Reaction process parameter predictive control method and system based on initial response constraint

Abstract

This invention relates to the field of data prediction technology, and particularly to a method and system for predicting and controlling reaction process parameters based on initial response constraints. The invention proposes the following approach: acquiring raw material detection information and initial reaction response information for the current batch of reactants; generating multiple candidate reaction parameter combinations based on target reaction requirements; inputting the above information into a reaction behavior prediction model to obtain the predicted cobalt conversion rate corresponding to the candidate combinations; and correcting the prediction results through stoichiometric constraints; when the predicted value exceeds the limit, determining the risk of side reactions by combining the excess residual amount and initial reaction response information, thereby screening the target reaction parameter combination. This application can improve the accuracy and stability of matching reaction parameters between different batches of recovered cobalt materials.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, and in particular to a method and system for predicting and controlling reaction process parameters based on initial response constraints. Background Technology

[0002] Cobalt-containing waste and recycled cobalt-manganese compounds are increasingly being reused to produce cobalt acetate, cobalt bromide, and cobalt-manganese-bromine composite catalysts. Compared to virgin cobalt salt raw materials with stable sources, recycled cobalt materials typically exhibit complex origins, significant batch-to-batch variations, unstable impurity entrainment states, and inconsistent material morphologies. Even batches with similar cobalt content may show significant differences in acid consumption rates, exothermic intensity, gas release rhythm, redox response, and impurity leaching tendency. In current production processes, reaction parameters often rely on fixed formulas, manual experience, or small-scale raw material testing results. For example, the amount of acid added is estimated based on cobalt content, and the amount of redox agent added, reaction temperature, stirring intensity, and reaction time are set according to existing processes. Such methods fail to fully reflect the dynamic performance of the current batch of reactants in the initial stages of the actual reaction, easily leading to problems such as mismatched acid or redox agent additions, insufficient cobalt conversion, hydrogen peroxide decomposition losses, increased acidic tail gas release, impurities entering the product, or excessively intensified reaction conditions. Although introducing model predictions can assist in parameter setting to some extent, if the cobalt conversion rate is predicted directly based solely on the initial reaction response, the model may overestimate the effective cobalt conversion capacity due to factors such as temperature rise, rapid pH changes, ORP mutations, or enhanced gas release, resulting in predictions that do not conform to the boundaries of acid supply, redox agent supply, and the amount of cobalt to be converted.

[0003] To address the above issues, this application presents a reaction process parameter prediction and control method and system based on initial response constraints. Summary of the Invention

[0004] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a method and system for predicting and controlling reaction process parameters based on initial response constraints. This method acquires raw material detection information and initial reaction response information for the current batch of reactants, and generates multiple candidate reaction parameter combinations based on target reaction requirements. The information is then input into a reaction behavior prediction model to obtain predicted cobalt conversion rates for the candidate combinations, which are then corrected using stoichiometric constraints. When predicted values ​​exceed limits, the risk of side reactions is determined by combining the excess residual amount and initial reaction response information, thereby selecting the target reaction parameter combination. This application can improve the accuracy and stability of matching reaction parameters between different batches of recovered cobalt.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A reaction process parameter prediction and control method based on initial response constraints, the method comprising: Obtain the raw material detection information of the current batch of reactants, and obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions; Based on the raw material detection information, the initial reaction response information, and the target reaction requirements, multiple candidate reaction parameter combinations are generated. The raw material detection information, the initial reaction response information, and the combinations of candidate reaction parameters are input into the reaction behavior prediction model to obtain the predicted reaction results corresponding to each combination of candidate reaction parameters. The reaction behavior prediction model is established based on historical batch production data and the predicted reaction results are constrained by stoichiometry. Based on the predicted reaction results corresponding to each candidate reaction parameter combination, the target reaction parameter combination for the current batch of reactants is determined from the multiple candidate reaction parameter combinations.

[0006] The raw material testing information includes at least one of the following: raw material source, target product type, cobalt content, impurity content, black particle content, raw material form, and historical recycling process label; the initial reaction conditions include contacting the current batch of reactant material with acid of a preset concentration according to a preset sampling amount, and ensuring that at least one of the following is within a preset range: acid addition amount, reaction temperature, stirring speed, and response acquisition time.

[0007] The methods for obtaining the initial response information include: A sample of the current batch of reactants is obtained according to the preset sampling amount, and the acid solution is added to the sample under the initial reaction conditions; During the response acquisition period, at least one of the following information is acquired: pH change information, temperature change information, redox potential change information, gas release information, and stirring power change information of the sample. The acid consumption response information of the current batch of reactants is determined based on the pH change information, the exothermic response information of the current batch of reactants is determined based on the temperature change information, the redox response information of the current batch of reactants is determined based on the redox potential change information, the gas release response information of the current batch of reactants is determined based on the gas release information, and the mass transfer response information of the current batch of reactants is determined based on the stirring power change information. At least one of the acid consumption response information, exothermic response information, redox response information, gas release response information, and mass transfer response information is determined as the initial reaction response information.

[0008] The methods for generating the candidate reaction parameter combinations include: Based on the target reaction requirements and the cobalt content in the raw material detection information, determine the basic reaction measurement range corresponding to the current batch of reactants; The basic reaction measurement range is corrected based on the initial reaction response information to obtain the reaction parameter value range corresponding to the current batch of reactants; Within the range of reaction parameter values, multiple candidate reaction parameter combinations are generated according to preset combination rules. Each candidate reaction parameter combination includes at least two of the following: acid addition parameter, oxidizing / reducing agent addition parameter, reaction temperature parameter, stirring parameter, and reaction time parameter.

[0009] The corrections include: The range of acid addition amount is adjusted based on acid consumption response information; the range of oxidizing and reducing agent addition amount is adjusted based on redox response information; the range of reaction temperature is adjusted based on exothermic response information; the range of feeding rate is adjusted based on gas release response information; and the range of stirring speed is adjusted based on mass transfer response information.

[0010] The main architecture of the reaction behavior prediction model is a gated recurrent unit neural network, and the reaction behavior prediction model includes: The response sequence input layer is used to receive the response change sequence corresponding to the initial response information; The gated loop unit layer is used to extract the temporal features of the response change sequence to obtain the response temporal features that reflect the initial reaction trend of the current batch of reactants. The parameter input layer is used to receive the raw material detection information and the parameter input information corresponding to the candidate reaction parameter combination; The feature fusion layer is used to fuse the response time-series features with the parameter input information to obtain the fusion prediction features of the current batch of reactants under the corresponding candidate reaction parameter combinations; A fully connected prediction layer is used to output the initial cobalt conversion rate prediction value corresponding to the candidate reaction parameter combination based on the fused prediction features. The constraint correction unit is used to constrain and correct the initial cobalt conversion rate prediction value based on the cobalt content in the raw material detection information, the acid addition parameters and redox agent addition parameters in the candidate reaction parameter combination, and the preset stoichiometric relationship, so as to obtain the cobalt conversion rate prediction result corresponding to the candidate reaction parameter combination.

[0011] The constraint correction includes: Based on the cobalt content in the raw material testing information, determine the amount of cobalt to be converted for the current batch of reactants; Based on the acid addition parameters in the candidate reaction parameter combinations and the preset acid-cobalt reaction stoichiometric relationship, determine the amount of cobalt that can be matched with the acid in the candidate reaction parameter combinations. Based on the redox agent addition parameters in the candidate reaction parameter combinations and the preset redox stoichiometric relationship, determine the redox-matchable cobalt amount corresponding to the candidate reaction parameter combinations; Based on the amount of cobalt to be converted, the amount of cobalt that can be matched by the acid solution, and the amount of cobalt that can be matched by the redox reaction, determine the allowable range of cobalt conversion rate corresponding to the candidate reaction parameter combination; The initial cobalt conversion rate prediction is compared with the allowable range of cobalt conversion rate. When the initial cobalt conversion rate prediction is greater than the allowable range of cobalt conversion rate, the initial cobalt conversion rate prediction is corrected to the allowable range of cobalt conversion rate to obtain the cobalt conversion rate prediction result.

[0012] After obtaining the predicted cobalt conversion rate, the method further includes: When the initial cobalt conversion rate prediction value is greater than the upper limit of the allowable range of cobalt conversion rate, the excess residual amount corresponding to the candidate reaction parameter combination is determined based on the difference between the initial cobalt conversion rate prediction value and the upper limit value. Based on the excess residual amount and the initial reaction response information, the side reaction risk information corresponding to the candidate reaction parameter combination is determined, wherein the side reaction risk information includes at least one of the following: redox agent decomposition risk, impurity leaching risk, acidic tail gas release risk, and local exothermic abnormal risk.

[0013] The methods for determining the target reaction parameter combination include: Based on the cobalt conversion rate prediction results corresponding to each candidate reaction parameter combination, the candidate reaction parameter combinations whose cobalt conversion rate prediction results meet the preset conversion rate requirements are determined as optional reaction parameter combinations; Based on the target reaction requirements, candidate reaction parameter combinations that do not meet the product quality constraints are screened out from the optional reaction parameter combinations to obtain the remaining optional reaction parameter combinations; Based on the acid addition parameters, redox agent addition parameters, reaction temperature parameters, stirring parameters, and reaction time parameters corresponding to the remaining optional reaction parameter combinations, and in conjunction with the side reaction risk information, determine the parameter matching evaluation value for each remaining optional reaction parameter combination; The remaining optional reaction parameter combination with the highest parameter matching evaluation value is determined as the target reaction parameter combination for the current batch of reactants.

[0014] A reaction process parameter prediction and control system based on initial response constraints, the system comprising: The candidate parameter generation module is used to obtain the raw material detection information of the current batch of reactants, and to obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions, and to generate multiple candidate reaction parameter combinations based on the raw material detection information, the initial reaction response information and the target reaction requirements; The reaction behavior prediction module is used to input the raw material detection information, the initial reaction response information and the combinations of candidate reaction parameters into the reaction behavior prediction model to obtain the initial target reaction prediction result corresponding to each combination of candidate reaction parameters. The constraint and risk determination module is used to determine the allowable range of the target reaction corresponding to each candidate reaction parameter combination based on the raw material detection information and the reactant addition parameters corresponding to each candidate reaction parameter combination, based on the stoichiometric relationship of the target reaction, constrain the initial target reaction prediction result according to the allowable range of the target reaction, and determine the prediction part that exceeds the allowable range of the target reaction as the out-of-bounds residual amount, and determine the non-target reaction risk corresponding to each candidate reaction parameter combination according to the out-of-bounds residual amount and the initial reaction response information; The parameter decision control module is used to determine the target reaction parameter combination from the multiple candidate reaction parameter combinations based on the constrained target reaction prediction results and the non-target reaction risks, and to control the reaction process of the current batch of reactants according to the target reaction parameter combination.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention characterizes the reaction properties of the current batch of reactants by combining raw material detection information and initial reaction response information, and uses a reaction behavior prediction model combined with stoichiometric constraints to determine the target reaction parameter combination, which can avoid the problem that fixed formulas are difficult to adapt to batch differences. At the same time, by identifying the risk of side reactions through the excess residual amount, it is beneficial to reduce the risk of redox agent decomposition, impurity leaching and tail gas release, and improve the stability of cobalt conversion and the consistency of product quality. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 An exemplary application scenario diagram provided for an embodiment of this application; Figure 2 A schematic diagram of the reaction parameter matching platform provided in the embodiments of this application; Figure 3 This is a flowchart illustrating the reaction process parameter prediction and control method based on initial response constraints provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of phrases in various places in the specification does not necessarily refer to the same embodiment, nor are they independent or alternative embodiments mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0019] In the embodiments of this application, the reaction process parameter prediction and control method based on initial response constraints is applicable to reaction control scenarios where recycled cobalt materials are used as raw materials to prepare cobalt salts or cobalt-based catalysts. The recycled cobalt materials can originate from lithium battery recycling systems, cobalt-containing catalyst recycling systems, cobalt salt production by-product recycling systems, or other cobalt-containing intermediate material recycling systems. Their forms can include recycled cobalt carbonate, cobalt hydroxide, cobalt-containing waste, recycled cobalt-manganese compounds, cobalt-containing filter residue, or pretreated cobalt-containing intermediates. Although these materials can all be reused as cobalt sources in the preparation processes of cobalt acetate, cobalt bromide, cobalt-manganese-bromine composite catalysts, etc., due to differences in their upstream sources, recycling processes, impurity entrainment states, and material morphology, different batches of recycled cobalt materials with the same nominal cobalt content may exhibit completely different acid consumption rates, exothermic intensities, gas release rhythms, redox responses, and impurity leaching tendencies in actual reactions.

[0020] In conventional production processes, recycled cobalt materials typically undergo basic testing, including cobalt content, impurity content, and appearance. Operators then determine the acid addition amount, redox agent addition amount, heating conditions, stirring conditions, and holding time based on the target product type and experienced formulation. For raw materials with stable sources and minimal compositional fluctuations, this can meet production needs to a certain extent. However, in recycled cobalt material systems, the basic testing results cannot fully reflect the actual behavior of the material after entering the reaction system. For example, some batches of recycled cobalt material may show high cobalt content in tests, but due to coatings, inclusions, black particles, insoluble phases, or localized agglomeration, the amount of cobalt effectively participating in the reaction is not equivalent to the total cobalt content detected. Similarly, some batches may exhibit significant temperature rises or gas releases in the initial stages of acid addition, superficially resembling high reactivity. However, a strong response does not necessarily contribute entirely to effective cobalt conversion; it may also be accompanied by hydrogen peroxide decomposition, acidic tail gas release, dissolution of impurity metals, or enhanced localized side reactions.

[0021] Understandably, matching reaction parameters for cobalt recovery is not simply a matter of calculating the theoretical feed amount based on cobalt content, nor is it a simple matter of migrating historical formulas to the current batch. The real difficulty lies in the fact that the initial reaction performance of the current batch of reactants contains information about effective cobalt conversion, but may also be mixed with non-target reaction information such as side reactions, limited mass transfer, and impurity release. If the reaction is carried out solely based on a fixed formula, problems such as insufficient reaction, excessive acid, loss of oxidizing and reducing agents, increased tail gas load, or excessive product impurities are likely to occur. If the cobalt conversion rate is predicted solely based on the model according to the initial response, the model may over-amplify early strong responses such as temperature rise, pH drop, ORP mutation, or gas release, resulting in a cobalt conversion rate prediction higher than that supported by actual stoichiometric conditions.

[0022] Based on the above understanding, this application does not treat the reaction behavior prediction model as an independent black-box decision-making tool, but rather interprets the model's prediction results within the stoichiometric boundaries of the cobalt recovery reaction. In other words, the reaction behavior prediction model can predict the cobalt conversion trend of a batch of reactants under different reaction conditions based on the raw material detection information, initial reaction response information, and candidate reaction parameter combinations of the current batch of reactants; however, the predicted trend is still constrained by the cobalt content, acid addition amount, redox agent addition amount, and corresponding reaction stoichiometric relationships. This avoids the model outputting prediction results that do not conform to material conservation or reaction stoichiometric relationships based solely on the initial strong response, making the prediction results closer to the effective conversion range achievable by the actual reaction system.

