Building group load forecasting method based on causal inference

Abstract

The application discloses a building group load prediction method based on causal inference. First, the building group load sequence is acquired and pretreated; the discrete wavelet transform is used to decompose the building load sequence to obtain a plurality of high-frequency components and a low-frequency component; the building load fluctuation characteristics are reconstructed by using the high-frequency components and the low-frequency component; then, the static and dynamic causal diagrams are constructed by taking the building as a node, the building load fluctuation characteristics are coded into latent representations, the building load state structure equation is constructed by using the latent representations, each building load state structure equation is solved by optimization, and the adjacency matrix of the causal diagram is obtained; finally, the load prediction model is constructed, the adjacency matrix of the static causal diagram and the dynamic causal diagram is taken as input, the building group load at a future time step is predicted according to the building group historical load sequence, and the building group load prediction sequence is obtained. The problems that it is difficult to capture the dynamic propagation effect of the load change among the buildings and the modeling of the complex interaction among the buildings is insufficient are solved.

Description

Technical Field

[0001] This invention belongs to the field of multivariate load forecasting technology, specifically relating to a building cluster load forecasting method based on causal inference. Background Technology

[0002] With the acceleration of urbanization and the widespread application of smart building technologies, buildings have become major consumers of urban energy. Faced with ever-increasing energy demand and carbon emission reduction pressures, accurate forecasting of building cluster electricity load is crucial for achieving building energy conservation and intelligent management. However, current building cluster load forecasting faces the following two key technical challenges: First, the quantification of the dynamic propagation effect of load is insufficient. Load changes between buildings do not occur in isolation, but are transmitted through power grid coupling, personnel movement, and other pathways, exhibiting dynamic and cumulative characteristics. For example, the off-hours rush hour in office buildings can cause load fluctuations in surrounding commercial facilities, and traditional causal inference methods struggle to accurately quantify the dynamic propagation effect of load changes between buildings. For instance, Granger causal analysis is limited to the assumption of linear relationships and only supports causal tests between paired variables, making it difficult to extend to high-dimensional scenarios involving a large number of buildings; while PC algorithms are suitable for high-dimensional data analysis, their ability to capture the dynamic characteristics of causal relationships evolving over time is limited. Therefore, existing models cannot effectively characterize the complex dynamic load propagation paths and intensity between buildings, thus affecting the model's predictive accuracy and stability.

[0003] Second, there is insufficient modeling of complex interactions between buildings. Current mainstream prediction models such as decision trees, recurrent neural networks (RNNs), and spatiotemporal graph convolutional networks (ST-GCNs) all have limitations in capturing complex interactions between buildings. RNNs can automatically learn temporal features, but they usually ignore spatial dependencies between buildings. Although decision trees can handle nonlinear data, they rely heavily on manual feature engineering and are difficult to extend to large-scale, complex building clusters. While ST-GCNs introduce a spatial dimension, their graph structures are usually built based on fixed geographical proximity and cannot reflect the real load propagation patterns driven by functions and behaviors. For example, in a complex consisting of office buildings, shopping malls, and residences, the electricity consumption patterns form a complex causal chain: the electricity consumption behavior of office buildings affects the customer flow and load of the shopping mall, while the activities of the shopping mall may change the electricity consumption habits of the residential area. This multimodal and dynamically coupled characteristic exceeds the modeling capabilities of existing prediction models based on simple graph structures or single temporal analysis. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a building cluster load prediction method based on causal inference, in order to solve the problems of difficulty in capturing the dynamic propagation effect of load changes between buildings and insufficient modeling of complex interactions between buildings.

