Machine learning-based indoor pollen concentration prediction method and system

Abstract

The application relates to the technical field of indoor environment monitoring and intelligent regulation and control, and discloses an indoor pollen concentration prediction method and system based on machine learning, which comprises S1, multi-source heterogeneous data acquisition, S2, multi-scale feature engineering and physical feature extraction, S3, construction of a physical and data double-driven prediction model, S4, model training and verification, S5, real-time prediction and dynamic correction, and S6, result output and regulation suggestion generation. The indoor pollen concentration prediction method and system based on machine learning introduce a dynamic self-adaptive correction and abnormal event correction mechanism, can directly reflect the actual sensitization risk, enable the model to continuously adapt to the environmental changes of different buildings and seasons, do not need to be retrained for specific scenes, and reduce the deployment and maintenance cost of the system. Time strict alignment of multi-source heterogeneous data and physical feature extraction provide accurate prior constraints for the model, and improve the interpretability and reliability of the prediction result.

Description

Technical Field

[0001] This invention relates to the field of indoor environmental monitoring and intelligent control technology, specifically to a method and system for predicting indoor pollen concentration based on machine learning. Background Technology

[0002] Pollen concentration prediction equipment can be widely used in indoor air quality management of various buildings such as residences, offices, and schools, providing accurate health protection guidance for people with pollen allergies, while also providing data support for the intelligent operation of building ventilation systems and air purification equipment.

