Equipment energy consumption dynamic evaluation method based on multi-source data fusion

Abstract

The invention provides an equipment energy consumption dynamic evaluation method based on multi-source data fusion, and relates to the field of energy consumption evaluation, and the method specifically comprises the steps: constructing an energy consumption evaluation data set, carrying out the statistics of the missing rate, jump rate and straightness rate of an energy consumption time sequence, and generating dimension-by-dimension credibility gating. Calculating the deviation degree of the multi-source data in the energy consumption time sequence based on the median vector to obtain the dynamic fluctuation intensity, differentiating the multi-source data and the median vector, normalizing the multi-source data and the median vector in combination with the robust scale vector and the dynamic fluctuation intensity, and performing element-by-element combination operation with dimension-by-dimension credibility gating to obtain normalized output, a soft division probability is obtained through a gating circulation unit, linear transformation and a Softmax activation function in sequence, weighted fusion is carried out on normalized output, global feature representation is formed through one-dimensional convolution and summary average, dynamic energy consumption evaluation values of the main equipment and the auxiliary equipment are obtained through mapping by combining fusion representation vectors, and the dynamic energy consumption evaluation values of the main equipment and the auxiliary equipment are fused to form a dynamic energy consumption evaluation value of the equipment. Accurate and dynamic evaluation of the energy consumption state of the equipment is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of energy consumption evaluation, and in particular to a device energy consumption dynamic evaluation method based on multi-source data fusion. BACKGROUND

[0002] In the production operation process of industrial enterprises, the energy consumption of equipment is directly related to the production cost and equipment utilization rate of enterprises. With the increase in the number of production equipment and the diversification of operating conditions, the proportion of enterprise energy expenditure in operating costs is increasing. If there is a lack of fine control of the energy consumption state of equipment, it is easy to cause energy waste, long-term high-load operation of equipment or abnormal energy consumption difficult to find in time, and thus affect the cost control ability, production plan execution efficiency and equipment service life of enterprises. Therefore, accurate and dynamic evaluation of equipment energy consumption is of great significance for enterprises to achieve fine operation management.

[0003] The existing equipment energy consumption evaluation technology mostly uses a single energy consumption indicator or a static statistical analysis method based on historical data, and usually relies on manual experience to set threshold values or simple models for judgment, which is difficult to adapt to the actual working conditions with frequent changes in equipment operating state. Some evaluation methods that introduce multi-source data do not fully consider the quality differences and time series fluctuation characteristics of different data sources in the modeling process, and are easily affected by missing data, abnormal jumps or noise interference, resulting in insufficient stability of the evaluation results, and it is difficult to provide reliable real-time energy consumption decision basis for enterprises.

[0004] Since the energy consumption of equipment is often driven by multi-source data, and presents the characteristics of continuous evolution and stage change in the time dimension, if a multi-source data fusion mechanism can be introduced in the energy consumption evaluation process, and the data quality, time series fluctuation and state change are modeled collaboratively, it will help to more comprehensively depict the real energy consumption state of equipment, improve the stability and reliability of the evaluation results, and thus provide more practical value technical support for enterprises to carry out energy consumption analysis and operation optimization. SUMMARY

[0005] Aiming at the problems of large quality difference, unstable time sequence fluctuation and insufficient energy consumption state representation of multi-source data in the process of equipment energy consumption monitoring, the application proposes a device energy consumption dynamic evaluation method based on multi-source data fusion. First, the multi-source data affecting the equipment energy consumption is obtained, and the energy consumption evaluation data set is constructed. Then, the missing rate, jump rate and flat rate of the energy consumption time sequence are counted to form a multi-source collection quality statistical vector, and a dimension-by-dimension credibility gate is generated. At the same time, the deviation degree of multi-source data is recursively modeled with the median vector as a reference to obtain dynamic fluctuation intensity. On this basis, the multi-source data is normalized by combining the robust scale vector and the dynamic fluctuation intensity to obtain the normalized output, and the time sequence fusion state vector is constructed to obtain the soft partition probability. Further, the normalized output is weighted and fused based on the soft partition probability to obtain a fusion representation vector, and a one-dimensional convolution and a full-time summary are performed to form a global feature representation. The global feature representation and the fusion representation vector are respectively nonlinearly mapped to obtain the main and auxiliary equipment energy consumption dynamic evaluation values, and the two values are aggregated to form the final equipment energy consumption dynamic evaluation value.

