Dynamic reference determination method for heat consumption / cold quantity of heating ventilation air-conditioning system

Abstract

The invention relates to a dynamic reference determination method for heat consumption / cold energy of a heating ventilation air-conditioning system. The method comprises the following steps: establishing a heat consumption (cold energy) classification model for processed operation data of the heating ventilation air-conditioning system; carrying out data processing before association rule mining by adopting a K-means clustering discrete method and an equal-width discrete method; mining the association rules by using an Apriori algorithm, and processing the mined rules; and according to the mined association rules, calculating the dynamic reference of the heat (cold) consumption of each energy station or heating station. Compared with the prior art, the method has the advantages that the energy stations or the heating stations of the heating ventilation and air conditioning system are taken as objects, and the reasonable dynamic energy consumption benchmark is established for the energy stations or the heating stations with different conditions and different levels by fusing the multi-factor characteristics. Compared with a traditional mechanism model, calculation is simple, time and cost are low, and a quantitative assessment basis and an energy-saving direction are provided for energy conservation and carbon reduction of the heating ventilation and air conditioning system.

Description

Technical Field

[0001] The invention relates to a method for determining a dynamic benchmark of heat consumption / cold capacity of a heating, ventilation and air conditioning system based on decision tree classification-association rule mining. Background Art

[0002] The HVAC system is an important energy consumption system in buildings, and its energy consumption level directly affects the energy consumption and carbon emissions of buildings. Since the heat (cold) consumption of the HVAC system is affected by many factors such as outdoor air temperature, indoor load demand, equipment operating status and system control strategy, and changes dynamically with the changes of these factors, mining the existing HVAC system operation data and establishing a reasonable dynamic heat (cold) consumption benchmark is the key basis for reducing building energy consumption and achieving energy-saving operation.

[0003] At present, there are many methods for determining the heat (cold) consumption benchmark of HVAC systems, but there are certain problems and shortcomings: First, most of the existing heat (cold) consumption benchmark determination methods are based on mechanism models, which are highly dependent on the input characteristics of the model. The uncertainty of the input data affects the reliability of the simulation results. The modeling process is complex, time-consuming and costly. Second, most of the existing studies focus on the building energy consumption benchmark, and few studies are conducted on the heat (cold) consumption benchmark of HVAC system energy stations or thermal power stations. Third, the various energy consumption benchmarks of HVAC systems recognized by current standards and specifications are mostly regional static benchmarks, which cannot effectively guide the operation and regulation of thermal power stations.

[0004] The method for determining the dynamic benchmark of heat (cold) consumption of a HVAC system of the present invention is based on decision tree classification-association rule mining, screens characteristic factors that affect the energy consumption of an energy station or a thermal power station of a HVAC system, establishes a heat (cold) consumption classification model, and establishes a dynamic benchmark model of heat (cold) consumption of a HVAC system based on association rule mining according to the classification results. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide a method for determining the dynamic benchmark of the heat (cold) consumption of a HVAC system. The method uses the historical operation data of the HVAC system based on decision tree classification-association rule mining to determine the dynamic benchmark of the heat (cold) consumption of the HVAC system energy station or thermal power station, providing a reliable reference for the energy-saving and emission-reduction operation of the HVAC system.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for determining a dynamic benchmark of heat consumption / cooling capacity of a heating, ventilation and air conditioning system, the method comprising the following steps:

[0008] Step 1: Feature Selection

[0009] The HVAC system data under different working conditions are obtained. After eliminating abnormal data, the Person correlation coefficient between each energy consumption feature and heat (cold) consumption is calculated, and the features with a Person correlation coefficient greater than or equal to 0.24 are retained. At the same time, the importance index of heat (cold) consumption is calculated using the ReliefF algorithm, and the features with an importance index greater than or equal to 0.003 are selected.

[0010] After merging the features selected by the Person correlation coefficient and the ReliefF algorithm, the random forest embedding method is used to select the features with random forest importance index exceeding 0.3 as the main features affecting energy consumption;

[0011] Step 2: Construct a heat (cooling) consumption classification model:

[0012] The heat (cold) consumption classification model is based on a decision tree model. The nodes of the decision tree include a root node, a leaf node, and a decision node. The root node of the decision tree is a collection of all thermal power station data. The decision node corresponds to the division rule of energy consumption characteristics. The decision node divides the data on the root node into several subsets according to the energy consumption characteristic rule. The leaf node is the final energy consumption classification node. The thermal power stations in the leaf node are classified into the same category to realize the classification of thermal power station heat consumption.

[0013] The input features of the decision tree model are the main features affecting energy consumption and the corresponding heat consumption determined in step 1. Bayesian optimization is used to find the optimal parameters of the decision tree model. At the same time, the packaging method is used to completely search the features to determine the decision tree regression feature combination with the best score exceeding 75%;

[0014] Then, the decision tree model is pruned using the cost complexity post-pruning method to remove unnecessary nodes, and the heat (cooling) consumption classification model and the final input features of the model are obtained;

[0015] Step 3: Use the heat (cooling) consumption classification model constructed in step 2 to classify the HVAC system data using the final input features;

[0016] Step 4: For each type of HVAC system data, the Person correlation coefficient between each energy consumption feature and heat (cooling) consumption is calculated again. All energy consumption features with a Person correlation coefficient greater than 0.7 are used as the rule antecedents for association rule mining, and the heat (cooling) consumption is used as the rule posterior;

[0017] Step 5: Discretize the HVAC system data, and after discretization, use the Apriori algorithm to mine association rules according to the rule antecedents and rule consequents determined in step 4 to obtain association rules for all classes;

[0018] Step 6: Perform rule processing on the association rules of all classes obtained in step 5 to obtain available rules, and use the available rules to determine the dynamic benchmark of the heat (cold) consumption of the HVAC system.