[0003] Existing technologies include various methods and systems for predicting indoor pollen concentration. While purely data-driven machine learning models can fit concentration changes under normal operating conditions, they are prone to significant deviations and insufficient generalization ability under extreme weather conditions such as rainfall and strong winds, as well as sudden scenarios such as dense crowds. Purely physical models, while simulating pollen movement based on fluid dynamics principles, are computationally complex and time-consuming, making them unsuitable for real-time prediction applications. Most existing prediction systems only predict the total concentration of pollen grains, indirectly inferring allergy risk through grain concentration. They fail to distinguish the differences in allergenicity among pollen morphologies and sizes, leading to a mismatch between warning information and the actual severity of allergic reactions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a machine learning-based method and system for predicting indoor pollen concentration. This addresses the technical problems of existing technologies, such as large prediction deviations and insufficient generalization ability of purely data-driven models under extreme weather conditions like rainfall and strong winds, as well as sudden scenarios like dense crowds, and the computational complexity and time consumption of purely physical models making them difficult to meet real-time prediction requirements. Furthermore, predicting only the total concentration of pollen grains cannot distinguish the differences in allergenicity of pollen of different morphologies and sizes, leading to a mismatch between early warning information and the actual degree of allergic harm.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based method for predicting indoor pollen concentration, comprising: S1. Multi-source heterogeneous data acquisition: Simultaneously collect outdoor meteorological, building ventilation, indoor environmental and indoor and outdoor pollen monitoring data to ensure strict alignment of each parameter in the time dimension, eliminate correlation analysis errors caused by asynchronous data, and provide a unified time benchmark for subsequent multi-dimensional feature fusion; S2. Multi-scale feature engineering and physical feature extraction: Data was cleaned, normalized, and missing values ​​were imputed. Outliers were removed using the 3σ criterion, and the Z-score method was used to unify data dimensions, improving the convergence speed and stability of subsequent model training. Spatiotemporal features, architectural features, and meteorological features were extracted to comprehensively cover various key factors affecting indoor pollen concentration. Based on the physical mechanism of pollen movement, gravity settling rate, airflow carrying coefficient, and I / O ratio dynamic coefficient (i.e., the ratio of indoor to outdoor pollen concentration) were calculated to characterize the efficiency of outdoor pollen transfer to indoor environments and the dynamic physical characteristics of the resuspension trigger threshold. The gravity settling rate was calculated using the Stokes equation. In the formula The diameter of the pollen grain. Pollen grain density, Where is the air density, and g is the acceleration due to gravity. The aerodynamic viscosity is used; the airflow carrying coefficient is determined based on the ratio of indoor wind speed to pollen settling rate, with a value range of 0.1-0.9; the resuspension trigger threshold is set to be triggered when the indoor activity intensity of people is greater than 0.5 m / s²; the motion characteristics of pollen particles under different environmental conditions are quantified to provide prior constraints that conform to objective laws for the data-driven model; S3. Construct a prediction model driven by both physics and data: The physical mechanism module uses CFD simulation to observe the diffusion, sedimentation, and resuspension of pollen under different ventilation conditions. The output dimension is a physical constraint tensor representing the time step, number of monitoring points, and number of physical features, accurately reproducing the interaction between the indoor flow field and pollen particles, thus overcoming the shortcomings of pure data-driven models in generalization under extreme conditions. The data-driven module employs a spatiotemporal attention-enhanced Transformer network, which performs element-wise weighted fusion of multi-scale features and the physical constraint tensor at the attention fusion layer. This effectively captures long-term temporal dependencies and complex nonlinear associations, addressing the problem of insufficient spatiotemporal prediction accuracy in traditional machine learning models. The network consists of an input layer, a spatiotemporal coding layer, a multi-head attention layer, an attention fusion layer, and a multi-task output layer. The spatiotemporal coding layer uses sinusoidal position coding combined with learnable spatial embedding coding to encode temporal and spatial location information, respectively. The multi-head attention layer contains eight parallel attention heads, which capture temporal dependencies at different scales. At the end of the network, a parallel allergen and pollen particle co-prediction branch is set. The two branches share the underlying feature extraction network and output the concentration of pollen particles of different sizes and the corresponding allergen concentrations for the next 1-24 hours, respectively. This enables the simultaneous prediction of particle concentration and sensitization risk, avoiding the limitation that predicting only particle concentration cannot accurately reflect the actual allergic harm. S4. Model Training and Validation: Historical data was divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer was used for supervised training of the model, with an initial learning rate of 0.001 and a batch size of 32. A physical loss function was introduced to constrain the output range, embedding fundamental physical laws into the training process. This forced the model output to conform to the laws of mass and momentum conservation, improving the physical rationality and interpretability of the prediction results. The physical loss function included mass conservation loss. Settlement rate constraint loss Compared to I / O range constraint loss The mass conservation loss is calculated as the absolute value of the change in total indoor pollen mass and the difference in pollen mass entering and leaving the room; the sedimentation rate constraint loss is calculated as the mean square error between the model-predicted sedimentation rate and the theoretically calculated sedimentation rate; the I / O ratio range constraint loss is calculated as the penalty term when the predicted I / O ratio exceeds the range [0.01, 0.9]; the total loss function is... ,in For mean square error data loss, For physical loss, These are dynamic weight coefficients, used in the early stages of training. It increases linearly with the number of training iterations. By optimizing hyperparameters through validation sets, Bayesian optimization methods are used to optimize hyperparameters such as learning rate, batch size, and number of attention heads to find the optimal model configuration, thereby further improving prediction accuracy and robustness. S5. Real-time prediction and dynamic correction: The preprocessed real-time data is input into the trained model, which outputs hourly predictions of pollen and allergen concentrations for the next 1-24 hours. Based on monitoring bias, the model's output layer parameters are updated online via a dynamic adaptive correction submodule. The model is continuously optimized based on the latest monitoring data to adapt to environmental changes in different buildings and seasons, addressing the issue of performance degradation over time in offline training models. The dynamic adaptive correction submodule employs a sliding window for online learning, with a window size of 24 hours and a learning rate of 0.0001. Parameters are updated hourly. When the mean absolute percentage error (MAPE) of prediction exceeds 1 for three consecutive time steps, the prediction is considered complete. The fully connected layer parameters are updated when the rainfall is between 5% and 20%. Simultaneously, anomaly event correction is triggered, identifying anomalies through real-time data mutation characteristics: rainfall exceeding 0.5 mm / h is classified as a rainfall event, and the I / O ratio dynamic coefficient is multiplied by a correction factor of 0.2; outdoor wind speed exceeding 10 m / s is classified as a strong wind event, and the airflow carrying capacity is multiplied by a correction factor of 1.5; indoor CO2 concentration change rate exceeding 50 ppm / 10 min is classified as a densely populated activity event, and the resuspension trigger threshold is lowered to 0.3 m / s². These specific adjustments significantly improve prediction accuracy under extreme events. S6. Result Output and Regulation Suggestion Generation: The system generates tiered early warning information, classifying allergy risk into four levels based on predicted allergen concentrations: low risk (allergen concentration <10 ng / m³), medium risk (10 ng / m³ ≤ allergen concentration <50 ng / m³), high risk (50 ng / m³ ≤ allergen concentration <200 ng / m³), and very high risk (allergen concentration ≥200 ng / m³). This facilitates quick identification and appropriate protective measures for users. It also combines ventilation system parameters to output optimal suggestions for door and window opening / closing, air purifier operation, and personnel activity. Door and window opening / closing strategies are based on outdoor pollen concentration, indoor CO2 concentration, and ventilation needs. Based on a comprehensive assessment: when outdoor pollen concentration is <100 pollen grains / m³ and indoor CO2 concentration is >1000ppm, it is recommended to open windows for ventilation for 30 minutes; when outdoor pollen concentration is ≥100 pollen grains / m³, it is recommended to close doors and windows and turn on the air purifier; the air purifier's operating parameters should be automatically adjusted according to the indoor allergen concentration: low setting for low risk, medium setting for medium risk, and high setting with negative ion function for high risk and above; the effectiveness of different control measures should be quantitatively analyzed to minimize energy consumption and allergy risk while ensuring that indoor CO2 concentration ≤1000ppm and other air quality indicators meet the standards.

