Multi-site VOCs prediction method based on statistical information assistance

Abstract

The invention discloses a multi-site VOCs prediction method based on statistical information assistance, and the method comprises the steps: 1, collecting multi-site VOCs monitoring data of a to-be-predicted region, and carrying out the preprocessing of the multi-site VOCs monitoring data, and obtaining the preprocessed multi-site VOCs monitoring data; step 2, performing statistical characteristic calculation and similarity analysis on the preprocessed multi-site VOCs monitoring data to obtain statistical characteristic similarity between different sites; and step 3, selecting site data with high similarity according to the similarity between different sites, calculating an autocorrelation statistic of site monitoring quantity, and selecting feature attributes for model training according to the statistic. Step 4, training an LSTM model by using the site data with high similarity in the step 3), and evaluating and optimizing the LSTM model to obtain a multi-site VOCs prediction model; and step 5, predicting the discharged VOCs by adopting the multi-site VOCs prediction model.

Description

Technical Field

[0001] The present invention relates to the field of industrial waste gas monitoring and prevention, and in particular to a multi-site VOCs prediction method assisted by statistical information. Background Art

[0002] The monitoring and early warning of volatile organic compounds (VOCs) in industrial parks are extremely important, and the lack of highly timely VOCs prediction technology has led to weak early warning and handling capabilities for VOCs pollution emergencies. Once a pollution emergency occurs, it may seriously endanger regional production safety and people's lives and health, causing irreparable losses.

[0003] The LSTM (Long Short-Term Memory) model [Graves A, Graves A. Long short-term memory. Supervised sequence labelling with recurrent neural networks, 2012: 37-45.] has made important research progress in predicting air quality [Van Houdt G, Mosquera C, Nápoles GA review on the long short-term memory model. Artificial Intelligence Review, 2020, 53 (8): 5929-5955.]. First, in terms of improving measurement accuracy, the LSTM model outperforms traditional statistical models and other deep learning models in predicting air quality. For example, some studies have shown that the LSTM model has higher accuracy and lower error than other models in predicting PM2.5 and PM10 concentrations [Tsai YT, Zeng YR, Chang Y S. Air pollution forecasting using RNN with LSTM / / 2018 IEEE 16th Intl Conf on Dependable, Autonomic and Secure Computing, 16th Intl Conf on Pervasive Intelligence and Computing, 4th Intl Conf on BigData Intelligence and Computing and Cyber ​​Science and Technology Congress (DASC / PiCom / DataCom / CyberSciTech). IEEE, 2018: 1074-1079.]. Secondly, in terms of multi-source data fusion, the LSTM model can process multi-source data, which is very important for air quality prediction.For example, some studies have integrated meteorological data (such as temperature, humidity, wind speed, etc.), traffic data (such as traffic flow, speed, etc.), geographic data (such as terrain, vegetation coverage, etc.) and incorporated other multi-source data into the LSTM model to further improve the prediction accuracy [Inam S, Mahmood A, Khatoon S, et al. Multisource data integration and comparative analysis of machine learning models for on-street parking prediction. Sustainability, 2022, 14 (12): 7317.]. Third, in terms of the expansion of the prediction time range, traditional air quality prediction models can only predict air quality in the short term (such as a few hours or a day). However, the memory capacity of the LSTM model enables it to process data sequences over a long period of time, so it can be used to predict air quality over a longer period of time (such as a few days or a week). Although deep learning models are generally considered a "black box", some studies have begun to explore the interpretability of LSTM models [Liu DR, Lee SJ, Huang Y, et al. Airpollution forecasting based on attention-based LSTM neural network and ensemble learning. Expert Systems, 2020, 37(3): e12511.].

