Machine learning-based kitchen garbage workshop odor concentration dynamic prediction and agent addition intelligent optimization system and method

Abstract

The invention discloses a kitchen garbage workshop odor concentration dynamic prediction and agent adding intelligent optimization system and method based on machine learning, and the system consists of a data collection module, a data preprocessing module, an odor concentration prediction module, an over-standard prediction module, a control execution module, a closed-loop updating module and the like. Real-time monitoring and key feature extraction of multi-component odor pollutants such as ammonia gas, hydrogen sulfide and volatile fatty acid can be realized. According to the method, a key pollutant set is dynamically determined through working condition recognition and feature importance analysis, an odor concentration multi-step prediction model integrating a mechanism model and a deep learning model is constructed, and the agent adding concentration, proportion and spraying duration are optimized in combination with a genetic algorithm and reinforcement learning. And meanwhile, the spraying rhythm preposed control is realized by utilizing the medicament action time efficiency. The system can automatically update the feature library and model parameters according to monitoring feedback, the odor concentration and the standard exceeding duration can be remarkably reduced, the agent consumption is reduced, and the intelligence and the stability of kitchen garbage workshop odor treatment are improved.

Description

Technical Field

[0001] This invention relates to the field of environmental governance technology, and more specifically, to a machine learning-based intelligent optimization system and method for dynamic prediction of odor concentration and intelligent optimization of reagent dosing in kitchen waste processing plants. Background Technology

[0002] Food waste treatment workshops continuously release various odorous pollutants, including ammonia, hydrogen sulfide, methanethiol, dimethyl disulfide, acetic acid, propionic acid, and butyric acid, during crushing, pressing, pretreatment, and temporary storage. These pollutants have multiple sources, fluctuate greatly, and have complex compositions, exhibiting significant non-steady-state characteristics. Their emissions are influenced by multiple factors, including the batch of feed, material moisture content, fermentation stage, temperature and humidity changes, and workshop operations. Existing odor control methods mostly rely on fixed-frequency or fixed-concentration chemical spraying, which is difficult to adapt to rapid changes in operating conditions. This often results in problems such as insufficient dosage leading to excessive odor, and excessive dosage causing chemical waste. The treatment effect is unstable, and the operating cost is high.

[0003] Traditional odor control systems typically rely on monitoring data from a limited number of pollutants for simple assessments. They cannot capture the coupling effects between dozens of odor-causing substances in real time, nor can they dynamically identify the changing trends of dominant odor factors under different pollution scenarios. Due to the lack of a holistic approach encompassing "multi-pollutant feature screening—odor trend prediction—dosing strategy optimization," traditional control models often employ experience-based or fixed solutions. This results in significant problems such as delayed response, high risk of rebound, and inability to plan reagent dosing schedules in advance. Furthermore, it is difficult to maintain continuous compliance in complex and unstable environments such as kitchen waste processing plants.

[0004] With the development of online monitoring technology, a large amount of real-time monitoring data has become available. However, how to quickly identify key pollutants from complex gas compositions and environmental parameters, accurately predict future odor trends, and intelligently optimize reagent dosing strategies based on the prediction results has become a common technical challenge in the field of odor control. Existing methods lack dynamic feature screening capabilities based on machine learning, making it difficult to build adaptive models applicable to multiple scenarios. Furthermore, they lack a multi-step prediction framework that combines mechanistic models and deep learning models, and also lack intelligent optimization methods that can achieve "proactive dosing" and "pre-intervention" in the odor control timeline.

