Abnormal Energy Consumption Monitoring Method and System for Base Stations Based on Wavelet Decomposition and Migration Discrimination

Abstract

The present invention proposes a method and system for abnormal energy consumption monitoring of base stations based on wavelet decomposition and transfer discrimination, including: performing clustering and classification of base station energy consumption samples based on multi-dimensional high and low frequency features obtained from an energy consumption feature set; constructing an abnormal energy consumption discrimination model with the high and low frequency features, archive feature data, meteorological feature data, and holiday feature data of a certain clustering energy consumption historical sample data set as inputs; performing nearest neighbor class model parameter tuning transfer learning on the constructed abnormal energy consumption discrimination model, adaptively adjusting the model parameters through reinforcement learning, and outputting a nearest neighbor clustering abnormal energy consumption discrimination model. Sequentially perform nearest neighbor transfer learning and parameter tuning until the construction of abnormal energy consumption discrimination models for all clustering categories is completed, forming a set of energy consumption abnormal discrimination models with generalization; using the set of energy consumption abnormal discrimination models with generalization to perform real-time monitoring of base station energy consumption. Thereby further improving the effectiveness and generalization of abnormal energy consumption discrimination.

Description

Technical Field

[0001] The invention belongs to the technical field of power information processing, and particularly relates to a method and system for abnormal monitoring of base station energy consumption based on wavelet decomposition and migration discrimination. Background Technique

[0002] The statements in this part only provide background technical information related to the present invention, and do not necessarily constitute prior art.

[0003] With the increasing number of communication base stations being built, the power consumption of base stations is also continuously growing. Due to the large number of base stations and the heavy operation and maintenance burden, there are currently common problems such as insufficient monitoring of base station power consumption and high power consumption management costs. It is necessary to combine the power consumption and load curve data of the power consumption information collection system to achieve the monitoring of the power consumption of each base station user.

[0004] However, whether it is a macro base station with a large increase in equipment power or a large number of small and micro base stations, they are all powered by a low-voltage power distribution system. The traditional monitoring data of base station energy consumption has a coarse data granularity and cannot achieve equipment-level energy consumption monitoring, so it cannot constitute the data preparation for effective evaluation; moreover, although the power consumption data collected by the power consumption information collection system can provide the hourly frozen power of the base station, there are significant differences in the site location, internal equipment, internal technological process, and load branch situation of different base stations. Therefore, it is difficult to conduct refined monitoring and discrimination for the characteristics of various base stations by using the energy over-limit discrimination method constructed in a unified mode. Summary of the Invention

[0005] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a method for abnormal monitoring of base station energy consumption based on wavelet decomposition and migration discrimination, which realizes adaptive energy consumption monitoring and over-limit warning.

[0006] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0007] In the first aspect, a method for abnormal monitoring of base station energy consumption based on wavelet decomposition and migration discrimination is disclosed, including:

[0008] Based on the historical energy consumption sample data set of the communication base station, complete the screening of energy consumption characteristics and output the energy consumption feature set;

[0009] For the curve energy consumption characteristic data in the energy consumption feature set, perform wavelet decomposition on the energy consumption characteristic curve by using wavelet decomposition to obtain multi-dimensional high and low frequency characteristics;

[0010] Based on the multi-dimensional high and low frequency characteristics and archive feature data obtained from the energy consumption feature set, perform clustering and classification of base station energy consumption samples;

[0011] Taking the high- and low-frequency characteristics, archive feature data, meteorological feature data, and holiday feature data of a certain cluster's energy consumption historical sample dataset as input, an energy consumption anomaly discrimination model is constructed;

[0012] Perform nearest neighbor class model parameter tuning transfer learning on the constructed energy consumption anomaly discrimination model, adaptively adjust the model parameters through reinforcement learning, and output the nearest neighbor cluster energy consumption anomaly discrimination model. Sequentially perform nearest neighbor transfer learning and parameter tuning until the construction of all cluster category energy consumption anomaly discrimination models is completed, forming a generalizable energy consumption anomaly discrimination model set;

[0013] Use the generalizable energy consumption anomaly discrimination model set to monitor the energy consumption of the base station in real time.

[0014] As a further technical solution, it also includes: obtaining the historical energy consumption related data of the communication base station and performing data preprocessing, combining the archive feature data, meteorological feature data, and holiday feature data, and outputting the historical energy consumption sample dataset of the communication base station.

[0015] As a further technical solution, for the curve energy consumption feature data, use the wavedec function in the wavelet decomposition algorithm to perform D-layer wavelet decomposition on the input curve data, and obtain the wavelet decomposition coefficient matrix and the number of coefficients in the matrix;

[0016] Divided into M layers, a total of M + 1 high- and low-frequency features are included.