[0005] The present invention solves the aforementioned technical problem by adopting the following technical solution: A building cluster load forecasting method based on causal inference, characterized by the following steps: Step 1: Obtain the load sequence of the building complex and perform preprocessing; Step 2: Use discrete wavelet transform to perform multi-scale decomposition on the building load sequence to obtain multiple high-frequency components and one low-frequency component of the building load sequence; use the high-frequency components and low-frequency components to reconstruct the building load fluctuation characteristics. Step 3: Construct static cause-effect graphs and dynamic cause-effect graphs using buildings as nodes; The multilayer perceptron encodes the building load fluctuation characteristics into a latent representation, and uses the latent representation to model the building load sequence into the load state structure equation of equation (8); the adjacency matrix of the static cause-effect graph is obtained by optimizing and solving each building load state structure equation. (8) In the formula, For the first A building load sequence, It is a nonlinear transformation function; The elements of the adjacency matrix of a static causal graph represent the first... The building faces the first The causal transmission strength of a building; For the first Potential representation of building load fluctuation characteristics For the number of buildings, For structural residuals; The preprocessed building cluster load sequence is divided into multiple building cluster load segments using a sliding window. A dynamic cause-effect graph is constructed for each building cluster load segment. The adjacency matrix of the dynamic cause-effect graph is constructed in the same way as the adjacency matrix of the static cause-effect graph, thus obtaining all dynamic cause-effect graphs and their adjacency matrices. Step 4: Construct a load forecasting model based on asymmetric spatiotemporal graph convolution, predict the load of the building cluster at future time steps based on the historical load sequence of the building cluster, and obtain the building cluster load forecast sequence; A geographic distance map is constructed using buildings as nodes, and its adjacency matrix is ​​defined as follows: (13) In the formula, These are the elements of the adjacency matrix of the geographic distance map. For the first The building and the first The Euclidean distance between the buildings. For distance attenuation bandwidth; The static cause-effect graph and the geographic distance graph are fused to obtain a static cause-effect-geographic distance graph; the adjacency matrices of the static cause-effect graph and the geographic distance graph are concatenated, and then... Convolution yields the adjacency matrix of the static causal-geographic distance graph; Modulate the adjacency matrix of the static causal-geographic distance map according to equation (15) to obtain the static adjacency matrix with dynamic causal enhancement. (15) In the formula, For the first A dynamic causal-enhanced static adjacency matrix, The adjacency matrix of a static causal-geographic distance graph. The gating coefficient matrix is ​​a learnable matrix. Represents the Hadama product; The load prediction model based on asymmetric spatiotemporal graph convolution consists of multiple stacked asymmetric spatiotemporal graph convolutional modules. Each asymmetric spatiotemporal graph convolutional module includes a local temporal convolutional layer, an asymmetric graph convolutional layer, and a higher-order temporal convolutional layer; for the th... For each asymmetric spatiotemporal graph convolutional module, the local temporal convolutional layer processes the high-order temporal features of each building load along the time dimension according to equation (16), extracting the local temporal features of the building load; and stacks all the local temporal features of the building load along the node dimension to obtain the first... Local temporal characteristics of building cluster load output from the local temporal convolutional layer of an asymmetric spatiotemporal graph convolutional module; (16) In the formula, For the first The output of the local temporal convolutional layer of the asymmetric spatiotemporal graph convolutional module is the first... Local time-series characteristics of building load For the first The output of the first asymmetric spatiotemporal graph convolution module is the... High-order time-series characteristics of building load It is the ReLU activation function. It is a one-dimensional convolution; The asymmetric convolutional layer utilizes the dynamic causal enhancement of the static adjacency matrix corresponding to the previous building group historical load segment of the current prediction time step. The asymmetric convolutional layer aggregates the local temporal features of the building group load at each time position according to Equation (17) to obtain the building group load features at each time position. The building group load features at all time positions are stacked along the time dimension to obtain the building group load features. (17) In the formula, For the first The temporal location of the output of the asymmetric spatiotemporal graph convolutional module is... Building load characteristics at the location For time position Local temporal characteristics of building load in the area. It is the Sigmoid activation function. This is the static adjacency matrix with dynamic causal enhancement corresponding to the historical load segment of the building cluster preceding the current prediction time step. For transpose operation, , The weight matrix is ​​a learnable weight matrix; The higher-order temporal convolutional layer processes each building load feature along the time dimension according to equation (18) to extract higher-order temporal features of building load; all higher-order temporal features of building load are stacked along the node dimension to obtain the first... High-order temporal features of building cluster load output by an asymmetric spatiotemporal graph convolutional module; (18) In the formula, For the first The output of the higher-order temporal convolutional layer of the asymmetric spatiotemporal graph convolutional module is the first... High-order time-series characteristics of building load For the first Building load characteristics; After passing through multiple asymmetric spatiotemporal graph convolutional modules, the final high-order temporal features of the building cluster load are obtained. The final high-order temporal features of the building cluster load are flattened into a one-dimensional vector to obtain the load vector of each building. The load vector of each building is mapped to the load prediction value of the future time step through a fully connected layer to obtain the load prediction sequence of the building cluster.

