Automatic door and window fitting equipment energy consumption prediction method

Abstract

The invention provides an automatic door and window fitting equipment energy consumption prediction method, and relates to the technical field of machine learning. The invention provides an EnerNet prediction model, a periodic component coding module, a trend component processing module and a prediction module, the periodic component coding module is used for processing the incidence relation among modeling data periodicity, long-term trend and characteristics, and the trend component processing module is used for processing data long-term trend information. And the prediction module is used for fusing the prediction component detailed solutions output by the periodic component coding module and the trend component processing module to obtain a final energy consumption prediction result of the automatic door and window fitting equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting energy consumption of automated door and window fitting equipment. Background Art

[0002] With the promotion of the concept of green buildings, energy consumption prediction of building door and window automation equipment has become an important part of energy-saving management. The automation function of doors and windows in modern buildings not only improves the convenience of use, but also effectively improves the comfort of the indoor environment. However, due to the complex motor drive, sensor control and environmental response involved in the operation of the equipment, its energy consumption is significantly dynamic and random. Energy consumption prediction can help building managers optimize equipment use strategies and reduce unnecessary energy consumption losses, thereby achieving efficient utilization of energy resources, while helping to achieve the dual carbon goals of the construction industry.

[0003] In the energy management of automated door and window equipment, accurate energy consumption prediction is the key to reducing energy costs. Current equipment energy consumption analysis mostly relies on static or local data, which is difficult to adapt to scenarios with long time series and complex interactions of multiple variables. Especially in industrial applications, equipment needs to be highly coordinated with the production process, and energy demand shows cyclical fluctuations and trend changes. It is equally important to achieve lower computational complexity and higher prediction accuracy in long-sequence predictions, which can reduce the energy cost of equipment operation and improve overall operation and maintenance efficiency.

[0004] Automated door and window fitting equipment generates a large amount of sensor data during operation, involving multiple variables such as ambient temperature, humidity, wind speed, and equipment status. Although these data can help accurately monitor equipment performance, traditional energy consumption prediction methods often cannot fully explore the interactions between these complex variables, resulting in a lack of sufficient transparency and interpretability in the prediction results. Existing methods usually only rely on simple statistical models or basic regression analysis, which cannot effectively quantify the specific impact of each variable on energy consumption, and it is difficult to automatically identify features that are critical to prediction. Summary of the invention

[0005] The present invention provides an energy consumption prediction method for automatic door and window fitting equipment. Aiming at fishery aquaculture water environment quality related data with long sequences and multiple variables, an EnerNet prediction model is proposed, which consists of a periodic component encoding module, a trend component processing module and a prediction module.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose specifically includes the following steps: S1. Collect data related to energy consumption of automated door and window fitting equipment, including features and target variables, and pre-process the collected data; S2, using the Max-Min method to normalize the energy consumption data of the automatic door and window fitting equipment, and divide the data set proportionally; S3. Construct a periodic component encoding module to process the relationship between the periodicity, long-term trend and features of the modeling data, which specifically includes the following steps: S31. Input the energy consumption data X of the automatic door and window fitting equipment, and use the average pooling filter to extract the long-term trend X of the energy consumption data of the automatic door and window fitting equipment. t and the periodic trend quantity X s ; S32. An enhanced Transformer is proposed, which introduces a multi-head hierarchical attention mechanism through the feature embedding layer to obtain a high-dimensional representation of the periodic trend, and performs weighted fusion after local attention and global attention calculation, and splices them to obtain the final output H'; S33, process the final output through the feedforward network, use residual connection and layer normalization to ensure gradient stability, and finally obtain the periodic trend prediction component through the mapping layer S4. Construct a trend component processing module for processing long-term trend information of data, which specifically includes the following steps: S41. Input long-term trend quantity X t Mapping to high-dimensional space through feature embedding layer to obtain high-dimensional representation of long-term trends S42. Update the high-dimensional representation of the long-term trend layer by layer through the fully connected layer to obtain the updated high-dimensional representation of the long-term trend of the l+1th layer S43. Obtaining long-term trend prediction components through mapping layers And build a prediction module, add the periodic trend prediction component and the long-term trend prediction component results to get the final energy consumption prediction value of the automatic door and window fitting equipment

[0007] Preferably, in S1, the energy consumption data of the automatic door and window fitting equipment is collected, including motor power data, equipment speed data, load factor data, real-time energy consumption data, processing time data and cutting mode data, and the mean method is used to fill the missing values ​​of the energy consumption data of the automatic door and window fitting equipment. The specific formula is: Where, X ave is the imputed mean, n is the number of missing values, x i is a non-missing value.

