Production enterprise energy consumption data prediction method based on machine learning

Abstract

The invention belongs to the technical field of industrial data processing, and particularly relates to a production enterprise energy consumption data prediction method based on machine learning. The method comprises the following steps: deploying an Internet of Things sensor to collect power and equipment operation data, and after three-stage preprocessing of anomaly detection, deletion repair and normalization, constructing a layered feature extraction network which comprises a time sequence feature layer, an equipment association feature layer and a modal fusion layer; the two features are respectively used for capturing power consumption features and equipment collaborative consumption features and fusing the features to generate depth feature vectors; and constructing a dynamic integrated prediction model, training three types of base models including a support vector machine and the like, screening a model entering an integrated pool by using JS divergence, and optimizing the weight through a genetic algorithm to obtain a final prediction result. The method can accurately capture the characteristics of energy consumption data, effectively improves the prediction precision, assists an enterprise in reasonably planning energy use, and reduces the cost.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial data processing, and particularly relates to a method for predicting energy consumption data of manufacturing enterprises based on machine learning. Background Art

[0002] In the operation of current manufacturing enterprises, energy costs account for a relatively large proportion. With the increasingly strict environmental protection requirements, it is crucial to reasonably plan and control energy consumption. However, current manufacturing enterprises face many challenges in energy consumption data management. On the one hand, traditional data collection methods are inefficient and inaccurate, making it difficult to obtain power consumption data and equipment operation data in real time, and unable to provide a reliable basis for energy consumption analysis. On the other hand, existing energy consumption prediction methods often lack accuracy, cannot effectively capture the periodic fluctuations and abnormal mutation patterns of power consumption, and are also difficult to consider the energy consumption correlation relationships between devices. This results in enterprises being unable to make reasonable arrangements for energy use in advance. Summary of the Invention

[0003] In view of the technical problems existing in the above background art, the present invention proposes a method for predicting energy consumption data of manufacturing enterprises based on machine learning.

[0004] To achieve the above object, the technical solution adopted by the present invention is as follows, including the following steps:

[0005] S1. Deploy Internet of Things sensors to collect power consumption data and equipment operation data in real time;

[0006] S2. Perform three-level preprocessing on the collected data, including anomaly detection, missing value repair, and normalization processing;

[0007] S3. Construct a hierarchical feature extraction network, which includes a time series feature layer, a device association feature layer, and a modality fusion layer;

[0008] Time series feature layer: Use an improved attention mechanism network to extract the time-dependent features of power consumption, and capture periodic fluctuations and abnormal mutation patterns;

[0009] Device association feature layer: Model the energy consumption association relationships between devices based on a graph structure, and extract the collaborative consumption features of device clusters;

[0010] Modality fusion layer: Fusion time and device features through a cross-modal interaction mechanism to generate a deep feature vector F;

[0011] S4. Construct a dynamic ensemble prediction model, and the specific implementation steps are as follows:

[0012] S41. First, sample L bootstrap subsets from F, which are respectively , each subset randomly contains 80% of the original samples;

[0013] S42. Train three types of prediction base models, namely support vector machine, random forest, and logistic regression, for each subset;

[0014] S43. For each type of base model, use the Jensen-Shannon (JS) divergence to measure the difference between the prediction distribution of the machine model and the true power consumption distribution. The calculation formula is , where represents the predicted value of the -th subset model in the i-th type of base model, is the true value, represents the Kullback-Leibler (KL) divergence;

[0015] S44. Set a threshold . For each type of base model, select the subset models of each type of base model with a JS divergence less than the threshold to enter the integration pool;

[0016] S5. Input new features and input them into the trained dynamic integration prediction model to generate prediction results.

[0017] Preferably, the device operation data in step S1 includes device real-time power, operation duration, and load rate.

[0018] Preferably, the specific implementation steps of anomaly detection in the three-level preprocessing in step S2 are as follows:

[0019] S21. Using the Isolation Forest, for each isolation tree, randomly select a subset from the standard space , recursively split until the leaf node. The splitting condition is to randomly select a feature dimension q and a splitting threshold at the current node, and divide the samples into and two parts, where is the multi-dimensional feature vector at the t-th moment;

[0020] S22. At each split, calculate the mean difference of the Mahalanobis distances of the samples in the left and right nodes , and select the feature and threshold that maximize the mean difference to enhance the separation ability of the anomaly points, where , is the mean vector of the samples in the current node, the inverse of the covariance matrix;

[0021] S23. For the sample , calculate the path length from the root node to the leaf node in each isolation tree, and obtain the Mahalanobis distance outlier score by integrating the path lengths of all trees. The calculation method is , where is the expected path length, is the path length correction coefficient when the sample size is n;

[0022] S24. Calculate the mean of the outlier scores of all samples and the standard deviation , set the threshold to , and remove the outlier scores greater than the threshold and the Mahalanobis distance greater than of the abnormal data, where is the overall standard deviation of the Mahalanobis distance.