[0005] Therefore, there is an urgent need for a clinker-free ecological cement mix design method based on the theory of closest particle packing to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems mentioned in the background section, and to provide a machine learning-based intelligent optimization system for dynamic prediction of odor concentration and intelligent dosing of chemicals in kitchen waste processing plants, comprising: The data acquisition module is used to collect real-time data including environmental parameters such as the concentration of various odor pollutants, air volume, temperature, humidity, and air pressure, and upload them to the central database. The data preprocessing module is used to remove outliers, fill in missing values, standardize, detrend, and smooth signals from the raw monitoring data in order to reduce the impact of sensor drift and random noise. The dynamic pollutant screening module is used to calculate the contribution of each pollutant to the odor concentration or odor level based on the machine learning feature importance ranking method, sort the pollutants and dynamically screen the top 10-20 key pollutants, and retrain and update the pollutant feature set according to a preset period or when the cumulative number of new samples reaches a threshold. The odor concentration prediction module is used to construct a time series prediction model based on the concentration of the key pollutants and environmental variables such as air volume, temperature, humidity, and air pressure, to predict the odor concentration and its changing trend in multiple future periods, and output the prediction results for reagent dosing decisions. The multi-objective prediction and optimization module is used to jointly predict the odor concentration, agent concentration and ratio, spraying time and the time of exceeding the standard after addition, based on the odor concentration prediction results. It also uses a multi-objective optimization algorithm to comprehensively determine the agent type, addition concentration, ratio and spraying time to generate the optimal agent addition scheme. The exceedance prediction module is used to predict the time point when the odor concentration will exceed the limit again after the pesticide has been applied, based on the pesticide application conditions and time series models, so as to trigger the next round of spraying strategy in advance. The control module is used to send control commands to the spraying system and execute corresponding pesticide dosing operations based on the dosing scheme output by the multi-target prediction and optimization module and the time information output by the over-standard prediction module. The closed-loop update module is used to update and iteratively train the pollutant set, odor prediction model parameters, and reagent dosing optimization strategy based on the actual monitoring feedback data after reagent dosing, so that the system has long-term self-adaptation and self-evolution capabilities.

[0007] As a preferred embodiment of the present invention, the dynamic pollutant screening module further includes: The first classification model is used to classify workshop operation data into working conditions or pollution scenarios, and the classification confidence score is output. The feature subset size threshold of the second feature screening model is dynamically determined based on the classification confidence level. The second feature screening model filters pollutant features under various pollution scenarios to obtain a scenario-specific set of key pollutants. The key pollutants are then re-identified based on operational feedback to ensure long-term prediction accuracy.

[0008] As a preferred technical solution of the present invention, the odor concentration prediction module adopts a hybrid structure combining a mechanistic model and a data-driven model, wherein: Mechanistic models are used to provide theoretical trends in odor concentration over time based on the mechanisms of odor generation and diffusion. The data-driven model is used to learn the residuals between actual monitoring data and the theoretical trend, and to predict the residuals. The fusion layer is used to fuse the output of the mechanistic model with the output of the data-driven model to obtain the final odor concentration prediction result, and at the same time characterize the change trend of the drug's effect over time.

[0009] As a preferred embodiment of the present invention, the multi-objective prediction and optimization module comprehensively considers the following objectives when generating the drug dosing plan: (1) Minimize the total consumption of medicines; (2) Minimize the duration of odor concentration exceeding the predicted value; (3) Minimize the energy consumption of the injection equipment; By searching for combinations of pesticide concentration, ratio, and spraying time, the optimal or second-best dosing scheme that balances compliance rate and economy can be obtained.

[0010] As a preferred technical solution of the present invention, the agents include at least two major categories: catalysts and oxidants. The system intelligently proportions and synergistically optimizes different agents according to the type and proportion of pollutants, operating conditions and prediction results to achieve multi-agent synergistic treatment.

[0011] This invention also provides a machine learning-based method for dynamic prediction of odor concentration and intelligent optimization of reagent dosing in kitchen waste processing plants, comprising the following steps: S1: Collect data on the concentration of various pollutants in the workshop exhaust gas, as well as environmental parameters such as air volume, temperature, humidity, and air pressure, and store them in the database; S2: Perform outlier removal, missing value completion, standardization, detrending, and signal smoothing on the collected data to form a clean dataset for modeling. S3: Use machine learning feature importance analysis to rank pollutants by contribution, dynamically select the top 10-20 key pollutants, and update the set of key pollutants according to a preset period or sample size threshold. S4: Based on the concentrations of the key pollutants and environmental parameters, a time series prediction model is constructed to predict the odor concentration and its changing trend for multiple future periods, and to generate prediction results for reagent dosing decisions. S5: Based on the prediction results, the agent type, concentration, ratio and spraying time are determined by a multi-objective optimization algorithm to obtain a multi-objective optimized agent dosing scheme; S6: Based on the current odor concentration, agent type and dosage, predict the odor concentration rebound time and the time of exceeding the standard after the agent's action, and trigger the next round of agent spraying in advance before the standard is exceeded; S7: Based on the actual monitoring feedback data after the addition of the reagent, the key pollutant set, odor prediction model parameters and reagent addition optimization strategy are dynamically updated to form a closed-loop control of "sensing-prediction-decision-feedback-re-optimization".

