CN114386142B - Building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework - Google Patents
Building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework Download PDFInfo
- Publication number
- CN114386142B CN114386142B CN202111568679.0A CN202111568679A CN114386142B CN 114386142 B CN114386142 B CN 114386142B CN 202111568679 A CN202111568679 A CN 202111568679A CN 114386142 B CN114386142 B CN 114386142B
- Authority
- CN
- China
- Prior art keywords
- feature
- prediction
- energy consumption
- fuzzy
- building energy
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/06—Power analysis or power optimisation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Geometry (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Architecture (AREA)
- Civil Engineering (AREA)
- Structural Engineering (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
本发明涉及一种基于多源融合特征选择和模糊差异性增强Stacking框架的建筑能耗预测方法，包括：采集建筑能耗历史数据，构建建筑能耗预测数据集，对建筑能耗预测数据集进行预处理，并将处理后的数据集划分为训练集和预测集；对预处理后的数据集进行特征工程，利用多源融合特征方式选出最优特征子集和最优特征筛选个数；选择多个基模型，对基模型进行超参数优化，以组成Stacking框架的第一层，进而构建模糊差异性增强Stacking框架，利用构建的模糊差异性增强Stacking框架进行建筑能耗预测。与现有技术相比，本发明具有提高预测性能等优点。
The present invention relates to a building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework, comprising: collecting historical building energy consumption data, constructing a building energy consumption prediction data set, preprocessing the building energy consumption prediction data set, and dividing the processed data set into a training set and a prediction set; performing feature engineering on the preprocessed data set, selecting an optimal feature subset and an optimal feature screening number using a multi-source fusion feature method; selecting multiple base models, performing hyperparameter optimization on the base models to form the first layer of the Stacking framework, and then constructing a fuzzy difference enhanced Stacking framework, and using the constructed fuzzy difference enhanced Stacking framework to predict building energy consumption. Compared with the prior art, the present invention has the advantages of improving prediction performance and the like.
Description
Technical Field
The invention relates to the technical field of building energy consumption prediction, in particular to a building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhancement Stacking framework.
Background
With the current advancement of the global sustainable development industry, building energy conservation has gradually been seen as a key issue by many countries and regions. Under the demand of promoting urban sustainable development, it is necessary to know the energy-saving characteristics of urban buildings. In order to meet the current energy-saving requirement, a corresponding building energy-saving policy is formulated, and a building energy consumption prediction model becomes an indispensable tool. The method can know the influence of different characteristics, such as weather, residence ratio, date and the like, on the energy consumption of the building, so as to guide how to improve the energy use efficiency of the building and reduce the energy consumption. The accurate energy consumption prediction can effectively reduce energy waste and regional financial expenditure.
Current predictive goals for building energy consumption mainly include building interior heat gain, building cold load, regional heat load, power demand, and peak power demand. In the energy management system, the service objects of the prediction model are mainly optimal control and fault detection. Optimizing control includes matching the supply and demand of building energy, maintaining indoor thermal comfort, and optimizing unit and system operation. Therefore, an accurate building energy consumption prediction model is built by a more efficient and reasonable means, and the method is an effective method for effectively controlling the current situation of energy consumption and improving the condition of energy waste. The continuous improvement of the accuracy, speed and generalization performance of the prediction algorithm is also a key for ensuring efficient building operation.
Although prediction methods based on machine learning prove to have good prediction accuracy and speed, different data sets have respective characteristics for different actual energy consumption scenarios. It is a very difficult procedure to directly select the corresponding individual predictive model from the data set. And the generalization performance is poor due to the large space of the hypothesis when the single algorithm predicts. Such a predictive model with poor generalization performance will greatly affect the subsequent management of building energy consumption and energy deployment. Thus, the idea of ensemble learning, which has gradually become popular for energy consumption prediction in recent years by virtue of high generalization performance and high accuracy, was introduced.
The integrated model integrates multiple single models through a combination strategy, thereby achieving better performance than any single model therein. The mechanism well compensates for the unilaterality of a single prediction algorithm, and the problem of over-expression in a hypothetical space can be solved. The integration algorithm can be divided into Boosting, bagging and Stacking according to the current combination strategy. Compared with Boosting and Bagging, stacking has been demonstrated to have better predictive and generalization performance. Stacking uses the outputs of the plurality of base learners to generate a new modified data set and then predicts via the meta-model. As an integrated framework, it can obtain better predictive performance than any base model, with high robustness. Most of current research on a Stacking framework mainly stays in application, and a Stacking prediction model is built by selecting a base model. However, because the Stacking framework is built, although cross-validation methods are used to reduce over-fitting, the similarity that exists between the output results of the base learner still leads to the over-fitting problem of the predictive model. This problem reduces the generalization of learning by the Stacking second-level learner, thereby affecting the final accuracy of the prediction model.