[0023] It should be noted that the stoichiometric constraints described in this application are not merely for simple boundary truncation of predicted values. In cobalt recovery reaction systems, when the initial predicted cobalt conversion exceeds the stoichiometric range, the excess does not necessarily imply meaningless model error. Rather, it typically reflects strong acid consumption, redox, exothermic, or gas release drives in the current batch of material under initial reaction conditions. These drives may shift to non-target reactions if the acid or redox agent stoichiometric conditions are insufficient to support the corresponding cobalt conversion. In other words, the deviation between the model prediction and the stoichiometric boundary precisely reveals the potential side reaction risks in candidate reaction parameter combinations.

[0024] Therefore, in the implementation of this application, when the initial cobalt conversion rate prediction value output by the reaction behavior prediction model is higher than the allowable range of cobalt conversion rate, it can be corrected to the allowable range to obtain a cobalt conversion rate prediction result that conforms to stoichiometric logic. Simultaneously, the corrected excess portion can be used as an out-of-bounds residual amount to characterize the remaining reaction driving force not absorbed by the effective cobalt conversion reaction. The out-of-bounds residual amount can be combined with the initial reaction response information to further determine whether the current candidate reaction parameter combination has the risk of hydrogen peroxide decomposition, impurity leaching, acidic tail gas release, or local exothermic anomaly. Thus, this application does not simply pursue the maximum predicted cobalt conversion rate, but rather establishes a matching relationship between effective cobalt conversion, reactant stoichiometric constraints, and side reaction risks.

[0025] For example, under a certain combination of candidate reaction parameters, the model predicts a high cobalt conversion rate based on the rapid temperature rise and drastic changes in ORP in the initial stage. However, the allowable range of cobalt conversion rate determined by the acid addition parameters and redox agent addition parameters does not support the predicted value, indicating that the current strong response is not entirely derived from an effective cobalt conversion reaction. If the redox response information further indicates a rapid redox agent consumption trend, it can be considered that the candidate parameter combination carries the risk of hydrogen peroxide decomposition or by-oxidation reactions. If the gas release information indicates a significantly high gas release rate, it can be considered that the candidate parameter combination may cause an increase in the acidic tail gas release load. If the impurity content, black particle content, and exothermic response indicate abnormal local reaction intensity, it can be considered that the candidate parameter combination carries the risk of impurity dissolution or abnormal local exothermic reactions. The above risks cannot be identified in advance by simply detecting the cobalt content or calculating a fixed formula; they need to be judged by combining model predictions, stoichiometric constraints, and the initial reaction response.

[0026] Based on this, the method described in this application can be applied to scenarios involving the reaction of recycled cobalt carbonate with acetic acid to prepare cobalt acetate, to the preparation of cobalt salts from cobalt-containing waste under the combined action of acetic acid and hydrogen peroxide, and to the preparation of cobalt-manganese-bromine composite catalysts from recycled cobalt-manganese materials. The commonality among these scenarios is not that the target products are completely identical, but rather that they all exhibit significant batch-to-batch variations in recycled cobalt materials, the initial reaction response having indicative significance for subsequent parameter settings, the boundary constraints of acid or redox agent metering conditions on the effective conversion of cobalt, and the risk of side reactions affecting product quality and resource utilization. The basic principle of this application is as follows: First, the initial reaction response information is used to capture the actual reaction tendency of the current batch of reactants. Then, the reaction behavior prediction model is used to predict the cobalt conversion trend under different combinations of candidate reaction parameters. The prediction results are limited to the range that the actual reaction resources can support by stoichiometric constraints. For the predicted part that exceeds the range, it is not simply discarded, but identified as the remaining reaction driving force that may be transferred to the side reaction. This avoids selecting the target reaction parameter combination based solely on the corrected cobalt conversion rate. The target reaction parameter combination can simultaneously take into account effective cobalt conversion, reactant utilization and side reaction suppression, which is more suitable for the reaction control scenario of recycled cobalt materials with complex sources and large composition fluctuations.

[0027] refer to Figure 1 , Figure 1 This is an exemplary application scenario diagram provided for an embodiment of this application.

[0028] like Figure 1 As shown in the embodiments of this application, the application scenario can be set in the production connection stage before the recycled cobalt material enters the formal reaction system. Recycled cobalt materials from different sources or different batches can be temporarily stored in silos 1, 2 to N, respectively. The recycled cobalt material can be recycled cobalt carbonate, cobalt-containing waste, cobalt hydroxide, recycled cobalt-manganese compounds or other cobalt-containing intermediate materials.

[0029] It should be noted that, Figure 1 The multiple silos represent different batches or sources of recycled cobalt material, and are not intended to limit the storage of recycled cobalt material to silos. It can also be stored in bins, tanks, packaging bags, turnover containers or other containers suitable for turnover in chemical production sites.

[0030] In actual production, although recycled cobalt materials from different silos can all be used as cobalt sources in subsequent cobalt salt or cobalt-based catalyst preparation processes, differences in their upstream recovery sources, pretreatment methods, impurity entrainment states, and material forms can easily lead to problems such as insufficient reaction, ineffective consumption of redox agents, increased impurity leaching, or increased tail gas release load if only acid addition, heating, stirring, or redox agent addition is performed according to a fixed formula for the target product. Therefore, in the embodiments of this application, the current batch of reactants does not directly enter the production reaction system, but first establishes a data association with the raw material detection station and the initial reaction acquisition station to obtain input information that reflects the basic composition and early reaction performance of the current batch of materials.

[0031] Specifically, the raw material testing station can be used to obtain raw material testing information for the current batch of reactants, such as raw material source, target product type, cobalt content, impurity content, black particle content, raw material form, or historical recycling process labels. This information is primarily used to characterize the basic material properties of the current batch of reactants, ensuring that subsequent reaction parameter matching no longer relies entirely on manual experience or fixed formulations. The initial reaction acquisition station can be used to conduct small-scale reactions or collect initial response data on samples of the current batch of reactants under initial reaction conditions to obtain information such as pH changes, temperature changes, redox potential changes, gas release changes, or stirring power changes. Unlike simple static detection, the information obtained from the initial reaction acquisition station reflects the actual reaction tendency of the current batch of materials under acidic, temperature, stirring, or redox environments.

[0032] Data obtained from the raw material testing station and the initial reaction acquisition station can be sent to the reaction parameter matching platform. This platform can be an industrial control computer, server, DCS host system, or other control platform with data processing capabilities located on the production site; it does not require specific hardware architecture.

[0033] The production reaction system is used to execute the formal reaction based on the target reaction parameter combination output by the reaction parameter matching platform. The production reaction system may include one or more of the following: a reaction vessel, an acid metering device, an oxidizing / reducing agent dosing device, a stirring device, a temperature control device, a tail gas absorption device, and a filtration device.

[0034] refer to Figure 2 , Figure 2 This is a schematic diagram of the reaction parameter matching platform provided in an embodiment of this application.

[0035] like Figure 2 As shown, the reaction parameter matching platform may include a reaction scheme generation module and a reaction scheme selection module. It should be noted that... Figure 2The module division shown is only for illustrating the functional processing logic of the reaction parameter matching platform and does not limit each module to be implemented as an independent hardware unit. In actual deployment, the reaction scheme generation module and the reaction scheme selection module can be integrated into the same industrial control computer, server, DCS host system or edge computing device, or they can be implemented by different software functional units in a collaborative manner.

[0036] The reaction scheme generation module generates multiple candidate reaction parameter combinations based on the raw material detection information, initial reaction response information, and target reaction requirements of the current batch of reactants. A reaction scheme is not the final production formula that has been determined and implemented, but rather a set of candidate parameters that may meet the reaction requirements under the current batch of materials. When generating candidate reaction parameter combinations, the reaction scheme generation module does not simply exhaustively list all possible process parameters. Instead, it first determines the basic reaction measurement range based on the target reaction requirements and the cobalt content in the raw material detection information, and then modifies the basic reaction measurement range by combining the initial reaction response information. For example, when the initial reaction response information shows a strong acid consumption response in the current batch of reactants, the candidate range for acid addition parameters can be appropriately expanded or shifted upwards; when the redox response changes rapidly, the amount of redox agent added or the staged addition method can be constrained; when the exothermic response is significant, the reaction temperature range can be narrowed or the heating rate reduced; when the gas release response is strong, the feed rate can be limited and the candidate range for tail gas absorption-related parameters can be increased; when the mass transfer response is weak, the candidate range for stirring parameters or reaction time can be increased.

[0037] The reaction scheme selection module is used to determine the target reaction parameter combination from multiple candidate reaction parameter combinations output by the reaction scheme generation module. Specifically, the reaction scheme selection module can input the raw material detection information, initial reaction response information, and each candidate reaction parameter combination of the current batch of reactants into the reaction behavior prediction model to obtain the predicted reaction results corresponding to each candidate reaction parameter combination. The predicted reaction results may include at least the cobalt conversion rate prediction results, which are used to characterize the effective reaction degree that the current batch of reactants can achieve under the corresponding candidate reaction parameter combinations.

[0038] In some implementations, the reaction scheme selection module can further assess the predicted excess portion truncated by stoichiometry constraints. This excess portion can be understood as the remaining reaction driving force that cannot be effectively absorbed by the cobalt conversion reaction under the current candidate reaction parameter combination. When the remaining reaction driving force is large, it indicates that although the candidate reaction parameter combination may exhibit a high cobalt conversion rate after correction, it may still be accompanied by side reaction risks such as redox agent decomposition, impurity dissolution, acidic tail gas release, or abnormal local exothermic reactions.

[0039] Finally, the reaction scheme selection module can determine the target reaction parameter combination for the current batch of reactants if the predicted cobalt conversion rate meets the preset conversion rate requirements, conforms to the target product quality constraints, and the side reaction risk is within the allowable range. The target reaction parameter combination can then be sent to the production reaction system to guide the formal reaction process, such as acid addition, oxidizing and reducing agent addition, heating, stirring, and heat preservation.

[0040] Next, with reference to the accompanying drawings, the reaction process parameter prediction and control method based on initial response constraints provided in the embodiments of this application will be further elaborated. Figure 3 The methods shown include: S1: Obtain the raw material detection information of the current batch of reactants, and obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions; In this embodiment, the current batch of reactants can be understood as the cobalt-containing material to be reacted in the cobalt salt or cobalt-based catalyst preparation process. Specifically, it can be recycled cobalt carbonate, cobalt-containing waste, cobalt hydroxide, recycled cobalt-manganese compounds, cobalt-containing filter residue, or pre-treated cobalt-containing intermediates. Although all of the above materials can be used as cobalt sources, their source processes, impurity entrainment states, black particle content, powder or block morphology, and historical recycling paths differ. Therefore, under the same target product, different batches of recycled cobalt materials will not respond consistently to acid, oxidizing and reducing agents, temperature, and stirring conditions. This step first obtains raw material testing information to determine the basic composition of the current batch of materials and the basic material boundaries that can participate in the reaction, such as raw material source, target product type, cobalt content, impurity content, black particle content, raw material morphology, and historical recycling process labels. Those skilled in the art will understand that raw material testing information can be selected according to actual production conditions, as long as it can reflect the basic material properties of the current batch of reactants to a minimum. This application does not impose further limitations on specific testing instruments or the number of testing items.

[0041] S2: Generate multiple candidate reaction parameter combinations based on the raw material detection information, the initial reaction response information, and the target reaction requirements; In this embodiment, the target reaction requirements can be determined based on the actual preparation material, such as the cobalt content, product concentration, free acid range, impurity limits, and stability requirements corresponding to cobalt acetate, cobalt bromide, or cobalt-manganese-bromine composite catalysts. Based on the target reaction requirements and the cobalt content in the raw material testing information, the basic reaction measurement range corresponding to the current batch of reactants can be determined first. The basic reaction measurement range is mainly used to limit the initial selectable range of acid addition parameters and redox agent addition parameters, so that the combination of candidate reaction parameters does not deviate from the basic material relationship required for the formation of the target product.

[0042] S3: Input the raw material detection information, the initial reaction response information, and the combinations of candidate reaction parameters into the reaction behavior prediction model to obtain the predicted reaction results corresponding to each combination of candidate reaction parameters. The reaction behavior prediction model is established based on historical batch production data and the predicted reaction results are constrained by stoichiometry constraints. In this embodiment, the reaction behavior prediction model is used to predict the cobalt conversion of the current batch of reactants under different combinations of candidate reaction parameters. Considering that the initial reaction response information is usually a time-varying sequence, such as a pH change sequence, temperature change sequence, redox potential change sequence, gas release change sequence, or stirring power change sequence, the reaction behavior prediction model can adopt a gated recurrent unit neural network as its main architecture. The gated recurrent unit layer can extract the temporal features of the response change sequence to obtain the response temporal features reflecting the initial reaction trend of the current batch of reactants; subsequently, the response temporal features are fused with the raw material detection information and the parameter input information corresponding to the candidate reaction parameter combinations, and the predicted value of the initial cobalt conversion rate is output through a fully connected prediction layer.

[0043] S4: Based on the predicted reaction results corresponding to each candidate reaction parameter combination, determine the target reaction parameter combination for the current batch of reactants from the multiple candidate reaction parameter combinations; In this embodiment, the determination of the target reaction parameter combination is not simply a matter of selecting the candidate reaction parameter combination with the highest predicted cobalt conversion rate. For the cobalt recovery reaction, excessive acid addition, excessive redox agent addition, excessive heating conditions, or excessively long reaction time may improve the cobalt conversion rate predicted by the model, but may also lead to hydrogen peroxide decomposition, impurity leaching, acidic tail gas release, or excessively high free acid content in the product.

[0044] Before delving into the specific technical details of the steps, the embodiments of this application need to be emphasized again.

[0045] The cobalt content of recycled cobalt materials is not necessarily equivalent to the effective amount of cobalt that can be stably converted under the target reaction conditions. For cobalt-containing recycled materials from complex sources, cobalt may be distributed in different phases, particle morphologies, or entrained structures. Although some cobalt can be included in the total cobalt content in the test, its release rate, dissolution path, and conversion degree will be affected by the material surface condition, black particle entrainment, insoluble phase coating, and local mass transfer conditions under actual reaction conditions such as acid contact, addition of oxidizing and reducing agents, heating, and stirring. Therefore, if the amount of acid or oxidizing and reducing agent is derived solely from the total cobalt content, it is easy to obtain parameter results that seem to satisfy the stoichiometric relationship but are inconsistent with the actual reaction process.

[0046] In conventional production, a common approach is to apply empirical formulas based on raw material testing results, or to adjust the amount of acid, redox agent, and reaction time according to historical batch data. The reliability of this approach relies on the similar reaction behavior between different batches of recycled cobalt. However, the unique characteristic of recycled cobalt lies in the fact that the initial reaction stage often exhibits rapid differences. For example, the rate of pH change after acid addition, the magnitude of temperature rise, fluctuations in redox potential, the rhythm of gas release, and changes in stirring power can all reflect the actual reaction tendency of the current batch. It is important to note that not all of these initial reaction behaviors indicate effective cobalt conversion; some may originate from acid consumption side reactions, redox agent decomposition, leaching of impurity metals, localized exothermic reactions, or changes in gas release load. If these response sources are not distinguished, and the initial response intensity is equated solely with cobalt conversion capacity, subsequent parameter matching may easily deviate towards excessive addition, excessively high temperatures, or excessively long reaction times.