[0004] The above research progress shows that LSTM models have broad application prospects in air quality prediction. However, the prediction performance of these models is still affected by factors such as data quality, model structure, and parameter selection, and further research and optimization are needed. Based on the literature review, there are few studies focusing on using multi-site VOCs monitoring data for prediction, and there are also few studies on using statistical data of VOCs monitoring data to assist deep learning models to improve prediction effects. Summary of the invention

[0005] The present invention provides a multi-site VOCs prediction method based on statistical information assistance. The present invention uniquely integrates statistical information into multi-site VOCs prediction, so that short-term, long-term and potential environmental changes can be accurately reflected.

[0006] A multi-site VOCs prediction method based on statistical information assistance includes the following steps:

[0007] 1) Collect multi-site VOCs monitoring data of the area to be predicted, and pre-process them to obtain pre-processed multi-site VOCs monitoring data;

[0008] 2) Calculate the statistical characteristics and perform similarity analysis on the preprocessed multi-site VOCs monitoring data to obtain the statistical characteristics similarity between different sites;

[0009] 3) Select site data with high similarity based on the similarity between different sites, calculate the autocorrelation statistics of the site monitoring data, and select feature attributes for model training based on the statistics.

[0010] 4) Using the site data with high similarity in step 3) to train the LSTM model, and evaluating and optimizing the LSTM model to obtain a multi-site VOCs prediction model;

[0011] 5) Use a multi-site VOCs prediction model to predict the emitted VOCs.

[0012] In step 1), multiple sites refer to three or more sites.

[0013] In step 1), preprocessing includes: data cleaning and format adjustment.

[0014] In step 1), the multi-site VOCs monitoring data is time series data of non-methane hydrocarbons (VOCs).

[0015] In step 2), the statistical characteristics of the pre-processed multi-site VOCs monitoring data are calculated and analyzed, specifically including:

[0016] The pre-processed multi-site VOCs monitoring data is used to obtain the mean value, standard deviation, variance, minimum value and maximum value of the VOCs data of each site;

[0017] The similarity of each site's statistics is calculated using the cosine similarity algorithm.

[0018] Cosine similarity algorithm:

[0019]

[0020] Among them, vector A = {a1, a2, a3, a4, a5}, vector B = {b1, b2, b3, b4, b5}. The mean, standard deviation, variance, minimum and maximum values ​​of VOCs at each site can be represented by vector A or B, where the five elements a1, a2, a3, a4, a5 or b1, b2, b3, b4, b5 represent the mean, standard deviation, variance, minimum and maximum values, respectively.

[0021] Therefore, the similarity value between different sites can be obtained, expressed as f ij (i.e., f in formula (1)), i.e., the cosine similarity between site i and site j. The closer its value is to 1, the more similar it is.

[0022] In step 3), according to the similarity values ​​between different sites obtained in step 2), a threshold greater than 0.6 or three or more sites with the highest similarity are set as sites for modeling the multi-site prediction model.

[0023] In step 3), the Spearman correlation coefficient matrix of the six monitoring quantities of non-methane hydrocarbons, exhaust gas, flue gas flow, flue gas temperature, flue gas pressure, and flue gas moisture content of the site is calculated. Based on the statistic, the monitoring quantities with an absolute value of the correlation coefficient with non-methane hydrocarbons greater than 0.5 or the top 2 rankings are selected as characteristic attributes for model training, such as non-methane hydrocarbons, flue gas flow, and flue gas pressure.

[0024] In step 4), the input data for training the LSTM model is the selected multiple sites, and the data is aligned according to the monitoring time. The format is:

[0025] Time, monitoring station 1 data 1, monitoring station 1 data 2, monitoring station 1 data 3, monitoring station 2 data 1, monitoring station 2 data 2, monitoring station 2 data 3, monitoring station 3 data 1, monitoring station 3 data 2, monitoring station 3 data 3, ...

[0026] In step 4), Keras is used to build the LSTM model, adding one hidden layer with 50 neurons; finally, an output layer with 4 neurons is added.