[0019] Furthermore, the process of rule processing is:

[0020] Discretize the outdoor temperature with equal width to obtain different outdoor temperature intervals. Group all the class association rules obtained in step 5 according to whether they are in the same outdoor temperature interval, and determine the confidence level of the rules in each group.

[0021] Set the minimum confidence level and extract all rules with confidence levels greater than the minimum confidence level for expert analysis;

[0022] The rules are analyzed through the knowledge of heating domain, and unreasonable rules are replaced by adjacent and reasonable rules with lower confidence.

[0023] Furthermore, the process of determining the dynamic benchmark of dynamic heat (cold) consumption is: statistics are obtained on the operating conditions and available heat consumption rules under all outdoor temperature ranges, and the upper and lower bounds of the heat (cold) consumption range after the rules are used to calculate the dynamic heat (cold) consumption benchmark 1 and benchmark 2 of the thermal power station respectively.

[0024] Furthermore, in step 5, the discretization process is: discretizing the data of the primary pipe network flow, pressure, outdoor air temperature, and secondary pipe network supply and return water average temperature parameters of the HVAC system; using K-means clustering discretization for operation data and energy consumption data, and using equal-width discretization method for indoor and outdoor air temperatures.

[0025] Furthermore, the decision tree model was pruned using the cost complexity post-pruning method. Random 5-fold cross-training was used for model verification, and the mean square error (MSE), mean absolute error (MAE), and determination coefficient (R) of the five training and test sets were calculated. 2 , the average value of the explained variance Evar, when the mean square error MSE and mean absolute error MAE of the model are less than 0.05, and the determination coefficient R 2 If the sum of the explained variance Evar is greater than 75%, the model is considered to have passed the test and is used to classify the heat (cooling) consumption.

[0026] Furthermore, the HVAC system is a thermal power station or an energy station.

[0027] Furthermore, in step 4, the rule antecedents of association rule mining are the primary flow rate of the energy station or thermal power station, the outdoor temperature, the secondary supply and return water average temperature, and the indoor temperature of typical users.

[0028] Furthermore, in the Apriori algorithm, the minimum support is set to 0.002, the minimum confidence is set to 60%, and the minimum lift is set to 1.

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

[0030] The outstanding essential features of the present invention are:

[0031] 1. The present invention relies on historical data when determining the dynamic benchmark of heat (cold) consumption of HVAC system energy stations or thermal power stations, and does not rely on complex mechanism models, which reduces time and cost. The method of decision tree classification-association rule mining is used to establish the dynamic benchmark of heat (cold) consumption of HVAC system. The present invention can mine the energy stations and thermal power stations of many HVAC systems, and integrate multi-factor characteristics to establish reasonable dynamic energy consumption benchmarks for energy stations or thermal power stations with different conditions and levels.

[0032] 2. According to the feature selection results, the present invention uses Bayesian optimization and cost-complexity post-pruning to refine the model and establish a HVAC system heat (cold) consumption classification model based on decision tree classification, providing a basis for the subsequent establishment of dynamic benchmark models for energy consumption of various types of energy stations or thermal power stations.

[0033] 3. The present invention mines the dynamic benchmark of heat (cold) consumption of HVAC systems of various energy stations or thermal power stations based on association rule algorithms. Association rule mining is a commonly used descriptive statistical method that can extract hidden rules from a large amount of data. It overcomes the shortcomings of traditional mechanism models that are affected by historical data quality, operating modes, abnormal operating conditions and human experience adjustments. It has strong interpretability and is used for HVAC system control optimization. The effective association rules (available rules) established through association rule mining can reasonably determine the dynamic benchmark of heat (cold) consumption of energy stations or thermal power stations, providing a reliable basis for optimizing system operation and achieving energy conservation and carbon reduction.

[0034] The method for determining the dynamic benchmark of heat (cold) consumption of the present invention provides a quantitative assessment basis and energy-saving operation direction for energy conservation and carbon reduction of various types of energy stations or thermal power stations in HVAC systems, and provides a prerequisite for calculating the system carbon reduction rate and evaluating the carbon reduction potential.

[0035] According to the selected energy consumption characteristics, a heat (cold) consumption classification model of the HVAC system energy station or thermal power station is established, and the association rules of the HVAC system are mined. According to the mined association rules, the dynamic benchmark of the heat (cold) consumption of the HVAC system is calculated. This method uses the primary pipe network flow, pressure, outdoor air temperature, secondary pipe network supply and return water average temperature and user indoor temperature of the HVAC system as independent variables to determine the dynamic benchmark of heat (cold) consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1The present invention is a flow chart of a method for determining a dynamic benchmark of heat consumption of a HVAC system based on decision tree classification-association rule mining.

[0037] Figure 2 A flow chart is selected for the characteristics of the present invention.

[0038] Figure 3 This is a graph showing the calculation results of the Person correlation coefficient of the heat consumption characteristics of the present invention.

[0039] Figure 4 This is a diagram showing the calculation results of the importance index of the heat consumption characteristic ReliefF algorithm of the present invention.

[0040] Figure 5 This is the calculation result of the importance index of the random forest feature selection of the heat consumption feature involved in the present invention.

[0041] Figure 6 This is an example diagram of the decision tree of the present invention.

[0042] Figure 7 Schematic diagram of the Bayesian optimization process of the present invention.

[0043] Figure 8 Schematic diagram of the cost complexity pruning process of the present invention.

[0044] Fig. 9 This is a schematic diagram of the 5-fold cross-validation principle of the present invention.

[0045] Fig.10 This is a post-processing flow chart of the decision tree regression heat consumption classification model of the present invention.

[0046] Figure 11-13 It is a box plot of heat consumption, flow rate and comprehensive physical property coefficient of each type of thermal power station of the present invention.

[0047] Fig.14 , Fig.15 They are respectively the dynamic heat consumption benchmarks of floor radiant heating and radiator heating thermal power stations in the embodiments of the present invention. The black curve in the figure represents the outdoor temperature in the heating season, and the other curves of different colors represent the dynamic heat consumption benchmarks of various types of floor radiant heating thermal power stations in the entire heating season. DETAILED DESCRIPTION

[0048] The present invention is further explained below in conjunction with the embodiments and drawings. The drawings are simplified schematic diagrams, but are not intended to limit the protection scope of the present application.

[0049] The present invention is a method for determining the dynamic reference of the heat (cold) consumption of the HVAC system (see Figure 1 ), the method comprising the following steps:

[0050] Step S1: Using filtering and embedding methods on HVAC system data, preliminarily select the main energy consumption features that affect the amount of heat (cooling) consumption;

[0051] The HVAC system data under different working conditions are obtained. After eliminating abnormal data, the Person correlation coefficient between each energy consumption feature and heat (cold) consumption is calculated, and the features with a Person correlation coefficient greater than or equal to 0.24 are retained. At the same time, the importance index of heat (cold) consumption is calculated using the ReliefF algorithm, and the features with an importance index greater than or equal to 0.003 are selected.

[0052] After merging the features selected by the Person correlation coefficient and the ReliefF algorithm, the random forest embedding method is used to select the features with random forest importance index exceeding 0.3 as the main features affecting energy consumption;

[0053] Step S2: Construct a heat (cooling) consumption classification model:

[0054] The heat (cold) consumption classification model is based on a decision tree model. The nodes of the decision tree include a root node, a leaf node, and a decision node. The root node of the decision tree is a collection of all thermal power station data. The decision node corresponds to the division rule of energy consumption characteristics. The decision node divides the data on the root node into several subsets according to the energy consumption characteristic rule. The leaf node is the final energy consumption classification node. The thermal power stations in the leaf node are classified into the same category to realize the classification of thermal power station heat consumption.

[0055] The input features of the decision tree model are the main features that affect energy consumption and the corresponding heat consumption. Bayesian optimization is used to find the optimal parameters of the decision tree model. At the same time, the packaging method is used to completely search the features to determine the best decision tree regression feature combination with a score of more than 75%;

[0056] Then, the decision tree model is pruned using the cost complexity post-pruning method to remove unnecessary nodes, and the heat (cooling) consumption classification model and the final input features of the model are obtained;

[0057] Through Bayesian optimization and cost-complexity post-pruning, the generalization ability of the model can be improved, thus generating a simpler and more efficient decision tree. In the cost-complexity post-pruning method, when the mean square error (MSE) and mean absolute error (MAE) of the regression model are less than 0.05, the determination coefficient R 2 If the and explained variance Evar is greater than 75%, the model is considered to have passed the test.

[0058] Step S3: using the heat (cooling) consumption classification model constructed in step S2 and taking the finally determined input features as input, classifying the HVAC system data;

[0059] Step S4: for each type of HVAC system data, the Person correlation coefficient between each energy consumption feature and heat (cooling) consumption is calculated again, and all energy consumption features with a Person correlation coefficient greater than 0.7 are used as the rule antecedents for association rule mining, and the heat (cooling) consumption is used as the rule posterior;

[0060] Step S5: discretize the HVAC system data, and after the discretization, use the Apriori algorithm to mine association rules according to the rule antecedents and rule consequents determined in step 4 to obtain association rules of all classes;

[0061] When determining the antecedent and consequent of the association rules for each type of energy station or thermal power station, the data used are all original data with abnormal data removed, and the discretely processed data are used when conducting rule mining.

[0062] Step S6: performing rule processing on the association rules of all classes obtained in step S5 to obtain available rules, and using the available rules to determine the dynamic benchmark of the heat (cold) consumption of the HVAC system.

[0063] In step S5, the discretization process is: discretize the primary pipe network flow, pressure, outdoor air temperature, and secondary pipe network supply and return water average temperature parameters of the HVAC system; use K-means clustering discretization for operation data and energy consumption data, and use equal width discretization method for indoor and outdoor air temperature. The data of discrete attributes required for association rule mining can be obtained as the antecedent and consequent of the rule.

[0064] The process of rule processing is:

[0065] The outdoor temperature is discretized with equal width to obtain different outdoor temperature intervals, and all the class association rules obtained in step S5 are grouped according to whether they are in the same outdoor temperature interval, and the confidence of the rules in each group is determined;

[0066] Set the minimum confidence level and extract all rules with confidence levels greater than the minimum confidence level for expert analysis;

[0067] The rules are analyzed through the knowledge of heating domain, and unreasonable rules are replaced by adjacent and reasonable rules with lower confidence.

[0068] After actually determining the benchmark, the benchmark range of heat (cooling) consumption is determined by the range of the outdoor temperature of the thermal power station.

[0069] Preferably, the minimum confidence level is set to 60%.

[0070] The Apriori algorithm is used to discover frequent item sets. The thresholds of minimum support, minimum confidence, and minimum lift are set in the algorithm to mine rules.

[0071] Example 1

[0072] This embodiment uses the operating status of a typical HVAC system in a cold region as the identification object. The HVAC system has a total of 169 thermal power stations, and the actual heating area is about 8.49 million square meters. The historical heat source operation data of the system is collected, including meteorological parameters, unit information, operation data, and indoor data of typical users under the unit. Among the 169 thermal power units, 63 units have buildings under their jurisdiction that use floor radiation heating, and 106 units have buildings under their jurisdiction that use radiators for heating.

[0073] The method for determining the dynamic benchmark of heat consumption of a centralized HVAC system based on decision tree classification-association rule mining in this embodiment includes the following steps:

[0074] Step S1: Use filtering method and embedding method to select the energy consumption characteristics related to the dynamic benchmark of heat consumption of the thermal power station, and screen out the main characteristics affecting energy consumption.

[0075] The present invention combines the filtering method with the embedding method, which can quickly screen and accurately optimize features through complementary advantages, improve the accuracy and efficiency of model prediction, and reduce the risk of overfitting, thereby providing a more robust feature selection scheme for the establishment of a dynamic benchmark for heat consumption of thermal power stations.

[0076] The feature selection process is shown in Figure 2 , first use the filtering method: 1) calculate the Pearson correlation coefficient between each energy consumption feature and heat consumption; 2) use the ReliefF algorithm to calculate the importance index of heat (cold) consumption to evaluate the nonlinear correlation of each energy consumption feature. The present invention implements the ReliefF algorithm through the "relieff" function in MATLAB R2021. Set the "regression" parameter to solve the regression problem, and set the maximum number of iterations to 1000 and the error threshold to 10 -6 , the learning rate is 0.01, and the feature selection of ReliefF algorithm is performed. The energy consumption characteristics of the factors affecting the heat consumption in the thermal power station are shown in Table 1.

[0077] Table 1 Factors affecting heat consumption and energy consumption characteristics

[0078]

[0079] Figure 3 The Person correlation coefficient calculation result diagram between the energy consumption characteristics and heat consumption of this embodiment is shown in FIG. Figure 3 It can be seen that the flow rate G and the secondary supply and return water temperature change rate std t2The correlation coefficient between the indoor temperature and the room temperature change rate and the heat consumption is 0.34, which is the strongest correlation among all energy consumption characteristics. The absolute value of the correlation coefficient between the indoor temperature and the room temperature change rate and the heat consumption is about 0.1, which is the weakest correlation among all characteristics. This embodiment selects the characteristics with a Person correlation coefficient greater than or equal to 0.24, namely, water consumption W, flow rate G, and secondary supply and return water temperature change rate std t2 , secondary supply and return water average temperature t2 and secondary supply and return water temperature difference Δt2 are taken as candidate features.

[0080] Figure 4 This is a diagram showing the calculation results of the importance index of the ReliefF algorithm in this embodiment. Figure 4 It can be seen that the importance index of the flow G index is 0.06, which is much higher than the other features, indicating that the flow index has an important influence on the heat consumption index. The features with importance index greater than or equal to 0.003 are selected, namely water consumption W, flow G, heating area A, building comprehensive physical coefficient Bc and secondary supply and return water temperature change rate std t2 as candidate features.

[0081] Based on the results of Person coefficient and ReliefF feature screening, water consumption W, heating area A, building comprehensive physical coefficient Bc, secondary supply and return water temperature change rate std are initially filtered out. t2 , flow rate G, secondary supply and return water average temperature t2 and secondary supply and return water temperature difference Δt2. Then, the random forest embedding method is used to select these 7 features.

[0082] Figure 5 This is a graph showing the calculation results of the importance of energy consumption features using the random forest embedding method in this embodiment. Figure 5 As shown in the figure, the flow characteristic has the greatest impact on the heat consumption, and the random forest importance index of 5 characteristics exceeds 0.3. After screening, the heating area A, building comprehensive physical coefficient Bc, flow index G, secondary supply and return water temperature change rate std t2 , and the average temperature t2 of secondary supply and return water are the main characteristics affecting energy consumption.

[0083] Step 2: Establish a heat consumption classification model based on a decision tree according to the five main energy consumption-influencing features that have been screened out:

[0084] Decision tree is a common machine learning algorithm, mainly used for classification and regression tasks. Its working principle is to segment data through a series of judgment rules, and ultimately achieve the purpose of classifying or predicting data.

[0085] Figure 6The decision tree diagram is shown by taking the decision tree classification model of the heat consumption of a thermal power station as an example. The nodes of the decision tree include root nodes, leaf nodes, and decision nodes. For the classification problem of the heat consumption of a thermal power station of the present invention, the root node of the decision tree is the set of all the data of the thermal power station; the root node in the figure includes four data D1-D4. The decision node corresponds to the division rule of the energy consumption characteristics, and the data on the root node is divided into several subsets according to the energy consumption characteristic rules; if the decision node G>1 divides D1 and D2 into a subset, and divides D3 and D4 into a subset. The leaf node is the final energy consumption classification node, and the thermal power stations within the leaf node are classified into the same category to realize the classification of the heat consumption of the thermal power station; the example in the figure divides the thermal power station data into 3 leaf nodes, and divides the thermal power station into three categories.

[0086] Bayesian optimization and cost-complexity post-pruning are two different optimization methods. Using only Bayesian optimization may not effectively reduce the complexity of the constructed model, while using only cost-complexity post-pruning may lack the precise search for the optimal parameters, resulting in suboptimal model performance. The present invention establishes a decision tree-based heat consumption classification model based on two-step optimization of Bayesian optimization and cost-complexity post-pruning, and selects a typical thermal power station for subsequent benchmark establishment. The two-step optimization method can accurately select parameters, enhance the generalization ability of the model, significantly reduce manual intervention, and improve overall training efficiency.

[0087] Figure 7 It is a schematic diagram of the Bayesian optimization process involved in the present invention.

[0088] by Figure 7 For example, Bayesian optimization mainly includes 5 steps:

[0089] Step 1. Define the objective function. The present invention selects the decision tree regression accuracy score as the objective function. The vertical coordinate of the example objective function in the figure is the decision tree regression accuracy score.

[0090] Step 2. Define the parameter space. For decision tree regression problems, the parameter space includes splitting criteria, maximum tree depth, minimum number of samples for internal node re-division, minimum number of samples for leaf nodes, splitting strategy, and weight parameters. The example in the figure uses the splitting criteria and the maximum tree depth of decision tree regression as the parameter space. The red and black curves represent the mean absolute error and mean square error of the splitting criteria respectively; the horizontal axis is the maximum tree depth, and the curves represent the accuracy scores obtained for different splitting criteria and different maximum tree depths. The curve in the figure is a black box function, and the maximum tree depth and splitting criteria corresponding to the highest score need to be determined through model optimization.

[0091] Step 3. Use the Gaussian mixture model as a surrogate model, which can effectively fit complex data distribution and perform probability density estimation and soft clustering.

[0092] Step 4. Explore the best parameter points and evaluate the objective function. The pink square points in the figure are the initial sampling points. After sampling, calculate the accuracy score of the sampling points and compare the difference between the score and the target score; if the target score cannot be reached, re-sample according to the proxy function.

[0093] Step 5. Get the optimal solution. Use the new parameter values ​​and the corresponding accuracy score evaluation results to update the proxy model and iterate until the preset convergence condition is reached to obtain the parameter combination with the best accuracy. In the example in the figure, the target score is set to 80%. After four iterations, the preset target is met as shown by the red five-pointed star.

[0094] Figure 8 Schematic diagram of the cost complexity pruning process involved in the present invention.

[0095] by Figure 8 For example, cost complexity post-pruning includes 5 steps:

[0096] Step 1. Model construction. The model is a decision tree regression model. The input features in the example in the figure are flow rate G and secondary supply and return water average temperature t2, and the objective function is heat consumption Q. After model construction, the blue decision tree diagram in the figure is generated. The tree diagram includes 3 decision nodes and 4 leaf nodes.

[0097] Step 2. Cost complexity calculation. Calculate the cost complexity of each internal node according to formula (1) (2). The example in the figure is calculated to obtain the cost complexity index after pruning three internal nodes.

[0098]

[0099] In the formula, α represents the cost complexity index; MSE represents the mean square error; MSE(a) represents the mean square error of the current node after pruning; MSE(A) represents the sum of the mean square errors of the nodes below the current node after pruning; N(A) represents the number of leaf nodes pruned by the current node. n is the number of nodes, y i is the ith actual observation value, is the i-th predicted value.

[0100] Step 3. Prune the tree graph and select the tree with the smallest cost complexity. The example in the figure is a green new tree graph obtained after comparative pruning.

[0101] Step 4. Repeat steps 2 and 3 until the cost function no longer decreases. The example in the figure repeatedly calculates the cost complexity of the pruned tree and finds that the cost complexity after the second calculation no longer decreases.

[0102] Step 5. Get the optimal pruned decision tree. The example in the figure is pruned to get a green tree diagram, retaining two decision nodes and three leaf nodes.

[0103] Fig. 9 This is a schematic diagram of the 5-fold cross-validation principle involved in the present invention.

[0104] by Fig. 9 For example, the 5-fold cross-validation model reliability assessment is divided into 4 steps:

[0105] Step 1. Dataset division. Divide the dataset randomly and evenly into 5 parts. The original dataset in the figure has 5 data items, which are divided into 5 parts, each with 1 data item.

[0106] Step 2. Model training. Select one subset as the validation set and the remaining four subsets as the training sets for model training. The example in the figure is divided into five training sets and test sets, and five decision tree models are obtained after training.

[0107] Step 3. Model validation. Calculate the mean square error (MSE), mean absolute error (MAE), and determination coefficient (R) of the test set and training set. 2 , and the explained variance Evar is used for model validation. The five regression models in the figure are verified and five model scores are obtained. The formulas for explained variance and mathematical expectation are shown in formulas (3) and (4).

[0108] Step 4. Calculate the average performance index. Calculate the average of the five model performance indicators to evaluate the performance of the model.

[0109]

[0110] Where Evar represents the explained variance; E x Represents the mathematical expectation of an independent variable (or feature) x.

[0111] The present invention performs the steps of complete search feature selection by the packaging method and finding the optimal parameters of the decision tree regression model by Bayesian optimization simultaneously. For model optimization, the decision tree regression (DecisionTreeRegressor) module in the "scikit-learn" library in Python3 is used to establish the decision tree regression model. The Bayesian search (BayesSearchCV) module is used for Bayesian optimization (the input training set data at this time is the data after being screened by the filtering method and the embedding method). Set the node splitting criterion (criterion) to Friedman mean square error (friedman_mse), mean square error (mse) or mean absolute error (mae); the maximum tree depth (max_depth) is 3 to 9; the minimum number of samples required for internal node re-division (min_samples_split) is 2 to 8; the minimum number of samples required for leaf nodes (min_samples_leaf) is 1 to 4; the parameter classification feature number (max_features) is to consider all features (None), the square root of the total number of features is the maximum number of features (sqrt) or the maximum number of features is the logarithm of the total number of features (log2); the weight parameter (class_weigh) is the same for all classes (None) or the weight parameter is inversely proportional to the number of samples in the class (balanced); the splitting strategy (splitter) is the best split (best) or random split (random).

[0112] For full search feature selection, use the "cross_val_score" module in Python3 to train and 5-fold cross-validate each feature combination, use the "for" loop to traverse all possible decision tree regression feature combinations, calculate the best cross-validation score of all feature combinations and parameter combinations, and implement full search feature selection and Bayesian optimization parameter tuning.

[0113] After calculation, the best scores of 6 groups of features exceeded 75%, which met the model accuracy requirements. They are {G,t2}, {G,B C}, {G,B C ,std t2}, {G,B C ,t2}、{G,B C ,t2 std t2} and {G, t2, A}, a CHART decision tree regression model was established for the six feature combinations. The model input parameters were set according to the model optimization results. The model was pruned after cost complexity, and the five-fold cross validation method was used for model verification.

[0114] In the post-pruning part of the cost complexity, the grid search method is used to specify the search range of the cost complexity (ccp_alpha) parameter from 0 to 1, and the search step is 0.01. Call the "fit" function to fit the feature factors and the objective function to find the optimal cost complexity parameter to control the degree of pruning. In the model validation part, the "cross_val_score" module is used to perform random 5-fold cross training, and the mean square error MSE, mean absolute error MAE, and determination coefficient R of the 5 training test sets are calculated. 2 , the average of the explained variance Evar is used to verify the model performance. When the mean square error MSE and mean absolute error MAE of the regression model are less than 0.05, the determination coefficient R 2 If the sum of the explained variance Evar is greater than 75%, the model is considered to have passed the test and is used for heat consumption classification calculation.

[0115] After model training, the feature combination {G, t2} and the feature combination {G, B C The two models (these two combinations can be used as leaf nodes for subsequent classification to perform preliminary classification of thermal power stations) passed the model verification, and the models have strong interpretability. The model verification results are shown in Table 2.

[0116] Table 2 Performance verification results of the decision tree regression model for heat consumption of thermal power stations

[0117]

[0118] Step 3: Decision tree heat consumption classification model verification and post-processing:

[0119] For the obtained decision tree regression results, the classification benchmark calculation cannot be performed directly. The model needs to be post-processed to eliminate unreasonable results. The obtained regression tree results are analyzed based on the knowledge in the heating field. According to the results, the thermal power stations are reasonably classified, and typical thermal power stations that can represent the operating characteristics of each type are selected to establish the subsequent benchmark model. The present invention establishes an 8-step decision tree heat consumption classification model post-processing process, such as Fig.10 shown.

[0120] Fig.10 This is a post-processing flow chart of the decision tree regression heat consumption classification model involved in the present invention.

[0121] The mean absolute percentage error (MAPE) of the prediction results of each thermal station in each regression model is calculated according to formula (5).

[0122]

[0123] It has been verified that the feature combination {G, t2} and the feature combination {G, B CThe prediction results of the decision tree regression model can effectively reflect the heat consumption of the thermal power station.

[0124] Based on the establishment of a decision tree regression model for the heat consumption of thermal power stations, the heat consumption of thermal power stations is classified through three steps: removing abnormal samples, initial classification of leaf nodes, and reclassification of model comparison heat consumption. The heat stations corresponding to the leaf nodes of the regression tree after the unreasonable samples are removed are counted, and the heat stations are initially classified according to the leaf nodes of each model. Then, the categories with a small number of heat stations in the initial classification results are counted, and these heat stations are divided into leaf node categories with more similar characteristics and more adjacent results. First, the heat stations that are individually classified into one category are merged, and then the categories with 2 heat stations are merged. The categories with a small number of heat stations are gradually merged to reduce the number of categories for heat consumption classification of thermal stations.

[0125] After calculation, all heating stations are finally divided into 13 categories, of which categories 1 to 6 are floor radiant heating heating stations; categories 7 to 13 are radiator heating heating stations. A heat consumption classification model based on a decision tree is established. The classification model uses the filtered features (feature combination {G, t2} or feature combination {G, B C}) Classify the heat consumption of thermal power stations;

[0126] For each type of thermal power station, the Person correlation coefficient between each energy consumption characteristic and the daily average heat consumption is calculated again, and factors with a correlation above 0.7 are screened. It is determined that the factors with significant correlation with the real-time heat consumption of the thermal power station are: primary flow, outdoor temperature, and secondary supply and return water average temperature;

[0127] Figure 11-13 It is a box-and-whisker diagram of heat consumption, flow rate and comprehensive physical property coefficient of each type of thermal power station involved in the present invention.

[0128] from Figures 11 to 13 It can be seen that the flow, heat consumption and average temperature of secondary supply and return water of heating stations from 1 to 13 show an increasing trend, while the comprehensive physical property coefficient of the building shows a decreasing trend. That is, the heat consumption, flow and average temperature of secondary supply and return water of the radiator heating heating station are generally higher than those of floor radiant heating, and the comprehensive physical property coefficient of the building is generally lower than that of floor radiant heating. According to the definition, B C The value is the ratio of the product of the heat dissipation area and heat transfer coefficient of the heating system under the thermal power station to the product of the building envelope area and heat transfer coefficient under the thermal power station. Therefore, the higher the comprehensive physical property coefficient of the building, the better the thermal insulation performance of the building under the thermal power station, the smaller the heat demand of the thermal power station, and the lower the corresponding secondary supply and return water average temperature, flow rate and heat consumption, which is consistent with the actual heating law. The classification results can effectively reflect the heat consumption characteristics of various types of thermal power stations, indicating that the constructed heat (cold) consumption classification model has a good classification effect, and this model is used to classify thermal power stations in the later stage.

[0129] The heat consumption data of each type of thermal power station were further analyzed to verify the rationality of the classification results. The daily operation data of each type of thermal power station after classification were statistically analyzed, and the Person correlation coefficient between the heat consumption of each type of thermal power station and parameters such as outdoor temperature and secondary supply and return water average temperature was calculated. After analysis, it was found that there was a significant correlation between heat consumption and primary flow, outdoor temperature and secondary supply and return water average temperature.

[0130] Step 4: Determine the antecedent and consequent of the rule:

[0131] According to the analysis results of step 3 (indicating that heat consumption is highly correlated with primary flow, outdoor temperature and secondary supply and return water average temperature, and can be used as rule antecedents), the primary flow, outdoor temperature, secondary supply and return water average temperature and typical user indoor temperature of the thermal power station are selected as the rule antecedents for association rule mining, and heat consumption is selected as the rule posterior.

[0132] Association rule mining requires the use of data with discrete attributes as the antecedent and consequent of the rule. The present invention uses the K-means clustering algorithm to discretize the primary flow rate, heat consumption, and secondary supply and return water average temperature data. The number of clusters is determined by the silhouette coefficient; the equal-width discretization method is used to discretize the indoor and outdoor temperature data, and a discrete interval is set every 2°C. Through calculation, the optimal cluster numbers of the primary flow rate, heat consumption, and secondary supply and return water average temperature are 13, 12, and 13, respectively. K-means clustering discretization is performed on these three types of data according to the optimal cluster number. At the same time, the indoor and outdoor temperature data are discretized with equal width, and the indoor temperature is finally discretized into 7 intervals and the outdoor temperature is discretized into 23 intervals. The discretization results are shown in Tables 3 and 4.

[0133] Table 3 Discrete intervals for mining association rules of indoor temperature in heating stations

[0134]

[0135] Table 4 Discrete intervals of association rules for mining outdoor temperature in heating stations

[0136]

[0137] Step 5: Use the Apriori algorithm to mine association rules on the discretized data of each type of thermal power station to determine the heat consumption benchmarks of 13 types of thermal power stations.

[0138] The Apriori algorithm is used to discover frequent item sets, with the minimum support set to 0.002, the minimum confidence set to 60%, and the minimum lift set to 1. The support, confidence, and lift are calculated according to formulas (6), (7), and (8) to generate association rules. Finally, a total of 17,167 original rules are generated for typical thermal power stations. The number of rules generated for each type of typical thermal power station is shown in Table 5.

[0139]

[0140] Among them, X and Y represent the antecedent and consequent of association rule mining respectively; Support(X→Y) represents the support of the association rule; P(Y∪X) is the probability that X and Y coincide in the data set, Count(XY) represents the number of times X and Y appear at the same time; D is the size of the total data set; Confidence(X→Y) represents the confidence of the association rule; Lift(X→Y) represents the lift of the association rule, and Support(X) represents the confidence of the antecedent of the rule.

[0141] Table 5 Original rules generated by mining association rules of heat consumption of thermal power stations

[0142]

[0143] Step 6: Rule post-processing:

[0144] The mined original association rules cannot be directly used for the dynamic benchmark calculation of heat consumption, and rule post-processing is required, including rule grouping, rule comparison and expert analysis. First, the association rules corresponding to all typical thermal power stations are grouped, and the rules with the same outdoor temperature range are grouped together; then the rules are compared to extract the rules with higher confidence (rules with confidence greater than the minimum confidence of 60%); finally, the rules are analyzed through the knowledge of heating domain, and the unreasonable rules are replaced with rules with slightly lower confidence (slightly lower confidence means rules with confidence lower than the current unreasonable rules) but reasonable (in line with the knowledge of heating domain), and the first rule that meets the confidence greater than the minimum confidence and conforms to the knowledge of heating domain is selected to replace the unreasonable rules. Finally, each thermal power station generates 23 rules, corresponding to 23 outdoor temperature ranges. Taking the category 3 floor heating thermal power station 168 and the category 10 radiator heating thermal power station 74 as examples, the intervals of the antecedent and contingent of the rules of the two thermal power stations and their support, confidence and promotion indicators are shown in Tables 6 and 7 respectively.

[0145] Table 6 Generation rules for mining association rules of heat consumption of type 3 thermal power station

[0146]

[0147]

[0148] Table 7 Generation rules for mining association rules of heat consumption of type 10 thermal power station

[0149]

[0150]

[0151]

[0152] Step 7: Dynamic Benchmarking of Heat Consumption:

[0153] The association rules generated by 13 typical thermal power stations were used to calculate the dynamic heat consumption benchmarks of each type of thermal power station. The operating conditions and heat consumption rules of the thermal power station in each outdoor temperature range were counted. The upper and lower bounds of the heat consumption range of the rule posterior term were used to calculate the dynamic heat consumption benchmarks 1 and 2 of the thermal power station respectively.

[0154] Figure 14-15 It is a heat consumption benchmark chart, and the selection is based on the outdoor temperature.

[0155] Fig.14 , Fig.15 They are respectively the dynamic heat consumption benchmarks of floor radiant heating and radiator heating thermal power stations in the embodiments of the present invention. The black curve in the figure represents the outdoor temperature in the heating season, and the other curves of different colors represent the dynamic heat consumption benchmarks of various types of floor radiant heating thermal power stations in the entire heating season.

[0156] Depend on Fig.14 It can be seen that the heat consumption benchmark has an obvious negative correlation with the outdoor temperature. The heat consumption benchmarks of type 1 and type 2 heating stations are the lowest, and the heat consumption benchmarks of type 3 and type 4 heating stations are at a medium level; the heat consumption benchmark of type 5 heating station is relatively high among the 6 types of floor heating heating stations, and the heat consumption benchmark of type 6 heating station is the highest. Fig.15 It can be seen that the heat consumption benchmark has an obvious negative correlation with the outdoor temperature. The heat consumption benchmark of type 8 thermal power station is the lowest, followed by type 11 thermal power station; the heat consumption benchmarks of type 7, type 9, type 10, and type 12 thermal power stations are relatively high; the heat consumption benchmark of type 13 thermal power station is the highest.

[0157] Based on the above description, relevant staff can make various changes and modifications without departing from the technical idea of ​​this invention.

[0158] Where the present invention is not described, reference is made to the prior art.

Claims

1. A method for determining a dynamic benchmark of heat consumption / cooling capacity of a HVAC system, characterized in that: The method comprises the following steps: Step 1: Feature Selection The HVAC system data under different working conditions are obtained. After eliminating abnormal data, the Person correlation coefficient between each energy consumption feature and heat (cold) consumption is calculated, and the features with a Person correlation coefficient greater than or equal to 0.24 are retained. At the same time, the importance index of heat (cold) consumption is calculated using the ReliefF algorithm, and the features with an importance index greater than or equal to 0.003 are selected. After merging the features selected by the Person correlation coefficient and the ReliefF algorithm, the random forest embedding method is used to select the features with random forest importance index exceeding 0.3 as the main features affecting energy consumption; Step 2: Construct a heat (cooling) consumption classification model: The heat (cold) consumption classification model is based on a decision tree model. The nodes of the decision tree include a root node, a leaf node, and a decision node. The root node of the decision tree is a collection of all thermal power station data. The decision node corresponds to the division rule of energy consumption characteristics. The decision node divides the data on the root node into several subsets according to the energy consumption characteristic rule. The leaf node is the final energy consumption classification node. The thermal power stations in the leaf node are classified into the same category to realize the classification of thermal power station heat consumption. The input features of the decision tree model are the main features affecting energy consumption and the corresponding heat consumption determined in step 1. Bayesian optimization is used to find the optimal parameters of the decision tree model. At the same time, the packaging method is used to completely search the features to determine the decision tree regression feature combination with the best score exceeding 75%; Then, the decision tree model is pruned using the cost complexity post-pruning method to remove unnecessary nodes, and the heat (cooling) consumption classification model and the final input features of the model are obtained; Step 3: Use the heat (cooling) consumption classification model constructed in step 2 to classify the HVAC system data using the final input features; Step 4: For each type of HVAC system data, the Person correlation coefficient between each energy consumption feature and heat (cooling) consumption is calculated again. All energy consumption features with a Person correlation coefficient greater than 0.7 are used as the rule antecedents for association rule mining, and the heat (cooling) consumption is used as the rule posterior; Step 5: Discretize the HVAC system data, and after discretization, use the Apriori algorithm to mine association rules according to the rule antecedents and rule consequents determined in step 4 to obtain association rules for all classes; Step 6: Perform rule processing on the association rules of all classes obtained in step 5 to obtain available rules, and use the available rules to determine the dynamic benchmark of the heat (cold) consumption of the HVAC system.

2. The method according to claim 1, characterized in that The process of rule processing is: Discretize the outdoor temperature with equal width to obtain different outdoor temperature intervals. Group all the class association rules obtained in step 5 according to whether they are in the same outdoor temperature interval, and determine the confidence level of the rules in each group. Set the minimum confidence level and extract all rules with confidence levels greater than the minimum confidence level for expert analysis; The rules are analyzed through the knowledge of heating domain, and unreasonable rules are replaced by adjacent and reasonable rules with lower confidence.

3. The method according to claim 1, characterized in that The process of determining the dynamic benchmark of dynamic heat (cold) consumption is: statistics the operating conditions and available heat consumption rules under all outdoor temperature ranges, and calculate the dynamic heat (cold) consumption benchmark 1 and benchmark 2 of the thermal power station based on the upper and lower limits of the heat (cold) consumption range after the rules.

4. The method according to claim 1, characterized in that In step 5, the discretization process is: discretizing the data of the primary pipe network flow, pressure, outdoor air temperature, and secondary pipe network supply and return water average temperature parameters of the HVAC system; using K-means clustering discretization for operation data and energy consumption data, and using equal-width discretization method for indoor and outdoor air temperatures.

5. The method according to claim 1, characterized in that The decision tree model is pruned using the cost complexity post-pruning method. Random 5-fold cross-training is used for model verification, and the mean square error (MSE), mean absolute error (MAE), and determination coefficient (R) of the five training and test sets are calculated. 2 , the average value of the explained variance Evar, when the mean square error MSE and mean absolute error MAE of the model are less than 0.05, and the determination coefficient R 2 If the sum of the explained variance Evar is greater than 75%, the model is considered to have passed the test and is used to classify the heat (cooling) consumption.

6. The method according to claim 1, characterized in that The HVAC system is a thermal power station or an energy station.

7. The method according to claim 1, characterized in that In step 4, the rule antecedents of association rule mining are the primary flow rate of the energy station or thermal station, the outdoor temperature, the secondary supply and return water average temperature, and the indoor temperature of typical users.

8. The method according to claim 1, characterized in that In the Apriori algorithm, the minimum support is set to 0.002, the minimum confidence is set to 60%, and the minimum lift is set to 1.

Images