[0006] Furthermore, in the process of optimizing the structural equations of building load state, the loss is calculated based on the following loss function: (9) (10) (11) (12) In the formula, For the total loss, For structural residuals, For sparsity constraints, It is an L1 norm. For the constraints of a directed acyclic graph, and These are the hyperparameters for the sparsity constraint term and the directed acyclic constraint term, respectively. This is the adjacency matrix of a static cause-effect graph. This indicates taking the absolute value. For matrix trace operations, This is the base of the exponential function.

[0007] Furthermore, in the second step, all high-frequency and low-frequency components are reconstructed by sequence length to obtain high-frequency and low-frequency reconstructed components; all reconstructed components are spliced ​​along the feature dimension to obtain multi-scale features of the building load sequence; at the same time, the reconstructed components are globally averaged pooled along the time dimension to obtain channel description features; using the channel description features, channel attention weights are generated according to equation (4). (4) In the formula, For the first Each channel attention weight, For the first Each channel describes a feature. , For weight parameters, , For bias parameters, The number of decomposition levels for the discrete wavelet transform; According to Equation (5), the multi-scale features of the building load sequence are weighted by channels to obtain the channel attention multi-scale features; (5) In the formula, For the first Multi-scale features of channel attention; The building load fluctuation characteristics are obtained by summing the multi-scale features of all channels along the feature dimension.

[0008] Compared with the prior art, the beneficial effects of the present invention are: 1. Load changes between buildings are transmitted through power grid coupling and personnel movement. Existing causal inference methods are limited by linear assumptions or static structures, making it difficult to quantify this dynamic propagation effect. Therefore, this invention improves the quantification capability of dynamic load propagation effects between buildings by constructing static and dynamic causal graphs. First, discrete wavelet transform and adaptive channel attention mechanism are introduced to decompose and weightedly fuse historical load sequences of buildings at multiple scales, overcoming the limitation of decreased causal identification accuracy caused by multi-scale information aliasing. Second, the adjacency matrix of the causal graph is parameterized as a continuously learnable variable. End-to-end optimization is performed by fusing loss functions of structural residual terms, sparsity constraint terms, and directed acyclic constraint terms to construct static and dynamic causal graphs respectively. The explicit graph structure quantitatively represents the path, direction, and time-varying intensity of load propagation, realizing the quantitative representation of causal propagation path and intensity between buildings in high-dimensional nonlinear time-varying scenarios. At the same time, static and dynamic causal graphs can provide structural priors for prediction models.

[0009] 2. Addressing the shortcomings of existing prediction models, which typically rely on fixed geographical proximity to construct graph structures or model only single temporal dependencies, making it difficult to effectively capture complex interactions between buildings driven by functional coupling and behavioral correlations, this invention constructs a load prediction model based on asymmetric spatiotemporal graph convolution, enhancing its ability to model complex interactions between buildings (including asymmetric and time-varying dependencies). First, learnable convolutional layers enable adaptive fusion of geographical distance graphs and static causal graphs, overcoming the limitation of a single graph structure in simultaneously considering spatial proximity and causal dependencies. Second, a dynamic causal graph is incorporated through a gating mechanism at each prediction time step, giving the graph structure time-varying adaptability. Finally, in the asymmetric graph convolution operation, adjacency matrices and their transposes are used to aggregate node features and assign independent weight matrices, explicitly distinguishing the different roles of buildings as causal influencing parties and receiving parties, thus achieving effective modeling of causal interactions between buildings with varying directions and time-varying intensity. This method enables joint modeling of geographical and causal dependencies and characterization of asymmetric time-varying transmission relationships, providing a multi-dimensional spatiotemporal modeling framework for building cluster load forecasting and improving the accuracy of building cluster load forecasting. Attached Figure Description

[0010] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a structural diagram of the load prediction model based on asymmetric spatiotemporal graph convolution of the present invention. Detailed Implementation

[0011] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail and are not intended to limit the scope of protection of this application.

[0012] like Figure 1-2As shown, this invention proposes a building cluster load forecasting method based on causal inference, comprising the following steps: Step 1: Obtain the load sequence of the building complex and perform preprocessing; The load sequence of the building complex is denoted as This includes multiple building load sequences; among them, It refers to the number of buildings. The sequence length is specified. Preprocessing includes: first, identifying missing and outlier values, and then using Lagrange interpolation to fill in missing values ​​and replace outliers; subsequently, normalizing the sequence and scaling it to a uniform scale to improve the stability and efficiency of model training, resulting in the preprocessed building cluster load sequence.