[0027] In step 4), epochs is set to 100, the adam optimizer is used for parameter optimization, and the root mean square error (RMSE), mean absolute error (MAE) and r-root prediction value are used to evaluate the prediction effect. Model training is performed by adding the number of hidden layers and the dimension of the hidden layers, and fine-tuning the learning rate to optimize the model structure, and finally a trained model is obtained.

[0028] In step 5), the trained multi-site VOCs prediction model is used, the required prediction time is input, prediction is performed, and the prediction values ​​of multiple sites are output respectively.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] 1) Through this method, the similar characteristics of VOCs data from multiple sites can be accurately extracted, providing a scientific basis for subsequent data processing and environmental quality assessment. Specifically, it includes but is not limited to obtaining statistical indicators such as the mean value, standard deviation, variance, minimum and maximum values ​​of the data. These indicators can fully reflect the changing trend, fluctuation range and distribution state of VOCs concentration. By comparing and analyzing the statistical characteristics of the data from each monitoring site, the present invention further reveals that by constructing a prediction model based on the statistical similarity of VOCs monitoring data between different sites, the accuracy of multi-site monitoring can be improved.

[0031] 2) The present invention is about a training method for selecting various attributes and performing data fusion after analyzing the autocorrelation statistical characteristics of multi-site VOCs monitoring data, specifically selecting key attributes with prediction information, and aligning and merging these attributes to generate a training data set. Through this method, we remove redundant information while maintaining the comprehensiveness of the data, effectively improving the processing efficiency of the training set and the accuracy of the prediction model. This is an environmental monitoring and prediction technology that combines statistical characteristic analysis with machine learning concepts, which is expected to significantly improve the accuracy and efficiency of environmental monitoring and prediction.

[0032] 3) The present invention uniquely integrates statistical information into multi-site VOCs prediction, so that short-term, long-term and potential environmental changes can be accurately reflected. Through the long short-term memory (LSTM) model, historical monitoring data can be effectively analyzed and learned to predict the VOCs concentration every hour in the future. All of this is done in the same model framework, and the implementation process is coherent and concise, which greatly reduces the design and operation costs. More importantly, this technology of the present invention is not only highly accurate, but also fast in execution and strong in real time. The design based on statistical information assistance makes the model highly flexible and adaptable, and can make accurate predictions for various environmental and meteorological conditions. This method has a much better error control ability than traditional machine learning methods when predicting hourly VOCs concentrations. In addition, this technology can also be extended to other similar environmental condition prediction scenarios, such as the concentration prediction of particulate matter PM2.5 and PM10, and the concentration prediction of gases such as sulfur dioxide and carbon monoxide. With the assistance of statistical information, the model is able to better understand and predict the complexity and dynamics of environmental data, representing a new direction for future environmental monitoring and prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present disclosure, but are not intended to limit the present disclosure.

[0034] Figure 1 It is a schematic diagram of the spatial distribution of VOCs monitoring stations in the area to be monitored in this application.

[0035] Figure 2 This is a framework diagram of a new multi-site VOCs prediction method based on statistical information assistance proposed in this application.

[0036] Figure 3 It is the statistical value of the historical multi-site VOCs monitoring data in Example 1.

[0037] Figure 4It is a time series comparison chart of multi-site VOCs predicted values ​​and true values ​​(monitoring values) in Example 1.

[0038] Figure 5 It is the performance evaluation index of the multi-site VOCs prediction model in Example 1. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0040] In the present application, unless otherwise specified or different meanings can be derived from the context, each term has the meaning generally understood in the art.

[0041] like Figure 2 As shown, a multi-site VOCs prediction method based on statistical information assistance includes the following steps:

[0042] 1) Collect multi-site VOCs monitoring data in the area to be predicted and preprocess them to obtain preprocessed multi-site VOCs monitoring data; multi-site refers to three or more sites. Preprocessing includes: data cleaning and format adjustment. Multi-site VOCs monitoring data is time series data of non-methane hydrocarbons (VOCs).

