An indoor personnel state feature and multiple energy demand estimation method based on information mining

Abstract

The application discloses an indoor personnel state feature and a variety of energy demand estimation method based on information mining, and belongs to the field of building energy consumption accurate prediction.The application applies data mining technology, applies data expression to the influence of "people", and adds the simulation platform based on the traditional physical model, so that the problem that the traditional physical model cannot depict human behaviors is overcome.Meanwhile, the method can better perform model migration on different buildings, so that the data-driven model can better perform generalization.

Description

Technical Field

[0001] This invention relates to a method for estimating various energy demands within a building, and more particularly to a method for estimating various energy demands based on indoor occupant status characteristics and information mining, belonging to the field of accurate building energy consumption prediction. Background Technology

[0002] As people's living standards continue to improve, the comfort of indoor environments is becoming increasingly important, leading to the deepening application of HVAC systems in buildings. However, this has also brought about serious energy problems. Statistics show that building energy consumption accounts for 40% of global energy consumption, and 40% of the energy consumption in public buildings comes from HVAC systems. Therefore, energy-saving optimization in the design and operation phases of buildings is of great significance. Reasonable energy consumption forecasting is a crucial aspect of these measures, playing a key role in matching the supply and demand of various energy uses in public buildings and in the refined management of HVAC systems. However, existing research on energy demand forecasting often differs significantly from actual energy consumption levels, which prevents subsequent energy-saving renovations and refined management of HVAC systems in public buildings from achieving the desired results. The reasons for this are twofold: First, due to individual differences in adaptability, the thermal comfort zone based on a group is a broad range rather than a set point. An unreasonable set point will inevitably lead to excessive cooling or heating, resulting in unnecessary energy waste. Second, existing building energy demand forecasting does not consider "humans" as a variable in its modeling, completely disconnecting subjective feelings and random behavior from the forecasting process. Therefore, incorporating information mining of indoor occupant status characteristics into building energy demand forecasting plays a crucial role in improving forecast accuracy and maintaining indoor environmental comfort. Existing building energy consumption forecasting primarily employs two traditional models: data-driven models and physical models. Data-driven models use machine learning or deep learning algorithms to model energy demand using historical time-series data. Physical models rely on thermodynamic rules and the building's inherent characteristics to model building components and various systems, thereby outputting building energy consumption and the indoor environment. However, both models have their limitations and advantages. The former can capture nonlinear relationships in complex systems, but it often requires a large amount of historical data, places high demands on sensors, and has poor interpretability when describing the relationship between energy consumption and buildings, failing to generalize well outside the training range. While the latter reduces the data detection work required for the model, it requires manual input of people's schedules for simulating human behavior, making it impossible to guarantee randomness. This invention proposes an information mining-based method for estimating indoor occupant state characteristics and various energy demands. It aims to improve the accuracy of building energy consumption prediction by combining data-driven models and physical models to mine information on indoor occupant state characteristics, thus laying the foundation for supply and demand matching in multi-energy systems and refined control of HVAC systems. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for estimating indoor occupant status characteristics and various energy demands based on information mining, so as to improve the accuracy and dynamic changes of building energy consumption prediction.

[0004] The technical solution of this invention is as follows: A method for estimating indoor occupant status characteristics and various energy demands based on information mining. Utilizing data mining techniques, it integrates the characteristics of physical models and data-driven models to characterize the impact of human behavior and accurately predict energy consumption. The specific steps are as follows:

[0005] S1. Data Acquisition and Preprocessing:

[0006] S1.1 Data Acquisition

[0007] The collected data is divided into two parts: one part is personnel physical information based on the individual, and the other part is personnel movement information based on the individual's behavior. Personnel physical information includes: human skin temperature (T). skin The time people spend in the room (RST), the proportion of people staying in different rooms (RSP), and whether people are actually in the room. Also, the air conditioning temperature setpoint (T) is... set Indoor temperature T in Indoor relative humidity Rh in Outdoor temperature T out Outdoor relative humidity Rh out Precipitation conditions, solar radiation I s Outdoor wind speed V d Continuous variables such as the Air Quality Index (AQI) are recorded, along with categorical variables based on weekday conditions and time of day. Personnel activity information includes: window opening and closing, shading opening and closing, air conditioning controller operation, thermostat adjustment, and building energy consumption is also collected.

[0008] S1.2 Data Preprocessing Techniques

[0009] Data preprocessing mainly includes data cleaning, data integration, data transformation, and data reduction.

[0010] Furthermore, the data cleaning process includes:

[0011] For consecutively missing data, the tuples are deleted directly. For non-consecutive missing data, an exponentially weighted moving average method is used to fill in the blanks.

[0012] P t =w·P t-1 +(1-w)·X t (1)

[0013]

[0014] Where P t P is the moving average at the current time. t-1 P is the moving average value from the previous time step. biased_t X is the corrected moving average when t is a smaller value than t, where t is the time. tis the current observation value; w is the decay weight, ranging from 0 to 1, which controls the model update speed. The larger the value, the more stable the model tends to be. It is taken as a constant close to 1.

[0015] The K-Means algorithm is used for outlier identification in abnormal data: the class label of a sample is unknown, and samples are classified based on their features. Convergence is achieved by minimizing the sum of Euclidean distances from each sample to its class center.

[0016]

[0017]

[0018]

[0019] Where x (i) For i training samples, μ j For j cluster centers, c (i) Let J be the class with the smallest Euclidean distance from the sample to the class center. Equation (4) is the update of the sample center, i.e., the sample mean of sample j. Let J(c,μ) be the class center with the smallest Euclidean distance to its class center, and let J(c,μ) be the loss function. Equation (5) is the sum of squared Euclidean distances from all samples to their class centers. Let J(c,μ) be the class center with the smallest Euclidean distance to its class center, and let J(c,μ) be the loss function. Let J(c,μ) be the sum of squared Euclidean distances from all samples to their class centers, (i) .

[0020] Furthermore, the aforementioned data integration and data transformation:

[0021] Redundancy is eliminated and avoided by calculating the Pearson correlation coefficient:

[0022]

[0023] Where, r A,B Here, A[n] represents the number of data points in class A, and B[n] represents the number of data points in class B, where n represents the number of data points. σ is the sample mean. A Let σ be the unbiased standard deviation of vector A. B Let be the unbiased standard deviation of vector B.

[0024] The data transformation employs the min-max normalization method, which is shown below:

[0025]

[0026] Furthermore, the data reduction described above:

[0027] By employing feature selection from dimensionality reduction, the optimal subset is selected for backward selection, thereby improving algorithm efficiency and optimizing the data-driven model.

[0028] Finally, the data is split into training and test sets according to a set ratio, which are used for subsequent model building and evaluation, respectively.

[0029] S2. Establishment of the personnel model:

[0030] S2.1 Establishment of Personnel Movement Model

[0031] A personnel movement model is established based on Markov chains and event mechanisms. The random transition process of Markov chains is analogous to the random movement of people. Based on the current position of indoor personnel and the transition matrix P, the position of personnel at the next moment is predicted based on conditional probability. This reveals the movement process of different personnel in the building space and describes the indoor situation of personnel. The model is shown below:

[0032]

[0033]

[0034]

[0035] P(RST a =k)=P aa k (1-P aa (11)

[0036] Among them, P ab For the transition probability, Let k represent the movement of people, and a and b represent different spatial states. RST a Let P(RST) be the time spent in state a, k be the time step, and P(RST) be the time spent in state a. a Let (k) be the probability that a Markov chain starts in state a and leaves state a after k time steps. When a person moves at time k, they must move from one state a to another state b. A value of 1 indicates a movement event has occurred; otherwise, no movement has occurred and the state remains unchanged.

[0037] S2.2 Establishment of Personnel Behavior Model

[0038] First, a correlation test is used to measure the correlation between the binary variable and the continuous variable. The correlation test uses the Point-biserial correlation coefficient test method, and the specific solution method is as follows:

[0039]

[0040] Where, r pb The correlation coefficient is... Let m1 and m0 be the means of the continuous variable when the binary variable is 1 and 0, respectively, and m be the frequency of the binary variable when the value is 1 and 0, respectively. Y represents the standard deviation of a continuous variable (where n is the number of continuous variables, Y is the standard deviation of the continuous variable). i For continuous variable sample values, (The mean of a continuous variable).

[0041] After assessing the correlation between different dependent and independent variables, a behavioral model is established for behavior prediction. The Support Vector Machine (SVM) algorithm is used to train the model. In the SVM algorithm, different kernel functions map data points into a high-dimensional space, and an optimal hyperplane is constructed within this space. This hyperplane can maximally separate data points of different categories and minimize classification error. The principle is as follows:

[0042] ω T x+β=0 (13)

[0043]

[0044]

[0045] st y g (ω T x g +β)≥1-ξ g (16)

[0046] ξ g ≥0, g=1,2,…,n (17)

[0047] Where x is the sample feature vector, ω is the normal vector, β is the displacement term, and γ is the interval. To optimize the objective and minimize it, C is a penalty parameter used to balance classification error and model complexity, ξ g The slack variable is used to allow some samples to be misclassified or very close to the hyperplane, and g is the subscript, i.e., the total number of samples is n.

