Building energy consumption prediction method and system based on deep feature fusion network

Abstract

The present invention relates to the technical field of energy consumption prediction, and discloses a method and system for predicting building energy consumption based on a deep feature fusion network. The method for predicting building energy consumption based on a deep feature fusion network of the present invention converts a one-dimensional sequence into a two-dimensional space graph according to a set time span, strengthens the periodic information of the building energy consumption sequence, and embeds a modified multi-scale convolution in the module, which is conducive to further mining the spatial information and time characteristics of the input sequence. At the same time, the embedded features of time factors, meteorological factors and historical energy consumption sequences are fused, which is conducive to capturing the correlation between various heterogeneous factors and their impact on future energy consumption, thereby extracting useful information from the influencing factors and avoiding interference of a large amount of redundant information on the prediction task.

Description

Technical Field

[0001] The present invention relates to the technical field of energy consumption prediction, and in particular to a method and system for predicting building energy consumption based on a deep feature fusion network. Background Art

[0002] Accurate energy consumption forecasting is crucial for power system design and planning, as well as for policy formulation, such as power distribution. It also assists with user profiling and anomaly detection. With the development of smart cities, big data has become a key feature of the era. Research on building energy consumption forecasting methods is no longer constrained by data volume, driving the development and application of data-driven models in this field. In recent years, neural network-based methods have garnered significant research attention, benefiting from their ability to model complex scenarios and adapt to diverse heterogeneous data. Artificial neural networks (ANNs), as a classic model, significantly outperform traditional statistical regression-based methods in building energy consumption forecasting and have been widely used. In recent years, researchers in this field have explored the application of deep network models, including convolutional neural networks (CNNs), long short-term memory (LSTMs), and their variant, bi-LSTMs. Some research has also constructed CNN-LSTMs, leveraging the local perception capabilities of CNNs and the temporal modeling capabilities of LSTMs to capture more information from the input sequence using deep network structures.

[0003] Furthermore, given that building energy consumption trends are significantly influenced by occupant behavior and exhibit regular variations, some studies have incorporated factors influencing occupant behavior (such as time and weather) as part of the model input. However, existing methods ignore the potential correlation between these factors and energy consumption, enriching the input features while introducing a large amount of interference information, limiting the improvement of prediction accuracy. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the problems of insufficient feature extraction and susceptibility to interference from redundant information in energy consumption prediction in the prior art.

[0005] To solve the above technical problems, the present invention provides a building energy consumption prediction method based on a deep feature fusion network, which includes the following steps:

[0006] S1. One-hot encode the historical time factor data and the future time factor data respectively, and generate and Among them, c t Indicates the number of time features after encoding, l i Indicates the length of the input sequence, l o Indicates the length of the predicted sequence;

[0007] S2. Normalize the continuous variables in the meteorological factor data to generate Perform one-hot encoding on the discrete variables in the meteorological factor data to generate Among them, c s and c d represent the number of continuous meteorological features and the number of discrete meteorological features respectively;

[0008] S3, T h 、T f 、W s 、W d Input the feature embedding network respectively and generate the time embedding feature f respectively th , time embedding feature f tf , meteorological embedded features f ws and meteorological embedding features f wd ;

[0009] S4. Normalize the historical energy consumption series data to generate a historical energy consumption series E;

[0010] S5. According to different time spans k, the historical energy consumption sequence E is converted into graphs of different sizes in two-dimensional space, and then the converted graphs are sent to the multi-scale convolutional network for feature extraction. Finally, the energy consumption embedding feature f is calculated through weighted merging operation. e ;

[0011] S6, connection feature f th 、f tf 、f ws 、f wd 、f e , and use the linear layer of the neural network to perform feature fusion to obtain feature F;

[0012] S7. Set the trainable weight parameters, perform weighted addition on the historical energy consumption sequence E and the feature F, and then input them into the multi-layer perceptron to obtain the final prediction result.