[0043] 2) Calculate the statistical characteristics and perform similarity analysis on the preprocessed multi-site VOCs monitoring data to obtain the statistical characteristics similarity between different sites;

[0044] Calculate and analyze the statistical characteristics of the pre-processed multi-site VOCs monitoring data, including:

[0045] The pre-processed multi-site VOCs monitoring data is used to obtain the mean value, standard deviation, variance, minimum value and maximum value of the VOCs data of each site;

[0046] The similarity of each site's statistics is calculated using the cosine similarity algorithm.

[0047] Cosine similarity algorithm:

[0048]

[0049] Among them, vector A = {a1, a2, a3, a4, a5}, vector B = {b1, b2, b3, b4, b5}. The mean, standard deviation, variance, minimum and maximum values ​​of VOCs at each site can be represented by vector A or B, where the five elements a1, a2, a3, a4, a5 or b1, b2, b3, b4, b5 represent the mean, standard deviation, variance, minimum and maximum values, respectively.

[0050] Therefore, the similarity value between different sites can be obtained, expressed as f ij , that is, the cosine similarity between site i and site j. The closer the value is to 1, the more similar they are.

[0051] 3) Select site data with high similarity based on the similarity between different sites, calculate the autocorrelation statistics of the site monitoring data, and select feature attributes for model training based on the statistics.

[0052] According to the similarity values ​​between different sites obtained in step 2), set the threshold greater than 0.6 or the three or more sites with the highest similarity as the sites for modeling the multi-site prediction model. Calculate the Spearman correlation coefficient matrix of the six monitoring quantities of non-methane hydrocarbons, exhaust gas, flue gas flow, flue gas temperature, flue gas pressure, and flue gas moisture content of the site. According to the statistic, select the monitoring quantities with an absolute value of the correlation coefficient with non-methane hydrocarbons greater than 0.5 or the top two rankings for the characteristic attributes of model training, such as non-methane hydrocarbons, flue gas flow, and flue gas pressure.

[0053] 4) Use the site data with high similarity in step 3) to train the LSTM model, evaluate and optimize the LSTM model, and obtain a multi-site VOCs prediction model; the input data for training the LSTM model is the selected multiple sites, and the data is aligned according to the monitoring time, and the format is:

[0054] Time, monitoring station 1 data 1, monitoring station 1 data 2, monitoring station 1 data 3, monitoring station 2 data 1, monitoring station 2 data 2, monitoring station 2 data 3, monitoring station 3 data 1, monitoring station 3 data 2, monitoring station 3 data 3, ...

[0055] Use Keras to build an LSTM model, add one hidden layer with 50 neurons; finally add an output layer with 4 neurons.

[0056] The number of epochs was set to 100, the adam optimizer was used for parameter optimization, the root mean square error (RMSE), mean absolute error (MAE) and r-square root prediction value were used to evaluate the prediction effect, and the model training was performed. The model structure was optimized by adding the number of hidden layers and the dimension of the hidden layers, and fine-tuning the learning rate to finally obtain a trained model.

[0057] 5) Use a multi-site VOCs prediction model to predict the emitted VOCs. Use the trained multi-site VOCs prediction model, input the required prediction time, make predictions, and output the predicted values ​​of multiple sites respectively.

[0058] The present application can be more easily understood through the following embodiments.

[0059] Example

[0060] Taking an industrial park in Jinshan District, Shanghai as an example, a case application is carried out. Figure 1 It is a schematic diagram of the spatial distribution of VOCs monitoring stations in the area to be monitored in this application. Figure 2 This is a framework diagram of a new multi-site VOCs prediction method based on statistical information assistance proposed in this application. Figure 3 It is the statistical value of the historical multi-site VOCs monitoring data in Example 1. Figure 4 It is a time series comparison chart of multi-site VOCs predicted values ​​and true values ​​(monitoring values) in Example 1. Figure 5 It is the performance evaluation index of the multi-site VOCs prediction model in Example 1.

