Method for constructing explainable urban building energy consumption model integrating inter-building effect

Abstract

The invention relates to a method for constructing an explainable urban building energy consumption model integrating inter-building effects, which comprises the following steps: collecting building electricity consumption information, carrying out data processing on missing values and abnormal values in original data, and constructing a feature project according to the electricity consumption of a time window; constructing a dynamic architectural drawing by taking the buildings as nodes in the architectural drawing and the inter-architectural effect as edges in the architectural drawing in the time window; constructing an encoder-decoder-based depth model taking the LSTM as a unit as a baseline model; on the basis of the baseline model, integrating the graph neural network, the attention mechanism and the LSTM to construct a GAT-LSTM model and training the GAT-LSTM model; and outputting an attention calculation result, explaining the black box model by adopting a clustering algorithm, distinguishing different modes of the inter-building effect and a distribution rule of the inter-building effect along with time, and further identifying key buildings in different modes.

Description

Technical Field

[0001] The invention belongs to the technical field of urban building energy conservation, and in particular relates to a method for constructing an interpretable urban building energy consumption model integrating inter-building effects. Background Art

[0002] At present, the energy consumption of urban buildings accounts for more than 60% of the world's total energy consumption. Predicting changes in urban building energy consumption and exploring the laws of its energy consumption are important ways to promote green, energy-saving and sustainable development of cities. The urban building energy consumption model has the functions of predicting and analyzing urban energy consumption, planning and controlling urban energy, etc., which is conducive to the formulation and implementation of energy-saving measures. It is a key technology to improve the energy performance of urban buildings. The urban building energy consumption model is based on the dynamic and complex feedback relationship between buildings and the urban environment, and expands the single building energy consumption model to the entire urban environment for energy performance simulation. The inter-building effect is particularly critical to the impact of building energy consumption.

[0003] Urban building energy consumption models can be divided into two types: "top-down" and "bottom-up". Although the "top-down" model only needs to input a small amount of data to obtain large-scale energy consumption prediction results, it lacks building space or time details, and it is difficult to go deep into the building unit level for specific analysis. In recent years, with the deepening of research, more and more scholars have begun to turn to the "bottom-up" method. The "bottom-up" model includes three types: data-driven, physical simulation and hybrid models. Although the physical simulation model can clearly calculate the energy consumption of the building, it requires a lot of computing resources, time and basic data input. The data-driven model is based on mathematical principles and historical data, and can quickly predict the energy consumption of the building. Therefore, many scholars use various time series prediction models to carry out urban building energy consumption prediction, such as SARIMA, SVM, RF, XGBoost, RNN, ANN, etc.

[0004] Although these time series algorithms can achieve rapid energy consumption forecasting, the complexity of the urban environment and the diversity of factors affecting building energy consumption, especially the existence of inter-building effects, result in data-driven models also having application limitations. Therefore, how to characterize and quantify inter-building effects and effectively integrate them into data-driven building energy consumption models still remains a technical challenge.

[0005] In addition, an inherent challenge facing data-driven models is their interpretability problem. Data-driven models, also known as black-box models, achieve high-accuracy predictions by fitting the relationship between input features and target values, but the internal decision-making mechanism and prediction results are difficult to interpret, which limits their practical application scope. Summary of the invention

[0006] The present invention aims to provide a method for constructing an interpretable urban building energy consumption model that integrates inter-building effects. The method is based on graph neural networks, attention mechanisms, and time series prediction algorithms, and aims to solve the technical problem of integrating inter-building effects into data-driven urban building energy consumption models. Furthermore, the interpretability of the black box model is significantly improved through the model's attention mechanism and clustering algorithm.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows: a method for constructing an interpretable urban building energy consumption model integrating inter-building effects, characterized in that the method comprises the following steps:

[0008] S1: Collect building electricity consumption information, process missing values ​​and outliers in the original data, and construct feature engineering based on electricity consumption in the time window;

[0009] S2: Construct a dynamic building graph within the time window, using buildings as nodes in the building graph and inter-building effects as edges in the building graph;

[0010] S3: Build an encoder-decoder based deep model with LSTM as the unit as the baseline model;

[0011] S4: Based on the baseline model, the GAT-LSTM model is constructed and trained by integrating graph neural network, attention mechanism and LSTM.