[0047] Based on the above understanding, the processing approach adopted in this embodiment is not simply to improve the accuracy of model predictions, but rather to interpret model predictions in relation to the stoichiometric relationship of reactants and the initial reaction behavior. The reaction behavior prediction model can learn the correspondence between raw material detection information, initial reaction response information, and reaction parameter combinations from historical batch production data, thereby providing a predicted cobalt conversion rate for the current batch under candidate reaction parameter combinations. However, the predicted value does not directly represent the final, reliable reaction result, because actual cobalt conversion cannot be separated from the amount of cobalt to be converted, the available acid supply capacity, and the available oxidizing and reducing agent supply capacity. In other words, the model prediction reflects the reaction trend of the current batch of materials under given conditions, while the stoichiometric relationship reflects the conversion boundary that reaction resources can support; both need to be considered simultaneously.

[0048] Furthermore, when the initial cobalt conversion rate prediction output by the model exceeds the range allowed by the stoichiometric relationship, this embodiment does not simply interpret it as invalid error. For cobalt recovery reactions, a predicted value exceeding the boundary usually indicates strong reaction-driven signs in the initial response, such as rapid acid consumption, significant exothermic reactions, drastic changes in redox potential, or enhanced gas release; however, due to limitations in the amount of acid added, the amount of redox agent added, or the amount of cobalt to be converted, these strong responses cannot all be converted into effective cobalt conversion contributions. The portion truncated by the stoichiometric boundary can be understood as residual reaction-driven force not absorbed by the main reaction. If this residual reaction-driven force continues to exist in the reaction system, it may manifest as hydrogen peroxide decomposition, increased release of acidic tail gas, enhanced impurity dissolution, or localized exothermic anomalies, rather than simply disappearing.

[0049] Therefore, in this embodiment, when constraining and correcting the initial cobalt conversion prediction, a cobalt conversion prediction result that conforms to the stoichiometric relationship is obtained, while retaining the out-of-bounds residual amount where the prediction value exceeds the allowable range. The significance of the out-of-bounds residual amount is not to further improve the cobalt conversion prediction value, but to provide a risk assessment basis for candidate reaction parameter combinations. For example, when the out-of-bounds residual amount is accompanied by a strong redox response, it indicates that the redox agent may undergo ineffective decomposition or be consumed by by-oxidation; when the out-of-bounds residual amount is accompanied by a strong gas release response, it indicates that the tail gas release load may be underestimated; when the out-of-bounds residual amount is accompanied by a high impurity content or black particle content, it indicates that local strong reaction conditions may drive impurities into the liquid phase. Through this interpretation, the model prediction value, stoichiometric boundary, and initial reaction response are no longer isolated data, but are used together to determine whether the candidate reaction parameter combination is suitable for the current batch of reactants.

[0050] Those skilled in the art will understand that the excess residual amount does not require a fixed physical unit form. It can be characterized by the difference between the initial predicted cobalt conversion rate and the upper limit of the allowable range of cobalt conversion rate, or by the normalized difference, the difference ratio, or the result weighted with the initial reaction response intensity. It only needs to reflect the degree to which the predicted reaction trend exceeds the stoichiometric boundary. This embodiment does not impose a unique limitation on the specific calculation form of the excess residual amount. The key is to continue using the information truncated by the constraint correction for side reaction risk assessment, rather than discarding it directly after numerical correction.

[0051] Through the above processing, this embodiment can avoid two common misjudgments: one is that candidate reaction parameter combinations are preferentially selected because of high model prediction values, but in reality, due to limitations in acid or redox agent dosage, they cannot achieve the corresponding effective conversion; the other is that candidate reaction parameter combinations, after constraint correction, show that the cobalt conversion rate meets the target, but the truncated reaction driving force may be transferred to the risk of side reactions. Therefore, the determination of the target reaction parameter combination no longer only pursues the maximum conversion rate, but forms a more stable matching relationship between effective cobalt conversion, reaction resource constraints, and side reaction risks, which is more suitable for cobalt recovery reaction scenarios with significant batch differences, complex initial responses, and high product quality requirements.

[0052] Next, we will further elaborate on the technical content of the method of this application regarding the initial reaction response information.

[0053] In one example, the raw material testing information includes at least one of the following: raw material source, target product type, cobalt content, impurity content, black particle content, raw material form, and historical recycling process label.

[0054] It is understood that the raw material testing information includes at least one of the aforementioned information, and does not mean that any single piece of raw material testing information can independently complete reaction parameter matching without the other information. Rather, it means that the types of information that can be used to characterize the basic material properties of the current batch of reactants may differ under different production scenarios. In other words, raw material testing information is used to provide batch-based constraints for the generation of subsequent candidate reaction parameter combinations and reaction behavior prediction. The specific items can be selected according to the incoming material management method, testing conditions, and target reaction requirements. For scenarios with complete incoming material testing conditions, information such as cobalt content, impurity content, black particle content, and raw material morphology can be obtained simultaneously. For scenarios where batch traceability has been established, the basic attribute range of the current batch of reactants can also be obtained based on the source of the raw materials or historical recycling process labels. For scenarios where the target product has been determined by the production task order, the target product type does not need to be collected again as raw material testing information, but rather as a target reaction requirement for subsequent processing.

[0055] It should be noted that this application's solution is designed for process scenarios where all materials participate in the reaction. "All materials participating in the reaction" means that the main cobalt-containing components, associated impurities, black particles, residual intermediates, and materials in different forms in the current batch of reactants can all affect the reaction process under conditions of acid, oxidizing / reducing agents, heating, and stirring. Therefore, raw material detection information is not only used to identify material names or sources, but also to describe the basic material state of the current batch of reactants before it enters the reaction as completely as possible. For production scenarios with detection capabilities, it is preferable to simultaneously obtain information such as raw material source, target product type, cobalt content, impurity content, black particle content, raw material form, and historical recycling process labels. This provides a more sufficient data foundation for determining the subsequent basic reaction measurement range, generating candidate reaction parameter combinations, and correcting stoichiometric constraints.

[0056] In one optional embodiment, the initial reaction conditions include contacting the current batch of reactants with an acid solution of a preset concentration according to a preset sampling amount, and ensuring that at least one of the following is within a preset range: acid addition amount, reaction temperature, stirring speed, and response acquisition time.

[0057] It should be noted that the preset sampling amount and preset concentration of acid solution can be determined by those skilled in the art based on the target product type, the detection results of cobalt content in the recovered cobalt material, historical batch reaction records, small-sample pre-experiment results, and production safety control requirements. The preset sampling amount does not need to be a fixed proportion to the formal production feed amount; it only needs to represent the basic material state of the current batch of reactants and be able to generate collectable pH changes, temperature changes, redox potential changes, or gas release changes during the initial reaction stage. The preset concentration of acid solution can be selected based on the type of acid solution used in the subsequent target reaction. For example, when the target product is cobalt acetate, an acetic acid solution corresponding to the acetic acid reaction system can be used; when the target product involves cobalt bromide or a cobalt-manganese-bromine composite catalyst, an acid solution matching the corresponding reaction system can be used. The acid concentration should be selected to stimulate the initial reaction response of the current batch of reactants without causing the sample to enter a complete reaction or violent side reaction state within a short period.

[0058] Similarly, the preset range can be understood as the operational boundary used to ensure that different batches of recycled cobalt produce an initial reaction response under relatively comparable conditions. The preset range can be determined based on the conventional reaction window corresponding to the target product, the initial response records of small samples from historical qualified batches, the effective acquisition range of the detection instrument, and the requirements for safe exothermic and venting control. Specifically, the operational range that can stably distinguish the response differences between different batches from historical qualified batches without causing violent exothermic or concentrated venting can be selected as the basic range, and appropriately widened or narrowed according to the current target product and acid system.

[0059] For example, for the initial reaction sampling of the cobalt-containing material and acetic acid system, the preset sampling amount can be 5g to 50g, and the preset acid concentration can be 5% to 30%. The amount of acid added can be determined according to 60% to 120% of the acid consumption corresponding to the theoretical amount of cobalt to be reacted in the sample. The reaction temperature can be controlled at 25℃ to 60℃, the stirring speed can be controlled at 200r / min to 600r / min, and the response sampling time can be 5min to 30min. If historical qualified batches can form stable and distinguishable pH, temperature, and gas release responses under the sampling conditions of 10g sample, 20% acetic acid, 40℃, 300r / min, and 10min, the above conditions can be used as the baseline conditions, and the amount of acid added can be allowed to fluctuate by 10% to 20% above and below the baseline amount, the temperature by 5℃ to 10℃ above and below, and the stirring speed by 50r / min to 100r / min above and below, to form the corresponding preset range.

[0060] In one example, the initial response information is obtained in the following ways: A sample of the current batch of reactants is obtained according to the preset sampling amount, and the acid solution is added to the sample under the initial reaction conditions; During the response acquisition period, at least one of the following information is acquired: pH change information, temperature change information, redox potential change information, gas release information, and stirring power change information of the sample. The acid consumption response information of the current batch of reactants is determined based on the pH change information, the exothermic response information of the current batch of reactants is determined based on the temperature change information, the redox response information of the current batch of reactants is determined based on the redox potential change information, the gas release response information of the current batch of reactants is determined based on the gas release information, and the mass transfer response information of the current batch of reactants is determined based on the stirring power change information. The acid consumption response information, exothermic response information, redox response information, gas release response information, and mass transfer response information are determined as the initial reaction response information.

[0061] Next, we will further elaborate on the technical content of the method of this application regarding the combination of candidate reaction parameters.

[0062] In one example, the candidate reaction parameter combination is generated in the following ways: S2.1: Determine the basic reaction measurement range corresponding to the current batch of reactants based on the target reaction requirements and the cobalt content in the raw material detection information; Specifically, the basic reaction measurement range can be understood as the pre-determined basic boundary of reaction parameters based on the required material composition of the target product and the calculable cobalt content in the current batch of reactants, before considering the initial reaction rate of the current batch of reactants. For processes that recover cobalt to prepare cobalt acetate, cobalt bromide, or cobalt-manganese-bromine composite catalysts, the amount of acid added, the amount of oxidizing and reducing agent added, and subsequent reaction temperature, stirring, and reaction time cannot be set without considering the target cobalt content in the target product and the actual cobalt content in the raw materials. If a fixed amount of acid is given directly according to an empirical formula, insufficient acid or incomplete reaction may occur when the cobalt content of the current batch is high, while excessive free acid and increased burden of subsequent neutralization or concentration may occur when the cobalt content of the current batch is low. Therefore, determining the basic reaction measurement range based on the target reaction requirements and cobalt content before generating candidate reaction parameter combinations ensures that the initial boundaries of the candidate parameters have a material basis, preventing the candidate parameter combinations from deviating from the basic reaction conditions required for the preparation of the target product from the outset.

[0063] In this embodiment, the target reaction requirements may include one or more of the following: target product type, target cobalt concentration, target free acid range, impurity control requirements, product stability requirements, and subsequent application evaluation requirements. The cobalt content in the raw material testing information can be obtained through conventional chemical analysis, titration analysis, spectroscopic analysis, or cobalt-containing material testing methods already used in the enterprise's production site. After obtaining the cobalt content, the amount of cobalt to be processed in the current batch can be determined by combining it with the planned feed amount of the current batch of reactants. Then, the corresponding acid type and reactant relationship can be selected according to the target product type. For example, when the target product is cobalt acetate, the basic range of acetic acid addition can be determined based on the salt formation relationship corresponding to the acetic acid system; when the target product is cobalt bromide or a cobalt-manganese-bromine composite catalyst, the basic range of acid addition or bromine-containing component addition can be determined based on the corresponding bromine-containing reaction system.

[0064] It should be noted that the basic reaction measurement range is not limited to a single value, but rather a certain range of fluctuation is set around the theoretical amount added. For example, the theoretical amount of acid added can be used as a central reference, and the range of acid added can be formed by fluctuating it up and down by 5% to 15%. For the amount of oxidizing and reducing agent added, the theoretical reference amount can be fluctuated up and down by 10% to 20% based on the effective utilization of oxidizing and reducing agents in historical qualified batches to form a basic range. The above values ​​are only examples for easy understanding and do not constitute a limitation on specific implementation.

[0065] S2.2: Correct the basic reaction measurement range based on the initial reaction response information to obtain the reaction parameter value range corresponding to the current batch of reactants; This can demonstrate the actual reaction behavior of the current batch of material after contact with acid, heating, stirring, or the addition of oxidizing and reducing agents. Recycled cobalt material differs from newly prepared cobalt salt raw materials with stable composition; its cobalt element may be present along with black particles, insoluble phases, inclusions, or different particle morphologies. The acid consumption rate, exothermic process, gas release rhythm, and redox potential changes after entering the reaction medium may all deviate from theoretical stoichiometric calculations. Therefore, if the basic reaction stoichiometric range is directly used as the candidate reaction parameter range, two types of deviations may occur: one is that the current batch of material has weak reactivity, and the reaction time, stirring intensity, or acid addition amount within the basic range is insufficient to support effective conversion; the other is that the initial response of the current batch of material is too strong, and if acid or oxidizing and reducing agents are added according to the upper limit of the basic range, side reactions, tail gas release, or impurity dissolution may easily be introduced.

[0066] In one example, the correction includes: correcting the acid addition range based on acid consumption response information, correcting the oxidizing agent addition range based on redox response information, correcting the reaction temperature range based on exothermic response information, correcting the feed rate range based on gas release response information, and correcting the stirring speed range based on mass transfer response information.

[0067] It should be noted that the above correction does not simply and mechanically map a certain type of initial reaction response information to a single reaction parameter. Instead, it uses the basic reaction stoichiometric range as the initial boundary, and combines this with the deviation exhibited by the current batch of reactants under initial reaction conditions to narrow, shift upwards, downwards, or adjust the parameter boundaries directly related to the reaction process in stages. The basic reaction stoichiometric range reflects the theoretical requirements between the target product and cobalt content, while the initial reaction response information reflects the dynamic behavior of the current batch of materials after actual contact with acid, heat, stirring, and redox environments. The resulting range of reaction parameter values, formed by combining the two, neither deviates from the basic material stoichiometric relationship nor ignores the actual reaction deviations caused by differences in the source, form, inclusion state, or activity of the current batch of reactants.

[0068] In this embodiment, a response benchmark range corresponding to the target product type and raw material category can be pre-established. This response benchmark range can be determined based on the data collected from historical qualified batches under the same initial reaction conditions, or based on the results of small-scale pre-experiments. For example, in the scenario of recovering cobalt carbonate to prepare cobalt acetate, multiple qualified batches can be sampled under preset sampling amounts, preset acid concentrations, preset temperatures, and preset stirring conditions, including pH decrease, temperature rise, ORP change, gas release per unit time, and stirring power change. The response range within a stable production state can be used as a reference. If the pH decrease of historical qualified batches within three minutes is typically between 0.3 and 0.6, while the current batch reaches above 0.9, it can be determined that the current batch has a strong acid consumption response; if the temperature rise of historical qualified batches typically does not exceed 2°C, while the current batch reaches above 4°C, it can be determined that the current batch has a strong exothermic response; if the gas release of historical qualified batches gradually increases and slowly decreases after acid addition, while the current batch shows a short-term concentrated release, it can be determined that the gas release response is strong. The above values ​​are only used to illustrate the method of establishing the response baseline range. Those skilled in the art can adjust them according to the type of acid, material type, sampling time and production control requirements.

