A commercial building central air conditioning energy consumption prediction system and method based on data analysis

Abstract

The application discloses a kind of based on data analysis's commercial building central air conditioning energy consumption prediction system and method, it is related to data analysis technical field, the present method includes the following steps: obtaining the multi-source data related to commercial building central air conditioning energy consumption;Data is preprocessed;Processing time sequence characteristic data, generate time sequence characteristic importance weight, and filter key time sequence characteristic;Processing static characteristic data, generate static characteristic importance weight, and filter key static characteristic;Splice key time sequence characteristic and key static characteristic, construct energy consumption prediction model;Predict key time sequence characteristic data, further predict energy consumption data;Define energy consumption cost and numerical / classification type key time sequence characteristic conversion cost, with total cost minimization as target, combined constraint condition is solved using mixed integer optimization algorithm, output key time sequence characteristic optimal target state.The present application can effectively improve the situation that energy consumption optimization scheme is difficult to effectively utilize influencing factor in prior art.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically a data analysis-based system and method for predicting the energy consumption of central air conditioning systems in commercial buildings. Background Technology

[0002] Central air conditioning in commercial buildings is a core component of building energy consumption. Its energy consumption level is influenced by multiple factors, including dynamic temporal characteristics (such as ambient temperature and humidity, solar radiation, pedestrian density, and equipment operating parameters) and fixed static characteristics (such as room orientation, floor level, and space size). To achieve energy-saving management, the industry generally uses data analysis techniques to build energy consumption prediction models, combining historical data to predict energy consumption trends and adjusting air conditioning operating parameters based on the prediction results. In related research, prediction models mostly revolve around time-series data mining and feature correlation analysis, while energy consumption optimization is one of the core research directions in the field of building energy conservation.

[0003] Existing energy consumption optimization technologies for central air conditioning in commercial buildings have significant limitations: they often focus solely on reducing energy consumption values, completely ignoring the conversion costs involved in parameter adjustments. For example, adjusting the air conditioning temperature (numerical) and switching fan speeds (categorical) incur costs such as equipment wear and energy consumption adjustments. Furthermore, existing algorithms struggle to balance the synergistic optimization of numerical and categorical parameters, often resulting in a situation where energy consumption decreases but conversion costs surge. This prevents the achievement of the optimal total cost of "energy consumption cost + conversion cost," severely impacting the practical application benefits of the optimization scheme. Summary of the Invention

[0004] The purpose of this invention is to provide a data analysis-based energy consumption prediction system and method for central air conditioning in commercial buildings, in order to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the energy consumption of central air conditioning in commercial buildings based on data analysis, the method comprising the following steps:

[0006] S1. Acquire multi-source data related to the energy consumption of central air conditioning in commercial buildings, including time-series characteristic data, static characteristic data, and energy consumption data; preprocess the data.

[0007] S2. Process time series feature data using the Transformer model, generate time series feature importance weights, and filter key time series features;

[0008] S3. Process static feature data through graph neural networks (GNNs), generate static feature importance weights, and filter key static features;

[0009] S4. Combine key time-series features and key static features to construct an energy consumption prediction model; based on the time-series model, predict key time-series feature data to further predict energy consumption data;

[0010] S5. Define energy consumption cost and numerical / categorical key time series feature conversion cost. With the goal of minimizing the total cost, use a mixed integer optimization algorithm to solve the problem in combination with constraints, and output the optimal target state of key time series features.

[0011] In S1, the time-series feature data includes, but is not limited to, the following features: ambient temperature and humidity, solar radiation, wind speed, air conditioning temperature, air conditioning wind speed, crowd density, and corresponding timestamps;

[0012] The static feature data includes, but is not limited to, the following features: room orientation, room floor, room location, and room size;

[0013] The preprocessing includes, but is not limited to, missing value handling, outlier removal, data standardization and alignment, and encoding.

