A data completeness verification and feature learning method for building load forecasting

Abstract

The present invention discloses a data completeness test and feature transfer learning method for building load prediction, comprising the following steps: proposing a feature screening method based on diffusion kernel density estimation and maximum correlation minimum redundancy, determining the optimal feature set under different load prediction models, establishing a method for data set feature completeness test by determining the correlation between features and feature distribution similarity between the new feature set and the optimal feature set; and establishing a load prediction model for feature transfer. The present invention can solve the problem of decreased building load prediction accuracy due to incomplete data features, and can still maintain the stability of the prediction results when there are slight changes in the model input, while reducing the computational complexity of the building load prediction model. The effects demonstrated by the present invention are to significantly improve the building load prediction accuracy, enhance the robustness of the load prediction model, and reduce the calculation time.

Description

Technical Field

[0001] The present invention belongs to the research field of building load forecasting and data mining, and specifically relates to a data completeness verification and feature transfer learning method for building load forecasting. Background Art

[0002] Air conditioning systems account for approximately 40% or even more of public building energy consumption, representing a significant potential for energy savings. To improve the energy efficiency of building air conditioning systems, various building energy-saving technologies and energy system optimization methods are emerging, including the coordinated optimization of renewable energy and building energy storage, building demand-side response, and distributed energy system optimization. However, the various control systems and optimization designs mentioned in these studies rely on accurate load forecasting, making efficient and accurate load forecasting crucial.

[0003] During the design and operation stages of a building, it is necessary to predict the building's load. From the perspective of the structure of the building cooling load prediction model, it is to build a mapping relationship between input features and output loads. The factors that play a decisive role in prediction accuracy include the quality of the data set itself, the data processing method, and the selection of the prediction algorithm. Since the data-driven model has certain requirements for the quality of the data, this limits the use of this method in actual projects to a certain extent. However, for some well-built building energy systems, their load-related data information is rich, but how to apply this part of the rich building energy information data to the information-deficient model and combine it with the corresponding prediction algorithm so that these data-deficient buildings can still achieve ideal prediction accuracy under small sample conditions is very meaningful in the process of promoting the application of load prediction in actual projects. Summary of the Invention

[0004] In view of this, the present invention provides a data completeness verification and feature learning method for building load forecasting. Under the condition of data scarcity, it is possible to mine transferable data information from the energy information of other buildings as a supplement to ensure that the load forecasting results for buildings with incomplete data information can still achieve load forecasting results that meet the prediction accuracy requirements.

[0005] To achieve the above objectives, the present invention proposes a data integrity verification and feature learning method for building load forecasting, comprising the following steps:

[0006] Step 1: First, a calculation method based on diffusion kernel density estimation (DKDE) is used to automatically generate the appropriate bandwidth. Furthermore, the mutual information value between discrete load variables is calculated:

[0007]

[0008] Where p(x) and p(y) are the probability density functions (PDFs) of x and y, respectively; x and y represent any variables in X and Y, respectively; and p(x, y) is the joint probability distribution of variables X and Y.

[0009] By introducing the maximum relevance minimum redundancy (mRMR) algorithm, the calculation method of mutual information between multiple variables is simplified to the method of constructing the optimal feature set:

[0010]

[0011] Where S n is a feature set consisting of n features, I(S n ; Y) represents the building cooling load Y and the characteristic set S n The mutual information value of Represents the information redundancy between each feature variable in the feature set; S\S h-1 Represents the entire feature set S and the selected feature set S h-1 The difference between ; h represents the number of steps of current feature selection, that is, the hth feature.

[0012] The incremental search process stops when the value of Equation (4) is non-positive, which indicates that the newly added features can only improve the redundancy of the feature set but cannot provide new information related to the load, thereby obtaining the optimal feature set.

[0013] Step 2: After data preprocessing, the dataset's feature completeness must be evaluated and determined by referencing the optimal feature set to determine whether the dataset can achieve the desired prediction accuracy. Based on the features in the new dataset, the load correlation between the features in the optimal feature set and each feature in the new dataset is calculated. Reaching a threshold indicates that the new dataset contains features that correspond to those in the optimal feature set. Observe whether each feature in the optimal feature set corresponds to each feature in the new dataset. If so, the new dataset is complete; otherwise, it is incomplete.