[0006] A device energy consumption dynamic evaluation method based on multi-source data fusion, characterized by comprising the following steps: S1, obtaining multi-source data affecting the equipment energy consumption, and constructing an energy consumption evaluation data set; S2, counting the missing rate, jump rate and flat rate of the energy consumption time sequence, and concatenating by dimension to generate a dimension-by-dimension credibility gate using linear transformation and ReLU, Sigmoid activation functions, and recursively modeling the deviation degree of multi-source data with the median vector as a reference to obtain dynamic fluctuation intensity; S3, differentiating the multi-source data and the median vector, combining the robust scale vector and the dynamic fluctuation intensity to scale the differentiated results by dimension, performing element-by-element combination operation with the dimension-by-dimension credibility gate to obtain normalized output, inputting the gate recurrent unit to form a time sequence fusion state vector, and obtaining soft partition probability after linear transformation and Softmax activation function; S4, associating and averaging the absolute value of the normalized output with the mapping weight by dimension, weighting and summarizing to form a fusion representation vector through the soft partition probability, and forming a sequence in time sequence and performing one-dimensional convolution along the time dimension to obtain a time sequence convolution feature. After summarizing and averaging in the full-time range, a global feature representation is constructed; S5, linearly transforming the global feature representation and the fusion representation vector respectively and nonlinearly mapping to obtain the main and auxiliary equipment energy consumption dynamic evaluation values, aggregating and averaging the two values to form the final equipment energy consumption dynamic evaluation value; S6, constructing a device energy consumption dynamic evaluation model, inputting the energy consumption evaluation data set, and iteratively training through steps S2 to S5 until convergence.

[0007] Preferably, the construction process of the equipment energy consumption evaluation dataset includes two stages of data acquisition and preprocessing. In the data acquisition stage, the collected multi-source data includes equipment input voltage, equipment input current, equipment instantaneous active power, cumulative power consumption, equipment operating speed, equipment load rate, equipment surface temperature, and environmental temperature. In the data preprocessing stage, the validity of the collected multi-source data is verified channel by channel, and missing records, continuous repeated records, and data exceeding the rated operating parameter range of the equipment are directly excluded. Local missing data is compensated by the previous valid sampling value to obtain multi-source energy consumption time series data. Then the upper and lower threshold values of the multi-source data are determined, and the equipment input voltage, equipment input current, equipment instantaneous active power, cumulative power consumption, equipment operating speed, equipment load rate, and equipment surface temperature are normalized accordingly. The average value of the environmental temperature data in a continuous time period is used as the reference for centering processing. After the scale is unified, all data is uniformly resampled and time-aligned. Finally, the time-aligned multi-source data is segmented by a fixed time window length, and the corresponding 8 types of data in each time window are combined and packaged in a fixed order as an energy consumption time series. All energy consumption time series are arranged in chronological order to form the equipment energy consumption evaluation dataset, which serves as the input data for the subsequent equipment energy consumption dynamic evaluation model.

[0008] Preferably, in the S2 step, the missing rate, jump rate, and flat rate of the energy consumption time series are calculated, and the linear transformation and ReLU and Sigmoid activation functions are used to generate a dimension-by-dimension credibility gate. The deviation of the multi-source data is recursively modeled with reference to the median vector to obtain the dynamic fluctuation intensity. The data integrity and change characteristics of the energy consumption time series in each dimension are statistically analyzed to obtain the missing rate, jump rate, and flat rate values of each dimension. The missing rate, jump rate, and flat rate are combined according to the dimension index order to form a multi-source acquisition quality statistical vector. The linear transformation and ReLU activation function are applied to the multi-source acquisition quality statistical vector to obtain an intermediate feature representation. Then, the intermediate feature representation is input to the second layer linear transformation and Sigmoid activation function to obtain a dimension-by-dimension credibility gate. The multi-source data at the current time step and the median vector are differenced, and the difference result is squared to form the deviation term at the current time step. Then, a learnable smoothing coefficient is introduced to recursively weight the deviation term and the dynamic fluctuation intensity at the previous time step according to a predetermined proportion to construct the dynamic fluctuation intensity corresponding to the current time step.

[0009] Further, in view of the inconsistent quality of multi-source data collection, abnormal fluctuations and stable state in the device energy consumption evaluation process, the application first performs statistical analysis on the data integrity and change characteristics of the energy consumption time series in each dimension, obtains the missing rate value, jump rate value and flat rate value of each dimension, and combines the missing rate, jump rate and flat rate according to the dimension index order respectively, and then splices them in a predetermined order to construct a multi-source collection quality statistical vector, thereby realizing unified quantitative description of the reliability of multi-source data collection; secondly, the linear transformation and ReLU activation function are applied to the multi-source collection quality statistical vector to obtain an intermediate feature representation, and further through the second layer linear transformation and Sigmoid activation function, a dimension-by-dimension credibility gate is generated, so that the data in different dimensions can be adaptively adjusted in the subsequent modeling process according to the collection quality Influence weight; finally, the multi-source data and the median vector at the current time step are subjected to difference processing, and the square operation is performed on the difference result to form a deviation term, a learnable smoothing coefficient is introduced to recursively weight the deviation term and the dynamic fluctuation intensity of the previous time step, and the dynamic fluctuation intensity corresponding to the current time step is constructed, so as to continuously depict the real change trend of the device energy consumption state while suppressing noise interference.