[0023] Preferably, the specific implementation of the time series feature layer in step S3 includes the following steps:

[0024] S311. Concatenate the preprocessed time series data with the time position encoding to obtain an input vector with time series position information , and the calculation method of the time position encoding is: , where is the number of encoding terms, and d is the feature dimension;

[0025] S312. Construct a self-attention network with a dual gating mechanism, where the trend gate is , and the period gate is , where, is the weight matrix, are the hidden states at t-11 and t-1 respectively, is the bias term; then calculate the self-attention coefficient , where represents the hidden layer dimension;

[0026] S313. Aggregate the historical states by weighting: Output the time series feature matrix .

[0027] Preferably, the extraction method of the collaborative consumption feature of the device cluster in the device association feature layer in step S3 is: , where is the finally output feature representation, is the activation function, K is the number of attention heads, represents the set of neighborhood nodes of device i, represents the learnable weight matrix corresponding to the kth attention head, is the original feature vector of the neighborhood device j; then output the device association matrix , where M is the total number of devices.

[0028] Preferably, the specific operations of the modal fusion layer in step S3 include:

[0029] S321. First, construct an interaction matrix of time features and device features , where is the modal interaction matrix;

[0030] S322. Incorporate time-dependent information into device features to obtain and embed device association information into the time series to obtain , where is element-wise multiplication;

[0031] S323. Concatenate the enhanced features and retain the original information through residual information connection. The calculation method is: , where is the weight fusion matrix.

[0032] Preferably, before making predictions in S5, the dynamic ensemble prediction model needs to be trained. The specific implementation is to first input training samples and use the three types of base models in the ensemble pool to make predictions respectively. For each type of base model, there is more than one subset model. Therefore, based on dynamic weighting, all the base models in each type are weighted. The weight optimization objective function is: , where Q is the total number of training samples, is the initial weight of the th subset model, is the prediction value of the th subset model, is the true value of the jth training sample; the constraint condition is ; Solve the weight vector through the genetic algorithm and obtain the optimal solution through iteration.

[0033] Preferably, after training, the prediction values of each type of base model are obtained, and then the prediction values of the three machine models are weighted and averaged to obtain the final prediction value.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are as follows. In terms of data collection, with the help of Internet of Things sensors, power and equipment operation data are collected in real time, overcoming the problems of low efficiency and poor accuracy in the traditional method, and providing reliable data for energy consumption analysis. In data processing, based on the anomaly detection method of isolation forest and Mahalanobis distance, outliers are accurately identified, misjudgment is reduced, and combined with appropriate missing value repair and normalization means, the data quality is improved. When extracting features, the hierarchical feature extraction network can comprehensively capture the features of energy consumption data. The time series feature layer captures the fluctuation and mutation patterns, the device association feature layer reveals the energy consumption relationship between devices, and the modal fusion layer deeply fuses multi-faceted information. The prediction model adopts a dynamic ensemble method, fusing the advantages of multiple types of base models, and through sampling, screening and weight optimization, effectively improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0036] Figure 1 It is a schematic structural flowchart of a production enterprise energy consumption data prediction method based on machine learning; Figure 2 It is a schematic structural diagram of a hierarchical feature extraction network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0037] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described below in conjunction with the drawings and embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0038] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in the following specification.

[0039] Embodiment: A large enterprise has multiple complex production lines, with numerous equipment and complex operating modes in each link. Under the dual pressures of continuously rising energy costs and increasingly strict environmental protection, it is urgently necessary to accurately control the energy consumption situation to optimize energy utilization and reduce costs. First, IoT sensors are deployed at key positions in the production enterprise. For example, high-precision power sensors are installed in the power distribution room and on each production equipment to collect power consumption data in real time; corresponding monitoring devices are installed on the equipment to collect operation data such as real-time power, operation duration, and load rate of the equipment. These sensors transmit the collected data to the data center in real time through wireless transmission technology to ensure the timeliness of the data.

