Method for establishing real time particulate matter concentration calculation Model, calculation model established thereby, and real time particulate matter concentration calculation method using the calculation Mode

Abstract

The present invention comprises a step of collecting basic fine dust data including indoor and outdoor environments and HVAC data inside the facility; A data analysis step that analyzes the correlation between fine dust concentration and input / output data regarding collected basic fine dust data; A method for establishing a real-time fine dust concentration calculation model is provided, comprising: a step of determining input variables (features) for machine learning; a machine learning step of constructing a dataset for the input variables and training a fine dust concentration calculation model; and a step of periodically retraining. In addition, the method provides a real-time fine dust concentration calculation model established by the above method and a method for calculating indoor fine dust concentration in real time using the same.

Description

Technology Field

[0001] The present invention relates to a data-based prediction system utilizing machine learning, which is one of the AI ​​techniques, and to a method for establishing an AI algorithm capable of calculating indoor fine dust concentration. Background Technology

[0002] With the recent rise in importance of indoor environment management, it is essential for all buildings operating HVAC systems to install various sensors to effectively control the indoor environment. However, existing indoor environment monitoring systems primarily rely on temperature and humidity sensors, and fail to monitor particulate matter concentration (PM2.5), a critical element of air quality. 2.5 There are many cases where real-time detection of ) is not performed.

[0003] With the recent expansion of AI technology, fine dust prediction systems utilizing it are being developed. For example, methods are being developed to predict indoor fine dust concentrations using machine learning and determine whether to operate air purifiers based on these results. Furthermore, technologies such as deep learning-based real-time air quality prediction devices and methods are also being developed. However, these technologies [regarding] current indoor fine dust (PM 10 , PM 2.5 Although it calculates the concentration, it utilizes a mathematical formula, so there is no function to calculate it flexibly when the corresponding variable value is not provided.

[0004] As such, since conventional technology calculates fine dust concentration based on a fixed mathematical formula, if essential variable values ​​(temperature, humidity, CO2 concentration, etc.) are not provided, it is impossible to predict or results in inaccurate results.

[0005] Furthermore, conventional technology has a problem in that prediction accuracy significantly decreases when air quality characteristics inside a building change over time. Prior art literature

[0006] Korean Patent Publication 10-2021-0123688 (October 14, 2021) (Smart environmental monitoring system based on big data analysis of fine dust and odor measurement) The problem to be solved

[0007] The present invention is an invention sought to solve the aforementioned problems, and aims to provide a method for establishing a real-time fine dust concentration calculation model that can provide flexible fine dust concentration calculation results in response to changes in the indoor environment and support the efficient operation of an air conditioning system, thereby maximizing the reduction of building operating costs and energy saving effects.

[0008] In addition, the present invention aims to provide a calculation model capable of calculating indoor fine dust in real time based on the established calculation model.

[0009] Furthermore, the present invention aims to provide a method for calculating real-time fine dust concentration that reflects the real-time environment based on the above-mentioned calculation model. means of solving the problem

[0010] A method for establishing a real-time fine dust concentration calculation model according to one embodiment of the present invention comprises: a step of collecting basic fine dust data including indoor and outdoor environments and HVAC data inside a facility; a data analysis step of analyzing the correlation between the fine dust concentration and the input / output data for the collected basic fine dust data; a step of determining input variables (features) for machine learning; a machine learning step of constructing a dataset for the input variables and training a fine dust concentration calculation model; and a step of periodically retraining.

[0011] To analyze the above correlation, the Pearson correlation analysis technique can be performed.

[0012] The above input variable is outdoor PM 2.5It includes concentration, AHU supply and exhaust operating airflow, AHU set temperature, air purifier operating status, EHP airflow, and indoor humidity.

[0013] The above machine learning can be performed using a DNN model-based method.

[0014] The above machine learning can be achieved through learning a data structure having two-dimensional input variables and one-dimensional output variables.

[0015] The above retraining step may include a sliding window-based learning step that deletes some of the data prior to a specific point in time and adds new data.

[0016] The above relearning step can be carried out on a weekly cycle.

[0017] A real-time fine dust concentration calculation model according to one embodiment of the present invention is established by a learning model comprising: a fine dust basic data collection step including indoor and outdoor environments and HVAC data inside a facility; a data analysis step analyzing the correlation between the fine dust concentration and the input / output data for the collected fine dust basic data; a step of determining input variables (features) for machine learning; a machine learning step of building a dataset for the input variables and training a fine dust concentration calculation model; and a step of periodically retraining.

[0018] A method for calculating indoor fine dust concentration according to one embodiment of the present invention is performed by the real-time fine dust concentration calculation model. Effects of the invention

[0019] This invention overcomes the limitations of existing indoor fine dust detection systems and utilizes a machine learning-based model to detect indoor fine dust (PM 2.5 It provides a technology capable of precisely calculating the concentration.

