An intelligent control system for optimizing energy consumption of central air conditioners based on big data analysis

Abstract

The present invention belongs to the technical field of intelligent control of air conditioners, and discloses an intelligent control system for optimizing energy consumption of central air conditioners based on big data analysis; the system comprises the following steps: collecting comprehensive operation data of central air conditioners; preprocessing the collected comprehensive operation data to obtain low-dimensional characteristic variables; clustering the low-dimensional characteristic variables to obtain several typical operation modes; constructing an operation directed graph model based on several typical operation modes; based on the operation directed graph model, using an improved partitioning algorithm to solve and obtain an optimal control strategy; applying the optimal control strategy to the real-time control of central air conditioners to achieve energy consumption optimization, thereby being able to guide the operation of air conditioners more scientifically and reasonably and avoiding the limitations of artificial rules of thumb; and significantly improving the energy-saving performance of air conditioning systems while taking into account comfort requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of air-conditioning intelligent control, and more specifically, to a central air-conditioning energy consumption optimization intelligent control system based on big data analysis. Background Art

[0002] The patent application with publication number CN116027700A discloses an intelligent optimization method and system based on central air-conditioning big data analysis, including a data acquisition module, a data processing module, an equipment model database module, a load and energy consumption database module, a global optimization module, and a control module; the Pearson correlation coefficient method is used for data analysis and processing, and a relationship model between the controllable variables and the uncontrollable variables of the central air-conditioning system is established, so that the intelligent optimization system can run the optimal working plan under different working conditions according to the changes in working conditions, so that the total energy consumption of the system is minimized and the system efficiency is optimized, thereby saving the energy consumption of the central air-conditioning system as much as possible without reducing the indoor comfort.

[0003] However, the existing control systems still lack the full utilization of a large amount of operating data and are unable to effectively mine potential laws and knowledge from massive data. With the development of building intelligence and Internet of Things technology, the operating data generated by central air-conditioning systems are characterized by large volume, multi-dimensionality and heterogeneity. It is difficult to accurately describe and control complex operating processes by relying solely on limited manual experience and knowledge. Secondly, most of the existing control strategies are based on simple rules or empirical models and lack the ability to adapt to complex and changeable operating environments. The operation of central air-conditioning systems is affected by many factors, and the interaction of influencing factors makes the system operation process highly dynamic and uncertain. Control strategies based on empirical models are difficult to fully characterize this complexity, resulting in limited control performance. Furthermore, existing control systems lack comprehensive balance considerations. Excessive pursuit of energy saving often sacrifices comfort, while excessive pursuit of comfort leads to increased energy consumption. In addition, existing control methods lack the analysis and utilization of the dynamic evolution laws of operation. The operating state of the central air-conditioning system will evolve dynamically over time. How to predict and grasp this evolution trend is crucial to the formulation of forward-looking control strategies, but existing methods are still insufficient in this regard.

[0004] In view of this, the present invention proposes a central air-conditioning energy consumption optimization intelligent control system based on big data analysis to solve the above problems. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: an intelligent control system for optimizing energy consumption of central air conditioners based on big data analysis, comprising: a data acquisition and processing module, a clustering module, a model building module and a strategy solving module; each module is connected by wire and / or wireless means;

[0006] The data acquisition and processing module includes an acquisition unit and a processing unit. The acquisition unit is used to acquire the comprehensive operation data of the central air conditioner; the processing unit is used to pre-process the acquired comprehensive operation data to obtain low-dimensional feature variables;

[0007] Clustering module, used to cluster low-dimensional feature variables and obtain several typical operation modes;

[0008] Model building module, which builds and runs directed graph models based on several typical operation modes;

[0009] The strategy solving module includes a solving unit and a control unit. The solving unit uses an improved partitioning algorithm to solve the optimal control strategy based on the running directed graph model; the control unit is used to apply the optimal control strategy to the real-time control of the central air conditioner.

[0010] Furthermore, the comprehensive operation data includes indoor environment data, outdoor environment data, air conditioning operation data and building data; the indoor environment data includes indoor temperature, indoor humidity and carbon dioxide concentration; the outdoor environment data includes outdoor temperature, outdoor humidity and solar radiation intensity;

[0011] Air conditioning operation data includes unit operating status, supply and return water temperature, supply and return water flow, compressor operation frequency and energy consumption per unit time; building data includes building orientation and building insulation rate.

[0012] Furthermore, the preprocessing method includes data cleaning, data standardization and data dimension reduction; the data dimension reduction method includes:

[0013] The comprehensive operation data after data cleaning and data standardization are recorded as the operation data set; variables are extracted from the operation data set to form a data matrix X, and the correlation coefficient matrix R of the data matrix X is calculated. The element R(i,j) in the correlation coefficient matrix R represents the correlation coefficient between the i-th variable and the j-th variable;

[0014] Based on the correlation coefficient matrix, calculate the sample fitness value KM of the running data set;

[0015] ; Where r_ij is the correlation coefficient between the ith variable and the jth variable, w_ij is the weight coefficient between the ith variable and the jth variable, a_ij is the partial correlation coefficient between the ith variable and the jth variable, N is the amount of data in the running data set, and p is the number of variables; f(N,p) is the partial correlation coefficient adjustment function;

[0016] Partial correlation coefficient adjustment function ; Among them, α is a parameter that controls the curvature of the function; β is the ratio influencing parameter;

[0017] A dimension reduction analysis threshold is preset. If the sample fitness value KM of the running data set is less than the dimension reduction analysis threshold, the data in the running data set is readjusted; the sample fitness value of the running data set is recalculated until the sample fitness value KM is greater than or equal to the dimension reduction analysis threshold; the running data set at this time is retained as the optimal data set, and the correlation coefficient matrix corresponding to the optimal data set is calculated and obtained; recorded as the optimal correlation coefficient matrix; the number of retained factors is obtained based on the optimal correlation coefficient matrix;