[0008] Preferably, in S2, the Max-Min method is used to normalize the energy consumption data of the automatic door and window fitting equipment, and the training set and the test set are divided proportionally. The specific formula is: In the formula, x is the energy consumption data of the automatic doors and windows equipment, x min is the minimum value of the energy consumption data of the automatic doors and windows equipment, x max It is the maximum value among the energy consumption data of the automated doors and windows equipment.

[0009] Preferably, the energy consumption data of the automatic door and window fitting equipment is input in S3 and S31 as X={x1, ..., x T}∈R T×N , where T is the time step, N is the number of variables of the energy consumption data of the automatic doors and windows equipment, and the average pooling filter is used to extract the long-term trend and periodic trend of the energy consumption data of the automatic doors and windows equipment. The specific formula is: X t =F(X); X s =XX t ; Where, X t is the long-term trend quantity, X s is the periodic trend quantity, X is the energy consumption data of the automatic doors and windows equipment, and F(.) is the average pooling filter.

[0010] Preferably, the long-term trend quantity and the periodic trend quantity are extracted respectively through the average pooling filter to generate trend components, which provide a basis for subsequent encoding. The long-term trend quantity reflects the overall change trend of the data, and the periodic trend quantity captures the periodic change characteristics. The combination of the two enhances the modeling ability of the overall structure of the data. The trend extraction process significantly improves the model's ability to capture the trend changes of different features, and provides clear trend information for subsequent feature embedding and modeling.

[0011] Preferably, the enhanced Transformer is proposed in S3 and S32, and a multi-head hierarchical attention mechanism is introduced. First, the periodic trend quantity is mapped to a high-dimensional space through a feature embedding layer to obtain a high-dimensional representation of the periodic trend of n time steps inputted in the first layer. The specific formula is: Where, X s is the periodic trend quantity, Embedding is the operation of embedding each time step of the periodic trend quantity into the high-dimensional space, PositionalEncoding is the position encoding operation, which is used to retain the timing information of the periodic trend quantity. s The local attention calculation is performed on each local window of , and the specific formula is: In the formula, Q h is the query vector, is the transpose of the key vector, V h is the value vector, Softmax is the normalization function, M local is the local mask matrix, d k is the dimension of the key vector, and then the dependency of the key positions in the periodic trend volume is calculated through global attention. The specific formula is: Where M global is the global sparse mask matrix. Finally, the local and global attention are weighted fused and the fused attention is calculated. The specific formula is: HierarchicalAttention h (Q,K,V)=α h LocalAttention h (Q,K,V)+(1-α h )·GlobalAttention h (Q, K, V); Where LocalAttention h (Q, K, V) is the local attention calculation output, GlobalAttention h (Q, K, V) global attention calculation output, α h is the learnable weight of local and global attention. The hierarchical attention results of all heads are concatenated and the final output is obtained through linear transformation. The specific formula is: H'=Concat(HierarchicalAttention1,...,HierarchicalAttention H )W o ; In the formula, Concat is a concatenation operation, HierarchicalAttention H is the hierarchical attention result of the Hth attention head, W o is the linear transformation matrix used to integrate the results of multiple heads.

[0012] Preferably, a multi-head hierarchical attention mechanism is introduced into the traditional Transformer. After mapping the feature embedding layer to a high-dimensional space, the weighted fusion features are calculated through local and global attention. The multi-head hierarchical attention mechanism can capture local and global dependencies, make full use of local changes and global patterns in periodic features, and significantly improve the ability to understand nonlinear relationships in complex time series.