[0013] Step 2: Extract multi-scale features of building load sequences based on multi-scale volatility perception, and adaptively fuse the multi-scale features to obtain building load volatility features; Discrete wavelet transform (DWT) is used to decompose the load sequences of each building at multiple scales, yielding multiple high-frequency components and one low-frequency component for each building's load sequence. The DWT employs a cascaded decomposition structure, directly retaining the high-frequency components obtained from each decomposition level, which characterize the fluctuation features at the corresponding frequency scale. The low-frequency components obtained from the previous level are further decomposed into high-frequency and low-frequency components in the current level. These low-frequency components are then input into the next level for further separation into even smaller high-frequency and low-frequency components. This process is repeated until the predicted decomposition level is reached. The low-frequency component obtained from the last level is used as the low-frequency component for the high-rise, reflecting the long-term trend of the load sequence and is retained. Historical load sequence of a building As the original sequence of discrete wavelet transform The decomposition recursive formula for the discrete wavelet transform is as follows: (1) (2) In the formula, , The first The building load sequence after the first The low-frequency and high-frequency components obtained from the layer decomposition For the first The building load sequence after the first Low-frequency components obtained from layer decomposition; These are the coefficients of the low-pass decomposition filter. These are the coefficients of the high-pass decomposition filter, and they satisfy the orthogonality condition. For sampling point index, For filter coefficient index, The number of decomposition levels for the discrete wavelet transform; Finally, the first Building load sequence One high-frequency component and one low-frequency component Among them, low-level high-frequency components Reflects hourly instantaneous high-frequency fluctuations, high-frequency components at higher levels. and low frequency components They respectively reflect the cyclical patterns and long-term trends at the daily to weekly levels.

[0014] Since the components obtained from each decomposition layer have unequal lengths after downsampling, direct interpolation alignment will introduce false information. Therefore, this implementation method reconstructs the components obtained from each decomposition layer using zero-filling inverse wavelet transform, that is, the first... High-frequency components are obtained through layer decomposition. Reconstruct to sequence length High-frequency reconstructed components are obtained. Similarly, the low-frequency reconstructed components are obtained. ; All reconstructed components are spliced ​​along the feature dimension to obtain the multi-scale features of the building load sequence; (3) In the formula, For the first Multi-scale characteristics of historical load sequences of buildings This indicates concatenation along the feature dimension; Each reconstructed component corresponds to a channel. Based on this, a channel attention mechanism is introduced to learn the attention weights of each channel. First, global average pooling is performed on each reconstructed component along the time dimension to obtain the descriptive features of each channel. The channel attention weights are generated according to equation (4): (4) In the formula, For the first Each channel attention weight, It is the Sigmoid activation function. For the first Each channel describes a feature; , For weight parameters, For compression ratio, it is usually taken as ; It is the ReLU activation function. , These are bias parameters; The channel attention multi-scale features are obtained by channel weighting of the multi-scale features of the building load sequence using the following formula; (5) In the formula, For the first Multi-scale features of channel attention Represents the Hadama product; The building load fluctuation characteristics are obtained by summing the multi-scale features of all channels along the feature dimension; (6) In the formula, For the first Individual building load fluctuation characteristics; Stack all building load fluctuation characteristics row by row to obtain the building group load fluctuation characteristics. .

[0015] Step 3: Construct static cause-effect graphs and dynamic cause-effect graphs using buildings as nodes; 3.1) Constructing a static cause-effect graph This reflects the long-term and stable causal transmission relationship between buildings; First, define a static cause-effect graph. The adjacency matrix is Use it as the parameter to be optimized, adjacency matrix It has the following properties: (1) Adjacency matrix It is an asymmetric matrix, and its elements and These are two independent parameters, representing the first and second lines respectively. The building faces the first The causal transmission strength of the building and the first The building faces the first (2) Adjacency matrix It should satisfy sparsity, since not all buildings in the building group have causal transmission relationships; (3) Adjacency matrix The elements are non-negative, that is To ensure that the causal transmission strength has physical interpretability; the adjacency matrix It is initialized to a fully connected state, that is, all elements are initialized to small positive values, indicating that there is a causal transmission relationship between any two buildings at the beginning.

