Non-equilibrium industrial load identification method

Abstract

The application discloses a non-balanced industrial load identification method, and is specifically implemented according to the following steps: industrial load data and corresponding switch state original samples are selected from a TMLD data set, a plurality of original samples are divided into a plurality of intervals, 27 time domain features containing 8 entropy features are extracted for each interval to construct an original feature set; the importance of features in the original feature set is calculated based on a predicted value change amount mode and is sorted, a CatBoost classification accuracy is taken as a decision variable, and the optimal classification feature subset of the original sample corresponding to the equipment is determined through forward feature selection; a Borderline-SMOTE method is used to synthesize sample data for balancing treatment to obtain balanced switch sample data; and a CatBoost classifier with a Bayesian superband hyperparameter optimization is constructed to identify the industrial load; and the non-balanced industrial load identification method can improve the identification precision of the non-balanced industrial load.

Description

Technical Field

[0001] This invention belongs to the field of load identification technology, specifically relating to a method for identifying unbalanced industrial loads. Background Technology

[0002] With the gradual development of smart electricity use and demand-side management technologies, and the increasing awareness of energy conservation and emission reduction, research on load identification technology is essential. Compared to residential electricity load, industrial electricity load accounts for a larger proportion, and industry is a major source of carbon emissions. To address the issues of rising temperatures, energy supply shortages, and to promote demand-side management and manage the electricity load of industrial enterprises, industrial load identification is a crucial solution. Industrial load identification enables industrial enterprises to understand the specific energy consumption of each piece of equipment at different times during production. This helps enterprises flexibly adjust production schedules and increases their willingness to actively participate in the power grid company's low-carbon demand response. Simultaneously, it enhances the energy-saving awareness of industrial enterprises, effectively mobilizing user potential and encouraging them to consciously optimize their electricity consumption behavior.

[0003] Existing load identification processes typically involve two key steps: feature extraction and load recognition. While time-frequency feature extraction methods such as S-transform can achieve good results, they suffer from high computational cost, difficulty in parameter selection, loss of high-frequency information, and high time and space complexity, which affect the classification performance of features extracted after signal processing. Furthermore, excessively high feature vector dimensionality increases classifier complexity and reduces classification efficiency. Secondly, industrial load data suffers from imbalance due to unequal accumulation of positive and negative samples, affecting the classification accuracy of minority classes. Finally, CatBoost's classification performance is closely related to its hyperparameter values; the hyperparameter search space is large, making manual parameter tuning, grid search, and random search methods inefficient. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying unbalanced industrial loads, which can improve the identification accuracy of unbalanced industrial loads.

[0005] The technical solution adopted in this invention is an unbalanced industrial load identification method, which is implemented according to the following steps:

[0006] Step 1: Industrial load data and its corresponding switch status raw samples. Select multiple raw samples from the TMLD dataset, divide each raw sample into multiple intervals, and extract 27 time-domain features containing 8 entropy features from each interval to construct the raw feature set.

[0007] Step 2: Calculate and sort the feature importance in the original feature set based on the change in predicted value. Use CatBoost classification accuracy as the decision variable and determine the optimal subset of classification features for the device corresponding to the original sample through forward feature selection.

[0008] Step 3: Use the Borderline-SMOTE method to synthesize sample data and perform balancing processing to obtain balanced switch sample data;

[0009] Step 4: Construct a Bayesian superband hyperparameter optimized CatBoost classifier to identify industrial loads.

[0010] The invention is further characterized by:

[0011] The specific process of dividing each original sample into multiple intervals in step 1 is as follows: Each original sample contains 40 sampling points, and each sample is divided into three intervals: the first 20 sampling points, the last 20 sampling points, and the total 40 sampling points.

[0012] In step 1, the sampling frequency of each original sample point is 1 / 30Hz, and the sampling duration is 1200s.

[0013] Step 2 is as follows:

[0014] A decision tree model is established, where each leaf node represents a subset of features from the original feature set. The importance of each subset feature is calculated by the change in predicted value, and the importance is sorted from largest to smallest. CatBoost classification accuracy is used as the decision variable, and forward feature selection is used to determine the optimal subset of classification features for the device corresponding to the original sample.