[0010] Preferably, in the S3 step, the multi-source data and the median vector are differentiated, the difference result is scaled by the robust scale vector and the dynamic fluctuation intensity, and the normalized output is obtained by performing element-by-element combination operation with the dimension-by-dimension credibility gate, and the time sequence fusion state vector is formed by inputting the gate recurrent unit, and after linear transformation and Softmax activation function, the soft partition probability is obtained. The multi-source data and the median vector at the first time step are subjected to difference processing to construct a deviation result; then, a robust scale vector based on quantile statistics is introduced, the square term of the robust scale vector and the dynamic fluctuation intensity at the first time step are added, and the square root processing is performed on the added result to form a dynamic normalization scale; then, the dynamic normalization scale is used to scale the deviation result in each dimension to obtain an intermediate result, and the normalized output is obtained by performing element-by-element combination operation with the dimension-by-dimension credibility gate. The normalized output is input into the gate recurrent unit to obtain the time sequence fusion state vector. The time sequence fusion state vector is subjected to linear transformation and nonlinear mapping by using the Softmax activation function to obtain the soft partition probability.

[0011] Further, the deviation result is obtained by differentiating the multi-source data from the median vector, and the dynamic normalization scale is constructed by combining the robust scale vector and the dynamic fluctuation intensity, so that the energy consumption related data with different dimensions and different fluctuation amplitudes are adaptively modulated under a unified scale, effectively reducing the interference of abnormal values and noise on the energy consumption evaluation result; the intermediate result is obtained by using the dynamic normalization scale to scale the deviation result dimension by dimension, and the normalized output is obtained by combining the element by element combination operation of the dimension by dimension reliability gate, avoiding the excessive influence of low reliability dimensions on the overall evaluation result, thereby improving the stability and reliability of the energy consumption dynamic evaluation; at the same time, the normalized output is time series fused by using the gating recurrent unit, and the soft partition probability is obtained by using the Softmax activation function, realizing the continuous and smooth description of the device energy consumption running state, so that the model can reflect the short-term energy consumption fluctuation and maintain the sensitivity to long-term change trend.

[0012] Preferably, in the S4 step, the absolute value of the normalized output is dimensionally associated with the mapping weight and averaged, and the fusion representation vector is formed by weighted summation through the soft partition probability, and the time series convolution feature is obtained by one-dimensional convolution along the time dimension after the sequence is composed in time sequence; after the global feature representation is summarized and averaged in the whole time range, the global feature representation is constructed; The absolute value of the normalized output of each dimension at the first time step is processed, and the mapping weight in the corresponding partition state is associated and calculated, and the fusion representation vector is constructed by weighted processing and summation through the soft partition probability; The absolute value of the normalized output of each dimension at the first time step is processed, and the mapping weight in the corresponding partition state is associated and calculated, and the fusion representation vector is constructed by weighted processing and summation through the soft partition probability; The fusion representation vectors of each time step are arranged in time sequence to construct a fusion representation vector sequence, and one-dimensional convolution operation is performed along the time dimension to form a time series convolution feature; The corresponding time series convolution features of each time step are sequentially processed in the time dimension, and the overall average operation is performed in the range of all time steps to construct the global feature representation.

[0013] Further, the deviation result is obtained by differentiating the multi-source data from the median vector, and the dynamic normalization scale is constructed by combining the robust scale vector and the dynamic fluctuation intensity, so that the energy consumption related data with different dimensions and different fluctuation amplitudes are adaptively modulated under a unified scale, effectively reducing the interference of abnormal values and noise on the energy consumption evaluation result; the intermediate result is obtained by using the dynamic normalization scale to scale the deviation result dimension by dimension, and the normalized output is obtained by combining the element by element combination operation of the dimension by dimension reliability gate, avoiding the excessive influence of low reliability dimensions on the overall evaluation result, thereby improving the stability and reliability of the energy consumption dynamic evaluation; at the same time, the normalized output is time series fused by using the gating recurrent unit, and the soft partition probability is obtained by using the Softmax activation function, realizing the continuous and smooth description of the device energy consumption running state, so that the model can reflect the short-term energy consumption fluctuation and maintain the sensitivity to long-term change trend. The normalized output of each dimension of the time step performs absolute value processing, and is associated with the mapping weight in the corresponding partition state for dimension-by-dimension calculation, average processing in each dimension range, and weighted aggregation combined with the soft partition probability, to construct a fusion representation vector capable of reflecting the multi-state collaborative characteristics, thereby realizing adaptive integration of different energy consumption state information. On this basis, the fusion representation vectors of each time step are arranged in time sequence to construct a fusion representation vector sequence, and one-dimensional convolution operation is performed along the time dimension to model the local change relationship between adjacent time steps, forming a time sequence convolution feature to enhance the expression ability of short-term energy consumption change patterns. Subsequently, the time sequence convolution features corresponding to each time step are sequentially aggregated in the time dimension, and overall average operation is performed in the range of all time steps to construct a global feature representation at the time series level, thereby extracting stable features reflecting the overall energy consumption level and change trend of the device while maintaining local dynamic information.

[0014] Preferably, in the S5 step, the global feature representation is linearly transformed and respectively obtains the main and auxiliary device energy consumption dynamic evaluation values through nonlinear mapping, the two are aggregated and averaged to form the final device energy consumption dynamic evaluation value; The global feature representation is linearly transformed and nonlinearly mapped using a ReLU activation function to obtain the main device energy consumption dynamic evaluation value. The fusion representation vector sequence is performed with a straight splicing operation to construct an overall fusion representation vector, which is linearly transformed and nonlinearly mapped using a ReLU activation function to form the auxiliary device energy consumption dynamic evaluation value. The main device energy consumption dynamic evaluation value and the auxiliary device energy consumption dynamic evaluation value are aggregated and averaged to form the final device energy consumption dynamic evaluation value.