[0014] S2 contains the following:

[0015] S201: The time step is represented by t, and the time series feature vector is... Inputting the Transformer encoder layer, the association weights between features are calculated through a multi-head self-attention mechanism, and the average importance weight of each temporal feature when used as a query is calculated: Where i represents a temporal feature for which average importance needs to be calculated, N is the total dimension of the temporal features, j represents the number of temporal features traversed, H is the number of attention heads, and h is a specific attention head. Let be the attention weight of temporal feature i (as query) on temporal feature j (as key) in the h-th attention head;

[0016] S202: Extract the attention weight matrix and generate a temporal feature importance weight vector through global average pooling; set a weight threshold θ1 (θ1 is determined by minimizing the prediction error on the validation set, with a value range of 0.1-0.3), and retain temporal features with weight vector values ​​greater than θ1 as key temporal features; key temporal features are represented as: [f1, f2, ..., f p ]; where p is a positive integer representing the number of key time-series features, f1, f2, ..., f p These represent the 1st, 2nd, ..., pth key time-series features, respectively.

[0017] S3 includes the following:

[0018] S301: Perform GNN modeling, constructing the static features into a node feature matrix H of the GNN. (0)The association relationships between static features are constructed into an original adjacency matrix A. After adding self-loops to A, symmetric normalization is performed to obtain... The formula is: in, I is the identity matrix; express The degree matrix;

[0019] S302: Transfer the node feature matrix H (0) and symmetric normalized adjacency matrix Input GAT and perform static feature attention weight calculation. The formula for calculating the feature matrix of layer l is: Among them, H (l) Let W be the feature matrix of the l-th layer. (l) Let σ be the trainable weight matrix, and σ be the ReLU activation function.

[0020] S303: Extract the attention weight matrix of GAT, and generate a static feature importance weight vector through global average pooling; set a weight threshold θ2, and retain static features with values ​​greater than θ2 in the weight vector as key static features; key static features are represented as: [F1, F2, ..., F q ]; where q is a positive integer representing the number of key static features, F1, F2, ..., F q These represent the 1st, 2nd, ..., qth key static features, respectively.

[0021] S4 includes the following:

[0022] S401: For each time step t, the key time series feature vector [f] is... 1,t ,f 2,t ,...,f p,t ] and key static feature vectors [F1,F2,...,F q The vectors are directly concatenated to form the total feature vector X at each time step. t ∈R p+q ;

[0023] S402: with {X t} t＝1 M (M is the total number of time steps) is the independent variable, and the energy consumption data for the same period {Y} is the independent variable. t} t＝1 M (Y t Using the actual energy consumption at time step t as the dependent variable, a fully connected neural network (FCN) regression model is constructed.

[0024] S403: The Adam optimizer and mean squared error (MSE) loss function are used. The training set, validation set and test set are divided in time order. An early stopping strategy is introduced to train the model and obtain the energy consumption prediction model.

[0025] S404: Long Short-Term Memory (LSTM) network is selected as the time series prediction model, using historical key time series feature data {[f 1,t ,f 2,t ,...,f p,t ]} t＝1 M As input, predict the key temporal feature data for the next time step {[f 1,t ,f 2,t ,...,f p,t ]} pre ;

[0026] S405: The predicted future key time-series feature data {[f 1,t ,f 2,t ,...,f p,t ]} pre Concatenate with key static feature data to form the total feature vector X for the next time step. pre ; X pre As input, the predicted energy consumption data Y for the next time step is obtained by combining it with the energy consumption prediction model. pre ;

[0027] S5 includes the following:

[0028] S501: Define energy cost C Y =λ·t·Y pre Where λ represents the energy consumption cost coefficient;

[0029] S502: When the key timing feature f a When the feature is numerical, the current state is represented as f. a,curr The target state is f a,target Characteristic change Δf a =f a,target -f a,curr Then the cost of transforming numerical features is: C(f a )=Fu a (Δf a ); where a∈{1,2,…,p}; Fu a () represents key time-series features f a The cost function;

[0030] When the key time series feature f b When the feature is categorical, the set of possible categories is represented as {U1, U2, ..., U...}. VDefine the transformation cost matrix M. b ∈R V×V M b [u][v] indicates that it comes from category U u Convert to Category U v The cost, and M b If [u][u] = 0, then the categorical feature conversion cost is: C(f b ) = M b [u][v]; where b∈{1,2,…,p}; u,v∈{1,2,…,V}; V represents the key temporal feature f. b Number of categories;

[0031] When key time-series features are transformed, the transformed key time-series features are concatenated with key static feature data and used as input. This is then combined with the energy consumption prediction model to obtain the predicted transformed energy consumption data Y. trans ;

[0032] S503: Minimize total cost: minC total =-λ·t·Y pre +λ·t·Y trans +Σ β＝1 pC(f β );

[0033] Define constraints: The transition states of numerical key time-series features are within a preset value range; the transition states of categorical key time-series features are within their set of selectable categories.