[0006] Preferably, in S1, multi-source heterogeneous data acquisition, the pollen monitoring data particle size range is divided into 0-2.5μm, 2.5-10μm, 10-20μm, 20-30μm, 30-40μm, and >40μm. The morphological types include monomers, fragmented, aggregated, and unknown pollen, covering the vast majority of allergenic pollen particle ranges, distinguishing pollen types with different movement characteristics and allergenic abilities, and providing a basis for accurate prediction. In S2, multi-scale feature engineering and physical feature extraction, the I / O ratio dynamic coefficient comprehensively considers ventilation mode, air exchange rate, outdoor meteorological conditions, and pollen particle size morphology, accurately reflecting the transmission relationship of indoor and outdoor pollen concentration under different conditions, solving the systematic bias caused by the traditional fixed I / O ratio assumption. In the fully closed mode, it is calculated as a function of building gap permeability and pollen particle size. After rainfall, it is corrected based on the rainwater flushing efficiency of pollen with different particle sizes, considering the difference in the flushing effect of rainwater on pollen with different particle sizes, and improving the accuracy of indoor pollen concentration prediction after rainfall.

[0007] Preferably, in S3, the construction of a dual-driven prediction model based on physics and data, the physics mechanism module uses the Lagrange particle tracking method to simulate the trajectory of pollen grains, tracking the complete movement path of a single pollen grain, accurately simulating the interaction between the grain and the wall and airflow, improving the accuracy of the physics simulation, considering wall collision rebound, adhesion, and resuspension caused by human activity, and pre-establishing a pollen concentration distribution database under different ventilation conditions to quickly call physical constraint terms, avoiding the high computational load of real-time CFD simulation, realizing millisecond-level calling of physical constraint terms, and meeting the performance requirements of real-time prediction; the spatiotemporal attention-enhanced Transformer network includes an input layer, a spatiotemporal coding layer, an attention fusion layer, and a multi-task output layer, which adaptively allocates weights through a multi-head attention mechanism, automatically identifies the importance of different influencing factors, focuses on abrupt changes that have a significant impact on pollen concentration, and focuses on capturing the impact of abrupt weather changes and sudden events on concentration.

[0008] Preferably, in S4, model training and validation, the physical loss function includes mass conservation, sedimentation rate constraint, and I / O ratio range constraint loss, which constrain the model output from three dimensions: total amount conservation, particle motion law, and concentration transfer relationship, respectively, to ensure that the prediction results conform to physical common sense. The total loss is the weighted sum of data loss and physical loss. The weight of physical loss is dynamically adjusted with training iterations. In the early stage of training, a higher weight is used to guide the model to learn basic physical laws, and in the later stage, the weight is reduced to allow the model to better fit the actual data, balancing physical rationality and data fitting accuracy. In S5, real-time prediction and dynamic correction, the dynamic adaptive correction submodule adopts online learning with a sliding window, using only the monitoring data of the most recent period to update parameters, which ensures the timeliness of the model and avoids excessive interference from historical data. When the prediction deviation exceeds 15%-20%, parameter updates are triggered. The abnormal event correction mechanism adjusts the model parameters for rainfall, strong winds, and densely populated activities.

[0009] Preferably, in S3, the construction of a dual-driven prediction model based on physics and data, the allergen and pollen particle collaborative prediction branch achieves independent prediction by learning the correlation between pollen breakage rate and allergen release amount and particle size distribution characteristics, directly inferring allergy risk from pollen physical characteristics without relying on expensive and time-consuming laboratory allergen testing; in S6, the result output and control suggestion generation, the door and window opening strategy and air purifier parameters are determined based on outdoor concentration, ventilation requirements and filtration efficiency, minimizing outdoor pollen entry and indoor pollen residue while ensuring that indoor air quality indicators such as CO2 concentration meet standards.