[0015] Further, to realize hierarchical characterization and comprehensive evaluation of the device energy consumption level, the global feature representation is linearly transformed and nonlinearly mapped using a ReLU activation function to obtain the main device energy consumption dynamic evaluation value, so that the evaluation result can mainly reflect the stable energy consumption characteristics and long-term change trend during the overall operation of the device. At the same time, the fusion representation vector sequence is performed with a straight splicing operation to construct an overall fusion representation vector, which is linearly transformed and nonlinearly mapped using a ReLU activation function to form the auxiliary device energy consumption dynamic evaluation value, thereby fully retaining the energy consumption change information of each time step and supplementing the characterization of local fluctuations and stage characteristics. Finally, the main device energy consumption dynamic evaluation value and the auxiliary device energy consumption dynamic evaluation value are aggregated and averaged to form the final device energy consumption dynamic evaluation value, so that the evaluation result takes into account the global stability and local dynamics while avoiding the deviation caused by a single perspective, improving the accuracy and reliability of the device energy consumption dynamic evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flow chart of a device energy consumption dynamic evaluation method based on multi-source data fusion provided by the application.

[0017] Figure 2 is a structure diagram of constructing dimension-by-dimension credibility gating and dynamic fluctuation intensity provided by the application.

[0018] Figure 3 is a structure diagram of constructing soft partition probability provided by the application.

[0019] Figure 4 is a structure diagram of constructing global feature representation provided by the application.

[0020] Figure 5 is a model training loss curve diagram provided by the application.

[0021] Figure 6 is a device energy consumption dynamic evaluation effect diagram provided by the application. DETAILED DESCRIPTION

[0022] The application provides a device energy consumption dynamic evaluation method based on multi-source data fusion, aiming at the problems of large quality difference of multi-source data, unstable time sequence fluctuation and insufficient energy consumption state representation in the device energy consumption monitoring process, the missing rate, jump rate and flat rate of the energy consumption time sequence in the energy consumption evaluation data set are counted, and after being spliced by dimension, the linear transformation is combined with the ReLU and Sigmoid activation functions to generate the dimension-by-dimension credibility gating, and the deviation degree of the multi-source data is recursively modeled with the median vector as a reference; on this basis, the multi-source data and the median vector are subjected to difference processing, the difference results are scaled by dimension by combining the robust scale vector and the dynamic fluctuation intensity, and the normalized output is obtained by performing element-by-element combination operation with the dimension-by-dimension credibility gating, and the time sequence fusion state vector is formed by further inputting the gated recurrent unit, the soft partition probability is obtained after linear transformation and Softmax activation function, the absolute value of the normalized output under each partition state is associated with the mapping weight, the fusion representation vector is formed by weighted summarizing through the soft partition probability, the time sequence convolution feature is obtained through one-dimensional convolution, and the global feature representation is constructed after summarizing and averaging; finally, the global feature representation and the fusion representation vector are subjected to linear transformation and nonlinear mapping respectively to obtain the main device energy consumption dynamic evaluation value and the auxiliary device energy consumption dynamic evaluation value, the two values are converged and averaged to form the device energy consumption dynamic evaluation value, and the stable, comprehensive and dynamic evaluation of the device energy consumption state is realized.

[0023] S1, acquiring multi-source data affecting device energy consumption, constructing an energy consumption evaluation data set.

[0024] Please refer to Figure 1As shown, the device energy consumption dynamic evaluation method based on multi-source data fusion in the embodiment of the application has the following specific steps.

[0025] In this embodiment, the construction process of the device energy consumption evaluation dataset includes two stages of data acquisition and preprocessing. In the data acquisition stage, the collected multi-source data includes device input voltage, device input current, device instantaneous active power, cumulative electric energy consumption, device operating speed, device load rate, device surface temperature, and environmental temperature. Among them, the device input voltage, device input current, device instantaneous active power, and cumulative electric energy consumption are directly collected by the electric energy metering unit, the device operating speed and device load rate are read in real time by the device control system, the device surface temperature is collected by the temperature sensor fixedly installed at the key structure position of the device, and the environmental temperature is collected by the environmental monitoring unit in the device operating area. UTC time stamp is uniformly attached to each type of data record, and each collection unit is time-synchronized based on a unified clock source. In the data preprocessing stage, the validity of the collected multi-source data is verified channel by channel, and the missing records, continuous repeated records, and data exceeding the rated operating parameter range of the device are directly excluded. The local missing data not exceeding 3 consecutive sampling points is compensated by the previous valid sampling value to obtain time-continuous and physically reasonable multi-source energy consumption time series data. Then, based on the device's nearly 30-day historical stable operation data, the upper and lower threshold values of the multi-source data are determined, and the device input voltage, device input current, device instantaneous active power, cumulative electric energy consumption, device operating speed, device load rate, and device surface temperature are respectively mapped to the [0, 1] interval by using the maximum and minimum normalization method. The average value of the environmental temperature data in a continuous 300-second time period is taken as the reference for centralization processing, so that it is represented as a temperature offset relative to the local environmental reference, thereby reducing the influence of overall environmental changes on energy consumption dynamic evaluation. After scaling, all data are uniformly resampled to 1-second time steps and time-aligned. Finally, the time-aligned multi-source data is segmented by sliding with a fixed time window length of 120 seconds, and the corresponding 8 types of data in each time window are combined and packaged in a fixed order as an energy consumption time series. All energy consumption time series are arranged in chronological order to form a device energy consumption evaluation dataset, which serves as input data for subsequent device energy consumption dynamic evaluation models.