[0034] A mixed-integer optimization algorithm (such as Mixed-Integer Particle Swarm Optimization (MIPSO) or Mixed-Integer Linear Programming (MILP) is employed to address both numerical features (continuous variables) and categorical features (integer variables), outputting a solution that satisfies C. total The minimum set of key temporal feature target states is fed back to the staff;

[0035] A data analysis-based energy consumption prediction system for central air conditioning in commercial buildings, comprising a data preprocessing module, a time-series feature module, a static feature module, a prediction model module, and an optimization decision module;

[0036] The data preprocessing module is used to acquire multi-source data related to the energy consumption of central air conditioning in commercial buildings, including time-series feature data, static feature data, and energy consumption data; and preprocesses the data. The time-series feature module is used to process the time-series feature data using a Transformer model, generate time-series feature importance weights, and filter key time-series features. The static feature module is used to process the static feature data using a Graph Neural Network (GNN), generate static feature importance weights, and filter key static features. The prediction model module is used to concatenate key time-series features and key static features to construct an energy consumption prediction model; and predicts key time-series feature data based on the time-series model to further predict energy consumption data. The optimization decision module is used to define energy consumption cost and the conversion cost of numerical / categorical key time-series features, aiming to minimize the total cost, and solves the problem using a mixed-integer optimization algorithm in combination with constraints, outputting the optimal target state of the key time-series features.

[0037] The data preprocessing module includes a data acquisition unit and a cleaning and alignment unit;

[0038] The data acquisition unit is used to collect time-series characteristic data, static characteristic data and energy consumption data of commercial building central air conditioning from multi-source systems; the cleaning and alignment unit is used to handle data quality issues, including filling missing values, removing outliers, standardizing numerical data, aligning timestamps, and encoding and converting categorical features.

[0039] The time-series feature module includes a weight calculation unit and a feature filtering unit;

[0040] The weight calculation unit is used to analyze the dynamic correlation between various temporal features and calculate the feature importance weights through the multi-head self-attention mechanism of the Transformer model; the feature filtering unit is used to filter key temporal features according to the importance weight threshold.

[0041] The static feature module includes a graph construction unit, a weight calculation unit, and a feature filtering unit;

[0042] The graph construction unit is used to construct the static features of the building into a graph structure, creating a feature node matrix and a feature association adjacency matrix; the weight calculation unit is used to analyze the spatial correlation between static features through a graph attention network and calculate the importance weight of the building attribute features; the feature filtering unit is used to filter key static features according to the importance weight threshold.

[0043] The prediction model module includes a feature splicing unit, a model building unit, a time series prediction unit, and an energy consumption prediction unit.

[0044] The feature concatenation unit is used to fuse key temporal features and key static features to construct a spatiotemporal joint feature vector; the model building unit is used to train a fully connected neural network regression model and establish a mapping relationship between features and energy consumption; the temporal prediction unit is used to predict the changing trend of key temporal features in the future through an LSTM model; the energy consumption prediction unit is used to combine the predicted features and static features to output the predicted energy consumption value for the future time step.

[0045] The optimization decision module includes a cost definition unit and an optimization solution unit;

[0046] The cost definition unit is used to construct the energy consumption cost model and the feature transformation cost model; the optimization solution unit is used to solve the total cost minimization problem under constraints and output the optimal target state of key time series features.

[0047] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention uses Transformer to capture the dynamic correlation between temporal features and GNN to mine the spatial correlation between static features. Then, it generates importance weights for the two types of features and selects key features, avoiding information distortion and redundant interference when traditionally processing multi-source features. This provides accurate core feature input for energy consumption prediction and solidifies the foundation for prediction accuracy. This invention uses a two-step prediction logic: first, it uses LSTM to predict future key temporal features, and then it combines static features with the energy consumption model. This dynamically adapts to changes in the operating conditions of temporal features and avoids the problem of ignoring the dynamic nature of temporal features when directly predicting energy consumption in the traditional way. This invention defines a total cost function of "energy consumption cost + numerical / categorical feature conversion cost" and uses a mixed integer optimization algorithm to take into account both types of parameter solving and constraints. This breaks through the limitation of the traditional single objective of only pursuing energy consumption reduction and can achieve the minimization of the total cost of "energy consumption-conversion", improving the practical application benefits of the optimization scheme. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the process of a data analysis-based energy consumption prediction system for central air conditioning in commercial buildings according to the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example: Figure 1As shown, this invention provides a technical solution: a method for predicting the energy consumption of central air conditioning in commercial buildings based on data analysis. The method includes the following steps:

[0051] S1. Acquire multi-source data related to the energy consumption of central air conditioning in commercial buildings, including time-series characteristic data, static characteristic data, and energy consumption data; preprocess the data.