[0018] According to the number of retained factors, construct the initial factor matrix, in which each column corresponds to a factor and each row corresponds to a variable;

[0019] The initial factor matrix is ​​orthogonally rotated to obtain the factor loading matrix; based on the factor loading matrix, the comprehensive scores of several samples on each factor are calculated; one sample corresponds to a set of comprehensive operation data;

[0020] The formula for calculating the comprehensive score is:

[0021] ; Where FZ_IJ is the comprehensive score of the I-th sample on the J-th factor, w'_JS is the weight coefficient of the J-th factor and the S-th variable in the sample; x_IS is the standardized value of the S-th variable of the I-th sample, ε_J is the eigenvalue corresponding to the J-th factor, and a_JS is the loading value of the S-th variable on the J-th factor;

[0022] The comprehensive scores of all samples on the Jth factor are weighted averaged to obtain the factor score of the Jth factor; a score threshold is preset, and factors with factor scores greater than or equal to the score threshold are retained as main factors, and the main factors are used as low-dimensional feature variables.

[0023] Furthermore, the method for obtaining the number of retained factors based on the optimal correlation coefficient matrix includes: generating several groups of random data with the same data volume and number of variables as the optimal data set, performing feature analysis on each group of random data, and extracting eigenvalues; calculating the mean and 95% percentile of each eigenvalue, and calculating the standard deviation of the eigenvalue; performing feature analysis on the optimal data set at the same time, and recording the obtained eigenvalues ​​as constrained eigenvalues;

[0024] When the constraint eigenvalue is greater than the comparison eigenvalue MK, the factor corresponding to the constraint eigenvalue is retained;

[0025] Comparison of feature values ; where μ is the mean of the eigenvalues, σ is the standard deviation of the eigenvalues, is the 95% percentile of the eigenvalue, α1 is an adjustable parameter;

[0026] The number of factors is defined as the serial number of the factor corresponding to the last eigenvalue retained, the explanatory power of the retained factors is calculated, and the explanatory threshold is preset. If the explanatory power is less than the explanatory threshold, the number of retained factors is increased or decreased, and the number of increases or decreases is 1; until the explanatory power of the retained factors is greater than or equal to the explanatory threshold.

[0027] Furthermore, the method of performing feature analysis on each group of random data includes:

[0028] Calculate the correlation coefficient matrix of random data, denoted as random correlation coefficient matrix M, and find the largest eigenvalue γ1 and its corresponding eigenvector a1 from M;

[0029] Calculate the explained variance contribution rate c1 of the factor F1 corresponding to the eigenvector a1; preset the factor retention threshold, if c1 is less than the factor retention threshold, abandon the factor F1, and modify M to obtain a new random correlation coefficient matrix M';

[0030] If c1 is greater than or equal to the factor retention threshold, factor F1 is used as the first factor, and the corresponding variables in M ​​and the variance of F1 are eliminated from M to obtain the residual matrix M1; the maximum eigenvalue γ2 and the corresponding eigenvector a2 are found in M1, and the explained variance contribution rate c2 of factor F2 corresponding to eigenvector a2 is calculated; based on the factor retention threshold, it is also determined whether to retain factor F2; if retained, factor F2 is used as the second factor;

[0031] Repeat until all factors are extracted or the remaining eigenvalues ​​are less than the preset eigenvalue threshold; the eigenvalues ​​corresponding to all extracted factors are the eigenvalues ​​obtained by feature analysis;

[0032] Ways to make corrections include:

[0033] Define a penalty matrix P, the diagonal elements of the penalty matrix P are 1, and the off-diagonal elements are -ω; where ω is a hyperparameter with a value range of [0, 1]; then calculate the new random correlation coefficient matrix ; Where t' is the number of judgment steps.

[0034] Furthermore, the step of performing cluster analysis includes:

[0035] Step 1: Sample the low-dimensional feature variables and extract some data points as representative point sets;

[0036] Step 2: For each data point p in the representative point set, calculate the distance between it and all other representative points, find a fixed number of representative points closest to p, and form a cluster with these representative points and p;

[0037] Step 3: For each formed cluster, calculate the center point of the cluster and calculate the sum of the squares of the distances from all data points in the cluster to the center point; select the cluster with the smallest sum of squares of distances and use its center point as the first true cluster center;

[0038] Step 4: For the remaining clusters, repeatedly select the cluster with the smallest sum of squared distances, and use its center point as the second true cluster center until kl true cluster centers are obtained; for each data point in the data set, calculate its minimum distance to the kl true cluster centers, and assign the corresponding data point to the cluster with the closest minimum distance;

[0039] Step 5: For each cluster, calculate the sum of the distances from all its data points to the true cluster center as the scatter parameter of the cluster; select cl clusters with the largest scatter parameters;

[0040] Step 6: For the selected cl clusters, perform steps 2 to 4 respectively to obtain new clusters; repeat steps 5 and 6 until the number of clusters reaches the preset cluster value; each cluster finally obtained corresponds to a typical operation mode.

