Park carbon emission dynamic prediction method based on BIM and CNN-GRU fusion

Abstract

The invention relates to a park carbon emission dynamic prediction method based on BIM and CNN-GRU fusion, and the method comprises the following steps: S1, constructing an integrated high-precision three-dimensional digital model, and accessing and integrating an external dynamic data source; s2, constructing the processed data into a feature matrix with a space-time structure; s3, performing sliding convolution operation on the feature matrix, and automatically learning and extracting spatial local correlation and combined features among the influence factors; s4, inputting the time sequence features which are extracted by the convolutional neural network and contain the spatial information into a gating circulation unit network, and capturing a long-term dependency relationship and a dynamic evolution rule by a gating circulation unit through an internal update gate and reset gate structure; s5, an end-to-end CNN-GRU fusion prediction model is constructed; and S6, deploying the combined prediction model to an actual application environment, and outputting a park carbon emission prediction value. According to the method, the limitation that a traditional model is difficult to process park complexity and multi-source data can be overcome, and the prediction accuracy and real-time performance are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of smart park operation management and carbon emission prediction technology, and more specifically, to a dynamic carbon emission prediction method for parks based on the fusion of BIM and CNN-GRU. Background Technology

[0002] As the core unit of urban energy consumption and carbon emissions, the refined management and low-carbon operation of industrial parks are crucial for achieving overall emission reduction targets. Carbon emissions from industrial parks are a complex system driven by multiple factors, including building physics, electromechanical systems, dynamic energy consumption, external meteorological conditions, and human activity patterns, exhibiting high dimensionality, strong time-varying characteristics, and nonlinearity. Therefore, achieving accurate and dynamic prediction of carbon emissions from industrial parks is a prerequisite for formulating scientific energy-saving strategies, optimizing energy dispatch, and evaluating emission reduction effectiveness, and its importance is increasingly evident.

[0003] Traditional carbon emission prediction relies heavily on physical models, empirical formulas, or statistical methods such as linear regression. While these methods offer relatively simple model structures and are easy to implement, they struggle to effectively capture the nonlinear coupling relationships between multiple variables and have stringent requirements for data quality and completeness. When faced with the complex and ever-changing operational conditions of industrial parks, their models exhibit poor generalization ability, and their prediction accuracy and dynamic adaptability fail to meet the refined management needs of modern smart parks.

[0004] With the development of artificial intelligence technology, deep learning has provided new solutions to the aforementioned problems. Recurrent Neural Networks (RNNs), represented by Gated Recurrent Units (GRUs), demonstrate advantages in processing time-series data and capturing long-term dependencies in dynamic information such as energy consumption and weather. However, they have limitations when processing gridded spatial data, failing to effectively analyze the complex influences of spatial factors such as building layout and equipment distribution. Conversely, Convolutional Neural Networks (CNNs) excel at extracting spatial local correlations in data and effectively uncovering implicit features in building morphology and physical attributes. However, they lack the ability to memorize time series data and struggle to independently capture the dynamic evolution of carbon emissions.

[0005] Building Information Modeling (BIM), as a digital platform carrying information throughout the entire building lifecycle, provides an ideal data source foundation for high-precision prediction. Although some research has begun to attempt to combine BIM with deep learning, it is still in its early stages: First, the data utilization level is relatively shallow, with most studies only using BIM as a static database of building parameters, failing to fully explore its inherent deep spatial topological relationships and physical attributes; second, the fusion strategies are relatively rudimentary, with existing methods mostly using simple feature stitching for data fusion, lacking an effective mechanism for deep interaction and coupling of spatial and temporal features, resulting in the data advantages of BIM not being fully transformed into the predictive performance of the model.