[0013] In one embodiment of the present invention, the feature embedding network is composed of a one-dimensional convolutional layer with a convolution kernel size of a×a and a fully connected layer FC, and each convolutional layer is followed by a ReLU activation function; wherein, a≥3.

[0014] In one embodiment of the present invention, step S5 includes:

[0015] According to different time spans k and preset graph heights H, the historical energy consumption sequence E is two-dimensionally transformed to generate a graph G. Then, the G generated based on different time spans is respectively sent to a multi-scale convolutional network consisting of a 1×1 convolutional layer, a 3×3 convolutional layer, and a pooling layer, where the pooling layer consists of an average pooling layer and a maximum pooling layer.

[0016] In one embodiment of the present invention, the historical energy consumption sequence E is subjected to a two-dimensional transformation to generate a graph G as follows:

[0017] G[i,:]=E[(i×k):(i×k+W)]

[0018] Where W = l i -H+1 is the width of the two-dimensional graph, and i×k+W≤l i During the conversion process, a sliding window is used to traverse the sequence. The time span determines the interval between each row of data in the two-dimensional graph, that is, the time difference between them.

[0019] In one embodiment of the present invention, the formula for the normalization process is as follows:

[0020]

[0021] Among them, x t Represents the input variable at time t, max and min represent the preset maximum and minimum values ​​respectively.

[0022] In one embodiment of the present invention, the time factors include month, hour, week, and day types; the discrete variables in the meteorological factor data include weather, outdoor temperature, relative humidity, relative humidity, and atmospheric pressure.

[0023] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the program.

[0024] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0025] The present invention also provides a processor, which is used to run a program, wherein the program executes any of the above methods when running.

[0026] The present invention also provides a building energy consumption prediction system based on a deep feature fusion network, which includes the following modules:

[0027] The time factor module is used to perform one-hot encoding on historical time factor data and future time factor data, and generate and Among them, c t Indicates the number of time features after encoding, l i Indicates the length of the input sequence, l o Indicates the predicted sequence length;

[0028] Meteorological factor module is used to normalize the continuous variables in meteorological factor data and generate Perform one-hot encoding on the discrete variables in the meteorological factor data to generate Among them, c s and cx represent the number of continuous and discrete meteorological features, respectively;

[0029] Time and weather embedding feature generation module is used to embed T h 、T f 、W s 、W d Input the feature embedding network respectively and generate the time embedding feature f respectively th , time embedding feature f tf , meteorological embedded features f ws and meteorological embedding features f wd ;

[0030] The historical energy consumption module is used to normalize the historical energy consumption series data and generate the historical energy consumption series E;

[0031] The multi-span time fusion module is used to convert the historical energy consumption sequence E into graphs of different sizes in two-dimensional space according to different time spans k, and then send the converted graphs into a multi-scale convolutional network for feature extraction. Finally, through the weighted merging operation, the energy consumption embedding feature f is calculated and generated. e ;

[0032] Feature fusion module, used to connect features f th , t ft 、f ws 、f wd 、f e , and use the linear layer of the neural network to perform feature fusion to obtain feature F;

[0033] The prediction module is used to set the trainable weight parameters, perform weighted addition of the historical energy consumption sequence E and the feature F, and then input them into the multi-layer perceptron to obtain the final prediction result.

[0034] The above technical solution of the present invention has the following advantages over the prior art:

[0035] The building energy consumption prediction method based on the deep feature fusion network in the present invention converts a one-dimensional sequence into a two-dimensional space graph according to the set time span, thereby strengthening the periodic information of the building energy consumption sequence. The modified multi-scale convolution is embedded in the module, which is conducive to further mining the spatial information and temporal characteristics of the input sequence.

[0036] At the same time, fusing the embedded features of time factors, meteorological factors and historical energy consumption series is conducive to capturing the correlation between various heterogeneous factors and their impact on future energy consumption, thereby extracting useful information from the influencing factors and avoiding the interference of a large amount of redundant information on the prediction task.

[0037] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following preferred embodiments are specifically cited and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein

[0039] Figure 1 is a flow chart of a building energy consumption prediction method based on a deep feature fusion network in an embodiment of the present invention;

[0040] Figure 2 is a schematic diagram of a building energy consumption prediction method based on a deep feature fusion network in an embodiment of the present invention;

[0041] Figure 3 is a structural diagram of a feature embedding network according to an embodiment of the present invention;

[0042] Figure 4 is a structural diagram of a multi-span time fusion module in an embodiment of the present invention;

[0043] Figure 5 is a schematic diagram of a pooling operation in an embodiment of the present invention;

[0044] Figure 6 Schematic diagram of the sequence conversion process in an embodiment of the present invention;

[0045] Figure 7 is the historical power consumption data of the experimental samples in the embodiment of the present invention;

[0046] Figure 8 This is a visualization of the test results of the building energy consumption prediction method based on the deep feature fusion network in an embodiment of the present invention on the sample. DETAILED DESCRIPTION

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0048] Example 1

[0049] Reference Figure 1-2 As shown, this embodiment discloses a building energy consumption prediction method based on a deep feature fusion network, which includes the following steps:

[0050] Step S1: One-hot encode the historical time factor data and the future time factor data, and generate and Among them, c t Indicates the number of time features after encoding, l i Indicates the length of the input sequence, l o Indicates the predicted sequence length;

[0051] Among them, the specific method of one-hot encoding is as follows:

[0052] To convert categorical variables into a form that is easily usable by the algorithm model, one-hot encoding is used to binarize discrete variables. Specifically, for a feature with n states, it is converted into a one-dimensional vector of length n. In each sample, only one bit of this vector is 1, and all other bits are 0, representing a state of the feature. For example, the day type can be defined as two states: working day and non-working day, encoded as (1, 0) and (0, 1), respectively.

[0053] Step S2: normalize the continuous variables in the meteorological factor data to generate Perform one-hot encoding on the discrete variables in the meteorological factor data to generate Among them, c s and c d represent the number of continuous meteorological features and the number of discrete meteorological features respectively;

[0054] Optionally, the normalization formula is as follows:

[0055]

[0056] Among them, x t Represents the input variable at time t, max and min represent the preset maximum and minimum values ​​respectively.

[0057] Optionally, referring to Table 1, the time factors include month, hour, week, day type, etc.; the discrete variables in the meteorological factor data include weather, outdoor temperature, relative humidity, relative humidity, atmospheric pressure, etc.

[0058] Table 1 Summary of input variables

[0059]

[0060] Step S3: T h 、Tf 、W s 、W d Input the feature embedding network EN respectively and generate the time embedding feature f respectively th , time embedding feature f tf , meteorological embedded features f ws and meteorological embedding features f wd ;

[0061] Reference Figure 3 Optionally, the feature embedding network consists of a one-dimensional convolution layer with a convolution kernel size of a×a and a fully connected layer FC, and each convolution layer is followed by a ReLU activation function; wherein, a≥3.

[0062] Step S4: Normalize the historical energy consumption sequence data to generate a historical energy consumption sequence The normalization formula may be as shown in the above step S2.

[0063] Step S5: According to different time spans k, the historical energy consumption sequence E is converted into graphs of different sizes in two-dimensional space, and then the converted graphs are fed into a multi-scale convolutional network for feature extraction. Finally, the energy consumption embedding feature f is calculated and generated through a weighted merging operation. e ;

[0064] In one embodiment, the structure diagram of the multi-span time fusion module for executing step S5 is shown in FIG. Figure 4 Step S5 specifically includes: performing a two-dimensional transformation on the historical energy consumption sequence E to generate a graph G according to different time spans k and preset graph heights H; then, the G generated based on different time spans is respectively fed into a multi-scale convolutional network consisting of a 1×1 convolutional layer, a 3×3 convolutional layer, and a pooling layer, wherein the pooling layer consists of an average pooling layer and a maximum pooling layer, which is used to perform a pooling operation, referring to Figure 5 .

[0065] Step S6: Connect feature f th 、f tf 、f ws 、f wd 、f e , generate (f th ,f tf ,f ws ,f wd ,f e ), and use the linear layer of the neural network to perform feature fusion to obtain the feature

[0066] In one embodiment of the present invention, the historical energy consumption sequence E is subjected to a two-dimensional transformation to generate a graph G as follows:

[0067] G[i,:]=E[(i×k):(i×k+W)]

[0068] Where W = l i -H+1 is the width of the two-dimensional graph, and i×k+W≤l i ; The conversion process is as follows Figure 6 As shown, taking the time span k1=1 and k2=2 as an example, a sliding window is used to traverse the sequence during the conversion process. The time span determines the interval between each row of data in the two-dimensional graph, that is, the time difference between them.

[0069] Step S7: Set the trainable weight parameter α, perform weighted addition on the historical energy consumption sequence E and the feature F, and then input them into the multi-layer perceptron to obtain the final prediction result.

[0070] The building energy consumption prediction method based on the deep feature fusion network in the present invention converts a one-dimensional sequence into a two-dimensional space graph according to the set time span, thereby strengthening the periodic information of the building energy consumption sequence. The modified multi-scale convolution is embedded in the module, which is conducive to further mining the spatial information and temporal characteristics of the input sequence.

[0071] At the same time, fusing the embedded features of time factors, meteorological factors and historical energy consumption series is conducive to capturing the correlation between various heterogeneous factors and their impact on future energy consumption, thereby extracting useful information from the influencing factors and avoiding the interference of a large amount of redundant information on the prediction task.

[0072] To verify the accuracy and robustness of the present invention, experiments were conducted based on the public dataset HUE (https: / / dataverse.harvard.edu / dataset.xhtml?persistentId=doi:10.7910 / DVN / N3HGRN).

[0073] The HUE dataset contains electricity data from 22 buildings in the Vancouver area and weather data from the nearest weather station. We selected the electricity data from 2018-11-01 to 2019-11-20 in Resident_2 and the weather data from the corresponding weather station in Vancouver as experimental samples to verify the model. The power load curve during this period is as follows: Figure 7 As shown. Among them, the maximum value of the outdoor temperature is max T =29.20℃, minimum value min T =-8.60℃, maximum relative humidity max RH =100%, minimum value min RH =18%, the maximum value of atmospheric pressure max P =103.94kPa, minimum value min P=98.15kPa; the maximum value of historical energy consumption E =1.09kWh, minimum value min T =0.09kWh; a=3; l i =168, k=(k1,k2,k3)=(1,12,24), H=72.

[0074] During training, we reconstructed the dataset using a sliding window with a one-hour sliding step size. The sliding window size was 168 + 24, so that the time span of the input historical series was 168 hours and the time span to be predicted was 24 hours. The experimental parameter settings are shown in Table 2. During training, the initial learning rate was 0.001 and was reduced to 0.0001 in the last 10 rounds.

[0075] Table 2 Experimental parameter settings

[0076] Number of training samples Number of test samples Learning rate Number of iterations 8124 732 0.001 40

[0077] Table 3 shows a comparative experiment conducted to verify the advantages of our method. The comparison targets were popular energy consumption prediction methods based on ANN, LSTM, CNN-LSTM, and CNN-BiLSTM. Mean absolute error (MAE) and root mean square error (RMSE) were selected as evaluation metrics. The proposed method achieved optimal performance on both metrics.

[0078] Table 3 Comparison of the present invention with other popular methods

[0079] method MAE RMSE The present invention 0.076 0.091 ANN 0.095 0.137 LSTM 0.099 0.147 CNN-LSTM 0.099 0.140 CNN-BiLSTM 0.103 0.142

[0080] Reference Figure 8 , which is a visualization of the test results of the building energy consumption prediction method based on the deep feature fusion network in an embodiment of the present invention on the sample. It can be seen that the error between the energy consumption value predicted by the method of the present invention and the true value is very small, and the prediction accuracy is very high.

[0081] Example 2

[0082] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the first embodiment when executing the program.

[0083] Example 3

[0084] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the method described in the first embodiment when the program is executed by a processor.

[0085] Example 4

[0086] The present invention also provides a processor, which is used to run a program, wherein the program executes the method described in embodiment 1 when running.

[0087] Example 5

[0088] This embodiment provides a building energy consumption prediction system based on a deep feature fusion network, which includes the following modules:

[0089] The time factor module is used to perform one-hot encoding on historical time factor data and future time factor data, and generate and Among them, c t Indicates the number of time features after encoding, l i Indicates the length of the input sequence, l o Indicates the length of the predicted sequence;

[0090] Meteorological factor module is used to normalize the continuous variables in meteorological factor data and generate Perform one-hot encoding on the discrete variables in the meteorological factor data to generate Among them, c s and c d represent the number of continuous meteorological features and the number of discrete meteorological features respectively;

[0091] Time and weather embedding feature generation module is used to embed T h 、T f 、W s 、W d Input the feature embedding network respectively and generate the time embedding feature f respectively th , time embedding feature f tf , meteorological embedded features f ws and meteorological embedding features f wd ;

[0092] The historical energy consumption module is used to normalize the historical energy consumption series data and generate the historical energy consumption series E;

[0093] The multi-span time fusion module is used to convert the historical energy consumption sequence E into graphs of different sizes in two-dimensional space according to different time spans k, and then send the converted graphs into a multi-scale convolutional network for feature extraction. Finally, through the weighted merging operation, the energy consumption embedding feature f is calculated and generated. e ;reference Figure 4 .

[0094] Feature fusion module, used to connect features f th 、f tf 、f ws 、f wd 、fe , and use the linear layer of the neural network to perform feature fusion to obtain feature F;

[0095] The prediction module is used to set the trainable weight parameters, perform weighted addition of the historical energy consumption sequence E and the feature F, and then input them into the multi-layer perceptron to obtain the final prediction result.

[0096] The building energy consumption prediction system based on the deep feature fusion network in the embodiment of the present invention is used to implement the aforementioned building energy consumption prediction method based on the deep feature fusion network. Therefore, the specific implementation method of the system can be seen in the embodiment part of the building energy consumption prediction method based on the deep feature fusion network in the previous article. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part, and will not be elaborated here.

[0097] In addition, since the building energy consumption prediction system based on deep feature fusion network in this embodiment is used to implement the aforementioned building energy consumption prediction method based on deep feature fusion network, its function corresponds to that of the above method and will not be repeated here.

[0098] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0099] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0100] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0102] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A building energy consumption prediction method based on a deep feature fusion network, characterized in that: The following steps are involved: S1. One-hot encode the historical time factor data and the future time factor data respectively, and generate and Among them, c t represents the number of time features after encoding, l i Indicates the length of the input sequence, l o Indicates the predicted sequence length; S2. Normalize the continuous variables in the meteorological factor data to generate One-hot encoding is performed on the discrete variables in the meteorological factor data to generate Among them, c s and c d They represent the number of continuous meteorological features and the number of discrete meteorological features respectively; S3, T h 、T f , W s , W d Input the feature embedding network respectively and generate the time embedding features f respectively th , time embedding feature f tf , meteorological embedded features f ws and meteorological embedding features f wd ; S4, normalizing the historical energy consumption series data to generate a historical energy consumption series E; S5. According to different time spans k, the historical energy consumption sequence E is converted into graphs of different sizes in two-dimensional space, and then the converted graphs are sent to the multi-scale convolutional network for feature extraction. Finally, the energy consumption embedding feature f is calculated through weighted merging operation. e ; According to different time spans k and preset graph heights H, the historical energy consumption sequence E is transformed into a two-dimensional graph G. Then, the graphs G generated based on different time spans are respectively sent to a multi-scale convolutional network consisting of a 1×1 convolutional layer, a 3×3 convolutional layer, and a pooling layer, where the pooling layer consists of an average pooling layer and a maximum pooling layer. The two-dimensional transformation of the historical energy consumption sequence E generates a graph G as follows: G[i,:]=E[(i×k):(i×k+W)] Where W = l i -H+1 is the width of the two-dimensional graph, and i×k+W≤l i ; During the conversion process, a sliding window is used to traverse the sequence. The time span determines the interval between each row of data in the two-dimensional graph, that is, the time difference between them; S6, connection feature f th 、f tf 、f ws 、f wd 、f e , and use the linear layer of the neural network to perform feature fusion to obtain feature F; S7. Set the trainable weight parameters, perform weighted addition on the historical energy consumption sequence E and the feature F, and then input them into the multi-layer perceptron to obtain the final prediction result.

2. The building energy consumption prediction method based on deep feature fusion network according to claim 1 is characterized in that: The feature embedding network consists of a one-dimensional convolutional layer with a convolution kernel size of a×a and a fully connected layer FC, and each convolutional layer is followed by a ReLU activation function; where a≥3.

3. The building energy consumption prediction method based on deep feature fusion network according to claim 1 is characterized in that: The formula for the normalization process is as follows: Among them, x t Represents the input variable at time t, max and min represent the preset maximum and minimum values ​​respectively.

4. The building energy consumption prediction method based on deep feature fusion network according to claim 1 is characterized in that: The time factors include month, hour, week, and day types; the discrete variables in the meteorological factor data include weather, outdoor temperature, relative humidity, relative humidity, and atmospheric pressure.

5. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

7. A processor, characterized in that: The processor is used to run a program, wherein the program executes the method according to any one of claims 1 to 4 when running.

8. Building energy consumption prediction system based on deep feature fusion network, characterized by: Includes the following modules: The time factor module is used to perform one-hot encoding on historical time factor data and future time factor data, and generate and Among them, c t represents the number of time features after encoding, l i Indicates the length of the input sequence, l o Indicates the predicted sequence length; The meteorological factor module is used to normalize the continuous variables in the meteorological factor data and generate One-hot encoding is performed on the discrete variables in the meteorological factor data to generate Among them, c s and c d They represent the number of continuous meteorological features and the number of discrete meteorological features respectively; The time and weather embedding feature generation module is used to embed T h 、T f , W s , W d Input the feature embedding network respectively and generate the time embedding features f respectively th , time embedding feature f tf , meteorological embedded features f ws and meteorological embedding features f wd ; The historical energy consumption module is used to normalize the historical energy consumption sequence data and generate the historical energy consumption sequence E; The multi-span time fusion module is used to convert the historical energy consumption sequence E into graphs of different sizes in two-dimensional space according to different time spans k, and then send the converted graphs to the multi-scale convolutional network for feature extraction. Finally, the energy consumption embedding feature f is calculated through weighted merging operations. e ; According to different time spans k and preset graph heights H, the historical energy consumption sequence E is transformed into a two-dimensional graph G. Then, the graphs G generated based on different time spans are respectively sent to a multi-scale convolutional network consisting of a 1×1 convolutional layer, a 3×3 convolutional layer, and a pooling layer, where the pooling layer consists of an average pooling layer and a maximum pooling layer. The two-dimensional transformation of the historical energy consumption sequence E generates a graph G as follows: G[i,:]=E[(i×k):(i×k+W)] Where W = l i -H+1 is the width of the two-dimensional graph, and i×k+W≤l i ; During the conversion process, a sliding window is used to traverse the sequence. The time span determines the interval between each row of data in the two-dimensional graph, that is, the time difference between them; Feature fusion module, used to connect the features f th 、f tf 、f ws 、f wd 、f e , and use the linear layer of the neural network to perform feature fusion to obtain feature F; The prediction module is used to set the trainable weight parameters, perform weighted addition of the historical energy consumption sequence E and the feature F, and then input them into the multi-layer perceptron to obtain the final prediction result.

Images