[0052] S2. Process time series feature data using the Transformer model, generate time series feature importance weights, and filter key time series features;

[0053] S3. Process static feature data through graph neural networks (GNNs), generate static feature importance weights, and filter key static features;

[0054] S4. Combine key time-series features and key static features to construct an energy consumption prediction model; based on the time-series model, predict key time-series feature data to further predict energy consumption data;

[0055] S5. Define energy consumption cost and numerical / categorical key time series feature conversion cost. With the goal of minimizing the total cost, use a mixed integer optimization algorithm to solve the problem in combination with constraints, and output the optimal target state of key time series features.

[0056] In S1, the time-series feature data includes, but is not limited to, the following features: ambient temperature and humidity, solar radiation, wind speed, air conditioning temperature, air conditioning wind speed, crowd density, and corresponding timestamps;

[0057] The static feature data includes, but is not limited to, the following features: room orientation, room floor, room location, and room size;

[0058] The preprocessing includes, but is not limited to, missing value handling, outlier removal, data standardization and alignment, and encoding.

[0059] S2 contains the following:

[0060] S201: The time step is represented by t, and the time series feature vector is... Inputting the Transformer encoder layer, the association weights between features are calculated through a multi-head self-attention mechanism, and the average importance weight of each temporal feature when used as a query is calculated: Where i represents a temporal feature for which average importance needs to be calculated, N is the total dimension of the temporal features, j represents the number of temporal features traversed, H is the number of attention heads, and h is a specific attention head. Let be the attention weight of temporal feature i (as query) on temporal feature j (as key) in the h-th attention head;

[0061] S202: Extract the attention weight matrix and generate a temporal feature importance weight vector through global average pooling; set a weight threshold θ1 (θ1 is determined by minimizing the prediction error on the validation set, with a value range of 0.1-0.3), and retain temporal features with weight vector values ​​greater than θ1 as key temporal features; key temporal features are represented as: [f1, f2, ..., f p ]; where p is a positive integer representing the number of key time-series features, f1, f2, ..., f p These represent the 1st, 2nd, ..., pth key time-series features, respectively.

[0062] S3 includes the following:

[0063] S301: Perform GNN modeling, constructing the static features into a node feature matrix H of the GNN. (0) The association relationships between static features are constructed into an original adjacency matrix A. After adding self-loops to A, symmetric normalization is performed to obtain... The formula is: in, I is the identity matrix; express The degree matrix;

[0064] S302: Transfer the node feature matrix H (0) and symmetric normalized adjacency matrix Input GAT and perform static feature attention weight calculation. The formula for calculating the feature matrix of layer l is: Among them, H (l) Let W be the feature matrix of the l-th layer. (l) Let σ be the trainable weight matrix, and σ be the ReLU activation function.

[0065] S303: Extract the attention weight matrix of GAT, and generate a static feature importance weight vector through global average pooling; set a weight threshold θ2, and retain static features with values ​​greater than θ2 in the weight vector as key static features; key static features are represented as: [F1, F2, ..., F q ]; where q is a positive integer representing the number of key static features, F1, F2, ..., F q These represent the 1st, 2nd, ..., qth key static features, respectively.

[0066] S4 includes the following:

[0067] S401: For each time step t, the key time series feature vector [f] is... 1,t ,f 2,t ,...,f p,t ] and key static feature vectors [F1,F2,...,F qThe vectors are directly concatenated to form the total feature vector X at each time step. t ∈R p+q ;

[0068] S402: with {X t} t＝1 M (M is the total number of time steps) is the independent variable, and the energy consumption data for the same period {Y} is the independent variable. t} t＝1 M (Y t Using the actual energy consumption at time step t as the dependent variable, a fully connected neural network (FCN) regression model is constructed.

[0069] S403: The Adam optimizer and mean squared error (MSE) loss function are used. The training set, validation set and test set are divided in time order. An early stopping strategy is introduced to train the model; thus, an energy consumption prediction model is obtained.

