene and improve ethylene production efficiency. Therefore, the objective function, which is the predicted overall energy consumption per unit of ethylene, is:

[0127]

[0128] It is worth noting that, based on the definition of the objective function above, maximizing the ethylene yield or minimizing the coking rate can also improve ethylene production efficiency. Therefore, the definition of this objective function depends on the performance index output by the prediction model.

[0129] Furthermore, in order to ensure that the solved operating parameters are physically feasible and comply with safe production regulations, a series of constraints must be imposed.

[0130] Specifically, taking the furnace tube outlet temperature and the steam-to-hydrocarbon mass ratio (gas-to-hydrocarbon ratio) as examples, boundary constraints for the manipulated variables can be set, namely:

[0131]

[0132]

[0133] in, and These represent the lower and upper limits of the furnace tube outlet temperature (COT), respectively. and These represent the minimum and maximum values ​​of the vapor-to-hydrogen ratio, respectively.

[0134] Specifically, taking the predicted ethylene yield and coking rate as examples, the predicted ethylene yield cannot be lower than the expected target value, and the predicted coking rate cannot be higher than the expected warning value. The resulting process performance constraints are as follows:

[0135]

[0136] in, For the desired target value (i.e., the desired ethylene yield), accordingly, This is the expected warning value (i.e., the expected coking rate).

[0137] Specifically, based on the temperature of the outer wall of the furnace tube For example, the temperature of the outer wall of the furnace tube must not exceed the material's tolerance limit, which is a safety constraint for the equipment:

[0138]

[0139] in, This is the upper limit of the furnace tube outer wall temperature;

[0140] In summary, the complete optimization problem can be expressed as:

[0141] in, This is the inequality form of the above constraints. The constraints represent the production operation parameters to be optimized, and the range of these parameters cannot exceed the lower limit of their own range. and upper limit , ;

[0142] Furthermore, once the optimization problem is defined, a nonlinear programming (NLP) optimization algorithm, such as Sequential Quadratic Programming (SQP) or Interior-Point Methods, is invoked through the client's computing node to solve it. Simultaneously, this algorithm uses a personalized prediction model as its core and iteratively calculates the gradients of the objective function and constraints with respect to the decision variables to find the optimal combination of production operation parameters that satisfies all constraints.

[0143] Specifically, the inventor uses the first operating parameter in the sequence as the operating parameter setting value at the current moment. Then, the setting value is sent to the distributed control system or process control system of each client (such as a factory A) to overwrite the original setting value. The actual process variables are then driven by calling the underlying PID controller to make the actual actuators (such as valves, heaters, etc.) track the setting value.

[0144] In addition, the present invention also includes a performance monitoring and adaptive triggering step to address the potential performance degradation of the model due to unmodeled changes in operating conditions (such as deep catalyst deactivation or significant changes in feed composition).

[0145] Specifically, for each client's personalized prediction model, its prediction performance metrics (such as...) are continuously... , , The model (etc.) is compared with the actual performance indicators obtained through mechanism calculations to calculate the prediction bias. By analyzing the statistical characteristics of this bias, its mean and standard deviation within the sliding time window are calculated. Based on these mean and standard deviation, corresponding mean thresholds and standard deviation thresholds are preset. When one of the following conditions occurs, it indicates that the model performance has significantly decreased:

[0146] Scenario 1: When the absolute value of the mean exceeds the mean threshold, it indicates a persistent systematic prediction bias.

[0147] Scenario 2: When the standard deviation exceeds the standard deviation threshold, it indicates that the predictive stability of the model decreases sharply.

[0148] It should be noted that the mean threshold and standard deviation threshold can be determined based on statistical analysis of the prediction deviation sequence generated in a stable operating cycle during the initial stage of model deployment. For example, the mean threshold can be set to twice the mean deviation in that cycle, and the standard deviation threshold can be set to three times the standard deviation of the deviation in that cycle (following the 3sigma principle) to ensure that normal fluctuations do not cause false triggers.