[0020] Furthermore, by applying sliding window-based relearning techniques, adaptability to changes in the indoor environment is enhanced, and the efficient operation of the HVAC system is supported, thereby maximizing the reduction of building operating costs and energy savings. This enables not only improved indoor air quality but also sustainable smart building operations. Brief explanation of the drawing

[0021] FIG. 1 is a flowchart conceptually explaining a method for establishing a real-time fine dust concentration calculation model according to an embodiment of the present invention. Figure 2 is a relationship diagram illustrating the results of a Pearson correlation analysis related to fine dust concentration based on 17 types of data. FIG. 3 is a structural diagram illustrating the data structure of a machine learning dataset according to one embodiment of the present invention. FIG. 4 is a structural diagram of a real-time fine dust concentration calculation model according to one embodiment of the present invention. FIG. 5 is a conceptual diagram illustrating a sliding window type relearning method in one embodiment of the present invention. Specific details for implementing the invention

[0022] Hereinafter, a method for establishing a real-time fine dust concentration calculation model according to an embodiment of the present invention will be described in detail with reference to the attached drawings. However, the following descriptions are exemplary descriptions intended to explain the embodied form of the technical concept of the present invention, and the technical concept of the present invention is not limited by the following descriptions. The technical concept of the present invention can only be interpreted and limited by the claims set forth below.

[0023] FIG. 1 is a flowchart conceptually explaining a method for establishing a real-time fine dust concentration calculation model according to an embodiment of the present invention.

[0024] Referring to FIG. 1, the method for establishing a real-time fine dust concentration calculation model (S100) includes a basic data collection step (S110), a correlation analysis step (S120), an input variable determination step (S130), a machine learning step (140), and a retraining step (S150).

[0025] In the above data collection step (S110), data correlated with fine dust, such as various indoor and outdoor environmental data and HVAC data within a building, is collected. The above indoor and outdoor environmental data may include various climate and weather data, and in this embodiment, outdoor fine dust concentration (PM 2.5 It includes ) etc. In addition, the above HVAC data is an abbreviation for Heating, Ventilation, and Air Conditioning, referring to a system responsible for heating, ventilation, and air control. Environmental HVAC data is data that measures and records environmental conditions inside a building or facility.

[0026] In this embodiment, a total of 17 types of indoor, outdoor, and HVAC data were collected from actual office facilities.

[0027] Meanwhile, the above data collection may be performed using Apache Kafka, etc., and various commercial data collection programs may be utilized depending on the form of the data, and the data collection program in the present invention is not particularly limited.

[0028] When the data balance step (S110) is completed, a correlation analysis step (S120) is performed to pre-analyze the correlation with fine dust concentration based on the collected data.

[0029] In this embodiment, the correlation analysis step (S120) is performed using Pearson correlation analysis.

[0030] In the correlation analysis step (S120), the correlation between the fine dust concentration and the input / output data regarding the collected fine dust basic data is analyzed.

[0031] Correlation analysis is an indicator that can intuitively and efficiently measure the linear relationship between variables; it is easy to interpret and fast to calculate as it is not affected by units. The Pearson correlation analysis technique used in this embodiment is a statistical method that measures the strength and direction of the linear relationship between two variables. In other words, it is advantageous for determining whether two variables tend to increase or decrease together, or whether one variable tends to decrease when the other increases. In Pearson correlation, a value of -0.4 or lower and 0.4 or higher indicates a moderate correlation, -0.6 or lower and 0.6 or higher indicates a strong correlation, and values ​​closer to 0 indicate that the variable has no influence.

[0032] Figure 2 is a relationship diagram illustrating the results of a Pearson correlation analysis related to fine dust concentration based on 17 types of data.

[0033] Based on the correlation analysis results as shown in Fig. 2, an input variable determination step (S130) is performed. The input variable determination step (S130) is a variable (feature) for constructing a dataset for performing machine learning, and depending on the input variable, there may be a significant difference in the reliability of the learning model established.

[0034] Through the above input variable determination step (S130), a total of 6 types of input variables were determined as follows, and the contents of the input variables and the correlation coefficients, which are the results of the Pearson correlation analysis, are listed below.

[0035] (1) AHU system supply and exhaust airflow (ahu_s, ahu_r): -0.23

[0036] (2) Outdoor PM2.5 concentration (out_pm2_5): 0.43

[0037] (3) Air purifier operating status (air_purifier): -0.49

[0038] (4) AHU set temperature (ahu_sp): 0.39

[0039] (5) Indoor humidity (humi): 0.61

[0040] (6) EHP airflow (indoor_flow): -0.57

[0041] According to the Pearson correlation analysis results, the above six input variables are indoor fine dust concentration (PM 2.5 It was determined to have a significant relationship with ). On the other hand, other data were excluded from model training due to low correlation. Consequently, the final input variable is PM 2.5 It was determined by concentration, AHU supply and exhaust airflow, AHU set temperature, air purifier operating status, EHP airflow, and indoor humidity.

[0042] In this way, when the input variable determination step (S130) for machine learning is completed, a machine learning step (S140) is performed to build a dataset for the input variables and train a fine dust concentration calculation model.

