Energy consumption data up-sampling method, system, device and medium based on transfer learning

Abstract

The application provides an energy consumption data up-sampling method, system, device and medium based on transfer learning, and relates to the technical field of artificial intelligence. The method comprises the following steps: reconstructing historical energy consumption data of multiple energy consumption types in a preset period into a multi-dimensional energy consumption feature tensor according to a preset period length; performing weighted pooling on energy consumption influence data associated with multiple energy consumption influence factors in the preset period based on preset correlation weights of the energy consumption influence factors for each energy consumption type, so as to select and aggregate information of each energy consumption influence factor and obtain a core feature vector; splicing the core feature vector and an energy consumption feature vector obtained by performing feature extraction on the multi-dimensional energy consumption feature tensor to obtain a fusion feature vector; and inputting the fusion feature vector into a linear mapping layer for mapping to output target energy consumption data, wherein the linear mapping layer is obtained by transfer learning training.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method, system, device, and medium for upsampling energy consumption data based on transfer learning. Background Technology

[0002] With the increasing demands for refined industrial carbon emission monitoring and energy consumption management, high temporal resolution energy consumption data has become a key foundation for carbon accounting and energy efficiency optimization. In related technologies, deep learning methods are typically used to upsample energy consumption data to meet the requirements for data refinement.

[0003] In the process of realizing the concept of this invention, it was found that at least the following problems exist in the related technologies: the deep learning method is insufficient in feature mining of energy consumption data, resulting in large errors in the upsampling results, which makes it difficult to meet the actual needs of high-precision carbon emission monitoring. Summary of the Invention

[0004] In view of this, the present invention provides a method, system, device and medium for upsampling energy consumption data based on transfer learning.

[0005] One aspect of the present invention provides an energy consumption data upsampling method based on transfer learning, comprising: reconstructing historical energy consumption data of multiple energy consumption types within a preset period into a multidimensional energy consumption feature tensor according to a preset period length, wherein the multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided according to the preset period length; performing weighted pooling on energy consumption impact data associated with multiple energy consumption impact factors within the preset period based on preset correlation weights for each energy consumption type of multiple energy consumption impact factors, so as to select and aggregate the information of each energy consumption impact factor to obtain a core feature vector; concatenating the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector; inputting the fused feature vector into a linear mapping layer for mapping, and outputting target energy consumption data, wherein the linear mapping layer is obtained by transfer learning training, and the transfer learning training includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

[0006] According to an embodiment of the present invention, historical energy consumption data of various energy consumption types within a preset period are reconstructed into a multidimensional energy consumption feature tensor based on a preset period length. This includes: for each energy consumption type, dividing the historical energy consumption data into multiple consecutive period segments in time according to the preset period length, and arranging the data in each period segment into row vectors in chronological order; stacking the multiple row vectors to form a two-dimensional matrix according to the chronological order between the multiple period segments, wherein the row direction of the two-dimensional matrix represents each moment in a single period, and the column direction represents the corresponding moment in different periods; and stacking the two-dimensional matrices corresponding to different energy consumption types along the third dimension to form a multidimensional energy consumption feature tensor.

[0007] According to an embodiment of the present invention, based on preset correlation weights for multiple energy consumption influencing factors for each energy consumption type, weighted pooling is performed on energy consumption influencing data associated with multiple energy consumption influencing factors within a preset time period to select and aggregate information of each energy consumption influencing factor to obtain a core feature vector. This includes: organizing energy consumption influencing data associated with multiple energy consumption influencing factors within a preset time period into an initial influencing tensor; mapping the initial influencing tensor from its original time length to a preset core length using a first multilayer perceptron in the time dimension to obtain a compressed influencing tensor; sampling the energy consumption influencing factors at each time step in the compressed influencing tensor for each energy consumption type according to preset correlation weights to obtain a sampling index for each time step; extracting information of the corresponding energy consumption influencing factors from the compressed influencing tensor according to the sampling index as the core feature representation for each time step; and combining the core feature representations of each time step in chronological order to form a core feature vector.

[0008] According to an embodiment of the present invention, the preset correlation weights are determined as follows: Reference energy consumption data for multiple energy consumption types within a reference period, and reference impact data associated with multiple energy consumption impact factors within the reference period; for each energy consumption type, based on the joint change trend and fluctuation amplitude of the reference energy consumption data and the reference impact data, the linear correlation coefficient between each energy consumption impact factor and the reference energy consumption data of the energy consumption type is determined; the linear correlation coefficients of each energy consumption impact factor under the same energy consumption type are normalized to obtain the preset correlation weights of multiple energy consumption impact factors for each energy consumption type.

[0009] According to an embodiment of the present invention, the method further includes: extracting features from the multidimensional energy consumption feature tensor using hierarchical initial convolution to obtain periodic feature sequences corresponding to each energy consumption type; splicing and truncating each periodic feature sequence in chronological order to obtain an energy consumption feature vector corresponding to each energy consumption type, wherein the time length of the energy consumption feature vector is the same as that of the historical energy consumption data.

[0010] According to an embodiment of the present invention, concatenating the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector includes: concatenating the core feature vector with the energy consumption feature vector corresponding to each energy consumption type in the time dimension to obtain a combined feature sequence corresponding to each energy consumption type; mapping each combined feature sequence to a fused feature sequence corresponding to each energy consumption type through a second multilayer perceptron, so as to use multiple fused feature sequences as a fused feature vector, wherein the time length of the fused feature vector is the same as that of the historical energy consumption data.

[0011] According to an embodiment of the present invention, the transfer learning training further includes: using sample energy consumption data of multiple energy consumption types corresponding to the region to which the historical energy consumption data belongs as a supervision signal, iteratively adjusting the weights and biases of the initial mapping layer until the difference between the total energy consumption value of the predicted energy consumption data output by the linear mapping layer accumulated on the time axis and the total energy consumption value in the sample energy consumption data satisfies a preset convergence condition; in response to the difference satisfying the preset convergence condition, using the adjusted initial mapping layer as the linear mapping layer; wherein, the initial mapping layer is trained by reference energy consumption data of multiple energy consumption types in the reference region during the historical period and reference influence data associated with multiple energy consumption influence factors during the historical period.

[0012] Another aspect of the present invention provides an energy consumption data upsampling system based on transfer learning, comprising: a data reconstruction module, configured to reconstruct historical energy consumption data of multiple energy consumption types within a preset period into a multidimensional energy consumption feature tensor according to a preset period length, wherein the multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided based on the preset period length; a core extraction module, configured to perform weighted pooling on energy consumption impact data associated with multiple energy consumption impact factors within a preset period based on preset correlation weights for each energy consumption type, so as to select and aggregate the information of each energy consumption impact factor to obtain a core feature vector; a vector concatenation module, configured to concatenate the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector; and a mapping output module, configured to input the fused feature vector into a linear mapping layer for mapping and output target energy consumption data, wherein the linear mapping layer is obtained by transfer learning training, and the transfer learning training includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

[0013] Another aspect of the present invention provides an electronic device comprising:

[0014] One or more processors;

[0015] Memory, used to store one or more programs.

[0016] Specifically, when one or more programs are executed by one or more processors, the one or more processors implement the above method.

[0017] Another aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0018] Another aspect of the present invention provides a computer program product including computer-executable instructions that, when executed, are used to implement the methods described above.

[0019] According to an embodiment of the present invention, historical energy consumption data is reconstructed into a multidimensional energy consumption feature tensor that can simultaneously characterize temporal changes within the same period and correlation changes between different periods by a preset period length. This explicitly encodes the intra-period correlation and inter-period correlation in the energy consumption sequence, allowing subsequent feature extraction to capture the cyclical evolution of energy consumption from two dimensions. Based on this, preset correlation weights are introduced between energy consumption influencing factors and each energy consumption type, and the influencing factor data is weighted and pooled. This mechanism dynamically selects the information of each energy consumption influencing factor based on its correlation and aggregates it into a core feature vector. The core feature vector is then concatenated and fused with the energy consumption feature vector extracted from the multidimensional energy consumption feature tensor. This allows the model to retain the cyclical patterns of energy consumption itself while adaptively incorporating the differentiated effects of external influencing factors. Finally, the target energy consumption data is output through linear mapping, achieving the fusion of the multi-period intrinsic patterns of historical energy consumption data and the dynamic interaction of external influencing factors, significantly improving the accuracy and robustness of upsampling historical energy consumption data. Attached Figure Description

[0020] The above and other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the invention with reference to the accompanying drawings.

[0021] Figure 1 The present invention illustrates an application scenario of an energy consumption data upsampling method, system, device, and storage medium based on transfer learning according to embodiments of the present invention.

[0022] Figure 2 A flowchart of an energy consumption data upsampling method based on transfer learning according to an embodiment of the present invention is shown.