[0070] S404: Long Short-Term Memory (LSTM) network is selected as the time series prediction model, using historical key time series feature data {[f 1,t ,f 2,t ,...,f p,t ]} t＝1 M As input, predict the key temporal feature data for the next time step {[f 1,t ,f 2,t ,...,f p,t ]} pre ;

[0071] S405: The predicted future key time-series feature data {[f 1,t ,f 2,t ,...,f p,t ]} pre Concatenated with key static feature data, forming the total feature vector X for the next time step. pre ; X pre As input, the predicted energy consumption data Y for the next time step is obtained by combining it with the energy consumption prediction model. pre ;

[0072] S5 includes the following:

[0073] S501: Define energy cost C Y =λ·t·Y pre Where λ represents the energy consumption cost coefficient;

[0074] S502: When the key timing feature f a When the feature is numerical, the current state is represented as f. a,curr The target state is f a,targetCharacteristic change Δf a =f a,target -f a,curr Then the cost of transforming numerical features is: C(f a )=Fu a (Δf a ); where a∈{1,2,…,p}; Fu a () represents key time-series features f a The cost function;

[0075] When the key time series feature f b When the feature is categorical, the set of possible categories is represented as {U1, U2, ..., U...}. V Define the transformation cost matrix M. b ∈R V×V M b [u][v] indicates that it comes from category U u Convert to Category U v The cost, and M b If [u][u] = 0, then the categorical feature conversion cost is: C(f b ) = M b [u][v]; where b∈{1,2,…,p}; u,v∈{1,2,…,V}; V represents the key temporal feature f. b Number of categories;

[0076] When key time-series features are transformed, the transformed key time-series features are concatenated with key static feature data and used as input. This is then combined with the energy consumption prediction model to obtain the predicted transformed energy consumption data Y. trans ;

[0077] S503: Minimize total cost: minC total =-λ·t·Y pre +λ·t·Y trans +Σ β＝1 pC(f β );

[0078] Define constraints: The transition states of numerical key time-series features are within a preset value range; the transition states of categorical key time-series features are within their set of selectable categories.

[0079] A mixed-integer optimization algorithm (such as Mixed-Integer Particle Swarm Optimization (MIPSO) or Mixed-Integer Linear Programming (MILP) is employed to address both numerical features (continuous variables) and categorical features (integer variables), outputting a solution that satisfies C. total The minimum set of key temporal feature target states is fed back to the staff;

[0080] A data analysis-based energy consumption prediction system for central air conditioning in commercial buildings, comprising a data preprocessing module, a time-series feature module, a static feature module, a prediction model module, and an optimization decision module;

[0081] The data preprocessing module is used to acquire multi-source data related to the energy consumption of central air conditioning in commercial buildings, including time-series feature data, static feature data, and energy consumption data; and preprocesses the data. The time-series feature module is used to process the time-series feature data using a Transformer model, generate time-series feature importance weights, and filter key time-series features. The static feature module is used to process the static feature data using a Graph Neural Network (GNN), generate static feature importance weights, and filter key static features. The prediction model module is used to concatenate key time-series features and key static features to construct an energy consumption prediction model; and predicts key time-series feature data based on the time-series model to further predict energy consumption data. The optimization decision module is used to define energy consumption cost and the conversion cost of numerical / categorical key time-series features, aiming to minimize the total cost, and solves the problem using a mixed-integer optimization algorithm in combination with constraints, outputting the optimal target state of the key time-series features.

[0082] The data preprocessing module includes a data acquisition unit and a cleaning and alignment unit;

[0083] The data acquisition unit is used to collect time-series characteristic data, static characteristic data and energy consumption data of commercial building central air conditioning from multi-source systems; the cleaning and alignment unit is used to handle data quality issues, including filling missing values, removing outliers, standardizing numerical data, aligning timestamps, and encoding and converting categorical features.

[0084] The time-series feature module includes a weight calculation unit and a feature filtering unit;

[0085] The weight calculation unit is used to analyze the dynamic correlation between various temporal features and calculate the feature importance weights through the multi-head self-attention mechanism of the Transformer model; the feature filtering unit is used to filter key temporal features according to the importance weight threshold.

[0086] The static feature module includes a graph construction unit, a weight calculation unit, and a feature filtering unit;

[0087] The graph construction unit is used to construct the static features of the building into a graph structure, creating a feature node matrix and a feature association adjacency matrix; the weight calculation unit is used to analyze the spatial correlation between static features through a graph attention network and calculate the importance weight of the building attribute features; the feature filtering unit is used to filter key static features according to the importance weight threshold.

[0088] The prediction model module includes a feature splicing unit, a model building unit, a time series prediction unit, and an energy consumption prediction unit.

[0089] The feature concatenation unit is used to fuse key temporal features and key static features to construct a spatiotemporal joint feature vector; the model building unit is used to train a fully connected neural network regression model and establish a mapping relationship between features and energy consumption; the temporal prediction unit is used to predict the changing trend of key temporal features in the future through an LSTM model; the energy consumption prediction unit is used to combine the predicted features and static features to output the predicted energy consumption value for the future time step.

[0090] The optimization decision module includes a cost definition unit and an optimization solution unit;

[0091] The cost definition unit is used to construct the energy consumption cost model and the feature transformation cost model; the optimization solution unit is used to solve the total cost minimization problem under constraints and output the optimal target state of key time series features.

[0092] In this embodiment, the central air conditioning energy management system of a large shopping mall adopts this solution to implement energy consumption prediction and optimization:

[0093] Data Acquisition and Preprocessing: Time-series data (hourly ambient temperature and humidity, solar radiation intensity, air conditioning set temperature, fan speed, real-time pedestrian flow and timestamp) and static data (orientation, floor, window location, and building area) of different areas on each floor of the shopping mall (such as the lobby on the first floor, the clothing area on the second floor, and the catering area on the third floor) were collected from different areas. Missing temperature and humidity data were imputed using mean interpolation. Outlier pedestrian density values ​​were removed using the IQR method. Numerical features were standardized. Categorical features such as room orientation were uniquely coded. All data were aligned according to timestamps.

[0094] Temporal feature processing: The preprocessed temporal data is input into the Transformer encoder. The correlation weights between each feature are calculated through the multi-head self-attention mechanism. The feature importance weight vector is obtained by global average pooling. The threshold θ1 is set to filter out ambient temperature, air conditioning setting temperature, and crowd density as key temporal features.

[0095] Static feature processing: Each region is treated as a GNN node, and the node features are static data. An adjacency matrix is ​​constructed based on the spatial location correlation of the regions. After normalization, it is input into the GAT model. Attention weights are calculated through multi-layer feature updates. A threshold θ2 is set to filter out the region size, floor, and orientation as key static features.

[0096] Energy consumption prediction: The key time-series features of each time step are concatenated with the key static features of the corresponding region to construct a fully connected neural network regression model. The dataset is divided by time and trained using an early stop strategy, with historical energy consumption data as the label. At the same time, the LSTM model is used to predict the ambient temperature, air conditioning set temperature and population density for the next hour based on historical key time-series features. These are then concatenated with the static features and input into the energy consumption model to obtain the predicted energy consumption for the next hour.

[0097] Optimization Solution: Define energy consumption cost as the product of predicted energy consumption and unit energy consumption cost coefficient; the conversion cost of numerical features (such as air conditioner set temperature) is set as a quadratic function of temperature change, and the conversion cost of categorical features (such as fan speed setting) is represented by a preset matrix (such as the cost value of switching from high to medium speed); with the goal of minimizing total cost, constrain the air conditioner temperature to 24-28℃ and the fan speed to three levels: low / medium / high, and use a mixed integer optimization algorithm to solve the problem, outputting the optimal air conditioner set temperature and fan speed setting suggestions for each region in the next hour.

[0098] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A data analysis-based method for predicting the energy consumption of central air conditioning systems in commercial buildings, characterized in that: The method includes the following steps: S1. Acquire multi-source data related to the energy consumption of central air conditioning in commercial buildings, including time-series characteristic data, static characteristic data, and energy consumption data; preprocess the data. S2. Process time series feature data using the Transformer model, generate time series feature importance weights, and filter key time series features; S3. Process static feature data through graph neural network (GNN), generate static feature importance weights, and filter key static features; S4. Combine key time-series features and key static features to construct an energy consumption prediction model; based on the time-series model, predict key time-series feature data to further predict energy consumption data; S5. Define energy consumption cost and numerical / categorical key time series feature conversion cost. With the goal of minimizing the total cost, use a mixed integer optimization algorithm to solve the problem in combination with constraints, and output the optimal target state of key time series features. In S1, the time-series feature data includes the following features: ambient temperature and humidity, solar radiation, wind speed, air conditioning temperature, air conditioning wind speed, crowd density, and corresponding timestamps; The static feature data includes the following features: room orientation, room floor, room location, and room size; S4 includes the following: S401: For each time step t, the key time series feature vector [f] is... 1,t ,f 2,t ,...,f p,t ] and key static feature vectors [F1,F2,...,F q The vectors are directly concatenated to form the total feature vector X at each time step. t ∈R p+q Where q is a positive integer, representing the number of key static features, F1, F2, ..., F q These represent the 1st, 2nd, ..., qth key static features, respectively. S402: with As the independent variable, the energy consumption data for the same period. Using Y as the dependent variable, construct a fully connected neural network regression model; where M represents the total number of time steps; Y... t This represents the actual energy consumption at time step t; S403: The model is trained using the Adam optimizer and mean squared error loss function, and the training set, validation set, and test set are divided in time order. An early stopping strategy is introduced to train the model, resulting in an energy consumption prediction model. S404: Long Short-Term Memory (LSTM) network is selected as the time series prediction model, based on historical key time series feature data. As input, predict the key temporal feature data for the next time step {[f 1,t ,f 2,t ,...,f p,t ]} pre Where p is a positive integer, representing the number of key time-series features; S405: The predicted future key time-series feature data {[f 1,t ,f 2,t ,...,f p,t ]} pre Concatenated with key static feature data, forming the total feature vector X for the next time step. pre ; X pre As input, the predicted energy consumption data Y for the next time step is obtained by combining it with the energy consumption prediction model. pre ; S5 includes the following: S501: Define energy cost C Y =λ·t·Y pre Where λ represents the energy consumption cost coefficient; S502: When the key timing feature f a When the feature is numerical, the current state is represented as f. a,curr The target state is f a,target Characteristic change Δf a =f a,target -f a,curr Then the cost of transforming numerical features is: C(f a )=Fu a (Δf a ); where a∈{1,2,…,p}; Fu a () represents key time-series features f a The cost function; When the key time series feature f b When the feature is categorical, the set of possible categories is represented as {U1, U2, ..., U...}. V Define the transformation cost matrix M. b ∈R V×V M b [u][v] indicates that it comes from category U u Convert to Category U v The cost, and M b If [u][u]=0, then the categorical feature transformation cost is: C(f b )=M b [u][v]; where b∈{1,2,…,p}; u,v∈{1,2,…,V}; V represents the key temporal feature f. b Number of categories; When key time-series features are transformed, the transformed key time-series features are concatenated with key static feature data and used as input. This is then combined with the energy consumption prediction model to obtain the predicted transformed energy consumption data Y. trans ; S503: Minimize total cost: ; Define constraints: The transition states of numerical key time-series features are within a preset value range; the transition states of categorical key time-series features are within their set of selectable categories. A mixed-integer optimization algorithm is employed to address both numerical and categorical features, outputting a solution that satisfies C. total The minimum set of key temporal features and target states is fed back to the staff.

2. The method for predicting the energy consumption of central air conditioning in commercial buildings based on data analysis according to claim 1, characterized in that: In S1, the preprocessing includes missing value handling, outlier removal, data standardization and alignment, and encoding.

3. The method for predicting the energy consumption of central air conditioning in commercial buildings based on data analysis according to claim 2, characterized in that: S2 contains the following: S201: The time step is represented by t, and the time series feature vector is... The input is the Transformer encoder layer, which calculates the correlation weights between features through a multi-head self-attention mechanism, and calculates the average importance weight of each temporal feature when used as a query. Where i represents a temporal feature for which average importance needs to be calculated, N is the total dimension of the temporal features, j represents the number of temporal features traversed, H is the number of attention heads, and h is a specific attention head. Let i be the attention weight of temporal feature i on temporal feature j in the h-th attention head; S202: Extract the attention weight matrix and generate a temporal feature importance weight vector through global average pooling; set a weight threshold θ1, and retain temporal features with values ​​greater than θ1 in the weight vector as key temporal features; key temporal features are represented as: [f1, f2, ..., f p ]; where p is a positive integer representing the number of key time-series features, f1, f2, ..., f p These represent the 1st, 2nd, ..., pth key time-series features, respectively.