[0006] Existing technologies generally suffer from insufficient spatiotemporal feature fusion, difficulties in integrating multi-source heterogeneous data, and room for improvement in the robustness and adaptability of models to complex working conditions. Therefore, there is an urgent need in this field for a novel method that can deeply couple the advantages of BIM's refined spatial data with the spatiotemporal feature learning capabilities of the CNN-GRU model. This would break through the bottlenecks of traditional models, achieve high-precision and timely dynamic prediction of carbon emissions in industrial parks, and provide key technical support for the scientific and low-carbon management of industrial parks. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a dynamic prediction method for carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, which can significantly improve the accuracy and real-time performance of predictions and provide reliable data support and decision-making basis for the optimization of energy systems and the planning of emission reduction pathways in industrial parks.

[0008] The technical solution adopted by this invention to solve its technical problem is: to construct a dynamic prediction method for carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, including the following steps: S1. Construct an integrated high-precision three-dimensional digital model, access and integrate external dynamic data sources, and form a basic database of carbon emissions in the park; S2. Perform preprocessing operations on the multi-source heterogeneous data in the basic database; normalize the continuous numerical data in the basic database to eliminate the influence of dimensions; construct a feature matrix with spatiotemporal structure from the processed data as input for subsequent deep learning models. S3. Input the feature matrix into the convolutional neural network, and use the convolution kernel of the convolutional neural network to perform sliding convolution operation on the feature matrix to automatically learn and extract the spatial local correlation and combination features between various influencing factors. S4. Input the time series features containing spatial information extracted by the convolutional neural network into the gated recurrent unit network. The gated recurrent unit captures the long-term dependence and dynamic evolution of the time series features in the time dimension through the internal update gate and reset gate structure, and learns the pattern of carbon emissions changing over time. S5. Stack convolutional neural network and gated recurrent unit network layers to build an end-to-end CNN-GRU fusion prediction model. Connect a fully connected layer at the end of the model to map the hidden state output by the gated recurrent unit to the final carbon emission prediction value. Use historical datasets to supervise the training of the fusion prediction model. Define a loss function to measure the difference between the predicted value and the true value. Use an optimization algorithm to continuously adjust the network parameters to minimize the loss function. S6. Deploy the trained CNN-GRU fusion prediction model to the actual application environment and output the predicted carbon emissions of the park for one or more future time steps.

[0009] According to the above scheme, in step S1, an integrated high-precision three-dimensional digital model is constructed using the BIM model; BIM model data integration includes: building geometry and physical information, electromechanical system information, and dynamic operation and environmental information; The data in the high-precision 3D digital model includes the geometric structure of buildings within the park, the materials of the building envelope, the HVAC system, lighting equipment, and the patterns of human activity. External dynamic data sources include real-time weather data, energy prices, and policy factors.

[0010] According to the above scheme, in step S2, the spatiotemporal feature matrix is ​​constructed. Specifically, it is expressed as follows:

[0011] In the formula, For the previous time step, and These represent the height and width of the gridded spatial layout of the park, respectively, used to characterize the spatial distribution of carbon emission-related factors. This represents the number of feature channels.

[0012] According to the above scheme, in step S3, the convolutional neural network extracts spatial features through multiple convolution and pooling operations. The calculation formula for the convolutional layer is as follows:

[0013] In the formula: For the first The first layer Each output feature map at location Activation value at that location, For the first The input feature map of the layer, For the first Layer In each convolutional kernel, there is a corresponding input channel. The weights, the kernel size is For the first Layer The bias term of each convolution kernel, It is a non-linear activation function.

[0014] According to the above scheme, in step S4, the gated recurrent unit network captures time series dependencies through its internal update and reset gates, expressed as: Reset door : Determines the hidden state at the previous moment How much information needs to be ignored:

[0015] Update Gate Control the hidden state of the previous moment. and the current candidate hidden state For the current final hidden state Degree of impact:

[0016] Candidate hidden state : Calculate the candidate information at the current time step, including the current input. And historical information after resetting the gate filtering:

[0017] Final hidden state By combining the update gate with historical information and current candidate information through linear interpolation, the final output at the current time step is generated.

[0018] In the formula: For the current time step Input; and These are the hidden states of the previous time step and the current time step, respectively. and These are the weight matrices and bias terms corresponding to each gating unit and candidate state calculation. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. This represents a vector concatenation operation. It represents the Hadamardi (or Hadama) stack.