[0014] Symmetric uncertainty (SU) is used as the metric for the correlation between features, that is, IFC (f i , f j )=SU(f i , f j ), by calculating the feature f i and f j The mutual information of , and then normalized, can be obtained as SU. The calculation method is shown in formula (5):

[0015]

[0016] Among them, H(f i) represents the uncertainty of the feature (i.e., value), and its definition is shown in formula (6).

[0017] H(f i )=-∑p(f′ i )×log2p(f′ i ) (6)

[0018] In the formula, p(f′ i ) represents the feature f i The prior probability of taking a particular value.

[0019] IG(f i |f j ) represents the information gain rate, and its calculation method is shown in formula (7):

[0020] IG(f i |f j )=H(f i )-H(f i |f j )=H(f j )-H(f j |f i ) (7)

[0021] H(f i |f j ) represents the feature f j Under certain circumstances, the feature f i The value of is calculated as shown in formula (8):

[0022]

[0023] In the formula, p(f′ j ) represents f j The prior probability of taking a particular value.

[0024] The KS (Kolmogorov-Sminov) parameter is used to test the similarity of data distributions corresponding to the same feature in two different data sets. This method determines whether the feature distributions are similar by analyzing whether the cumulative frequency distributions of the feature in the two data sets are close. The steps of the KS test are as follows:

[0025] 1) Set two hypotheses: H0: S1(x) = S2(x), H1: S1(x) ≠ S2(x).

[0026] 2) Determine the cumulative probability distribution function. Let S1(x) and S2(x) represent the cumulative probability distribution function of a feature in the two data sets respectively.

[0027] 3) Calculate the test statistic. Test statistic D = max|S1(x)-S2(x)|.

[0028] Determine the distribution similarity. If the test statistic for each feature is small (the threshold is set to 0.05 in this paper), it indicates that the two data sets have a high degree of fit, and the hypothesis H0 is accepted. Otherwise, it is rejected.

[0029] Step 3: Based on the completeness judgment, if the result is that the dataset is complete, then the dataset can be used as the input of the model through feature engineering to train the prediction model; if the result is that the dataset is incomplete, it is necessary to migrate features from other complete datasets to fill in the missing features. The present invention uses the migration component analysis method to perform feature migration, and the steps of this method are as follows:

[0030] 1) Use K s，s , K t，t and K s，t Represent the kernel matrices of the source domain, target domain, and cross domain, define a composite kernel matrix, and convert it into the style of empirical kernel mapping, as shown in formula (2-47).

[0031]

[0032] 2) A (n1+n2)×m transformation matrix W is introduced to transform the corresponding eigenvector into the m-dimensional space, where m=n1+n2, which simplifies the problem of minimizing the difference between domains into an equivalent matrix trace maximization problem, as shown in Equation (10).

[0033]

[0034] Where μ represents the trade-off parameter;

[0035] And it is the identity matrix; 1

[0036] and is the covariance matrix, where is a column vector whose elements are all 1, is the identity matrix.

[0037] Solve the matrix (I+μKLK) -1 The eigenvector corresponding to the eigenvalue of KHK is the solution of formula (10), where at most n s +n t -1 eigenvector.

[0038] Beneficial effects

[0039] (1) The present invention can solve the problem of decreased building load prediction accuracy caused by incomplete data features, and greatly improve the prediction accuracy of building load.

[0040] (2) The present invention can maintain the stability of the prediction results when there are slight changes in the model input, thereby enhancing the robustness of the load forecasting model.

[0041] (3) The present invention reduces the computational complexity of the building load prediction model and shortens the computation time. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a technical roadmap for a data integrity verification and feature learning method for building load forecasting in the present invention;

[0043] Figure 2 is the kernel density estimation result of the outdoor temperature probability distribution in one embodiment of the present invention;

[0044] Figure 3 A distribution diagram of the mutual information between building cooling load and external disturbance parameters in one embodiment of the present invention;

[0045] Figure 4 The calculation results of feature correlation and feature distribution similarity of a new data set in one embodiment of the present invention;

[0046] Figure 5 The figure shows the R-Squared and CV_RMSE distribution box plots of the prediction results under each model in one embodiment of the present invention.

[0047] Figure 6 This is a robustness diagram of the prediction results of models A, B, and C under different samples in one embodiment of the present invention.