[0016] The load fluctuation characteristics of each building are input into a multilayer perceptron with shared parameters, and encoded into a latent representation: (7) In the formula, For potential representation, For the first Building load fluctuation characteristics The parameter is Multilayer perceptron; Based on the latent representation, each building load sequence is modeled as the following load state structure equation: (8) In the formula, For the first A building load sequence, This is a nonlinear transformation function implemented by a fully connected layer; For structural residuals, it represents the components that cannot be explained by causal transmission relationships between buildings; In the above structural equation, the adjacency matrix Control information transmission; when When approaching zero, the first Information about each building is not included in the discussion of the first building. Explanation of building load status; when When it is larger, it indicates that the first The building to the first There is a significant causal relationship between the buildings; 3.2) Using the Adam optimizer, the load state structure equation for each building is optimized according to the following loss function to obtain the adjacency matrix of the static cause-effect graph. ; To ensure the learned adjacency matrix To construct an effective static cause-effect graph, the following loss function is established: (9) In the formula, For the total loss, For structural residuals, For sparsity constraints, It is an L1 norm. For directed acyclic graph constraints; and These are the hyperparameters for the sparse constraint term and the directed acyclic constraint term, respectively, used to control the degree of sparsity penalty and the penalty strength for the directed acyclic constraint. Structural residuals This term is used to measure the explanatory power of the load state structure equation for the dynamic changes in the load of a building group. Minimizing this term allows the cause-effect graph to use information from other buildings to explain the load changes of each building. If the _th_ term is used... Information about the building helps explain the first... The load status of each building, During the optimization process, it tends towards a non-zero value; the expression for the structural residual term is: (10) The sparsity constraint term uses L1 regularization to sparsify the adjacency matrix. This term works in opposition to the structural residual term, which tends to retain all possible propagation paths, while the sparsity constraint term forces the causal graph to eliminate connections that do not contribute sufficiently to explaining load variations. When the contribution to reducing structural residuals is insufficient to compensate for its L1 regularization cost, the connection will be compressed to zero, i.e. The expression for the sparsity constraint term is: (11) In the formula, Indicates taking the absolute value; The directed acyclic constraint term uses the trace exponential function to ensure that the learned causal graph satisfies the directed acyclic property, then: (12) In the formula, For matrix trace operations, The base of the exponential function is given if and only if there are no loops in the causal graph. ,at this time When a loop exists in the causal graph, .

[0017] The structural residual term preserves causal connections with significant explanatory power, the sparsity constraint term eliminates weak spurious associations, and the directed acyclic constraint eliminates logical causal loops. Through the synergistic effect of these three terms, the model can automatically output a sparse, directed, and acyclic causal graph based on historical load sequences.

[0018] 3.3) Construct a dynamic cause-effect graph to reflect the short-term time-varying causal transmission relationship; For the construction of the dynamic cause-effect graph, a sliding window is used to divide the preprocessed building cluster load sequence into multiple building cluster load segments, with a window width of [missing information]. The sliding step size is The total number of fragments is A dynamic cause-effect graph is constructed based on the load segments of the building cluster. Each segment corresponds to a dynamic cause-effect graph, and within each segment, the adjacency matrix of the static cause-effect graph is used. The adjacency matrix of the dynamic cause-effect graph is initialized, and then optimized using the same method as for constructing the adjacency matrix of the static cause-effect graph, ultimately yielding all dynamic cause-effect graphs. and its adjacency matrix , .

[0019] The role of dynamic causal graphs is to quantify the short-term fluctuation characteristics that deviate from static causal graphs. When the physical interaction pattern of building load changes within a local time period (for example, due to differences in the work and rest of different building functional areas, the original causal flow direction is reversed at a specific time), this sliding fine-tuning mechanism can capture the reversal of local causal direction and the dynamic evolution of transmission intensity. Static causal graphs and dynamic causal graphs together constitute the causal structure prior of "global benchmark + local fine-tuning". Static causal graphs provide a benchmark for the network transmission path of long-term temperature, while dynamic causal graphs capture real-time interaction changes under non-stationary operating conditions.