[0012] As a preferred embodiment of the present invention, the screening of key pollutants in step S3 includes: The first classification model was used to identify pollution scenarios in workshop operation data. Based on the classification confidence level under different pollution scenarios, the feature subset size of the second feature screening model is adaptively determined. The second feature screening model is used to screen pollutant features under various pollution scenarios to obtain a scenario-specific set of key pollutants, thereby improving the targeting of odor prediction.

[0013] As a preferred technical solution of the present invention, the time series prediction model used in S4 is a hybrid model that combines a mechanistic model and a deep learning model, which is used to simultaneously predict the trend of odor concentration and the time-effect of drug action, and improve the prediction accuracy through residual learning.

[0014] As a preferred technical solution of the present invention, the feature importance analysis model is selected from the XGBoost model and / or the random forest model; The time series prediction model uses a Long Short-Term Memory (LSTM) network model and / or a Transformer structure. The multi-objective optimization algorithm uses reinforcement learning and / or genetic algorithms. This enables the method to possess strong nonlinear expression and self-learning capabilities in feature selection, time series prediction, and dosing decision-making.

[0015] As a preferred technical solution of the present invention, the closed-loop update in S7 simultaneously applies to the pollutant set, odor prediction model parameters and reagent dosing strategy, enabling the system to be updated in conjunction with seasonal changes, material ratio changes and workshop operating conditions, and to maintain the predictive accuracy and adaptability of the control strategy over a long period of time.

[0016] Compared with existing technologies, the present invention has the following advantages: The machine learning-based odor concentration dynamic prediction and intelligent optimization system and method for chemical dosing in kitchen waste workshops proposed in this invention have significant technical advantages over traditional fixed chemical dosing mode, single pollutant monitoring mode and experience-based treatment methods.

[0017] First, this invention achieves "dynamic perception" of odor emissions and "real-time identification" of key pollutants, overcoming the problems of fixed pollutant types and inability to adapt to complex fluctuations in workshops in traditional odor control methods. Through machine learning feature importance calculation and operating condition classification models, the system can automatically screen 10-20 key odor factors under different pollution scenarios and dynamically update the feature library as it operates, fundamentally improving the targeting of odor control and the accuracy of predictive data.

[0018] Secondly, this invention constructs a multi-step odor concentration prediction framework that integrates a mechanistic model and a deep learning model. This framework can accurately predict odor change trends over the next few tens of minutes to several hours, effectively addressing the shortcomings of traditional treatment strategies that rely on "lagging response and passive application." By inputting the prediction results into a multi-objective optimization algorithm, this invention can simultaneously consider multiple indicators such as odor compliance rate, reagent cost, energy consumption, and spraying duration to generate an optimal or near-optimal reagent application plan. Furthermore, by combining the prediction of reagent effectiveness, it can trigger the next round of spraying in advance, upgrading the treatment strategy from "passive control" to "active intervention," significantly reducing the probability of odor rebound.

[0019] Finally, this invention possesses continuous adaptive evolution capabilities. Through a closed-loop update module, it continuously adjusts the pollutant feature set, prediction model parameters, and optimization strategies, enabling the system to adapt to long-term factors such as seasonal changes, waste composition variations, and operational disturbances. In actual operation verification, the application of this system reduced the average odor concentration by 21%, decreased the duration of odor exceedances by 80.4%, reduced reagent costs by 32%, and decreased the number of complaints by 80%. It also successfully identified three typical pollution scenarios and updated the feature database twice, fully demonstrating the reliability, stability, and engineering application value of this invention in complex, non-steady-state odor control scenarios. Attached Figure Description

[0020] Figure 1 Flowchart of the odor concentration prediction model of this invention; Figure 2 Flowchart of the drug dosing prediction model of this invention; Figure 3 Flowchart of the time-series model for predicting odor concentration in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with embodiments and appendices. Figures 1-3 The present invention will be further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to these embodiments. Equivalent modifications made by those skilled in the art without departing from the principles of the present invention should fall within the protection scope of the present invention.