[0015] Step 3 is as follows:

[0016] Step 3.1: Assume that the optimal subset S of classification features is the sample set, S min It is a minority class sample set, S maxj It is the set of nearest majority class samples, where m is the number of nearest samples, and x is the nearest neighbor. i It represents all the attributes of the sample, x ij It represents all the attributes of neighboring samples, x n It is a neighboring sample, R ij The value can be either 0.5 or 1, and each sample x i ∈S min Determine the set of samples that is closest to it, and let the dataset be S. NN And S NN ∈S;

[0017] Step 3.2: For each sample x i Determine the number of nearest neighbor samples that belong to the majority class sample set, i.e., |S NN ∩S maxj |;

[0018] Step 3.3: Select the option that satisfies x i : The standard number of neighboring samples, x i With neighboring sample x n The difference in attribute j is denoted as d. ij =x i -x ij ; A newly synthesized sample h was obtained ij =x i +d ij ×rand(0,R) ij This refers to the sample data of the balance switch.

[0019] Step 4 specifically involves: using the Bayesian overband optimization algorithm to optimize the hyperparameter function x∈χ of the CatBoost classifier. With the goal of globally minimizing or maximizing the hyperparameter optimization function, the objective function is constructed as follows:

[0020]

[0021] h represents the balanced sample data as input, and x represents the hyperparameter.

[0022] f(h;x) can be replaced by the function y(h;x), expressed as:

[0023] y(h;x)=f(h;x)+∈

[0024] in

[0025] The sample data of the balancing switch is input into the objective function to obtain the CatBoost classifier with optimized hyperparameters. The industrial load data to be identified is input into the CatBoost classifier with optimized hyperparameters to obtain the switch state.

[0026] The beneficial effects of this invention are:

[0027] 1) After performing time-domain segmentation on the original samples, the temporal features containing entropy are directly extracted from the three time-domain intervals, which effectively shortens the feature extraction time and improves the efficiency of feature extraction;

[0028] 2) Calculate and rank the feature importance based on the change in predicted values, and use CatBoost classification accuracy as the decision variable. Determine the optimal subset of classification features for the five devices through forward feature selection, further reducing feature extraction time, classifier training, and recognition time.

[0029] 3) The Borderline-SMOTE method was used to synthesize sample data for balancing, which improved the model's classification performance on unbalanced data and thus improved the identification accuracy of unbalanced industrial loads.

[0030] 4) A CatBoost classifier for Bayesian superband BOHB hyperparameter optimization was constructed to identify industrial loads, thereby improving the efficiency of hyperparameter optimization. Attached Figure Description

[0031] Figure 1 This is a flowchart of the unbalanced industrial load identification method of the present invention;

[0032] Figure 2 This is a diagram showing the individual sample division of the five devices of this invention;

[0033] Figure 3 This is a schematic diagram illustrating the accuracy of different dimensional feature combinations in the forward feature selection process of this invention.

[0034] Figure 4 This is a flowchart illustrating the specific construction process of the BOHB algorithm for optimizing the CatBoost model in this invention.

[0035] Figure 5 This is a schematic diagram illustrating the experimental results of different hyperparameter optimization algorithms in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0037] The present invention provides a method for identifying unbalanced industrial loads, such as... Figure 1 As shown, please follow these steps:

[0038] Step 1: Industrial load data and its corresponding switch status raw samples. Select multiple raw samples from the TMLD dataset. Each raw sample contains 40 sampling points. The sampling frequency of the sampling points in each raw sample is 1 / 30Hz and the sampling duration is 1200s. Divide each sample into three intervals: the first 20 sampling points, the last 20 sampling points, and the total 40 sampling points. Extract 27 time-domain features containing 8 entropy features from each interval to construct the raw feature set.

[0039] Step 2: Based on the Predicted Value Change (PVC), we calculate the observed change in classification result when the value corresponding to the feature changes. The feature set assigned to the machine learning model is represented as:

[0040] F = {f1, f2, f3, ..., f} n}

[0041] F represents a set of input features, β is a numerical factor assigned to the input features, and P is the prediction step.

[0042] P i =β i F j

[0043] P i β represents the predicted value of the substitution factor. i F represents the numerical factor. j This represents a specific feature selected from a given feature set.

[0044] P i+1 =β i+1 F j

[0045] P i+1 β represents the predicted value when the numerical factor changes. i+1 This indicates the modified numerical factor.

[0046] P i＝0 ≠P i ≠P i+1

[0047] This indicates that this specific feature is necessary when the numerical factor changes and alters the predicted value.