[0026] S2, the missing rate, jump rate, and flat rate of the energy consumption time series are calculated, and the linear transformation and ReLU, Sigmoid activation functions are used to generate a dimension-by-dimension credibility gate. Meanwhile, the deviation of the multi-source data is recursively modeled with reference to the median vector to obtain a dynamic fluctuation intensity.

[0027] Further, in the S2 step, the dimension-by-dimension credibility gate and the dynamic fluctuation intensity are generated, and the process is as shown in Figure 2 The specific steps are as follows.

[0028] Statistical analysis was performed on the data integrity and variation characteristics of energy consumption time series in various dimensions to obtain the missing rate, jump rate and flatness value of each dimension. The missing rate, jump rate and flatness value were combined according to the dimension index order to form the missing rate, jump rate and flatness value respectively, and then spliced ​​in a predetermined order to construct a multi-source acquisition quality statistical vector. In this embodiment, the energy consumption time series from the energy consumption assessment dataset is input. , ,in For the first Multi-source data affecting device energy consumption collected at time steps The length of the time series. Using data dimensions, statistical analysis is performed on the data integrity and variation characteristics of energy consumption time series across various dimensions. This is achieved by analyzing the data at each time step of the energy consumption time series. The missing points in each dimension are summarized to form a missing rate value; the missing values ​​in adjacent time steps are then analyzed. The data dimensionality changes exceeding the jump threshold are statistically analyzed to obtain the corresponding jump rate value; and the jump rate value is calculated for the 1st jump in adjacent time steps. Cases where the variation in dimensional data is less than the flatness threshold are statistically analyzed to obtain the corresponding flatness value. After completing the statistical analysis of the missing rate, jump rate, and flatness value for all dimensions, they are combined according to the dimensional order to form the missing rate, jump rate, and flatness value. The missing rate, jump rate, and flatness value are then concatenated to obtain the multi-source acquisition quality statistical vector. The specific mathematical model is as follows: ; ; ; in, , , These are the first time steps in the energy consumption time series. The missing rate, jump rate, and flatness values ​​of each dimension are combined in order to obtain the missing rate of the energy consumption time series. Jump rate Flatness The missing rate, jump rate, and flatness are concatenated to obtain the multi-source acquisition quality statistics vector. , , To count the number of elements in a set, For the absolute value operation, For the jump threshold, The flat threshold, , They represent the first , time step multi-source data dimensional value, for all time steps satisfying the first dimensional value changes more than a jump threshold , for all time steps satisfying the first dimensional value changes less than a flat threshold ; In this embodiment, the time series length , the data dimension , the number of types of multi-source data collected determines the first dimensional energy consumption time series first-order difference standard deviation , wherein is a standard deviation function, the jump threshold , which can effectively distinguish between normal fluctuations and abnormal jumps, the flat threshold , to exclude quantization noise and retain the true no-change state; Apply a linear transformation and a ReLU activation function to the multi-source collection quality statistics vector to obtain an intermediate feature representation, and then input the intermediate feature representation into a second layer linear transformation and a Sigmoid activation function to obtain a per-dimension credibility gate. In this embodiment, a linear transformation is applied to the multi-source collection quality statistics vector through a weight matrix and a bias term, and a non-linear mapping is performed through a ReLU activation function to obtain an intermediate feature representation. Then the intermediate feature representation is input into a second layer linear transformation, and finally a per-dimension credibility gate is output through a Sigmoid activation function non-linear mapping. The specific mathematical model is: ; wherein, is the per-dimension credibility gate, is the first layer learnable weight matrix, is the first layer learnable bias term, is the second layer learnable weight matrix, is the second layer learnable bias term, is the dimensional value; In this embodiment, the dimensional value , to cover the interaction between the missing rate, jump rate and flat rate, while avoiding overfitting caused by too large a dimension; ​​The multi-source data and median vector at the current time step are differentially processed, and the differential result is squared to form the deviation term at the current time step. Then, a learnable smoothing coefficient is introduced, and the deviation term and the dynamic fluctuation intensity of the previous time step are recursively weighted according to a preset ratio to construct the dynamic fluctuation intensity corresponding to the current time step. In this embodiment, based on the multi-source data of the energy consumption time series at the current time step, the median vector of the energy consumption time series in each dimension is introduced. The deviation between the multi-source data at the current time step and the median vector is measured in a squared form. A learnable smoothing coefficient is used to weight and combine the deviation with the dynamic fluctuation intensity of the previous time step to form the dynamic fluctuation intensity of the current time step. The specific mathematical model is as follows: ; in, , The first , The intensity of dynamic fluctuations at the time step. Let be the median vector of the energy consumption time series in each dimension. The smoothing coefficient is a learnable factor. In this embodiment, the smoothing coefficient is initialized. To maintain consistency with the equipment's characteristics of high energy consumption inertia and few sudden changes, noise is smoothed, and true abnormal trends are preserved during initialization. It is a zero vector.