4. The method for predicting the energy consumption of central air conditioning in commercial buildings based on data analysis according to claim 3, characterized in that: S3 includes the following: S301: Perform GNN modeling, constructing the static features into a node feature matrix of the GNN. The association relationships between static features are constructed into an original adjacency matrix A. After adding self-loops to A, symmetric normalization is performed to obtain... The formula is: ,in, I is the identity matrix; express The degree matrix; S302: Convert the node feature matrix and symmetric normalized adjacency matrix Input GAT and perform static feature attention weight calculation. The formula for calculating the feature matrix of layer l is: , where H (l) Let W be the feature matrix of the l-th layer. (l) For trainable weight matrix, It is the ReLU activation function; S303: Extract the attention weight matrix of GAT, and generate a static feature importance weight vector through global average pooling; set a weight threshold θ2, and retain static features with values ​​greater than θ2 in the weight vector as key static features; key static features are represented as: [F1, F2, ..., F q ]; where q is a positive integer representing the number of key static features, F1, F2, ..., F q These represent the 1st, 2nd, ..., qth key static features, respectively.

5. A data analysis-based energy consumption prediction system for central air conditioning in commercial buildings, applied to the data analysis-based energy consumption prediction method for central air conditioning in commercial buildings as described in any one of claims 1-4, characterized in that: The system includes a data preprocessing module, a time-series feature module, a static feature module, a prediction model module, and an optimization decision module; The data preprocessing module is used to acquire multi-source data related to the energy consumption of central air conditioning in commercial buildings, including time-series feature data, static feature data, and energy consumption data; and to preprocess the data; the time-series feature module is used to process the time-series feature data through the Transformer model, generate time-series feature importance weights, and filter key time-series features. The static feature module is used to process static feature data through a graph neural network (GNN), generate static feature importance weights, and filter key static features. The prediction model module is used to concatenate key time-series features and key static features to construct an energy consumption prediction model. Based on the time series model, it predicts key time-series feature data and further predicts energy consumption data. The optimization decision module is used to define energy consumption cost and numerical / categorical key time-series feature conversion cost, aiming to minimize the total cost. It uses a mixed-integer optimization algorithm to solve the problem based on constraints and outputs the optimal target state of the key time-series features.

6. The data analysis-based energy consumption prediction system for central air conditioning in commercial buildings according to claim 5, characterized in that: The data preprocessing module includes a data acquisition unit and a cleaning and alignment unit; The data acquisition unit is used to collect time-series characteristic data, static characteristic data and energy consumption data of commercial building central air conditioning from multi-source systems; the cleaning and alignment unit is used to handle data quality issues, including filling missing values, removing outliers, standardizing numerical data, aligning timestamps, and encoding and converting categorical features.

7. The data analysis-based energy consumption prediction system for central air conditioning in commercial buildings according to claim 6, characterized in that: The time-series feature module includes a weight calculation unit and a feature filtering unit; The weight calculation unit is used to analyze the dynamic correlation between various temporal features and calculate the feature importance weights through the multi-head self-attention mechanism of the Transformer model. The feature filtering unit is used to filter key time-series features based on importance weight thresholds; The static feature module includes a graph construction unit, a weight calculation unit, and a feature filtering unit; The graph construction unit is used to construct the static features of the building into a graph structure, creating a feature node matrix and a feature association adjacency matrix; the weight calculation unit is used to analyze the spatial correlation between static features through a graph attention network and calculate the importance weight of the building attribute features. The feature filtering unit is used to filter key static features based on importance weight thresholds.

8. The data analysis-based energy consumption prediction system for central air conditioning in commercial buildings according to claim 7, characterized in that: The prediction model module includes a feature splicing unit, a model building unit, a time series prediction unit, and an energy consumption prediction unit. The feature concatenation unit is used to fuse key temporal and static features to construct a spatiotemporal joint feature vector; the model building unit is used to train a fully connected neural network regression model to establish a mapping relationship between features and energy consumption; the temporal prediction unit is used to predict the future changing trends of key temporal features using an LSTM model; the energy consumption prediction unit is used to combine predicted features and static features. Output the predicted energy consumption value for the future time step; The optimization decision module includes a cost definition unit and an optimization solution unit; The cost definition unit is used to construct the energy consumption cost model and the feature transformation cost model; the optimization solution unit is used to solve the total cost minimization problem under constraints and output the optimal target state of key time series features.

Images