In addition, in the prediction of building energy consumption, the data acquisition process often involves multiple aspects, and the formed data set also often has a large dimension. In this dataset, not all features typically contribute forward to the prediction result. The existence of useless features in the data set not only affects the precision of the prediction model, but also greatly reduces the construction speed of the model. Whereas the single feature selection algorithms typically used often only represent one aspect of the dataset, this one-sidedness results in a certain limitation to the generalization performance of the predictive model. Moreover, how to obtain the optimal number of features, i.e. how many features to keep as input to the predictive model, is always a difficulty in feature selection.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhancement Stacking framework.
The aim of the invention can be achieved by the following technical scheme:
a building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhancement Stacking framework comprises the following steps:
s1, data acquisition and preprocessing, wherein the specific steps comprise:
s101, collecting historical data of building energy consumption, determining and collecting weather factors, human factors and the like which influence the prediction of the building energy consumption, and constructing a prediction data set of the building energy consumption by taking the weather factors, the human factors and the like as main influence factors;
in the step, as a preferable mode, building energy consumption historical data is collected through an intelligent energy management system;
S102, carrying out normalization processing on each feature according to various features in the building energy consumption prediction data set, so that different features are in the same dimension for subsequent comparison of importance degrees of the different features;
And S103, dividing the processed data set into a training set and a test set, training the prediction model by using training data, and evaluating different prediction models by using the test data.
S2, selecting multisource fusion characteristics, wherein the specific steps include:
S201, carrying out feature engineering on the normalized data set, and selecting an optimal feature subset and an optimal feature screening number aiming at the current data set;
S202, performing preliminary evaluation of each characteristic of the data set through a variance selection method (VT, varianceThreshold), recursive characteristic elimination (RFE, recursive feature elimination), XGBoost and linear characteristic selection (LR, linear Regression) respectively;
the variance selection method filters out corresponding features through variances of the features, namely when the variances of one feature are smaller than the variances of other features, the information content contained in the feature is smaller than the variances of other features, the influence on the final prediction result is smaller, but the running time of the prediction model is greatly increased, and therefore the feature is removed. The specific variance filtering formula is as follows:
Where S 2 is the variance of the feature itself, x i is the value of each sample in the feature, x mean is the average of all samples in the feature, and n is the number of samples. In the present invention, a sample is a specific sampling point under a certain characteristic, i.e., a sample point acquired in time.
Recursive feature elimination as a basic model of fusion feature selection essentially uses a reverse selection technique. The recursive feature elimination starts its search process starting with the entire feature network and filters the effect of each feature on the predicted outcome. Some ranking criteria are computed that assess feature importance, feature ordering and model performance are stored for final feature selection. The least relevant features are gradually eliminated by each iteration of the loop and a subset of the input variables is updated. This process is iterated until no further variables need to be deleted.
S203, after importance evaluation is carried out on each feature according to the self mechanism of different feature selection algorithms, feature importance scores of four different feature selection algorithms are obtained, feature importance under different algorithms are respectively ordered, and output results aiming at different feature selection algorithms are obtained under different feature subset number requirements:
RF＝[XF(1),...,XF(k),...XF(m)]
Wherein F represents the four different feature selection algorithms selected above, X F (k) is the optimal sub-feature selected by the current feature selection algorithm, and m is the total number of features to be screened out by the optimal feature subset.
S204, comparing the prediction performance under different feature selection numbers in sequence, and when the current feature selection number is determined, sequentially selecting corresponding features through sequencing values of different algorithms and integrating the features into an optimal feature pre-selection set S all, wherein the expression is as follows:
Wherein R VT、RRFE、RXG and R LR are feature subsets screened by a variance selection method, recursive feature elimination, XGBoost and linear feature selection respectively, and the optimal pre-selection set is fused with a plurality of feature selection algorithm results, and the occurrence frequency of the internal optimal features is at most 4 and at least 1.
At this time, the occurrence frequency of each feature in the optimal prediction set is counted, and then the fusion importance value R S of the features screened by the four algorithms is obtained.
Where X S(1),...,XS(k),...XS (d) is the number of occurrences of each feature in the optimal feature pre-set.
The fusion importance value R S is used for later screening out an optimal feature subset to be laid, the optimal feature subset is screened through ranking of fusion importance values of different features, and the top m features with large fusion importance values are selected as the optimal feature subset.
And S205, finally traversing the prediction performance under all the different feature selection numbers to obtain the optimal feature screening number so as to optimize the original data set, and taking the original data set as the input of a subsequent prediction model. The predictive model is a selected base model.
And S3, carrying out single-model super-parameter optimization on the optimal feature subset obtained in the S2, and building a fuzzy-difference-based enhanced Stacking framework.
S301, selecting a base model for the whole framework of the Stacking framework to form a first layer of the Stacking framework. And performing performance evaluation of different base models through the optimal feature subset after the S2 feature engineering, and selecting XGboost, random Forest and LightGBM as the base models of the frames.