[0029] S3. Difference the multi-source data with the median vector, and perform dimension-wise scaling of the difference results by combining the robust scaling vector and dynamic fluctuation intensity. Perform element-wise combination operations with dimension-wise confidence gating to obtain the normalized output, which is input into the gating loop unit to form the temporal fusion state vector. After linear transformation and Softmax activation function, the soft partition probability is obtained.

[0030] Furthermore, in step S3, soft partition probabilities are generated, the process of which is as follows: Figure 3 As shown, the specific steps are as follows.

[0031] The first The multi-source data at each time step are differentially processed with the median vector to construct the deviation result; subsequently, a robust scaling vector based on quantile statistics is introduced, and the squared term of the robust scaling vector is compared with the first... The dynamic fluctuation intensity of the time step is summed, and the summation result is squared to form a dynamic normalized scale. Then, the deviation result is scaled dimension by dimension using the dynamic normalized scale to obtain an intermediate result, which is then combined with the dimension-wise confidence gating element by element to form a normalized output. In this embodiment, for the first The multi-source data of the time step is differentially processed with the median vector to obtain a deviation result; a robust scale vector is introduced, and a square term of the robust scale vector is added to the The dynamic fluctuation intensity of the time step is added and square-rooted to form a dynamic normalization scale; the deviation result is scaled dimension by dimension using the dynamic normalization scale to obtain an intermediate result, and finally, the intermediate result is multiplied element by element with a dimension-by-dimension credibility gate to obtain a normalized output, and the specific mathematical model is: ; Wherein, is the normalized output of the t time step, is the robust scale vector, , wherein , are the 75% and 25% median vectors of the energy consumption time series in each dimension, respectively, is the quartile range value of the standard normal distribution, is a learnable energy compensation coefficient, is element-wise multiplication, is a very small positive number; In this embodiment, the quartile range value of the standard normal distribution is In this embodiment is taken , the energy compensation coefficient is initialized to make the model first rely on the long-term steady-state scale and then gradually introduce the influence of dynamic fluctuations, and the very small positive number is to prevent the denominator from being 0; The normalized output is input into a gated recurrent unit to obtain a time series fusion state vector; In this embodiment, the normalized output and the time series fusion state vector of the previous time step are jointly input into the gated recurrent unit, and the historical state information and the current time step information are selectively retained and updated through the gating structure to form the time series fusion state vector corresponding to the current time step, and the specific mathematical model is: ; Wherein, is the time series fusion state vector of the t time step, is the gated recurrent unit; The time series fusion state vector is linearly transformed and nonlinearly mapped using a Softmax activation function to obtain a soft partition probability; In this embodiment, the time series fusion state vector is linearly transformed using a mapping weight matrix and a mapping bias term and nonlinearly mapped using a Softmax activation function to obtain a soft partition probability, and the specific mathematical model is: ; wherein, is the soft partition probability of the time step, is the learnable mapping weight matrix, is the learnable mapping bias term.

[0032] S4, the absolute value of the normalized output is dimensionally associated with the mapping weight and averaged, and the fusion representation vector is formed by weighted summarization through the soft partition probability, and the sequence is composed in time sequence to obtain the time sequence convolution feature through one-dimensional convolution along the time dimension. After summarization and averaging in the whole time range, the global feature representation is constructed.

[0033] Further, in the S4 step, the global feature representation is generated, and the flow is as shown in Figure 4 , and the specific steps are as follows.

[0034] The absolute value of the normalized output of each dimension of the time step is processed, and the associated calculation is performed with the mapping weight in the corresponding partition state. The fusion representation vector is obtained by weighted processing and summarization through the soft partition probability. In this embodiment, after the absolute value of the normalized output of each dimension of the time step is taken, the mapping weight in the corresponding partition state is multiplied to obtain the intermediate representation under each partition state; then, the soft partition probability corresponding to the time step is used to weight modulate and summarize the intermediate representation under each partition state to obtain the fusion representation vector, and the specific mathematical model is as follows: ; wherein, is the fusion representation vector of the time step, is the soft partition probability value of the dimension of the time step, is the mapping weight of the normalized output value in the dimension partition state; The fusion representation vectors of each time step are arranged in time sequence to construct a fusion representation vector sequence, and one-dimensional convolution operation is performed along the time dimension to form a time sequence convolution feature, is the element-wise multiplication; In the embodiment, the fusion representation vectors of each time step are sequentially composed into a fusion representation vector sequence, a one-dimensional convolution operation is performed on the fusion representation vector sequence along the time dimension, local correlation modeling of the fusion representation vectors of adjacent time steps is performed within a time window, different time sequence change patterns are extracted in parallel through multiple groups of convolution kernels, and a time sequence convolution feature is obtained. The specific mathematical model is: ; wherein, is the time sequence convolution feature, is a one-dimensional convolution operation performed on the fusion representation vector of the i-th time step along the time dimension, the convolution kernel size is , the number of convolution kernels is , , is a fusion representation vector sequence obtained by combining the fusion representation vectors of all time steps, ; In the embodiment, the convolution kernel size is used to characterize the change pattern of the energy consumption state in a short time scale, and the number of convolution kernels is , which is beneficial to the subsequent understanding of each dimension corresponding to an energy consumption sub-pattern by the model while ensuring the expression ability; The time sequence convolution features corresponding to each time step are sequentially summarized in the time dimension, and an overall average operation is performed within the range of all time steps to construct a global feature representation; In the embodiment, the time sequence convolution features corresponding to each time step are accumulated, and an overall average operation is performed within the range of all time steps to form a global feature representation at the time sequence level. The specific mathematical model is: ; wherein, is the global feature representation at the time sequence level.