[0023] Example 1: This example focuses on a machine learning-based intelligent optimization system for dynamic prediction of odor concentration and reagent dosing in a food waste processing plant. It emphasizes the system's module composition, data flow between modules, and its operation in an engineering environment. The system consists of a data acquisition module, a data preprocessing module, a dynamic pollutant screening module, an odor concentration prediction module, a multi-objective optimization module, an exceedance prediction module, a control module, and a closed-loop update module. These modules are interconnected at high frequency via a data bus, forming a complete intelligent odor control system.

[0024] During system operation, online monitoring equipment collects concentrations of various odor pollutants, including ammonia, hydrogen sulfide, methanethiol, and volatile fatty acids, in key areas such as the unloading port, crushing zone, pretreatment zone, and temporary storage zone. Simultaneously, environmental parameters such as airflow, temperature, humidity, and air pressure are also collected. After anomaly removal, missing data completion, and signal smoothing by the preprocessing module, the collected data enters the dynamic pollutant screening module for real-time feature identification. This module uses a machine learning model to calculate the feature importance of pollutants and dynamically adjusts the set of key pollutants based on operating conditions, providing high-quality input for subsequent prediction modules.

[0025] The system's data flow, pollutant screening logic, and prediction model training path are as follows: Figure 1 As shown in the figure, pollutant characteristics are mutually verified through dual-channel data from gas sensors and GC-MS. The screening results, along with environmental parameters, are input into the machine learning model to ultimately generate an odor concentration prediction model, forming the core data and model flow path of the system.

[0026] In the prediction phase, the system employs a hybrid structure that integrates mechanistic models and deep learning models to predict odor trends for the next tens of minutes to several hours. The prediction results are input into a multi-objective optimization module to calculate the dosage concentration, ratio, and spraying duration of the pesticide, while simultaneously input into an exceedance prediction module to determine the duration of pesticide action. The system ultimately uses a control module for automatic dosing and a closed-loop update module to continuously refine the feature library and model parameters, ensuring long-term stable operation under different seasons and material disturbances.

[0027] In the system deployment, multiple monitoring points are set up in the workshop using online detection equipment to monitor dozens of odor pollutants, including ammonia, hydrogen sulfide, methanethiol, dimethyl disulfide, acetic acid, propionic acid, and butyric acid. Simultaneously, environmental parameters such as airflow, temperature, humidity, and air pressure are collected. The collected data is uploaded every 5 minutes and undergoes outlier removal, missing value completion, standardization, and signal smoothing by a preprocessing module to maintain the stability and continuity of the input data. Subsequently, the data is transmitted to a dynamic pollutant screening module. This module uses random forest or XGBoost models to calculate the contribution of each pollutant to the odor concentration and dynamically determines the feature screening range based on real-time operating conditions, thereby selecting 10–20 of the most critical pollutants as input variables for the prediction model.

[0028] The odor prediction module employs a hybrid structure combining mechanistic and deep learning models. By fusing mechanistic trends with historical data residuals, it outputs a predicted sequence of odor concentrations for the next tens of minutes to several hours. These predictions are then input into a multi-objective optimization module to generate the optimal pesticide dosing strategy. The optimization module uses pesticide consumption, odor compliance rate, spraying energy consumption, and equipment safety constraints as optimization objectives, and employs a combined approach of genetic algorithms and reinforcement learning to solve for the optimal pesticide concentration, ratio, and spraying duration.

[0029] The prediction results are simultaneously fed into the exceedance prediction module to assess the odor rebound time after the agent's action and to pre-set the next round of spraying rhythm. Finally, the control module controls the spraying system to execute the treatment according to the optimized instructions. This system has a closed-loop update function, which can automatically initiate model correction based on the error between the actual monitored values ​​and the predicted values ​​after spraying, dynamically adjusting the pollutant set, prediction model weights, and dosage strategy to ensure the system maintains high efficiency over the long term. This embodiment demonstrates that the system has a complete modular structure, a stable data flow mechanism, and highly automated decision-making capabilities, making it effectively applicable to workshop odor control scenarios.