[0023] Figure 3 A schematic diagram of a multidimensional energy consumption feature tensor according to an embodiment of the present invention is shown.

[0024] Figure 4 A schematic diagram of energy consumption upsampling data processing according to an embodiment of the present invention is shown.

[0025] Figure 5AA heatmap of cosine similarity analysis results according to an embodiment of the present invention is shown.

[0026] Figure 5B A heatmap of correlation analysis results according to an embodiment of the present invention is shown.

[0027] Figure 6 A flowchart of an energy consumption data upsampling method based on transfer learning according to another embodiment of the present invention is shown.

[0028] Figure 7A An error graph of coal consumption upsampling according to an embodiment of the present invention is shown.

[0029] Figure 7B An error graph of oil consumption upsampling according to an embodiment of the present invention is shown.

[0030] Figure 7C An error graph of natural gas consumption upsampling according to an embodiment of the present invention is shown.

[0031] Figure 7D Radar graphs showing different algorithmic errors according to embodiments of the present invention are illustrated.

[0032] Figure 7E A radar chart illustrating the algorithm error under extreme conditions according to an embodiment of the present invention is shown.

[0033] Figure 8A A graph showing different input errors for coal consumption according to an embodiment of the present invention is shown.

[0034] Figure 8B A graph showing different input errors for oil consumption according to an embodiment of the present invention is shown.

[0035] Figure 8C A graph showing different input errors for natural gas consumption according to an embodiment of the present invention is shown.

[0036] Figure 8D Radar plots of different input scale errors according to embodiments of the present invention are shown.

[0037] Figure 9 A block diagram of an energy consumption data upsampling system based on transfer learning according to an embodiment of the present invention is shown.

[0038] Figure 10 A block diagram of an electronic device suitable for implementing a transfer learning-based energy consumption data upsampling method according to an embodiment of the present invention is shown. Detailed Implementation

[0039] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0040] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0041] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0042] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0043] In the embodiments of this invention, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.

[0044] In the embodiments of the present invention, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0045] An embodiment of the present invention provides an energy consumption data upsampling method based on transfer learning, comprising: reconstructing historical energy consumption data of multiple energy consumption types within a preset period into a multidimensional energy consumption feature tensor according to a preset period length, wherein the multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided according to the preset period length; performing weighted pooling on energy consumption impact data associated with multiple energy consumption impact factors within the preset period based on preset correlation weights for each energy consumption type of multiple energy consumption impact factors, so as to select and aggregate the information of each energy consumption impact factor to obtain a core feature vector; concatenating the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector; inputting the fused feature vector into a linear mapping layer for mapping, and outputting target energy consumption data, wherein the linear mapping layer is obtained by transfer learning training, and the transfer learning training includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

[0046] Figure 1 This paper illustrates an application scenario of an energy consumption data upsampling method, system, device, and storage medium based on transfer learning according to embodiments of the present invention. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to embodiments of the present invention, in order to help those skilled in the art understand the technical content of the present invention, but do not mean that embodiments of the present invention cannot be used in other devices, systems, environments or scenarios.

[0047] like Figure 1 As shown, the system architecture 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0048] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social media platform software, etc. (for example only).

[0049] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0050] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0051] It should be noted that the energy consumption data upsampling method based on transfer learning provided in this embodiment of the invention can generally be executed by server 105. Correspondingly, the energy consumption data upsampling system based on transfer learning provided in this embodiment of the invention can generally be located in server 105. The energy consumption data upsampling method based on transfer learning provided in this embodiment of the invention can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the energy consumption data upsampling system based on transfer learning provided in this embodiment of the invention can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Alternatively, the energy consumption data upsampling method based on transfer learning provided in this embodiment of the invention can also be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103, or it can be executed by other terminal devices different from the first terminal device 101, the second terminal device 102, or the third terminal device 103. Accordingly, the energy consumption data upsampling system based on transfer learning provided in this embodiment of the invention can also be set in the first terminal device 101, the second terminal device 102 or the third terminal device 103, or in other terminal devices different from the first terminal device 101, the second terminal device 102 or the third terminal device 103.

[0052] It should be understood that Figure 1 The number of first terminal devices, second terminal devices, third terminal devices, networks, and servers shown in the diagram is merely illustrative. Depending on implementation needs, any number of first terminal devices, second terminal devices, third terminal devices, networks, and servers can be included.

[0053] Figure 2A flowchart of an energy consumption data upsampling method based on transfer learning according to an embodiment of the present invention is shown.

[0054] like Figure 2 As shown, the method includes operations S210 to S240.

[0055] In operation S210, based on the preset period length, the historical energy consumption data of various energy consumption types within the preset time period are reconstructed into a multidimensional energy consumption feature tensor. The multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided based on the preset period length.

[0056] In operation S220, based on the preset correlation weights of multiple energy consumption influencing factors for each energy consumption type, the energy consumption impact data associated with multiple energy consumption influencing factors within a preset time period are weighted and pooled to select and aggregate the information of each energy consumption influencing factor to obtain the core feature vector.

[0057] In operation S230, the core feature vector is concatenated with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain the fused feature vector.

[0058] In operation S240, the fused feature vector is input into the linear mapping layer for mapping, and the target energy consumption data is output. The linear mapping layer is obtained by transfer learning. The transfer learning training includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

[0059] According to an embodiment of the present invention, historical energy consumption data of multiple energy consumption types (e.g., coal, oil, and natural gas) within a preset time period are obtained. This historical energy consumption data is typically a statistical value with low time resolution (e.g., annual total). To explore the inherent multi-period characteristics of the historical energy consumption data, during the model training phase, a fast Fourier transform can be used to perform frequency domain analysis on the reference energy consumption data of multiple energy consumption types within the reference time period, calculate the amplitude corresponding to each frequency, and select the period length corresponding to the top few frequencies with the largest amplitudes as the preset period length.

[0060] For each type of energy consumption, the time series is divided into multiple consecutive periodic segments according to the preset period length. The energy consumption data in each periodic segment is arranged into a row vector in chronological order. Then, the row vectors of all periodic segments are stacked in chronological order to form a two-dimensional matrix. The row direction represents each moment in a single period (e.g., each month in a year), and the column direction represents the corresponding moments in different periods (e.g., the same month in different years).

[0061] By stacking the two-dimensional matrices corresponding to different energy consumption types along the third dimension, a multidimensional energy consumption feature tensor is formed. This multidimensional energy consumption feature tensor can simultaneously characterize the time-varying pattern within the same period (time-series changes within the period) and the correlation of the same phase position between different periods (corresponding changes between periods).

[0062] Multiple energy consumption influencing factors are introduced, such as temperature, energy price, and electricity consumption. For each energy consumption type, the Pearson correlation coefficient between each energy consumption influencing factor and the reference energy consumption data for that type is pre-calculated, and then normalized as a preset correlation weight. The preset correlation weight reflects the importance of different energy consumption influencing factors to actual energy consumption.

[0063] A weighted pooling operation is performed on the energy consumption impact data (which also has a time series structure) associated with these energy consumption impact factors within a preset time period. Specifically, the energy consumption impact data can be organized into an initial impact tensor (dimensions of time × energy consumption impact factor × energy consumption type), and then compressed in the time dimension by a first multilayer perceptron, mapping the original time length to a smaller preset core length to obtain a compressed impact tensor.

[0064] At each compressed time step, based on the probability distribution determined by the preset relevance weights, information on an energy consumption influencing factor is randomly extracted as the output value for that time step (factors with larger weights have a higher probability of being extracted), thereby achieving dynamic selection and aggregation of information on different energy consumption influencing factors. Finally, the extraction results from all time steps are combined in chronological order to obtain the core feature vector.

[0065] To extract useful periodic patterns from the multidimensional energy consumption feature tensor, a hierarchical initial convolution is used to perform multi-scale feature extraction on the multidimensional energy consumption feature tensor, obtaining periodic feature sequences corresponding to each energy consumption type. These periodic feature sequences are concatenated in chronological order and truncated to the same time length as the original historical energy consumption data, thereby obtaining the energy consumption feature vectors for each energy consumption type.

[0066] The previously obtained core feature vectors are fused with these energy consumption feature vectors, concatenating them along the time dimension to obtain a fused feature vector containing fused feature sequences of different energy consumption types. This fused feature vector is then input into a linear mapping layer for linear transformation, outputting the target energy consumption data.

[0067] The linear mapping layer is obtained through fine-tuning pre-training. Specifically, the initial mapping layer can be trained using fine-grained energy consumption data and corresponding influencing factor data from a reference region (such as a developed region with complete data). Then, for the target region of the actual application (i.e., the region to which historical energy consumption data belongs), known sample energy consumption data (such as annual statistics) of the target region are used as supervision signals to iteratively adjust the weights and biases of the initial mapping layer, thereby obtaining a linear mapping layer adapted to the target region.