[0012] S5: Output the attention calculation results and use clustering algorithms to interpret the black box model, visualizing the different patterns of inter-building effects and their distribution over time.

[0013] Among them, S1 collects building electricity consumption information, processes missing values ​​and outliers in the original data, and constructs feature engineering based on electricity consumption in the time window, as follows:

[0014] S1.1: Collect building electricity consumption information through energy-saving platforms, sensors and field survey methods;

[0015] S1.2: Linear interpolation is used to supplement missing values ​​to ensure the integrity and consistency of the data. The interpolation calculation formula is as follows:

[0016]

[0017] Where y is the interpolated value of the unknown data point, y1 and y2 are the known data points, and x1 and x2 are the known x values ​​of the unknown data points;

[0018] S1.3: Draw a box plot of the data after missing values ​​are supplemented, and mark the data outside the range of Q_L-1.5IQR and Q_U+1.5IQR in the box plot as outliers, treat them as missing values, and use linear interpolation to supplement them. U , Q L , IQR are the upper and lower quartiles and interquartile range of the data set respectively;

[0019] S1.4: In order to eliminate the adverse effects caused by the large difference in electricity consumption between different buildings and ensure the effectiveness and stability of model training, the present invention adopts the standard deviation standardization method. The standard deviation standardization calculation formula is as follows:

[0020]

[0021] where x i For each feature in the dataset X, μ is the mean of the dataset, σ is the standard deviation of the dataset, and z i is the normalized value for each feature;

[0022] S1.5: Feature engineering of building graph nodes using time window sliding.

[0023] Among them, S2, in the time window, uses buildings as nodes in the building graph and uses the effects between buildings as edges in the building graph, thereby constructing a dynamic building graph that integrates the effects between buildings. This step innovatively transforms the buildings and their mutual influence effects into a dynamic building graph as input data for model training, thereby more accurately capturing and reflecting the actual impact of the complex relationship between buildings on the urban building energy consumption model. The details are as follows,

[0024] S2.1: Each building is a node in the building diagram. The power consumption in a single time window is used as the building node feature. The full connection between building nodes is used to characterize the inter-building effect. The relevant formula is as follows:

[0025] G=(V,E)

[0026] V={v i |i∈{1, 2, ..., n}}

[0027] E={(v i , v j )|{v i , v j ∈V}}

[0028]

[0029] A ij =1,(v i , v j )∈E and i≠j

[0030] Among them, G, V, E, X∈R F Represents graph, node, edge and node feature vector set respectively; n is the number of nodes in the graph, v i , v j is a node in the graph, is node v i The initial feature vector of F is the initial node feature dimension; the adjacency matrix A∈R n×n A is used to represent the relationship between nodes in the graph. ij Represents the value in the adjacency matrix. Since the initial building graph is constructed in a fully connected manner, A ij Then it is 1;

[0031] S2.2: Construct a static building map in each time window, and construct a continuous dynamic building map as the time window slides.

[0032] Among them, S3, builds a deep model based on encoder-decoder with LSTM as the unit as the baseline model: the baseline model is a 4-layer LSTM model and a 1-layer fully connected layer with Relu as the activation function, the first 2 layers of LSTM model are feature encoders (Encoder), and the last 2 layers of LSTM model and 1 layer of fully connected layer are reconstruction encoders (Decoder);

[0033] Among them, S4, based on the baseline model, integrates the graph neural network and attention mechanism to build and train the GAT-LSTM model. A graph convolution layer with a multi-head attention mechanism is added before the baseline model to calculate the dynamic building graph constructed in step S2. During the training process, the attention value is continuously calculated and the adjacency matrix is ​​updated. By introducing the edge attention learning mechanism to characterize the mutual influence between buildings, this step can accurately distinguish the contribution of different buildings to the energy consumption prediction of the target building, thereby significantly improving the accuracy of energy consumption prediction. The details are as follows,