[0048] The importance of features in the original feature set is calculated and ranked based on the Predicted Value Change (PVC) method. CatBoost classification accuracy is used as the decision variable, and forward feature selection is employed to determine the optimal subset of classification features for the device corresponding to the original samples. The specific process is as follows:

[0049] A decision tree model is established, where each leaf node represents a subset of features from the original feature set. The importance of each subset feature is calculated by the change in predicted value, and the importance is sorted from largest to smallest. CatBoost classification accuracy is used as the decision variable, and forward feature selection is used to determine the optimal subset of classification features for the device corresponding to the original sample.

[0050] Step 3: The sample data is synthesized using the Borderline-SMOTE method and then balanced to obtain the balanced switch sample data; the specific process is as follows:

[0051] Step 3.1: Assume that the optimal subset S of classification features is the sample set, S min It is a minority class sample set, S maxj It is the set of nearest majority class samples, where m is the number of nearest samples, and x is the nearest neighbor. i It represents all the attributes of the sample, x ij It represents all the attributes of neighboring samples, x n It is a neighboring sample, R ij The value can be either 0.5 or 1, and each sample x i ∈S min Determine the set of samples that is closest to it, and let the dataset be S. NN And S NN ∈S;

[0052] Step 3.2: For each sample x i Determine the number of nearest neighbor samples that belong to the majority class sample set, i.e., |S NN ∩S maxj |;

[0053] Step 3.3: Select the option that satisfies x i : The standard number of neighboring samples, x i With neighboring sample x n The difference in attribute j is denoted as d. ij =x i -x ij ; A newly synthesized sample h was obtained ij =x i +d ij ×rand(0,R) ij This refers to the sample data of the balance switch.

[0054] Step 4: Construct a Bayesian overband hyperparameter-optimized CatBoost classifier to identify industrial loads. Specifically, the Bayesian overband optimization algorithm is used to optimize the hyperparameter function x∈χ of the CatBoost classifier. With the goal of globally minimizing or maximizing the hyperparameter optimization function, the objective function is constructed as follows:

[0055]

[0056] h represents the balanced sample data as input, and x represents the hyperparameter.

[0057] f(h;x) can be replaced by the function y(h;x), expressed as:

[0058] y(h;x)=f(h;x)+∈

[0059] in

[0060] The sample data of the balancing switch is input into the objective function to obtain the CatBoost classifier with optimized hyperparameters. The industrial load data to be identified is input into the CatBoost classifier with optimized hyperparameters to obtain the switch state.

[0061] In this invention, the BOHB hyperparameter optimization algorithm combines the Bayesian optimization algorithm (BO) and the superband (HB) algorithm.

[0062] BO uses various probabilistic surrogate models based on the calculated data sample D = {(x0,y0),(x1,y1),...,(x n-1 ,y n-1Define the prediction distribution p(f|D). This paper uses the expected improvement (EI) criterion as the acquisition function, expressed as:

[0063]

[0064] Where y * =min{y0,y1,...,y n}. BO employs the Tree Parzen Estimator (TPE) method. Instead of directly modeling the function f using p(f|D), TPE models the density of the input configuration using a kernel density estimator, expressed as:

[0065] l(x) = p(y < y * |x,D)

[0066] g(x) = p(y ≥ y) * |x,D)

[0067] In order to select new optimization candidates, the maximum The ratio is equivalent to optimizing the EI function.

[0068] BOHB employs a tree-structured Parzen estimator (TPE) Bayesian optimization method using a single multidimensional kernel density estimator (KDE). In BOHB, a minimum number of data points N are required to obtain a useful KDE. min Set the evaluation for d+1 executions, where d is the number of hyperparameters. The algorithm is initialized to N. min +2 configurations until D is satisfied B >N min +2 condition. The model densities for the worst and best configurations are expressed as:

[0069] N B,l =max(N) min ,q·N B )

[0070] N B,g =max(N) min N B -N B ,l)

[0071] Where, N B q is the number of data points in budget B, and q is N. B Percentiles. The EI of the sample S in the optimized BOHB algorithm is obtained from l(x). The bandwidth of the density in l(x) is multiplied by a factor b to find and explore more promising configurations.

[0072] Example 1

[0073] 1. Selection of dataset

[0074] The data used in this invention is the Industrial Load Dataset TMLD. The dataset contains the total load sequence and equipment status of a textile mill. The sampling frequency is 1 / 30Hz, and the sampling duration is 28 days. The total load sequence includes four different machine tools and one sewing machine.

[0075] 2. A time-domain segmentation method is used to efficiently extract time-domain features containing entropy.