S302, performing super-parameter optimization on the three base models through random search, so that a single model with excellent performance is obtained and is used as the base model of the first layer of the Stacking framework to build the integration framework.
And S303, for building the Stacking integrated frame, integrating N different basic models in a first layer in a two-layer structure of the Stacking frame to form a basic model layer of the integrated frame, and taking one prediction model as a meta model to fit the first layer output and serve as a second layer of the integrated frame. If the input feature is X i, the kth base model of the first layer is F k and the predictive model of the second layer is F. The output of the kth basic model of the first layer is F k(Xi), the specific expression of the output is as follows:
Yi＝F(F1(Xi),…,Fk(Xi)…,FN(Xi))
the base model is trained by means of cross-validation. Firstly, dividing an original training set into k parts, wherein each part comprises a verification set and a test set, wherein k-1 parts are taken for training each time, the other 1 part is used for verification, the k times are circulated, and the prediction results of the training set and the test set are taken as S and M:
For basic model 1 to basic model 3, training set S 1～S3 and test sets M 1-M3.S1～S3 and M 1-M3 of each basic model are obtained respectively, and combined training set and combined test set are obtained respectively. The combined training set S formula is shown below, which applies equally to the combined test set M:
s304, obtaining an error e i (n) of the current sample through the difference between the predicted value and the true value output by each base model.
Wherein the method comprises the steps ofFor the predicted value of the i-th base model on the current sample value, n is the number of samples, and y (n) is the tag realism value for the current sample.
S305, after obtaining the predicted value error output by each base model aiming at the training set, judging the influence of the current error value on the predicted sample. Therefore, the error rate of each sample is further calculated to measure the magnitude of the above-mentioned effect.
eci(n)＝ei(n)/y(n)
Where e ci (n) is the prediction error rate for each sample of each base model for the current training set.
And taking the calculated error and error rate as two input parameters of the fuzzy difference enhancement layer. By corresponding study of the output of the Stacking first layer, the resulting error threshold range for the strong learner was found to be approximately the same. The same fuzzy rule is set to differentially enhance the output of the base model. The fuzzy difference enhancement layer is arranged into a two-dimensional structure, and is provided with double inputs and single outputs.
S306, inputting two input variables of error and error rate into a fuzzy difference enhancement layer, and carrying out fuzzy reasoning through fuzzy rules to fuzzify the two variables into E and Ec. And finally obtaining the predicted differential enhancement coefficient b i(0<bi < 0.6). And multiplying the error of each sample with the self differential enhancement coefficient to obtain a differential enhancement value, and integrating the differential enhancement value with the output of the base model to obtain a new training set S 'and a new testing set M'.
S307, a second layer of metamodel is used to train S 'and predict M' to obtain the final output:
Compared with the prior art, the building energy consumption prediction method based on the multisource fusion feature selection and fuzzy difference enhancement Stacking framework provided by the invention at least has the following beneficial effects:
1) The invention merges four different types of single feature selection algorithms in the feature selection stage, thereby comprehensively considering the importance degree of a certain feature in the feature selection process, and improving the generalization performance of the whole feature engineering; in addition, the fuzzy difference enhancement layer is added on the basis of the traditional Stacking framework in a targeted manner, and the overfitting risk of the integrated framework is further reduced by increasing the output difference of the base model, so that the prediction performance of the integrated framework is further improved;
2) The multi-source fusion feature selection in the invention can automatically screen out the optimal feature number and the optimal feature subset aiming at the data set, thereby greatly improving the construction speed of the integrated frame on the premise of ensuring the accuracy of the prediction model.
Drawings
FIG. 1 is a diagram of a multi-source fusion feature selection framework of a building energy consumption prediction method based on the multi-source fusion feature selection and fuzzy diversity enhancement Stacking framework in an embodiment;
FIG. 2 is a fuzzy layer dual input single output membership function setup in an embodiment, wherein sub-graph (a) is a membership function established with error values, sub-graph (b) is a membership function established with error rates, and sub-graph (c) is an output fuzzy difference correction coefficient;
FIG. 3 is a diagram of a fuzzy variability enhancing Stacking overall framework in an embodiment.
Detailed Description
The invention will now be described in detail with reference to the drawings and specific examples. It will be apparent that the described embodiments are some, but not all, embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present invention without making any inventive effort, shall fall within the scope of the present invention.
Examples
In order to increase the variability of a Stacking framework base model and reduce the risk of model overfitting, thereby improving the generalization capability and the prediction performance of the whole framework, the invention provides a building energy consumption prediction method based on multi-source fusion feature selection and fuzzy variability enhancement Stacking framework, and the method provides a novel variability enhancement fuzzy Stacking framework.