[0040] Considering the problem that the collected data may contain outliers, the existing processing method is to identify outliers through simple threshold judgment. This method often ignores the distribution characteristics and potential laws of the data, resulting in a high misjudgment rate. For example, when the power of a production equipment fluctuates due to a short-term commissioning, it is easily misjudged as abnormal data and excluded, affecting the integrity of the data. The present invention uses Isolation Forest. For each isolation tree, a subset is randomly selected from the standard space and recursively split until the leaf nodes. The splitting condition is to randomly select a feature dimension q and a splitting threshold at the current node, and divide the samples into ​ and two parts, where is the multi-dimensional feature vector at the t-th moment. At each split, calculate the mean difference of the Mahalanobis distances of the samples within the left and right nodes , select the feature and threshold that maximize the mean difference to enhance the separation ability of outliers, where , is the mean vector of the samples within the current node, the inverse of the covariance matrix. For the sample , calculate the path length from the root node to the leaf node in each isolation tree, and obtain the Mahalanobis distance outlier score by integrating the path lengths of all trees. The calculation method is , where is the expected path length, is the path length correction coefficient when the sample is n. Calculate the mean of the outlier scores of all samples and the standard deviation , set the threshold to , and remove the outlier data with the outlier score greater than the threshold and the Mahalanobis distance greater than , where is the overall standard deviation of the Mahalanobis distance. The present invention adopts an outlier detection method based on isolation forest and Mahalanobis distance, considering the differences between data points and the normal data distribution from multiple dimensions. It can not only more accurately identify true outliers, but also reduce the misjudgment of normal fluctuating data, effectively improving the data quality, providing a more reliable data basis for subsequent energy consumption analysis and prediction, and ensuring the accuracy and stability of the prediction results. Then, missing value repair and normalization processing are carried out. Specifically, to repair the missing values in the collected data, an interpolation method based on time series is adopted. According to the time series characteristics of the data, if the power consumption data at a certain moment is missing, first determine the data of the adjacent time points before and after it. For short-term missing values, linear interpolation is used to fill the missing values by calculating the linear relationship between the front and back data; if the missing time interval is long, polynomial interpolation is used to fit a curve that better conforms to the data trend for missing value filling and normalize the repaired data. The Min-Max normalization method is used for normalization processing.

[0041] Considering that existing energy consumption prediction methods are difficult to effectively capture the periodic fluctuations, abnormal mutation patterns of power consumption, and the energy consumption correlation relationships between devices. The present invention constructs a hierarchical feature extraction network, and the hierarchical extraction network includes a time series feature layer, a device association feature layer, and a modal fusion layer.

[0042] Specifically, the time series feature layer extracts the time-dependent features of power consumption using an improved attention mechanism network to capture periodic fluctuations and abnormal mutation patterns. Specifically, the preprocessed time series data is concatenated with the time position encoding to obtain an input vector with time sequence position information , and the calculation method of the time position encoding is as follows: , where is the number of encoded terms, and d is the feature dimension; construct a self-attention network with a dual gating mechanism, where the trend gate is , and the period gate is , where is the weight matrix, are the hidden states at t-11 and t-1 respectively, is the bias term; then calculate the self-attention coefficient , where represents the hidden layer dimension; aggregate the historical states through weighting: Output the time series feature matrix , where represents the weight matrix.

[0043] The device association feature layer models the energy consumption association relationship between devices based on the graph structure and extracts the collaborative consumption features of the device cluster. Specifically, the extraction method of the collaborative consumption features of the device cluster in the device association feature layer is as follows: , where is the final output feature representation, is the activation function, K is the number of attention heads, represents the set of neighborhood nodes of device i, represents the learnable weight matrix corresponding to the k-th attention head, The original feature vector of the neighboring device j; then output the device association matrix , where M is the total number of devices.