[0043] The machine learning step (S140) described above is performed based on a Deep Neural Network (DNN) model.

[0044] DNN models can be designed to learn meaningful patterns even from time-series data and have the advantage of effectively reflecting non-linear relationships between various input variables. In particular, unlike existing recurrent neural network (RNN, GRU, LSTM)-based models that directly process sequential data, DNNs adopt an approach of learning input variables of a fixed size. Therefore, feature engineering was performed to transform time-series data into a form suitable for DNNs, thereby configuring the model to learn changes over time. The transformed data structure consists of two-dimensional input variables (features) and one-dimensional output variables (targets).

[0045] FIG. 3 is a structural diagram illustrating the data structure of a machine learning dataset according to an embodiment of the present invention. FIG. 4 is a structural diagram of a real-time fine dust concentration calculation model according to an embodiment of the present invention.

[0046] Meanwhile, the input variable was designed to convert data from past time points (n-1), such as 5, 10, and 15 minutes, into a feature set in accordance with the ventilation system control cycle, utilizing the information from each time point as an independent feature. This configuration enables the model to effectively learn changes over time, and the output variable (n) represents the indoor PM at the current time point. 2.5 It is concentration. The trained model used evaluation metrics MAE and CvRMSE to verify performance.

[0047] When the machine learning step (S140) is completed, a basic real-time fine dust calculation model (AI) can be established. However, in order to establish a more accurate prediction system by reflecting the latest indoor environmental conditions, the calculation model (A1) can be updated by periodically retraining the calculation model.

[0048] To this end, the present method (S100) includes a relearning step (S150).

[0049] In the above retraining step (S150), for example, retraining for performance improvement is performed on the above output model at 00:00 AM on Sunday.

[0050] Retraining uses a sliding window-based method that deletes some of the previous data and adds new data to train the system.

[0051] FIG. 5 is a conceptual diagram illustrating a sliding window type relearning method in one embodiment of the present invention.

[0052] According to Figure 5, the above retraining method allows for continuous learning by retaining only the latest data and deleting old data, and as new data is introduced, the model reflects the latest state and performance can be improved. For example, when an initial model is created using data from week n to week n+4 and retraining is performed in week n+5, the data from week n is excluded, and retraining is performed using data from week n+1 to n+5.

[0053] To continuously verify the performance of the retrained model, evaluation metrics are analyzed, and if performance degradation is detected, the model prior to retraining is reused. When new data is added, data quality management is performed, including outlier detection and missing value handling, and the model update strategy is optimized through data drift detection.

[0054] A real-time fine dust concentration calculation model established as described above can be established, and the said real-time fine dust concentration calculation model continuously reflects changes in the indoor fine dust environment through periodic learning.

[0055] By utilizing this real-time fine dust concentration calculation model, By supporting the efficient operation of building HVAC systems, it is possible to maximize reductions in building operating costs and energy savings. This enables not only improved indoor air quality but also sustainable smart building operations.

Claims

Claim 1 A method for establishing a real-time fine dust concentration calculation model, comprising: a step of collecting basic fine dust data including indoor and outdoor environments and HVAC data inside a facility; a data analysis step of analyzing the correlation between fine dust concentration and input / output data regarding the collected basic fine dust data; a step of determining input variables (features) for machine learning; a machine learning step of constructing a dataset for the input variables and training a fine dust concentration calculation model; and a step of periodically retraining. Claim 2 A method for establishing a real-time fine dust concentration calculation model according to claim 1, characterized by performing a Pearson correlation analysis technique to analyze the above correlation. Claim 3 A method for establishing a real-time fine dust concentration calculation model according to claim 1, wherein the input variables include outdoor PM2.5 concentration, AHU supply and exhaust operating airflow, AHU set temperature, air purifier operating status, EHP airflow, and indoor humidity. Claim 4 A method for establishing a real-time fine dust concentration calculation model according to claim 1, characterized in that the machine learning is performed in a DNN model-based manner. Claim 5 A method for establishing a fine dust concentration calculation model according to claim 4, characterized in that the machine learning is performed through learning a data structure having two-dimensional input variables and one-dimensional output variables. Claim 6 A method for establishing a fine dust concentration calculation model according to claim 1, wherein the retraining step comprises a sliding window-based learning step that deletes some of the data prior to a specific point in time and adds new data. Claim 7 A method for establishing a fine dust concentration calculation model according to claim 1, characterized in that the relearning step is performed on a weekly basis. Claim 8 A step for collecting basic fine dust data including indoor and outdoor environments and HVAC data within the facility; a data analysis step for analyzing the correlation between fine dust concentration and input / output data regarding the collected basic fine dust data; a step for determining input variables (features) for machine learning; a machine learning step for constructing a dataset for the input variables and training a fine dust concentration calculation model; and A real-time fine dust concentration calculation model established by a learning model that includes a step of periodic retraining. Claim 9 A method for calculating indoor fine dust concentration in real time using the real-time fine dust concentration calculation model of paragraph 8.

Images