[0030] Example 2: This example describes a machine learning-based method for dynamic prediction of odor concentration and intelligent optimization of chemical dosing in a food waste processing plant. The main process structure of this invention is as follows: Figure 2 As shown. This method includes six main steps: data acquisition and preprocessing, dynamic pollutant screening, odor concentration prediction, multi-objective input optimization, exceedance early warning judgment, and closed-loop update, forming a complete process from "data perception" to "intelligent addition" and then to "self-evolutionary review".

[0031] In the initial stage of the method's operation, the concentrations of various odorous pollutants in the workshop's exhaust gas were collected using online monitoring equipment, and external environmental parameters such as air volume, temperature, humidity, and air pressure were recorded and stored in a database. To address the volatility and complexity of the raw data, this method employs preprocessing steps to remove outliers, fill in missing data, unify the dimensions of the data, and detrend the data. Furthermore, techniques such as moving averages are used to improve signal stability, providing high-quality input information for subsequent machine learning models.

[0032] After obtaining the cleaned data, this method performs a dynamic pollutant screening step. First, the classification model identifies the operating conditions of the data, such as determining whether it is in a "high sulfur content scenario" or a "high fatty acid fluctuation scenario". Then, the importance score of the pollutants is calculated using a random forest or XGBoost model, and the size of the screening subset is dynamically determined according to the operating condition confidence. Finally, 10-20 key pollutants are selected from dozens of pollutants as the core input variables of the odor prediction model.

[0033] The odor prediction step employs a hybrid structure model. A mechanistic model calculates the theoretical trends of pollutant release and diffusion, while a deep learning model learns the nonlinear characteristics of the residuals. The two models are fused to output predicted odor concentrations for multiple future time periods. The prediction results are used for subsequent multi-objective optimization. This method uses a genetic algorithm and a reinforcement learning module to solve for the reagent dosing strategy. It comprehensively considers odor compliance, reagent costs, spray equipment energy consumption, and spraying time safety constraints. It optimizes the concentration combinations, mixing ratios, and spraying durations of different reagents, outputting a set of optimal dosing schemes that satisfy multiple objectives.

[0034] Once the dosing strategy is generated, the method further executes an over-limit warning step. Based on the "pesticide action time curve" formed by the prediction model and the agent action model, the duration of the agent's effect is determined, the rebound time point when the odor may reach the limit again is given, and the start time of the next round of spraying is set in advance, so that odor control changes from a passive response to a proactive control, fundamentally reducing the risk of odor rebound.

[0035] During implementation, this method continuously updates the pollutant set, model parameters, and optimization strategies by incorporating post-spray monitoring feedback. When significant deviations occur between monitoring data and model predictions, the system automatically triggers model retraining, gradually enabling the method to develop adaptive update capabilities and achieve continuous evolution. This embodiment demonstrates the completeness, logic, and feasibility of the method's process.

[0036] Example 3: In this example, the system of the present invention is deployed in a kitchen waste treatment workshop with a daily processing capacity of approximately 300 tons. Continuous operation verifies the system's predictive capabilities, optimization capabilities, and treatment effects under real-world conditions. Odor emissions from the workshop exhibit significant fluctuations and non-steady-state characteristics, with pollutant composition dynamically changing depending on factors such as the batch of feed, waste type, and degree of fermentation. This example demonstrates the entire process from data acquisition, scene recognition, odor prediction, reagent optimization to closed-loop updates, and further provides a quantitative performance comparison based on actual operation.

[0037] During system operation, the online monitoring equipment collects the concentrations of pollutants such as ammonia, hydrogen sulfide, methanethiol, dimethyl disulfide, and volatile fatty acids every 5 minutes, while simultaneously collecting data on airflow, temperature, humidity, and air pressure. After being cleaned by the preprocessing module, the data enters the dynamic pollutant screening module. During a high-load feeding period in the afternoon, the system identified fatty acids such as acetic acid, propionic acid, and butyric acid as key pollutants through feature importance analysis, and dynamically updated the set of key pollutants to ensure the predictive model maintains high accuracy.

