Two-stage electrolytic copper foil electrodeposition energy consumption optimization method based on federal learning

Abstract

The invention belongs to the field of energy consumption optimization, and particularly relates to a two-stage electrolytic copper foil electrodeposition energy consumption optimization method based on federal learning. The method comprises the steps that each client side collects multi-dimensional technological parameters in the electrolytic copper foil electro-deposition process, and technological parameter vectors are constructed; constructing a federated optimization framework, training a local energy consumption predictor by each client in a local autonomous optimization stage, and calling an intelligent optimization algorithm for optimization; in the federated collaborative optimization stage, each client updates a local energy consumption predictor based on the structure parameters of the global energy consumption predictor generated by the server, and calls an intelligent optimization algorithm to perform collaborative optimization; and based on the latest local energy consumption predictor, carrying out backtracking evaluation on the optimal process parameter vector set, and screening out the process parameter vector which has the lowest predicted energy consumption and is finally applied to production control. According to the method, flexible support for various intelligent optimizers is realized, and generalization, stability and reliability of the model in a real industrial scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of energy consumption optimization, specifically involving a two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning. Background Technology

[0002] Traditional energy consumption optimization mainly revolves around two aspects: electrolyte parameters and non-electrolyte processes. Electrolyte control involves variables such as additives, copper ion concentration, acid concentration, and temperature. Energy consumption is reduced by improving deposition behavior and suppressing side reactions. However, complex coupling exists between these parameters, and improper settings can exacerbate polarization or ohmic losses, and even accelerate equipment corrosion. Non-electrolyte processes involve current intensity, electrode spacing, electrode material selection, waste heat recovery, and power supply strategies. A balance must be struck between improving current efficiency and maintaining stable production, and the effectiveness still heavily relies on human experience.

[0003] While the aforementioned methods have shown initial success in energy conservation, they are highly dependent on the experience of technical personnel and difficult to implement systematic and adaptive optimization. With the development of artificial intelligence technology, energy consumption modeling methods for electrolytic copper foil are gradually shifting from mechanism-driven to data-driven approaches. Existing research has improved energy consumption prediction accuracy by constructing hybrid prediction models through stacked ensemble learning, or by combining intelligent algorithms with optimization strategies to search for optimal process parameters, effectively promoting the development of this field towards digitalization and intelligence.

[0004] However, in real-world industrial scenarios, data from a single enterprise is often limited in scale and heterogeneously distributed, making it difficult to support the construction of high-precision global models. While federated learning can break down data silos through local training and parameter aggregation, enabling collaborative modeling with privacy protection, existing research is largely limited to Bayesian optimization frameworks, making it difficult to adapt to various optimizers such as genetic algorithms and particle swarm optimization, thus limiting its versatility and application scope in complex industrial optimization tasks. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing technologies and propose a two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning. By combining local autonomous optimization and federated collaborative optimization, it achieves flexible support for various intelligent optimizers. Furthermore, by employing adaptive parameter selection, dynamic learning rate decay, and optimal preservation strategies, it improves the generalization, stability, and reliability of the model in real industrial scenarios, ultimately constructing a high-precision federated optimization method with cross-client collaboration capabilities.

[0006] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0007] A two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning includes:

[0008] Each client collects multi-dimensional process parameters during the electrolytic copper foil electrodeposition process, constructs a process parameter vector, and forms a local dataset.

[0009] A federated optimization framework is used to train the local energy consumption predictor for each client. This framework includes a local autonomous optimization phase and a federated collaborative optimization phase, wherein:

[0010] During the local autonomous optimization phase, each client uses a parameter locking mechanism to train a local energy consumption predictor based on local data and uploads it to the server. Based on the trained local energy consumption predictor, a local objective function is constructed, and an intelligent optimization algorithm is called to find the optimal local process parameter vector set.

[0011] During the federated collaborative optimization phase, each client updates its local energy consumption predictor using a parameter adaptive filtering strategy and uploads it to the server. The server aggregates the structural parameters uploaded by each client to generate the structural parameters of the global energy consumption predictor for the current round, and calls the intelligent optimization algorithm to perform collaborative optimization and generate a collaborative optimal process parameter vector set.

[0012] In each round of the local autonomous optimization phase and the federated collaborative optimization phase, each client selects the optimal structural parameters for retention;

[0013] Based on the optimal structural parameters in each round, the local optimal process parameter vector set and the collaborative optimal process parameter vector set are back-evaluated to select the process parameter vector with the lowest predicted energy consumption that is finally applied to production control.

[0014] Furthermore, each client utilizes a parameter locking mechanism to train a local energy consumption predictor based on local data, including:

[0015] Each client builds a local energy consumption predictor based on a neural network;

[0016] In each round of the local autonomous optimization phase, the structure parameters updated in the previous round are used as the initial parameters for the current round, and the local energy consumption predictor is updated based on the local training set.

[0017] Furthermore, the step of constructing a local objective function based on the trained local energy consumption predictor and calling an intelligent optimization algorithm for optimization includes:

[0018] Each client builds the target function:

[0019]

[0020] in, This represents the total number of clients. Indicates the first A vector of process parameters for each client. Indicates the first Local energy consumption predictor for each client, Indicates the first The structural parameters of the local energy consumption predictor for each client, Indicates the first Wheel of Time Energy consumption prediction for each client, Indicates the first The lower bound of a vector of client-side process parameters. Indicates the first The upper bound of a vector of client-side process parameters. Indicates the number of communication rounds in the local autonomous optimization phase;

[0021] The objective function constructed by each client is optimized using an optimization algorithm to obtain the local optimal process parameter vector for each client under the current objective function.

[0022] Furthermore, each client updates its local energy consumption predictor using a parameter adaptive filtering strategy, including:

[0023] In each round of local training during the federated collaborative optimization phase, the structural parameters of the global energy consumption predictor from the previous round are used as the initial values ​​of the local energy consumption predictors for each client.

[0024] Each client performs multiple iterative updates, generating multiple sets of candidate structure parameters;

[0025] Using the independent validation sets built locally on each client, the prediction accuracy of each group of candidate structure parameters is calculated.

[0026] The candidate structure parameters with the highest prediction accuracy are selected as the parameters to be uploaded by the current client in this round.

[0027] Furthermore, in each round of the local autonomous optimization phase and the federated collaborative optimization phase, each client selects and retains the optimal structural parameters, including:

[0028] In each round of the local autonomous optimization phase and the federated collaborative optimization phase, the server distributes the structural parameters uploaded by all clients to every client except itself. Each client uses its local validation set to evaluate the distributed structural parameters, obtains the mean squared error corresponding to each structural parameter, and feeds it back to the server. The server then weights and aggregates the mean squared errors corresponding to each client to obtain the weighted mean squared error for each client.

[0029] If the weighted mean square error of the structural parameters after this round of updates is less than the weighted mean square error of the structural parameters retained in the previous round, then the structural parameters after this round of updates are selected as the optimal structural parameters and retained; otherwise, the structural parameters retained in the previous round are selected as the optimal structural parameters and retained.

[0030] Furthermore, in each round of training during the federated collaborative optimization phase, the learning rate is adjusted according to the following rules:

[0031]

[0032] in, Indicates the first Learning rate of the round, The initial learning rate, For the first The attenuation coefficient of the wheel, This is the number of rounds at which the learning rate begins to decay.

[0033] Furthermore, based on the optimal structural parameters of each round, a backtracking evaluation is performed on the local optimal process parameter vector set and the collaborative optimal process parameter vector set to select the process parameter vector with the lowest predicted energy consumption that is ultimately applied to production control, including:

[0034] The local optimal process parameter vector and the collaborative optimal process parameter vector of each round are substituted into the local energy consumption predictor corresponding to the optimal structural parameters retained in each round to obtain the historical energy consumption value sequence.

[0035] The minimum energy consumption value in the energy consumption value sequence is selected to obtain the corresponding process parameter vector that is finally applied to production control.

[0036] Furthermore, the local energy consumption predictor and the global energy consumption predictor are constructed using neural networks that include convolutional layers and fully connected layers.

[0037] Furthermore, the intelligent optimization algorithm employs any one of the following: particle swarm optimization, simulated annealing, whale optimization, stochastic gradient descent, Bayesian optimization, and genetic algorithm.

[0038] Compared with the prior art, the significant advantages of this invention are:

[0039] This invention constructs a general federated optimization framework that deeply decouples "energy consumption predictor training" from "process parameter optimization." The system first updates the structural parameters of the local energy consumption predictor by aggregating data from multiple plants through federated learning. Then, during the optimization phase, the predictor is locked and encapsulated into a fixed objective function, which can be called by any intelligent optimizer such as particle swarm optimization or genetic algorithms to search for the optimal process parameters. This design allows the optimization algorithm to avoid intervening in the gradient update details of the underlying federated learning, achieving modularity at the algorithm level, thus enabling the replacement of different optimizers like plug-ins. It provides flexible support for various intelligent optimizers such as particle swarm optimization, genetic algorithms, and Bayesian optimization.

[0040] To further enhance the adaptability and stability of the method, the framework integrates three core strategies: an adaptive parameter selection strategy that selects the optimal parameters for aggregation based on the performance of the local validation set, effectively suppressing the propagation of low-quality parameters caused by differences in client data distribution; a dynamic learning rate decay strategy that adjusts the learning rate exponentially according to the number of communication rounds, improving training stability through a balance between rapid initial convergence and fine-tuning in the later stages; and an optimal solution preservation strategy that ensures the effective preservation and consistent evaluation of historical optimal solutions during model iteration by establishing a unified evaluation benchmark and a monotonically decreasing selection mechanism.

[0041] This invention primarily addresses the contradiction between data silos (the inability to share data between different plants) and poor model generalization in industrial scenarios (a model trained solely by one plant may not be suitable for data from another plant). Through a federated collaboration mechanism, this invention integrates heterogeneous data from multiple clients to construct a more accurate and better generalized energy consumption prediction model, avoiding optimization biases caused by insufficient data from a single plant. Furthermore, by combining adaptive parameter filtering and optimal data preservation strategies, this invention not only eliminates low-quality model parameters but also ensures that historical optimal solutions are not lost during iterations. Thus, while protecting data privacy, it achieves a continuous, stable, and reliable reduction in energy consumption for complex electrodeposition processes. Attached Figure Description

[0042] Figure 1 This is a flowchart of a two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning, according to the present invention.

[0043] Figure 2 This is a communication diagram of the present invention;

[0044] Figure 3 This is a diagram illustrating the improvement in the determination coefficient of the local energy consumption predictor of this invention.

[0045] Figure 4 The diagram shows the optimized effects of each client in this invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0047] like Figures 1-2 As shown, this embodiment proposes a two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning, including the following steps:

[0048] Step 1: Each client deploys sensors to collect multi-dimensional process parameters, including current, copper ion concentration, acid concentration, temperature, cathode roller linear speed, foil production equipment number and specifications, to form a complete process parameter vector.

[0049] Step 2: Each client learns the complex mapping between process parameter vectors and energy consumption through a neural network model, thereby accurately capturing the deep coupling relationship between parameters such as current, concentration, and temperature. This model overcomes the limitations of traditional mechanistic and linear methods in handling nonlinear interactions, significantly improving the accuracy of energy consumption prediction.

[0050] This embodiment uses a convolutional neural network as the neural network model to construct process parameter vectors. Energy consumption The mapping relationship is used to calculate the predicted energy consumption value through forward propagation, and the neural network model parameters are iteratively updated using the backpropagation algorithm based on the mean square error between the predicted and actual values. The formula is expressed as follows:

[0051]

[0052] in, Represents a convolutional neural network. This represents the weights of the convolutional neural network.

[0053] Step 3: Based on the constructed neural network model, use an intelligent optimization algorithm as the optimizer to optimize the constructed objective function.

[0054] Step 3.1, for each client The constructed neural network model and the corresponding search space Construct the objective function:

[0055]

[0056] in, This is the total number of clients. Indicates the first A vector of process parameters for each client. It is the first Local energy consumption predictor for each client, Indicates the first The structural parameters of the local energy consumption predictor for each client, Indicates the first Energy consumption prediction for each client, Indicates the first The lower bound of a vector of client-side process parameters. Indicates the first The upper bound of the vector of client process parameters.

[0057] Step 3.2: Optimize the objective function constructed by each client using an optimization algorithm (such as particle swarm optimization, simulated annealing, whale optimization, stochastic gradient descent, Bayesian optimization, and genetic algorithm) to obtain the local optimal process parameter vector for each client under the current objective function.

[0058]

[0059] in, Indicates the first The local optimal process parameter vector for each client.

[0060] Step 4: In the local autonomous optimization phase, each client independently conducts model training and parameter optimization based on local data, calls the local optimizer to perform preliminary optimization on the neural network updated in each round, and adopts a structural parameter locking strategy to maintain training continuity, accumulating a high-quality initial solution set for the subsequent federated collaborative optimization phase.

[0061] Step 4.1: The number of samples of process parameter vectors uploaded by each client in the first round of local autonomous optimization. To the server;

[0062] Step 4.2, the client is in the... In round-robin communication, each client is based on its local dataset. implement Sub-local iteration training generates the structural parameters of the local energy consumption predictor. .

[0063] A parameter locking mechanism is employed, meaning that the client's initial parameters are not obtained from the server, but rather are reused from the parameters generated in the previous round of local communication.

[0064]

[0065] in, Indicates the first The initial iterative structure parameters of round-robin communication, This represents the structural parameters after the previous update. The initial iterative structure parameters for the first round of communication are randomly generated by each client. This parameter locking mechanism avoids frequent global parameter aggregation by reusing the results of the previous round, reducing communication overhead and maintaining the temporal continuity of local training.

[0066] Step 4.3: After completing the local model update, use the updated local energy consumption predictor... The client dynamically adjusts the process parameter vector through the optimizer and obtains the local optimal process parameter vector for this round. The corresponding objective function is as follows:

[0067]

[0068] in, Indicates the first Round communication Energy consumption prediction for each client, This represents the total number of communication rounds during the local autonomous optimization phase. The established objective function... It can be adapted to a variety of optimizers, including local energy consumption predictors. With this as the core, within the search space corresponding to the client, the process parameter vector is continuously adjusted. Iterate towards the parameter region with lower energy consumption, gradually approaching the local optimal process parameter vector that minimizes energy consumption, thereby keeping energy consumption in an optimal state.

[0069] In this embodiment, a round of communication is a complete closed-loop process from local computation on the client to the completion of parameter updates: First, each client performs iterative training based on local data to update the local energy consumption predictor and uses the predictor to search for the current local optimal process parameters; then, the client uploads the updated model structure parameters and validation metrics to the server, which receives them and performs aggregation processing or weight calculation before distributing them; finally, the client receives feedback from the server and updates the local model parameters according to preset rules (such as weighted mean square error comparison or aggregation update), thus marking the end of a round of communication.

[0070] Step 4.4: Each client will input the structural parameters of its local energy consumption predictor. The client's verification set obtained Uploaded to the server. After the server receives it, it will... The message is then sent to all other clients. Each client verifies the received message using an independent verification set. Verification was performed. Finally, the data is uploaded back to the server and weighted as follows:

[0071]

[0072] in, Indicates the first Round communication The weighted mean square error of the local energy consumption predictor for each client. , This indicates the number of communication rounds in the local autonomous optimization phase. Indicates the first Number of samples per client express Round communication The local energy consumption predictor for each client in the first... Mean squared error on client data This indicates the total amount of data.

[0073] Step 4.5, for , No. Wheel of Time The structural parameters that each client needs to retain are selected from the optimal structural parameters retained in the previous round and the structural parameters obtained in the current round, choosing the structural parameters with the smaller weighted mean square error on the validation set:

[0074]

[0075] in, Indicates the first Time-of-use client The optimal structural parameters that are retained. Indicates the first In the previous round, the client... The optimal structural parameters that are retained. Indicates the first Time-of-use client The obtained structural parameters, Indicates the first Time-of-use client The weighted mean square error of the obtained structural parameters, Indicates the first Time-of-use client The weighted mean square error of the retained optimal structural parameters.

[0076] Repeat steps 4.2-4.5 until the preset number of rounds is reached in this stage. Round communication The local optimal process parameter vectors are composed of local optimal process parameter vector sets.

[0077] This embodiment effectively maintains the continuity of the training process through a parameter locking mechanism, significantly reduces communication overhead, generates a high-quality initial solution set through multiple rounds of iterative optimization, ensures the generalization performance of the model parameters through a weighted evaluation mechanism, lays a solid foundation for subsequent federated collaborative optimization, and protects the data privacy of all participants.

[0078] Step 5: In the federated collaborative optimization phase, the server coordinates cross-domain knowledge fusion among various clients. Through a multi-round parameter aggregation and distribution mechanism, a global energy consumption predictor with stronger generalization ability is built. On this basis, an intelligent optimizer is combined to achieve collaborative optimization of process parameters. This phase integrates parameter adaptive screening and dynamic learning rate decay strategies, which work together from two dimensions: parameter quality assurance and training stability improvement, to comprehensively improve the optimization performance and convergence stability of the framework under non-independent and identically distributed industrial data.

[0079] Step 5.1: Employ a dynamic synchronization mechanism between global and local parameters. During client training, the client downloads the latest global energy consumption predictor's structure parameters from the server as initial values, replacing the strategy of using local historical parameters from the local autonomous optimization phase.

[0080]

[0081] in, Indicates the first The structural parameters of the global energy consumption predictor at round -1 are generated by the server using a weighted aggregation algorithm:

[0082]

[0083] in, The initial value is the optimal structure parameter retained in the last round of the local autonomous optimization phase.

[0084] Step 5.2: During the training process, let the learning rate of each client be... And adjust the learning rate in each training round according to the following rules (in The learning rate remained constant initially, then began to change.

[0085]

[0086] in, Indicates the first Learning rate of the round, The initial learning rate, For the first The attenuation coefficient of the wheel, This is the number of rounds at which the learning rate begins to decay.

[0087] Step 5.3: Each client generates several sets of structural parameters during local iteration (one set is generated per round of communication), and based on the verification set independently constructed locally by the client. The accuracy of the E candidate structural parameters is evaluated to reflect their suitability. The parameter selection criteria are as follows:

[0088]

[0089] in, Indicates the client In the Round communication Candidate structure parameters generated in the next iteration. Candidate structure parameters generated by local iteration In the independent verification set On the prediction accuracy, Generate the number of candidate structure parameters for each round of local training.

[0090] Step 5.4: Select the obtained structural parameters Substitute the updated local energy consumption predictor into the objective function and construct the following objective function to obtain the local optimal process parameter vector:

[0091]

[0092] in, Indicates the first Round communication Energy consumption prediction for each client, Indicates the number of communication rounds in the federated collaborative optimization phase. This is the total number of communication rounds.

[0093] Step 5.5: After each client optimizer has completed its optimization, each client updates its structure parameters based on the updated parameters. Calculate and will and Upload them together to the server, and the server will receive them. The parameter is then distributed to all other clients, who evaluate it using their local validation sets to obtain the corresponding mean squared error. And it feeds back to the server, the server responds to the first... The mean square errors of all client-side structural parameters are weighted and aggregated into a weighted mean square error. .

[0094] Step 5.6: Based on the calculated weighted mean square error, the server selects the better-performing structural parameter from the structural parameters retained by the client in the previous round and the structural parameters obtained in the current round as the optimal structural parameter for this round. :

[0095]

[0096] in, Indicates the first Time-of-use client The optimal structural parameters that are retained. Indicates the first In the previous round, the client... The optimal structural parameters that are retained. Indicates the first Time-of-use client The obtained structural parameters, Indicates the first Time-of-use client The weighted mean square error of the obtained structural parameters, Indicates the first Time-of-use client The weighted mean square error of the retained optimal structural parameters.

[0097] After the filtering is completed, the server processes the structural parameters uploaded by each client in this round. By performing aggregated weighted averaging, the structural parameters of the new global energy consumption predictor are obtained. .

[0098] Repeat steps 5.1-5.6 until the preset number of rounds is reached in this stage.

[0099] In this embodiment, during the federated collaborative optimization phase, the server integrates the local model parameters uploaded by each client using a weighted aggregation algorithm to construct a global energy consumption predictor based on the collaborative knowledge of each client. On this basis, each client uses this global energy consumption predictor as a shared evaluation function to call its local intelligent optimizer for iterative search. Since the global energy consumption predictor incorporates data features from multiple plants, the optimization process of each client is essentially guided by global knowledge, thus achieving cross-client process parameter optimization. This process also incorporates two strategies, enhancing the framework's optimization performance and generalization ability under non-independent and identically distributed industrial data. Round communication The three collaborative optimal process parameter vectors form a collaborative optimal process parameter vector set.

[0100] Step 6: After the two phases are completed, to address the inconsistency in the evaluation benchmark caused by the model's iterative changes, an optimal preservation strategy is adopted to take into account the knowledge of both clients and servers during the two phases.