[0069] Furthermore, when adjusting the acid addition range based on acid consumption response information, the pH decrease rate, the pH change per unit acid addition, and the initial acid consumption trend can be used as criteria. When the acid consumption response is higher than the baseline response range, it indicates that there are more acid-consuming components in the current batch of reactants, or factors such as carbonates, alkaline impurities, and soluble metal components that lead to rapid acid consumption. In this case, the lower limit of the acid addition range can be appropriately raised to prevent a significant acid shortage in the candidate reaction parameter combinations. Simultaneously, the upper limit of the acid addition range can be limited based on the target product's free acid requirements to avoid simply increasing the acid amount and causing excessively high free acid levels in the product. For example, if the theoretical reference range for acid determined based on cobalt content is 95% to 110% of the theoretical reference amount, and the current batch's acid consumption response is strong, the acid addition range can be adjusted to 100% to 115% of the theoretical reference amount. If the target product has strict requirements for free acid, the upper limit can remain unchanged, only the lower limit can be increased.

[0070] In this embodiment, when adjusting the range of oxidant addition based on redox response information, attention can be paid to the rate of ORP change, ORP stabilization time, response amplitude after oxidant addition, and whether the response shows an abnormal decline. When the ORP changes slowly and it is difficult to reach the redox state required for the target reaction, the range of oxidant addition can be appropriately increased or a segmented addition candidate method can be added. When the ORP changes too quickly, fluctuates greatly in a short period of time, or a violent response occurs after adding a small amount of oxidant, it indicates that the current batch of material may contain components that promote the rapid consumption or decomposition of the oxidant. In this case, it is not advisable to simply increase the total amount of oxidant, but rather to narrow the range of oxidant addition or adjust the candidate parameter from one-time addition to multi-segment low-rate addition. For example, if the basic oxidant addition range is 90% to 120% of the theoretical reference amount, and the current batch of ORP jumps rapidly in a short period of time with enhanced exothermic response, the range can be adjusted to 80% to 105% of the theoretical reference amount, and the segmented addition interval can be used as one of the candidate parameters to reduce ineffective decomposition caused by excessively high local concentrations.

[0071] Furthermore, when adjusting the reaction temperature range based on exothermic response information, the initial temperature rise amplitude, temperature rise rate, and temperature drop can be combined for judgment. If the current batch of reactants exhibits a strong exothermic response under initial reaction conditions, it indicates that continuing to use a higher temperature range in the formal reaction may accelerate local reactions, increase the decomposition of oxidizing and reducing agents, or enhance the dissolution of impurities. In this case, the upper limit of the reaction temperature range can be lowered, or the candidate temperature can be adjusted from the higher temperature range to the medium-low temperature range, and effective conversion can be achieved by extending the reaction time or enhancing stirring. If the current batch has a weak exothermic response, and the pH change, gas release, and ORP change are all lower than the corresponding baseline range, it may indicate that the material's reactivity is insufficient, and the reaction temperature range or reaction time range can be appropriately increased. For example, if the basic reaction temperature range is 70°C to 85°C, and the initial temperature rise of the current batch is significantly higher, it can be adjusted to 60°C to 75°C; if the initial temperature rise of the current batch is insufficient and the acid consumption response is slow, it can be maintained at 70°C to 85°C, and candidate combinations with longer holding times can be added.

[0072] In this embodiment, when adjusting the feeding rate range based on gas release response information, the gas release start time, peak occurrence time, peak duration, and gas release volume per unit time can be considered. For recovered cobalt carbonate or recovered cobalt materials containing carbonate impurities, gas release is prone to occur during acid addition. Excessive gas release can increase foaming, tail gas absorption load, and the risk of local entrainment. If the initial reaction data shows that the current batch's gas release peak is higher than the response baseline range, or if the release peak occurs too early and lasts for a long time, the feeding rate or acid addition rate limit in the formal reaction can be reduced, and a higher tail gas absorption capacity can be used as a candidate parameter. For example, if the basic feeding rate range is 3% to 6% of the planned feeding amount per minute, and the current batch's gas release response is strong, the feeding rate range can be adjusted to 1% to 3% per minute, while ensuring that the tail gas absorption water flow rate or absorbent circulation volume is in a higher candidate range. Through the above adjustments, excessively rapid gas release in the early stages of the formal reaction can be avoided, preventing concentrated tail gas treatment pressure.

[0073] Furthermore, when adjusting the stirring speed range based on mass transfer response information, variations in stirring power, sample dispersion, response lag, and gas release distribution can all be considered. If the stirring power fluctuates significantly in the initial reaction stage, or if pH and temperature changes show a marked lag, it indicates that the current batch of material may exhibit agglomeration, flake-like stacking, difficult-to-disperse lumps, or localized coating, and the reaction is strongly affected by mass transfer. In this case, the lower limit of the stirring speed range can be increased, or a combination of enhanced stirring and extended reaction time can be included in the candidate parameter range. If the stirring power changes smoothly and the response curve reaches stability quickly, the conventional stirring range can be maintained to avoid excessive stirring causing increased foaming, gas entrainment, or increased equipment energy consumption. For example, if the basic stirring speed range is 200 rpm to 350 rpm, and the current batch's mass transfer response shows insufficient sample dispersion, it can be adjusted to 300 rpm to 450 rpm. If the gas release response is also strong, it is better to avoid directly increasing the speed to an excessively high level and instead use a candidate range with a moderate stirring speed and a longer reaction time to balance mass transfer improvement and gas release control.

[0074] It should be further noted that the above-mentioned modifications can be implemented individually or in combination. Because there may be coupling relationships between acid consumption response, exothermic response, redox response, gas release response, and mass transfer response, the final determination of the reaction parameter range should not be based solely on information from a single response. For example, if both the acid consumption response and the gas release response are strong, the lower limit of the acid addition range can be appropriately increased, but the feed rate range needs to be reduced to avoid concentrated gas release caused by rapid acid addition; if both the redox response and the exothermic response are strong, the ranges of both the oxidizing agent addition range and the reaction temperature range can be narrowed to reduce the accumulation of side reactions; if the mass transfer response is weak but the gas release response is strong, the lower limit of the stirring speed can be increased, but the upper limit of the feed rate should be limited simultaneously.

[0075] S2.3: Within the range of the reaction parameter values, multiple candidate reaction parameter combinations are generated according to preset combination rules, wherein each candidate reaction parameter combination includes acid addition parameters, redox agent addition parameters, reaction temperature parameters, stirring parameters, and reaction time parameters; Specifically, after determining the range of reaction parameters, it is still necessary to generate candidate combinations of reaction parameters for model prediction and evaluation. It is important to note that the acid addition parameters, oxidant addition parameters, reaction temperature parameters, stirring parameters, and reaction time parameters are not independent of each other. When the acid addition amount is too high, the free acid, gas release, and acidity of the subsequent products may change simultaneously; when the oxidant addition amount is too high, a lower temperature or a slower addition pace may be required; at higher reaction temperatures, the reaction time can be appropriately shortened, but this will also increase the possibility of side reactions and tail gas release; with low stirring intensity, even if the total amount of acid and oxidant meets the requirements, insufficient mass transfer may lead to incomplete conversion.

[0076] In this embodiment, the preset combination rules can be implemented using a combination of stratified value selection and compatibility screening. Specifically, several representative value points can be set within the value ranges of acid addition parameters, redox agent addition parameters, reaction temperature parameters, stirring parameters, and reaction time parameters. Representative value points can include conservative values ​​near the lower limit of the range, intermediate values, and enhanced values ​​near the upper limit of the range. For example, the acid addition parameter can be set to three levels: low, medium, and high within the corrected range; the redox agent addition parameter can be set to three levels: low dose, normal dose, and segmented enhanced dose; the reaction temperature parameter can be set to three levels: mild temperature, normal temperature, and higher temperature; the stirring parameter can be set to two levels: normal stirring and enhanced stirring; and the reaction time parameter can be set to two levels: normal holding time and extended holding time. For scenarios where the number of combinations needs to be controlled, an orthogonal combination method can be used to select representative combinations from the above value points, so that different values ​​of each type of parameter can be covered, without having to exhaustively enumerate all values. For example, when the five types of parameters are set to three, three, three, two, and two levels respectively, the number of complete combinations is large. Twelve to twenty-four candidate reaction parameter combinations can be retained through orthogonal design or stratified sampling. When production experience has excluded certain high-risk ranges, the number of candidates can be controlled to eight to sixteen groups to facilitate rapid evaluation by the subsequent reaction behavior prediction model. The above numbers are merely examples, and this application does not limit the specific number of candidate reaction parameter combinations.

[0077] Furthermore, to generate multiple candidate reaction parameter combinations based on preset combination rules, one can refer to the factor-level combination concept in orthogonal experimental design and combine it with the process compatibility relationship in the cobalt material recycling reaction process.

[0078] Specifically, the acid addition parameter, oxidizing / reducing agent addition parameter, reaction temperature parameter, stirring parameter, and reaction time parameter can each be considered as a parameter factor, and each parameter factor can be divided into several parameter levels within the aforementioned reaction parameter value range. For the acid addition parameter, low acid level, medium acid level, and high acid level can be set according to the modified acid addition amount range, where the low acid level is close to the lower limit of the range, the medium acid level is close to the middle of the range, and the high acid level is close to the upper limit of the range; for the oxidizing / reducing agent addition parameter, low dosage level, normal dosage level, and segmented enhancement level can be set according to the oxidizing / reducing agent addition amount range and addition method, where the segmented enhancement level does not necessarily mean the highest total amount, but can also mean adding the oxidizing / reducing agent in multiple times at a low rate; for the reaction temperature parameter, mild temperature level, normal temperature level, and enhanced temperature level can be set; for the stirring parameter, normal stirring level and enhanced stirring level can be set; for the reaction time parameter, normal heat preservation level and extended heat preservation level can be set. The above parameter levels can be determined based on the upper and lower limits of the value range, the median value, the commonly used values ​​of historical qualified batches, and the equipment adjustment accuracy. It is sufficient to cover the low-intensity, medium-intensity, and enhanced reaction conditions that the current batch of reactants may use.

[0079] In this embodiment, after the parameter levels are divided, a basic candidate combination can be formed first. This basic candidate combination can consist of the intermediate levels of each parameter factor. Then, using this basic candidate combination as the center, extended combinations are generated according to single-factor perturbation and two-factor combined perturbation methods. Single-factor perturbation refers to adjusting the acid addition parameter, redox agent addition parameter, reaction temperature parameter, stirring parameter, or reaction time parameter to low or high levels while keeping other parameter factors at intermediate levels, thereby observing the impact of changes in individual reaction parameters on the cobalt conversion trend. Two-factor combined perturbation refers to adjusting parameters with process correlations in pairs, such as combining a higher acid addition level with a lower feed rate or stronger tail gas absorption conditions, combining a higher redox agent addition level with a lower reaction temperature or segmented addition method, combining a lower reaction temperature with a longer reaction time, and combining a lower stirring intensity with a longer reaction time.

[0080] Next, we will further elaborate on the technical content of the reaction behavior prediction model of the method in this application.

[0081] It is understood that the reaction behavior prediction model of this application is a prediction model used to characterize the "correspondence between the raw material state, initial reaction response, candidate reaction parameter combinations, and cobalt conversion results of the current batch of reactants". In other words, the reaction behavior prediction model is not equivalent to a simple empirical formula, nor is it merely a lookup of historical formulas. Instead, it learns the conversion trends of different recycled cobalt materials under different reaction conditions through historical batch production data, and combines the acid consumption, exothermic reaction, redox reaction, gas release, and mass transfer states exhibited by the current batch in the initial reaction stage to predict the cobalt conversion rate under various candidate reaction parameter combinations. Those skilled in the art will understand that any method that can receive raw material detection information, initial reaction response information, and candidate reaction parameter combinations, and output corresponding reaction prediction results, can be used as a specific implementation of the reaction behavior prediction model of this application. This application does not limit the model type to a single structure.

[0082] In optional embodiments, the reaction behavior prediction model can be implemented using a multilayer perceptron model, a backpropagation (BP) neural network model, a recurrent neural network model, a long short-term memory (LSTM) neural network model, a gated recurrent unit (ROU) neural network model, a temporal convolutional network model, or a time-series prediction model based on an attention mechanism. Among these, the BP neural network model and the multilayer perceptron model are suitable for handling static input parameters, facilitating nonlinear fitting of raw material detection information and candidate reaction parameter combinations; the RNU, LSTM, and GU models are suitable for handling initial reaction response sequences that change over time; the temporal convolutional network model is suitable for extracting local temporal features from response curves with fixed sampling lengths; and the time-series prediction model based on an attention mechanism is suitable for scenarios with large data volumes, long response sequences, and sufficient computational resources. Therefore, the reaction behavior prediction model described in this application can be selected based on the scale of historical batch data, sampling frequency, response sequence length, and on-site computational conditions.

[0083] In this embodiment, a gated recurrent unit neural network is used as the main architecture of the reaction behavior prediction model. This is because the initial reaction response information of the current batch of reactants is usually not an isolated value, but rather a process of change with a clear temporal sequence. For example, after acid addition, the pH does not reach a stable value instantaneously, but exhibits a rate of decrease, a magnitude of decrease, and a recovery process; temperature changes may also manifest as a rapid rise, a slow rise, or a short-term peak; redox potential may show abrupt changes, fluctuations, or delayed changes; gas release may manifest as concentrated release, continuous release, or delayed release. The value of this response information lies not only in its numerical value at a particular moment, but also in its trend over time. The gated recurrent unit neural network can continuously memorize and filter these short time series, enabling the model to extract the initial reaction trend of the current batch of reactants from the response change sequence, rather than relying solely on single-point detection values ​​for judgment.

[0084] Furthermore, compared to ordinary feedforward neural networks, gated recurrent unit neural networks can preserve the sequential relationships in the response sequence, making them more suitable for recognizing process-meaning reaction behaviors such as "rapid acid consumption followed by slow stabilization," "strong exothermic reaction followed by rapid decline," and "short-term ORP mutation followed by continuous shift." Compared to long short-term memory neural networks, gated recurrent unit neural networks have a relatively simple gating structure and fewer parameters, making them easier to train and deploy in production scenarios with limited historical batch data, limited on-site computing resources, or the need for rapid evaluation of candidate parameters. Compared to complex attention models, gated recurrent unit neural networks have lower requirements for sample size and computing resources, making them more suitable for chemical production scenarios such as cobalt recovery where batch data accumulates gradually and response sampling time is short. Therefore, considering temporal representation ability, training stability, and ease of engineering deployment, gated recurrent unit neural networks are more suitable as the model structure of this application.

[0085] It should be noted that the reaction behavior prediction model is not based solely on the model name, but is composed of a response sequence input layer, a gated recurrent unit layer, a parameter input layer, a feature fusion layer, a fully connected prediction layer, and a constraint correction unit. The gated recurrent unit layer is used to extract the temporal change features in the initial reaction response process, the parameter input layer is used to receive raw material detection information and candidate reaction parameter combinations, the fully connected prediction layer is used to output the initial cobalt conversion rate prediction value, and the constraint correction unit is used to correct the initial cobalt conversion rate prediction value according to the stoichiometric relationship.