[0019] According to the above scheme, the network architecture of the CNN-GRU fusion model in step S5 is specifically as follows: Input layer: Receives the data constructed in step S2 with... Dimensional data, in which For time step; Convolutional module: Consists of one or more convolutional layers, activation function layers, and pooling layers, performing operations at each time step. Spatial features are extracted from the feature matrix of dimension 1, and the compressed feature vector sequence is output. Recurrent module: Consists of one or more stacked GRU layers, receives the feature vector sequence output by the CNN module, and performs temporal modeling; Output layer: One or more fully connected layers that map the final hidden state of the GRU layer to the carbon emission prediction for one or more future time steps.

[0020] According to the above scheme, model training is performed by minimizing the error between the predicted and the true values. The loss function and parameter update method are as follows: Loss function: The mean squared error is used to quantify the model's predictive performance. The expression is:

[0021] In the formula, The total number of training samples; For the model to the first Predicted carbon emissions for each sample; This corresponds to the actual carbon emissions; This represents the set of all parameters to be optimized in the model; Parameter optimization: An adaptive moment estimation optimization algorithm is used, which calculates the loss function with respect to the parameters. gradient The learning rate for each parameter is dynamically adjusted by combining first-order and second-order moment estimates to update the parameters:

[0022] In the formula, For the first The parameters for the next iteration; The learning rate; and These are the bias correction estimates for the first and second moments of the gradient, respectively; It is a very small constant.

[0023] According to the above scheme, in step S6, the CNN-GRU fusion prediction model receives real-time updated campus data online, and after preprocessing, it is input into the CNN-GRU fusion prediction model for forward propagation calculation.

[0024] According to the above scheme, the predicted carbon emissions of the park are serialized data of the total carbon emissions of the park at a specified future time scale.

[0025] This invention also provides a dynamic prediction system for carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, comprising: The model building module is used to build an integrated high-precision 3D digital model, access and integrate external dynamic data sources, and form a basic database of carbon emissions in the park. The preprocessing module is used to preprocess multi-source heterogeneous data in the basic database; normalize continuous numerical data in the basic database to eliminate the influence of dimensions; and construct a feature matrix with spatiotemporal structure from the processed data. The correlation learning module is used to input the feature matrix into the convolutional neural network, and use the convolution kernel of the convolutional neural network to perform sliding convolution operation on the feature matrix to automatically learn and extract the spatial local correlation and combination features between various influencing factors. The carbon emission change pattern learning module is used to input the time series features containing spatial information extracted by the convolutional neural network into the gated recurrent unit network. The gated recurrent unit captures the long-term dependence and dynamic evolution law of the time series features in the time dimension through the internal update gate and reset gate structure, and learns the pattern of carbon emission change over time. The prediction module stacks convolutional neural network and gated recurrent unit network layers to build an end-to-end CNN-GRU fusion prediction model. A fully connected layer is connected at the end of the model to map the hidden state output by the gated recurrent unit to the final carbon emission prediction value. The fusion prediction model is trained under supervision using historical datasets. A loss function is defined to measure the difference between the predicted value and the true value. An optimization algorithm is used to continuously adjust the network parameters to minimize the loss function. The output module is used to deploy the trained CNN-GRU fusion prediction model to a real-world application environment and output the predicted carbon emissions of the park.