[0068] Through the above processing, the final output target energy consumption data has a higher time resolution than the original input historical energy consumption data (e.g., refining annual data into monthly data), thus providing more refined data support for carbon emission monitoring and energy consumption prediction.

[0069] According to an embodiment of the present invention, historical energy consumption data is reconstructed into a multidimensional energy consumption feature tensor that can simultaneously characterize temporal changes within the same period and correlation changes between different periods by a preset period length. This explicitly encodes the intra-period correlation and inter-period correlation in the energy consumption sequence, allowing subsequent feature extraction to capture the cyclical evolution of energy consumption from two dimensions. Based on this, preset correlation weights are introduced between energy consumption influencing factors and each energy consumption type, and the influencing factor data is weighted and pooled. This mechanism dynamically selects the information of each energy consumption influencing factor based on its correlation and aggregates it into a core feature vector. The core feature vector is then concatenated and fused with the energy consumption feature vector extracted from the multidimensional energy consumption feature tensor. This allows the model to retain the cyclical patterns of energy consumption itself while adaptively incorporating the differentiated effects of external influencing factors. Finally, the target energy consumption data is output through linear mapping, achieving the fusion of the multi-period intrinsic patterns of historical energy consumption data and the dynamic interaction of external influencing factors, significantly improving the accuracy and robustness of upsampling historical energy consumption data.

[0070] According to an embodiment of the present invention, historical energy consumption data of various energy consumption types within a preset period are reconstructed into a multidimensional energy consumption feature tensor based on a preset period length. This includes: for each energy consumption type, dividing the historical energy consumption data into multiple consecutive period segments in time according to the preset period length, and arranging the data in each period segment into row vectors in chronological order; stacking the multiple row vectors to form a two-dimensional matrix according to the chronological order between the multiple period segments, wherein the row direction of the two-dimensional matrix represents each moment in a single period, and the column direction represents the corresponding moment in different periods; and stacking the two-dimensional matrices corresponding to different energy consumption types along the third dimension to form a multidimensional energy consumption feature tensor.

[0071] Regarding the drawbacks of the preset period length, during the training phase, a fast Fourier transform can be performed on the reference energy consumption data for each energy consumption type, and the amplitude corresponding to each frequency can be calculated in the frequency domain. The specific calculation method is shown in formula (1).

[0072] (1)

[0073] in, Represents the Fast Fourier Transform. Indicates amplitude calculation, This represents the average of the calculated results of reference energy consumption data for different energy consumption types. X represents the sample energy consumption data, and A represents the amplitude of the sample energy consumption data at each frequency. A= .

[0074] By analyzing the amplitude distribution and eliminating high-frequency noise, the top k amplitude values ​​arranged from largest to smallest were selected. and the corresponding frequency The period lengths corresponding to the first k amplitudes are calculated and used as the preset period lengths. The specific calculation methods are shown in formulas (2) and (3).

[0075] (2)

[0076] (3)

[0077] in, This indicates the frequencies corresponding to the first k amplitude values ​​arranged from largest to smallest. This indicates rounding up, where T represents the sequence length of the sample energy consumption data, and f i The i-th term representing the frequency, t i This represents the i-th term of the period, which is also the predetermined preset period length, and .

[0078] After obtaining the preset cycle length, for each energy consumption type (e.g., coal, oil, natural gas), according to the selected preset cycle length t i The original sequence of historical energy consumption data is divided into multiple consecutive periodic segments. If the sequence length of the historical energy consumption data is not divisible by the preset period length, zeros are padded at the end of the sequence to make it completely divisible.

[0079] The data within each periodic segment is arranged into a row vector in chronological order. All the row vectors corresponding to each periodic segment are then stacked according to their chronological order (i.e., the temporal order of the periodic segments) to form a two-dimensional matrix. The row indices of this two-dimensional matrix correspond to specific moments within a single period (e.g., January to December of a year), while the column indices correspond to the same phase position in different periods (e.g., January of the first year, January of the second year, etc.). In this way, the row direction of the two-dimensional matrix captures the changes within a period, i.e., the patterns of evolution over time within the same period, while the column direction captures the changes between periods, i.e., the correlation between different periods at the same phase.

[0080] The above operation is repeated for each type of energy consumption to obtain several independent two-dimensional matrices (each matrix corresponds to a type of energy). Finally, these two-dimensional matrices are stacked along the third dimension (i.e., the depth direction, which represents the type of energy consumption) to form a multidimensional energy consumption feature tensor, as shown in formula (4).

[0081] (4)

[0082] in, This represents historical energy consumption data. This indicates a zero-padding operation. This step is unnecessary if the sequence length of the historical energy consumption data is divisible by the preset period length. This represents the process of constructing a multidimensional energy consumption feature tensor. This represents the multidimensional energy consumption feature tensor.

[0083] The three dimensions of the multidimensional energy consumption feature tensor represent: time within the period, period number, and energy consumption type. This structure not only fully preserves all information from historical energy consumption data but also supports subsequent extraction of multi-scale periodic features. If the original sequence increases in length due to zero-padding, the actual data points need to be truncated during subsequent feature sequence concatenation to ensure time alignment.

[0084] Figure 3 A schematic diagram of a multidimensional energy consumption feature tensor according to an embodiment of the present invention is shown.

[0085] like Figure 3 As shown, a multidimensional energy consumption feature tensor is presented using historical energy consumption data of three types: coal (C), oil (P), and natural gas (G). Specifically, the historical energy consumption data of each type is divided into multiple continuous periodic segments (such as periodic segment 1, periodic segment 2, periodic segment 3, etc.) according to a preset period length. The data within each periodic segment are arranged in chronological order, and the positions of the same phase in each periodic segment (such as the same moment in each periodic segment) are vertically aligned to form the direction of change within the period; while the horizontal direction shows the temporal changes between different moments within the same periodic segment. Figure 3 This intuitively illustrates the construction logic of the row direction (changes within the period) and column direction (changes during the period) in the multidimensional energy consumption feature tensor.

[0086] By dividing historical energy consumption data of each energy consumption type into row vectors according to periodic segments and stacking them into a two-dimensional matrix in periodic order, and then stacking the two-dimensional matrices of different energy consumption types along the third dimension into a multi-dimensional energy consumption feature tensor, a unified structured representation of temporal changes within a period and inter-period correlation changes is realized, providing an explicit two-dimensional modeling foundation for subsequent multi-scale periodic feature extraction.

[0087] According to an embodiment of the present invention, the energy consumption data upsampling method based on transfer learning further includes: extracting features from the multidimensional energy consumption feature tensor using hierarchical initial convolution to obtain periodic feature sequences corresponding to each energy consumption type; splicing and truncating each periodic feature sequence in chronological order to obtain energy consumption feature vectors corresponding to each energy consumption type, wherein the time length of the energy consumption feature vector is the same as that of the historical energy consumption data.

[0088] The multi-dimensional energy consumption feature tensor is extracted using hierarchical initial convolution to obtain the intra- and inter-periodic variation features at multiple scales, and this information is then aggregated. The multi-dimensional energy consumption feature tensor is processed by convolution to obtain the periodic feature sequence, and the specific calculation method is shown in formula (5).

[0089] (5)

[0090] in, This represents a hierarchical structure, where different two-dimensional matrices representing different energy consumption types within the multidimensional energy consumption feature tensor are considered as different layers. This indicates that an initial convolution is performed on each layer. It represents a periodic characteristic sequence.

[0091] To obtain the energy consumption feature representation for the entire preset time period, the periodic feature sequences of each segment are reassembled for the historical energy consumption data of each energy consumption type in the multidimensional energy consumption feature tensor to form a complete time series, i.e., the energy consumption feature vector. Since the zero-padding operation may have been used before, the sequence length will increase accordingly. Therefore, during the splicing process, the sequence needs to be truncated to ensure that it is consistent with the original sequence length, so as to obtain the deformed energy consumption feature vector. The specific calculation process is shown in formula (6).

[0092] (6)

[0093] in, This indicates that the characteristic sequences of each period are spliced ​​together. This indicates that the sequence length is truncated. This represents the energy consumption feature vector.

[0094] The amplitude of historical energy consumption data at various frequencies reflects the importance of the corresponding period and frequency, and correspondingly, also reflects the importance of multidimensional energy consumption feature tensors constructed based on different frequencies and periods. Therefore, by using amplitude information to aggregate energy consumption feature vectors of various energy consumption types, a comprehensive energy consumption tensor that aggregates multiple periodic variation information is obtained. The calculation method is shown in formulas (7) and (8).

[0095] (7)

[0096] (8)

[0097] in, The values ​​represent the amplitude of historical energy consumption data at various frequencies, and Softmax represents the normalization function. include This represents the result after normalizing the amplitude of historical energy consumption data at various frequencies. , express elements in (where h represents the number of energy consumption types).