Referring to fig. 1, 2 and 3, in the method, a new framework uses a fuzzy rule to strengthen the difference of output results of a first layer of base model, calculates errors and error rates between the prediction results and the true values output by the base model, judges the two as inputs of a fuzzy difference enhancement layer, and obtains a model difference enhancement coefficient through the fuzzy rule. And finally, integrating the differential enhancement coefficient and the error to obtain enhancement values, and respectively superposing the output of each base model and the respective enhancement values. In addition, in the prediction of building energy consumption, the data acquisition process often involves multiple aspects, and the formed data set also often has a large dimension. In this dataset, not all features typically contribute forward to the prediction result. The existence of useless features in the data set not only affects the precision of the prediction model, but also greatly reduces the construction speed of the model. Whereas the single feature selection algorithms typically used often only represent one aspect of the dataset, this one-sidedness results in a certain limitation to the generalization performance of the predictive model. Moreover, how to obtain the optimal number of features, i.e. how many features to keep as input to the predictive model, is always a difficulty in feature selection. For this purpose, the method of the invention proposes a multisource fusion feature selection algorithm. By fusing four feature selection algorithms with different tendencies, the optimal feature number aiming at the data set is automatically selected, and then a feature selection result with high generalization performance is obtained. The method can eliminate unimportant features and retain the feature subset with the largest information quantity.
Based on the improvement thought, during actual prediction, building energy consumption historical data, weather data and building residence ratio are collected as main factors to form an original data set for building energy consumption prediction. And carrying out normalization processing on each feature according to each feature in the data set, so that different features are in the same dimension for subsequent comparison of importance degrees of different features. The processed data set is divided into a training set and a test set, training of the prediction model is performed by using training data, and the test data are used for evaluating different prediction models. And carrying out feature engineering on the preprocessed data set to select the optimal feature subset and the optimal feature screening number aiming at the current data set. For the normalized dataset, preliminary evaluations of the features of the dataset were performed by variance selection (VT, varianceThreshold), recursive feature elimination (RFE, recursive feature elimination), XGBoost (eXtreme Gradient Boosting), and linear feature selection (LR, linear Regression), respectively. After the feature importance value scores of the four different feature selection algorithms are obtained, the feature importance under the different algorithms are respectively ranked. And comparing the prediction performance under different feature selection numbers sequentially. When the current feature selection number is determined, corresponding features are sequentially selected through sorting values of different algorithms and integrated into an optimal feature pre-selection set. The optimal pre-selection set fuses a plurality of characteristic selection algorithm results, and the occurrence frequency of the internal optimal characteristics is at most 4 and at least 1. At this time, the occurrence frequency of each feature in the optimal prediction set is counted, and then the fusion importance value of the features screened by the four algorithms is obtained. And finally traversing the prediction performance under all different feature selection numbers to obtain the optimal feature screening number so as to optimize the original data set, and taking the original data set as the input of a subsequent prediction model.
For the overall architecture of the Stacking framework, the selection of the base model is required to make up the first layer of the Stacking framework. And performing performance evaluation of different models through the optimal subset after feature engineering, and selecting XGboost, random Forest and LightGBM as base models of the frames.
XGBoost consists of a series of weak learners, and thus perfects the gradient lifting algorithm. Gradient boosting is based on iterative estimation of the tree on the residuals obtained at each step and adaptive updating of the estimates. Gradient lifting uses gradient descent techniques, for which a segmentation is chosen that facilitates approaching the minimum of the objective function. XGBoost is much faster than other common machine learning methods because it can efficiently process large amounts of data in parallel. The present invention selects XGBoost as one of the basic models of the Stacking framework.
The decision tree submodel in LightGBM is to split the nodes by leaf splitting, so its computation cost is relatively small. The decision tree algorithm based on the Histogram is selected, the characteristic value is divided into a plurality of bins, and the optimal splitting points are searched on the bins, so that the storage cost and the calculation cost are reduced. In addition, the processing of LightGBM on the category characteristics also enables the effect of the category characteristics to be better improved under specific data. Compared with XGBoost, the memory occupation is less and the accuracy is higher. Since LightGBM and XGBoost have different algorithmic patterns, they have different tendencies for the information of the dataset. Therefore, the two types of Boosting strong learners are used as complementary parts of the base model, so that the learning capacity and the final prediction effect of the whole Stacking framework are improved.
As a representative algorithm of the Bagging integrated model, there is a different operation mode from the Boosting algorithm. The RF's internals integrate multiple trees and each tree is trained using independently sampled random vectors. During the training phase of RF, a plurality of different sub-training data sets are collected from the input training data set using boottrap technology. And training a plurality of different decision trees in turn, and then fitting these decision trees to each sub-sample of the dataset. In the prediction stage, the final predicted value of the model is obtained by calculating the average value of the prediction results of CART in RF. By the method, the prediction accuracy can be effectively improved, and the risk of overfitting is reduced. RF has the advantages of less parameter setting, fast convergence and the like, and can still show stronger generalization performance when processing a large amount of data. The method can effectively avoid the problem of overfitting, and is suitable for the energy consumption prediction scene with large data volume. Thus, the present invention uses RF as one of the base models of the Stacking framework.