[0086] In this embodiment, the response change sequence received by the response sequence input layer can be obtained by organizing the data collected under the initial reaction conditions. For example, within the response acquisition duration, one or more of pH, temperature, redox potential, gas release rate, and stirring power can be collected at fixed sampling intervals. For instance, the response acquisition duration can be set to 5 to 30 minutes, and the sampling interval can be set to 5 seconds, 10 seconds, 30 seconds, or 1 minute. For the initial stage of acid addition where data changes rapidly, a shorter sampling interval can be used; for materials with slower response changes, a longer sampling interval can be used. To facilitate model processing, response sequences from different batches can be uniformly organized to the same length. For example, the response data for the first 10 minutes after acid addition can be taken, forming 60 time points with a sampling point every 10 seconds. Each time point contains five types of channel data: pH change value, temperature change value, redox potential change value, gas release rate, and stirring power change value. If some channels are not collected at a certain site, null value marking, historical averages can be used to fill in the gaps, or only the collected channels can be used to build the corresponding model. This application does not impose a unique limitation on the number of response channels.

[0087] In this embodiment, the initial reaction response information collected can be preprocessed before entering the response sequence input layer. Specifically, this can include outlier removal, time alignment, sampling interval standardization, baseline subtraction, and normalization. Outlier removal can remove isolated outliers clearly caused by sensor jitter, bubble obstruction, or short-term electrical signal interference; time alignment can use the moment the acid first contacts the sample as the response starting point, making response sequences from different batches comparable; baseline subtraction converts pH, temperature, ORP, and stirring power into changes relative to the initial moment, allowing the model to focus more on response trends rather than absolute deviations in equipment or the environment; normalization can be performed based on the maximum, minimum, mean, or standard deviation in historical training data, ensuring that data of different dimensions are within a suitable numerical range for neural network training. For gas release information, the release amount per unit time, cumulative release amount, or change in release rate can be used as input, as long as it reflects the initial gas release behavior of the current batch of material.

[0088] In this embodiment, the parameter input information received by the parameter input layer can include two categories. The first category consists of raw material parameters obtained from raw material testing information, such as cobalt content, impurity content, black particle content, raw material form, raw material source, and historical recycling process labels. For continuous data such as cobalt content, impurity content, and black particle content, normalization processing can be used before input; for categorical data such as raw material source, raw material form, and historical recycling process labels, number coding, unique thermal coding, or embedded coding can be used. The second category consists of process parameters obtained from candidate reaction parameter combinations, such as acid addition amount, acid addition method, redox agent addition amount, redox agent dosing method, reaction temperature, stirring speed, and reaction time. For non-numerical parameters such as acid addition method and redox agent dosing method, methods such as one-time addition, segmented addition, and continuous dripping can be converted into category codes, or further broken down into numerical parameters such as the number of additions, the proportion of each addition, and the addition interval.

[0089] In this embodiment, the gated recurrent unit layer can extract temporal features from the response change sequence provided by the response sequence input layer. Exemplarily, the gated recurrent unit layer can be configured as one or two layers. When the response sequence is short and the number of training samples is small, a single gated recurrent unit layer can be used, with the number of hidden units set to 32 or 64. When the response sequence is long and there are many acquisition channels, two gated recurrent unit layers can be used, with the first layer having 64 hidden units and the second layer having 32 hidden units. To reduce the risk of overfitting, a dropout layer can be set after the gated recurrent unit layer, with a dropout ratio set to 0.1 to 0.3. The above network parameters illustrate an implementable configuration; those skilled in the art can adjust them according to the historical batch sample size, response sequence length, and computing resources.

[0090] In this embodiment, the response time-series features output by the gated recurrent unit layer can be fused with the parameter features output by the parameter input layer. The fusion method can be either concatenation or weighted concatenation. A more direct implementation is to concatenate the response time-series features output at the last time point of the gated recurrent unit layer with the raw material parameter features and candidate reaction parameter features to form a fused prediction feature. This fused prediction feature is then input into a fully connected prediction layer. The fully connected prediction layer can include one or two fully connected layers. For example, the first fully connected layer can be set to 64 neurons, and the second fully connected layer can be set to 32 neurons. Nonlinear activation processing can be applied after each layer to enhance the ability to express the nonlinear relationship of the reaction behavior of the recycled cobalt. The output layer can be set to one neuron to output the predicted initial cobalt conversion rate corresponding to the combination of candidate reaction parameters. If the output value is in percentage form, the output range can be limited to 0 to 100; if the output value is in normalized form, it can be converted to percentage form in subsequent processing.

[0091] In this embodiment, the historical batch production data required to train the reaction behavior prediction model can consist of historical raw material testing records, historical initial reaction response records, historical reaction parameter records, and historical product testing records. Each training sample can correspond to a historical batch or a historical small-sample test batch. For a training sample, its input data can include the raw material testing information of the historical batch, the initial reaction response sequence collected under the initial reaction conditions, and the actual combination of reaction parameters used; its output label can be the cobalt conversion rate after the historical batch reaction is completed. The cobalt conversion rate can be determined based on the amount of cobalt added before the reaction, the amount of cobalt in the liquid phase product after the reaction, the amount of residual cobalt in the filter residue, or the cobalt content result in the product testing report.

[0092] In this embodiment, to improve the coverage of candidate reaction parameter combinations by the training samples, historical batch production data can include formal production data and small-scale validation data. Formal production data reflects the reaction results under actual production conditions, while small-scale validation data can be used to supplement response changes under different acid addition amounts, redox agent addition amounts, temperatures, stirring, and reaction times. For example, for recycled cobalt materials from the same or similar sources, several small-scale reaction conditions can be set within a safe range, allowing the same type of raw material to correspond to multiple reaction parameter combinations and cobalt conversion results. The model trained in this way can learn the impact of reaction parameter changes on cobalt conversion trends, rather than simply memorizing a fixed production formula.

[0093] In this embodiment, the training of the reaction behavior prediction model can be performed using supervised learning. Historical batch samples can be divided into training, validation, and test sets. For example, 70% of the samples can be used as the training set, 15% as the validation set, and 15% as the test set; when the historical sample size is small, cross-validation can also be used. During training, the error between the predicted cobalt conversion rate and the actual cobalt conversion rate can be used as the training objective, allowing the model to gradually adjust the network parameters in the gated recurrent unit layer and the fully connected prediction layer. Optimization methods can include the Adam optimizer, RMSProp optimizer, or stochastic gradient descent. For example, the initial learning rate can be set to 0.001, the batch size can be set to 16, 32, or 64, the number of training epochs can be set to 100 to 300 epochs, and an early stopping mechanism can be set based on changes in the validation set error. When the validation set error no longer decreases in several consecutive training epochs, training can be stopped early to reduce the risk of overfitting.

[0094] In this embodiment, after the model training is completed, the model's prediction performance can be evaluated using test set samples. During evaluation, the deviation between the model's predicted cobalt conversion rate and the historical actual cobalt conversion rate can be compared. It can also be used to statistically analyze whether the prediction results can correctly distinguish between reaction parameter combinations that meet and do not meet the conversion rate requirements. For samples with large errors in the test set, their raw material source, impurity content, black particle content, initial response curve, and actual reaction records can be further examined to determine whether the error stems from abnormal sample records, insufficient coverage of material types by the training data, or abnormal operating conditions during the reaction process. After evaluation confirms that the model meets the usage requirements, the trained gated recurrent unit neural network can be deployed as a reaction behavior prediction model in the reaction parameter matching platform.

[0095] In this embodiment, the deployed reaction behavior prediction model can process each candidate reaction parameter combination for the current batch of reactants separately during prediction. Specifically, for the same current batch, the raw material detection information and initial reaction response sequence remain unchanged, and the candidate reaction parameter combinations are sequentially replaced in the input model. The model outputs an initial cobalt conversion rate prediction value for each candidate reaction parameter combination, thereby obtaining initial cobalt conversion rate prediction results corresponding to multiple candidate reaction parameter combinations. Since the acid addition amount, redox agent addition amount, temperature, stirring, and time differ in different candidate reaction parameter combinations, the initial cobalt conversion rate prediction value output by the model can reflect the conversion differences of the current batch of materials under different parameter conditions.

[0096] In this embodiment, the initial cobalt conversion rate prediction value output by the model still needs to be fed into a constraint correction unit. The constraint correction unit is not a general structure in neural networks, but rather a rule-based processing unit that executes according to a preset stoichiometric relationship. The constraint correction unit can determine the amount of cobalt to be converted in the current batch based on the cobalt content in the raw material detection information, determine the amount of cobalt that can be matched with the acid solution based on the acid addition parameters in the candidate reaction parameter combinations, determine the amount of cobalt that can be matched with the redox reaction based on the redox agent addition parameters, and thus determine the allowable range of cobalt conversion rate under the corresponding candidate reaction parameter combinations. When the initial cobalt conversion rate prediction value exceeds the allowable range, the constraint correction unit corrects it to the allowable range, obtaining the cobalt conversion rate prediction result.

[0097] In one example, the constraint correction includes: Based on the cobalt content in the raw material testing information, determine the amount of cobalt to be converted for the current batch of reactants; Based on the acid addition parameters in the candidate reaction parameter combinations and the preset acid-cobalt reaction stoichiometric relationship, determine the amount of cobalt that can be matched with the acid in the candidate reaction parameter combinations. Based on the redox agent addition parameters in the candidate reaction parameter combinations and the preset redox stoichiometric relationship, determine the redox-matchable cobalt amount corresponding to the candidate reaction parameter combinations; Based on the amount of cobalt to be converted, the amount of cobalt that can be matched by the acid solution, and the amount of cobalt that can be matched by the redox reaction, determine the allowable range of cobalt conversion rate corresponding to the candidate reaction parameter combination; The initial cobalt conversion rate prediction is compared with the allowable range of cobalt conversion rate. When the initial cobalt conversion rate prediction is greater than the allowable range of cobalt conversion rate, the initial cobalt conversion rate prediction is corrected to the allowable range of cobalt conversion rate to obtain the cobalt conversion rate prediction result.

[0098] Understandably, constraint correction is not simply truncating the neural network output; rather, it involves re-verifying the model's output reaction trend within the material boundaries provided by the current combination of candidate reaction parameters. In practice, the amount of cobalt to be converted can be determined first based on the current batch's feed amount and cobalt content. For example, if the current batch plans to feed 100 kg of material and the detected cobalt content is 20%, approximately 20 kg of cobalt can be used as the base cobalt amount for subsequent measurement verification. If the cobalt content in the test report is expressed as metallic cobalt mass fraction, cobalt salt equivalent content, or target product equivalent content, it can be converted to the same cobalt content standard used in the company's production. The amount of cobalt that can be matched with the acid solution can be determined according to the salt formation relationship corresponding to the target product. For example, when preparing divalent cobalt salts, the amount of acid radicals required to form the target salt type per mole of cobalt can be calculated based on the acid solution concentration, acid solution addition volume, and acid solution effective utilization rate to determine the maximum amount of cobalt that the current acid solution addition parameters can support for dissolution or salt formation. The acid solution effective utilization rate can be derived from the acid consumption records of historical qualified batches. For example, in the acetic acid system, 90% to 98% can be used as the effective utilization rate correction range, and a lower value can be used in scenarios where acid volatilization or side reactions are more obvious. The redox-matchable cobalt amount can be determined based on the ability of the redox agent to effectively participate in the adjustment of the target valence state. For example, the cobalt conversion scale that the redox agent can support can be determined based on the redox agent concentration, addition amount, historical effective utilization rate, and the initial redox response of the current batch. When historical records show that a certain type of recycled cobalt material has significant hydrogen peroxide decomposition, the effective utilization rate of the redox agent can be reduced according to the historical average of qualified batches, for example, from 80% to 60% to 70%, to avoid including the redox agent that has been consumed by side reactions in the effective conversion capacity.

[0099] When determining the allowable range of cobalt conversion rate, the amount of cobalt to be converted can be considered as the total resource boundary of the current batch, and the amount of cobalt that can be matched by the acid solution and the amount of cobalt that can be matched by the redox reaction can be considered as the reaction condition boundary that the combination of candidate reaction parameters can support. For reactions that require the joint participation of acid solution and redox agent, the actual supportable effective cobalt conversion amount should not exceed the most restrictive one of the three; for reactions that do not require the participation of redox agent, the amount of cobalt that can be matched by the redox reaction can be set as not constituting a limitation, or the allowable range can be determined solely based on the amount of cobalt that can be matched by the acid solution and the amount of cobalt to be converted. For example, if the amount of cobalt to be converted in the current batch is 20 kg, and the candidate acid solution addition parameters, after effective utilization conversion, can support 18 kg of cobalt to complete salt formation, and the candidate redox agent addition parameters, after effective utilization conversion, can support 16 kg of cobalt to complete the corresponding valence state adjustment, then the upper limit of the allowable range of cobalt conversion rate can be determined according to the ratio of 16 kg to 20 kg, that is, the upper limit is approximately 80%; if the redox agent is not a limiting factor, and the acid solution can support the conversion of 18 kg of cobalt, then the upper limit can be approximately 90%. The lower limit of the allowable range can typically be set to 0, or it can be set according to the minimum qualified conversion requirement of the target product. For example, if production management requires a cobalt conversion rate of no less than 92%, 92% can be used as the screening threshold for subsequent schemes, rather than as the natural lower limit of the stoichiometric allowable range. After this processing, even if the initial cobalt conversion rate output by the model is too high due to initial exothermic reaction, gas release, or rapid changes in ORP, it still needs to be verified by the actual supply capacity of the current candidate parameters.

[0100] Furthermore, when the initial cobalt conversion rate prediction exceeds the upper limit of the allowable range, it can be corrected to the upper limit or a conservative value close to the upper limit. For production lines with stable historical data and small detection errors, the upper limit of the allowable range can be directly used as the cobalt conversion rate prediction result. For production lines with fluctuating acid concentration, fluctuating redox agent utilization rate, or unstable cobalt recovery sources, a safety margin can be reserved based on the upper limit of the allowable range, for example, by lowering the upper limit of the allowable range by 1% to 5% as the correction result. The above safety margin can be determined based on the average deviation between the theoretical achievable conversion rate and the actual detected conversion rate in historical qualified batches. For example, if the theoretical achievable conversion rate of the most recent 30 qualified batches is generally about 2.5% higher than the measured conversion rate, then 2% to 3% can be used as a correction margin. The cobalt conversion rate prediction result obtained in this way is no longer just a direct extrapolation of the initial reaction trend by the neural network, but is simultaneously constrained by the total amount of cobalt source, acid supply capacity, and redox agent supply capacity. This can reduce the situation where the model overestimates the effective cobalt conversion due to short-term strong response, and also facilitates subsequent comparable evaluation among multiple candidate reaction parameter combinations.

[0101] In one example, this application embodiment also provides another method for predictive control of reaction process parameters based on initial response constraints, the method comprising: A1: Obtain the raw material detection information of the current batch of reactants, and obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions. Generate multiple candidate reaction parameter combinations based on the raw material detection information, the initial reaction response information and the target reaction requirements. A2: Input the raw material detection information, initial reaction response information and combinations of candidate reaction parameters into the reaction behavior prediction model to obtain the initial target reaction prediction results corresponding to each combination of candidate reaction parameters; A3: Based on the raw material detection information and the reactant addition parameters corresponding to each candidate reaction parameter combination, determine the target reaction allowable range corresponding to each candidate reaction parameter combination based on the stoichiometric relationship of the target reaction, constrain the initial target reaction prediction result according to the target reaction allowable range, and determine the prediction part exceeding the target reaction allowable range as the out-of-bounds residual amount, and determine the non-target reaction risk corresponding to each candidate reaction parameter combination according to the out-of-bounds residual amount and the initial reaction response information; A4: Based on the constrained target reaction prediction results and the non-target reaction risks, determine the target reaction parameter combination from the multiple candidate reaction parameter combinations, and control the reaction process of the current batch of reactants according to the target reaction parameter combination.