[0035] S5, respectively, linearly transforming the global feature representation and the fusion representation vector and obtaining the main and auxiliary device energy consumption dynamic evaluation values through nonlinear mapping, converging and averaging the two to form the final device energy consumption dynamic evaluation value.

[0036] Further, in the S5 step, the device energy consumption dynamic evaluation value is calculated, and the specific steps are as follows.

[0037] The global feature representation is linearly transformed, and the ReLU activation function is used for nonlinear mapping to obtain the main device energy consumption dynamic evaluation value; In the embodiment, the global feature representation is linearly transformed, and the ReLU activation function is used for nonlinear mapping to output the main device energy consumption dynamic evaluation value. The specific mathematical model is: ; wherein, is a main device energy consumption dynamic evaluation value, is a learnable weight matrix, is a learnable bias term; performing a straightening splicing operation on the fusion representation vector sequence to construct an overall fusion representation vector, performing linear transformation, and performing non-linear mapping using a ReLU activation function to form a secondary device energy consumption dynamic evaluation value; In this embodiment, the overall fusion representation vector formed by performing a straightening splicing operation on the fusion representation vector sequence in time sequence, performing linear transformation on the overall fusion representation vector, and performing non-linear mapping using a ReLU activation function to output a secondary device energy consumption dynamic evaluation value, the specific mathematical model is: ; wherein, is a secondary device energy consumption dynamic evaluation value, is a learnable weight matrix, is a learnable bias term, is an overall fusion representation vector obtained by performing a straightening splicing operation on the fusion representation vector sequence; performing a convergence operation on the main device energy consumption dynamic evaluation value and the secondary device energy consumption dynamic evaluation value, and forming a final device energy consumption dynamic evaluation value by averaging; In this embodiment, the main device energy consumption dynamic evaluation value and the secondary device energy consumption dynamic evaluation value are uniformly converged, and the final device energy consumption dynamic evaluation value is formed by averaging, and the specific mathematical model is: ; wherein, is a final device energy consumption dynamic evaluation value.

[0038] Further, in the S6 step, the device energy consumption dynamic evaluation based on multi-source data fusion proposed by the present application is implemented by using Python programming language, and the model construction and training are completed based on the PyTorch deep learning framework. The model input is the energy consumption time series in the energy consumption evaluation data set, and the dimension is 120x8. In the model training process, the mean square error loss function is selected as the model optimization target, the Adam optimization algorithm is used to iteratively update each learnable parameter, the initial learning rate is set to 0.001, the batch size is set to 16, and the maximum training number is set to 120 rounds. The model gradually learns the change rule of multi-source energy consumption data, and finally obtains a stable model for device energy consumption dynamic evaluation.

[0039] Further, in the S6 step, the collected energy consumption evaluation data set is input into the constructed model for training and evaluation, and the model training loss curve is as follows:Figure 5 As shown in the figure, the abscissa is the training round, and the ordinate is the mean square error loss value. It can be observed from the figure that the loss value decreases rapidly in the early stage and gradually stabilizes in the later stage, indicating that the model realizes stable convergence in the iteration process, and the device energy consumption dynamic evaluation effect diagram is as shown in Figure 6 As shown in the figure, the abscissa is time, and the ordinate is the device energy consumption evaluation value (kilo Watt), it can be observed from the figure that the model prediction value is close to the actual change trend as a whole, and there is a slight deviation in part of the data, indicating that the model has good generalization ability and energy consumption evaluation ability.

[0040] The above is only the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the inventive concept, can make a number of deformation and improvement, these all belong to the protection scope of the present application.