[0020] Step 4: Construct a load forecasting model based on asymmetric spatiotemporal graph convolution, predict the load of the building cluster at future time steps based on the historical load sequence of the building cluster, and obtain the building cluster load forecast sequence; Construct a geographic distance map using buildings as nodes. Its adjacency matrix Reflecting the spatial proximity between buildings, the matrix elements are defined as follows: (13) In the formula, For the first The building and the first The Euclidean distance between the buildings. For distance attenuation bandwidth; By fusing the static cause-effect graph with the geographic distance graph, a static cause-effect-geographic distance graph is obtained. This diagram integrates complementary information from a static cause-and-effect diagram and a geographic distance diagram; The adjacency matrix of the static cause-effect graph and the geographic distance graph is concatenated, and then... Convolutional operations are used for compression to obtain a static causal-geographical distance map. adjacency matrix ; (14) In the formula, for The convolution operation, with its learnable convolutional kernel, weights and compresses two channels into one channel; Indicates two sizes of The matrices are stacked along the channel dimension. tensor; Since the adjacency matrix of the static causal-geographic distance graph only reflects long-term stable interaction patterns and cannot adapt to short-term changes in causal relationships, in order to incorporate time-varying information into the model, the adjacency matrix of the dynamic causal graph is used to modulate the adjacency matrix of the static causal-geographic distance graph to obtain a static adjacency matrix with dynamic causal enhancement. This matrix integrates long-term stable interaction patterns and short-term causal changes, providing a foundation for subsequent accurate predictions. (15) In the formula, For the first A dynamic causal-enhanced static adjacency matrix; It is a learnable gating coefficient matrix used to balance the influence of short-term causal relationships. When a certain short-term causal effect is significant, the corresponding gating coefficient will increase, thereby strengthening the expression of the effect in the graph structure; conversely, it will weaken. This design enables the model to balance long-term structural dependencies and short-term sudden effects, improving the model's adaptability to complex dynamic scenarios.

[0021] The model's prediction task is based on Building load observations at historical time steps to predict future... The building cluster load values ​​at each time step are used to extract the continuous load values ​​before the current prediction time step from the preprocessed building cluster load sequence. The historical load sequence of the building complex is obtained by analyzing load observations at each historical time step. , and use it as the model input.

[0022] The load prediction model based on asymmetric spatiotemporal graph convolution includes multiple asymmetric spatiotemporal graph convolutional modules. The output features of the previous asymmetric spatiotemporal graph convolutional module (i.e., high-order temporal features of building cluster load) serve as the input features of the next asymmetric spatiotemporal graph convolutional module. Each asymmetric spatiotemporal graph convolutional module includes a local temporal convolutional layer, an asymmetric graph convolutional layer, and a high-order temporal convolutional layer, which are used to extract local temporal features of building load, building cluster load features, and high-order temporal features of building load, respectively. Specifically, assuming the first... The input features of the asymmetric spatiotemporal graph convolutional module are , The length is the time dimension. The width of the convolution kernel. Number of feature channels; The local temporal convolutional layer independently processes the high-order temporal features of each building load along the time dimension, extracting the local temporal features of the building load; for the th Each asymmetric spatiotemporal graph convolutional module takes the high-order temporal features of building loads output by the previous asymmetric spatiotemporal graph convolutional module as input, and passes them through a local temporal convolutional layer to obtain the local temporal features of building loads: (16) In the formula, For the first The output of the local temporal convolutional layer of the asymmetric spatiotemporal graph convolutional module is the first... Local time-series characteristics of building load For the first High-order time-series characteristics of building load It is a one-dimensional convolution; Stacking all local temporal features of building loads along the node dimension yields the... Local temporal characteristics of building cluster loads output from the local temporal convolutional layer of an asymmetric spatiotemporal graph convolutional module. ;in, The length of the time dimension after convolution. Number of feature channels; The asymmetric convolutional layer utilizes the dynamically causally enhanced static adjacency matrix and its transpose corresponding to the historical load segment of the building cluster from the previous time step in the current prediction time step. It aggregates information from both the causal receiving direction and the causal influencing direction, assigning independent learnable weight matrices to each direction to explicitly distinguish the different roles of buildings in causal transmission. The asymmetric convolutional layer independently spatially aggregates the local temporal features of the building cluster load at each time location in the time dimension, obtaining the building cluster load features at each time location. The computation process of the asymmetric convolutional layer is as follows: (17) In the formula, For the first The temporal location of the output of the asymmetric convolutional layer of the asymmetric spatiotemporal graph convolutional module Building load characteristics at the location; For the first The input features of the asymmetric convolutional layer of the asymmetric spatiotemporal graph convolutional module are the temporal location. Local temporal characteristics of building load at the location; For transpose operation, , For learnable weight matrix, This is the static adjacency matrix with dynamic causal enhancement corresponding to the historical load segment of the building cluster preceding the current prediction time step; By stacking the building load features at all time locations along the time dimension, the output features of the asymmetric convolutional layer are obtained, which are the building load features. ; In equation (17), the first term It aggregates information obtained by each building as a causal receiver from its causal influencing parties; the second item It aggregates the information propagated by each building as a causal influence direction and its causal receiver; this operation enables the model to distinguish the different roles of buildings as information sources and information targets while learning spatial dependencies; the two directions use independent learnable weight matrices, enabling the model to learn different feature transformation relationships between causal influencers and receivers respectively, thereby achieving modeling of the asymmetry of causal transmission. Similar to local temporal convolutional layers, higher-order temporal convolutional layers process each building load feature independently and further extract higher-order temporal features of building load along the time dimension; (18) In the formula, For the first The output of the higher-order temporal convolutional layer of the asymmetric spatiotemporal graph convolutional module is the first... High-order time-series characteristics of building load For the first Building load characteristics; Stacking all high-order temporal features of building loads along the node dimension yields the... High-order temporal features of building cluster load output by an asymmetric spatiotemporal graph convolutional module This is used as the input to the next asymmetric spatiotemporal graph convolution module. ; Repeating the aforementioned process, after passing through multiple asymmetric spatiotemporal graph convolutional modules, the final high-order temporal features of building cluster load are obtained. The final high-order temporal features of the building cluster load are flattened into one-dimensional vectors along the time and channel dimensions to obtain the load vectors of each building; each building load vector is mapped to the future load vector through a fully connected layer. The load forecast values ​​at each time step are used to obtain the building cluster load forecast sequence. ;in, For the first Load forecast vector for each building.