[0013] Preferably, in S3 and S33, the linear transformation is further processed by a feedforward network to obtain a final output H, and a feature extraction output is obtained. The specific formula is: FFN(H')=ReLU(H'W1+b1)W2+b2; In the formula, W1 and W2 are the weight matrices of the feedforward network, b1 and b2 are bias terms, H' is the linear transformation to obtain the final output, ReLU is the activation function, and then residual connections and layer normalization are added to each layer to ensure gradient stability. The specific formula is: In the formula, is the high-dimensional representation of the periodic trend of the lth layer, LayerNorm is the layer normalization operation, HierarchicalAttention is the hierarchical attention calculation operation, FFN is the feedforward network operation, and It is the residual connection part, and finally the periodic trend prediction component is obtained through the mapping layer. The specific formula is: In the formula, Projection s For mapping layer operations, is the high-dimensional representation of the last layer of n time-step periodic trends, It is the periodic trend prediction component.

[0014] Preferably, a feedforward network is introduced to perform nonlinear processing on periodic features, and residual connections and layer normalization are used to ensure gradient stability. The feedforward network enhances the nonlinear expression ability of features, and residual connections and layer normalization improve the stability of model training, avoid the gradient vanishing problem in deep network training, and enhance the prediction accuracy of periodic trend components.

[0015] Preferably, the long-term trend quantity X is input in S4 and S41. t Through the feature embedding layer mapping to the high-dimensional space, the high-dimensional representation of the long-term trend is obtained. The specific formula is: Where, X t is the long-term trend quantity, Embedding is the operation of embedding each time step of the long-term trend quantity into the high-dimensional space, and PositionalEncoding is the position encoding operation, which is used to retain the time series information of the long-term trend quantity.

[0016] Preferably, the long-term trend quantity is mapped to a high-dimensional space through a feature embedding layer, and the time series information is retained, which can more fully capture the long-term trend characteristics and enhance the understanding of time dependence. The high-dimensional representation of long-term trends significantly improves the model's ability to capture long-term dependencies and provides a solid foundation for subsequent processing.

[0017] Preferably, in S4 and S42, the high-dimensional representation of the long-term trend is updated layer by layer through the fully connected layer, and the specific formula is: H” t =RuLU(H' t ); Where W3 and W4 are weight matrices of linear transformation, is the high-dimensional representation of the long-term trend of the lth layer, b3 and b4 are bias terms, ReLU is the activation function, and H' t is the output feature representation of the first layer linear transformation, H” t It is the feature representation of the first layer after the activation function, and LayerNorm is the layer normalization operation.

[0018] Preferably, the high-dimensional representation of the long-term trend is updated layer by layer through the fully connected layer, and the activation function and layer normalization are used to improve the nonlinear expression ability and model stability. The layer-by-layer update mechanism can gradually optimize the long-term trend representation, ensuring that the final high-dimensional representation contains more accurate long-term features, enhancing the feature refinement capability, and making the prediction component of the long-term trend more accurate.

[0019] Preferably, in said S4 and S43, the long-term trend prediction component is obtained through the mapping layer, and the specific formula is: In the formula, Projection s For mapping layer operations, is the high-dimensional representation of the long-term trend at the l+1th layer, is the long-term trend prediction component. The final energy consumption prediction value of the automatic door and window fitting equipment is obtained by adding the periodic trend prediction component and the long-term trend prediction component. The specific formula is: In the formula, is the periodic trend forecast component, Provide energy consumption forecast for the final automated doors and windows.

[0020] Preferably, the high-dimensional representation of the long-term trend is converted into a prediction component through a mapping layer and added to the periodic trend component to generate the final prediction result. The fusion of trend components ensures that the model can comprehensively consider periodic and long-term characteristics, thereby improving the comprehensiveness of the prediction and significantly improving the overall accuracy and stability of the prediction results.

[0021] In summary, due to the adoption of the present technical solution, the beneficial effects of the present invention are as follows: the present invention proposes an EnerNet prediction model, which is applied to the energy consumption prediction scenario of automatic door and window fitting equipment, and includes a periodic component encoding module, a trend component processing module and a prediction module. The periodic component encoding module is used to process the correlation between the periodicity, long-term trends and features of the modeling data, the trend component processing module is used to process the long-term trend information of the data, and the prediction module is used to fuse the prediction components output by the periodic component encoding module and the trend component processing module to obtain the final energy consumption prediction result of the automatic door and window fitting equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A step diagram of a method for predicting energy consumption of automated door and window fitting equipment.

[0023] Figure 2 This is the structure diagram of the EnerNet prediction model.

[0024] Figure 3 The EnerNet prediction model realizes the energy consumption prediction fitting effect diagram of automated door and window fitting equipment. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0026] See also Figure 1-Figure 3 The present invention provides a technical solution: a method for predicting energy consumption of automatic door and window fitting equipment, by proposing an EnerNet prediction model, constructing a periodic component encoding module for processing data long-term trend information, constructing a trend component processing module for processing data long-term trend information, constructing a prediction module for fusing the prediction components output by the periodic component encoding module and the trend component processing module to obtain the final automatic door and window fitting equipment energy consumption prediction result, the specific steps are as follows Figure 1 shown.

[0027] Construct the EnerNet prediction model, whose structure is as follows Figure 2As shown, the specific steps are as follows:

[0028] S1. Collect energy consumption data of automated door and window equipment, including characteristics and target variables, and pre-process the collected data.

[0029] Furthermore, the energy consumption data of the automatic door and window cutting equipment are collected, including motor power data, equipment speed data, load factor data, real-time energy consumption data, processing time data and cutting mode data. The mean method is used to fill the missing values ​​of the energy consumption data of the automatic door and window cutting equipment. The specific formula is: Where, X ave is the imputed mean, n is the number of missing values, x i is a non-missing value.

[0030] S2. The Max-Min method is used to normalize the energy consumption data of the automated door and window fitting equipment and divide the data set proportionally.

[0031] Furthermore, the Max-Min method is used to normalize the energy consumption data of the automatic door and window fitting equipment, and the training set and the test set are divided proportionally. The specific formula is: In the formula, x is the energy consumption data of the automatic doors and windows equipment, x min is the minimum value of the energy consumption data of the automatic doors and windows equipment, x max It is the maximum value among the energy consumption data of the automated doors and windows equipment.

[0032] S31. Input the energy consumption data X of the automatic door and window fitting equipment, and use the average pooling filter to extract the long-term trend X of the energy consumption data of the automatic door and window fitting equipment. t and the periodic trend quantity X s .

[0033] Furthermore, the energy consumption data of the automatic door and window fitting equipment is inputted as X={x1,…,x T}∈R T×N , where T is the time step, N is the number of variables of the energy consumption data of the automatic doors and windows equipment, and the average pooling filter is used to extract the long-term trend and periodic trend of the energy consumption data of the automatic doors and windows equipment. The specific formula is: X t =F(X); X s =XX t ; Where, X t is the long-term trend quantity, X sis the periodic trend quantity, X is the energy consumption data of the automatic doors and windows equipment, and F(.) is the average pooling filter.

[0034] S32. An enhanced Transformer is proposed, which introduces a multi-head hierarchical attention mechanism through the feature embedding layer to obtain a high-dimensional representation of the periodic trend, and performs weighted fusion after local attention and global attention calculations, and concatenates them to obtain the final output H'.