And performing super-parameter optimization on the three base models through random search, so that a single model with excellent performance is obtained and is used as the base model of the first layer of the Stacking framework to build the integration framework. And constructing the whole frame by taking three single algorithms after super-parameter optimization as a base model of the Stacking frame. Five-fold cross-validation was performed against the base model to reduce the risk of overfitting of the entire framework. After the five-fold cross validation is carried out on the base model, the obtained output result is compared with the real predicted value, and the error rate of each sample are calculated.
And calculating the ambiguity enhancement coefficient by taking the errors and error rates of different samples of different obtained algorithms as two inputs of the ambiguity enhancement layer. Wherein, setting the fuzzy rule for the fuzzy difference enhancement layer through expert experience. The present invention maps errors and error rates into five-level membership functions. Under the condition that the positive value and the negative value of the error are considered independently, the absolute value membership function of the error is taken to be 60, the membership of the error rate is taken to be 1, and the membership function is set to be a triangle membership function. Under the fuzzy rule, the model difference enhancement coefficient is taken to be 0 to 0.6.
After obtaining the corresponding fuzzy differential enhancement coefficient, multiplying the error of each sample with the differential enhancement coefficient to obtain a differential enhancement value, and integrating the differential enhancement value with the output of the base model to obtain a new training set and a new testing set.
The overall learning mode of the Stacking framework easily results in overfitting, so the learner of the second layer typically selects a simpler model for integrating the output results of the base model. The invention selects the logistic regression for carrying out regression prediction on the output result of the base model with enhanced diversity.
The method is verified according to actual data, preprocessing of building energy consumption data sets is selected based on the multi-source fusion characteristics, and short-term prediction of building energy consumption is conducted through a fuzzy difference-based reinforced Stacking framework. According to the embodiment, six evaluation indexes are used for carrying out a comparison experiment, and the experiment shows that the optimal feature screening number aiming at different data sets can be automatically selected by the multi-source fusion feature selection, and finally the optimal feature subset of the original data set is obtained. The method also creates corresponding conditions for subsequently improving the accuracy of the building energy consumption prediction model and reducing the model construction time. The accuracy and generalization performance of the proposed fuzzy-difference-enhanced Stacking framework are further enhanced for the prediction of building energy consumption.
The method specifically comprises the following steps:
step one, multisource fusion feature selection
And fusing the variability of the single feature selection algorithm through the multi-source fusion feature selection algorithm, and calculating the contribution degree of the features. The algorithm uses the RMSE as an evaluation index to compare the prediction effect of the optimal feature number screened by the algorithm. Aiming at the campus building energy consumption data set of the embodiment, the optimal feature number obtained by the multisource fusion feature selection algorithm is 13. The prediction results of the different feature numbers are compared to obtain table 1.
TABLE 1 comparison of predictive Performance under different feature screenings
Table 1 shows that the prediction model achieves the best prediction effect when the feature selection number is 13. This further verifies the effectiveness and convenience of the multi-source fusion feature selection. Then, as the feature number is increased, the performance of the prediction model is not improved, but the prediction accuracy of the prediction model is reduced. In addition, as the number of samples of the original data set is large, the construction speed of the prediction model is greatly reduced with the increase of the feature number. This illustrates that for a general dataset, prediction accuracy is often affected by redundant and irrelevant features, and feature selection is an indispensable key step in the prediction model building process.
In summary, the multi-source fusion feature selection algorithm provided by the invention can screen out the optimal feature subset in the data set, and shortens the construction time of the model while reducing the calculation cost. In addition, the algorithm can automatically select the feature number of the optimal feature subset, so that the unreliability of manually setting the feature number is also compensated, the overall operation time of the feature engineering is shortened, and the workload of the feature engineering is simplified.
Step two, fuzzy difference enhanced Stacking framework
And (3) performing single-model super-parameter optimization on the optimal feature subset obtained in the step one. And optimizing the super parameters of the XGboost, RF, lightGBM base models to obtain a prediction strong learner aiming at the current data set. Aiming at excellent prediction performance shown by RF, XGBoost, lightGBM integrated models, the three integrated models are used as a base learner of a first layer of a Stacking framework to build the whole integrated framework. Because the prediction performance of the Stacking framework is affected by the prediction precision of each base model, each base model is subjected to super-parameter optimization before the framework is built, so that the prediction precision of each base model is further improved. After the super-parameter optimization, the evaluation indexes of the three algorithms are shown in table 2.