[0010] An indoor pollen concentration prediction system based on machine learning, employing the aforementioned machine learning-based indoor pollen concentration prediction method, includes: The multi-source data acquisition module is used to simultaneously collect outdoor meteorological and pollen monitoring data, realize the integrated acquisition of multi-dimensional environmental parameters, and provide the system with comprehensive and accurate input data; The pollen monitoring unit is equipped with a multi-size graded sampler and an allergen detection device, which can simultaneously detect pollen particles of different sizes and morphologies and their corresponding allergen concentrations, and simultaneously acquire the physical characteristics and biological sensitization characteristics of pollen, providing data support for dual-dimensional prediction. The data preprocessing module is used for data cleaning, normalization, and missing value completion, eliminating data noise and inconsistencies, and providing high-quality data for subsequent model training and prediction. The prediction module is driven by both physics and data, including a physical mechanism unit and a data-driven unit. The physical mechanism unit simulates pollen diffusion, sedimentation and resuspension processes based on CFD and outputs physical constraint terms, providing prior constraints that conform to physical laws and improving model interpretability. The data-driven unit uses a spatiotemporal attention-enhanced Transformer network to leverage the advantages of data-driven models in fitting complex nonlinear relationships and improve prediction accuracy. The model training and validation module is used for supervised training of the model using historical data. It introduces physical loss function to constrain the output and optimize hyperparameters. Through large-scale historical data training and rigorous validation, it ensures the model's generalization ability in different scenarios. The real-time prediction and dynamic correction module is used to input real-time data and output prediction results. It processes input data in real time and outputs results quickly to meet users' requirements for timeliness. The output and control module is used to generate tiered early warnings and activity recommendations, transforming professional forecast results into intuitive and easy-to-understand information and actionable suggestions for user convenience.

[0011] Preferably, the multi-source data acquisition module includes meteorological monitoring, building ventilation monitoring, indoor environment monitoring, and pollen monitoring units, which respectively collect corresponding parameters. The meteorological unit acquires key parameters affecting pollen diffusion, the ventilation unit acquires indoor and outdoor air exchange parameters, and the indoor environment unit acquires parameters affecting pollen resuspension, forming a complete parameter acquisition system. The pollen monitoring unit is equipped with a Durham gravity sampler, an Andersen six-level graded sampler, and an ELISA detection device. The data preprocessing module includes a multi-dimensional pollen feature extraction unit, which is used to calculate gravity settling rate, airflow carrying coefficient, I / O ratio dynamic coefficient, and resuspension trigger threshold, automatically calculating key physical characteristics related to pollen movement without manual intervention, thus improving the system's automation level.

[0012] Preferably, the physical mechanism unit has a built-in CFD simulation database for different apartment types and ventilation conditions, which can quickly call physical constraint terms, pre-simulate common scenarios, significantly reduce system calculation latency, meet real-time requirements, and support online simulation of custom apartment types and ventilation conditions. It also supports users to upload custom apartment diagrams and parameters to generate personalized physical constraints, which are applicable to various special building structures. The multi-head attention mechanism of the data-driven unit includes time, space and feature attention heads, which capture the influencing factors of corresponding dimensions. The time head captures periodic patterns such as daily and seasonal changes, the space head captures the concentration differences at different locations in the room, and the feature head automatically identifies the weight of each influencing factor.

[0013] Preferably, the real-time prediction and dynamic correction module includes an abnormal event detection submodule, which can identify abnormal events such as rainfall, strong winds, and dense crowds entering and exiting the area and trigger corresponding correction strategies. It automatically identifies anomalies through data mutation characteristics without manual intervention and triggers the correction mechanism in a timely manner. The allergen and pollen particle collaborative prediction branch has built-in allergen release models for different pollen types. It calculates the concentration of free and bound allergens based on the proportion of broken pollen and the particle size distribution, distinguishes the sensitizing ability of allergens in different states, and more accurately reflects the actual allergy risk.

[0014] Preferably, the output and control module includes a closed-loop ventilation control feedback unit, which can communicate with the intelligent ventilation system and air purifier to achieve active control, forming a closed loop of prediction, control, monitoring and re-prediction, automatically adjusting equipment operating parameters without manual operation, and further improving control effect and convenience.

[0015] Compared with existing technologies, this invention provides a machine learning-based method and system for predicting indoor pollen concentration, which has the following advantages: This machine learning-based indoor pollen concentration prediction method and system employs a dual-driven prediction architecture of physics and data, combined with a constraint mechanism of physical loss function. This ensures that the model output conforms to objective physical laws and can adapt to complex and ever-changing indoor environments, solving the problems of insufficient generalization ability of purely data-driven models under extreme conditions and high computational complexity of purely physics models. By setting up a co-prediction branch for allergens and pollen particles, it can directly reflect the actual sensitization risk, avoiding the limitations of inferring allergic hazards solely from particle concentration. The introduction of dynamic adaptive correction and abnormal event correction mechanisms enables the model to continuously adapt to environmental changes in different buildings and seasons without requiring retraining for specific scenarios, reducing system deployment and maintenance costs. Strict temporal alignment and physical feature extraction of multi-source heterogeneous data provide accurate prior constraints for the model, improving the interpretability and reliability of the prediction results. Attached Figure Description

[0016] Figure 1This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0017] This invention provides a technical solution; please refer to [link / reference]. Figure 1 A machine learning-based method for predicting indoor pollen concentration includes: S1. Multi-source heterogeneous data acquisition: Simultaneously collect outdoor meteorological, building ventilation, indoor environmental and indoor and outdoor pollen monitoring data to ensure strict alignment of each parameter in the time dimension, eliminate correlation analysis errors caused by asynchronous data, and provide a unified time benchmark for subsequent multi-dimensional feature fusion; S2. Multi-scale feature engineering and physical feature extraction: The data is cleaned, normalized, and missing values ​​are filled in. Outliers are removed and data units are standardized to improve the convergence speed and stability of subsequent model training. Spatiotemporal, architectural, and meteorological features are extracted to comprehensively cover various key factors affecting indoor pollen concentration. Based on the physical mechanism of pollen movement, the dynamic physical characteristics of gravity settling rate, airflow carrying coefficient, I / O ratio dynamic coefficient, and resuspension trigger threshold are calculated to quantify the movement characteristics of pollen particles under different environmental conditions and provide prior constraints that conform to objective laws for data-driven models. S3. Construct a prediction model driven by both physics and data: The physical mechanism module simulates the diffusion, sedimentation, and resuspension of pollen under different ventilation conditions using CFD, outputting physical constraints to accurately recreate the interaction between the indoor flow field and pollen particles. This compensates for the lack of generalization ability of pure data-driven models under extreme conditions. The data-driven module uses a spatiotemporal attention-enhanced Transformer network to integrate multi-scale features and physical constraints, effectively capturing long-term temporal dependencies and complex nonlinear associations. This solves the problem of insufficient spatiotemporal prediction accuracy of traditional machine learning models. A co-prediction branch for allergens and pollen particles is set up to achieve simultaneous prediction of particle concentration and sensitization risk, avoiding the limitation that predicting only particle concentration cannot accurately reflect the actual allergic hazards. S4. Model Training and Validation: The model is trained using historical data, and a physical loss function is introduced to constrain the output range. Basic physical laws are embedded into the training process, forcing the model output to conform to the laws of mass conservation and momentum conservation, thereby improving the physical rationality and interpretability of the prediction results. Hyperparameters are optimized through validation sets, and the optimal model configuration is found through Bayesian optimization and other methods, further improving prediction accuracy and robustness. S5. Real-time prediction and dynamic correction: Input real-time data and output prediction results for the next 1-24 hours. Based on monitoring deviation, parameters are updated online through a dynamic adaptive correction submodule. The model is continuously optimized according to the latest monitoring data to adapt to environmental changes in different buildings and seasons, solve the problem of performance decay of offline training models over time, trigger abnormal event correction, and make special adjustments for emergencies such as rain erosion, strong winds, and disturbances caused by human activities, which greatly improves the prediction accuracy under extreme events. S6. Result Output and Regulation Suggestion Generation: It generates graded early warning information, classifies allergy risks into different levels based on predicted concentrations, and facilitates users to quickly identify and take corresponding protective measures. It combines ventilation system parameters to output optimal suggestions for opening and closing doors and windows, operating air purifiers, and managing personnel activities. It quantitatively analyzes the effects of different control measures, minimizing energy consumption and allergy risks while ensuring indoor air quality meets standards.

[0018] S1. In multi-source heterogeneous data acquisition, the pollen monitoring data is divided into particle size ranges of 0-2.5μm, 2.5-10μm, 10-20μm, 20-30μm, 30-40μm, and >40μm. The morphological types include monomers, fragmented, aggregated, and unknown pollen, covering the vast majority of allergenic pollen grains. This distinguishes pollen types with different movement characteristics and allergenic abilities, providing a basis for accurate prediction. S2. In multi-scale feature engineering and physical feature extraction, the I / O ratio dynamic coefficient comprehensively considers ventilation mode, air exchange rate, outdoor meteorological conditions, and pollen grain size morphology, accurately reflecting the transfer relationship of indoor and outdoor pollen concentration under different conditions. This solves the systematic bias caused by the traditional fixed I / O ratio assumption. In the fully closed mode, the calculation is performed by using the permeability of building gaps as a function of pollen grain size. After rainfall, the calculation is corrected based on the rainwater flushing efficiency of pollen with different grain sizes, taking into account the difference in the flushing effect of rainwater on pollen with different grain sizes, thus improving the accuracy of indoor pollen concentration prediction after rainfall.

[0019] In S3, the physical mechanism module uses the Lagrange particle tracking method to simulate the trajectory of pollen grains, tracking the complete movement path of a single pollen grain and accurately simulating the interaction between the grain and the wall and airflow, thus improving the accuracy of the physical simulation. It considers wall collision rebound, adhesion, and resuspension caused by human activities. A pollen concentration distribution database under different ventilation conditions is pre-established to quickly call physical constraint terms, avoiding the high computational load of real-time CFD simulation and achieving millisecond-level calling of physical constraint terms to meet the performance requirements of real-time prediction. The spatiotemporal attention-enhanced Transformer network includes an input layer, a spatiotemporal coding layer, an attention fusion layer, and a multi-task output layer. It adaptively allocates weights through a multi-head attention mechanism, automatically identifies the importance of different influencing factors, focuses on abrupt changes that have a significant impact on pollen concentration, and focuses on capturing the impact of abrupt weather changes and sudden events on concentration.

[0020] S4. In model training and validation, the physical loss function includes mass conservation, sedimentation rate constraint, and I / O ratio range constraint loss. These constraints constrain the model output from three dimensions: total quantity conservation, particle motion law, and concentration transfer relationship, ensuring that the prediction results conform to physical common sense. The total loss is the weighted sum of data loss and physical loss. The weight of physical loss is dynamically adjusted with training iterations. In the early stage of training, a higher weight is used to guide the model to learn basic physical laws. In the later stage, the weight is reduced to allow the model to better fit the actual data, balancing physical rationality and data fitting accuracy. S5. In real-time prediction and dynamic correction, the dynamic adaptive correction submodule adopts online learning with a sliding window. It only uses the monitoring data of the most recent period to update parameters, which ensures the timeliness of the model and avoids excessive interference from historical data. When the prediction deviation exceeds 15%-20%, parameter updates are triggered. The abnormal event correction mechanism adjusts the model parameters for rainfall, strong winds, and densely populated activities.

[0021] S3. In the construction of the dual-driven prediction model of physics and data, the allergen and pollen grain co-prediction branch achieves independent prediction by learning the correlation between pollen breakage rate and allergen release amount and particle size distribution characteristics. It directly infers the allergy risk from the physical characteristics of pollen without relying on expensive and time-consuming laboratory allergen testing. S6. In the result output and control suggestion generation, the door and window opening and closing strategy and air purifier parameters are determined based on the outdoor concentration, ventilation demand and filtration efficiency. Under the premise of ensuring that indoor air quality indicators such as CO2 concentration meet the standards, it minimizes the entry of outdoor pollen and indoor pollen residue.

[0022] An indoor pollen concentration prediction system based on machine learning, comprising the following: The multi-source data acquisition module is used to simultaneously collect outdoor meteorological and pollen monitoring data, realize the integrated acquisition of multi-dimensional environmental parameters, and provide the system with comprehensive and accurate input data; The pollen monitoring unit is equipped with a multi-size graded sampler and an allergen detection device, which can simultaneously detect pollen particles of different sizes and morphologies and their corresponding allergen concentrations, and simultaneously acquire the physical characteristics and biological sensitization characteristics of pollen, providing data support for dual-dimensional prediction. The data preprocessing module is used for data cleaning, normalization, and missing value completion, eliminating data noise and inconsistencies, and providing high-quality data for subsequent model training and prediction. The prediction module is driven by both physics and data, including a physical mechanism unit and a data-driven unit. The physical mechanism unit simulates pollen diffusion, sedimentation and resuspension processes based on CFD and outputs physical constraint terms, providing prior constraints that conform to physical laws and improving model interpretability. The data-driven unit uses a spatiotemporal attention-enhanced Transformer network to leverage the advantages of data-driven models in fitting complex nonlinear relationships and improve prediction accuracy. The model training and validation module is used for supervised training of the model using historical data. It introduces physical loss function to constrain the output and optimize hyperparameters. Through large-scale historical data training and rigorous validation, it ensures the model's generalization ability in different scenarios. The real-time prediction and dynamic correction module is used to input real-time data and output prediction results. It processes input data in real time and outputs results quickly to meet users' requirements for timeliness. The output and control module is used to generate tiered early warnings and activity recommendations, transforming professional forecast results into intuitive and easy-to-understand information and actionable suggestions for user convenience.

[0023] The multi-source data acquisition module includes meteorological monitoring, building ventilation monitoring, indoor environment monitoring, and pollen monitoring units, which collect corresponding parameters. The meteorological unit acquires key parameters affecting pollen diffusion, the ventilation unit acquires indoor and outdoor air exchange parameters, and the indoor environment unit acquires parameters affecting pollen resuspension, forming a complete parameter acquisition system. The pollen monitoring unit is equipped with a Durham gravity sampler, an Andersen six-level graded sampler, and an ELISA detection device. The data preprocessing module includes a multi-dimensional pollen feature extraction unit, which is used to calculate gravity settling rate, airflow carrying coefficient, I / O ratio dynamic coefficient, and resuspension trigger threshold. It automatically calculates key physical characteristics related to pollen movement without manual intervention, improving the system's automation level.

[0024] The physical mechanism unit has a built-in CFD simulation database for different apartment types and ventilation conditions. It can quickly call up physical constraints and pre-simulate common scenarios, significantly reducing system calculation latency and meeting real-time requirements. It also supports online simulation of custom apartment types and ventilation conditions, allowing users to upload custom floor plans and parameters to generate personalized physical constraints, applicable to various special building structures. The data-driven unit's multi-head attention mechanism includes time, space, and feature attention heads, which capture the influencing factors of corresponding dimensions. The time head captures periodic patterns such as daily and seasonal changes, the space head captures concentration differences at different locations indoors, and the feature head automatically identifies the weight of each influencing factor.

[0025] The real-time prediction and dynamic correction module includes an abnormal event detection submodule, which can identify abnormal events such as rainfall, strong winds, and dense crowds and trigger corresponding correction strategies. It automatically identifies anomalies through data mutation characteristics without manual intervention and triggers the correction mechanism in a timely manner. The allergen and pollen particle collaborative prediction branch has built-in allergen release models for different pollen types. It calculates the concentration of free and bound allergens based on the proportion of broken pollen and the particle size distribution, distinguishes the sensitizing ability of allergens in different states, and more accurately reflects the actual allergy risk.

[0026] The output and control module includes a closed-loop ventilation control feedback unit, which can communicate with intelligent ventilation systems and air purifiers to achieve active control, forming a closed loop of prediction, control, monitoring and re-prediction, automatically adjusting equipment operating parameters without manual operation, further improving control effect and convenience.

Claims

1. A machine learning-based method for predicting indoor pollen concentration, characterized in that, include: S1. Multi-source heterogeneous data acquisition: Simultaneously collect outdoor meteorological, building ventilation, indoor environmental, and indoor and outdoor pollen monitoring data; S2. Multi-scale feature engineering and physical feature extraction: The data is cleaned, normalized, and missing values ​​are filled in. Spatiotemporal, architectural, and meteorological features are extracted. Based on the physical mechanism of pollen movement, the dynamic physical features of gravity settling rate, airflow carrying coefficient, I / O ratio dynamic coefficient, and resuspension trigger threshold are calculated. S3. Construct a prediction model driven by both physics and data: The physical mechanism module is based on CFD simulation of the diffusion, sedimentation and resuspension of pollen under different ventilation conditions, and outputs physical constraint terms. The data-driven module adopts a spatiotemporal attention-enhanced Transformer network, integrates multi-scale features and physical constraints, and sets up a co-prediction branch for allergens and pollen particles. S4. Model Training and Validation: The model is trained under supervision using historical data, and the output range is constrained by a physical loss function. The hyperparameters are then optimized using a validation set. S5. Real-time prediction and dynamic correction: Input real-time data and output prediction results for the next 1-24 hours. Based on the monitoring deviation, the parameters are updated online through the dynamic adaptive correction submodule, and abnormal event correction is triggered. S6. Result Output and Regulation Suggestion Generation: Generate tiered early warning information and, in conjunction with ventilation system parameters, output optimal suggestions for opening and closing doors and windows, operating air purifiers, and managing personnel activities.

2. The indoor pollen concentration prediction method based on machine learning according to claim 1, characterized in that: In S1, multi-source heterogeneous data acquisition, the pollen monitoring data is divided into particle size ranges of 0-2.5μm, 2.5-10μm, 10-20μm, 20-30μm, 30-40μm and >40μm, and the morphological types include monomers, fragments, aggregates and unknown morphological pollen. In S2, multi-scale feature engineering and physical feature extraction, the I / O ratio dynamic coefficient comprehensively considers ventilation mode, air exchange rate, outdoor meteorological conditions and pollen particle size morphology. In the fully closed mode, it is calculated as a function of building gap permeability and pollen particle size. After rainfall, it is corrected based on the rainwater washout efficiency of pollen with different particle sizes.

3. The indoor pollen concentration prediction method based on machine learning according to claim 1, characterized in that: In the S3 section, which constructs a dual-driven prediction model based on physics and data, the physical mechanism module uses the Lagrange particle tracking method to simulate the trajectory of pollen grains, considering wall collisions and rebounds, adhesion, and resuspension caused by human activities. A pollen concentration distribution database under different ventilation conditions is pre-established to quickly call physical constraint terms. The spatiotemporal attention-enhanced Transformer network includes an input layer, a spatiotemporal coding layer, an attention fusion layer, and a multi-task output layer. It adaptively allocates weights through a multi-head attention mechanism, focusing on capturing the impact of sudden weather changes and unexpected events on concentration.

4. The indoor pollen concentration prediction method based on machine learning according to claim 1, characterized in that: In S4, model training and validation, the physical loss function includes mass conservation, settlement rate constraint, and I / O ratio range constraint loss. The total loss is the weighted sum of data loss and physical loss, and the physical loss weight is dynamically adjusted with training iterations. In S5, real-time prediction and dynamic correction, the dynamic adaptive correction submodule uses sliding window online learning. When the prediction deviation exceeds 15%-20%, parameter updates are triggered. The abnormal event correction mechanism adjusts the model parameters for rainfall, strong winds, and densely populated activities.

5. The indoor pollen concentration prediction method based on machine learning according to claim 1, characterized in that: In S3, the construction of a dual-driven prediction model based on physics and data, the allergen and pollen grain collaborative prediction branch achieves independent prediction by learning the correlation between pollen breakage rate and allergen release amount and particle size distribution characteristics; in S6, the result output and control suggestion generation, the door and window opening and closing strategy and air purifier parameters are determined based on outdoor concentration, ventilation demand and filtration efficiency.

6. A machine learning-based indoor pollen concentration prediction system, employing the machine learning-based indoor pollen concentration prediction method according to any one of claims 1-5, characterized in that, include: A multi-source data acquisition module is used to simultaneously collect outdoor meteorological and pollen monitoring data; The pollen monitoring unit is equipped with a multi-size graded sampler and an allergen detection device, which can simultaneously detect pollen particles of different sizes and shapes and their corresponding allergen concentrations. The data preprocessing module is used for data cleaning, normalization, and missing value completion. The dual-drive prediction module, which combines physics and data, includes a physical mechanism unit and a data-driven unit. The physical mechanism unit simulates pollen diffusion, sedimentation, and resuspension processes using CFD and outputs physical constraints. The data-driven unit employs a spatiotemporal attention-enhanced Transformer network. The model training and validation module is used for supervised training of the model using historical data, introducing physical loss function constraints to constrain the output and optimize hyperparameters; The real-time prediction and dynamic correction module is used to input real-time data and output prediction results; The output and control module is used to generate tiered early warnings and activity recommendations.

7. The indoor pollen concentration prediction system based on machine learning according to claim 6, characterized in that: The multi-source data acquisition module includes meteorological monitoring, building ventilation monitoring, indoor environment monitoring, and pollen monitoring units, which collect corresponding parameters respectively. The pollen monitoring unit is equipped with a Durham gravity sampler, an Andersen six-level graded sampler, and an ELISA detection device. The data preprocessing module includes a multi-dimensional pollen feature extraction unit, which is used to calculate gravity settling rate, airflow carrying coefficient, I / O ratio dynamic coefficient, and resuspension trigger threshold.

8. The indoor pollen concentration prediction system based on machine learning according to claim 6, characterized in that: The physical mechanism unit has a built-in CFD simulation database for different apartment types and ventilation conditions, which can quickly call up physical constraint terms and support online simulation of customized apartment types and ventilation conditions; the multi-head attention mechanism of the data-driven unit includes time, space and feature attention heads, which capture the influencing factors of the corresponding dimensions respectively.

9. The indoor pollen concentration prediction system based on machine learning according to claim 6, characterized in that: The real-time prediction and dynamic correction module includes an abnormal event detection submodule, which can identify abnormal events such as rainfall, strong winds, and dense crowds entering and exiting and trigger corresponding correction strategies; the allergen and pollen grain collaborative prediction branch has built-in allergen release models for different pollen types, and calculates the concentration of free and bound allergens based on the proportion of broken pollen and the particle size distribution.

10. The indoor pollen concentration prediction system based on machine learning according to claim 6, characterized in that: The output and control module includes a closed-loop ventilation control feedback unit, which can communicate with the intelligent ventilation system and air purifier to achieve active control.

Images