[0048] Figure 7 A comparison of CPU usage during the training process of models A, B, and C in one embodiment of the present invention DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0050] Take buildings A and B as an example. Building A is the source domain with complete data. The external interference data includes 4 types and 12 dimensions, as shown in Table 1. The internal interference data includes 8 types and 63 dimensions, as shown in Table 2.

[0051] Table 1 Summary of external interference parameters

[0052]

[0053] Table 2 Summary of internal interference parameters

[0054]

[0055] Building B is the target domain with incomplete data. The optimal feature set to be transferred from Building A, the complete feature set that achieves the best prediction accuracy, and the features included in the new dataset for Building B are shown in Table 3.

[0056] Table 3 Comparison of features included in the new dataset

[0057]

[0058] The present invention provides a data integrity test and feature learning method for building load forecasting, the flow chart of which is as follows: Figure 1 As shown, the following steps are included:

[0059] Step 1: First, for building A, a calculation method based on diffusion kernel density estimation (DKDE) is used to automatically generate a suitable bandwidth. Figure 2 As shown in the figure, taking outdoor temperature as an example, the kernel density estimation of the probability distribution of outdoor temperature is shown. Further, the mutual information value between discrete load variables is calculated: Figure 3 The figure shows the calculated mutual information distribution between the building cooling load and the external disturbance parameters. By introducing the maximum relevance minimum redundancy (mRMR) algorithm, the calculation method for the mutual information value between multiple variables is simplified to the construction method of the optimal feature set. Within each feature category, the feature with the largest mutual information with the building cooling load is selected to form the feature set as the starting point for the incremental search. The results of each incremental search are shown in Table 3.

[0060] Table 3 mRMR feature selection results

[0061]

[0062] From the feature selection results in Table 3, we can see that when the feature increment is updated to feature set FS11, according to the method described in step 1), it can be determined that when the value of formula (4) is non-positive, the newly added features can no longer provide new information. At this time, the feature increment update is stopped and the optimal feature set FS11 is obtained.

[0063] Step 2: Based on the complete feature set FS11, we compared the impact of missing feature variables on the accuracy of short-term building cooling load predictions to clarify the contribution of each feature variable to data completeness. CV_RMSE was used as an indicator of prediction accuracy. The results of the completeness contribution ranking are shown in Table 4.

[0064] Table 4 Ranking of completeness contribution rate of characteristic variables in short-term cooling load prediction model

[0065]

[0066] It can be seen from Table 4 that when conducting short-term prediction of building cooling load, the absence of each type of characteristic variable will result in a decrease in prediction accuracy, that is, an increase in CV_RMSE. The different rates of change in prediction accuracy indirectly reflect the contribution rate of each feature type to the completeness of the data features. The data feature completeness contribution rate of the characteristic variables is ranked from high to low as follows: building indoor temperature, solar radiation intensity, outdoor ambient temperature, building indoor lighting power, building indoor illumination, building indoor air humidity, building power consumption, outdoor dew point temperature, and heat pump system power consumption.

[0067] Using the evaluation method combining feature correlation and feature distribution similarity in step 2), select features with high correlation and distribution similarity with each feature in the new dataset from the best feature set under the loose standard. The selection results are as follows: Figure 4 shown.

[0068] Figure 4 In the figure, the x-axis represents the features in the new dataset, and the y-axis on the left represents the correlation between features. Symmetric uncertainty (SU) is used as a metric, with values ​​in the range [0, 1]. Values ​​closer to 1 indicate a higher correlation between the two features, corresponding to the bar chart in the figure. A threshold of 0.6 is selected; values ​​greater than 0.6 indicate correlation between the two features. The y-axis on the right represents the similarity of feature distributions. Based on the description in step 2), the KS test method will be used, with the p-value threshold set to 0.05. When the calculated value is greater than 0.05, the null hypothesis H0 (i.e., the two features have distributional similarity) is accepted; otherwise, the hypothesis is rejected. The value on the y-axis on the right is the p-value, corresponding to the scatter plot in the figure.

[0069] Step 3: Based on the completeness check from Step 2, the dataset for Building B is incomplete. Therefore, feature transfer from the complete dataset for Building A is required to complete the missing features. Transfer Component Analysis (TCA) is used to select the features to be transferred from the source domain feature set for Building A. This new dataset is then constructed along with the target domain features from Building B. This new dataset is then fed into the input layer of the LSTM model for training.