[0098] By performing multi-scale feature extraction on the multi-dimensional energy consumption feature tensor through hierarchical initial convolution, and concatenating the feature sequences of each period in chronological order to extract an energy consumption feature vector of the same length as the original data, it is possible to preserve the multi-scale change information within and between periods while aligning the feature dimension with the input time axis. This lays a foundation for complete information and temporal matching for subsequent accurate fusion with the core feature vector.

[0099] According to an embodiment of the present invention, based on preset correlation weights for multiple energy consumption influencing factors for each energy consumption type, weighted pooling is performed on energy consumption influencing data associated with multiple energy consumption influencing factors within a preset time period to select and aggregate information of each energy consumption influencing factor to obtain a core feature vector. This includes: organizing energy consumption influencing data associated with multiple energy consumption influencing factors within a preset time period into an initial influencing tensor; mapping the initial influencing tensor from its original time length to a preset core length using a first multilayer perceptron in the time dimension to obtain a compressed influencing tensor; sampling the energy consumption influencing factors at each time step in the compressed influencing tensor for each energy consumption type according to preset correlation weights to obtain a sampling index for each time step; extracting information of the corresponding energy consumption influencing factors from the compressed influencing tensor according to the sampling index as the core feature representation for each time step; and combining the core feature representations of each time step in chronological order to form a core feature vector.

[0100] The energy consumption impact data associated with multiple energy consumption influencing factors (such as temperature, energy price, and electricity consumption) within a preset time period are organized into an initial impact tensor. Assume there are h types of energy consumption (such as coal, oil, and natural gas), each type of energy consumption corresponds to a set of energy consumption impact data, the original time length of which is L (e.g., 60 months), and the number of energy consumption influencing factors is p (p=3 in this embodiment). These data are stacked according to energy consumption type, as shown in formula (9).

[0101] (9)

[0102] in, , and ( ) represent data points for temperature, energy price, and electricity consumption, respectively; L represents the original time length; Y represents the data points for temperature, energy price, and electricity consumption, respectively. u This indicates the impact of energy consumption data.

[0103] The matrices of energy consumption influencing factors corresponding to the h energy consumption types are stacked to construct a tensor Y containing the initial influence of all energy consumption influencing factors, allowing for simultaneous processing of this data. This is achieved through a function... Obtain the core representation of the initial influence tensor of various energy consumption types. (in The specific calculation method is shown in formula (10).

[0104] (10)

[0105] To obtain the core representation, the initial influence tensor Y needs to be passed through a first multilayer perceptron (MLP1) with an activation function to map the initial influence tensor Y from its original time length L to a preset core length L in the time dimension. c The compression effect tensor is obtained, and the specific calculation method is shown in formula (11).

[0106] (11)

[0107] in, This indicates the effect of compression on the tensor.

[0108] Obtain the preset relevance weight w j (Here, j is taken as {1,2,3}), and the preset correlation weights corresponding to each energy consumption impact factor are copied to L. c The above steps are repeated for each of the preset correlation weights corresponding to historical energy consumption data of different energy consumption types, and these weights are combined to form a complete correlation weight. , Represents the space of real numbers.

[0109] For each type of energy consumption, a preset correlation weight is used to sample each column according to the preset correlation weight. The larger the value of the preset correlation weight, the higher the probability of being sampled, indicating that more attention needs to be paid to the influence of the energy consumption influencing factor on the historical energy consumption data at the corresponding time step. The value returned by the sampling is the corresponding sampling index, and the specific representation of the sampling index is shown in formula (12).

[0110] (12)

[0111] in, This indicates a probability-based sampling operation. This indicates the sampling index.

[0112] The corresponding element value is taken from each matrix of the compression influence tensor according to the sampling index, and the specific calculation method is shown in formula (13).

[0113] (13)

[0114] in, This indicates an operation based on the value of the sampling index. This represents the core feature vector corresponding to the u-th energy consumption type.

[0115] By organizing energy consumption impact data into an initial impact tensor, compressing the time dimension using a first multilayer perceptron, and then performing probability sampling at each time step based on preset correlation weights, the corresponding energy consumption impact factor information is extracted according to the sampling index and combined into a core feature vector in chronological order. This achieves adaptive screening and compression aggregation of multi-source external impact factors, enabling the core feature vector to focus on the factors most relevant to energy consumption and reducing interference from redundant or irrelevant factors.

[0116] According to an embodiment of the present invention, concatenating the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector includes: concatenating the core feature vector with the energy consumption feature vector corresponding to each energy consumption type in the time dimension to obtain a combined feature sequence corresponding to each energy consumption type; mapping each combined feature sequence to a fused feature sequence corresponding to each energy consumption type through a second multilayer perceptron, so as to use multiple fused feature sequences as a fused feature vector, wherein the time length of the fused feature vector is the same as that of the historical energy consumption data.

[0117] The core feature vector is combined with the comprehensive energy consumption tensor of the u-th energy consumption type, which has extracted its own energy consumption variation features within multiple periods and during the week. The features of the two sequences are concatenated and further fused to obtain combined feature sequences corresponding to each energy consumption type. A second multilayer perceptron maps each combined feature sequence to a fused feature sequence F corresponding to each energy consumption type. u The specific calculation method is shown in formula (14).

[0118] (14)

[0119] in, This indicates a splicing operation. This represents the second multilayer perceptron, used to convert the stitched length from T+L. c Mapped to T, Represents linear layer mapping Where H represents the length of the target energy consumption data obtained after mapping. .

[0120] By concatenating the core feature vector with the energy consumption feature vectors corresponding to each energy consumption type in the time dimension, and mapping them through a second multilayer perceptron into a fusion feature sequence of the same length as historical data, the dynamic selection information of external influencing factors and the periodic features of energy consumption itself are precisely aligned and deeply coupled in the time axis. This allows the fusion feature vector to simultaneously carry the intrinsic laws of energy consumption and the differentiated effects of external factors, providing a feature representation with complete information and unified dimensions for subsequent high-resolution mapping.

[0121] According to an embodiment of the present invention, the transfer learning training further includes: using sample energy consumption data of multiple energy consumption types corresponding to the region to which the historical energy consumption data belongs as a supervision signal, iteratively adjusting the weights and biases of the initial mapping layer until the difference between the total energy consumption value of the predicted energy consumption data output by the linear mapping layer accumulated on the time axis and the total energy consumption value in the sample energy consumption data satisfies a preset convergence condition; in response to the difference satisfying the preset convergence condition, using the adjusted initial mapping layer as the linear mapping layer; wherein, the initial mapping layer is trained by reference energy consumption data of multiple energy consumption types in the reference region during the historical period and reference influence data associated with multiple energy consumption influence factors during the historical period.

[0122] For the data upsampling task of the region to which the historical energy consumption data belongs (i.e. the target region), the corresponding sample energy consumption data still adopts the existing fine-grained energy consumption data of developed regions that are highly similar to the target region, and the energy consumption influence data belonging to the target region is added to the energy consumption influence factor. Through the fine-tuning method in transfer learning, the parameters of the encoder in the model are frozen, and the parameters of the decoder (including the initial mapping layer) are fine-tuned. That is, considering the differences in energy consumption levels in different regions, the weights and biases of the initial mapping layer in the decoder are adjusted by using the existing sample energy consumption data of the two regions, so as to obtain the linear mapping layer, thereby obtaining the target energy consumption data of the target region more accurately. Specifically, the specific fine-tuning process during training is shown in formula (15).

[0123] (15)

[0124] in, and These represent the optimal model mapping parameters for the developed region and the target region, respectively. This indicates a fine-tuning operation. and ( The numbers ) represent the energy consumption of energy consumption type u in the target area and the developed area, both of which are low-frequency, respectively, and s represents the number of low-frequency energy consumption data of each type. , This represents the target energy consumption data corresponding to the historical energy consumption data, obtained by using the adjusted linear mapping layer output.

[0125] By using low-frequency sample energy consumption data from the target region as a supervision signal, the initial mapping layer of the pre-trained model is iteratively adjusted to match the accumulated prediction total with the supervision signal. This enables the model to adaptively migrate from the reference region to the target region with low resources. While preserving the general energy consumption pattern, it accurately adapts to the local total energy consumption scale, thereby improving the regional adaptability and engineering practicality of the upsampling results.

[0126] Figure 4 A schematic diagram of energy consumption upsampling data processing according to an embodiment of the present invention is shown.