[0048] S3. Establishment of the hybrid model

[0049] The hybrid model is built by overlaying a physical model and a data-driven model. The physical simulation model serves as the foundation, providing building thermal parameters and weather conditions. Human movement and behavior models are input into the physical model. The physical model provides the foundation, while the data-driven model provides human-related random variables. Finally, the integrated hybrid model predicts the building's energy demand over a future period.

[0050] S4, Performance Evaluation of Hybrid Model

[0051] The predicted energy consumption needs to be compared with the original data to determine the accuracy of the prediction. The hybrid model is evaluated using Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Coefficient of Variation (CV) to determine if it meets the accuracy prediction criteria. The specific calculation formulas are as follows:

[0052]

[0053]

[0054]

[0055] in, Represents the predicted value, y i This represents the true value. The closer the MAPE is to 0%, the more perfect the mixture model; the smaller the RMSE, the better the calculation result; the smaller the CV, the smaller the dispersion between the predicted value and the actual value, and the more successful the mixture model.

[0056] This invention utilizes data mining techniques to express the influence of "human beings" through data and incorporates it into a simulation platform based on traditional physical models, overcoming the problem that traditional physical models cannot characterize human behavior. Furthermore, this method allows for better model transfer across different buildings, enabling data-driven models to generalize more effectively.

[0057] Compared with existing technologies, the beneficial effects of this invention are: it provides a method for estimating indoor occupant status characteristics and various energy demands based on information mining, integrates the characteristics of physical models and data-driven models, incorporates human influence into the model building, and makes the prediction results more accurate. Attached Figure Description

[0058] Figure 1 This is a framework diagram for an indoor occupant status characteristic and multiple energy demand estimation method based on information mining.

[0059] Figure 2 This is a flowchart of a method for estimating indoor occupant status characteristics and various energy demands based on information mining.

[0060] Figure 3 This is a schematic diagram of data acquisition.

[0061] Figure 4 This is a diagram illustrating the principle of random movement of people, analogous to a Markov chain.

[0062] Figure 5 A framework diagram of the principle of personnel behavior model Detailed Implementation

[0063] The specific embodiments of the present invention are described in detail below with reference to the invention content, accompanying drawings, and formulas.

[0064] Reference Figure 2 This invention is a data-driven method for estimating the lag time of HVAC systems based on an information theory framework. Taking the energy demand forecasting of a public building in Dalian as an example, the specific steps are as follows:

[0065] S1. Data Acquisition and Preprocessing Techniques: Data acquisition and preprocessing techniques, the specific steps are as follows:

[0066] S1.1 Data Acquisition

[0067] Reference Figure 3 Data collection was conducted, and the data was divided into two parts: one part was personnel physical information based on the individuals themselves, and the other part was personnel movement information based on their behavior. The collection of personnel physical information utilized intelligent infrared sensor technology to measure human skin temperature (T0). skin The following data were collected: Personnel dwell time (RST) and the proportion of personnel staying in different rooms (RSP), including whether personnel were actually in the room. Simultaneously, the air conditioning temperature setpoint (T) was monitored using instruments. set Indoor temperature T in Indoor relative humidity Rh in Outdoor temperature T out Outdoor relative humidity Rh out Precipitation conditions, solar radiation I s Outdoor wind speed V d Continuous variables such as the Air Quality Index (AQI) and categorical variables such as weekday conditions and time periods are recorded to prepare for the correlation testing and prediction of human behavior factors in subsequent data-driven models. Human action information includes: window opening and closing, shading opening and closing, air conditioning controller opening and closing, and thermostat adjustment. Building energy consumption is also collected to facilitate the establishment of subsequent data-driven models.

[0068] S1.2 Data Preprocessing Techniques

[0069] Due to inherent characteristics of the data or objective conditions during data collection, the data itself may contain problems that prevent it from meeting the needs of subsequent data mining. Therefore, preprocessing is necessary to ensure the collected data conforms to standards. Data preprocessing mainly includes data cleaning, data integration, data transformation, and data reduction.