[0076] In the feature extraction process, after temporal domain segmentation of the original samples, temporal features are directly extracted. First, each sample is divided into three intervals: 20 sampling points before and after the sample, and a total of 40 sampling points. Then, 27 temporal features, including 8 types of entropy features, are extracted from each interval. Finally, an 81-dimensional original feature set is constructed. The individual sample segmentation diagram for the five devices is shown below. Figure 2 As shown.

[0077] 3. Selection of Industrial Load Characteristics Based on the Importance of PVC

[0078] Feature importance is calculated and ranked based on PVC. CatBoost classification accuracy is used as the decision variable, and forward feature selection is employed to determine the optimal subset of classification features for five devices. The accuracy of different feature combinations in forward feature selection is shown below. Figure 3 As shown.

[0079] Feature importance (FI) is calculated using the PVC method, and features are sorted in descending order. Forward feature selection is performed, and each feature is added to a feature subset sequentially. For each added feature, the recognition accuracy of the CatBoost classifier within that subset is calculated. This process is repeated until all features are added to the feature set. Then, the optimal feature subset is determined based on the highest recognition accuracy of the classifier. During the classification process, the classifier's recognition accuracy continuously changes. Devices 1 through 5 achieved the highest classification accuracy when the feature dimensions were 10, 23, 23, 30, and 22, respectively.

[0080] 4. Construct a Bayesian superband hyperparameter-optimized CatBoost classifier to identify industrial loads. The specific construction process is as follows: Figure 4 As shown, first, the optimal feature subset after balancing is input, parameter settings and initialization are optimized, and a search space is defined for use. Then, the BOHB algorithm is used to optimize and train the CatBoost parameters. Finally, the optimal parameter values ​​and classification accuracy are obtained while satisfying the maximum number of iterations. The maximum number of iterations is set to 50. The final output is the BOHB-CatBoost model.

[0081] 5. Evaluation Indicators

[0082] To more comprehensively evaluate the classification performance of the CatBoost model, this paper selects accuracy, precision, recall, and F1 score as evaluation metrics. Accuracy is the proportion of correctly predicted samples out of all samples; precision is the proportion of samples predicted as "open" that were actually "open"; recall is the proportion of samples that were actually "open" that were predicted as "open"; and the F1 score is the harmonic mean of accuracy and recall. The specific calculation formulas are shown below:

[0083]

[0084]

[0085]

[0086]

[0087] Among them: A cc For accuracy; P re For accuracy; R re For recall rate; F score This is the F-score, also known as the balanced F-score. T p This represents the number of samples where a certain type of device is actually in the "on" state, and the model also predicts it to be "on". T N This represents the number of samples where a certain type of device is actually off, and the model also predicts it to be off. F p This represents the number of samples where a certain type of device is actually off but predicted to be on. F N This represents the number of samples where a certain type of device is actually on but predicted to be off.

[0088] 6. Analysis of Experimental Results

[0089] To demonstrate the beneficial effects of the model designed in this invention, the specific experimental results are as follows:

[0090] 6.1 Efficiency Comparison of Temporal Segmentation-Based Feature Extraction Methods with Traditional Methods

[0091] The efficiency of the temporal segmentation-based feature extraction method is compared with that of S-transform (ST), wavelet transform (WT), and empirical mode decomposition (EMD). The feature extraction time required for the original sample under different feature extraction methods is shown in Table 1.

[0092] Table 1

[0093]

[0094] Experiments show that the feature extraction time of this invention is reduced by 36% compared to the ST method, 61.4% compared to the WT method, and 37.7% compared to the EMD method. Therefore, the feature extraction method based on temporal segmentation is more efficient than the ST, WT, and EMD methods.

[0095] 6.2 Comparison of Experimental Results of Different Imbalanced Data Processing Methods

[0096] By comparing the present invention with those without a synthesis algorithm and the SMOTE algorithm, the effectiveness of the present invention in improving the identification accuracy of unbalanced industrial loads in unbalanced scenarios is verified. The number of switching samples for the five devices in the experiment is shown in Table 2. The experimental results of different unbalanced data processing methods are shown in Table 3.

[0097] Table 2

[0098]

[0099] Table 3

[0100]

[0101] Experiments show that the metrics of this invention are higher than those of the original data and the SMOTE method. The average values ​​of the four metrics for devices 1 to 5 were calculated. The accuracy, precision, recall, and F1 score of this method are improved by 0.73%, 0.91%, 0.54%, and 0.73% respectively compared to SMOTE. Compared to the original data, they are improved by 0.92%, 1.12%, 0.71%, and 0.92% respectively. This effectively verifies the effectiveness of using the synthesized sample data of this invention for balancing, thereby improving the model's classification performance on imbalanced datasets.