[0023] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A building cluster load forecasting method based on causal inference, characterized in that, Includes the following steps: Step 1: Obtain the load sequence of the building complex and perform preprocessing; Step 2: Use discrete wavelet transform to perform multi-scale decomposition on the building load sequence to obtain multiple high-frequency components and one low-frequency component of the building load sequence; use the high-frequency components and low-frequency components to reconstruct the building load fluctuation characteristics. Step 3: Construct static cause-effect graphs and dynamic cause-effect graphs using buildings as nodes; The multilayer perceptron encodes the building load fluctuation characteristics into a latent representation. Using the latent representation, the building load sequence is modeled as the load state structure equation of equation (8). The adjacency matrix of the static cause-effect graph is obtained by optimizing and solving each building load state structure equation. （8） In the formula, For the first A building load sequence, It is a nonlinear transformation function; The elements of the adjacency matrix of a static causal graph represent the first... The building faces the first The causal transmission strength of a building; For the first Potential representation of building load fluctuation characteristics For the number of buildings, For structural residuals; The preprocessed building cluster load sequence is divided into multiple building cluster load segments using a sliding window. A dynamic cause-effect graph is constructed for each building cluster load segment. The adjacency matrix of the dynamic cause-effect graph is constructed in the same way as the adjacency matrix of the static cause-effect graph, thus obtaining all dynamic cause-effect graphs and their adjacency matrices. Step 4: Construct a load forecasting model based on asymmetric spatiotemporal graph convolution, predict the load of the building cluster at future time steps based on the historical load sequence of the building cluster, and obtain the building cluster load forecast sequence; A geographic distance map is constructed using buildings as nodes, and its adjacency matrix is ​​defined as follows: （13） In the formula, These are the elements of the adjacency matrix of the geographic distance map. For the first The building and the first The Euclidean distance between the buildings. For distance attenuation bandwidth; The static cause-effect graph and the geographic distance graph are fused to obtain a static cause-effect-geographic distance graph; the adjacency matrices of the static cause-effect graph and the geographic distance graph are concatenated, and then... Convolution yields the adjacency matrix of the static causal-geographic distance graph; Modulate the adjacency matrix of the static causal-geographic distance map according to equation (15) to obtain the static adjacency matrix with dynamic causal enhancement. （15） In the formula, For the first A dynamic causal-enhanced static adjacency matrix, The adjacency matrix of a static causal-geographic distance graph. The gating coefficient matrix is ​​a learnable matrix. Represents the Hadama product; The load prediction model based on asymmetric spatiotemporal graph convolution consists of multiple stacked asymmetric spatiotemporal graph convolutional modules. Each asymmetric spatiotemporal graph convolutional module includes a local temporal convolutional layer, an asymmetric graph convolutional layer, and a higher-order temporal convolutional layer; for the th... For each asymmetric spatiotemporal graph convolutional module, the local temporal convolutional layer processes the high-order temporal features of each building load along the time dimension according to equation (16), extracting the local temporal features of the building load; and stacks all the local temporal features of the building load along the node dimension to obtain the first... Local temporal characteristics of building cluster load output from the local temporal convolutional layer of an asymmetric spatiotemporal graph convolutional module; （16） In the formula, For the first The output of the local temporal convolutional layer of the asymmetric spatiotemporal graph convolutional module is the first... Local time-series characteristics of building load For the first The output of the first asymmetric spatiotemporal graph convolution module is the... High-order time-series characteristics of building load It is the ReLU activation function. It is a one-dimensional convolution; The asymmetric convolutional layer utilizes the dynamic causal enhancement of the static adjacency matrix corresponding to the previous building group historical load segment of the current prediction time step. The asymmetric convolutional layer aggregates the local temporal features of the building group load at each time position according to Equation (17) to obtain the building group load features at each time position. The building group load features at all time positions are stacked along the time dimension to obtain the building group load features. （17） In the formula, For the first The temporal location of the output of the asymmetric spatiotemporal graph convolutional module is... Building load characteristics at the location For time position Local temporal characteristics of building load in the area. It is the Sigmoid activation function. This is the static adjacency matrix with dynamic causal enhancement corresponding to the historical load segment of the building cluster preceding the current prediction time step. For transpose operation, , The weight matrix is ​​a learnable weight matrix; The higher-order temporal convolutional layer processes each building load feature along the time dimension according to equation (18) to extract the higher-order temporal features of the building load; Stacking all high-order temporal features of building loads along the node dimension yields the... High-order temporal features of building cluster load output by an asymmetric spatiotemporal graph convolutional module; （18） In the formula, For the first The output of the higher-order temporal convolutional layer of the asymmetric spatiotemporal graph convolutional module is the first... High-order time-series characteristics of building load For the first Building load characteristics; After passing through multiple asymmetric spatiotemporal graph convolutional modules, the final high-order temporal features of building cluster load are obtained; The final high-order temporal features of the building cluster load are flattened into a one-dimensional vector to obtain the load vector of each building. The load vector of each building is then mapped to the load prediction value of the future time step through a fully connected layer to obtain the load prediction sequence of the building cluster.