[0044] The modality fusion layer fuses the time and device features through a cross-modal interaction mechanism to generate a deep feature vector F. Specifically, first construct an interaction matrix of the time feature and the device feature , where is the modality interaction matrix; incorporate the time-dependent information into the device feature to obtain and embed the device association information into the time series to obtain , where is element-wise multiplication; concatenate the enhanced features and retain the original information through residual information connection, and the calculation method is: , where is the weight fusion matrix. Compared with the traditional method, the hierarchical feature extraction network can more comprehensively and accurately mine the characteristics of the energy consumption data of manufacturing enterprises. The traditional method often only analyzes the surface characteristics of the data simply and is difficult to capture the deep complex information. The time series feature layer of this network can accurately capture the periodic fluctuations and abnormal mutations of power consumption; the device association feature layer can effectively reveal the energy consumption association relationship between devices; the modal fusion layer deeply fuses the time and device features, and the generated deep feature vector contains key information in many aspects. Through this series of operations, the hierarchical feature extraction network provides richer and more valuable inputs for the subsequent prediction model.

[0045] Finally, considering that the traditional single model cannot comprehensively capture the complex characteristics of energy consumption data and has poor adaptability to different production conditions and external factor changes. For example, before and after the production rhythm adjustment and equipment maintenance, the prediction error of the single model will increase significantly. At the same time, the model is prone to overfitting or underfitting, and the stability is poor. The dynamic integrated prediction model of the present invention integrates multiple different types of base models, generates multiple subsets by sampling to train different models, screens the models with excellent performance into the integration pool, and optimizes the weights by combining the genetic algorithm. The implementation process is as follows: First, input the training samples, and use the three types of base models in the integration pool to make predictions respectively. For each type of base model, there is more than one subset model. Therefore, based on dynamic weighting, all the base models in each type are weighted. The weight optimization objective function is: , where Q is the total number of training samples, is the initial weight of the th subset model, is the th subset model's predicted value, is the true value of the jth training sample; the constraint condition is ; The weight vector is solved by a genetic algorithm, and the optimal solution is obtained through iteration. After training, the predicted values of each type of base model are obtained, and then the predicted values of the three machine models are weighted and averaged to obtain the final predicted value. Specifically, first, from the deep feature vector F generated by the hierarchical feature extraction network, L subsets containing 80% of the original samples are generated using the bootstrap sampling method. During the sampling process, each sample has the same probability of being drawn, so that the subsets have diversity. Then, for each subset, three types of prediction base models, namely support vector machine, random forest, and logistic regression, are trained respectively. Different base models have different learning characteristics. The support vector machine is good at dealing with small sample and non-linear problems; the random forest can effectively process high-dimensional data and has strong anti-noise ability; the logistic regression is more efficient in modeling linear relationships. By training these three types of models, the characteristics and laws of energy consumption data can be learned from multiple perspectives. Then, the JS divergence is used to measure the difference between the prediction distribution of each type of base model and the true electricity consumption distribution. A reasonable threshold is set, and the subset models of each type of base model with a JS divergence less than the threshold are selected into the integration pool. This can screen out models with better prediction performance and ensure the overall quality of the integrated model. Input the training samples and use the three types of base models in the integration pool to make predictions respectively. Since there is more than one subset model for each type of base model, the weight optimization objective function is used, and under the constraint conditions, the weight vector is solved by a genetic algorithm. The genetic algorithm simulates the natural evolution process, and through multiple rounds of selection, crossover, and mutation operations, the optimal solution is obtained through iteration to determine the weights of each subset model in each type of base model. After training is completed, the predicted values of the three types of base models are weighted and averaged to obtain the final predicted value of the energy consumption data. Through this dynamic integrated prediction model, the advantages of different base models are fully utilized, and the prediction accuracy of complex energy consumption data is improved.

[0046] The above is only a preferred embodiment of the present invention, and it is not intended to limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting energy consumption data of production enterprises based on machine learning, characterized in that: The following steps are involved: S1. Deploy IoT sensors to collect power consumption data and equipment operation data in real time; S2, perform three-level preprocessing on the collected data, including anomaly detection, missing repair and normalization; S3. Constructing a hierarchical feature extraction network, wherein the hierarchical feature extraction network includes a time series feature layer, a device association feature layer, and a modality fusion layer; Time series feature layer: an improved attention mechanism network is used to extract the time-dependent features of power consumption and capture periodic fluctuations and abnormal mutation patterns; Equipment association feature layer: Based on the graph structure, the energy consumption association relationship between devices is modeled and the collaborative consumption characteristics of the equipment cluster are extracted; Modal fusion layer: fuses time and device features through cross-modal interaction mechanism to generate a deep feature vector F; S4. Build a dynamic integrated prediction model. The specific implementation steps are as follows: S41. First, generate L bootstrap subsets by sampling from F, which are , each subset randomly contains 80% of the original samples; S42, training three types of prediction base models, support vector machine, random forest and logistic regression, for each subset; S43. For each type of base model, the difference between the predicted distribution of the JS divergence metric model and the actual power consumption distribution is calculated using the formula: ,in Represents the first The predicted values ​​of the subset models are is the true value, represents KL divergence; S44. Setting threshold , for each type of base model, select the subset model of each type of base model whose JS divergence is less than the threshold to enter the integrated pool; S5. Input new features and input them into the trained dynamic integrated prediction model to generate prediction results.