[0101] Step 6.1, Client The optimal process parameter set obtained during the communication process (including the local optimal process parameter vector set and the collaborative optimal process parameter vector set) The local energy consumption predictor is constructed by substituting the optimal structural parameters retained in each round. From this, a historical energy consumption value sequence can be obtained. Among them, the client In the The energy consumption generated during the iteration of round-robin communication is , .

[0102] Step 6.2: To ensure that the result of each round is the optimal energy consumption value, the sequence after saving the optimal energy consumption value is as follows: , among which, the The optimal energy consumption value of the wheel is the front The formula for the minimum energy consumption value after round screening is as follows:

[0103]

[0104] in, Indicates the first The optimal energy consumption value of the wheel.

[0105] Backtracking The process parameters are then used to obtain the final vector of process parameters applied to production control. .

[0106] In the early stages of federated learning, the accuracy of the local energy consumption predictor is low, and some process parameters may be mistakenly recorded as "extremely low energy consumption" due to model prediction errors. Directly using these parameters may not be able to reproduce the low energy consumption effect in actual production. The backtracking strategy of this method involves uniformly substituting the candidate optimal process parameters generated in each historical round into the final, most accurate, and most generalized global energy consumption predictor for re-evaluation. This is equivalent to conducting a fair horizontal comparison of all candidate solutions under a unified, robust, and high-precision benchmark, thereby effectively eliminating early "pseudo-optimal solutions" and selecting the truly optimal process parameters that achieve the lowest energy consumption.

[0107] This embodiment employs an optimal retention strategy to evaluate the optimal process parameters from each iteration under the same robust optimal predictor. This mitigates evaluation bias and insufficient optimization caused by incomplete model convergence during federated learning, thereby selecting the optimal energy consumption value and its corresponding process parameters from the entire iteration history, providing a reliable basis for process parameter optimization in actual production. It should be noted that in the two-stage federated optimization mechanism, steps 4 and 5 of the model training process typically require multiple rounds of communication until the preset convergence condition is met; while step 6 of the process parameter optimization and application process can be executed as needed based on actual production requirements. The communication process and the optimization application process can be implemented independently or synchronously in an online learning scenario, thereby achieving continuous updating of model parameters and dynamic optimization of process parameters.

[0108] This invention proposes a two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning. By constructing a general federated optimization framework compatible with multiple intelligent optimizers, the optimization algorithm and the federated training process are decoupled. The framework design includes a two-stage collaborative mechanism of local autonomous optimization and federated collaborative optimization: In the local autonomous optimization stage, each client independently trains its local energy consumption predictor based on local data and uses multiple intelligent optimizers in parallel to search for high-quality initial solution sets; in the federated collaborative optimization stage, cross-plant knowledge fusion is achieved through server-side weighted parameter aggregation and multi-round interactions to construct a global energy consumption predictor with stronger generalization capabilities. To further improve the adaptability and stability of the method, the framework integrates three core strategies: an adaptive parameter selection strategy that selects the optimal parameters for aggregation based on the performance of the local validation set, effectively suppressing the propagation of low-quality parameters caused by differences in client data distribution; a dynamic learning rate decay strategy that exponentially adjusts the learning rate according to the number of communication rounds, improving training stability through a balance between rapid initial convergence and fine-tuning in the later stages; and an optimal solution preservation strategy that ensures the effective preservation and consistent evaluation of historical optimal solutions during model iteration by establishing a unified evaluation benchmark and a monotonically decreasing selection mechanism. This invention effectively solves the problem of single-domain optimization limitations caused by data silos in industrial scenarios while protecting the data privacy of all participants. It significantly improves the model's generalization ability, convergence stability, and cross-client collaborative optimization efficiency under non-independent and identically distributed data.