[0070] S1.2.1 Data Cleaning

[0071] Due to numerous uncontrollable factors, data may be missing to varying degrees in both spatial and temporal dimensions, necessitating data cleanup. Since the original operational status and environmental conditions are unknown, data imputation is not feasible. Therefore, for consecutively missing data, tuples are directly deleted. For non-consecutive missing data, an exponentially weighted moving average method is used to fill in the blanks.

[0072] P t =w·P t-1 +(1-w)·X t (1)

[0073]

[0074] Where P t P is the moving average at the current time. t-1 P is the moving average value from the previous time step. biased_t X is the corrected moving average when t is a smaller value than t, where t is the time. t is the current observation value; w is the decay weight, ranging from 0 to 1, which controls the model update speed. The larger the weight, the more stable the model tends to be. It is a constant close to 1, such as 0.99 or 0.999.

[0075] The K-Means algorithm is used for outlier identification in abnormal data. K-Means clustering is a type of unsupervised learning where the class label of a sample is unknown. It classifies samples based on their features and converges by minimizing the sum of Euclidean distances from each sample to its class center.

[0076]

[0077]

[0078]

[0079] Where x (i) For i training samples, μ j For j cluster centers, c (i) Let J be the class with the smallest Euclidean distance from the sample to the class center. Equation (4) is the update of the sample center, i.e., the sample mean of sample j. Let J(c,μ) be the class center with the smallest Euclidean distance to its class center, and let J(c,μ) be the loss function. Equation (5) is the sum of squared Euclidean distances from all samples to their class centers. Let J(c,μ) be the class center with the smallest Euclidean distance to its class center, and let J(c,μ) be the loss function. Let J(c,μ) be the sum of squared Euclidean distances from all samples to their class centers, (i) .

[0080] S1.2.2 Data Integration and Data Transformation

[0081] Due to the different sources of data, the data may exist in different formats and expressions. Therefore, it is necessary to unify the data processing and transform it into a consistent form. Data redundancy and data transformation are the two most important issues. Data redundancy can be eliminated and avoided by analyzing correlations and calculating the Pearson correlation coefficient.

[0082]

[0083] Where, r A,B Here, A[n] represents the number of data points in class A, and B[n] represents the number of data points in class B, where n represents the number of data points. σ is the sample mean. A Let σ be the unbiased standard deviation of vector A. B Let be the unbiased standard deviation of vector B.

[0084] The data transformation primarily employs the min-max normalization method, mapping variables such as air conditioner setpoint temperatures to the [0,1] interval, thereby normalizing their magnitude and simplifying the subsequent establishment and operation of data-driven models. The min-max normalization method is shown below:

[0085]

[0086] S1.2.3 Data Reduction

[0087] Because data often contains many attributes and variables irrelevant to the task, data mining on the entire dataset would be extremely time-consuming and costly. Therefore, data reduction is used to simplify the representation of the original dataset while achieving similar mining results to the entire dataset. This invention utilizes feature selection from dimensionality reduction to select the optimal subset and perform backward selection, improving algorithm efficiency and thus optimizing the data-driven model.

[0088] Finally, the data was split randomly in an 8:2 ratio, with 80% used as the training set and 20% as the test set, for subsequent model building and evaluation.

[0089] S2. Establishment of Personnel Models: The establishment of personnel movement models and personnel behavior models, with specific steps as follows:

[0090] S2.1 Establishment of Personnel Movement Model

[0091] Reference Figure 4The personnel movement model of this invention is based on Markov chains and event mechanisms. It uses the random transition process of a Markov chain as an analogy to the random movement of people. Based on the current position of indoor personnel and the transition matrix P, it predicts the position of personnel at the next moment based on conditional probability, revealing the movement process of different personnel in the building space and characterizing their indoor situation. The model is shown below:

[0092]

[0093]

[0094]

[0095] P(RST a =k)=P aa k (1-P aa (11)

[0096] Among them, P ab For the transition probability, Let k represent the movement of people, and a and b represent different spatial states. RST a Let P(RST) be the time spent in state a, k be the time step, and P(RST) be the time spent in state a. a Let (k) be the probability that a Markov chain starts in state a and leaves state a after k time steps. When a person moves at time k, they must move from one state a to another state b. A value of 1 indicates a movement event has occurred; otherwise, no movement has occurred and the state remains unchanged.