[0041] Furthermore, the data points are extracted in the following manner:

[0042] Grid the entire data set, that is, divide the data space where the data set is located into a number of equal grid cells, and calculate the number of data points in each grid cell as the density value of the grid cell;

[0043] Calculate the density value range [MI_min, MI_max] of the entire data set, where MI_min is the minimum density value and MI_max is the maximum density value; divide the density value range into several density levels;

[0044] Define the mapping function of the number of representative points ; Where d_T is the average density value of the Tth density level, ω1, ω2, ω3 and θ are mapping adjustable parameters;

[0045] The total number of representative points is preset to N1. For the Tth density level, the density value range is counted; and the average density value d_T of the Tth density level is calculated; based on the corresponding average density value, the number of representative points of the Tth density level is set to ;

[0046] A corresponding number of representative points are extracted according to the density level; for the extracted representative points, the average distance to other representative points is calculated, and a distance threshold is preset. Representative points whose average distance is greater than the distance threshold are treated as noise points and removed.

[0047] Furthermore, the method of constructing and running the directed graph model includes:

[0048] Take each cluster as a node, fix a time window of size tw, and for each data point in cluster f, find the cluster f' to which its next data point in the time window belongs; count the number of times N_ff' that all data points in cluster f are transferred to cluster f' in the time window;

[0049] Calculating transition probabilities ; where N_f is the total number of data points in cluster f, is a smoothing parameter greater than 0; K' is the total number of clusters;

[0050] The transition probability p_ff' between each pair of clusters (f, f') is used as the weight of the directed edge, that is, a directed edge is connected from cluster f to cluster f', and the weight of the directed edge is p_ff';

[0051] For two clusters with a transition probability of 0, no edge is connected or an edge with a preset weight is connected; finally, a running directed graph model with K' nodes is constructed.

[0052] Furthermore, the solution method of the optimal control strategy includes:

[0053] Each node in the running directed graph model is regarded as a corresponding state in the state space, and the preset control action is regarded as the action space; the state transition probability is given by the weight of the directed edge;

[0054] Define an instant cost function, the formula of the instant cost function is:

[0055] ; Where ce(s,a) represents the energy consumption when executing control action a in state s, δ1 is the penalty coefficient; cc(s,a) is the impact value on comfort when executing control action a in state s; θ is the comfort threshold; C(s,a) is the immediate cost when executing control action a in state s; is the indicator function;

[0056] For each state s, define its optimal value function V(s) at infinite time; solve the optimal control strategy based on the optimal value function V(s);

[0057] ;

[0058] in, To find a strategy ρ among all strategies so that its corresponding expected value E_ρ is minimized; C(s_t,a_t) is the immediate cost when executing the control action a_t in the state s_t; s_t is the state at time t, a_t is the control action executed at time t; bu(t) is the time-varying weight function; fp(s_t,a_t) is the constraint function; s_0 represents the initial state, s_0=s represents the initial state s; ∈ is the discount factor.

[0059] Furthermore, the method of solving the optimal control strategy based on the optimal value function V(s) includes:

[0060] For each state s, solve its optimal value function V(s), and express V(s) as the convex combination of V(s') of its adjacent state s';

[0061] Right now ; Where max_{s',s} represents the maximum value of all states s' adjacent to state s; C(s,ρ(s)) represents the immediate cost of executing the control action ρ(s) given by strategy ρ in state s; δ3 is the weight discount factor; V(s') is the optimal value function of state s'; solve the coefficients of this convex combination through linear programming method, that is, get V(s), solve V(s) for all states; get the optimal control strategy by inverse solution of V(s) ; Among them, argmin_a[] means taking the minimum value for all actions a; γ3 is the balance weight; pu(s'|s,a) means the probability of transferring to state s' after executing control action a in state s, that is, the state transition probability.

[0062] The technical effects and advantages of the central air-conditioning energy consumption optimization intelligent control system based on big data analysis of the present invention are as follows:

[0063] The present invention conducts in-depth mining of massive operating data, automatically identifies typical operating modes, and does not require manual modeling, so it can fully adapt to complex and changeable operating environments. At the same time, it constructs an operating directed graph model to characterize the dynamic transfer rules between operating modes, laying the foundation for formulating forward-looking control strategies and effectively improving control performance. Secondly, it combines the energy consumption minimization goal with the comfort constraint, and solves the optimal control strategy based on the improved partitioning algorithm to seek the best balance between energy saving and comfort, so as to achieve both energy saving and comfort. In the solution process, the solution efficiency is improved by effectively processing large-scale state spaces. Furthermore, it can be applied to central air-conditioning systems of different types and sizes, providing strong support for achieving energy-saving control. In addition, the control strategy has a basis to rely on, and is based on the analysis and mining of massive data rather than subjective experience, so it can guide the air-conditioning operation more scientifically and reasonably, avoiding the limitations of artificial rules of experience. It can significantly improve the energy-saving performance of the air-conditioning system while taking into account comfort requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of an intelligent control system for optimizing energy consumption of central air conditioners based on big data analysis according to the present invention;

[0065] Figure 2 This is a schematic diagram of an intelligent control method for optimizing central air-conditioning energy consumption based on big data analysis according to the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Example 1

[0068] See also Figure 1 As shown, the central air-conditioning energy consumption optimization intelligent control system based on big data analysis described in this embodiment includes: a data acquisition and processing module, a clustering module, a model building module and a strategy solving module; each module is connected by wired and / or wireless means to realize data transmission between modules;

[0069] The data acquisition and processing module includes an acquisition unit and a processing unit. The acquisition unit is used to acquire the comprehensive operation data of the central air conditioner; the processing unit is used to pre-process the acquired comprehensive operation data to obtain low-dimensional feature variables;

[0070] Clustering module, used to cluster low-dimensional feature variables and obtain several typical operation modes;

[0071] Model building module, which builds and runs directed graph models based on several typical operation modes;

[0072] The strategy solving module includes a solving unit and a control unit. The solving unit uses an improved partitioning algorithm to solve the optimal control strategy based on the running directed graph model; the control unit is used to apply the optimal control strategy to the real-time control of the central air conditioner to achieve energy consumption optimization.