[0109] In one specific embodiment, this embodiment is verified using on-site data from a copper foil manufacturing company, as detailed below:

[0110] During the data acquisition phase, high-precision sensors were used to monitor key process parameters of the production line, with a data acquisition frequency of once every 15 minutes for 18 months, resulting in a total of 133,660 valid samples. The collected parameters included seven categories of key process parameters: tank voltage, current, copper ion concentration, acid concentration, temperature, cathode roller speed, and foil-making equipment number and specifications. Feature vectors were constructed using these parameters, with tank voltage as the energy consumption output and the remaining parameters as model input features. A federated learning environment with two clients was constructed: 24 foil-making machines in plant A were designated as Client0, and 24 foil-making machines in plant B were designated as Client1, to simulate the characteristics of non-independent and identically distributed data. A neural network architecture containing convolutional and fully connected layers was used as the local and global energy consumption predictors. The first fully connected layer had an input dimension of 7 and an output dimension of 32. Subsequent layers used two 1D convolutional layers with 16 and 32 channels respectively, a kernel size of 3, and fully connected layers. The activation function was ReLU, the loss function was mean squared error, the optimizer was SGD, and the initial learning rate was set to... The batch size is 128. Downstream optimization tasks employ six types of optimizers: PSO (Particle Swarm Optimization), SA (Simulated Annealing), WOA (Whale Optimization), SGD (Stochastic Gradient Descent), BO (Bayesian Optimization), and GA (Genetic Algorithm) to optimize process parameters. The federated optimization framework consists of a 150-round local autonomous optimization phase and a 200-round federated collaborative optimization phase, with 5 local iterations per round. From the 200th round onwards, a dynamic learning rate decay of 0.95 is implemented, integrating adaptive parameter selection and optimal value preservation strategies. Training and validation sets are constructed based on actual production data. Through two-stage collaborative optimization, cross-client knowledge fusion and improved energy consumption predictor accuracy are achieved, ultimately outputting the optimal combination of process parameters that minimizes tank voltage. This effectively verifies the optimization performance and generalization ability of the proposed framework in a real industrial environment.

[0111] Figure 3 This paper demonstrates the improvement process of the determination coefficient of the cell voltage predictor for Client 0 and Client 1 under this method. The horizontal axis represents the number of communication rounds, and the vertical axis represents the determination coefficient reflecting the prediction accuracy. The solid line represents the mean of the determination coefficient of 5 experiments under the federated optimization framework, and the dashed line represents the mean of the determination coefficient of 5 experiments under the traditional centralized method. The shaded area corresponds to the fluctuation range of the determination coefficient under the two methods to reflect the stability of the results. The experimental results show that in the local autonomous optimization stage, the prediction accuracy of Client 0 and Client 1 is similar to that of the traditional centralized framework. After entering the federated collaborative optimization stage, their scores are improved to 0.9185 and 0.9200 respectively, which is much higher than 0.6765 and 0.3697 of the traditional centralized method. This advantage is due to the federated weighted aggregation algorithm and the parameter adaptive screening mechanism. Under the premise of ensuring that the data of each client is not shared, it effectively integrates the local model information of multiple clients, improves the prediction accuracy of the global energy consumption predictor for the voltage of the electrolytic copper foil electrodeposition cell, and lays the foundation for the subsequent accurate acquisition of the optimal process parameter combination.

[0112] Figure 4 Six intelligent optimizers—PSO, SA, WOA, SGD, BO, and GA—embedded with the method of this patent were demonstrated. These optimizers were used in the dynamic process of optimizing the slot voltage on client 0 and client 1, respectively. The horizontal axis represents the number of communication rounds, and the vertical axis represents the slot voltage (unit: V), a core indicator of energy consumption. A cross-client optimization environment was constructed. All optimizers achieved stable convergence under the federated framework. After 150 communication rounds, the voltage value tended to stabilize. Although there were slight differences in the convergence speed and final voltage among different optimizers, all effectively reduced the slot voltage. This verifies the good adaptability of this invention to non-independent and identically distributed data in industrial scenarios, as well as its versatility in compatibility with multiple intelligent optimizers. The energy consumption optimization goal was successfully achieved through cross-plant knowledge fusion.