[0034] S4.1: Construct a GAT-LSTM model based on graph neural network, attention mechanism and LSTM. The graph convolution layer with multi-head attention mechanism first calculates the attention value between different nodes and aggregates the feature information of adjacent building nodes. The specific formula is as follows:

[0035]

[0036] Among them, W is the weight matrix, which maps the original features to the high-dimensional space; is the shared self-attention performed on the node; e ij yes and The inner product value of two vectors, representing v j Node feature pair v iThe importance of the node; n i Finger and v i The set of node numbers with edge connections. Refers to the concatenation operation of the feature vector; K is the number of attention heads in the multi-head attention mechanism. In order to make the e between different pairs of nodes ij The coefficients can be easily compared. The softmax function is used to normalize the importance calculation results of all selected values ​​of j to X′.

[0037] S4.2: In order to enable the aggregated information to be operated by the encoder-decoder based deep model with LSTM as the unit, this method constructs a sequence creation module, which creates the convolution X′ into continuous three-dimensional data for input into the baseline model;

[0038] S4.3: The Adam optimizer is used to train the baseline model and GAT-LSTM model constructed in steps S3 and S4.1 respectively with mean square error (MSE) as the loss function, and the loss is back-propagated to adjust the hyperparameters.

[0039] Among them, S5 outputs the attention calculation results and uses the clustering algorithm to interpret the black box model, visualizing the different patterns of inter-building effects and their distribution over time, as follows:

[0040] S5.1: Derive the attention calculation results and use clustering algorithms to divide different types of inter-building effect patterns; S5.2: Use an electronic device to visualize different patterns of inter-building effects and their distribution patterns over time, and mine key influencing buildings in different patterns. This technology significantly enhances the interpretability of black box models, thereby promoting their wider application and providing strong technical support for scientific management and control of energy supply.

[0041] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, a method for constructing an interpretable urban building energy consumption model integrating inter-building effects is implemented.

[0042] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement a method for constructing an interpretable urban building energy consumption model integrating inter-building effects.

[0043] Compared with the prior art, the present invention has the following advantages: through experiments, when the same information is input, the prediction index R of the model constructed by the present method is 2The performance was improved by 9.75%, the MAE index was improved by 27.38%, and the RMSE index was improved by 24.92%. At the same time, this method proposed a device containing multiple algorithmic procedures to improve the interpretability of the black box model, visualize the effect patterns between different types of buildings and their distribution over time, and identify important buildings under different modes. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of a method for constructing an interpretable urban building energy consumption model integrating inter-building effects of the present invention.

[0045] Figure 2 is a flow chart of constructing the S2 dynamic graph of the present invention,

[0046] Figure 3 is the baseline model structure constructed by the present invention,

[0047] Figure 4 It is the GAT-LSTM model structure constructed by the present invention,

[0048] Figure 5 is a visualization example diagram of the clustering results of the present invention,

[0049] Figure 6 is a visualization example diagram of different modes of the present invention,

[0050] Figure 7 This is an example diagram of visualization of the time distribution of different modes of the present invention. DETAILED DESCRIPTION

[0051] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.

[0052] Example 1: See Figure 1-Figure 7 , a method for constructing an interpretable urban building energy consumption model integrating inter-building effects, the method comprising the following steps:

[0053] S1: Collect building electricity consumption information, process missing values ​​and outliers in the original data, and construct feature engineering based on electricity consumption in the time window;

[0054] S2: Construct a dynamic building graph within the time window, using buildings as nodes in the building graph and inter-building effects as edges in the building graph;

[0055] S3: Build an encoder-decoder based deep model with LSTM as the unit as the baseline model;

[0056] S4: Based on the baseline model, the GAT-LSTM model is constructed and trained by integrating graph neural network, attention mechanism and LSTM.