2. The building cluster load forecasting method based on causal inference according to claim 1, characterized in that, In the optimization solution of the building load state structural equation, the loss is calculated according to the following loss function: （9） （10） （11） （12） In the formula, For the total loss, For structural residuals, For sparsity constraints, It is an L1 norm. For the constraints of a directed acyclic graph, and These are the hyperparameters for the sparsity constraint term and the directed acyclic constraint term, respectively. This is the adjacency matrix of a static cause-effect graph. This indicates taking the absolute value. For matrix trace operations, This is the base of the exponential function.

3. The building cluster load forecasting method based on causal inference according to claim 1 or 2, characterized in that, In the second step, all high-frequency and low-frequency components are reconstructed by sequence length to obtain high-frequency and low-frequency reconstructed components; all reconstructed components are spliced ​​along the feature dimension to obtain multi-scale features of the building load sequence; at the same time, the reconstructed components are globally averaged pooled along the time dimension to obtain channel description features; using the channel description features, channel attention weights are generated according to equation (4). （4） In the formula, For the first Each channel attention weight, For the first Each channel describes a feature. , For weight parameters, , For bias parameters, The number of decomposition levels for the discrete wavelet transform; According to Equation (5), the multi-scale features of the building load sequence are weighted by channels to obtain the channel attention multi-scale features; （5） In the formula, For the first Multi-scale features of channel attention; The building load fluctuation characteristics are obtained by summing the multi-scale features of all channels along the feature dimension.

Images