[0113] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning, characterized in that, include: Each client collects multi-dimensional process parameters during the electrolytic copper foil electrodeposition process, constructs a process parameter vector, and forms a local dataset. A federated optimization framework is used to train the local energy consumption predictor for each client. This framework includes a local autonomous optimization phase and a federated collaborative optimization phase, wherein: During the local autonomous optimization phase, each client uses a parameter locking mechanism to train a local energy consumption predictor based on local data and uploads it to the server. Based on the trained local energy consumption predictor, a local objective function is constructed, and an intelligent optimization algorithm is called to find the optimal local process parameter vector set. During the federated collaborative optimization phase, each client updates its local energy consumption predictor using a parameter adaptive filtering strategy and uploads it to the server. The server aggregates the structural parameters uploaded by each client to generate the structural parameters of the global energy consumption predictor for the current round, and calls the intelligent optimization algorithm to perform collaborative optimization and generate a collaborative optimal process parameter vector set. In each round of the local autonomous optimization phase and the federated collaborative optimization phase, each client selects the optimal structural parameters for retention; Based on the optimal structural parameters in each round, the local optimal process parameter vector set and the collaborative optimal process parameter vector set are back-evaluated to select the process parameter vector with the lowest predicted energy consumption that is finally applied to production control.

2. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, Each client utilizes a parameter locking mechanism to train a local energy consumption predictor based on local data, including: Each client builds a local energy consumption predictor based on a neural network; In each round of the local autonomous optimization phase, the structure parameters updated in the previous round are used as the initial parameters for the current round, and the local energy consumption predictor is updated based on the local training set.

3. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, The process of constructing a local objective function based on the trained local energy consumption predictor and calling an intelligent optimization algorithm for optimization includes: Each client builds the target function: in, This represents the total number of clients. Indicates the first A vector of process parameters for each client. Indicates the first Local energy consumption predictor for each client, Indicates the first The structural parameters of the local energy consumption predictor for each client, Indicates the first Wheel of Time Energy consumption prediction for each client, Indicates the first The lower bound of a vector of client-side process parameters. Indicates the first The upper bound of a vector of client-side process parameters. Indicates the number of communication rounds in the local autonomous optimization phase; The objective function constructed by each client is optimized using an optimization algorithm to obtain the local optimal process parameter vector for each client under the current objective function.

4. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, Each client updates its local energy consumption predictor using a parameter adaptive filtering strategy, including: In each round of local training during the federated collaborative optimization phase, the structural parameters of the global energy consumption predictor from the previous round are used as the initial values ​​of the local energy consumption predictors for each client. Each client performs multiple iterative updates, generating multiple sets of candidate structure parameters; Using the independent validation sets built locally on each client, the prediction accuracy of each group of candidate structure parameters is calculated. The candidate structure parameters with the highest prediction accuracy are selected as the parameters to be uploaded by the current client in this round.

5. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, In each round of the local autonomous optimization phase and the federated collaborative optimization phase, each client selects and retains the optimal structural parameters, including: In each round of the local autonomous optimization phase and the federated collaborative optimization phase, the server distributes the structural parameters uploaded by all clients to every client except itself. Each client uses its local validation set to evaluate the distributed structural parameters, obtains the mean squared error corresponding to each structural parameter, and feeds it back to the server. The server then weights and aggregates the mean squared errors corresponding to each client to obtain the weighted mean squared error for each client. If the weighted mean square error of the structural parameters after this round of updates is less than the weighted mean square error of the structural parameters retained in the previous round, then the structural parameters after this round of updates are selected as the optimal structural parameters and retained; otherwise, the structural parameters retained in the previous round are selected as the optimal structural parameters and retained.

6. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, In each round of training during the federated collaborative optimization phase, the learning rate is adjusted according to the following rules: in, Indicates the first Learning rate of the round, The initial learning rate, For the first The attenuation coefficient of the wheel, This is the number of rounds at which the learning rate begins to decay.

7. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, Based on the optimal structural parameters of each round, the local optimal process parameter vector set and the collaborative optimal process parameter vector set are back-evaluated to select the process parameter vector with the lowest predicted energy consumption that is ultimately applied to production control, including: The local optimal process parameter vector and the collaborative optimal process parameter vector of each round are substituted into the local energy consumption predictor corresponding to the optimal structural parameters retained in each round to obtain the historical energy consumption value sequence. The minimum energy consumption value in the energy consumption value sequence is selected to obtain the corresponding process parameter vector that is finally applied to production control.

8. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, The local energy consumption predictor and the global energy consumption predictor are constructed using neural networks containing convolutional layers and fully connected layers.

9. The two-stage energy consumption optimization method for electrolytic copper foil electrodeposition based on federated learning according to claim 1, characterized in that, The intelligent optimization algorithm adopts any one of the following: particle swarm optimization, simulated annealing, whale optimization, stochastic gradient descent, Bayesian optimization, and genetic algorithm.

Images