[0057] S5: Output the attention calculation results and use clustering algorithms to interpret the black box model, visualizing the different patterns of inter-building effects and their distribution over time.

[0058] Among them, S1 collects building electricity consumption information, processes missing values ​​and outliers in the original data, and constructs feature engineering based on electricity consumption in the time window, as follows:

[0059] S1.1: Collect building electricity consumption information through energy-saving platforms, sensors and field survey methods;

[0060] S1.2: Linear interpolation is used to supplement missing values ​​to ensure the integrity and consistency of the data. The interpolation calculation formula is as follows:

[0061]

[0062] Where y is the interpolated value of the unknown data point, y1 and y2 are the known data points, and x1 and x2 are the known x values ​​of the unknown data points;

[0063] S1.3: Draw a box plot of the data after missing values ​​are supplemented, and mark the data outside the range of Q_L-1.5IQR and Q_U+1.5IQR in the box plot as outliers, treat them as missing values, and use linear interpolation to supplement them. U , Q L , IQR are the upper and lower quartiles and interquartile range of the data set respectively;

[0064] S1.4: In order to eliminate the adverse effects caused by the large difference in electricity consumption between different buildings and ensure the effectiveness and stability of model training, the present invention adopts the standard deviation standardization method. The standard deviation standardization calculation formula is as follows:

[0065]

[0066] where x i For each feature in the dataset X, μ is the mean of the dataset, σ is the standard deviation of the dataset, and z i is the normalized value for each feature;

[0067] S1.5: Feature engineering of building graph nodes using time window sliding.

[0068] In S2, buildings are used as nodes in the architectural graph within the time window, and the effects between buildings are used as edges in the architectural graph, so as to construct a dynamic architectural graph integrating the effects between buildings, see Figure 2 ,in Figure 2 (a) is an urban building of an embodiment, Figure 2(b) is a building map constructed in a single time window to represent the effects between buildings and buildings. Figure 2 (c) is a dynamic building map constructed as the time window slides. The details are as follows:

[0069] S2.1: Each building is a node in the building diagram. The power consumption in a single time window is used as the building node feature. The full connection between building nodes is used to characterize the inter-building effect. The relevant formula is as follows:

[0070] G=(V,E)

[0071] V={v i |i∈{1, 2, ..., n}}

[0072] E={(v i , v j )|{v i , v j ∈V}}

[0073]

[0074] A ij =1,(v i , v j )∈E and i≠j

[0075] Among them, G, V, E, X∈R F Represents graph, node, edge and node feature vector set respectively; n is the number of nodes in the graph, v i , v j is a node in the graph, is node v i The initial feature vector of F is the initial node feature dimension; the adjacency matrix A∈R n×n A is used to represent the relationship between nodes in the graph. ij Represents the value in the adjacency matrix. Since the initial building graph is constructed in a fully connected manner, A ij Then it is 1;

[0076] S2.2: Construct a static building map in each time window, and construct a continuous dynamic building map as the time window slides.

[0077] See also Figure 3 , where S3 constructs an encoder-decoder based deep model with LSTM as the unit as the baseline model: the baseline model is a 4-layer LSTM model and a 1-layer fully connected layer with Relu as the activation function, the first 2 layers of LSTM model are feature encoders (Encoder), and the last 2 layers of LSTM model and 1-layer fully connected layer are reconstruction encoders (Decoder);

[0078] Among them, S4, based on the baseline model, integrates the graph neural network and attention mechanism to build the GAT-LSTM model and trains it. A graph convolution layer with a multi-head attention mechanism is added before the baseline model to calculate the dynamic building map constructed in step S2. During the training process, the attention value is continuously calculated and the adjacency matrix is ​​updated. See Figure 4 , specifically as follows,