[0073] The comprehensive operation data includes indoor environment data, outdoor environment data, air conditioning operation data and building data; the collected comprehensive operation data is in several groups;

[0074] Indoor environmental data include indoor temperature, indoor humidity and carbon dioxide concentration; outdoor environmental data include outdoor temperature, outdoor humidity and solar radiation intensity, which are obtained using corresponding sensors;

[0075] Air conditioning operation data include unit operation status, supply and return water temperature, supply and return water flow rate, compressor operation frequency and energy consumption per unit time; unit operation status includes unit startup status, unit normal operation status, unit fault status, unit stop status, unit energy-saving mode status and defrost status (for heat pump units); obtained through experimental statistics; building data include building orientation (eight orientations, such as west, northwest, etc.) and building insulation rate (obtained by table lookup).

[0076] Furthermore, the preprocessing methods include data cleaning, data standardization (such as Z-score standardization) and data dimension reduction; in the process of data standardization, the numerical data is normalized and the categorical data is encoded.

[0077] The data cleaning method includes missing value processing, outlier processing and noise data filtering: missing value processing is to process missing data points by interpolation, mean interpolation or deletion; outlier processing is to identify and eliminate abnormal data points through statistical analysis; noise data filtering is to apply filtering algorithms (such as Kalman filtering, wavelet transform, etc.) to remove noise in the comprehensive operation data.

[0078] Methods of data dimensionality reduction include:

[0079] The comprehensive operation data after data cleaning and data standardization is recorded as the operation data set;

[0080] Calculate the correlation coefficient matrix between the variables in the running data set (such as indoor temperature, indoor humidity, etc.). Specifically, extract the variables from the running data set to form the data matrix X, and calculate the correlation coefficient matrix R of the data matrix X. The element R(i,j) in the correlation coefficient matrix R represents the correlation coefficient between the i-th variable and the j-th variable. The correlation coefficient reflects the degree of linear correlation between the two variables, and its value range is [-1, 1].

[0081] Correlation coefficient ; where x_ip is the value of the i-th variable of the p-th running data set, μ_i is the mean of the i-th variable, x_jp is the value of the j-th variable of the p-th running data set, and μ_j is the mean of the j-th variable.

[0082] Based on the correlation coefficient matrix, calculate the sample fitness value KM of the running data set;

[0083] ; Among them, r_ij is the correlation coefficient between the ith variable and the jth variable, w_ij is the weight coefficient between the ith variable and the jth variable, a_ij is the partial correlation coefficient between the ith variable and the jth variable (reflecting the correlation coefficient after eliminating the influence of other variables), N is the amount of data in the running data set, and p is the number of variables; f(N,p) is the partial correlation coefficient adjustment function;

[0084] Partial correlation coefficient adjustment function ; Among them, α is a parameter that controls the curvature of the function, and its value range is (0, +∞); β is the ratio influence parameter, that is, a parameter that controls the influence of the ratio of sample size and number of variables, and its value range is (-∞, +∞).

[0085] Preset the dimensionality reduction analysis threshold (by analyzing historical data, calculate a reasonable sample fitness value range, and use the lower limit of the range as the dimensionality reduction analysis threshold). If the sample fitness value KM of the operating data set is less than the dimensionality reduction analysis threshold, readjust the data in the operating data set (such as changing a set of comprehensive operating data to re-clean and standardize the data); recalculate the sample fitness value of the operating data set until the sample fitness value KM is greater than or equal to the dimensionality reduction analysis threshold; retain the operating data set at this time as the optimal data set, calculate and obtain the correlation coefficient matrix corresponding to the optimal data set; record it as the optimal correlation coefficient matrix; obtain the number of retained factors based on the optimal correlation coefficient matrix;

[0086] Specifically, several groups of random data with the same data volume and number of variables as the optimal data set are generated, and each group of random data is subjected to feature analysis to extract eigenvalues; the mean and 95% percentile of each eigenvalue are calculated, and the standard deviation of the eigenvalue is calculated; at the same time, the optimal data set is subjected to feature analysis, and the eigenvalues ​​obtained are recorded as constrained eigenvalues.

[0087] When the constraint eigenvalue is greater than the comparison eigenvalue MK, the factor corresponding to the constraint eigenvalue is retained; the comparison eigenvalue ; where μ is the mean of the eigenvalues, σ is the standard deviation of the eigenvalues, is the 95% percentile of the eigenvalue, α1 is an adjustable parameter used to control the strictness of the retention factor; the mean, standard deviation and 95% percentile of the eigenvalues ​​of random data are used in the eigenvalue calculation process.

[0088] The number of factors is defined as the serial number of the factor corresponding to the last eigenvalue retained, and the explanatory power of the retained factors is calculated (the sum of the eigenvalues ​​corresponding to the retained factors divided by the sum of the total variance of the random data). The explanatory threshold is preset (usually set to 0.8-0.9, which requires that the retained factors can explain 80%-90% of the total variance). If the explanatory power is less than the explanatory threshold, the number of retained factors is increased or decreased, and the number of increases or decreases is 1; until the explanatory power of the retained factors is greater than or equal to the explanatory threshold.

[0089] The methods for performing feature analysis on each set of random data include:

[0090] Calculate the correlation coefficient matrix of random data, denoted as random correlation coefficient matrix M, find the largest eigenvalue γ1 and its corresponding eigenvector a1 from M; calculate the explained variance contribution rate c1 of factor F1 corresponding to eigenvector a1 (divide the eigenvalue γ1 by the total variance of all eigenvalues ​​of the random correlation coefficient matrix M).