[0097] S2.2 Establishment of Personnel Behavior Model

[0098] Reference Figure 5 Establishing a human behavior model is crucial for predicting future human behavior. The aforementioned human behaviors influencing building energy consumption primarily include window opening and closing, shading opening and closing, air conditioning controller operation, and thermostat adjustment. The factors influencing these behaviors include psychological, physiological, environmental, and customary factors. Therefore, multiple variables across time and space influence human behavior, with varying degrees of influence. Thus, correlation testing of multiple variables is essential. The Point-biserial correlation coefficient test is employed, effectively measuring the correlation between binary and continuous variables. The specific solution method is as follows:

[0099]

[0100] Where, rpb The correlation coefficient is... Let m1 and m0 be the means of the continuous variable when the binary variable is 1 and 0, respectively, and m be the frequency of the binary variable when the value is 1 and 0, respectively. Y represents the standard deviation of a continuous variable (where n is the number of continuous variables, Y is the standard deviation of the continuous variable). i For continuous variable sample values, (The mean of a continuous variable).

[0101] After assessing the correlation between different dependent and independent variables, a behavioral model is established for behavior prediction. The Support Vector Machine (SVM) algorithm is used to train the model. The core idea of ​​SVM is to map data points into a high-dimensional space using different kernel functions and construct an optimal hyperplane within that space. This hyperplane can maximally separate data points of different categories and minimize classification error. The principle is as follows:

[0102] ω T x+β=0 (13)

[0103]

[0104]

[0105] st y g (ω T x g +β)≥1-ξ g (16)

[0106] ξ g ≥0, g=1,2,…,n (17)

[0107] Where x is the sample feature vector, ω is the normal vector, β is the displacement term, and γ is the interval. To optimize the objective and minimize it, C is a penalty parameter used to balance classification error and model complexity, ξ g The slack variable is used to allow some samples to be misclassified or very close to the hyperplane, and g is the subscript, i.e., the total number of samples is n.

[0108] S3. Establishment of the hybrid model

[0109] The hybrid model is built by overlaying a physical model and a data-driven model. The physical simulation model serves as the foundation, providing building thermal parameters and weather conditions. Human movement and behavior models are input into the physical model. The physical model provides the foundation, while the data-driven model provides human-related random variables. Finally, the integrated hybrid model predicts the building's energy demand over a future period.

[0110] S4. Model Performance Evaluation

[0111] The predicted energy consumption needs to be compared with the original data to determine the accuracy of the prediction. The model is evaluated using Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Coefficient of Variation (CV) to determine if it meets the accuracy prediction criteria. The specific calculation formulas are as follows:

[0112]

[0113]

[0114]

[0115] The predicted building energy demand was compared with actual data, and the evaluation criteria were calculated using the above formula. The MAPE values ​​for each room were 0.860%, 1.360%, 0.530%, and 0.420%, respectively; the RMSE values ​​for each room were 1.322 kWh, 0.541 kWh, 0.687 kWh, and 0.556 kWh, respectively; and the CV values ​​for each room were 7.211%, 5.257%, 6.534%, and 5.448%, respectively. Therefore, it can be concluded that the MAPE values ​​are all below 2%, the RMSE values ​​are all less than 2 kWh, and the CV values ​​are all controlled below 8%, indicating that the model fitting effect for each room meets the requirements for accurate prediction.