[0127] like Figure 4 As shown, the processing of energy consumption data is mainly divided into two branches. The first branch processes the energy consumption influencing factor tensor corresponding to energy consumption. The three dimensions of this tensor correspond to time, energy consumption influencing factor type, and energy consumption type, respectively. The figure labels three specific factors at the energy consumption influencing factor type: y1, y2, and y3, representing temperature, energy price, and electricity consumption, respectively. After this tensor is expanded along the time dimension, it is input into the first multilayer perceptron and, based on pooling with preset correlation weights, outputs a core feature vector (labeled "core" at the top of the figure). During the pooling process, the feature extraction results are probabilistically extracted based on preset correlation weights to extract features from the feature map (including y1, y2, and y3). 11 y 12 y 12 ...y 36 ) is mapped to local features (including y) 11 y 32 y 13 y 24 y 35 y 16 The entire first branch unfolds along the timeline, reflecting the dynamic selection and compression of three energy consumption influencing factors.

[0128] The second branch processes the energy consumption tensor convolution aggregation result, which includes energy consumption amount and type. This result, along with the core feature vector output from the first branch, is fed into the concatenation module for concatenation, yielding the concatenated result. The concatenated result then passes through a second multilayer perceptron and a linear mapping layer, and is further refined in the fine-tuning module. With the participation of [the relevant entity], the final output is the target energy consumption data.

[0129] Through a dual-branch structure, the first branch dynamically extracts and compresses multi-source influencing factors such as temperature, energy prices, and electricity consumption based on preset correlation weights, effectively mapping high-dimensional feature maps into low-dimensional local features, thus achieving focused aggregation of key external factors. The second branch fuses the energy consumption tensor convolution result with the core feature vector, and after passing through a multilayer perceptron and linear mapping, combines fine-tuning to output target energy consumption data, thereby improving the accuracy and adaptability of upsampling.

[0130] According to an embodiment of the present invention, the preset correlation weights are determined as follows: Reference energy consumption data for multiple energy consumption types within a reference period, and reference impact data associated with multiple energy consumption impact factors within the reference period; for each energy consumption type, based on the joint change trend and fluctuation amplitude of the reference energy consumption data and the reference impact data, the linear correlation coefficient between each energy consumption impact factor and the reference energy consumption data of the energy consumption type is determined; the linear correlation coefficients of each energy consumption impact factor under the same energy consumption type are normalized to obtain the preset correlation weights of multiple energy consumption impact factors for each energy consumption type.

[0131] The similarity between energy consumption data in the same industry is calculated using cosine similarity. The specific calculation method is shown in formula (16). The closer the result is to 1, the higher the similarity between the data. If the idea of ​​transfer learning is adopted, the existing fine-grained energy consumption data can be used as prior knowledge to provide high-value intrinsic information for upsampling low-frequency energy consumption data in the target area.

[0132] (16)

[0133] in, and These represent the energy consumption vectors of the two regions respectively. and The i-th data point in This indicates the number of energy consumption data points for each region.

[0134] Further analysis was conducted on the correlation between multiple energy consumption influencing factors and energy consumption data. The consumption of energy types such as coal, oil, and natural gas is affected by a variety of factors, such as climate, energy prices, and consumption distribution. To verify and quantify the correlation between different energy consumption influencing factors and energy consumption data, Pearson Correlation Analysis was used. The specific analysis process is shown in formula (17).

[0135] (17)

[0136] in, This indicates reference energy consumption data. This indicates the reference influence data; Cov indicates the calculation of the covariance between two variables. and These represent the standard deviations of the reference energy consumption data and the reference impact data, respectively. This represents the linear correlation coefficient.

[0137] Pearson analysis further shows that there are differences in the correlation between different energy consumption influencing factors and energy consumption data of different energy consumption types. The stronger the correlation, the higher the contribution of the energy consumption influencing factor to the energy consumption data. Therefore, it is necessary to calculate the preset correlation weight based on the correlation analysis results. Considering that the Softmax function can convert a real number vector into a probability distribution vector, where each element takes values ​​between 0 and 1 and the sum is 1, the preset correlation weight w corresponding to various energy consumption influencing factors and energy consumption data of different energy consumption types can be obtained through formula (18). j The original positive / negative correlation can be learned through subsequent model training.

[0138] (18)

[0139] Where, r j Let represent the linear correlation coefficient of the j-th energy consumption influencing factor, p represent the number of energy consumption influencing factors, which is also the number of linear correlation coefficients, and r represent the number of linear correlation coefficients. i Let i be the i-th term of the p linear correlation coefficients.

[0140] By acquiring reference energy consumption data and related influencing factor data for various energy consumption types within a reference period, linear correlation coefficients are calculated and normalized for each energy consumption type. This yields the preset correlation weights of each influencing factor for each energy consumption type, providing a quantifiable basis for the degree of correlation in weighted pooling. This ensures that the selection of information on energy consumption influencing factors accurately reflects their actual contribution, thereby improving the targeting and reliability of the upsampling process.

[0141] Figure 5A A heatmap of cosine similarity analysis results according to an embodiment of the present invention is shown.

[0142] Figure 5B A heatmap of correlation analysis results according to an embodiment of the present invention is shown.

[0143] like Figure 5A As shown, the similarity of natural gas consumption among regions A, B, C, D, and E is all above 0.97, close to 1, indicating that the energy consumption of these regions has common characteristics.

[0144] like Figure 5B As shown, the negative correlation between energy prices and temperature and natural gas consumption is as high as 0.8 or more, while the correlation between electricity consumption and natural gas consumption is -0.654, which is a strong inverse correlation.

[0145] Figure 6 A flowchart of an energy consumption data upsampling method based on transfer learning according to another embodiment of the present invention is shown.

[0146] like Figure 6 As shown, the method includes operations S601 to S615.

[0147] In the tensor modeling phase, in operation S601, historical energy consumption data and energy consumption impact data are input. In operation S602, historical energy consumption data is analyzed in the frequency domain. In operation S603, the top k values ​​with the largest amplitudes and their corresponding frequencies and periods are extracted. In operation S604, it is determined whether the length of the period is divisible. If yes, operation S606 is executed; otherwise, operation S605 is executed. In operation S605, zero-padding is performed. In operation S606, a multidimensional energy consumption feature tensor is constructed. In operation S607, periodic variation features are extracted using hierarchical initial convolution. In operation S608, multiple sets of periodic information from the multidimensional energy consumption feature tensor are aggregated using amplitudes to obtain an energy consumption feature vector. In operation S609, a third-order tensor corresponding to the energy consumption impact factors in the energy consumption impact data is constructed.

[0148] In the model processing stage, in operation S610, the third-order tensor is pooled based on preset correlation weights to obtain the core feature vector. In operation S611, the core feature vector and the energy consumption feature vector are concatenated and fused. In operation S612, the fitted fine-grained energy consumption data is output.

[0149] In operation S613, the error is determined to be minimized. If so, operation S615 is executed; otherwise, operation S614 is executed. In operation S614, the k value is updated, and then the process returns to operation S603. In operation S615, a linear mapping is performed on the fine-grained energy consumption data to obtain the target energy consumption data.

[0150] According to embodiments of the present invention, frequency domain analysis is performed on historical energy consumption data to extract the top few dominant frequencies with the largest amplitudes and their corresponding periods. Combined with integer division judgment and zero-padding, a multi-dimensional energy consumption feature tensor reflecting changes within and between periods can be accurately constructed. Subsequently, hierarchical initial convolution is used to capture multi-scale periodic change patterns, and multiple sets of periodic information are weighted and fused according to amplitude magnitude, significantly enhancing the completeness and robustness of energy consumption feature expression. Simultaneously, energy consumption impact data is organized into a third-order tensor and pooled based on preset correlation weights, effectively filtering out the most strongly correlated impact factors and generating a compact core feature vector.

[0151] After concatenation and feature fusion of the core feature vector and the energy consumption feature vector, the output fine-grained energy consumption data dynamically adjusts the number of periods (k value) through an error feedback mechanism until the root mean square error, mean absolute error, and mean absolute percentage error reach their minimum. Finally, high-resolution target energy consumption data is generated through linear mapping. The entire process achieves adaptive fusion of the inherent periodicity of energy consumption data and external influencing factors, and ensures upsampling accuracy through iterative optimization, significantly improving the accuracy and robustness of data refinement.

[0152] According to another embodiment of the present invention, in order to verify the effectiveness of the proposed improved pooling method, random pooling, max pooling, and average pooling were selected for comparison to study the impact of different pooling methods on the experimental results. The experimental results are shown in Table 1, and the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used to illustrate the comparison results.

[0153] Table 1

[0154]

[0155] In the sampling results for coal consumption, the maximum pooling method had the largest error, which was higher than the errors of the other pooling methods. In the sampling results for oil consumption, although the RMSE of the maximum pooling method was 0.7299, which was 3.44% lower than that of the pooling method in this embodiment, its MAE and MAPE were 5.85% and 5.78% higher than those of the pooling method in this paper, respectively. In the sampling results for natural gas consumption, the RMSE of the maximum pooling method was slightly lower than that of the pooling method in this embodiment by 0.74%, but its MAE and MAPE were 1.12% and 0.87% higher than those of the pooling method in this paper, respectively.

[0156] The three error metrics of random pooling and average pooling were all higher than the pooling method in this embodiment in the upsampling results of the three energy consumption scenarios. This is because random pooling selects input data randomly, max pooling only focuses on the maximum value in the data, and average pooling calculates the average value from the input data as the output. These three pooling methods fail to reflect the differences in the correlation between different external influencing factors and energy consumption. The pooling method in this embodiment takes this into account, thereby achieving accurate interaction between variables, and the experimental results obtained are generally better. Therefore, the experiments in this section prove the effectiveness of the pooling method proposed in this paper.

[0157] According to another embodiment of the present invention, this embodiment establishes a multidimensional energy consumption feature tensor of historical energy consumption data to simultaneously mine the periodic variation characteristics of historical energy consumption data of various energy consumption types. To verify the effectiveness of this paradigm, the algorithm proposed in this embodiment (denoted as M1) and a method that does not use the tensor paradigm, i.e., directly identifies the interaction relationship between each historical energy consumption data and energy consumption influencing factors (denoted as M2), are subjected to performance testing. The experimental results are shown in Table 2.

[0158] Table 2

[0159]

[0160] In the coal consumption upsampling results, although M2's MAPE was the same as M1's, its RMSE and MAE were 2.19% and 0.55% higher than M1's, respectively. In the oil consumption upsampling results, the three indicators in this embodiment were the lowest compared to M2. In the natural gas consumption upsampling results, the three error indicators of M2 increased by 6.01%, 7.95%, and 7.79% compared to those in this paper, respectively. This is because M2 lacks in-depth analysis of the periodic information of the energy consumption data itself, and ignoring the periodic variation characteristics will reduce the accuracy of the upsampling results.

[0161] Figure 7A An error graph of coal consumption upsampling according to an embodiment of the present invention is shown.

[0162] Figure 7B An error graph of oil consumption upsampling according to an embodiment of the present invention is shown.

[0163] Figure 7C An error graph of natural gas consumption upsampling according to an embodiment of the present invention is shown.

[0164] According to another embodiment of the present invention, in order to verify the effect of the algorithm proposed in this embodiment compared with other algorithms, three methods proposed in recent years that can also be used to solve timing problems—Inverted Transformer (iTransformer), Pyramidal Transformer (Pyraformer), and Multi-scale Gated Transformer Unit (M-GTU)—were selected as comparison algorithms for comparison with the method in this paper. The experimental results are shown in Table 3 and as shown in Figure 4. Figure 7A , Figure 7B , Figure 7C As shown.

[0165] Table 3

[0166]

[0167] Specifically analyzing the error indices in the table, in the coal consumption upsampling results, M-GTU had a lower error, but compared to M-GTU, the three error indices in this embodiment were reduced by 15.63%, 22.88%, and 23.66%, respectively. In the oil consumption upsampling results, Pyraformer had a lower experimental error, but compared to Pyraformer, the errors in this embodiment were still 29.34%, 35.70%, and 36.73% lower. The RMSE, MAE, and MAPE of iTransformer and M-GTU were significantly higher than the error indices in this embodiment. In the natural gas consumption upsampling results, Pyraformer had the highest three error indices, followed by M-GTU, with iTransformer having a lower error, but the error indices in this embodiment were reduced by 13.92%, 1.57%, and 0.86% compared to iTransformer, respectively. In summary, compared to the three comparison algorithms, this embodiment showed the best experimental results and the lowest error.

[0168] This is because, while iTransformer can extract relevant information between variables, it lacks in-depth mining of the rich periodic variation characteristics of the energy consumption sequence itself; Pyraformer and M-GTU can study the temporal variation characteristics of the energy consumption sequence at multiple scales, but their mining of periodic variation information is still insufficient, and they have significant limitations in the interaction between external influencing factors and the energy consumption sequence. This embodiment, however, can fully mine the periodic variation characteristics within the energy consumption sequence and promote the accurate interaction between energy consumption and external influencing factors, thus achieving better experimental results.

[0169] Figure 7D Radar graphs showing different algorithmic errors according to embodiments of the present invention are illustrated.

[0170] like Figure 7D As shown, the error radar chart is plotted by averaging the percentage error between the fitted value and the actual value for the month after upsampling the three types of energy consumption data. The inverted converter, pyramid converter, and multi-scale gated converter units generally have larger monthly absolute errors after upsampling the three types of energy consumption, failing to achieve high-precision upsampling of all three types simultaneously, and occupying a larger area in the error radar chart. In contrast, the upsampling results of this embodiment show lower overall monthly absolute errors in energy consumption, and occupy the smallest area in the error radar chart, indicating relatively higher upsampling accuracy.

[0171] According to another embodiment of the present invention, in order to verify the effectiveness of this embodiment in advanced carbon emission accounting, this section focuses on the carbon emission monitoring scenario of multiple energy sources such as coal, oil and natural gas used in the steel heavy industry. Based on the carbon emission factors corresponding to various energy sources, a comparison is made between carbon emission calculations and monthly data time-series upsampling carbon emission calculations performed using the algorithm of this embodiment. The experimental results are shown in Table 4, where the unit of carbon emission is tons of CO2, SQ represents before upsampling using the algorithm of this embodiment, and SH represents after upsampling using the algorithm of this embodiment.

[0172] Table 4

[0173]

[0174] The experimental results show that for industrial carbon emission monitoring, the algorithm upsampling in this embodiment can generate fine-grained monthly carbon emission data by upsampling the annual cumulative carbon emissions. Comparing the two methods, the errors are only: 1.11% for coal consumption, 1.63% for oil consumption, and only 0.69% for natural gas. This ensures that the carbon emissions before and after upsampling will not cause significant deviations. The fine-grained data will provide strong support for studying the changing patterns of carbon emissions in industry production and formulating carbon reduction strategies.

[0175] According to another embodiment of the present invention, this embodiment considers a more complex scenario, verifying the effectiveness of the proposed industrial energy consumption data upsampling algorithm under external influencing factors such as sudden temperature changes (e.g., severe short-term weather effects from hurricanes in summer and autumn) and complex fluctuations in energy prices. The algorithm is also compared with three other methods: iTransformer, Pyraformer, and M-GTU. Experimental results are shown in Table 5 and... Figure 7E As shown.

[0176] Table 5

[0177]

[0178] In the sampling results for coal consumption, the RMSE and MAE of this embodiment are both lower than those of the other algorithms. The MAPE is 9.28%, which is 17.07% lower than iTransformer, 15.64% lower than Pyraformer, and 41.41% lower than M-GTU. In the sampling results for oil consumption, the RMSE and MAE of this embodiment are the lowest compared to the comparison algorithms. The MAPE is 6.55%, which is 44.54% lower than iTransformer, 44.4% lower than Pyraformer, and 35.66% lower than M-GTU. In the sampling results for natural gas consumption, the errors of Pyraformer and M-GTU are relatively large. The three error indicators of iTransformer are slightly lower, but compared with iTransformer, this embodiment is still 0.64%, 11.73%, and 5.84% lower.

[0179] Figure 7E A radar chart illustrating the algorithm error under extreme conditions according to an embodiment of the present invention is shown.

[0180] like Figure 7E As shown, the three comparison algorithms occupy a large area, indicating that their monthly energy consumption error percentage after upsampling is generally large; while the area occupied by this embodiment is the smallest, and its experimental results are more accurate.

[0181] In summary, in complex scenarios, the algorithm proposed in this embodiment can still maintain a lower error compared to the comparison algorithm. This is due to the robustness of the model and the reasonable consideration of the correlation between external influencing factors and energy consumption.

[0182] According to another embodiment of the present invention, in determining the selection of the preset period length with respect to the value of k, this embodiment selects the top k principal values ​​with the largest amplitude in the frequency domain of the sample energy consumption data, along with their corresponding frequencies and periods, and constructs a set of multidimensional energy consumption feature tensors containing different period lengths. The optimal value of k is determined by performing sensitivity analysis on the k value. The experimental results are shown in Table 6.

[0183] Table 6

[0184]

[0185] In the sampling results for coal consumption, when k is 2, the three error indices are 0.0513, 0.0367, and 2.02%, respectively. As the value of k increases, the error decreases, reaching its lowest point at k 4, where the errors are reduced by 2.14%, 0.82%, and 0.99% compared to the results with k 2. However, when k is 5, the error begins to increase again. In the sampling results for oil consumption, the experimental error again decreases and then increases with increasing k, reaching its lowest point at k 4. In the sampling results for natural gas consumption, the error is smallest when k is 4, with reductions of 20.98%, 17.78%, and 17.79% compared to the results with k 2, and reductions of 21.79%, 17.54%, and 17.5% compared to the results with k 5. Therefore, in the experimental example of this embodiment, the experimental effect is the best and the error is lower when k is 4. At this time, the constructed multidimensional energy consumption feature tensor not only contains the frequency information with physical meaning of the energy consumption data itself, but also minimizes the impact of meaningless high-frequency noise on the experimental results.