[0017] As a further technical solution, the construction of the energy consumption anomaly discrimination model is as follows:

[0018] Based on the multi-dimensional frequency division feature data of a certain cluster's energy consumption historical sample set obtained, as well as the archive feature data, meteorological feature data, and holiday feature data of this cluster's energy consumption historical sample set, jointly constitute the sample feature data;

[0019] Based on a certain cluster's historical energy consumption sample data, use the attention mechanism to update the weighted average vector of a certain cluster's energy consumption anomaly discrimination sample data, as the input data of the LSTM neuron at time t;

[0020] Input the weighted average vector of sample i into the energy consumption anomaly discrimination model, use LSTM to construct the energy consumption anomaly discrimination model, and finally the neuron outputs a one-dimensional vector H of the energy consumption anomaly discrimination state t ;

[0021] Based on the training set data of a certain cluster's energy consumption historical sample, respectively obtain the output vectors H of the positive and negative sample data. Through the softmax function, obtain the probability that a certain cluster's energy consumption historical sample training set data is of category j, and determine whether it is an abnormal sample or a normal sample based on the probability threshold, and obtain y′ as the category data discriminated by the model.

[0022] As a further technical solution, constructing an energy consumption anomaly discrimination model further includes:

[0023] Using cross-entropy as the loss function, defining y as the true class data, feeding back the gradient of the model loss function loss(y′, y) to the energy consumption anomaly discrimination model constructed by LSTM, and iteratively solving by adjusting parameters to continuously minimize loss(y′, y) until loss(y′, y) is lower than the set threshold;

[0024] Testing the energy consumption anomaly discrimination model on the historical test set samples. If the accuracy requirement of the test set is not met, re-train after initializing the model parameters; if the accuracy requirement of the test set is met, fix the model parameters and output an energy consumption anomaly discrimination model for a certain clustering.

[0025] As a further technical solution, the process of forming a generalizable energy consumption anomaly discrimination model set is as follows:

[0026] Based on the obtained historical sample data set of base station energy consumption, randomly extract a certain proportion of historical samples as the training set respectively, and the remaining historical samples as the test set;

[0027] Use the energy consumption anomaly discrimination model of a certain clustering obtained by training to discriminate energy consumption anomalies in the training sample set and output discriminated abnormal samples;

[0028] Compare the actual abnormal samples and the discriminated abnormal samples in each clustering historical sample set to obtain the discrimination accuracy of the sample set, and screen out the clustering sample set with the highest discrimination accuracy as the nearest neighbor clustering sample set;

[0029] Based on the nearest neighbor clustering sample set, use reinforcement learning to perform reward feedback tuning on the parameters of the LSTM and attention mechanism of the energy consumption anomaly discrimination model of a certain clustering, and use the tuned nearest neighbor clustering energy consumption anomaly discrimination model to discriminate the training sample set of the nearest neighbor clustering sample set and output discriminated abnormal samples;

[0030] Judge whether the accuracy of the nearest neighbor clustering energy consumption anomaly discrimination model for the nearest neighbor clustering training sample set is higher than the threshold. If not, iteratively adjust the LSTM parameters by reward feedback. If so, fix the parameters and output the nearest neighbor clustering energy consumption anomaly discrimination model.

[0031] The nearest neighbor clustering energy consumption anomaly discrimination model discriminates anomalies in the nearest neighbor clustering test sample set and calculates the test accuracy based on the true abnormal samples;

[0032] As a further technical solution, it further includes: judging whether the test accuracy of the nearest neighbor clustering energy consumption anomaly discrimination model is higher than the set threshold. If not, perform iterative tuning optimization; if so, fix the model parameters and output the nearest neighbor clustering energy consumption anomaly discrimination model;

[0033] Output the nearest neighbor clustering energy consumption anomaly discrimination model set, and combine a certain clustering energy consumption anomaly discrimination model in the output to jointly form a generalized energy consumption anomaly discrimination model set for subsequent differential energy consumption anomaly discrimination according to the base station classification.

[0034] In a second aspect, a base station energy consumption anomaly monitoring system based on wavelet decomposition and migration discrimination is disclosed, including:

[0035] An energy consumption feature set construction module, configured to: based on the historical energy consumption sample data set of communication base stations, complete the screening of energy consumption features and output an energy consumption feature set;

[0036] An energy consumption anomaly discrimination model construction module, configured to: for the curve energy consumption feature data in the energy consumption feature set, perform wavelet decomposition on the energy consumption feature curve by wavelet decomposition to obtain multi-dimensional high and low frequency features;

[0037] Perform base station energy consumption sample clustering and classification based on the multi-dimensional high and low frequency features and archive feature data obtained from the energy consumption feature set;

[0038] Using the high and low frequency features, archive feature data, meteorological feature data, and holiday feature data of a certain clustering energy consumption historical sample data set as inputs, construct an energy consumption anomaly discrimination model;

[0039] An energy consumption anomaly discrimination model set construction module, configured to: perform nearest neighbor class model tuning transfer learning on the constructed energy consumption anomaly discrimination model, adaptively adjust the model parameters through reinforcement learning, output a nearest neighbor clustering energy consumption anomaly discrimination model, and perform nearest neighbor transfer learning tuning in sequence until the construction of all clustering category energy consumption anomaly discrimination models is completed, forming a generalized energy consumption anomaly discrimination model set;

[0040] A real-time monitoring module, configured to: use the generalized energy consumption anomaly discrimination model set to perform real-time monitoring of the base station energy consumption.

[0041] The above one or more technical solutions have the following beneficial effects:

[0042] In the present invention, aiming at the coarse granularity of the monitoring data of base station energy consumption data, wavelet decomposition is used to decompose the one-dimensional curve energy consumption data into multi-dimensional high and low frequency data. On the one hand, by refining the features, the fineness and difference degree of the base station clustering category division are provided. On the other hand, through the sensitivity of the high frequency data to the singularity and mutation of the curve data, and using the attention mechanism to increase the weight of the sensitive data features, thereby improving the accuracy of the abnormal energy consumption discrimination model.

[0043] In the present invention, there are significant differences in the site locations, internal equipment, internal technological processes, and load shunting situations of different base stations. It is difficult to achieve refined discrimination of abnormal energy consumption using an energy consumption over-line discrimination method constructed in a unified mode. Therefore, an abnormal energy consumption discrimination model is constructed based on a classification sample set of base station clustering. To simplify the process of adjusting the model parameters, transfer learning is used to adaptively adjust the parameters of the nearest neighbor model, thereby further improving the effectiveness and generalization of abnormal energy consumption discrimination.

[0044] In the present invention, wavelet decomposition is used to refine the decomposition of energy consumption characteristics, and then the k-means algorithm is used for base station clustering and classification to refine the difference degree between base station classes. Secondly, the LSTM algorithm is used to construct an abnormal energy consumption discrimination model. Finally, nearest neighbor clustering transfer learning is performed on the constructed abnormal energy consumption discrimination model to achieve refined discrimination of energy consumption over-line.

[0045] Advantages of additional aspects of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0047] Figure 1 It is the overall flowchart of the base station energy consumption abnormal monitoring method based on wavelet decomposition and transfer discrimination provided by the embodiments of the present disclosure;

[0048] Figure 2 It is the schematic diagram of 7-layer wavelet decomposition provided by the embodiments of the present disclosure;

[0049] Figure 3 It is the schematic diagram of the 7-layer detail coefficients of the wavelet decomposition mutation curve as high and low frequency features provided by the embodiments of the present disclosure;

[0050] Figure 4 It is the flowchart of constructing an abnormal energy consumption discrimination model using the LSTM algorithm provided by the embodiments of the present disclosure;

[0051] Figure 5 It is the flowchart of transfer learning for adjusting the parameters of the nearest neighbor model using the abnormal energy consumption discrimination model provided by the embodiments of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further explanations of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0053] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0054] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0055] Embodiment 1

[0056] As Figure 1 shown, this embodiment discloses an abnormal energy consumption monitoring method for base stations based on wavelet decomposition and migration discrimination, including:

[0057] A. Extract communication base station energy consumption acquisition data, archive data, and base station-related data provided by the tower company from the databases of the power marketing system, power consumption information acquisition system, and 95598 system, obtain historical energy consumption-related data of communication base stations and perform data preprocessing, and combine meteorological feature data and holiday feature data to output a historical energy consumption sample data set of communication base stations, including a normal energy consumption sample set and an abnormal energy consumption sample set;

[0058] Among them, both the normal energy consumption sample set and the abnormal energy consumption sample set include electrical energy measurement curve data such as current, voltage, and power, archive data such as user classification and power consumption address, base station-related data such as communication base station category, importance level, coordinates, UPS, base station equipment power, and sustainable duration, meteorological data, holiday data, and abnormal energy consumption marking data, and the above data are all structured data;

[0059] Specifically, based on the base station-related data of 10,120 base stations provided by the tower company in a certain region, and extracting the communication base station archive data from January to June 2020 from the databases of the power marketing system, power consumption information acquisition system, and 95598 system, and 1,831,720 energy consumption acquisition data samples measured in days, 120,000 samples are obtained through available sample screening. As shown in Table 1, it is a partial feature example of the samples:

[0060] Table 1 Partial Feature Examples of Low-Voltage Electric Energy Meter Acquisition Data

[0061]

[0062] B. Based on the historical energy consumption sample data set of communication base stations, use the information gain ratio algorithm to complete energy consumption feature screening and output an energy consumption feature set, including curve energy consumption feature data, archive type feature data, and influence type feature data;

[0063] Among them, the process of using the information gain ratio algorithm to complete energy consumption feature screening is as follows:

[0064] The numbers of abnormal energy consumption samples and normal energy consumption samples in the historical energy consumption sample data set of communication base stations are q and The amount of information required to achieve the classification of normal and abnormal categories is as follows:

[0065]

[0066] Classify the historical energy consumption sample dataset of communication base stations according to feature A into v subsets Then the subset contains Q ι abnormal energy consumption samples and normal energy consumption samples. Then the information entropy of each subset is

[0067]

[0068] Among them, the curve data realizes the classification of the historical energy consumption sample dataset of communication base stations by setting data segments. The archive feature data and influence feature data classify the historical energy consumption sample dataset of communication base stations according to the sample values, ι ∈ [1, 2, …, v];

[0069] Therefore, the information gain Gain(A) of feature A:

[0070]

[0071] The information gain ratio Gain-Ratio(A):

[0072]

[0073] Among them, the split information rate Split(A):

[0074]

[0075] Traverse the information gain ratios of all feature attributes as described above and set the information gain ratio threshold. If the information gain ratio is greater than the threshold, then feature A is used as an energy consumption feature, thus forming an energy consumption feature set;

[0076] Among them, the archive feature data includes user classification, electricity consumption address, communication base station category, importance level, coordinates, base station equipment power, sustainable duration, etc., which are used in step D to achieve clustering and classification of base station energy consumption samples and are used as input data for the energy consumption anomaly discrimination model in step E; the influence feature data includes meteorological data and holiday data, which are used in step E to characterize the influence of factors such as meteorology and holidays on the energy consumption data of each base station;

[0077] C. For the curve energy consumption feature data, perform wavelet decomposition on the energy consumption feature curve using wavelet decomposition to obtain multi-dimensional high and low frequency features;

[0078] For the energy consumption characteristic data of the curve, the wavedec function in the wavelet decomposition algorithm is used to perform D-layer wavelet decomposition on the input curve data to obtain the wavelet decomposition coefficient matrix C and the number L of coefficients in the matrix.

[0079] [C, L] = wavedec(A, M,'sym4')

[0080] In the formula, A is the input curve data, M = 7 is the decomposition layer number, C is the wavelet decomposition coefficient, sym4 represents the wavelet transform method, and L is the number of wavelet decomposition coefficients (i.e., the number of coefficients in the matrix).

[0081] See the appendix Figure 2 、 3 As shown, it is divided into 7 layers, then a total of 8 high and low frequency features are included, including 1 low frequency approximation eigenvalue and 7 high frequency detail eigenvalues.

[0082] D. Based on the energy consumption feature set in step B and the multi-dimensional high and low frequency features obtained in step C, use the k-means algorithm based on the Fréchet distance to perform clustering and classification on the base station energy consumption samples, and divide the energy consumption historical sample data set into B = 6 categories.

[0083] Among them, the number of base stations included in each of the B categories (initially 6 categories) of the training set is 1612, 3216, 213, 162, 2859, and 2058 respectively. According to the sample data volume collected daily from January to June, they are 291772 cases, 582096 cases, 38553 cases, 29322 cases, 517479 cases, and 372498 cases respectively. Based on the positive and negative sample balance, re-adjust the available samples obtained in step A to obtain 120,000 cases of samples, so that each category of base stations contains 20,000 cases of historical samples, and all 425 cases of base station abnormal sample data are included, that is, the number of abnormal base stations included in each of the 6 categories of training sets is 65, 125, 20, 15, 110, and 90 respectively, and the corresponding abnormal sample quantities are 2084 cases, 5251 cases, 1423 cases, 1234 cases, 6057 cases, and 6039 cases respectively.

[0084] E. Using the high and low frequency features, file feature data, meteorological feature data, and holiday feature data of a certain clustering energy consumption historical sample data set as inputs, see the appendix Figure 4 As shown, use the LSTM algorithm to construct an energy consumption anomaly discrimination model. After the loss function is lower than the threshold, fix the parameters and output a certain clustering energy consumption anomaly discrimination model, specifically:

[0085] E1. Based on the multi-dimensional frequency division feature data of a certain clustering energy consumption historical sample set obtained in step C, and based on the file feature data, meteorological feature data, and holiday feature data of this type of energy consumption historical sample set obtained in step B, jointly constitute the sample feature data X i(i = 1, 2, …, d), where i represents the i-th sample in a certain clustering energy consumption historical sample set, and d represents the number of samples in a certain clustering energy consumption historical sample;

[0086] Considering constructing an energy consumption anomaly discrimination model based on positive and negative sample balance, a clustering sample set of 2058 base stations is selected for a certain clustering energy consumption anomaly discrimination model. The historical sample set is 20,000 cases, the training sample set is 14,000 cases, and the test sample set is 6,000 cases. The training sample set contains 4227 cases of abnormal sample data, and the test sample set contains 1812 cases of abnormal sample data;

[0087] E2. Based on the historical energy consumption sample data X of a certain clustering i , the attention mechanism (Attention) is used to update the weight value of the energy consumption anomaly discrimination sample data of a certain clustering, that is:

[0088]

[0089] In the formula, is the feature weight value of sample i, ω c represents the output state of sample i at time t - 1 the influence parameter of the input data at time t, represents the neuron output vector of sample i at time t - 1, ω x represents the basic parameter matrix;

[0090] Pass to the softmax function for normalization processing to obtain the normalized calculation weight μ i

[0091]

[0092] Calculate the weighted average value of the eigenvalues of sample X

[0093]

[0094] In the formula, V i represents the weighted average vector of the eigenvalues of sample i;

[0095] Replace the obtained V i with the original input X i , as the input data of the LSTM neuron at time t;

[0096] E3. Input the feature vector V of sample i i into the energy consumption anomaly discrimination model, and use LSTM to construct the energy consumption anomaly discrimination model. The neural unit includes an input gate, a memory unit, an output gate, and a forget gate. The data processing process is as follows:

[0097] ft = sigmoid(W f ·[H t-1 , V t i + b f ) (9)

[0098] i t = sigmoid(W i ·[H t-1, V t i + b i ) (10)

[0099] υ t = tanh(W υ ·[H t-1, V t i + b υ ) (11)

[0100] υ t = f t × υ t-1 + i t × υ t (12)

[0101] o t = sigmoid(W o ·[H t-1 , V t i + b o ) (13)

[0102] H t = o t × tanh(υ t ) (14)

[0103] In the formula, V t i represents the input vector of sample i at time t, H t-1 represents the neuron output vector of sample i at time t - 1, {W f , W i , W c , W o} represents the weight coefficient matrix of the neuron, {b f , b i , b c , b o} represents the offset vector of the neuron, f t represents the data processing process of the forget gate; i t represents the data processing process of the input gate, υ tRepresents the data retention processing process of the memory unit; outputs data f through the forget gate t and outputs data i through the input gate t Adjusts the values of the memory units υ t-1 and υ t ; o t Represents the values of V t i and H t-1 after passing through the sigmoid activation function, and combines with the tanh function to implement the internal loop value update. Finally, the neuron outputs a one-dimensional vector H for abnormal energy consumption discrimination status t ;

[0104] E4. Based on the training set data of a certain clustering energy consumption historical sample, respectively obtain the output vectors H of positive and negative sample data. Through the softmax function, the historical sample X i generates a discriminant vector H i The probability of being classified as category j is:

[0105]

[0106] In the formula, W c and b c are the learning parameters of the softmax function;

[0107] Obtain the probability that a certain clustering energy consumption historical sample training set data is category j, and determine whether it is an abnormal sample or a normal sample based on the probability threshold, and obtain y′ as the category data discriminated by the model;

[0108] E5. Use cross-entropy as the loss function, define y as the true category data, then the model loss function is as follows:

[0109]

[0110] Feed back the gradient of the model loss function loss(y′, y) to steps E3 and E4. Use LSTM to construct an energy consumption abnormal discrimination model, and solve it through parameter adjustment and iteration to continuously minimize loss(y′, y) until loss(y′, y) is lower than the set threshold, then enter step E6;

[0111] E6. Test the energy consumption abnormal discrimination model on the historical test set samples. If the test set accuracy requirement is not met, feedback to step E2 for model parameter initialization and retraining; if the test set accuracy requirement is met, fix the model parameters and output the energy consumption abnormal discrimination model for a certain clustering sample set;

[0112] F. See the appendix Figure 5As shown, perform nearest neighbor class model tuning and transfer learning on the energy consumption anomaly discrimination model constructed in step E. Adaptive adjust the model parameters through reinforcement learning, and output the nearest neighbor clustering energy consumption anomaly discrimination model. Perform nearest neighbor transfer learning and tuning in sequence until the construction of the energy consumption anomaly discrimination models for all clustering categories is completed, forming a generalizable set of energy consumption anomaly discrimination models;

[0113] F1. Based on the historical energy consumption sample dataset of type B base stations obtained in step D, randomly extract 70% of the historical samples as the training set and the remaining 30% as the test set. According to the actual historical energy consumption sample dataset of base stations, the number of base stations included in the training sets of class B-1 (initially 5 classes) are 1,612, 3,216, 213, 162, and 2,859 respectively, and the amounts of abnormal sample data are 2,084 cases, 5,251 cases, 1,423 cases, 1,234 cases, and 6,057 cases respectively. The amounts of abnormal sample data included in the training sets are 1,459 cases, 3,676 cases, 996 cases, 864 cases, and 4,240 cases respectively, and the amounts of abnormal sample data included in the test sets are 625 cases, 1,575 cases, 427 cases, 370 cases, and 1,817 cases respectively;

[0114] F2. Use the energy consumption anomaly discrimination model of a certain clustering obtained by training in step E to perform energy consumption anomaly discrimination on the remaining training sample sets of class B-1, initially 5 classes, and output the discriminated abnormal samples;

[0115] F3. Compare the actual abnormal samples and the discriminated abnormal samples in each clustering historical sample set, so as to obtain the discrimination accuracy rate of class B-1, that is, the initial 5-class sample set, and select the clustering sample set with the highest discrimination accuracy rate as the nearest neighbor clustering sample set;

[0116] F4. Based on the nearest neighbor clustering sample set, use reinforcement learning to perform reward feedback tuning on the parameters of the LSTM and attention mechanism of the energy consumption anomaly discrimination model of a certain clustering, and use the tuned nearest neighbor clustering energy consumption anomaly discrimination model to discriminate the training sample set of the nearest neighbor clustering sample set, and output the discriminated abnormal samples;

[0117] The tuning iteration process τ of the parameters of the LSTM and attention mechanism of the energy consumption anomaly discrimination model of a certain clustering is as follows:

[0118] τ = {s 1 , a 1 , r 1 , s 2 , a 2 , r 2 , …, s t , a t , r t , …, s T , a T , r T} (17)

[0119] Among them, τ is the process of adjusting a certain parameter once to change the discrimination accuracy of the energy consumption anomaly discrimination model of a certain clustering. s represents the parameter values of the current model LSTM and the attention mechanism. a represents the adjustment actions of increasing or decreasing the parameter values of LSTM and the attention mechanism. θ represents the parameter set of LSTM and the attention mechanism of the current model. r is the discrimination reward value obtained by adjusting the parameters to improve the discrimination accuracy of the model. T is the total number of adjustments in this discrimination;

[0120] Among them, R(τ) represents the reward obtained in the process of adjusting the parameter τ this time;

[0121]

[0122] Among them, is the average reward expectation of N discrimination processes, that is, to maximize the training objective. P(τ / θ) is the conditional probability of the τ parameter adjustment process when the parameter is θ;

[0123] Maximize using gradient ascent and update the parameter θ

[0124]

[0125] Among them, θ n is the model parameter of the nth discrimination process, n ∈ [1, N], and η is the learning rate for setting the upper threshold of the threshold; is the gradient value fed back according to the discrimination result after the parameter adjustment in the nth time:

[0126]

[0127] Among them, τ n is the nth discrimination process, is the model parameter value after the (t - 1)th model parameter adjustment in the nth discrimination process, is the parameter adjustment action after the (t - 1)th model parameter adjustment in the nth discrimination process, and b is the set sample balance threshold;

[0128] Set N discrimination processes as a batch, and use the nearest neighbor clustering energy consumption anomaly discrimination model after parameter adjustment to discriminate the training sample set of the nearest neighbor clustering sample set, and output the discriminated abnormal samples;

[0129] F5. Determine whether the accuracy of the nearest neighbor clustering energy consumption anomaly discrimination model for the nearest neighbor clustering training sample set is higher than the threshold. If not, return to step F4, and iteratively reward and feedback to adjust the parameters of LSTM and the attention mechanism. If so, fix the parameters and output the nearest neighbor clustering energy consumption anomaly discrimination model;

[0130] F6. The nearest neighbor clustering energy consumption anomaly discrimination model performs anomaly discrimination on the nearest neighbor clustering test sample set, calculates the test accuracy based on the true anomaly samples. Among the remaining 5 types of test sample sets, the actual number of anomaly samples is 34, 10, 4, 21, and 29 respectively.

[0131] F7. Determine whether the test accuracy of the nearest neighbor clustering energy consumption anomaly discrimination model is higher than the set threshold. If not, go to step F4 for iterative parameter tuning and optimization. If so, fix the model parameters, output the nearest neighbor clustering energy consumption anomaly discrimination model, B ← B - 1, and determine whether B is zero. If so, go to step F8; if not, go to step F2 and replace the energy consumption anomaly discrimination model of a certain clustering with the nearest neighbor clustering energy consumption anomaly discrimination model, and perform anomaly discrimination on the B - 1 type of clustering sample set.

[0132] F8. Output the B - 1 type of nearest neighbor clustering energy consumption anomaly discrimination model, and combine it with the energy consumption anomaly discrimination model of a certain clustering output in step E to jointly form a generalized energy consumption anomaly discrimination model set for subsequent differential energy consumption anomaly discrimination according to the base station classification.

[0133] Table 1 Energy Consumption Anomaly Discrimination Accuracy

[0134]

[0135] As shown in Table 1, it is the comparison of energy consumption anomaly discrimination data before and after transfer learning training:

[0136] Before transfer learning:

[0137]

[0138] After transfer learning:

[0139]

[0140] As shown above, the recognition recall rate before and after transfer learning has been significantly improved. A total of 425 base stations with energy consumption anomalies are involved. Based on the above sample data, 338 base stations with abnormal energy consumption were identified before transfer learning, and 406 were identified after transfer learning, indicating that this method improves the discrimination performance of energy consumption anomaly samples in different clustering categories and verifies the usability of the near - neighbor transfer learning parameter tuning of the model proposed in this patent.

[0141] G. Use the generalized energy consumption anomaly discrimination model set to monitor the energy consumption of base stations in real - time, dispatch personnel to conduct on - site verification of base stations with over - limit energy consumption, and feedback the verification data to step A, so as to continuously enrich the sample library and improve the energy consumption anomaly discrimination model set.

[0142] The technical solution of the present disclosure realizes adaptive energy consumption monitoring and over-line warning by refining monitoring data and optimizing the monitoring model. Thus, during the daily maintenance process of base stations, the operation and maintenance personnel can timely and accurately master the energy consumption data and abnormal energy consumption information of the entire network of base stations, quickly and specifically locate the positions of abnormal base stations and judge faults, effectively reducing the operation and maintenance workload of the operation and maintenance personnel.

[0143] Embodiment 2

[0144] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of the above method are implemented.

[0145] Embodiment 3

[0146] The purpose of this embodiment is to provide a computer-readable storage medium.

[0147] A computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, the steps of the above method are executed.

[0148] Embodiment 4

[0149] The purpose of this embodiment is to provide a base station energy consumption abnormal monitoring system based on wavelet decomposition and migration discrimination, including:

[0150] An energy consumption feature set construction module, configured to: based on the historical energy consumption sample data set of communication base stations, complete the screening of energy consumption features and output an energy consumption feature set;

[0151] An energy consumption abnormal discrimination model construction module, configured to: for the curve energy consumption feature data in the energy consumption feature set, perform wavelet decomposition on the energy consumption feature curve by using wavelet decomposition to obtain multi-dimensional high and low frequency features;

[0152] Perform base station energy consumption sample clustering and classification based on the multi-dimensional high and low frequency features and file feature data obtained from the energy consumption feature set;

[0153] Taking the high and low frequency features, file feature data, meteorological feature data, and holiday feature data of a certain clustering energy consumption historical sample data set as inputs, construct an energy consumption abnormal discrimination model;

[0154] An energy consumption abnormal discrimination model set construction module, configured to: perform nearest neighbor class model parameter tuning and transfer learning on the constructed energy consumption abnormal discrimination model, adaptively adjust the model parameters through reinforcement learning, output a nearest neighbor clustering energy consumption abnormal discrimination model, and sequentially perform nearest neighbor transfer learning and parameter tuning until the construction of all clustering category energy consumption abnormal discrimination models is completed, forming a generalizable energy consumption abnormal discrimination model set;

[0155] A real-time monitoring module, configured to: use a generalized abnormal energy consumption discrimination model set to perform real-time monitoring on the energy consumption of a base station.

[0156] In the devices of the above second, third, and fourth embodiments, the steps involved correspond to those in the first method embodiment. For the specific implementation manners, reference may be made to the relevant description part of the first embodiment. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and cause the processor to execute any method in the present invention.

[0157] Those skilled in the art should understand that the above-mentioned modules or steps of the present invention can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0158] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A method for abnormal energy consumption monitoring of base stations based on wavelet decomposition and transfer discrimination, characterized in that, it includes: S1: Based on the historical energy consumption sample data set of communication base stations, complete the screening of energy consumption characteristics and output the energy consumption characteristic set; S2: For the curve energy consumption characteristic data in the energy consumption characteristic set, use wavelet decomposition to decompose the energy consumption characteristic curve and obtain multi-dimensional high and low frequency characteristics; S3: Based on the multi-dimensional high and low frequency characteristics and archive characteristic data obtained from the energy consumption characteristic set, perform clustering and classification of base station energy consumption samples; S4: Use the high and low frequency characteristics, archive characteristic data, meteorological characteristic data, and holiday characteristic data of a certain cluster of energy consumption historical sample data set as inputs to construct an abnormal energy consumption discrimination model; Constructing an abnormal energy consumption discrimination model specifically includes: S401: Based on the multi-dimensional frequency-divided feature data of a certain clustered energy consumption historical sample set obtained, as well as the file feature data, meteorological feature data, and holiday feature data of this type of energy consumption historical sample set, jointly constitute the sample feature data X i (i = 1, 2, …, d); i represents the i-th sample in a certain clustered energy consumption historical sample set, and d represents the number of samples in a certain clustered energy consumption historical sample; S402: Based on a certain cluster of historical energy consumption sample data, use the attention mechanism to update the characteristic weight values of a certain cluster of abnormal energy consumption discrimination sample data; Transfer the characteristic weight values to the softmax function for normalization processing, obtain the normalized calculation weights, and calculate the weighted average vector of the characteristic values of sample X; S403: Replace the original input X with the obtained weighted average vector i , which serves as the input data of the LSTM neuron at time t. Specifically, input the weighted average vector of sample i into the energy consumption anomaly discrimination model. The energy consumption anomaly discrimination model is constructed using LSTM. Finally, the neuron outputs a one-dimensional vector H representing the energy consumption anomaly discrimination status t ; S404: Based on the training set data of a certain clustered energy consumption history sample, obtain the output vectors of the positive and negative sample data respectively. Through the softmax function, for the historical sample X i Generate the discriminant vector H i Classify it into category j; Obtain the probability that a certain cluster of energy consumption historical sample training set data belongs to category j, and determine whether it is an abnormal sample or a normal sample based on the probability threshold to obtain the category data discriminated by the model; S405: Use cross-entropy as the loss function, define y as the true category data, Feed back the gradient of the model loss function to the abnormal energy consumption discrimination model constructed by LSTM, and solve by parameter tuning iteration to continuously minimize the loss function until the loss function is lower than the set threshold; S406: Test the abnormal energy consumption discrimination model on the historical test set samples. If the accuracy requirement of the test set is not met, perform model parameter initialization and retraining; if the accuracy requirement of the test set is met, fix the model parameters and output a certain cluster of abnormal energy consumption discrimination models; S5: Perform nearest neighbor class model parameter tuning transfer learning on the constructed abnormal energy consumption discrimination model, adaptively adjust the model parameters through reinforcement learning, output the nearest neighbor cluster abnormal energy consumption discrimination model, and perform nearest neighbor transfer learning parameter tuning in turn until the construction of all cluster category abnormal energy consumption discrimination models is completed, forming a general abnormal energy consumption discrimination model set; S6: Use the general abnormal energy consumption discrimination model set to perform real-time monitoring of base station energy consumption.