TABLE 2 comparison of the predicted Performance of the base models and the Stacking framework after Supermarameter optimization
The 3 optimized prediction models are used as a base model of a first layer of a Stacking framework. To prevent the risk of overfitting, the second layer of the Stacking framework chooses to use linear regression as a metamodel to fit the output of the first layer base model. The prediction results of the Stacking framework are also set forth in table 2 to compare with the performance of each base model.
As can be seen from table 2, the Stacking framework fuses the three super-parameter optimized strong learners, thereby obtaining better prediction performance than any one of the models. Aiming at the prediction of building energy consumption, the optimal prediction effect of the three base models is LightGBM. Whereas XGBoost has an overall prediction accuracy slightly inferior to LightGBM, the prediction effect is relatively poor RF. Comparing each evaluation index of the Stacking prediction performance with LightGBM with the optimal base model prediction performance, the RMSE is reduced by 5.36%, and SMAPE is reduced by 5.84%. This also further demonstrates that the Stacking integration framework has an optimal level of accuracy compared to other advanced prediction models.
After verifying the excellent prediction performance of the Stacking framework, the invention provides a new differential enhanced fuzzy Stacking framework in order to reduce the overfitting risk during the Stacking framework construction and further improve the prediction performance of the framework. The new framework reduces the risk of model overfitting by increasing the variability of the output of the Stacking first-layer base model, so that the generalization capability and the prediction performance of the whole framework are further improved. The performance versus performance of the output results after processing through the blur difference enhancement layer for the three base models of the present invention are shown in Table 3, where F-XGBoost refers to the XGBoost model processed through the blur difference enhancement layer, F-LightGBM refers to the LightGBM model processed through the blur difference enhancement layer, and F-RF refers to the RF model processed through the blur difference enhancement layer.
TABLE 3 comparison of predictive variability between models before and after blurring
The comparison results of the various evaluation indexes output by the base models before and after the base models pass through the fuzzy difference enhancement layer are shown in table 3. It is noted that the results of the three base models in table 3 before the blur difference enhancement are not matched with the optimized base model results in table 2. This is because the results in table 2 are the final predictions made for the test set, while the model in table 3 is the result of the Stacking first layer base model obtained by cross-validation. The two training sets for the models are different, so the evaluation indexes cannot be directly compared, and the output result of the base model in table 3 is not the final prediction result.
As shown in Table 3, XGBoost, lightGBM and RF are subjected to fuzzy difference enhancement treatment, and the effect of each index is reduced to a certain extent. This means that the differences between the processed base models are enhanced to a certain extent along with the change trend of the errors and error rates of the base models, so that the information entropy of the new data set input to the Stacking second layer is increased, and the learning capacity of the whole prediction framework is enhanced. The final Fuzzy-Stacking and conventional Stacking are shown in table 4 for each evaluation index pair for the building energy consumption prediction.
TABLE 4 comparison of Fuzzy-Stacking versus Stacking predicted Performance
After processing the output of the base model by the blur difference enhancement layer, the processing result is fitted with linear regression as the second layer of the Stacking framework. As shown in Table 4, the prediction performance of the Fuzzy-Stacking system provided by the invention is further improved on the basis of the traditional Stacking framework. Wherein RMSE can be reduced to 12.349 and R2 reaches 0.979. The accuracy is further improved because after the fuzzy difference enhancement layer is passed, the coverage of the assumed space between the outputs of each base model is enlarged, so that the integral learning capacity of the model is improved. In addition, the blur difference enhancement process and the blur rule are obtained by the error and error rate of the base model. The basis leads the performance of each evaluation index output by the base model to be reduced to a certain extent, but does not reduce the prediction precision of the final Fuzzy-Stacking framework, and leads the advantages among the base models to be further complemented.
The building energy consumption prediction method based on the multisource fusion feature selection and fuzzy difference enhancement Stacking framework improves the multisource fusion feature selection process, automatically selects the optimal feature number aiming at a data set through the feature selection algorithm fusing four different tendencies, and further obtains a feature selection result with high generalization performance. The method can eliminate unimportant features and retain the feature subset with the largest information quantity. On the basis, in order to increase the variability of the Stacking framework base model and reduce the risk of model overfitting, the generalization capability and the prediction performance of the whole framework are improved, and the invention also provides a novel variability enhanced fuzzy Stacking framework. The new framework uses fuzzy rules to strengthen the difference of the output results of the first layer base model, and the error and error rate between the prediction result and the true value output by the base model are calculated, and the prediction result and the true value are used as the input of the fuzzy difference enhancement layer to judge, and then the model difference enhancement coefficient is obtained through the fuzzy rules. And finally, integrating the differential enhancement coefficient and the error to obtain enhancement values, and respectively superposing the output of each base model and the respective enhancement values. In this way, the information content of the meta-model input data generated by the Stacking first layer is increased, so that the risk of model overfitting is reduced, and the overall learning rate of the meta-model is improved. The method provides a novel prediction integration framework with high precision and high generalization performance for the field of building energy consumption prediction, and enriches the diversity of prediction models.