[0061] The data comes from monitoring data from four automatic monitoring stations in an industrial park in Shanghai from June 15, 2020 to June 15, 2021. The monitoring items include six items: non-methane hydrocarbons, exhaust gas, flue gas flow, flue gas temperature, flue gas pressure, and flue gas moisture content. The monitoring principle of the automatic monitoring station is to use gas chromatography hydrogen flame ionization (GC-FID), and the detector is a high-sensitivity FID detector. The detector is deployed at the exhaust port of the factory for 24-hour monitoring, and the monitoring project data is recorded every 60 minutes. Therefore, there are more than 8,000 records for the six monitoring items monitored by each station.

[0062] Considering the protection of factory privacy, the above data was cleaned. The original monitoring place names were hidden, and A, B, C, and D were used to represent the monitoring stations respectively. Figure 1 In order to test the advantages of this method, sites A, B, and C adopted the method proposed in this patent, while site D adopted the traditional method without using statistical information to assist VOCs prediction.

[0063] The data is then divided into a training set and a test set. Then, the statistics of the training data are calculated, including the mean, standard deviation, variance, and minimum and maximum values. The specific process of the method of the present invention is as follows Figure 1 As shown, the operation steps are as follows:

[0064] 1) Data collection and preprocessing. We first collect data from each monitoring station and preprocess it. The preprocessing steps include data cleaning (e.g., handling missing values ​​and outliers), feature engineering (e.g., normalization and standardization), and other possible steps to ensure that the data is suitable for subsequent model training.

[0065] 2) Statistical feature calculation and similarity analysis. Calculate various statistical data of the prediction results, such as mean, variance, full range, etc., and then use the cosine similarity calculation formula to calculate the similarity values ​​between different sites. Set the threshold greater than 0.6 or the three or more sites with the highest similarity as the sites for modeling the multi-site prediction model. In this case, they are sites A, B, and C.

[0066] 3) Data fusion: The selected multi-site monitoring quantities include non-methane hydrocarbons, exhaust gas, flue gas flow, flue gas temperature, flue gas pressure, and flue gas moisture content. By calculating the Spearman correlation coefficient matrix, the monitoring quantities with an absolute value of correlation coefficient with non-methane hydrocarbons greater than 0.5 or ranked in the top 2 are selected according to the statistic as the characteristic attributes for model training. In this case, they are non-methane hydrocarbons, flue gas flow, and flue gas pressure. Through this method, redundant information is removed, which effectively improves the processing efficiency of the training set and the accuracy of the prediction model.

[0067] 4) Training set and test set division. The data of each monitoring station is divided into a training set and a test set. 90% of the data is used as the training set, and the remaining 10% of the data is used as the test set. The training set data is used to train the model, and the test set data is used to evaluate the prediction performance of the model.

[0068] 5) LSTM model training. Use the training set data to train the LSTM model. During the training process, adjust the parameters and structure of the model to find the optimal model.

[0069] 6) Model prediction. Use the trained model to predict the test set data. Compare the generated prediction results with the actual values ​​to evaluate the prediction performance of the model, such as root mean square error (RMSE), mean absolute error (MAE) and r-squared prediction value, such as Figure 3 If the prediction performance of the model is not satisfactory, return to the "Training LSTM Model" step and optimize the model structure by increasing the number of hidden layers and the dimensions of the hidden layers and fine-tuning the learning rate.

[0070] 7) Model performance test: Compare the prediction performance of the multi-station prediction model. Analyze the advantages and disadvantages of the model and explore the possible reasons. This will help us understand how to improve the model to improve its prediction performance.