[0102] It is understood that, after obtaining the initial target reaction prediction result, this application does not directly select candidate reaction parameter combinations based on the prediction result's level, nor does it simply truncate the portion exceeding the stoichiometric boundary when the prediction result exceeds it. For each candidate reaction parameter combination, the target reaction range that the candidate reaction parameter combination can support at the material supply level is first determined using the actual amount of reactants that the candidate reaction parameter combination can provide, the amount of effective components that can participate in the target reaction in the current batch of reactants, and the corresponding stoichiometric relationship. Then, the initial target reaction prediction result given by the reaction behavior prediction model is compared with the allowable range of the target reaction. The prediction portion within the allowable range can continue to be used as the basis for predicting the degree of the target reaction, while the prediction portion exceeding the allowable range is retained as an out-of-bounds residual amount and further interpreted in conjunction with the initial reaction response information.

[0103] Those skilled in the art will understand that the initial target reaction prediction result and the target reaction allowable range have different technical meanings. The initial target reaction prediction result mainly reflects the statistical mapping relationship between the raw material characteristics, initial response characteristics, candidate reaction parameters, and target reaction result established by the reaction behavior prediction model based on historical production data; the target reaction allowable range, on the other hand, reflects the target reaction boundary that the current batch of materials and the corresponding candidate reaction parameters can support under the reactant supply relationship and stoichiometry. Therefore, when the initial target reaction prediction result exceeds the target reaction allowable range, it cannot be concluded solely from this that the reactants can actually obtain a target reaction result that exceeds the stoichiometric boundary.

[0104] Furthermore, the input to the reaction behavior prediction model includes initial reaction response information, and the reaction activity reflected in the initial reaction response information does not necessarily all originate from the target reaction. For example, in the acidification, salt formation, or redox reaction process for recovering cobalt-containing materials, rapid pH changes may originate from the acid consumption of the target cobalt-containing component, or may include reactions of associated alkaline impurities or other acid-consuming components; rapid temperature increases may accompany the target reaction, or may originate from the decomposition of redox agents, impurity reactions, or localized concentrated reactions; rapid changes in redox potential can reflect the adjustment of the target valence state, or may be caused by the ineffective consumption of redox agents; enhanced gas release may be related to the reaction of the target material, or may correspond to the concentrated release of non-target gases. Therefore, when the model obtains a high initial target reaction prediction result based on the above strong responses, it may contain response contributions that cannot be explained by the stoichiometric relationship of the target reaction.

[0105] Based on the above relationship, this embodiment retains the portion of the model prediction result that exceeds the allowable range of the target reaction, instead of discarding it directly after numerical correction. The out-of-bounds residual can be understood as: under the current combination of candidate reaction parameters, the portion of the target reaction tendency identified by the reaction behavior prediction model based on the initial response of the current batch that cannot be explained by the target reaction resource boundary that the candidate reaction parameters can actually provide.

[0106] It should be noted that the aforementioned out-of-bounds residual amount is a response characterization quantity used for subsequent control judgment, and does not mean that this value is directly equivalent in physical terms to the actual mass of acid consumed, redox agent consumed, actual heat released, or actual by-product generated by a certain non-target reaction. In other words, if the initial target reaction prediction result is 94%, and the upper limit of the target reaction allowable range is 86%, the resulting 8 percentage point out-of-bounds residual amount does not mean that 8% of the material will necessarily undergo a certain side reaction; this 8 percentage point mainly characterizes a portion of the reaction tendency identified by the model that cannot be absorbed by the current target reaction resource boundary. Subsequently, by analyzing the deviation of different types of initial reaction responses from the normal response benchmark, this unexplained portion of the predicted response is allocated and attributed, thereby obtaining non-target reaction risk information that can be used for candidate reaction parameter screening.

[0107] In one optional implementation, A3 may further include: A3.1: Based on the raw material detection information of the current batch of reactants and the reactant addition parameters in the candidate reaction parameter combinations, determine the target reaction resource boundary corresponding to the current batch of reactants; A3.2: Based on the target reaction resource boundary and the stoichiometric relationship corresponding to the target reaction, determine the allowable range of the target reaction corresponding to the current candidate reaction parameter combination; A3.3: Compare the initial target reaction prediction result with the allowable range of the target reaction. If the initial target reaction prediction result exceeds the allowable range of the target reaction, constrain the initial target reaction prediction result and obtain the corresponding out-of-bounds residual amount. A3.4: Based on the direction and degree of deviation of each response type from the corresponding response baseline in the initial response information, determine the response deviation information corresponding to each response type; A3.5: Based on the out-of-bounds residual amount and the response deviation information corresponding to each response type, perform residual response allocation on the out-of-bounds residual amount to determine the degree of residual correlation between the out-of-bounds residual amount and each non-target response type; A3.6: Determine the non-target reaction risk corresponding to the candidate reaction parameter combination based on the residual correlation degree corresponding to each non-target reaction type.

[0108] The target reaction resource boundary can be determined based on the limiting reactants necessary for the actual reaction system. Taking the reaction process of recovering cobalt-containing materials to prepare cobalt salts as an example, the amount of cobalt to be converted can be determined based on the cobalt content in the raw material testing information and the mass of the planned reaction material in the current batch; the target cobalt conversion amount that the acid can support can be determined based on the acid concentration and acid addition amount in the candidate reaction parameter combination and the corresponding acid-cobalt stoichiometric relationship; for target reactions that require the participation of redox agents to adjust the valence state, the target cobalt conversion amount that the redox agent can support can also be determined based on the redox agent concentration, addition amount, effective utilization degree, and redox stoichiometric relationship.

[0109] When multiple reaction resources constitute the necessary conditions for the target reaction, the upper limit of the allowable range of the target reaction can be determined based on the reaction resource with the highest limiting degree. For example, if the current batch of cobalt to be converted is 20 kg, and the amount of acid added in the candidate reaction parameter combination, after being converted according to the acid concentration, acid-cobalt stoichiometry, and historical effective utilization rate, can support 18 kg of cobalt to complete the target salt formation reaction, and the amount of redox agent added, after being converted according to the redox stoichiometry and effective utilization rate, can support 17.2 kg of cobalt to complete the target valence state adjustment, then the effective target conversion amount supported by the current candidate reaction parameter combination is limited by the redox agent conditions. The upper limit of its allowable range of the target reaction can be determined according to the ratio of 17.2 kg to 20 kg, which is 86%.

[0110] Next, the technical content of this application regarding the distribution of residual reaction response will be further elaborated.

[0111] In this embodiment, the initial reaction response information may include multiple types of information such as acid consumption response, exothermic response, redox response, gas release response, and mass transfer response. Different types of response information can be extracted from pH change sequences, temperature change sequences, redox potential change sequences, gas release change sequences, and stirring power change sequences, respectively.

[0112] For each response type, a corresponding response baseline can be pre-established. The response baseline is preferably established from historical qualified batches that have the same or similar target product, reactant type, initial feed ratio, and initial reaction conditions as the current batch. When the number of historical batches allows, the most recent 30 to 100 batches with stable production status, qualified product testing, and no recorded abnormal reaction events can be selected from historical production data, and a baseline can be established based on the response characteristics of these batches within the same initial data collection period.

[0113] For example, the rate of pH decrease per unit time can be statistically analyzed for acid consumption response, the rate of temperature rise per unit time can be statistically analyzed for exothermic response, the rate of ORP change, peak-to-valley difference or short-term fluctuation amplitude can be statistically analyzed for redox response, the rate of gas release per unit mass of material or the peak release value can be statistically analyzed for gas release response, and the change amplitude of stirring power relative to steady-state power can be statistically analyzed for mass transfer response.

[0114] Considering the potential for outliers in historical response data, the response benchmark can be determined using the median and mean absolute deviation. For example, for a specific response indicator, first calculate the median value of the indicator corresponding to multiple historical qualified batches, then calculate the absolute value of the difference between each historical indicator and the median value, and further obtain the median or average of these absolute values ​​as the response dispersion. For indicators with stable reaction processes and relatively concentrated distributions, the interval covered by the two response dispersion values ​​above and below the median value can be considered as the normal response interval; for indicators with relatively significant production fluctuations, the normal response interval can be formed using the three response dispersion values ​​above and below the median value.

[0115] For example, for the most recent 40 qualified batches, the median pH decrease rate within the first three minutes of the initial reaction was 0.14 pH / min, with a corresponding mean absolute deviation of 0.02 pH / min. Therefore, 0.18 pH / min can be used as the upper reference boundary for a strong acid consumption response. The median temperature rise rate was 0.55 °C / min, with a mean absolute deviation of 0.10 °C / min. Therefore, by adding twice the mean absolute deviation to the median value, 0.75 °C / min can be used as the upper reference boundary for a strong exothermic response. The median ORP change rate was 28 mV / min, with a mean absolute deviation of 6 mV / min. Therefore, 40 mV / min can be used as the upper reference boundary for a strong redox response.

[0116] Accordingly, if the median gas release rate per unit mass of material in historical qualified batches is 0.80 L / (min·kg) and the average absolute deviation is 0.15 L / (min·kg), then 1.10 L / (min·kg) can be used as the upper reference boundary for a strong gas release response; if the median fluctuation of stirring power relative to steady-state power is 4.5% and the average absolute deviation is 1.5%, then 7.5% can be used as the upper reference boundary for an abnormal mass transfer response.

[0117] The above response boundaries do not need to be fixed for all materials. Whenever there is a significant change in the raw material category, target reaction system, or initial reaction conditions, the corresponding historical batches can be recalled to re-establish the response baseline. If the number of historical qualified batches on the current production line is insufficient, for example, if only 10 or fewer batches of a new type of recycled material have been completed, an initial response baseline can be established based on small-scale experiments and gradually updated as more qualified batches are added.

[0118] Furthermore, after obtaining the response benchmarks corresponding to each response type, response deviation information can be formed based on the deviation direction and degree of the current batch response relative to the response benchmarks. The deviation direction is determined by calculating the response difference between the current batch response feature value and the corresponding response benchmark, and based on the sign of the response difference.

[0119] For response indicators that primarily use excessively strong responses to indicate the likelihood of non-target reactions, the positive deviation can be set to zero when the current response does not exceed the corresponding normal upper boundary; when the current response exceeds the corresponding normal upper boundary, the degree of response deviation is determined based on the ratio between the excess portion and the normal upper boundary.

[0120] For example, if the pH decrease rate of the current batch in the first three minutes is 0.24 pH / min, while the reference upper limit for acid consumption response is 0.18 pH / min, then the relative excess is approximately 0.333. If the temperature rise rate of the current batch is 1.05 °C / min, while the reference upper limit for exothermic response is 0.75 °C / min, then the corresponding deviation is approximately 0.40. If the ORP change rate of the current batch is 68 mV / min, while the reference upper limit for redox response is 40 mV / min, then the corresponding deviation is approximately 0.70. If the gas release rate per unit material of the current batch is 1.65 L / (min·kg), while the corresponding reference upper limit is 1.10 L / (min·kg), then the gas release response deviation is 0.50. If the stirring power fluctuation of the current batch is 11%, while the historical reference upper limit is 7.5%, then the mass transfer response deviation is approximately 0.467.

[0121] Therefore, the response deviation vector corresponding to the current candidate reaction parameter combination can be obtained, which can be expressed in the order of acid consumption response, exothermic response, redox response, gas release response, and mass transfer response as: 0.333, 0.400, 0.700, 0.500, and 0.467. For response indices with directional meaning, the response direction can also be retained. For example, rapid rise, rapid fall, and high-frequency fluctuation of ORP may correspond to different reaction behaviors, so ORP rise deviation, ORP fall deviation, and ORP fluctuation deviation can be formed respectively, instead of using only a single absolute value of ORP change. Similarly, continuous increase of stirring power and short-term high-frequency fluctuation of stirring power can reflect viscosity increase and local agglomeration or mass transfer inhomogeneity, respectively. By preserving the deviation direction, responses with different physical meanings can be uniformly compressed into a single value.

[0122] After obtaining the response deviation information corresponding to each response type, this embodiment further performs residual reaction response allocation on the out-of-bounds residual amount.

[0123] The residual response allocation does not involve evenly distributing the out-of-bounds residual amount to various response types. Instead, it determines the allocation ratio of the out-of-bounds residual amount across different response channels based on the degree of deviation of each response type from the historical normal response.

[0124] For example, we can first calculate the sum of the deviations of each positive response. Taking the data above as an example, the sum of the deviations of each response is approximately 2.400. Therefore, the residual allocation ratio corresponding to the acid consumption response is approximately 13.9%; the residual allocation ratio corresponding to the exothermic response is approximately 16.7%; the residual allocation ratio corresponding to the redox response is approximately 29.2%; the residual allocation ratio corresponding to the gas release response is approximately 20.8%; and the residual allocation ratio corresponding to the mass transfer response is approximately 19.5%. Taking the aforementioned 8.5 percentage point out-of-bounds residual amount as an example, after allocating this out-of-bounds residual amount according to the above ratios, we can obtain approximately 1.18 percentage points of acid consumption residual response, 1.42 percentage points of exothermic residual response, 2.48 percentage points of redox residual response, 1.77 percentage points of gas release residual response, and 1.66 percentage points of mass transfer residual response. Thus, the out-of-bounds residual amount, which originally only represented an 8.5 percentage point value, is transformed into a set of residual reaction responses with distinct response types. This set does not represent the actual conversion amount of the five independent chemical reactions, but rather the explanatory proportion of the out-of-bounds portion of the current model across the five abnormal response channels. In the example above, the residual amount corresponding to the redox response is the largest, indicating that a large part of the strong reaction trends that exceed the stoichiometric boundary identified by the current model occur simultaneously with the anomalous redox response; this can be further used to determine the possibility of redox agent decomposition or other non-target redox processes.

[0125] In other implementations, to avoid individual response metrics being affected by transient noise, the allocation ratio can be determined based on deviation information from multiple consecutive sampling periods. For example, the initial sampling period can be divided into three response phases: 0 to 1 minute, 1 to 3 minutes, and 3 to 5 minutes, and the degree of response deviation in each phase can be calculated. If a response produces a spike only at a single sampling point, while the two preceding and following phases are within the normal range, the allocation weight corresponding to that response can be reduced; if a response exceeds the response benchmark in two or more consecutive phases, the corresponding allocation ratio can be maintained or increased.

[0126] The continuity criterion can also be determined based on historical sampling noise. For example, if a single point of anomaly in a historically stable batch typically lasts no more than two sampling periods, while a true response anomaly typically lasts more than five sampling periods, then with a sampling interval of 10 seconds, a deviation lasting more than 30 seconds can be considered a valid continuous deviation criterion. If the current ORP has only one 10-second sampling point exceeding the baseline, while the remaining sampling points return to normal, this single point of anomaly does not need to be fully included in the redox residual response; if the ORP exceeds the baseline for 60 consecutive seconds, it can be included in the allocation according to the degree of complete deviation.