[0079] S4.1: Construct a GAT-LSTM model based on graph neural network, attention mechanism and LSTM. The graph convolution layer with multi-head attention mechanism first calculates the attention value between different nodes and aggregates the feature information of adjacent building nodes. The specific formula is as follows:

[0080]

[0081] Among them, W is the weight matrix, which maps the original features to the high-dimensional space; is the shared self-attention performed on the node; e ij yes and The inner product value of two vectors, representing v j Node feature pair v i The importance of the node; n i Finger and v i The set of node numbers with edge connections. Refers to the concatenation operation of the feature vector; K is the number of attention heads in the multi-head attention mechanism. In order to make the e between different pairs of nodes ij The coefficients can be easily compared. The softmax function is used to normalize the importance calculation results of all selected values ​​of j to X′.

[0082] S4.2: In order to enable the aggregated information to be operated by the encoder-decoder based deep model with LSTM as the unit, this method constructs a sequence creation module, which creates the convolution X′ into continuous three-dimensional data for input into the baseline model;

[0083] S4.3: The Adam optimizer is used to train the baseline model and GAT-LSTM model constructed in steps S3 and S4.1 respectively with mean square error (MSE) as the loss function, and the loss is back-propagated to adjust the hyperparameters.

[0084] Among them, S5 outputs the attention calculation results and uses the clustering algorithm to interpret the black box model, visualizing the different patterns of inter-building effects and their distribution over time, as follows:

[0085] S5.1: Export the attention calculation results and use clustering algorithm to divide different types of inter-building effect patterns; S5.2: Use a device to visualize different patterns of inter-building effects and their distribution over time, and explore the key influencing buildings in different patterns. Figure 5 This is the clustering result of the attention of the embodiment. Figure 6 The five distinguished inter-building effect modes are visualized and important buildings under different modes are identified. Figure 7 The time distribution and quantitative statistical visualization results of the five types of inter-building effects are shown in Table 1. The evaluation index results of the experimental case of the method of the present invention are as follows:

[0086] Table 1

[0087]

[0088] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A method for constructing an interpretable urban building energy consumption model integrating inter-building effects, characterized in that: The method comprises the following steps: S1: Collect building electricity consumption information, process missing values ​​and outliers in the original data, and construct feature engineering based on electricity consumption in the time window; S2: Construct a dynamic building graph within the time window, using buildings as nodes in the building graph and inter-building effects as edges in the building graph; S3: Build an encoder-decoder based deep model with LSTM as the unit as the baseline model; S4: Based on the baseline model, the GAT-LSTM model is constructed and trained by integrating graph neural network, attention mechanism and LSTM. S5: Output the attention calculation results and use clustering algorithms to interpret the black box model, visualizing the different patterns of inter-building effects and their distribution over time.

2. The method for constructing an interpretable urban building energy consumption model integrating inter-building effects according to claim 1 is characterized in that: S1, collects building electricity energy consumption information, performs data processing on missing values ​​and outliers in the original data, and constructs feature engineering based on electricity consumption in the time window, as follows: S1.1: Collects building electricity energy consumption information through energy-saving platforms, sensors, and field survey methods; S1.2: Linear interpolation is used to supplement missing values ​​to ensure the integrity and consistency of the data. The interpolation calculation formula is as follows: Where y is the interpolated value of the unknown data point, y1 and y2 are the known data points, and x1 and x2 are the known x values ​​of the unknown data points; S1.3: Draw a box plot of the data after missing values ​​are supplemented, and mark the data outside the range of Q_L-1.5IQR and Q_U+1.5IQR of the box plot as outliers, treat them as missing values, and use linear interpolation to supplement them, where Q U , Q L , IQR are the upper and lower quartiles and interquartile range of the data set respectively; S1.4: The standard deviation standardization method is used. The standard deviation standardization calculation formula is as follows: where x i For each feature in the dataset X, μ is the mean of the dataset, σ is the standard deviation of the dataset, and z i is the normalized value for each feature; S1.5: Feature engineering of buildings using sliding time windows.