[0071] In the case of the present invention, the prediction performance of sites A, B, and C is significantly higher than the prediction accuracy of site D. Figure 4 , Figure 5 As shown. It is explained that the present invention calculates the similarity of the statistical characteristics of VOCs monitoring data from multiple sites to select sites with high similarity. At the same time, after calculating the autocorrelation statistical characteristics of the monitoring data from multiple sites, various attributes are selected and data alignment and fusion are performed. This training data construction method specifically selects key attributes with predictive information. Through this method, we remove redundant information while maintaining the comprehensiveness of the data, effectively improving the processing efficiency of the training set and the accuracy of the prediction model. This is an environmental monitoring and prediction technology that combines statistical characteristic analysis with machine learning concepts, which is expected to significantly improve the accuracy and efficiency of environmental monitoring and prediction.

[0072] The above description is merely an exemplary embodiment of the present disclosure and is not intended to limit the protection scope of the present disclosure. The protection scope of the present disclosure is determined by the appended claims.

Claims

1. A multi-site VOCs prediction method based on statistical information, characterized in that: The following steps are involved: 1) Collect multi-site VOCs monitoring data of the area to be predicted, and pre-process them to obtain pre-processed multi-site VOCs monitoring data; 2) Calculate the statistical characteristics and perform similarity analysis on the preprocessed multi-site VOCs monitoring data to obtain the statistical characteristics similarity between different sites; 3) According to the similarity of statistical characteristics between different sites, select the site data with high similarity, calculate the autocorrelation statistics of the site monitoring data, and select the characteristic attributes for model training according to the autocorrelation statistics; 4) Using the site data with high similarity in step 3) to train the LSTM model, and evaluating and optimizing the LSTM model to obtain a multi-site VOCs prediction model; 5) Use a multi-site VOCs prediction model to predict the emitted VOCs.

2. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 1), multiple sites refer to three or more sites.

3. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 1), preprocessing includes: data cleaning and format adjustment.

4. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 1), the multi-site VOCs monitoring data is time series data of non-methane VOCs.

5. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 2), statistical feature calculation and similarity analysis are performed on the pre-processed multi-site VOCs monitoring data, specifically including: The pre-processed multi-site VOCs monitoring data is used to obtain the mean value, standard deviation, variance, minimum value and maximum value of the VOCs data of each site; The similarity of the statistics of each site is calculated using the cosine similarity algorithm to obtain the statistical feature similarity between different sites.

6. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 3), based on the similarity of statistical features between different sites, site data with high similarity is selected, specifically including: According to the statistical feature similarity between different sites obtained in step 2), a threshold greater than 0.6 or three or more sites with the highest similarity is set, and the data corresponding to the site is selected as the site data with high similarity.

7. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 3), the autocorrelation statistics of the site monitoring data are calculated, and the feature attributes for model training are selected according to the autocorrelation statistics, specifically including: The Spearman correlation coefficient matrix of the six monitoring quantities of non-methane hydrocarbons, exhaust gas, flue gas flow, flue gas temperature, flue gas pressure, and flue gas moisture content of the site was calculated, and the monitoring quantities with an absolute value of the correlation coefficient with non-methane hydrocarbons greater than 0.5 or ranked in the top 2 were selected as the characteristic attributes for model training.

8. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 4), the LSTM model is constructed using Keras, adding one hidden layer and finally adding an output layer.

9. The multi-site VOCs prediction method based on statistical information assistance according to claim 1 is characterized in that: In step 4), the LSTM model is evaluated and optimized to obtain a multi-site VOCs prediction model, which specifically includes: The adam optimizer was used for parameter optimization, and the root mean square error, mean absolute error, and R-square root prediction value were used to evaluate the prediction effect. Model training was performed by adding the number of hidden layers and the dimension of the hidden layers and fine-tuning the learning rate to optimize the model structure. Finally, a trained model was obtained as a multi-site VOCs prediction model.

10. The multi-site VOCs prediction method based on statistical information assistance according to claim 1, characterized in that: In step 5), a multi-site VOCs prediction model is used to predict the emitted VOCs, including: A multi-site VOCs prediction model is used to input the required prediction time, perform prediction, and output the predicted values ​​of multiple sites respectively.

Images