[0127] Next, the residual reaction responses are further attributed based on the correlation between each type of residual reaction response and different non-target reactions.

[0128] In this embodiment, a response correlation between response types and non-target reaction types can be established in advance. The non-target reaction types may include at least one of the following: decomposition of redox agents, dissolution of impurities, release of acidic tail gas, and local exothermic anomalies.

[0129] The aforementioned response correlation does not require manual specification and can be established based on historical production batches or small-sample validation data with clear anomaly records. For example, for batches historically confirmed to have significant decomposition of redox agents, the average deviation of acid consumption response, exothermic response, redox response, gas release response, and mass transfer response before and after the anomaly occurs can be statistically analyzed, and the corresponding correlation weights can be determined based on the ability of each type of response to distinguish between abnormal and normal batches.

[0130] For example, for 20 historical batches with clear records of redox agent decomposition, the anomaly distinguishing factors for the five types of responses were calculated respectively. The average distinguishing factors for acid consumption response, exothermic response, redox response, gas release response, and mass transfer response were 0.08, 0.34, 0.61, 0.27, and 0.07, respectively. The sum of the above distinguishing factors is 1.37. After normalizing each distinguishing factor, the association weights are approximately 0.06, 0.25, 0.45, 0.20, and 0.04.

[0131] Therefore, in one example, the redox decomposition type can assign a high weight to the redox residual response, a second-highest weight to the exothermic residual response and the gas release residual response, and a low weight to the acid consumption residual response and the mass transfer residual response.

[0132] Using the same method, the correlation weights between impurity leaching and each response type can be determined based on historical batches exceeding impurity limits. For example, the normalized correlation weights for acid consumption response, exothermic response, redox response, gas release response, and mass transfer response can be 0.35, 0.25, 0.15, 0.05, and 0.20, respectively.

[0133] For the risk of acidic exhaust gas release, the gas release response and acid consumption response usually have higher identification contributions and can be set to 0.50 and 0.30, respectively, while the remaining exothermic response, redox response and mass transfer response can be set to 0.10, 0.05 and 0.05, respectively.

[0134] For local exothermic anomalies, the corresponding correlation can be obtained based on historical local temperature anomalies, stirring power anomalies, or reactant agglomeration batches. For example, exothermic response, mass transfer response, and redox response have relatively high weights, at 0.45, 0.30, and 0.15, respectively, while acid consumption response and gas release response have a weight of 0.05.

[0135] The weights described above are only for illustrating one feasible method of determination. In practical applications, the associated weights are preferably obtained from historical abnormal batches of the corresponding production line or from pre-conducted small-scale control experiments. When a certain type of non-target reaction lacks sufficient historical abnormal samples, the initial weights can also be determined based on small-scale experiments and gradually updated as subsequent abnormality confirmation results are obtained.

[0136] Based on the aforementioned association weights, the residual response responses corresponding to each of the aforementioned response types can be weighted and combined to obtain the initial residual association quantity corresponding to each non-target response type.

[0137] Taking the aforementioned residual responses of acid consumption (1.18), exothermic reaction (1.42), redox reaction (2.48), gas release reaction (1.77), and mass transfer reaction (1.66) as examples, for the decomposition of redox agents, using 0.05, 0.25, 0.45, 0.20, and 0.05 as example correlation weights, the corresponding initial residual correlation amount is approximately 1.97.

[0138] For impurity dissolution, using correlation weights of 0.35, 0.25, 0.15, 0.05, and 0.20, an initial residual correlation of approximately 1.56 can be obtained.

[0139] For acidic exhaust gas release, using correlation weights of 0.30, 0.10, 0.05, 0.50, and 0.05, an initial residual correlation of approximately 1.59 can be obtained.

[0140] For localized exothermic anomalies, using correlation weights of 0.05, 0.45, 0.15, 0.05, and 0.30, an initial residual correlation of approximately 1.66 can be obtained.

[0141] Furthermore, the initial residual correlation values ​​corresponding to the four types of non-target reactions can be normalized to determine the degree of residual correlation between the out-of-bounds residual values ​​and each type of non-target reaction. The sum of the above four initial residual correlation values ​​is approximately 6.78, so the residual correlation value corresponding to the decomposition of the redox agent is approximately 29.1%, that of impurity dissolution is approximately 23.0%, that of acidic tail gas release is approximately 23.5%, and that of local exothermic anomaly is approximately 24.5%.

[0142] Therefore, the 8.5 percentage points of out-of-bounds residual amount can be further allocated to different non-target reaction types according to the degree of residual correlation, resulting in type residual amounts of approximately 2.47 percentage points, 1.96 percentage points, 2.00 percentage points, and 2.08 percentage points, respectively.

[0143] It should be understood that the aforementioned types of residual quantities are also characterization values ​​in the control analysis process. Their role is to indicate which one or more non-target reaction pathways the out-of-bounds residual information is more likely to be explained by, rather than being used to directly calculate the mass of the actual products of the corresponding non-target reaction. Using the above two-level allocation method, first assigning the out-of-bounds residual quantity to the response type, and then attributing the response type according to its association with different non-target reactions, can preserve the correspondence between the out-of-bounds prediction information and the original response information, avoiding the direct determination of a specific side reaction type based solely on a single out-of-bounds value.

[0144] After obtaining the residual correlation degree corresponding to each non-target reaction type, the risk degree corresponding to each non-target reaction type can be further determined.

[0145] Risk assessment boundaries can be determined based on the residual amounts of types corresponding to historical normal and abnormal batches. For non-target reaction types with clearly defined abnormal labels, the distribution of type residual amounts in normal and abnormal batches can be statistically analyzed, and the boundary value that can distinguish between the two types of batches can be selected as the risk assessment benchmark. For cases where the number of abnormal samples is small, a temporary risk boundary can be established using the median value of type residual amounts and its dispersion in historical qualified batches.

[0146] For example, regarding the risk of redox agent decomposition, the median value of the corresponding type of residual amount in historical qualified batches is 0.75 percentage points, the average absolute deviation is 0.35 percentage points, and most batches with obvious redox agent decomposition in history have a percentage point higher than 1.8 percentage points. Therefore, 1.8 percentage points can be used as the boundary for judging higher risk. Below 0.9 percentage points can be defined as the low-risk range, 0.9 to 1.8 percentage points as the medium-risk range, and more than 1.8 percentage points as the higher-risk range.

[0147] Regarding impurity leaching, if historical data shows that the residual amount of the type of product with no abnormal impurities is usually below 1.2 percentage points, while the residual amount of products with a significant increase in impurity leaching is usually above 2.2 percentage points, then 1.2 percentage points and 2.2 percentage points can be used as the dividing line between medium risk and high risk, respectively.

[0148] For acidic exhaust gas releases, a judgment benchmark can be established based on historical batches where the exhaust gas inlet concentration or exhaust gas load per unit time exceeds the normal operating range. For example, if the proportion of historical batches with a residual amount of the type exceeding 2.5 percentage points significantly increases, and the peak exhaust gas concentration exceeds the normal load of the absorption system, then 2.5 percentage points can be used as a higher risk boundary.

[0149] For localized exothermic anomalies, calibration can be performed based on historical temperature differences in the reaction system, peak values ​​of local temperature probes, or records of abnormal stirring power. For example, if the residual amount of a batch with a localized temperature difference exceeding 5°C or a reaction temperature change rate exceeding the process allowable value generally exceeds 1.6 percentage points, then 1.6 percentage points can be used as the higher risk boundary for that type.

[0150] Taking the aforementioned example, the residual amount of the type corresponding to the decomposition of redox agents is approximately 2.47 percentage points, which is higher than the higher risk threshold of 1.8 percentage points, and can be identified as a high risk of redox agent decomposition; the residual amount of the type corresponding to impurity dissolution is approximately 1.96 percentage points, which is between 1.2 and 2.2 percentage points, and can be identified as a medium risk of impurity dissolution; the acidic tail gas release corresponds to approximately 2.00 percentage points, which is below the higher risk threshold of 2.5 percentage points, and can be identified as a medium risk of acidic tail gas release; the local exothermic anomaly corresponds to approximately 2.08 percentage points, which is higher than the higher risk threshold of 1.6 percentage points, and can be identified as a high risk of local exothermic anomaly.

[0151] Furthermore, when a candidate reaction parameter combination corresponds to multiple non-target reaction types, the risks of each non-target reaction can be retained to form a risk vector, rather than immediately combining them into a single total risk value. For example, the non-target reaction risks corresponding to a certain candidate reaction parameter combination can be represented as a combination of redox agent decomposition risk, impurity leaching risk, acidic tail gas release risk, and localized exothermic anomaly risk. Therefore, in the subsequent parameter selection process, different risk emphases can be set for different target products or different production safety requirements.

[0152] For example, for production tasks with strict requirements on product impurities, the impact of impurity leaching risk in parameter evaluation can be increased; for production stages where the tail gas absorption system is close to its design load, the impact of acidic tail gas release risk can be increased; and for scenarios where the heat exchange capacity of reaction equipment is limited, the constraint priority of local exothermic anomaly risk can be increased. The corresponding risk weights can be determined based on the production task order, equipment design capacity, and the degree of impact of historical anomalies, without requiring all non-target reaction risks to adopt the same weight.

[0153] It is important to emphasize that the aforementioned proportional allocation of out-of-bounds residues based on the degree of deviation of each response type is a calculation method for the structured interpretation of residue prediction information, and does not imply that the actual reaction process corresponding to the current batch of reactants satisfies a linear relationship. In the actual reaction process, there are mutual influences among the amount of acid added, the amount of oxidizing and reducing agent added, the reaction temperature, the mass transfer conditions, and the material composition. Furthermore, the target reaction and non-target reactions such as oxidizing and reducing agent decomposition, impurity dissolution, gas release, and localized exothermic reactions may occur simultaneously. Therefore, the change in the same response index under different reaction conditions does not necessarily correspond to the same proportional change in the degree of reaction. The aforementioned proportional allocation is mainly used to establish a comparable initial mapping relationship between out-of-bounds residues and different response channels, that is, based on the degree of deviation of each response from the corresponding historical response benchmark, it characterizes the relative explanatory power of different response channels for the current out-of-bounds phenomenon. The acid consumption residual response, exothermic residual response, oxidizing and reducing residual response, gas release residual response, and mass transfer residual response obtained in this way are intermediate characterization quantities for subsequent non-target reaction attribution, rather than linear conversion results of the actual amounts occurring in the corresponding reaction process.

[0154] In an optional implementation, a nonlinear coupling correction can be used to adjust the residual correlation degree corresponding to each non-target reaction type. This nonlinear coupling correction characterizes the enhancing effect of multiple response anomalies on a specific non-target reaction when they co-occur. Specifically, based on historical qualified batches and historical abnormal batches with clear anomaly types, response co-occurrence combinations corresponding to each non-target reaction type and their corresponding coupling correction relationships can be determined. When multiple response deviation information constituting a certain response co-occurrence combination in the current batch simultaneously satisfies the corresponding anomaly deviation condition, the corresponding coupling correction amount is added to the initial residual correlation degree obtained based on a single response residual. For example, historical data shows that in redox agent decomposition batches, when the redox response deviation exceeds 0.5 and the exothermic response deviation exceeds 0.3, the proportion of abnormal redox agent consumption events increases from approximately 35% in cases of isolated ORP anomalies to approximately 70%. Therefore, a coupling correction coefficient can be determined based on these historical changes in occurrence rates. When the redox response deviation of the current batch is 0.70 and the exothermic response deviation is 0.40, and the amount of redox agent added in the candidate reaction parameter combination is within the top 20% of its allowable range, a coupling correction of 1.2 to 1.5 times can be applied based on the initial residual correlation corresponding to redox agent decomposition. The abnormal deviation conditions and coupling correction coefficients can be determined based on the statistical relationship between different response combinations and the actual occurrence of corresponding non-target reactions in historical abnormal batches, and updated with subsequent production batches' actual abnormal records. This ensures that residual response attribution reflects the characteristics of multivariate coupling and nonlinear amplification during the actual reaction process.

[0155] Next, we will further elaborate on the technical content of this application regarding the determination of the target reaction parameter combination.

[0156] In one example, the determination of the target reaction parameter combination includes: Based on the target reaction prediction results after constraints corresponding to each candidate reaction parameter combination, the candidate reaction parameter combinations that meet the target reaction requirements are determined as optional reaction parameter combinations; Based on the target reaction requirements, candidate reaction parameter combinations that do not meet the product quality constraints are screened out from the optional reaction parameter combinations to obtain the remaining optional reaction parameter combinations; Based on the constrained target reaction prediction results, reaction parameters, and non-target reaction risks corresponding to each remaining optional reaction parameter combination, determine the parameter matching evaluation results corresponding to each remaining optional reaction parameter combination. Based on the parameter matching evaluation results, the target reaction parameter combination is determined from the remaining optional reaction parameter combinations.

[0157] Specifically, the target reaction requirement can be determined based on the minimum effective reaction level of the production task or target product. Taking cobalt salt preparation as an example, if historical production and product quality control requirements indicate that a cobalt conversion rate below 92% easily leads to an increase in unreacted solid phase or a greater load on subsequent filtration, then 92% can be used as the minimum target reaction requirement for the current production task. This 92% is not a fixed setting for all reaction systems, but can be determined by the technical requirements of the target product and the minimum stable conversion level of historically qualified production batches.

[0158] Product quality constraints may include target product concentration range, free acid range, impurity content limits, product stability requirements, or other quality conditions that can be predicted by the current candidate reaction parameter combinations and reaction prediction results. Before performing parameter matching evaluation, it is preferable to delete candidate reaction parameter combinations that are clearly unable to meet the target product quality requirements, thereby avoiding the problem of offsetting product quality non-compliance with high-target reaction prediction results later.

[0159] For the remaining optional reaction parameter combinations, target reaction evaluation information, parameter margin information, and non-target reaction risk information can be determined separately.

[0160] The parameter margin information reflects the degree to which candidate response parameter combinations are close to the allowable boundary of the corresponding parameter. For example, for a parameter with a selectable range of 80 to 120, when the candidate value is 100, the parameter is located in the middle of the range and has a large margin from both the upper and lower boundaries; when the candidate value is 118, although it is still within the allowable range, it is close to the upper boundary. The position of the candidate parameter within the corresponding value range can be normalized, and the parameter margin can be determined based on the normalized distance of the candidate parameter to the nearest boundary.

[0161] For example, for a parameter with a normalization value of 0.5, its distance to the nearest boundary is 0.5, so the corresponding parameter margin can be determined as 1; for a parameter with a normalization value of 0.8, its distance to the nearest boundary is 0.2, so the corresponding parameter margin can be determined as 0.4; for parameters located at the boundary, their parameter margin is close to 0. When the candidate reaction parameter combination includes acid addition parameters, redox agent addition parameters, reaction temperature parameters, stirring parameters, and reaction time parameters, the margins of each parameter can be averaged to obtain the comprehensive parameter margin of the corresponding candidate reaction parameter combination.

[0162] For parameters with obvious unidirectional risks, such as the amount of acid added or the reaction temperature, the risk increases mainly when approaching the upper boundary. In such cases, it is possible to calculate only the remaining space from the high-risk boundary, without requiring a symmetrical evaluation of the upper and lower boundaries.