[0035] Furthermore, an enhanced Transformer is proposed, which introduces a multi-head hierarchical attention mechanism. First, the periodic trend quantity is mapped to a high-dimensional space through a feature embedding layer to obtain a high-dimensional representation of the periodic trend of n time steps input to the first layer. The specific formula is: Where, X s is the periodic trend quantity, Embedding is the operation of embedding each time step of the periodic trend quantity into the high-dimensional space, PositionalEncoding is the position encoding operation, which is used to retain the timing information of the periodic trend quantity. s The local attention calculation is performed on each local window of , and the specific formula is: In the formula, Q h is the query vector, is the transpose of the key vector, V h is the value vector, Softmax is the normalization function, M local is the local mask matrix, d k is the dimension of the key vector, and then the dependency of the key positions in the periodic trend volume is calculated through global attention. The specific formula is: Where M global is the global sparse mask matrix. Finally, the local and global attention are weighted fused and the fused attention is calculated. The specific formula is: HierarchicalAttention h (Q,K,V)=α h LocalAttention h (Q,K,V)+(1-α h )·GlobalAttention h (Q, K, V); Where LocalAttention h (Q, K, V) is the local attention calculation output, GlobalAttention h (Q, K, V) global attention calculation output, αh is the learnable weight of local and global attention. The hierarchical attention results of all heads are concatenated and the final output is obtained through linear transformation. The specific formula is: H'=Concat(HierarchicalAttention1,...,HierarchicalAttention H )W o ; In the formula, Concat is a concatenation operation, HierarchicalAttention H is the hierarchical attention result of the Hth attention head, W o is the linear transformation matrix used to integrate the results of multiple heads.

[0036] S33, process the final output through the feedforward network, use residual connection and layer normalization to ensure gradient stability, and finally obtain the periodic trend prediction component through the mapping layer

[0037] Furthermore, the linear transformation is further processed by the feedforward network to obtain the final output H', and the feature extraction output is obtained. The specific formula is: FFN(H')=ReLU(H'W1+b1)W2+b2; In the formula, W1 and W2 are the weight matrices of the feedforward network, b1 and b2 are bias terms, H' is the linear transformation to obtain the final output, ReLU is the activation function, and then residual connections and layer normalization are added to each layer to ensure gradient stability. The specific formula is: In the formula, is the high-dimensional representation of the periodic trend of the lth layer, LaterNorm is the layer normalization operation, HierarchicalAttention is the hierarchical attention calculation operation, FFN is the feedforward network operation, and It is the residual connection part, and finally the periodic trend prediction component is obtained through the mapping layer. The specific formula is: In the formula, Proojection s For mapping layer operations, is the high-dimensional representation of the last layer of n time-step periodic trends, It is the periodic trend prediction component.

[0038] S41. Input long-term trend quantity X t Mapping to high-dimensional space through feature embedding layer to obtain high-dimensional representation of long-term trends

[0039] Further, input the long-term trend quantity X t Through the feature embedding layer mapping to the high-dimensional space, the high-dimensional representation of the long-term trend is obtained. The specific formula is: Where, X t is the long-term trend quantity, Embedding is the operation of embedding each time step of the long-term trend quantity into the high-dimensional space, and PositionalEncoding is the position encoding operation, which is used to retain the time series information of the long-term trend quantity.

[0040] S42. Update the high-dimensional representation of the long-term trend layer by layer through the fully connected layer to obtain the updated high-dimensional representation of the long-term trend of the l+1th layer

[0041] Furthermore, the high-dimensional representation of the long-term trend is updated layer by layer through the fully connected layer. The specific formula is: H” t =RuLU(H' t ); Where W3 and W4 are weight matrices of linear transformation, is the high-dimensional representation of the long-term trend of the lth layer, b3 and b4 are bias terms, ReLU is the activation function, and H' t is the output feature representation of the first layer linear transformation, H” t It is the feature representation of the first layer after the activation function, and LayerNorm is the layer normalization operation.

[0042] S43. Obtaining long-term trend prediction components through mapping layers And build a prediction module, add the periodic trend prediction component and the long-term trend prediction component results to get the final energy consumption prediction value of the automatic door and window fitting equipment

[0043] Furthermore, the long-term trend prediction component is obtained through the mapping layer, and the specific formula is: In the formula, Projection s For mapping layer operations, is the high-dimensional representation of the long-term trend at the l+1th layer, is the long-term trend prediction component. The final energy consumption prediction value of the automatic door and window fitting equipment is obtained by adding the periodic trend prediction component and the long-term trend prediction component. The specific formula is: In the formula, is the periodic trend forecast component, Provide energy consumption forecast for the final automated doors and windows.

[0044] Furthermore, the EnerNet prediction model is written in Python, the experiment runs on the Windows operating system, Pytorch is selected as the framework in the CUDA11.27 environment, and training is performed on the GeForce RTX 3090. The optimizer uses Adam, the initial learning rate is set to 0.001, the training batch is set to 64, the training cycle is set to 100, and the data set is 90 days of automated door and window equipment energy consumption related data, which is input into the EnerNet prediction model after preprocessing.

[0045] Furthermore, the EnerNet prediction model realizes the energy consumption prediction fitting effect of automatic door and window fitting equipment. Figure 3 As shown in the figure, the horizontal axis is time (days), the vertical axis is equipment energy consumption (tce, tons of standard coal), the black solid line represents the actual energy consumption data, and the gray dotted line represents the model prediction data. It can be seen from the figure that the model can better follow the overall fluctuation trend of the actual energy consumption, especially in the stage with large fluctuations, such as the 5th to the 15th day and the 20th to the 30th day. The predicted value of the model can accurately reflect the rising and falling trends of energy consumption, indicating that the model has a strong ability to capture the overall trend. In the stable fluctuation stage (such as the 15th to the 20th day), the model can also better predict the energy consumption changes of the equipment, indicating that its performance in the stable area is more reliable. The experiment predicted a time range of 30 days. The model can maintain a high prediction accuracy throughout the entire time period. Whether in the rapid fluctuation stage or the relatively stable stage, it shows good adaptability, which shows that the model has a strong ability to deal with long time series prediction problems.

Claims

1. A method for predicting energy consumption of automated door and window fitting equipment, characterized in that: The following steps are involved: S1. Collect data related to energy consumption of automated door and window fitting equipment, including features and target variables, and pre-process the collected data; S2, using the Max-Min method to normalize the energy consumption data of the automatic door and window fitting equipment, and divide the data set proportionally; S3. Construct a periodic component encoding module to process the relationship between the periodicity, long-term trend and features of the modeling data, which specifically includes the following steps: S31. Input the energy consumption data X of the automatic door and window fitting equipment, and use the average pooling filter to extract the long-term trend X of the energy consumption data of the automatic door and window fitting equipment. t and the periodic trend quantity X s ; S32. An enhanced Transformer is proposed, which introduces a multi-head hierarchical attention mechanism through the feature embedding layer to obtain a high-dimensional representation of the periodic trend, and performs weighted fusion after local attention and global attention calculation, and splices them to obtain the final output H'; S33, process the final output through the feedforward network, use residual connection and layer normalization to ensure gradient stability, and finally obtain the periodic trend prediction component through the mapping layer S4. Construct a trend component processing module for processing long-term trend information of data, which specifically includes the following steps: S41. Input long-term trend quantity X t Mapping to high-dimensional space through feature embedding layer to obtain high-dimensional representation of long-term trends S42. Update the high-dimensional representation of the long-term trend layer by layer through the fully connected layer to obtain the updated high-dimensional representation of the long-term trend of the l+1th layer S43. Obtaining long-term trend prediction components through mapping layers And build a prediction module, add the periodic trend prediction component and the long-term trend prediction component results to get the final energy consumption prediction value of the automatic door and window fitting equipment 2. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 1, characterized in that: In S3 and S31, the energy consumption data of the automatic door and window fitting equipment is inputted X={x1, ..., x T }∈R T×N , where T is the time step, N is the number of variables of the energy consumption data of the automatic doors and windows equipment, and the average pooling filter is used to extract the long-term trend and periodic trend of the energy consumption data of the automatic doors and windows equipment. The specific formula is: X t =F(X); X s =X-X t ; Where, X t is the long-term trend quantity, X s is the periodic trend quantity, X is the energy consumption data of the automatic doors and windows equipment, and F(.) is the average pooling filter.

3. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 2, characterized in that: The enhanced Transformer is proposed in S3 and S32, and a multi-head hierarchical attention mechanism is introduced. First, the periodic trend quantity is mapped to a high-dimensional space through a feature embedding layer to obtain a high-dimensional representation of the periodic trend of n time steps inputted in the first layer. The specific formula is: Where, X s is the periodic trend quantity, Embedding is the operation of embedding each time step of the periodic trend quantity into the high-dimensional space, PositionalEncoding is the position encoding operation, which is used to retain the timing information of the periodic trend quantity. s The local attention calculation is performed on each local window of , and the specific formula is: In the formula, Q h is the query vector, is the transpose of the key vector, V h is the value vector, Softmax is the normalization function, M local is the local mask matrix, d k is the dimension of the key vector, and then the dependency of the key positions in the periodic trend volume is calculated through global attention. The specific formula is: Where M global is the global sparse mask matrix. Finally, the local and global attention are weighted fused and the fused attention is calculated. The specific formula is: HierarchicalAttention h (Q,K,V)=α h ·LocalAttention h (Q,K,V)+(1-α h )·GlobalAttention h (Q,K,V); Where LocalAttention h (Q, K, V) is the local attention calculation output, GlobalAttention h (Q, K, V) global attention calculation output, α h is the learnable weight of local and global attention. The hierarchical attention results of all heads are concatenated and the final output is obtained through linear transformation. The specific formula is: H’=Concat(HierarchicalAttention1,…,HierarchicalAttention H )W o ; In the formula, Concat is a concatenation operation, HierarchicalAttention H is the hierarchical attention result of the Hth attention head, W o is the linear transformation matrix used to integrate the results of multiple heads.

4. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 3, characterized in that: In S3 and S33, the linear transformation is further processed by the feedforward network to obtain the final output H', and the feature extraction output is obtained. The specific formula is: FFN(H') = ReLU(H'W1+b1)W2+b2; In the formula, W1 and W2 are the weight matrices of the feedforward network, b1 and b2 are bias terms, H' is the linear transformation to obtain the final output, ReLU is the activation function, and then residual connections and layer normalization are added to each layer to ensure gradient stability. The specific formula is: In the formula, is the high-dimensional representation of the periodic trend of the lth layer, LayerNorm is the layer normalization operation, HierarchicalAttention is the hierarchical attention calculation operation, FFN is the feedforward network operation, and It is the residual connection part, and finally the periodic trend prediction component is obtained through the mapping layer. The specific formula is: In the formula, Projection s For mapping layer operations, is the high-dimensional representation of the last layer of n time-step periodic trends, It is the periodic trend prediction component.

5. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 4, characterized in that: In S4 and S41, the long-term trend quantity X is input. t Through the feature embedding layer mapping to the high-dimensional space, the high-dimensional representation of the long-term trend is obtained. The specific formula is: Where, X t is the long-term trend quantity, Embedding is the operation of embedding each time step of the long-term trend quantity into the high-dimensional space, and PositionalEncoding is the position encoding operation, which is used to retain the time series information of the long-term trend quantity.

6. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 5, characterized in that: In S4 and S42, the high-dimensional representation of the long-term trend is updated layer by layer through the fully connected layer. The specific formula is: H" t =RuLU(H' t ); Where W3 and W4 are weight matrices of linear transformation, is the high-dimensional representation of the long-term trend of the lth layer, b3 and b4 are bias terms, ReLU is the activation function, and H' t is the output feature representation of the first layer linear transformation, H” t It is the feature representation of the first layer after the activation function, and LayerNorm is the layer normalization operation.

7. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 6, characterized in that: In S4 and S43, the long-term trend prediction component is obtained through the mapping layer, and the specific formula is: In the formula, Projection s For mapping layer operations, is the high-dimensional representation of the long-term trend at the l+1th layer, is the long-term trend prediction component. The final energy consumption prediction value of the automatic door and window fitting equipment is obtained by adding the periodic trend prediction component and the long-term trend prediction component. The specific formula is: In the formula, is the periodic trend forecast component, Provide energy consumption forecast for the final automated doors and windows.

8. The method for predicting energy consumption of automatic door and window fitting equipment according to claim 1, characterized in that: In order to predict the energy consumption of automatic door and window fitting equipment, the energy consumption data of automatic door and window fitting equipment were collected, including motor power data, equipment speed data, load factor data, real-time energy consumption data, processing time data and cutting mode data. The collected relevant data were preprocessed to ensure that there were no missing values ​​in the data. Then the processed data were divided into training set and test set for training and evaluating the energy consumption prediction model of automatic door and window fitting equipment.

Images