[0038] The prediction module, based on a specific hybrid model for a "high fatty acid scenario," predicts odor concentration changes over the next 30 minutes to 3 hours. The prediction results show that without intervention, the odor concentration will rebound to 18 mg / m³ within 30 minutes, exceeding the limit of 15 mg / m³. The multi-objective optimization module generates a treatment plan based on the prediction results, recommending a combined application strategy of "emulsifier + oxidant." The control module then automatically completes the spraying operation.

[0039] After application, the system continues to track monitoring values. The exceedance prediction module analyzes the efficacy of the pesticide and determines that the current application will show signs of decay after 120 minutes. Therefore, the next round of spraying is brought forward to 110 minutes to avoid a rebound peak. During long-term operation, the closed-loop update module continuously corrects the feature set and prediction model based on monitoring feedback, making the overall treatment more stable.

[0040] To objectively evaluate system performance, we selected one month's worth of operational data before and after system installation for comparative analysis, resulting in the following key performance indicator table:

[0041] The performance evaluation results above demonstrate that the system of this invention achieves significant results under actual working conditions: the average odor concentration in the workshop decreased from 16.2 mg / m³ to 12.8 mg / m³; the duration of odor exceedance decreased from 9.2% to 1.8%; reagent costs decreased by 32%; and the number of complaints decreased from 5 to 1, significantly improving the stability of pollution control. Furthermore, the system can automatically identify three typical pollution scenarios and update the feature database twice during operation, fully demonstrating its adaptive and self-learning capabilities in complex odor emission environments.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based dynamic prediction of odor concentration and intelligent optimization of reagent dosing system for kitchen waste workshop, characterized in that, The method comprises the following steps: A data acquisition module is used to collect real-time data including concentrations of various odor pollutants, air volume, temperature, humidity, and atmospheric pressure environmental parameters, and upload them to a central database; A data preprocessing module is used to perform outlier rejection, missing value completion, standardization, detrending, and signal smoothing processing on the original monitoring data to reduce the influence of sensor drift and random noise; A dynamic pollutant screening module is used to calculate the contribution of each pollutant to odor concentration or odor grade based on a machine learning feature importance ranking method, sort and dynamically screen the top 10-20 key pollutants, and retrain and update the pollutant feature set according to a preset period or when the cumulative number of new samples reaches a threshold; An odor concentration prediction module is used to construct a time series prediction model based on the concentrations of the key pollutants and the air volume, temperature, humidity, and atmospheric pressure environmental variables, predict the odor concentration and change trend in future time periods, and output the prediction results for medicament dosing decision-making; A multi-objective prediction and optimization module is used to jointly predict the odor concentration, medicament concentration and proportion, spraying time, and over-limit time after dosing based on the odor concentration prediction results, and comprehensively determine the medicament type, dosing concentration, ratio, and spraying time through a multi-objective optimization algorithm to generate an optimal medicament dosing scheme; An over-limit prediction module is used to predict the time point when the odor concentration exceeds the limit value again after the action of the medicament based on the medicament dosing conditions and the time series model, so as to trigger the next round of spraying strategy in advance; A control module is used to issue control instructions to the spraying system according to the dosing scheme output by the multi-objective prediction and optimization module and the time information output by the over-limit prediction module, and perform the corresponding medicament dosing operation; A closed-loop update module is used to update and iteratively train the pollutant set, odor prediction model parameters, and medicament dosing optimization strategy based on the actual monitoring feedback data after medicament dosing, so that the system has long-term self-adaptation and self-evolution capabilities.

2. The system of claim 1, wherein, The dynamic pollutant screening module further comprises: A first classification model is used to classify the operating data of the workshop or the pollution scene, and output a classification confidence; A feature subset size threshold of a second feature screening model is dynamically determined based on the classification confidence; The second feature screening model is used to screen the pollutant features under each type of pollution scene to obtain a scene-specific key pollutant set, and re-identify the key pollutants combined with the operating feedback to ensure long-term prediction accuracy. 3.The machine learning based intelligent system for dynamic prediction of odour concentration and optimization of dosing of odour control agents in a kitchen waste plant according to claim 1, wherein, The odor concentration prediction module adopts a hybrid structure combining mechanism models and data-driven models, wherein: The mechanism model is used to give the theoretical trend of the change of odor concentration with time based on the mechanism of odor generation and diffusion; The data-driven model is used to learn the residual between the actual monitoring data and the theoretical trend, and predict the residual; The fusion layer is used to fuse the outputs of the mechanism model and the data-driven model to obtain the final odor concentration prediction result, and simultaneously represent the change trend of the medicament action time with time.