[0163] The parameter matching evaluation results can be determined jointly based on the constrained target response prediction results, the comprehensive parameter margin, and the risk of non-target responses. The optimal evaluation weights for each component are determined based on historical production verification results.

[0164] For example, 50 recent qualified batches with complete reaction data and several batches with clear anomaly records can be selected. Candidate reaction parameter combinations can be reconstructed for each batch, and different evaluation weights can be used for retrospective selection. The deviation between the selected parameter combinations and the actual qualified parameters, the actual target reaction results, and the occurrence of anomalies can be compared. Under the premise of meeting product quality constraints and safety limitations, a set of weights that achieves a higher success rate in matching historical parameters can be selected.

[0165] For example, after backtracking historical data, if the evaluation weights corresponding to the target response prediction result, parameter margin, and non-target response risk are determined to be 0.55, 0.20, and 0.25, respectively, then the constrained target response prediction result can be normalized to 0 to 1, the comprehensive parameter margin can be normalized to 0 to 1, and the risk of each type of non-target response risk can be combined to form a risk quantity between 0 and 1. Then, the target response prediction evaluation is multiplied by 0.55, the parameter margin is multiplied by 0.20, and the risk safety evaluation is multiplied by 0.25 to form the parameter matching evaluation result.

[0166] The overall risk level for non-target reactions can be determined based on the target product and equipment operating requirements. For example, risk weights can be assigned to the risks of redox agent decomposition, impurity leaching, acidic tail gas release, and localized exothermic anomalies. When all risks are equally important, the average value of each risk can be used; when any one high risk is unacceptable, the maximum value among all risks can be used as the overall risk level.

[0167] For example, there are three candidate reaction parameter combinations P1, P2 and P3 that have been screened by target reaction requirements and product quality constraints. Their target reaction prediction results after constraints are 94.2%, 93.3% and 92.5% respectively, all of which are higher than the minimum target reaction requirement of 92%.

[0168] The comprehensive parameter margins for P1, P2, and P3, calculated based on the degree to which each reaction parameter is far from the allowable range boundary, are 0.52, 0.78, and 0.81, respectively.

[0169] Further analysis of the residual reaction response allocation results corresponding to each candidate reaction parameter combination yielded comprehensive non-target reaction risks of 0.63, 0.28, and 0.17, respectively.

[0170] Using the example weights of 0.55, 0.20, and 0.25 mentioned above, the parameter matching evaluation result for P1 is approximately 0.715; the parameter matching evaluation result for P2 is approximately 0.849; and the parameter matching evaluation result for P3 is approximately 0.878.

[0171] Understandably, although P1 has the highest target response prediction result after constraints, its response parameters are closer to the parameter range boundary and it corresponds to a higher risk of non-target response. Therefore, it did not obtain the highest parameter matching evaluation result. Although the target response prediction result of P3 is slightly lower than that of P1 and P2, it still meets the target response requirements and has a high parameter margin and a low risk of non-target response. Therefore, P3 can be identified as the target response parameter combination.

[0172] Therefore, the selection of candidate reaction parameter combinations in this application is not equivalent to maximizing the target reaction prediction results, nor is it equivalent to simply selecting the combination with the lowest reactant addition amount or the lowest risk. Instead, it is based on meeting the target reaction requirements and product quality constraints, and jointly evaluating the effectiveness of the target reaction, the parameter operating margin, and the non-target reaction risk characterized by the out-of-bounds residual response.

[0173] Once the target reaction parameter combination is determined, the formal reaction process of the current batch of reactants can be controlled based on this combination. For example, the acid addition parameter in the target reaction parameter combination can be sent to the acid metering device, the oxidizing / reducing agent addition parameter can be sent to the corresponding dosing device, the reaction temperature parameter can be sent to the temperature control device, and the stirring parameter can be sent to the stirring drive control device. The duration of the corresponding reaction stage can be controlled based on the reaction time parameter.

[0174] For target reaction parameter combinations that include staged addition parameters, the addition amount and stage switching time for each stage can be determined according to the target reaction parameter combination. For parameter combinations that include changes in heating rate or stirring speed, corresponding timing control commands can be generated and sent to the production control system. Thus, parameter prediction and non-target reaction risk analysis can ultimately be applied to the actual reaction equipment for reactant addition, temperature regulation, stirring, and reaction time control, rather than just obtaining prediction results detached from the production process.

[0175] Compared to selecting parameters solely based on model predictions, the above approach reduces parameter overemphasis caused by the model interpreting a strong initial response as a high target reactivity. Compared to simply truncating the model output with stoichiometric limits, it retains the information on anomalous reactions in the current batch contained in the truncated portion. Especially when different batches of reactants differ in their sources, impurity states, particle morphology, and pretreatment processes, even if multiple candidate reaction parameter combinations yield similar target reaction predictions after stoichiometric constraints, the distribution of their excess residual amounts across different response pathways such as acid consumption, exothermic reactions, redox reactions, gas release, and mass transfer may still show significant differences. This allows for further differentiation of the non-target reaction risks corresponding to each candidate reaction parameter combination.

[0176] Therefore, the embodiments of this application form a target reaction explainable part and an out-of-bounds residual part between the reaction behavior prediction result and the stoichiometric boundary, and perform multi-channel allocation and non-target reaction attribution on the out-of-bounds residual part through the initial reaction response, so that the evaluation of candidate parameter combination is simultaneously constrained by the target reaction resource boundary and the actual reaction behavior of the current batch, thereby improving the adaptability of reaction parameter prediction control to batch differences and non-target reaction interference.

[0177] In yet another example, embodiments of this application provide a reaction process parameter prediction and control system based on initial response constraints, the system comprising: The candidate parameter generation module is used to obtain the raw material detection information of the current batch of reactants, and to obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions, and to generate multiple candidate reaction parameter combinations based on the raw material detection information, the initial reaction response information and the target reaction requirements; The reaction behavior prediction module is used to input the raw material detection information, the initial reaction response information and the combinations of candidate reaction parameters into the reaction behavior prediction model to obtain the initial target reaction prediction result corresponding to each combination of candidate reaction parameters. The constraint and risk determination module is used to determine the allowable range of the target reaction corresponding to each candidate reaction parameter combination based on the raw material detection information and the reactant addition parameters corresponding to each candidate reaction parameter combination, based on the stoichiometric relationship of the target reaction, constrain the initial target reaction prediction result according to the allowable range of the target reaction, and determine the prediction part that exceeds the allowable range of the target reaction as the out-of-bounds residual amount, and determine the non-target reaction risk corresponding to each candidate reaction parameter combination according to the out-of-bounds residual amount and the initial reaction response information; The parameter decision control module is used to determine the target reaction parameter combination from the multiple candidate reaction parameter combinations based on the constrained target reaction prediction results and the non-target reaction risks, and to control the reaction process of the current batch of reactants according to the target reaction parameter combination.

[0178] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A reaction process parameter prediction and control method based on initial response constraints, characterized in that, The method includes: Obtain the raw material detection information of the current batch of reactants, and obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions. Generate multiple candidate reaction parameter combinations based on the raw material detection information, the initial reaction response information and the target reaction requirements. The raw material detection information, initial reaction response information and combinations of candidate reaction parameters are input into the reaction behavior prediction model to obtain the initial target reaction prediction results corresponding to each combination of candidate reaction parameters. Based on the raw material detection information and the reactant addition parameters corresponding to each candidate reaction parameter combination, the allowable range of the target reaction corresponding to each candidate reaction parameter combination is determined based on the stoichiometric relationship of the target reaction. The initial target reaction prediction result is constrained according to the allowable range of the target reaction, and the prediction part exceeding the allowable range of the target reaction is determined as the out-of-bounds residual amount. The non-target reaction risk corresponding to each candidate reaction parameter combination is determined according to the out-of-bounds residual amount and the initial reaction response information. Based on the constrained target reaction prediction results and the non-target reaction risks, a target reaction parameter combination is determined from the multiple candidate reaction parameter combinations, and the reaction process of the current batch of reactants is controlled according to the target reaction parameter combination.

2. The method according to claim 1, characterized in that, The raw material testing information includes the source of the raw material, the type of the target product, the cobalt content, the impurity content, the black particle content, the raw material form, and the historical recycling process label; the initial reaction conditions include contacting the current batch of reactants with acid of a preset concentration according to the preset sampling amount, and ensuring that the amount of acid added, the reaction temperature, the stirring speed, and the response acquisition time are within preset ranges.

3. The method according to claim 2, characterized in that, The methods for obtaining the initial response information include: A sample of the current batch of reactants is obtained according to the preset sampling amount, and the acid solution is added to the sample under the initial reaction conditions; Within the specified response acquisition time, information on pH changes, temperature changes, redox potential changes, gas release, and stirring power changes of the sample are acquired. The acid consumption response information of the current batch of reactants is determined based on the pH change information, the exothermic response information of the current batch of reactants is determined based on the temperature change information, the redox response information of the current batch of reactants is determined based on the redox potential change information, the gas release response information of the current batch of reactants is determined based on the gas release information, and the mass transfer response information of the current batch of reactants is determined based on the stirring power change information. The acid consumption response information, exothermic response information, redox response information, gas release response information, and mass transfer response information are determined as the initial reaction response information.

4. The method according to claim 1, characterized in that, The methods for generating the candidate reaction parameter combinations include: Based on the target reaction requirements and the cobalt content in the raw material detection information, determine the basic reaction measurement range corresponding to the current batch of reactants; The basic reaction measurement range is corrected based on the initial reaction response information to obtain the reaction parameter value range corresponding to the current batch of reactants; Within the range of reaction parameter values, multiple candidate reaction parameter combinations are generated according to preset combination rules. Each candidate reaction parameter combination includes acid addition parameters, oxidizing and reducing agent addition parameters, reaction temperature parameters, stirring parameters, and reaction time parameters.

5. The method according to claim 4, characterized in that, The corrections include: The range of acid addition amount is adjusted based on acid consumption response information; the range of oxidizing and reducing agent addition amount is adjusted based on redox response information; the range of reaction temperature is adjusted based on exothermic response information; the range of feeding rate is adjusted based on gas release response information; and the range of stirring speed is adjusted based on mass transfer response information.

6. The method according to claim 1, characterized in that, The main architecture of the reaction behavior prediction model is a gated recurrent unit neural network, and the reaction behavior prediction model includes: The response sequence input layer is used to receive the response change sequence corresponding to the initial response information; The gated recurrent unit layer is used to extract the temporal features of the response change sequence to obtain the response temporal features that reflect the initial reaction trend of the current batch of reactants. The parameter input layer is used to receive raw material detection information and parameter input information corresponding to candidate reaction parameter combinations; The feature fusion layer is used to fuse response time-series features with parameter input information to obtain the fused prediction features of the current batch of reactants under the corresponding candidate reaction parameter combinations. The fully connected prediction layer is used to output the initial cobalt conversion rate prediction value corresponding to the candidate reaction parameter combination based on the fused prediction features. The constraint correction unit is used to constrain and correct the initial cobalt conversion rate prediction value based on the cobalt content in the raw material detection information, the acid addition parameters and redox agent addition parameters in the candidate reaction parameter combination, and the preset stoichiometric relationship, so as to obtain the cobalt conversion rate prediction result corresponding to the candidate reaction parameter combination.

7. The method according to claim 6, characterized in that, The constraint correction includes: Based on the cobalt content in the raw material testing information, determine the amount of cobalt to be converted for the current batch of reactants; Based on the acid addition parameters in the candidate reaction parameter combinations and the preset acid-cobalt reaction stoichiometric relationship, determine the amount of cobalt that can be matched with the acid in the candidate reaction parameter combinations. Based on the redox agent addition parameters in the candidate reaction parameter combinations and the preset redox stoichiometric relationship, determine the redox-matchable cobalt amount corresponding to the candidate reaction parameter combinations; Based on the amount of cobalt to be converted, the amount of cobalt that can be matched by the acid solution, and the amount of cobalt that can be matched by the redox reaction, determine the allowable range of cobalt conversion rate corresponding to the candidate reaction parameter combination; The initial cobalt conversion rate prediction is compared with the allowable range of cobalt conversion rate. When the initial target reaction prediction result exceeds the upper limit of the allowable range of the target reaction, the initial cobalt conversion rate prediction is corrected to the allowable range of cobalt conversion rate to obtain the cobalt conversion rate prediction result.

8. The method according to claim 7, characterized in that, The methods for determining the risk of non-target reactions include: Based on the direction and degree of deviation of each response type from the corresponding response baseline in the initial response information, determine the response deviation information corresponding to each response type; Based on the out-of-bounds residual amount and the response deviation information corresponding to each response type, the degree of residual correlation between the out-of-bounds residual amount and each non-target reaction type is determined, wherein the non-target reaction type includes at least one of redox agent decomposition, impurity dissolution, acidic tail gas release, and local exothermic anomaly. Based on the residual correlation degree corresponding to each non-target reaction type, the non-target reaction risk corresponding to the candidate reaction parameter combination is determined.

9. The method according to claim 8, characterized in that, The methods for determining the target reaction parameter combination include: Based on the target reaction prediction results after constraints corresponding to each candidate reaction parameter combination, the candidate reaction parameter combinations that meet the target reaction requirements are determined as optional reaction parameter combinations; Based on the target reaction requirements, candidate reaction parameter combinations that do not meet the product quality constraints are screened out from the optional reaction parameter combinations to obtain the remaining optional reaction parameter combinations; Based on the constrained target reaction prediction results, reaction parameters, and non-target reaction risks corresponding to each remaining optional reaction parameter combination, determine the parameter matching evaluation results corresponding to each remaining optional reaction parameter combination. Based on the parameter matching evaluation results, the target reaction parameter combination is determined from the remaining optional reaction parameter combinations.

10. A reaction process parameter prediction and control system based on initial response constraints, implemented based on the reaction process parameter prediction and control method based on initial response constraints as described in any one of claims 1-9, characterized in that, The system includes: The candidate parameter generation module is used to obtain the raw material detection information of the current batch of reactants, and to obtain the initial reaction response information of the current batch of reactants under the initial reaction conditions, and to generate multiple candidate reaction parameter combinations based on the raw material detection information, the initial reaction response information and the target reaction requirements; The reaction behavior prediction module is used to input the raw material detection information, the initial reaction response information and the combinations of candidate reaction parameters into the reaction behavior prediction model to obtain the initial target reaction prediction result corresponding to each combination of candidate reaction parameters. The constraint and risk determination module is used to determine the allowable range of the target reaction corresponding to each candidate reaction parameter combination based on the raw material detection information and the reactant addition parameters corresponding to each candidate reaction parameter combination, based on the stoichiometric relationship of the target reaction, constrain the initial target reaction prediction result according to the allowable range of the target reaction, and determine the prediction part that exceeds the allowable range of the target reaction as the out-of-bounds residual amount, and determine the non-target reaction risk corresponding to each candidate reaction parameter combination according to the out-of-bounds residual amount and the initial reaction response information; The parameter decision control module is used to determine the target reaction parameter combination from the multiple candidate reaction parameter combinations based on the constrained target reaction prediction results and the non-target reaction risks, and to control the reaction process of the current batch of reactants according to the target reaction parameter combination.

Images