[0149] Specifically, once one of the above situations is triggered, the central server will receive a "model retraining" request from the client. When the central server receives enough requests from the client (at least two based on the star topology of the present invention), it will automatically restart a new round of federated learning process, that is, starting from step S1, using the latest data accumulated by all parties to retrain a new global prediction model that is more adaptable to the current working conditions.

[0150] It should be noted that by adaptively updating the global prediction model, the production operating parameters were ensured to remain in optimal condition during the ethane-to-ethylene process, thereby systematically improving ethylene production efficiency.

[0151] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0152] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0155] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0156] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for improving the efficiency of ethane-to-ethylene production in a refining scenario, characterized in that, The method is based on a federated learning framework that incorporates physical mechanism constraints to train and iterate a global prediction model for predicting ethane cracking processes. Specific steps include: In multiple clients associated with multiple ethane cracking production unit data sources, a local model is trained based on the local operating data of each production unit through a hybrid loss function that includes data-driven loss components and physical constraint loss components, generating a local model update quantity. A central server receives local model updates from multiple clients and assigns an aggregation weight to each local model update based on a preset contribution evaluation strategy. The received local model updates are then weighted and aggregated to update the global prediction model. Based on the updated global prediction model, operating parameter settings are generated to guide the ethane cracking process.

2. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1, characterized in that, The local operating data based on each production unit is processed into feature vectors and performance tag vectors. The feature vectors include at least one of raw material fingerprint features, process operation features, and equipment status features. The performance tag vectors include values ​​related to ethylene yield, unit ethylene comprehensive energy consumption, or coking rate.

3. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1 or 2, characterized in that, The physical constraint loss component is calculated in the following ways: Obtain the partial derivative values ​​of the predicted output of the local model with respect to one or more process input variables, and determine the magnitude of the physical constraint loss component based on the deviation between the partial derivative values ​​and the target values ​​determined by the ethane cracking chemical engineering mechanism.

4. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1, characterized in that, The contribution evaluation strategy includes generating a set of contribution metadata on the client side and uploading it along with the local model update, wherein the contribution metadata includes at least one of the following: The decrease in the hybrid loss function during this local model training process; Data quality metrics used to characterize the quality or distribution characteristics of the locally running data; And, a physical compliance score that quantifies the degree to which the local model satisfies the physical constraints after training.

5. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 4, characterized in that, The process of determining aggregation weights follows these rules: The aggregate weight value assigned to any client is positively correlated with the rate of decline, data quality, or physical compliance score contained in the contribution metadata uploaded by that client.

6. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1, characterized in that, Before updating the global prediction model, the following is also included: On the client side, a homomorphic encryption algorithm is used to encrypt the local model update amount; On the central server, weighted aggregation is performed directly on encrypted local model update amounts.

7. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1, characterized in that, The method further includes an iterative loop step, in which, after updating the global prediction model, the updated global prediction model is distributed to the multiple clients as the initial model for the next round of client model training.

8. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1, characterized in that, The process of generating operation parameter settings includes: The updated global prediction model is used as a prediction simulator and embedded into an optimization problem with comprehensive energy consumption as the objective function. The optimization algorithm is then used to solve the problem and obtain the optimal sequence of operating parameters. Use the first operation parameter in the sequence as the setting value of the current operation parameter.

9. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 8, characterized in that, Before using the global prediction model as a prediction simulator, a localization fine-tuning step is also included: On the client side, the received global prediction model is trained using its associated local runtime data to generate a personalized prediction model adapted to the characteristics of the local data, and this personalized prediction model is used as the prediction simulator.

10. The method for improving the efficiency of ethane-to-ethylene production in a refining scenario as described in claim 1, characterized in that, The method also includes a performance monitoring and adaptive triggering step: After deploying the global prediction model on the client side, continuously monitor the deviation between its prediction performance metrics and actual production data; When the statistical characteristics of the deviation exceed a preset threshold, the training and iteration steps of the global prediction model of the federated learning framework are automatically re-executed.

Images