[0186] According to another embodiment of the present invention, in the process of upsampling low-frequency energy consumption data, it is necessary to determine a suitable data input scale. If the length of the input data is too short, it is insufficient to provide complete feature information, resulting in a large error between the output energy consumption fitted value and the actual value; however, when the input data is very long, redundant information may be introduced, increasing the fitting difficulty and making the output error larger. Therefore, this embodiment studies the influence of different input scales on experimental results, setting four input scales of 36, 48, 60, and 72 (unit: months) to determine the optimal input length.

[0187] Figure 8A A graph showing different input errors for coal consumption according to an embodiment of the present invention is shown.

[0188] Figure 8B A graph showing different input errors for oil consumption according to an embodiment of the present invention is shown.

[0189] Figure 8C A graph showing different input errors for natural gas consumption according to an embodiment of the present invention is shown.

[0190] Figure 8D Radar plots of different input scale errors according to embodiments of the present invention are shown.

[0191] like Figure 8A , Figure 8B , Figure 8C , Figure 8DAs shown, for input lengths of 36 and 48, the absolute errors of the three energy consumption methods are generally large after upsampling, and the areas they occupy in the error radar chart are also large. When the input length is 60, the monthly absolute errors of the three energy consumption methods are basically small, and the area in the radar chart is the smallest, indicating that the upsampling result is good. When the input length is increased to 72, the absolute errors of the energy consumption begin to increase again, and the area in the error radar chart also increases. A detailed analysis of the error indicators follows.

[0192] For an input length of 60 months, the MAPE values ​​after upsampling of coal, oil, and natural gas were 2.00%, 3.29%, and 2.31%, respectively, which were the lowest compared to other input scales. The corresponding RMSE and MAE were also the lowest. Therefore, the experimental results were best when the input data length was 60 months. This was determined as the input length for subsequent experiments in this embodiment, ensuring sufficient data feature information is provided without introducing a large amount of redundant information.

[0193] Figure 9 A block diagram of an energy consumption data upsampling system based on transfer learning according to an embodiment of the present invention is shown.

[0194] like Figure 9 As shown, the energy consumption data upsampling system 900 based on transfer learning includes a data reconstruction module 910, a core extraction module 920, a vector concatenation module 930, and a mapping output module 940.

[0195] The data reconstruction module 910 is used to reconstruct historical energy consumption data of various energy consumption types within a preset period into a multidimensional energy consumption feature tensor based on a preset period length. The multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided based on the preset period length.

[0196] The core extraction module 920 is used to perform weighted pooling on the energy consumption impact data associated with multiple energy consumption impact factors within a preset time period based on the preset correlation weights of multiple energy consumption impact factors for each energy consumption type, so as to select and aggregate the information of each energy consumption impact factor and obtain the core feature vector.

[0197] The vector concatenation module 930 is used to concatenate the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector.

[0198] The mapping output module 940 is used to input the fused feature vector into the linear mapping layer for mapping and output the target energy consumption data. The linear mapping layer is obtained by transfer learning training, which includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

[0199] According to an embodiment of the present invention, the data reconstruction module 910 includes a periodic partitioning submodule, a vector stacking submodule, and a matrix stacking submodule.

[0200] The period segmentation submodule is used to divide historical energy consumption data into multiple consecutive period segments in time according to a preset period length for each energy consumption type, so that the data in each period segment can be arranged into a row vector in chronological order.

[0201] The vector stacking submodule is used to stack multiple row vectors into a two-dimensional matrix according to the order of multiple period segments. The row direction of the two-dimensional matrix represents the time points within a single period, and the column direction represents the corresponding time points in different periods.

[0202] The matrix stacking submodule is used to stack two-dimensional matrices corresponding to different energy consumption types along the third dimension to form a multidimensional energy consumption feature tensor.

[0203] According to an embodiment of the present invention, the core extraction module 920 includes a data organization submodule, a vector compression submodule, an information sampling module, an information extraction submodule, and a feature combination submodule.

[0204] The data organization submodule is used to organize energy consumption impact data associated with multiple energy consumption impact factors within a preset time period into an initial impact tensor.

[0205] The vector compression submodule is used to map the initial influence tensor from the original time length to a preset core length in the time dimension using the first multilayer perceptron, thereby obtaining the compressed influence tensor.

[0206] The information sampling module is used to sample the energy consumption influence factors at each time step in the compressed influence tensor for each energy consumption type, based on preset correlation weights, and obtain the sampling index for each time step.

[0207] The information extraction submodule is used to extract the information of the corresponding energy consumption influence factor from the compressed influence tensor according to the sampling index, and use it as the core feature representation of each time step.

[0208] The feature combination submodule is used to combine the core feature representations of each time step in chronological order to form a core feature vector.

[0209] According to an embodiment of the present invention, the energy consumption data upsampling system 900 based on transfer learning further includes a reference acquisition module, a linear determination module, and a weight generation module.

[0210] The reference acquisition module is used to acquire reference energy consumption data for various energy consumption types within the reference period, as well as reference impact data associated with multiple energy consumption impact factors within the reference period.

[0211] The linear determination module is used to determine the linear correlation coefficient between each energy consumption influencing factor and the reference energy consumption data of each energy consumption type, based on the joint trend of the reference energy consumption data and the reference influence data and their respective fluctuation amplitudes.

[0212] The weight generation module is used to normalize the linear correlation coefficients of various energy consumption influencing factors under the same energy consumption type, and obtain the preset correlation weights of multiple energy consumption influencing factors for each energy consumption type.

[0213] According to an embodiment of the present invention, the energy consumption data upsampling system 900 based on transfer learning further includes a feature extraction module and a feature processing module.

[0214] The feature extraction module is used to extract features from the multidimensional energy consumption feature tensor using hierarchical initial convolution to obtain periodic feature sequences corresponding to each energy consumption type.

[0215] The feature processing module is used to splice and truncate the feature sequences of each period in chronological order to obtain the energy consumption feature vector corresponding to each energy consumption type. The time length of the energy consumption feature vector is the same as that of the historical energy consumption data.

[0216] According to an embodiment of the present invention, the vector splicing module 930 includes a vector splicing submodule and a sequence fusion submodule.

[0217] The vector concatenation submodule is used to concatenate the core feature vector with the energy consumption feature vectors corresponding to each energy consumption type in the time dimension, so as to obtain the combined feature sequence corresponding to each energy consumption type.

[0218] The sequence fusion submodule is used to map each combined feature sequence into a fusion feature sequence corresponding to each energy consumption type through a second multilayer perceptron, so as to use multiple fusion feature sequences as a fusion feature vector, wherein the time length of the fusion feature vector is the same as that of the historical energy consumption data.

[0219] According to an embodiment of the present invention, the energy consumption data upsampling system 900 based on transfer learning further includes a mapping adjustment module and a mapping determination module.

[0220] The mapping adjustment module is used to use sample energy consumption data of various energy consumption types corresponding to the regions to which the historical energy consumption data belongs as supervision signals to iteratively adjust the weights and biases of the initial mapping layer until the difference between the total energy consumption value of the predicted energy consumption data output by the linear mapping layer and the total energy consumption value in the sample energy consumption data meets the preset convergence condition.

[0221] The mapping determination module is used to take the adjusted initial mapping layer as a linear mapping layer in response to the difference satisfying the preset convergence condition; wherein, the initial mapping layer is trained by reference energy consumption data of various energy consumption types in the reference area during the historical period and reference influence data associated with multiple energy consumption influence factors during the historical period.

[0222] Any one or more of the modules, submodules, units, and subunits according to embodiments of the present invention, or at least part of the functions of any one or more of them, can be implemented in a single module. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present invention can be implemented by being divided into multiple modules. Any one or more of the modules, submodules, units, and subunits according to embodiments of the present invention can be at least partially implemented as hardware circuits, such as field-programmable gate arrays (FPGAs), programmable logic arrays (PLAs), systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits (ASICs), or implemented in hardware or firmware by any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, and firmware, or in any suitable combination of any of these three implementation methods. Alternatively, one or more of the modules, submodules, units, and subunits according to embodiments of the present invention can be at least partially implemented as computer program modules, which, when run, can perform corresponding functions.

[0223] For example, any plurality of the data reconstruction module 910, core extraction module 920, vector splicing module 930, and mapping output module 940 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of the present invention, at least one of the data reconstruction module 910, core extraction module 920, vector splicing module 930, and mapping output module 940 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the data reconstruction module 910, core extraction module 920, vector splicing module 930, and mapping output module 940 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0224] It should be noted that the energy consumption data upsampling system part based on transfer learning in the embodiments of the present invention corresponds to the energy consumption data upsampling method part based on transfer learning in the embodiments of the present invention. For a detailed description of the energy consumption data upsampling system part based on transfer learning, please refer to the energy consumption data upsampling method part based on transfer learning, which will not be repeated here.