[0026] The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, as described in this invention, has the following beneficial effects: 1. This invention constructs a CNN-GRU cascaded model, utilizing CNN to mine the complex spatial relationships such as building orientation and equipment layout contained in BIM data, and then feeding these feature sequences into GRU for temporal dependency modeling. This deep coupling mechanism of "space first, time later" can more accurately capture the driving factors of carbon emissions. It takes BIM as the core of building a high-fidelity 3D digital model and innovatively transforms its spatial information into a gridded feature matrix suitable for CNN processing. It not only utilizes the geometric and physical properties of BIM, but also transforms its deep spatial topological relationships into effective inputs for the model, turning the data potential of BIM into improved predictive performance. 2. This invention, through end-to-end training of a deep learning model, can learn complex relationships between multiple variables that are highly nonlinear, overcoming the shortcomings of traditional physical models or statistical methods in terms of poor generalization ability and low accuracy. At the same time, once the model is trained, it can be deployed for online dynamic prediction, and can receive the latest data in real time and quickly output highly timely prediction results for multiple future time steps, providing the possibility for "pre-event" refined management of the park, such as optimizing energy scheduling and early warning intervention. 3. The model of this invention learns the inherent laws of carbon emissions in the park by training on historical datasets containing various operating conditions. This makes the model more robust and adaptive to complex and ever-changing actual operating conditions, and can more stably and reliably serve the long-term low-carbon operation of smart parks. 4. This invention is a novel technology that deeply integrates the advantages of BIM's fine spatial data with the spatiotemporal feature learning capabilities of the CNN-GRU model. It effectively solves the core problems of difficulty in fusing multi-source heterogeneous data and incomplete capture of spatiotemporal coupling characteristics, thereby achieving high-precision and high-timeliness dynamic prediction of carbon emissions in the park and providing technical support for achieving scientific energy conservation and emission reduction. Attached Figure Description

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the dynamic prediction method for carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, which is based on the present invention. Detailed Implementation

[0028] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] Example 1 like Figure 1 As shown, the present invention provides a dynamic prediction method for carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, comprising the following steps: S1. BIM Model Construction and Multi-Source Data Integration: Utilizing Building Information Modeling (BIM) technology, a high-precision 3D digital model is constructed that integrates information such as the geometric structure, building envelope materials, HVAC systems, lighting equipment, and human activity patterns of each building within the park. Simultaneously, external dynamic data sources, such as real-time meteorological data, energy prices, and policy factors, are accessed and integrated to form a unified and refined basic database of carbon emissions for the park. Real-time meteorological data includes temperature, humidity, and light intensity.

[0030] BIM model data integration includes building geometry and physical information, electromechanical system information, and dynamic operation and environmental information.

[0031] First, a high-precision 3D digital model of the park was constructed using BIM software such as Autodesk Revit. This high-precision 3D digital model includes the geometric information of the buildings, such as their shape, orientation, and dimensions, as well as detailed physical attribute information, such as: the building envelope, including the material, thickness, thermal conductivity, and heat capacity of walls, roofs, and windows; the equipment models, rated power, energy efficiency ratio (EER / COP), and duct and water pipe layout of the HVAC system; the location, type, and power of the lighting fixtures; and the density patterns of personnel activity in each functional area.

[0032] Subsequently, the BIM model was used as a data platform to integrate multi-source heterogeneous dynamic data. Real-time meteorological data for the park, such as temperature, humidity, solar radiation intensity, and wind speed, was obtained from the meteorological service platform via API interfaces. Real-time operational data for each area and major energy-consuming device was collected through Building Automation System (BAS) or Internet of Things (IoT) platforms using protocols such as Modbus and BACnet, including the start / stop status, actual operating power, and return air temperature of air conditioning units; the status of lighting switches and electricity consumption in each area; and water and gas consumption. This data was then correlated with the spatial information in the BIM model, providing a high-fidelity data foundation for refined and dynamic carbon emission prediction, forming a unified, spatiotemporally labeled basic database.

[0033] S2. Data Preprocessing and Feature Matrix Construction: Preprocessing operations such as cleaning, denoising, and missing value imputation are performed on the multi-source heterogeneous data integrated in S1. Continuous numerical data, such as temperature and power, are normalized to eliminate the influence of dimensions. The processed data are then used to construct a feature matrix with a spatiotemporal structure. The spatial dimension reflects the distribution of different areas or equipment within the park, while the temporal dimension constitutes sequential data, which serves as the input for subsequent deep learning models.

[0034] Constructed spatiotemporal feature matrix Specifically, it is expressed as follows:

[0035] In the formula, Represents the current time step. and These represent the height and width of the park's spatial layout after it has been gridded, and are used to characterize the spatial distribution of carbon emission-related factors. This represents the number of feature channels, with each channel corresponding to a specific data point affecting carbon emissions. This multidimensional feature matrix aligns data from different sources in the spatiotemporal dimension, forming a structured input suitable for CNN models.

[0036] S3. Spatial Feature Extraction Based on Convolutional Neural Network (CNN): The feature matrix generated in step S2 for each time step is input into the CNN. The CNN's convolutional kernels are used to perform sliding convolution operations on the feature matrix, automatically learning and extracting the spatial local correlations and combined features among various influencing factors. For example, the combined effect of building orientation, window-to-wall ratio, and light intensity on energy consumption in a specific area.

[0037] The spatial feature extraction process of a Convolutional Neural Network (CNN) is accomplished through multiple layers of convolution and pooling operations. The calculation formula for its convolutional layers is as follows:

[0038] In the formula: For the first The first layer Each output feature map at location The activation value at that location; For the first The input feature map for the first layer is the one constructed in step S2. . For the first Layer In each convolutional kernel, there is a corresponding input channel. The weights. The kernel size is... . For the first Layer Bias terms for each convolution kernel. For nonlinear activation functions, the modified linear unit (ReLU) function is typically used.

[0039] Each of the feature matrix sequences generated in step S2 The input is fed into the CNN module. In this embodiment, the convolutional neural network (CNN) module consists of two 2D convolutional layers (Conv2D) and two max pooling layers (MaxPooling2D) stacked alternately.

[0040] The first convolutional layer uses 32 3x3 convolutional kernels, employing a sliding window operation to capture basic local spatial features such as "high light intensity" and "large window-to-wall ratio." This is followed by a 2x2 max-pooling layer to downsample the feature map, preserving the most salient features and reducing computational cost. The second convolutional layer uses 64 3x3 convolutional kernels, operating on higher-level feature maps to learn more complex spatial features, such as the combined effects of "high population density," "high equipment power," and "low building envelope insulation performance" within a given region. Finally, another 2x2 max-pooling layer completes the process. ReLU is used as the activation function after all convolutional layers to introduce non-linearity. Through this process, the CNN extracts the high-dimensional spatial feature matrix at each time step t. This effectively compresses the data into a one-dimensional feature vector containing key spatial combination information. Through this process, the model can automatically learn and extract local spatial combination features that have a significant impact on carbon emissions from the raw data.

[0041] S4. Time-series dependency modeling based on Gated Recurrent Unit (GRU): The time-series features containing spatial information extracted by the CNN, i.e., a feature vector corresponding to each time step, are input into the GRU network. The GRU effectively captures the long-term dependencies and dynamic evolution patterns of time-series features in the time dimension through its internal update and reset gate structures, thereby learning the pattern of carbon emissions changing over time.

[0042] Gated Recurrent Unit (GRU) networks capture time-series dependencies through their internal update and reset gates, calculated as follows: Reset Gate : Determines the hidden state at the previous moment How much information needs to be ignored:

[0043] Update Gate Control the hidden state of the previous moment. and the current candidate hidden state For the current final hidden state Degree of impact:

[0044] Candidate Hidden State : Calculate the candidate information for the current time step, including the current input. And historical information after resetting the gate filtering:

[0045] Final Hidden State By combining the update gate with historical information and current candidate information through linear interpolation, the final output at the current time step is generated.

[0046] In the formula: For the current time step The input is the feature vector extracted by the CNN in step S3. and These are the hidden states of the previous time step and the current time step, respectively. and These are the weight matrices and bias terms corresponding to each gating unit and candidate state calculation. It is the Sigmoid activation function. It is the hyperbolic tangent activation function. This represents a vector concatenation operation. This represents the Hadamard product (element-wise product).