Claims

1. A device energy consumption dynamic evaluation method based on multi-source data fusion, characterized in that, The method comprises the following steps: S1, acquiring multi-source data affecting device energy consumption, and constructing an energy consumption evaluation dataset; S2, calculating the missing rate, jump rate and flat rate of the energy consumption time series, and generating a dimension-by-dimension credibility gate by using linear transformation and ReLU and Sigmoid activation functions after splicing by dimensions, and recursively modeling the deviation degree of the multi-source data from the median vector to obtain dynamic fluctuation intensity; S3, differentiating the multi-source data from the median vector, combining the robust scale vector with the dynamic fluctuation intensity to scale the differentiated results dimension by dimension, performing element-by-element combination operation on the dimension-by-dimension credibility gate to obtain normalized output, inputting the gate recurrent unit to form a time series fusion state vector, and obtaining soft partition probability after linear transformation and Softmax activation function; S4, associating and averaging the absolute value of the normalized output with the mapping weight by dimensions, forming a fusion representation vector by weighted summation through the soft partition probability, and forming a sequence in time sequence to obtain time series convolution features in the time dimension, and constructing a global feature representation after summarizing and averaging in the full time range; S5, linearly transforming the global feature representation and the fusion representation vector respectively and obtaining the main and auxiliary device energy consumption dynamic evaluation values through nonlinear mapping, merging and averaging the two to form the final device energy consumption dynamic evaluation value; S6, constructing a device energy consumption dynamic evaluation model, inputting the energy consumption evaluation dataset, and sequentially performing steps S2 to S5 for iterative training until convergence.

2. The device energy consumption dynamic evaluation method based on multi-source data fusion according to claim 1, characterized in that, The construction process of the device energy consumption evaluation dataset includes two stages of data acquisition and preprocessing. In the data acquisition stage, the multi-source data collected includes device input voltage, device input current, device instantaneous active power, cumulative power consumption, device operating speed, device load rate, device surface temperature and environmental temperature. In the data preprocessing stage, the validity of the multi-source data collected is verified channel by channel, and the missing records, continuous repeated records and data exceeding the rated operating parameter range of the device are directly excluded, and the local missing data is compensated by the previous valid sampling value to obtain multi-source energy consumption time series data. Then the upper and lower threshold values of the multi-source data are determined, and the device input voltage, device input current, device instantaneous active power, cumulative power consumption, device operating speed, device load rate and device surface temperature are normalized respectively, and the environmental temperature data is centered by taking the average value in the continuous time period as the reference; After the scale is unified, all the data are uniformly resampled and time-aligned; finally, the time-aligned multi-source data are segmented by sliding with a fixed time window length, and the corresponding 8 types of data in each time window are combined and packaged in a fixed order as an energy consumption time series, and all the energy consumption time series are arranged in chronological order to form the device energy consumption evaluation dataset as the input data of the subsequent device energy consumption dynamic evaluation model. 3.The method of claim 2, wherein, Statistical analysis is performed on the data integrity and change characteristics of the energy consumption time series in each dimension, and the missing rate value, jump rate value and flat rate value of each dimension are obtained. The missing rate, jump rate and flat rate are combined according to the dimension index order respectively, and are spliced in a predetermined order to construct a multi-source acquisition quality statistical vector; A linear transformation and a ReLU activation function are applied to the multi-source acquisition quality statistical vector to obtain an intermediate feature representation. Then, the intermediate feature representation is input into a second layer linear transformation and a Sigmoid activation function to obtain a dimension-by-dimension credibility gate.

4. The method of claim 3, wherein, The multi-source data of the current time step and the median vector are subjected to difference processing, and the difference result is subjected to square operation to form the deviation term of the current time step. Then, a learnable smoothing coefficient is introduced to recursively weight the deviation term and the dynamic fluctuation intensity of the previous time step according to a predetermined proportion to construct the dynamic fluctuation intensity corresponding to the current time step.

5. The method of claim 4, wherein, The first The multi-source data of the time step is subjected to difference processing with the median vector to construct a deviation result. Then, a robust scale vector based on the quantile statistics is introduced, and the square term of the robust scale vector is added to the first The dynamic fluctuation intensity of the time step is added, and the square root of the added result is taken to form a dynamic normalization scale. Then, the dynamic normalization scale is used to scale the deviation result dimension by dimension to obtain an intermediate result, which is combined with the dimension-by-dimension credibility gate element by element to construct a normalized output.

6. The method of claim 5, wherein, The normalized output is input into the gating recurrent unit to obtain a time series fusion state vector. The time series fusion state vector is subjected to linear transformation and non-linear mapping using a Softmax activation function to obtain a soft partition probability.

7. The method of claim 6, wherein, To the first The normalized output of each dimension at the time step is subjected to absolute value processing, and is associated with the mapping weight in the corresponding partition state to perform weighted processing through the soft partition probability and is summarized to construct a fusion representation vector.

8. The device energy consumption dynamic evaluation method based on multi-source data fusion according to claim 7, characterized in that, The fusion representation vectors of each time step are arranged in time sequence to construct a fusion representation vector sequence, and one-dimensional convolution operation is performed along the time dimension to form time series convolution features. The time series convolution features corresponding to each time step are sequentially summarized in the time dimension, and overall average operation is performed within the range of all time steps to construct a global feature representation.

9. The method of claim 8, wherein, The global feature representation is subjected to linear transformation and non-linear mapping using a ReLU activation function to obtain a main device energy consumption dynamic evaluation value. The fusion representation vector sequence is subjected to straightening and splicing operation to construct an overall fusion representation vector, which is subjected to linear transformation and non-linear mapping using a ReLU activation function to form an auxiliary device energy consumption dynamic evaluation value. The main device energy consumption dynamic evaluation value and the auxiliary device energy consumption dynamic evaluation value are subjected to aggregation operation, and the final device energy consumption dynamic evaluation value is formed by taking the average.

Images