2. The method for abnormal energy consumption monitoring of base stations based on wavelet decomposition and transfer discrimination according to claim 1, characterized in that it also includes: Obtain the historical energy consumption related data of the communication base station and perform data preprocessing, and combine the archive characteristic data, meteorological characteristic data and holiday characteristic data to output the historical energy consumption sample data set of the communication base station.

3. The method for abnormal energy consumption monitoring of base stations based on wavelet decomposition and transfer discrimination according to claim 1, characterized in that, For the curve energy consumption characteristic data, use the wavedec function in the wavelet decomposition algorithm to perform D-layer wavelet decomposition on the input curve data to obtain the wavelet decomposition coefficient matrix and the number of coefficients in the matrix; If it is divided into M layers, it includes a total of M + 1 high and low frequency characteristics.

4. The method for abnormal energy consumption monitoring of base stations based on wavelet decomposition and transfer discrimination according to claim 1, characterized in that, The process of forming a generalization-enabled energy consumption anomaly discrimination model set is as follows: Based on the obtained historical sample data set of the base station's energy consumption, a certain proportion of historical samples are randomly extracted as the training set, and the remaining historical samples are used as the test set; Use the energy consumption anomaly discrimination model of a certain cluster obtained through training to perform energy consumption anomaly discrimination on the training sample set and output the discriminated abnormal samples; Compare the actual abnormal samples and the discriminated abnormal samples in each cluster's historical sample set to obtain the discrimination accuracy rate of the sample set, and select the cluster sample set with the highest discrimination accuracy rate as the nearest neighbor cluster sample set; Based on the nearest neighbor cluster sample set, use reinforcement learning to perform reward feedback tuning on the parameters of the LSTM and attention mechanism of the energy consumption anomaly discrimination model of a certain cluster, and use the tuned nearest neighbor cluster energy consumption anomaly discrimination model to discriminate the training sample set of the nearest neighbor cluster sample set and output the discriminated abnormal samples; Determine whether the accuracy rate of the nearest neighbor cluster energy consumption anomaly discrimination model for the nearest neighbor cluster training sample set is higher than the threshold. If not, iteratively perform reward feedback to adjust the LSTM parameters. If so, fix the parameters and output the nearest neighbor cluster energy consumption anomaly discrimination model.

5. The base station energy consumption anomaly monitoring method based on wavelet decomposition and transfer discrimination according to claim 4, characterized in that it also includes: The nearest neighbor cluster energy consumption anomaly discrimination model performs anomaly discrimination on the nearest neighbor cluster test sample set and calculates the test accuracy rate based on the true abnormal samples; Determine whether the test accuracy rate of the nearest neighbor cluster energy consumption anomaly discrimination model is higher than the set threshold. If not, perform iterative tuning and optimization; if so, fix the model parameters and output the nearest neighbor cluster energy consumption anomaly discrimination model; Output the nearest neighbor cluster energy consumption anomaly discrimination model, and combine it with the output energy consumption anomaly discrimination model of a certain cluster to jointly form a generalization-enabled energy consumption anomaly discrimination model set for subsequent differential energy consumption anomaly discrimination according to the base station classification.

6. The base station energy consumption anomaly monitoring system based on wavelet decomposition and transfer discrimination adopts the base station energy consumption anomaly monitoring method based on wavelet decomposition and transfer discrimination according to claim 1, characterized in that it includes: The energy consumption feature set construction module is configured to: complete energy consumption feature screening based on the historical energy consumption sample data set of the communication base station and output the energy consumption feature set; The energy consumption anomaly discrimination model construction module is configured to: perform wavelet decomposition on the energy consumption feature curve using wavelet decomposition for the curve energy consumption feature data in the energy consumption feature set to obtain multi-dimensional high and low frequency features; Perform base station energy consumption sample clustering and classification based on the multi-dimensional high and low frequency features and archive feature data obtained from the energy consumption feature set; Construct an energy consumption anomaly discrimination model with the high and low frequency features, archive feature data, meteorological feature data, and holiday feature data of a certain cluster's energy consumption historical sample data set as the input. The energy consumption anomaly discrimination model set construction module is configured to: perform nearest neighbor class model tuning and transfer learning on the constructed energy consumption anomaly discrimination model, adaptively adjust the model parameters through reinforcement learning, output the nearest neighbor clustering energy consumption anomaly discrimination model, and sequentially perform nearest neighbor transfer learning and tuning until the construction of the energy consumption anomaly discrimination models for all clustering categories is completed, forming a generalizable energy consumption anomaly discrimination model set; The real-time monitoring module is configured to: perform real-time monitoring of the energy consumption of the base station using the generalizable energy consumption anomaly discrimination model set.

7. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein, when the processor executes the program, the steps of the method according to any one of claims 1-5 are implemented.

8. A computer-readable storage medium, having a computer program stored thereon, wherein, when the program is executed by the processor, the steps of the method according to any one of claims 1-5 are executed.

Images