[0091] Preset the factor retention threshold (set a lower limit, such as 0.05, that is, the explained variance contribution rate must be greater than 5% to retain the factor). If c1 is less than the factor retention threshold, factor F1 is abandoned, and M is modified to obtain a new random correlation coefficient matrix M'; if c1 is greater than or equal to the factor retention threshold, factor F1 is used as the first factor, and the corresponding variables in M ​​and the variance of F1 are eliminated from M to obtain the residual matrix M1; find the maximum eigenvalue γ2 and the corresponding eigenvector a2 in M1, and calculate the explained variance contribution rate c2 of factor F2 corresponding to eigenvector a2; based on the factor retention threshold, also determine whether to retain factor F2; if retained, factor F2 is used as the second factor; it should be noted that the "first" and "second" mentioned in the first factor and the second factor are the serial numbers of the factors.

[0092] Repeat until all factors are extracted or the remaining eigenvalues ​​are less than the preset eigenvalue threshold (usually set to 1); the eigenvalues ​​corresponding to all extracted factors are the eigenvalues ​​obtained by feature analysis;

[0093] Ways to make corrections include:

[0094] Define a penalty matrix P, the diagonal elements of the penalty matrix P are 1, and the off-diagonal elements are -ω; where ω is a hyperparameter with a value range of [0, 1], which is used to control the intensity of the penalty;

[0095] Then the new random correlation coefficient matrix is ​​calculated ; Where t' is the number of judgment steps.

[0096] According to the number of retained factors, an initial factor matrix is ​​constructed, in which each column corresponds to a factor and each row corresponds to a variable (the variable of the optimal data set); the initial factor matrix is ​​orthogonally rotated (such as maximum variable rotation) to obtain the factor loading matrix.

[0097] According to the factor loading matrix, the comprehensive scores of several samples on each factor are calculated; one sample corresponds to a set of comprehensive operating data.

[0098] The formula for calculating the comprehensive score is:

[0099] ; Among them, FZ_IJ is the comprehensive score of the I-th sample on the J-th factor, w'_JS is the weight coefficient of the J-th factor and the S-th variable in the sample; x_IS is the standardized value of the S-th variable of the I-th sample, ε_J is the eigenvalue corresponding to the J-th factor, and a_JS is the loading value of the S-th variable on the J-th factor. It should be noted that the factor can be regarded as a linear combination of variables, and the coefficient of the linear combination is the loading value.

[0100] Take a weighted average of the comprehensive scores of all samples on the Jth factor to obtain the factor score of the Jth factor; preset a score threshold (to decide whether to retain it as the main factor, which is set manually), retain the factors with factor scores greater than or equal to the score threshold as the main factors, and use the main factors as low-dimensional feature variables.

[0101] Furthermore, the steps of cluster analysis include:

[0102] Step 1: Sample the low-dimensional feature variables and extract some data points as a set of representative points. It should be explained that the values ​​of all samples (i.e., comprehensive operating data) on the low-dimensional feature variables are combined to form a multidimensional data point. All multidimensional data points constitute a data set, and some multidimensional data points are extracted from the data set as a set of representative points.

[0103] The methods for extracting data points include:

[0104] The entire data set is gridded, that is, the data space where the data set is located is divided into a number of equal grid cells, and the number of data points in each grid cell is calculated as the density value of the grid cell.

[0105] Calculate the density value range [MI_min, MI_max] of the entire data set, where MI_min is the minimum density value and MI_max is the maximum density value; divide the density value range into several density levels, such as [low density, lower density, medium density, higher density, high density].

[0106] Define the mapping function of the number of representative points ; Where d_T is the average density value of the Tth density level, ω1, ω2, ω3 and θ are mapping adjustable parameters that control the curvature and convergence speed of the mapping function representing the number of points.

[0107] The total number of representative points is preset to N1. For the Tth density level, the density value range is counted; and the average density value d_T of the Tth density level is calculated; based on the corresponding average density value, the number of representative points of the Tth density level is set to .

[0108] Extract the corresponding number of representative points according to the density level; ensure that the proportion of representative points at different density levels satisfies the representative point quantity mapping function.

[0109] For the extracted representative points, the average distance to other representative points is calculated, and a distance threshold is preset. The representative points whose average distance is greater than the distance threshold are treated as noise points and are removed.

[0110] Step 2: For each data point p in the representative point set, calculate the distance (Euclidean distance) between it and all other representative points, find a fixed number (such as a3) of representative points closest to p, and form these representative points and p into a cluster.

[0111] Step 3: For each formed cluster, calculate the center point of the cluster and calculate the sum of the squares of the distances from all data points in the cluster to the center point; select the cluster with the smallest sum of squared distances and use its center point as the first true cluster center.

[0112] Step 4: For the remaining clusters, repeatedly select the cluster with the smallest sum of squared distances, and use its center point as the second true cluster center until kl true cluster centers are obtained; for each data point in the data set, calculate its minimum distance (Euclidean distance) to the kl true cluster centers, and assign the corresponding data point to the cluster with the smallest distance to it.

[0113] Step 5: For each cluster, calculate the sum of the distances from all its data points to the true cluster center as the scatter parameter of the cluster; select cl clusters with the largest scatter parameters.

[0114] Step 6: For the selected cl clusters, perform steps 2 to 4 respectively to obtain new clusters; repeat steps 5 and 6 until the number of clusters reaches the preset cluster value; each cluster finally obtained corresponds to a typical operation mode.

[0115] Furthermore, the methods for building and running directed graph models include:

[0116] Take each cluster as a node and fix a time window of size tw (e.g. 1 hour). For each data point in cluster f, find the cluster f' to which its next data point in the time window belongs; count the number of times N_ff' that all data points in cluster f are transferred to cluster f' in the time window.

[0117] Calculating transition probabilities ; where N_f is the total number of data points in cluster f, is a smoothing parameter greater than 0; K' is the total number of clusters.

[0118] The transition probability p_ff' between each pair of clusters (f, f') is used as the weight of the directed edge, that is, a directed edge is connected from cluster f (node) to cluster f' (node), and the weight of the directed edge is p_ff'.

[0119] For two clusters with a transition probability of 0, no edge is connected or an edge with a preset weight (such as 1e-6) is connected; finally, an operation directed graph model with K' nodes is constructed; the operation directed graph model describes the dynamic transfer relationship and its probability between different operation modes, which is used to analyze the evolution law of the operation mode and is used in subsequent optimization control.

[0120] Furthermore, the optimal control strategy can be solved by:

[0121] Each node (typical operation mode) in the operation directed graph model is regarded as a corresponding state in the state space, and the preset (optional) control action (such as changing the supply and return water temperature, compressor frequency, etc.) is regarded as the action space; the state transition probability is given by the weight of the directed edge.

[0122] Define an instant cost function, the formula of the instant cost function is:

[0123] ; Wherein, ce(s,a) represents the energy consumption when executing control action a in state s, δ1 is the penalty coefficient, which is used to penalize the state-action pair that violates the comfort constraint; cc(s,a) is the impact value on comfort when executing control action a in state s; θ is the comfort threshold. When cc(s,a)>θ, it is considered that the comfort constraint is violated when executing control action a in state s; C(s,a) is the immediate cost when executing control action a in state s; It is an indicator function. If the value in the brackets is satisfied, the value is 1; otherwise, the value is 0.

[0124] It should be noted that the method for obtaining the impact value is: taking the state, control action, and comfort index as features, using an unsupervised machine learning algorithm to learn a machine learning model of the comfort impact value from historical data, and then applying the machine learning model to the new state-action pair to predict the impact value on the comfort when any control action is performed in any state.

[0125] For each state s, define its optimal value function V(s) at infinite time;

[0126] ;

[0127] in, To find a strategy ρ among all strategies so that its corresponding expected value E_ρ is minimized; C(s_t,a_t) is the instantaneous cost when executing the control action a_t in state s_t; s_t is the state at time t, and a_t is the control action executed at time t; bu(t) is a time-varying weight function (such as an exponential decay function), which is used to dynamically adjust the importance of the constraint item; fp(s_t,a_t) is a constraint function, that is, when the constraint is violated when the control action a_t is executed in state s_t, it takes a positive value, otherwise it takes 0 or a negative value; the constraint function can be in any form, as long as it takes a positive value when the constraint is violated, otherwise it takes 0 or a negative value; it can describe various constraints more flexibly, not limited to a single comfort constraint; for example, the constraint function is a weighted combination of multiple constraints, or a more complex nonlinear function, etc.; s_0 represents the initial state, s_0=s represents the initial state is s; ∈ is the discount factor, 0<∈<1, which is used to convert future costs to the present.

[0128] For each state s, solve its optimal value function V(s), and express V(s) as the convex combination of V(s') of its adjacent state s';

[0129] Right now ; Where max_{s',s} represents the maximum value of all states s' adjacent to state s (which can be directly transferred from s); C(s,ρ(s)) represents the immediate cost of executing the control action ρ(s) given by the strategy ρ in state s; δ3 is the weight discount factor, which is a number between 0 and 1 and is used to balance the weights of immediate rewards and future rewards; V(s') is the optimal value function of state s'; the coefficients of this convex combination are solved by the linear programming method, that is, V(s) is obtained, and V(s) of all states is solved.

[0130] The optimal control strategy is obtained by inverse solution of V(s) , that is, the optimal action given by the optimal control strategy in state s; ; Among them, argmin_a[] means taking the minimum value for all actions a, that is, finding an action that minimizes the expression in the square brackets []; γ3 is the balance weight, between 0 and 1, which is used to balance the weight of immediate rewards and future rewards; pu(s'|s,a) represents the probability of transferring to state s' after executing control action a in state s, that is, the state transition probability; the optimal control strategy obtained is Applied to actual control, it can minimize energy consumption while satisfying comfort constraints.

[0131] This embodiment, through in-depth mining of massive operating data, automatically identifies typical operating modes without manual modeling, and can fully adapt to complex and changeable operating environments; at the same time, constructs an operating directed graph model to characterize the dynamic transfer rules between operating modes, lays the foundation for formulating forward-looking control strategies, and effectively improves control performance; secondly, combines the energy consumption minimization goal with the comfort constraint, and solves the optimal control strategy based on the improved partitioning algorithm, seeks the best balance between energy saving and comfort, and achieves both energy saving and comfort; in the solution process, the solution efficiency is improved by effectively processing large-scale state space; furthermore, it can be applied to central air-conditioning systems of different types and sizes, providing strong support for achieving energy-saving control; in addition, the control strategy is based on a basis, based on the analysis and mining of massive data rather than subjective experience, and can guide air-conditioning operation more scientifically and reasonably, avoiding the limitations of artificial rules of thumb; it can significantly improve the energy-saving performance of the air-conditioning system while taking into account comfort requirements.

[0132] Example 2

[0133] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a central air-conditioning energy consumption optimization intelligent control method based on big data analysis, including:

[0134] S1. Collect comprehensive operation data of central air conditioners; pre-process the collected comprehensive operation data to obtain low-dimensional feature variables;

[0135] S2. Cluster analysis is performed on low-dimensional feature variables to obtain several typical operation modes;

[0136] S3. Construct an operation directed graph model based on several typical operation modes;

[0137] S4. Based on the operation directed graph model, an improved partitioning algorithm is used to solve the optimal control strategy; the optimal control strategy is applied to the real-time control of the central air conditioner to achieve energy consumption optimization.

[0138] Example 3

[0139] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the above-mentioned intelligent control method for optimizing energy consumption of central air conditioners based on big data analysis is implemented.

[0140] Since the electronic device introduced in this embodiment is an electronic device used to implement a central air-conditioning energy consumption optimization intelligent control method based on big data analysis in the embodiment of this application, based on the central air-conditioning energy consumption optimization intelligent control method based on big data analysis introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not introduced in detail here. As long as the technical personnel of this field implement the electronic device used in the central air-conditioning energy consumption optimization intelligent control method based on big data analysis in the embodiment of this application, it belongs to the scope of protection of this application.

[0141] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0142] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A central air conditioning energy consumption optimization intelligent control system based on big data analysis, characterized in that: include: Data collection and processing module, clustering module, model building module and strategy solving module; each module is connected by wired and / or wireless means; The data acquisition and processing module includes an acquisition unit and a processing unit. The acquisition unit is used to acquire the comprehensive operation data of the central air conditioner; the processing unit is used to pre-process the acquired comprehensive operation data to obtain low-dimensional feature variables; Clustering module, used to cluster low-dimensional feature variables and obtain several typical operation modes; Model building module, which builds and runs directed graph models based on several typical operation modes; The strategy solving module includes a solving unit and a control unit. The solving unit uses an improved partitioning algorithm to solve the optimal control strategy based on the running directed graph model. The control unit is used to apply the optimal control strategy to the real-time control of the central air conditioner. The preprocessing method includes data cleaning, data standardization and data dimension reduction; the data dimension reduction method includes: The comprehensive operation data after data cleaning and data standardization is recorded as the operation data set; variables are extracted from the operation data set to form a data matrix X, and the correlation coefficient matrix of the data matrix X is calculated , the correlation coefficient matrix Elements in Indicates variables and The correlation coefficient between the variables; Calculate the sample fitness value of the running data set based on the correlation coefficient matrix ; ;in, For the variables and The correlation coefficient between the variables, For the variables and The weight coefficient of the variable, For the variables and The partial correlation coefficient between the variables is is the amount of data in the running dataset, is the number of variables; is the partial correlation coefficient adjustment function; Partial correlation coefficient adjustment function ;in, is a parameter that controls the curvature of the function; is the ratio influencing parameter; Preset dimensionality reduction analysis threshold, if the sample fitness value of the running data set If the value is less than the dimensionality reduction analysis threshold, the data in the running data set is readjusted; the sample fitness value of the running data set is recalculated until the sample fitness value is greater than or equal to the dimensionality reduction analysis threshold; retain the running data set at this time as the optimal data set, calculate and obtain the correlation coefficient matrix corresponding to the optimal data set; record it as the optimal correlation coefficient matrix; obtain the number of retained factors based on the optimal correlation coefficient matrix; According to the number of retained factors, construct the initial factor matrix, in which each column corresponds to a factor and each row corresponds to a variable; The initial factor matrix is ​​orthogonally rotated to obtain the factor loading matrix; based on the factor loading matrix, the comprehensive scores of several samples on each factor are calculated; one sample corresponds to a set of comprehensive operation data; The formula for calculating the comprehensive score is: ;in, For the The sample in The combined score on the factors For the factors and the first The weight coefficient of each variable; For the The sample The standardized value of the variable, For the The eigenvalues ​​corresponding to the factors are For the The variable in The loading values ​​on the factors; All samples in The weighted average of the comprehensive scores on the factors is obtained. The factor scores of the factors are calculated; a score threshold is preset, and the factors with factor scores greater than or equal to the score threshold are retained as the main factors, and the main factors are used as low-dimensional feature variables; The steps of performing the cluster analysis include: Step 1: Sample the low-dimensional feature variables and extract some data points as representative point sets; Step 2: For each data point p in the representative point set, calculate the distance between it and all other representative points, find a fixed number of representative points closest to p, and form a cluster with these representative points and p; Step 3: For each formed cluster, calculate the center point of the cluster and calculate the sum of the squares of the distances from all data points in the cluster to the center point; select the cluster with the smallest sum of squares of distances and use its center point as the first true cluster center; Step 4: For the remaining clusters, repeatedly select the cluster with the smallest sum of squared distances, and use its center point as the second true cluster center until kl true cluster centers are obtained; for each data point in the data set, calculate its minimum distance to the kl true cluster centers, and assign the corresponding data point to the cluster with the closest minimum distance; Step 5: For each cluster, calculate the sum of the distances from all its data points to the true cluster center as the scatter parameter of the cluster; select cl clusters with the largest scatter parameters; Step 6: For the selected cl clusters, perform steps 2 to 4 respectively to obtain new clusters; repeat steps 5 and 6 until the number of clusters reaches the preset cluster value; each cluster finally obtained corresponds to a typical operation mode; The method of constructing the running directed graph model includes: Take each cluster as a node, fix a time window of size tw, and For each data point in the time window, find the cluster to which the next data point in the time window belongs. ; Statistical cluster All data points in the time window are transferred to the cluster The number of ; Calculating transition probabilities ;in It is a cluster The total number of data points in is a smoothing parameter greater than 0; is the total number of clusters; Each pair of clusters The transition probability between As the weight of the directed edge, that is, from cluster Pointing to Cluster Connect a directed edge, the weight of the directed edge is ; For two clusters with a transition probability of 0, no edge is connected or an edge with a preset weight is connected; finally, a The operation directed graph model of each node; The solution method of the optimal control strategy includes: Each node in the running directed graph model is regarded as a corresponding state in the state space, and the preset control action is regarded as the action space; the state transition probability is given by the weight of the directed edge; Define an instant cost function, the formula of the instant cost function is: ;in, Indicates in status Execute control actions Energy consumption is the penalty coefficient; For the status Execute control actions The impact value on comfort level; is the comfort threshold; For the status Execute control actions The immediate cost of is the indicator function; For each state , define its optimal value function at infinite time ; Based on the optimal value function Solve and obtain the optimal control strategy; ； in, To find a strategy among all strategies , so that its corresponding expected value Minimum; For the status Execute control actions The immediate cost of For the moment status, For the moment the control actions performed; is the time-varying weight function; is the constraint function; represents the initial state, Indicates that the initial state is ; is the discount factor.

2. According to claim 1, a central air conditioning energy consumption optimization intelligent control system based on big data analysis is characterized in that: The comprehensive operation data includes indoor environment data, outdoor environment data, air conditioning operation data and building data; the indoor environment data includes indoor temperature, indoor humidity and carbon dioxide concentration; the outdoor environment data includes outdoor temperature, outdoor humidity and solar radiation intensity; Air conditioning operation data includes unit operation status, supply and return water temperature, supply and return water flow, compressor operation frequency and energy consumption per unit time; Building data includes building orientation and building insulation.

3. According to claim 2, a central air conditioning energy consumption optimization intelligent control system based on big data analysis is characterized in that: The method for obtaining the number of retained factors based on the optimal correlation coefficient matrix includes: generating several groups of random data with the same data volume and number of variables as the optimal data set, performing feature analysis on each group of random data, and extracting eigenvalues; calculating the mean and 95% percentile of each eigenvalue, and calculating the standard deviation of the eigenvalue; performing feature analysis on the optimal data set at the same time, and recording the obtained eigenvalues ​​as constrained eigenvalues; When the constraint eigenvalue is greater than the comparison eigenvalue When , the factors corresponding to the constrained eigenvalues ​​are retained; Comparison of feature values ;in, is the mean of the eigenvalues, is the standard deviation of the eigenvalues, is the 95% percentile of the eigenvalue, It is an adjustable parameter; The number of factors is defined as the serial number of the factor corresponding to the last eigenvalue retained, the explanatory power of the retained factors is calculated, and the explanatory threshold is preset. If the explanatory power is less than the explanatory threshold, the number of retained factors is increased or decreased, and the number of increases or decreases is 1; until the explanatory power of the retained factors is greater than or equal to the explanatory threshold.

4. According to claim 3, a central air conditioning energy consumption optimization intelligent control system based on big data analysis is characterized in that: The method of performing feature analysis on each set of random data includes: Calculate the correlation coefficient matrix of random data, denoted as random correlation coefficient matrix ,from Find the largest eigenvalue in and its corresponding eigenvector ; Calculate the eigenvector The corresponding factor The contribution of explained variance ; Preset factor retention threshold, if If the factor is less than the factor retention threshold, the factor is discarded. , and Make corrections and get a new random correlation coefficient matrix ; like If the factor is greater than or equal to the factor retention threshold, the factor As the first factor, The corresponding variables in The variance from It is eliminated and the residual matrix is ​​obtained ;exist Find the largest eigenvalue in and the corresponding eigenvector , and calculate the eigenvector The corresponding factor The contribution of explained variance ; Based on the factor retention threshold, it is also determined whether to retain the factor ; If retained, the factor As the second factor; Repeat until all factors are extracted or the remaining eigenvalues ​​are less than the preset eigenvalue threshold; the eigenvalues ​​corresponding to all extracted factors are the eigenvalues ​​obtained by feature analysis; Ways to make corrections include: Define a penalty matrix , the penalty matrix The diagonal elements of are 1, and the off-diagonal elements are ;in, is a hyperparameter with a value range of [0, 1]; then the new random correlation coefficient matrix is ​​calculated ;in, is the number of steps to be judged.

5. According to claim 4, a central air conditioning energy consumption optimization intelligent control system based on big data analysis is characterized in that: The data point extraction method includes: Grid the entire data set, that is, divide the data space where the data set is located into a number of equal grid cells, and calculate the number of data points in each grid cell as the density value of the grid cell; Calculate the density value range [MI_min, MI_max] of the entire data set, where MI_min is the minimum density value and MI_max is the maximum density value; divide the density value range into several density levels; Define the mapping function of the number of representative points ;in, For the The average density value of density levels, , , and For mapping adjustable parameters; The total number of representative points is preset to , for the density level, count its density value range; and calculate the The average density value of density levels ; Based on the corresponding average density value, The number of representative points of each density level is set to ; A corresponding number of representative points are extracted according to the density level; for the extracted representative points, the average distance to other representative points is calculated, and a distance threshold is preset. Representative points whose average distance is greater than the distance threshold are treated as noise points and removed.

6. According to claim 1, a central air conditioning energy consumption optimization intelligent control system based on big data analysis is characterized in that: The optimal value function The methods for solving the optimal control strategy include: For each state , solve its optimal value function ,Will Indicates the state of its neighbors of Convex combination of ; Right now ;in, Represents all and status Adjacent state Take the maximum value; Indicates in status Execution strategy The control action given the immediate costs incurred; is the weight discount factor; Status The optimal value function of ; solve the coefficients of this convex combination by linear programming method, and get , solve for all states ;pass The optimal control strategy is obtained by inverse solution ;in, For all actions Take the minimum value; To balance the weight; Indicates in status Execute control actions After that, transfer to the state , that is, the state transition probability.

Images