4. The machine learning-based dynamic prediction and reagent dosing intelligent optimization system for kitchen waste plant odor concentration according to claim 1, characterized in that, The multi-objective prediction and optimization module considers the following objectives when generating the medicament dosing scheme: (1) minimizing the total consumption of reagents; (2) minimizing the duration of odor concentration prediction value exceeding the standard; (3) minimizing the energy consumption of reagent dosing equipment; and by searching for the combination of reagent dosing concentration, ratio and spraying time, the optimal or suboptimal dosing scheme considering both compliance rate and economy is obtained.

5. The machine learning-based intelligent optimization system for dynamic prediction of kitchen waste plant odor concentration and dosing of reagents according to claim 1, characterized in that, The reagents include at least two categories of catalysts and oxidants, and the system intelligently matches and optimizes different reagents according to the types and proportions of pollutants, operating conditions and prediction results, achieving collaborative management of multiple reagents.

6. A method for dynamic prediction of odor concentration and intelligent optimization of reagent dosage in a kitchen waste plant based on the system of any one of claims 1-5, characterized in that, The method comprises the following steps: S1: collecting the concentration of multiple pollutants in the waste gas of the workshop, air volume, temperature, humidity, and air pressure environmental parameter data, and storing them in a database; S2: performing outlier rejection, missing value completion, standardization, detrending, and signal smoothing on the collected data to form a cleaned data set for modeling; S3: using machine learning feature importance analysis method to sort the contribution of pollutants, dynamically selecting the top 10-20 key pollutants, and updating the key pollutant set according to the preset period or sample size threshold; S4: based on the key pollutant concentration and environmental parameters, a time series prediction model is constructed to predict the odor concentration and trend in future periods, and the prediction results for reagent dosing decision are formed; S5: based on the prediction results, the type, concentration, ratio and spraying time of the reagent are determined by a multi-objective optimization algorithm to obtain a multi-objective optimized reagent dosing scheme; S6: based on the current odor concentration, reagent type and dosing amount, the rebound time and over-standard time of the odor after the reagent is applied are predicted, and the next round of reagent spraying is triggered in advance at a preset time before the odor exceeds the standard; S7: based on the actual monitoring feedback data after reagent dosing, the key pollutant set, odor prediction model parameters and reagent dosing optimization strategy are dynamically updated to form a closed-loop control of "perception-prediction-decision-feedback-reoptimization".

7. The method of claim 6, wherein, The selection of key pollutants in S3 includes: using a first classification model to identify pollution scenarios based on workshop operation data; adaptively determining the feature subset size of a second feature selection model according to the classification confidence under different pollution scenarios; using the second feature selection model to select the pollutant features under each pollution scenario to obtain a scenario-specific key pollutant set.

8. The method of claim 6, wherein, The time series prediction model used in S4 is a hybrid model combining mechanism model and deep learning model, which is used to simultaneously realize odor concentration trend prediction and reagent action time prediction, and improve prediction accuracy through residual learning.

9. The method of claim 6, wherein: the feature importance analysis model selects an XGBoost model and / or a random forest model; the time series prediction model selects a long short-term memory network (LSTM) model and / or a Transformer structure; the multi-objective optimization algorithm selects a reinforcement learning algorithm and / or a genetic algorithm, so that the method has nonlinear expression and self-learning ability at the levels of feature selection, time series prediction and dosing decision.

10. The method of claim 6, wherein, The closed loop update in S7 acts on the pollutant set, odor prediction model parameters and reagent dosing strategy, so that the system is updated in linkage with seasonal changes, material ratio changes and workshop working condition changes, and long-term self-adaptability of prediction accuracy and control strategy is maintained.

Images