[0070] Specifically, this example compares and analyzes the building cooling load prediction results of the following four prediction models: Model A: Using source domain feature selection, the optimal source domain feature set is used as input for feature transfer; Model B: Using source domain feature selection, the optimal source domain feature set is used as input for feature transfer; Model C: Not using source domain feature selection, all source domain features are used as input for feature transfer; Model D: Not using feature transfer, the LSTM model is trained directly using target domain data. Models A, B, and C all use the transfer component analysis (TCA) method introduced in step 3) to perform feature transfer in the source domain.

[0071] Figure 5 The CV_RMSE and R-Squared distribution box plots of the prediction results under each model are shown. Models A, B, and C are all better than model D, indicating that after the introduction of feature migration, the prediction accuracy has been improved to varying degrees, and the CV_RMSE values ​​have all reached below 30%.

[0072] To further compare the robustness of models A, B, and C, the changes in prediction accuracy of the three models are compared in each sample, as shown in the following example: Figure 6 As shown in the figure, the horizontal axis corresponds to the target domain of gradually reducing the feature variables, the vertical axis is the R-Squared value, and (a) to (g) correspond to samples I to VI established in Table 5 respectively.

[0073] Table 5. Feature variables included in the target domain

[0074]

[0075] from Figure 6 It can be seen that model A can still ensure relatively high prediction accuracy when two feature variables are missing in the target domain, and as the target domain features decrease, the trend of prediction accuracy decreases the most slowly, showing its good robustness.

[0076] The complexity of the model can be directly reflected in the CPU usage. Further comparison of the differences in computational complexity among the three feature migration prediction models A, B, and C can be converted into a comparison of the CPU usage of the three feature migration prediction models during training, thereby comparing the computation time of the three models. Figure 7 shown.

[0077] Depend on Figure 7 From the comparison, it can be seen that model A can greatly reduce the CPU usage and training time during training while improving the prediction accuracy, which reflects the superiority of model A in reducing training time and improving model operation efficiency.

[0078] It should be understood that the embodiments and cases discussed here are for illustration only and may be improved or modified by those skilled in the art, and all such improvements and modifications should fall within the scope of protection of the claims appended to this invention.

Claims

1. A data integrity verification and feature transfer learning method for building load forecasting, comprising the following steps: Step 1): A calculation method based on diffusion kernel density estimation (DKDE) is used to automatically generate an appropriate bandwidth and calculate the mutual information value between discrete loads and characteristic variables. The maximum relevance minimum redundancy (mRMR) algorithm is introduced to transform the calculation method of mutual information between multiple variables into a method for constructing the optimal feature set. Where S n is a feature set consisting of n features, I(S n ; Y) represents the building cooling load Y and the characteristic set S n The mutual information value of Indicates the information redundancy between each feature variable in the feature set; S\S h-1 Represents the entire feature set S and the selected feature set S h-1 The difference of The incremental search process is The value is not stopped at a timely manner, which indicates that the newly added features can only improve the redundancy of the feature set, but cannot provide new information related to the load, so as to obtain feature variables with strong correlation with the load and thus obtain the optimal feature set; Step 2): Based on the optimal feature set obtained in step 1), for the newly obtained data set, after data preprocessing, it is first necessary to refer to the optimal feature set to evaluate and determine the feature completeness of the data set to determine whether the corresponding prediction accuracy can be achieved using the data set; based on the features in the new data set, the features contained in the optimal feature set are respectively correlated and distributed similarly with each feature in the new data set. If the threshold is reached, it indicates that the new data set contains features corresponding to the optimal feature set. Observe whether each feature in the optimal feature set corresponds to the features in the new data set. If so, it indicates that the features of the new data set are complete, otherwise it is incomplete. Step 3): Based on the completeness judgment in step 2), if the dataset is complete, the dataset can be used as the input of the model through feature engineering to train the prediction model; if the dataset is incomplete, it is necessary to migrate features from other complete datasets to fill in the missing features; Based on the transfer component analysis (TCA) method, the features that need to be migrated are selected from the constructed source domain feature set, and a new dataset is constructed together with the target domain features. The new dataset is then input into the input layer of the LSTM model for training. The effects of the migrated model on prediction accuracy, robustness and CPU usage are compared.

Images