3. The method for constructing an interpretable urban building energy consumption model integrating inter-building effects according to claim 1 is characterized in that: S2, in the time window, buildings are used as nodes in the building graph, and the effects between buildings are used as edges in the building graph, so as to construct a dynamic building graph that integrates the effects between buildings, as follows: S2.1: Each building is a node in the building diagram. The power consumption in a single time window is used as the building node feature. The full connection between building nodes is used to characterize the inter-building effect. The relevant formula is as follows: G=(V,E) V={v i |i∈{1,2,...,n)}} E={(v i ,v j )|{v i ,v j ∈V}} A ij =1,(v i ,v j )∈E and i≠j Among them, G, V, E, X∈R F Represents graph, node, edge and node feature vector set respectively; n is the number of nodes in the graph, v i , v j is a node in the graph, is node v i The initial feature vector of F is the initial node feature dimension; the adjacency matrix A∈R n×n A is used to represent the relationship between nodes in the graph. ij Represents the value in the adjacency matrix. Since the initial building graph is constructed in a fully connected manner, A ij Then it is 1; S2.2: Construct a static building map in each time window, and construct a continuous dynamic building map as the time window slides.

4. The method for constructing an interpretable urban building energy consumption model integrating inter-building effects according to claim 1 is characterized in that: S3, construct an encoder-decoder based deep model with LSTM as the unit as the baseline model: the baseline model is a 4-layer LSTM model and a 1-layer fully connected layer with Relu as the activation function, the first 2 layers of LSTM model are feature encoders (Encoder), and the last 2 layers of LSTM model and 1 fully connected layer are reconstruction encoders (Decoder).

5. The method for constructing an interpretable urban building energy consumption model integrating inter-building effects according to claim 1 is characterized in that: S4, based on the baseline model, integrates the graph neural network and attention mechanism to build the GAT-LSTM model and trains it. A graph convolution layer with a multi-head attention mechanism is added before the baseline model to calculate the dynamic building map constructed in step S2. During the training process, the attention value is continuously calculated and the adjacency matrix is ​​updated. The details are as follows: S4.1: Construct a GAT-LSTM model based on graph neural network, attention mechanism and LSTM. The graph convolution layer with multi-head attention mechanism first calculates the attention value between different nodes and aggregates the feature information of adjacent building nodes. The specific formula is as follows: Among them, W is the weight matrix, which maps the original features to the high-dimensional space; is the shared self-attention performed on the node; e ij yes and The inner product value of two vectors, representing v j Node feature pair v i The importance of the node; n i Finger and v i The node number set with edge connections, || refers to the concatenation operation of the feature vector; K is the number of attention heads in the multi-head attention mechanism, in order to make the e between different pairs of nodes ij The coefficients can be easily compared. The softmax function is used to normalize the importance calculation results of all selected values ​​of j to X′. S4.2: Construct a sequence creation module, which creates continuous three-dimensional data after convolution X′ for input into the baseline model; S4.3: The Adam optimizer is used to train the baseline model and GAT-LSTM model constructed in steps S3 and S4.1 respectively with mean square error (MSE) as the loss function, and the loss is back-propagated to adjust the hyperparameters.

6. The method for constructing an interpretable urban building energy consumption model integrating inter-building effects according to claim 1 is characterized in that: S5 outputs the attention calculation results and uses the clustering algorithm to interpret the black box model, visualizing the different patterns of inter-building effects and their distribution over time, as follows: S5.1: Derive the attention calculation results and use clustering algorithm to classify different types of inter-building effect patterns; S5.2: Visualize the different modes of inter-building effects and their temporal distribution patterns through an electronic device, and mine the key influencing buildings in different modes.

7. An electronic 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, a method for constructing an interpretable urban building energy consumption model integrating inter-building effects as described in any one of claims 1 to 6 is implemented.

8. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, a method for constructing an interpretable urban building energy consumption model integrating inter-building effects as described in any one of claims 1 to 6 is implemented.

Images