Claims

1. An information mining-based indoor personnel state feature and multiple energy demand estimation method, characterized in that, The specific steps are as follows: S1, data acquisition and data preprocessing: S1.1, data acquisition The collected data is divided into two parts, one part is the personnel physical information based on the personnel themselves, and the other part is the personnel action information based on the personnel behavior; The personnel body information includes: human skin temperature , personnel staying time in the room , personnel staying proportion in different rooms , and whether the personnel are in the room; meanwhile, the air conditioner temperature setting value , indoor temperature , indoor relative humidity , outdoor temperature , outdoor relative humidity , precipitation condition, solar radiation intensity , outdoor wind speed , air quality index These continuous variables and working day condition, time period classification variables are recorded; the personnel action information includes: window opening and closing condition, sunshade opening and closing condition, air conditioner controller opening and closing condition, thermostat adjustment condition, and meanwhile, the building cooling, heating and power consumption are collected; S1.2, data preprocessing technology Data preprocessing mainly includes data cleaning, data integration, data conversion and data reduction; S2, establishment of personnel model: S2.1, establishment of personnel movement model The personnel movement model is established based on Markov chain and event mechanism. The random jump process of Markov chain is used to analogize the random movement of people. According to the current position of indoor personnel and the transition matrix P, the position of personnel at the next time is predicted based on conditional probability, the movement process of different personnel in building space is revealed, and the in-room situation of personnel is described. The model is shown as follows: ； wherein, is the transition probability, is the time at which a movement of the person occurs, , is a different spatial state, is the time spent in state, is the time step, is the probability that the Markov chain leaves state after time steps, starting in state ; when a movement of the person occurs at time , it must move from one state point to another state point , then at this time is 1, the movement event occurs, otherwise the person has not moved and the state remains unchanged; S2.2, establishment of personnel behavior model Firstly, the correlation test is used to measure the correlation between the binary variable and the continuous variable. The Point-biserial correlation coefficient test method is used for correlation test. The specific solving method is as follows: ； wherein, is the correlation coefficient, , is the mean of the continuous variable when the dichotomous variable is 1, 0, , is the frequency of the dichotomous variable when it is 1, 0, respectively, is the number of variables of the dichotomous variable, is the standard deviation of the continuous variable, wherein is the number of variables of the continuous variable, is the sample value of the continuous variable, is the sample mean of the continuous variable; After measuring the correlation between different dependent variables and independent variables, the personnel behavior model is established, and the behavior prediction is carried out. The support vector machine algorithm SVM is used for model training. In the SVM algorithm, different kernel functions are used to map data points to high-dimensional space, and an optimal hyperplane is constructed in the space. This hyperplane can maximize the separation of different categories of data points and minimize the classification error. The principle is as follows: ； where is the sample feature vector, is the normal vector, is the displacement term, is the margin, is the optimization objective, minimized, is the penalty parameter used to balance misclassification error and model complexity, is the slack variable used to allow certain samples to be misclassified or close to the hyperplane, is the index, i.e. the total number of samples is ; S3, establishment of hybrid model The hybrid model is established by superimposing the physical model and the data-driven model. The building thermal parameters and weather conditions are provided based on the physical simulation model. The personnel movement model and the personnel behavior model are input into the physical model. The physical model provides the basis, and the data-driven model provides the random variables related to people. Finally, the integrated hybrid model is used to predict the energy demand of the building in the future period; S4, performance evaluation of hybrid model The predicted energy consumption needs to be compared with the original data to determine the prediction situation. The mean absolute percentage error MAPE, the root mean square error RMSE and the coefficient of variation CV are used to evaluate the hybrid model and determine whether it meets the accurate prediction. The specific calculation formula is as follows: ； wherein, represents the predicted value, represents the true value; the closer the MAPE to 0%, the more perfect the mixture model; the smaller the RMSE, the better the calculation result; the smaller the CV, the smaller the degree of dispersion between the predicted value and the actual value, and the more successful the mixture model. 2.The method of estimating indoor occupant state features and multiple energy use demands based on information mining according to claim 1, wherein, The data cleaning: For continuous missing data, the tuple is directly deleted. For non-continuous missing data, the exponential weighted moving average method is used to fill in the blank values: ； wherein is the current time moving average, is the previous time moving average, is the is the modified moving average with a small correction, is the time, is the observation value at the current time; is the decay weight, ranging between 0-1, controlling the model update speed, the larger the more stable, take a constant close to 1; The outlier identification of abnormal data uses the K-Means algorithm: the class label of the sample is unknown, the sample is classified by the sample features, and the sum of the Euclidean distances of the samples to the class centers is minimized, and finally the convergence is achieved: ； wherein is a training sample, is a cluster center, is the class with the smallest Euclidean distance from the sample to the class center, and formula (4) is the update of the sample center, i.e., the sample is the sample mean, is the class center with the smallest Euclidean distance, is a loss function, and formula (5) is the sum of the squared Euclidean distances of all samples to their class centers, so that it reaches convergence, thereby determining the optimal . 3.The method of claim 1, wherein, The data integration and data conversion: Pears correlation coefficient is calculated to eliminate and avoid redundancy: ； wherein, is a correlation coefficient, denotes data sharing , denotes data sharing , denotes the number of data, , is a sample mean, is an unbiased standard deviation of the A vector, is an unbiased standard deviation of the B vector; The data conversion uses the data minimum-maximum normalization method, and the minimum-maximum normalization method is shown as follows: ； 4.The method of claim 1, wherein, The data reduction: Feature selection in dimension reduction is used to select the best subset, perform backward selection, improve the algorithm efficiency, and optimize the data-driven model; Finally, the data is split, and the data is split into training set and test set in proportion at random, respectively used for subsequent model establishment and evaluation.

Images