While the invention has been described with reference to certain preferred embodiments, it will be understood by those skilled in the art that various changes and substitutions may be made without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the protection scope of the invention is subject to the protection scope of the claims.
Claims (7)
1. The building energy consumption prediction method based on the multisource fusion feature selection and fuzzy difference enhancement Stacking framework is characterized by comprising the following steps of:
Building energy consumption historical data are collected, a building energy consumption prediction data set is constructed, the building energy consumption prediction data set is preprocessed, and the processed data set is divided into a training set and a prediction set;
Performing feature engineering on the preprocessed data set, and selecting an optimal feature subset and an optimal feature screening number by utilizing a multi-source fusion feature mode;
Selecting a plurality of base models, performing super-parameter optimization on the base models to form a first layer of a Stacking framework, further constructing a fuzzy difference enhancement Stacking framework, and performing building energy consumption prediction by using the constructed fuzzy difference enhancement Stacking framework;
the specific content for constructing the fuzzy difference enhanced Stacking framework is as follows:
1) Firstly, performing super-parameter optimization on three base models through random search to obtain a single model with excellent performance as a base model of a first layer of a Stacking framework;
2) In the two-layer structure of the Stacking framework, a first layer is integrated by three different basic models to form a basic model layer of the integration framework, and then one model is used as a meta model to fit the output of the first layer and is used as a second layer of the integration framework;
3) Performing five-fold cross validation on each base model;
4) After performing five-fold cross validation on the base model, comparing the obtained output result with a real predicted value to obtain an error and an error rate of each sample;
5) Aiming at the obtained errors and error rates of different algorithms under different samples, the errors and error rates are used as two inputs of a fuzzy difference enhancement layer, and a fuzzy enhancement coefficient is calculated, wherein a fuzzy rule is set for the fuzzy difference enhancement layer through expert experience, and the errors and error rates are mapped into five-level membership functions;
6) After the corresponding fuzzy differential enhancement coefficient is obtained, multiplying the error of each sample with the differential enhancement coefficient to obtain a differential enhancement value, and integrating the differential enhancement value with the output of the base model to obtain a new training set and a new testing set.
2. The building energy consumption prediction method based on the multisource fusion feature selection and fuzzy diversity enhancement Stacking framework according to claim 1, wherein building energy consumption historical data are collected through an intelligent energy management system, and normalization preprocessing is carried out on a building energy consumption prediction data set.
3. The building energy consumption prediction method based on the multisource fusion feature selection and fuzzy diversity enhancement Stacking framework according to claim 1, wherein the specific contents of the optimal feature subset and the optimal feature screening number selected by utilizing a multisource fusion feature mode are as follows:
after carrying out feature engineering on the preprocessed data set, carrying out preliminary evaluation on each feature in the data set through different feature selection algorithms to obtain feature importance value scores of four different feature selection algorithms;
sorting the feature importance values under different feature selection algorithms respectively, and obtaining output results aiming at different feature selection algorithms under the requirement of different feature subset numbers;
The prediction performance comparison under different feature selection numbers is sequentially carried out, when the current feature selection number is determined, corresponding features are sequentially selected through the sequencing values of different feature selection algorithms and are fused into an optimal feature pre-selection set;
Traversing the prediction performance under all different feature selection numbers, sequentially selecting different m features with the front importance value as the optimal feature subset of the data set, comparing the prediction performance to obtain the optimal feature screening number, carrying out element integration optimization on the optimal feature subset to obtain the optimal feature subset, and taking the optimal feature subset as the input of a subsequent prediction model, wherein the prediction model is a selected base model.
4. The method for prediction of building energy consumption based on a multisource fusion feature selection and fuzzy diversity enhancement Stacking framework of claim 3, wherein the different feature selection algorithms include variance selection, recursive feature elimination, XGBoost and linear feature selection.
5. The method for predicting building energy consumption based on multi-source fusion feature selection and fuzzy diversity enhanced Stacking framework of claim 1, wherein the selected base model comprises three of XGboost, RF and LightGBM.
6. The method for predicting building energy consumption based on multi-source fusion feature selection and fuzzy diversity enhancing Stacking framework according to claim 5, wherein the fuzzy diversity enhancing Stacking framework is of a two-dimensional structure and is provided with double input and single output.