[0225] Figure 10 A block diagram of an electronic device suitable for implementing energy consumption data processing according to an embodiment of the present invention is shown. Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0226] like Figure 10 As shown, an electronic device 1000 according to an embodiment of the present invention includes a processor 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory ROM 1002 or a program loaded from a storage portion 1008 into a random access memory RAM 1003. The processor 1001 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1001 may also include onboard memory for caching purposes. The processor 1001 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0227] RAM 1003 stores various programs and data required for the operation of electronic device 1000. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Processor 1001 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 1002 and / or RAM 1003. It should be noted that programs may also be stored in one or more memories other than ROM 1002 and RAM 1003. Processor 1001 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in one or more memories.

[0228] According to an embodiment of the present invention, the electronic device 1000 may further include an input / output (I / O) interface 1005, which is also connected to a bus 1004. The electronic device 1000 may also include one or more of the following components connected to the input / output (I / O) interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the input / output (I / O) interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1010 as needed so that computer programs read from it can be installed into the storage section 1008 as needed.

[0229] According to embodiments of the present invention, the method flow according to embodiments of the present invention can be implemented as a computer software program. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by processor 1001, it performs the functions defined in the system of the embodiments of the present invention. According to embodiments of the present invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0230] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0231] According to embodiments of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0232] For example, according to embodiments of the present invention, a computer-readable storage medium may include the ROM 1002 and / or RAM 1003 described above and / or one or more memories other than ROM 1002 and RAM 1003.

[0233] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of the present invention. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the energy consumption data upsampling method based on transfer learning provided in the embodiments of the present invention.

[0234] When the computer program is executed by the processor 1001, it performs the functions defined in the system / apparatus of this embodiment of the invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0235] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1009, and / or installed from a removable medium 1011. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0236] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0237] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or pairings fall within the scope of this invention.

[0238] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. A method for upsampling energy consumption data based on transfer learning, characterized in that, The method includes: Based on a preset period length, historical energy consumption data of various energy consumption types within a preset time period are reconstructed into a multidimensional energy consumption feature tensor. The multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided based on the preset period length. The energy consumption types include coal, oil, and natural gas. Based on the preset correlation weights of multiple energy consumption influencing factors for each energy consumption type, the energy consumption impact data associated with the multiple energy consumption influencing factors within the preset time period are weighted and pooled to select and aggregate the information of each energy consumption influencing factor, and obtain the core feature vector focusing on the energy consumption influencing factor most related to the historical energy consumption data. The core feature vector is concatenated with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fusion feature vector that simultaneously carries the periodic pattern of the historical energy consumption data and the energy consumption influencing factor. The fused feature vector is input into a linear mapping layer for mapping, and the target energy consumption data is output to monitor carbon emissions and predict energy consumption. The linear mapping layer is obtained by transfer learning training, which includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

2. The energy consumption data upsampling method according to claim 1, characterized in that, The process of reconstructing historical energy consumption data of various energy consumption types within a preset time period into a multidimensional energy consumption feature tensor based on a preset period length includes: For each of the energy consumption types, the historical energy consumption data is divided into multiple consecutive period segments in time according to the preset period length, so that the data in each period segment is arranged into a row vector in chronological order. Based on the sequential order of the multiple periodic segments, the multiple row vectors are stacked to form a two-dimensional matrix, wherein the row direction of the two-dimensional matrix represents each moment within a single period, and the column direction represents the corresponding moment in different periods; The two-dimensional matrices corresponding to different energy consumption types are stacked along the third dimension to form the multidimensional energy consumption feature tensor.

3. The energy consumption data upsampling method according to claim 1, characterized in that, The method involves weighted pooling of energy consumption impact data associated with multiple energy consumption impact factors within a preset time period, based on preset correlation weights for each energy consumption type, to select and aggregate information from each energy consumption impact factor, thereby obtaining a core feature vector focusing on the energy consumption impact factor most relevant to the historical energy consumption data, including: The energy consumption impact data associated with multiple energy consumption impact factors within the preset time period are organized into an initial impact tensor; The initial influence tensor is mapped from the original time length to a preset core length in the time dimension by the first multilayer perceptron, thus obtaining the compressed influence tensor; For each of the energy consumption types, the energy consumption influence factors at each time step in the compressed influence tensor are sampled according to the preset correlation weights to obtain the sampling index of each time step. Based on the sampling index, information on the corresponding energy consumption influence factor is extracted from the compressed influence tensor and used as the core feature representation of each time step. The core feature representations of each time step are combined in chronological order to form the core feature vector.

4. The energy consumption data upsampling method according to claim 1, characterized in that, The preset relevance weights are determined in the following way: Obtain reference energy consumption data for various energy consumption types within a reference time period, as well as reference impact data associated with multiple energy consumption impact factors within the reference time period; For each of the energy consumption types, the linear correlation coefficient between each of the energy consumption influencing factors and the reference energy consumption data of the energy consumption type is determined based on the joint change trend and fluctuation range of the reference energy consumption data and the reference influence data. The linear correlation coefficients of the energy consumption influencing factors under the same energy consumption type are normalized to obtain the preset correlation weights of the multiple energy consumption influencing factors for each energy consumption type.

5. The energy consumption data upsampling method according to claim 1, characterized in that, The method further includes: The multidimensional energy consumption feature tensor is used to extract features by hierarchical initial convolution to obtain periodic feature sequences corresponding to each of the energy consumption types; The periodic feature sequences are spliced ​​and truncated in chronological order to obtain energy consumption feature vectors corresponding to each energy consumption type, wherein the time length of the energy consumption feature vectors is the same as that of the historical energy consumption data.

6. The energy consumption data upsampling method according to claim 5, characterized in that, The step of concatenating the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fused feature vector that simultaneously carries the periodic pattern of the historical energy consumption data and the energy consumption influencing factors includes: The core feature vector and the energy consumption feature vectors corresponding to each of the energy consumption types are concatenated in the time dimension to obtain the combined feature sequence corresponding to each of the energy consumption types. The combined feature sequences are mapped to fused feature sequences corresponding to each energy consumption type by a second multilayer perceptron, so that multiple fused feature sequences are used as the fused feature vector, wherein the time length of the fused feature vector is the same as that of the historical energy consumption data.

7. The energy consumption data upsampling method according to claim 1, characterized in that, The transfer learning training also includes: Using sample energy consumption data of various energy consumption types corresponding to the region to which the historical energy consumption data belongs as supervision signals, the weights and biases of the initial mapping layer are iteratively adjusted until the difference between the total energy consumption value of the predicted energy consumption data output by the linear mapping layer accumulated on the time axis and the total energy consumption value in the sample energy consumption data satisfies the preset convergence condition. In response to the difference satisfying the preset convergence condition, the adjusted initial mapping layer is used as the linear mapping layer; The initial mapping layer is trained from reference energy consumption data of various energy consumption types in the reference region during the historical period and reference impact data associated with multiple energy consumption impact factors during the historical period.

8. An energy consumption data upsampling system based on transfer learning, characterized in that, The system includes: The data reconstruction module is used to reconstruct historical energy consumption data of various energy consumption types within a preset time period into a multidimensional energy consumption feature tensor based on a preset period length. The multidimensional energy consumption feature tensor is used to characterize the temporal changes within the same period and the correlation changes between different periods in multiple periods divided based on the preset period length. The energy consumption types include coal, oil, and natural gas. The core extraction module is used to perform weighted pooling on the energy consumption impact data associated with the multiple energy consumption impact factors within the preset time period based on the preset correlation weights of multiple energy consumption impact factors for each energy consumption type, so as to select and aggregate the information of each energy consumption impact factor and obtain the core feature vector focusing on the energy consumption impact factor most related to the historical energy consumption data. The vector concatenation module is used to concatenate the core feature vector with the energy consumption feature vector obtained by feature extraction from the multidimensional energy consumption feature tensor to obtain a fusion feature vector that simultaneously carries the periodic pattern of the historical energy consumption data and the energy consumption influencing factor. The mapping output module is used to input the fused feature vector into the linear mapping layer for mapping and output target energy consumption data, so as to use the target energy consumption data for carbon emission monitoring and energy consumption prediction. The linear mapping layer is obtained by transfer learning training, which includes fine-tuning the pre-trained initial mapping layer using sample energy consumption data corresponding to the region to which the historical energy consumption data belongs.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more programs. The characteristic is that, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium having executable instructions stored thereon, characterized in that, When executed by the processor, this instruction causes the processor to implement the method of any one of claims 1 to 7.

Images