[0102] 6.3 Comparison of Different Hyperparameter Optimization Algorithms

[0103] To verify the superiority of the BOHB optimization algorithm in determining hyperparameters, it was compared with the default hyperparameters, random search algorithm, grid search algorithm, and Bayesian optimization algorithm. The metrics in the experiments are the average values ​​of five devices under different hyperparameter optimization algorithms. Experimental results for different hyperparameter optimization methods are shown below. Figure 5 As shown.

[0104] Experiments show that the BOHB hyperparameter optimization method significantly outperforms the other five hyperparameter optimization methods in all metrics. Compared to the default parameters, the BOHB optimization algorithm improves accuracy by 1.18%, precision by 0.76%, recall by 1.57%, and F1 score by 1.18%. This method effectively determines the optimal values ​​of the CatBoost algorithm's hyperparameters, significantly improving the accuracy of industrial load identification.

Claims

1. A method for identifying unbalanced industrial loads, characterized in that, The specific steps are as follows: Step 1: Industrial load data and its corresponding switch status raw samples. Select multiple raw samples from the TMLD dataset, divide each raw sample into multiple intervals, and extract 27 time-domain features containing 8 entropy features from each interval to construct the raw feature set. Step 2: Calculate and sort the feature importance in the original feature set based on the change in predicted value. Use CatBoost classification accuracy as the decision variable and determine the optimal subset of classification features for the device corresponding to the original sample through forward feature selection. Step 3: Use the Borderline-SMOTE method to synthesize sample data and perform balancing processing to obtain balanced switch sample data; Step 4: Construct a Bayesian superband hyperparameter optimized CatBoost classifier to identify industrial loads; Step 4 involves optimizing the hyperparameters of the CatBoost classifier using the Bayesian overband optimization algorithm. function With the goal of globally minimizing or maximizing the hyperparameter optimization function, the objective function is constructed as follows: h This represents the balanced sample data that was used as input. x Indicates hyperparameters; Use function To replace, represented as: in ; Input the sample data of the balancing switch into the objective function to obtain the CatBoost classifier with optimized hyperparameters, and input the industrial load data to be identified into the CatBoost classifier with optimized hyperparameters to obtain the switch state; Bayesian optimization algorithm (BO) uses various probabilistic surrogate models based on the calculated data samples. Define the prediction distribution This paper uses the Expected Improvement (EI) criterion as the acquisition function, expressed as: in BO employs the Tree Parzen Estimator (TPE) method, where TPE is not directly derived from... For functions Instead of performing modeling, a kernel density estimator is used to model the density of the input configuration, expressed as: 。 2. The method for identifying unbalanced industrial loads according to claim 1, characterized in that, The specific process of dividing each original sample into multiple intervals in step 1 is as follows: each original sample contains 40 sampling points, and each sample is divided into three intervals: the first 20 sampling points, the last 20 sampling points, and the total 40 sampling points.

3. The method for identifying unbalanced industrial loads according to claim 1, characterized in that, In step 1, the sampling frequency of each sampling point in the original sample is 1 / 30Hz, and the sampling duration is 1200s.

4. The method for identifying unbalanced industrial loads according to claim 1, characterized in that, Step 2 is as follows: A decision tree model is established, where each leaf node represents a subset of features from the original feature set. The importance of each subset feature is calculated by the change in predicted value, and the importance is sorted from largest to smallest. CatBoost classification accuracy is used as the decision variable, and forward feature selection is used to determine the optimal subset of classification features for the device corresponding to the original sample.

5. The method for identifying unbalanced industrial loads according to claim 1, characterized in that, Step 3 is as follows: Step 3.1: Assume the optimal subset of classification features It is a sample set. It is a minority class sample set. It is the nearest majority class sample set. It is the number of neighboring samples. It represents all the attributes of the sample. It is all the attributes of the neighboring samples. It is a neighboring sample. The value can be either 0.5 or 1, for each sample. Determine the set of samples that is closest to it, and let the dataset be... ,and ; Step 3.2: For each sample Determine the number of nearest neighbor samples that belong to the majority class sample set, i.e. ; Step 3.3, Select the one that satisfies the requirements. The standard number of neighboring samples will with neighboring samples Corresponding attributes The difference in is denoted as ; Newly synthesized samples were obtained This refers to the sample data of the balance switch.

Images