7. The building energy consumption prediction method based on the multi-source fusion feature selection and fuzzy diversity enhancement Stacking framework of claim 1, wherein the constructed fuzzy diversity enhancement Stacking framework is utilized to select logistic regression to carry out regression prediction on the output result of the diversity enhanced base model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111568679.0A CN114386142B (en) | 2021-12-21 | 2021-12-21 | Building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111568679.0A CN114386142B (en) | 2021-12-21 | 2021-12-21 | Building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114386142A CN114386142A (en) | 2022-04-22 |
| CN114386142B true CN114386142B (en) | 2025-05-30 |
Family
ID=81197260
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111568679.0A Active CN114386142B (en) | 2021-12-21 | 2021-12-21 | Building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114386142B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114970720A (en) * | 2022-05-26 | 2022-08-30 | 河钢数字技术股份有限公司 | An application method of industrial anomaly prediction in steel energy public assistance |
| CN115081763B (en) * | 2022-08-24 | 2022-11-11 | 山东鲁晟精工机械有限公司 | Energy consumption prediction system for turning process of numerical control lathe |
| CN115544879B (en) * | 2022-09-30 | 2025-07-08 | 三峡大学 | Method for predicting ground composite electric field based on improved Stacking algorithm |
| CN115860173A (en) * | 2022-10-21 | 2023-03-28 | 国网电力科学研究院武汉能效测评有限公司 | Construction and prediction method and medium of carbon emission prediction model based on Stacking algorithm |
| CN118656907A (en) * | 2024-08-21 | 2024-09-17 | 国网（天津）综合能源服务有限公司 | Small sample prediction method of building heat load based on simulation data and meta-learning method |
| CN120278563B (en) * | 2025-06-11 | 2025-08-29 | 西华师范大学 | Building energy consumption prediction method, system, equipment and medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| AU2020101854A4 (en) * | 2020-08-17 | 2020-09-24 | China Communications Construction Co., Ltd. | A method for predicting concrete durability based on data mining and artificial intelligence algorithm |
| CN112947300A (en) * | 2021-02-22 | 2021-06-11 | 西安交通大学 | Virtual measuring method, system, medium and equipment for processing quality |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111199343B (en) * | 2019-12-24 | 2023-07-21 | 上海大学 | A multi-model fusion data mining method for abnormal data in tobacco market supervision |
-
2021
- 2021-12-21 CN CN202111568679.0A patent/CN114386142B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| AU2020101854A4 (en) * | 2020-08-17 | 2020-09-24 | China Communications Construction Co., Ltd. | A method for predicting concrete durability based on data mining and artificial intelligence algorithm |
| CN112947300A (en) * | 2021-02-22 | 2021-06-11 | 西安交通大学 | Virtual measuring method, system, medium and equipment for processing quality |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114386142A (en) | 2022-04-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN114386142B (en) | Building energy consumption prediction method based on multi-source fusion feature selection and fuzzy difference enhanced Stacking framework | |
| CN111738477B (en) | Prediction method of new energy consumption capacity of power grid based on deep feature combination | |
| US8706670B2 (en) | Relative variable selection system and selection method thereof | |
| CN112700124B (en) | Multi-layer traffic network MRWC node importance ordering method, system, electronic equipment and computer-readable storage medium | |
| CN111583031A (en) | Application scoring card model building method based on ensemble learning | |
| Tembusai et al. | K-nearest neighbor with k-fold cross validation and analytic hierarchy process on data classification | |
| Gong et al. | Ensemble models of TCN-LSTM-LightGBM based on ensemble learning methods for short-term electrical load forecasting | |
| CN110674965A (en) | Multi-time step wind power prediction method based on dynamic feature selection | |
| CN112765415A (en) | Link prediction method based on relational content joint embedding convolution neural network | |
| CN113139570A (en) | Dam safety monitoring data completion method based on optimal hybrid valuation | |
| CN119669298B (en) | Adaptive enhancement method and system for streaming query semantic graph based on cognitive computing | |
| Li et al. | Automatic fuzzy architecture design for defect detection via classifier-assisted multiobjective optimization approach | |
| CN114997377B (en) | Architecture searching method, device, equipment and storage medium for super-large scale graph | |
| CN109063759A (en) | A kind of neural network structure searching method applied to the more attribute forecasts of picture | |
| CN114386697A (en) | Ship main engine spare part prediction method based on improved random forest | |
| Li et al. | Hydrological time series prediction model based on attention-LSTM neural network | |
| Wei et al. | Boosting offline reinforcement learning with residual generative modeling | |
| CN116777056A (en) | Method and device for training prediction model and determining material demand | |
| CN109741597B (en) | Bus section operation time prediction method based on improved deep forest | |
| CN113032612B (en) | Construction method of multi-target image retrieval model, retrieval method and device | |
| CN115345303A (en) | Convolutional neural network weight tuning method, device, storage medium and electronic equipment | |
| CN118690985B (en) | A method for evaluating the quality of formulated products | |
| Guo et al. | New algorithms of feature selection and big data assignment for CBR system integrated by bayesian network | |
| CN118674092B (en) | Carbon emission prediction method considering main equipment differential metering characteristics | |
| CN116821407B (en) | Cross-modal remote sensing image retrieval method and device based on HGR maximum correlation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |