植物活墙净化VOCs效能:机器学习预测与优化OACHSSCD

Indoor volatile organic compounds(VOC)pose a serious threat to human health.Hence,living walls(also known as vertical gardens)have significant potential as a green purification technology.Yet their purification efficiency is subject to complex influences from multiple factors that are difficult to measure and predict quantitatively through traditional research methods.This study aims to overcome the limitations of single-variable analysis.It constructs a data-driven research framework to quantify the purification capacity of living walls,develop a dynamic prediction model,analyse the parameter influence mechanism,and propose an intelligent optimisation strategy,thus facilitating the paradigm shift of this technology from empirical to data-driven design. The research follows the path of"experiment construction-model prediction-mechanism interpretation-strategy optimisation".First,a multi-parameter experimental framework was designed to encompass the environmental dimensions(background concentration,temperature and humidity)and the living wall dimensions(plant species,living wall area,and plant density).The framework was systematically designed in a controlled test chamber using a representative dataset obtained from stratified sampling.The VOC purification data of three living walls(Scindapsus aureus,Chlorophytum comosum and Mentha haplocalyx)under various working conditions were measured.Second,the performances of six machine learning algorithms-random forest,support vector regression,LightGBM,XGBoost,CatBoost,and multi-layer perceptron-were compared.A dynamic prediction model was constructed using the optimal algorithm and nested cross-validation.The parameter mechanism was then interpreted by SHAP values and partial dependence plots.Finally,a comprehensive efficiency evaluation function was built using purification capacity,economic cost,and maintenance difficulty as objectives.The configuration schemes were refined through multi-objective optimisation and sensitivity analysis. 1)Model performance:among the six algorithms,LightGBM showed the best performance(RMSE=0.0542±0.0046,R2=0.8529±0.0472).The constructed dynamic prediction model achieved significantly higher accuracy(R2>0.85)on the independent test set compared to traditional methods.2)Parameter influence:The influence ranking was as follows:background concentration>temperature>humidity>living wall area>plant density>plant species.Mechanism interpretation revealed that background concentration showed a stable positive correlation with plant purification efficiency;temperature had a threshold effect(significantly positive correlation when<14℃ or>25℃,with mild influence in the 14℃～25℃ range);humidity exerted a non-linear influence(rising first and then falling,with the optimal level at～65％);both area and density presented positive correlations.Mentha haplocalyx showed the best plant purification efficiency,followed by Chlorophytum comosum and Scindapsus aureus.3)Optimisation strategies and configuration schemes:three engineering-oriented living wall configuration strategies were proposed,which were"short-term high-efficiency type"(focusing on maximising purification capacity;recommending the large-area,high-density Mentha haplocalyx or Scindapsus aureus),"cost-effective practical type"(balancing purification,cost and maintenance,dominated by Scindapsus aureus),and"clean low-maintenance type"(prioritising low-maintenance cost,preferring Chlorophytum comosum).A specific parameter configuration table was generated for different indoor VOC concentration levels(low-concentration daily environments/high-concentration short-term pollution)and seasonal temperature conditions(low-temperature season/high-temperature season). This study constructed a dynamic prediction model for VOCs purification by living walls based on the LightGBM machine learning algorithm,which realised accurate and dynamic quantification of purification effects.Through interpretability analysis,the complex mechanism of environmental and structural parameters was quantified and revealed.In particular,the threshold effect of temperature and the non-linear influence of humidity were discovered.The final multi-objective optimisation framework and three configuration strategies elevated living wall design from empirical judgement to data-driven decision-making,providing critical methodological support and technical tools for the scientific application of this technology to healthy buildings. This research is based on controlled experiments,and its generalisability to real buildings with variable ventilation and multi-pollutant source scenarios needs to be further verified.Moreover,the model neither involves the air-exchange rate nor interprets the multi-pollutant interaction mechanisms.Multi-scenario field verification is needed,and a comprehensive purification prediction framework of multiple pollutants shall be explored in future.