2. The method for predicting energy consumption data of production enterprises based on machine learning according to claim 1 is characterized in that: The equipment operation data in step S1 includes the real-time power, operation time and load rate of the equipment.

3. The method for predicting energy consumption data of production enterprises based on machine learning according to claim 1 is characterized in that: The specific implementation steps of abnormality detection in the three-level preprocessing in step S2 are: S21. Using the isolation forest, for each isolated tree, from the standard space Randomly select a subset , recursively split until the leaf node, the splitting condition is to randomly select the feature dimension q and splitting threshold at the current node , divide the samples into and Two parts, including is the multidimensional feature vector at the tth moment; S22. At each split, calculate the mean difference of the Mahalanobis distances between the samples in the left and right nodes. , select the features and thresholds that maximize the mean difference and enhance the ability to separate outliers, where , is the mean vector of samples in the current node, Inverse of the covariance matrix; S23. For samples , in each isolated tree, the path length from the root node to the leaf node is calculated, and the Mahalanobis distance outlier score is obtained by integrating the path lengths of all trees, which is calculated as ,in is the expected path length, is the path length correction coefficient when the number of samples is n; S24. Calculate the mean outlier score of all samples and standard deviation , set the threshold to , remove outlier scores Greater than the threshold and the Mahalanobis distance Greater than The abnormal data, is the population standard deviation of the Mahalanobis distance.

4. The method for predicting energy consumption data of a production enterprise based on machine learning according to claim 1 is characterized in that: The specific implementation of the time series feature layer in step S3 includes the following steps: S311, concatenate the preprocessed time series data with the time position code to obtain an input vector with time series position information , the time position coding is calculated as follows: ,in is the number of items encoded, d is the feature dimension; S312, construct a self-attention network with a dual gating mechanism, where the trend gate is , the periodic gate is ,in, is the weight matrix, are the hidden states at time t-11 and t-1 respectively, is the bias term; then calculate the self-attention coefficient ,in represents the hidden layer dimension; S313, by weighted aggregation of historical states: Output time series feature matrix .

5. The method for predicting energy consumption data of production enterprises based on machine learning according to claim 1 is characterized in that: The method for extracting the collaborative consumption features of the device cluster in the device association feature layer in step S3 is: ,in is the feature representation of the final output, is the activation function, K is the number of attention heads, represents the set of neighboring nodes of device i, represents the learnable weight matrix corresponding to the kth attention head, The original feature vector of the neighborhood device j; then the device association matrix is ​​output , where M is the total number of devices.

6. The method for predicting energy consumption data of a production enterprise based on machine learning according to claim 1, characterized in that: The specific operations of the modality fusion layer in step S3 include: S321. First, construct the interaction matrix of time features and device features ,in is the modal interaction matrix; S322, integrating the time-dependent information into the device characteristics to obtain And embed the device association information into the time series to obtain ,in is element-wise multiplication; S323, concatenate the enhanced features and retain the original information through residual information connection, the calculation method is: ,in is the weight fusion matrix.

7. The method for predicting energy consumption data of a production enterprise based on machine learning according to claim 1, characterized in that: Before making predictions in S5, the dynamic integrated prediction model needs to be trained. Specifically, the training samples are first input, and predictions are made using the three types of base models in the integrated pool. There is more than one subset model for each type of base model, so all base models in each type are weighted based on dynamic weighting. The weight optimization objective function is: , where Q is the total number of training samples, For the The initial weights of the subset models, For the The predicted values ​​of the subset models are is the true value of the jth training sample; the constraints are ; Solve the weight vector through genetic algorithm and get the optimal solution through iteration.

8. The method for predicting energy consumption data of production enterprises based on machine learning according to claim 7 is characterized in that: After training, the prediction value of each type of base model is obtained, and then the prediction values ​​of the three machine models are weighted averaged to obtain the final prediction value.

Images