[0047] S5. Training and Optimization of the CNN-GRU Fusion Model: CNN and GRU network layers are stacked to construct an end-to-end CNN-GRU fusion prediction model. A fully connected layer is connected at the end of the model to map the hidden states of the GRU output to the final carbon emission prediction value. The fusion model is trained under supervision using a historical dataset. A loss function is defined to measure the difference between the predicted and true values, and an optimization algorithm is used to continuously adjust the network parameters (weights and biases) to minimize the loss function.

[0048] The network architecture of the CNN-GRU fusion model is as follows: Input layer: Receives the data constructed in step S2, which has... Dimensional data, in which For time step.

[0049] Convolutional Module (CNN): Consists of one or more convolutional layers, activation function layers, and pooling layers, responsible for processing data at each time step. Spatial features are extracted from the feature matrix of dimension 1, and the compressed feature vector sequence is output.

[0050] Recurrent Module (GRU): Consists of one or more stacked GRU layers, which receive the feature vector sequence output by the CNN module and perform temporal modeling on it.

[0051] Output layer: Typically one or more fully connected (dense) layers that map the final hidden state of the GRU layer to carbon emission predictions for one or more future time steps.

[0052] Stack the CNN and GRU modules mentioned above, and connect one or more fully connected (Dense) layers at the end of the GRU module to finally output one or more predicted values, corresponding to the carbon emissions at one or more future time steps.

[0053] During model training, the historical dataset was divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The model was trained using the training set, with the weighted mean squared error (WMSE) loss function as defined above, Adam optimizer selected, and an initial learning rate of 0.001. During training, the validation set was used to monitor model performance, prevent overfitting, and adjust hyperparameters (such as learning rate and batch size) accordingly.

[0054] Model training is performed by minimizing the error between the predicted and true values. The loss function and parameter update methods are as follows: Loss Function: The mean squared error (MSE) is used to quantify the model's predictive performance. Its expression is:

[0055] In the formula, The total number of training samples; For the model to the first Predicted carbon emissions for each sample;

[0056] This corresponds to the actual carbon emissions; This represents the set of all parameters to be optimized in the model.

[0057] Parameter optimization: The Adaptive Moment Estimation (Adam) optimization algorithm is used to calculate the loss function with respect to the parameters. gradient The learning rate for each parameter is dynamically adjusted by combining first-order and second-order moment estimates, thus updating the parameters:

[0058] In the formula, For the first The parameters for the next iteration; The learning rate; and These are the bias correction estimates for the first and second moments of the gradient, respectively; It is a very small constant to prevent the denominator from being zero.

[0059] S6. Online Dynamic Prediction and Result Output: Deploy the trained CNN-GRU fusion prediction model to the real-world application environment. Receive real-time updated park data online, such as BIM and meteorological data. After undergoing the same preprocessing steps as in step S2, input the data into the model for forward propagation calculations. Finally, output the predicted carbon emissions for one or more future time steps, providing park managers with dynamic and forward-looking carbon emission data to support energy conservation control and low-carbon decision-making.

[0060] The online dynamic prediction module receives real-time data streams and uses a trained model for forward propagation to output prediction results. The output of the online dynamic prediction is serialized data of the total carbon emissions (unit: kgCO2e) of the park at a specified future time scale (such as the next 1 hour or 24 hours). This data can be directly used to guide the optimized scheduling of the park's energy system, evaluate the effectiveness of energy-saving measures, and issue early warnings of carbon emission exceedances.

[0061] Example 2 This embodiment takes a machinery manufacturing plant in the Yangtze River Delta region as the research object. The plant covers an area of ​​8,000 square meters and has a building area of ​​12,000 square meters. It mainly consists of production workshops (areas A, B, and C), office buildings, a research and development center, and a supporting power station. The core production processes are metal cutting, welding, and assembly. Carbon emissions mainly come from electricity consumption of production equipment, energy consumption of the HVAC system, and natural gas usage (welding processes and winter heating). The plant operates on a three-shift system with an average of 320 workers per day. The actual carbon emissions for the entire year of 2024 were 2,860 tons of CO2e, of which energy consumption of production equipment accounted for 62%, HVAC for 23%, and natural gas consumption for 15%.

[0062] 1. The key basic parameters of the factory are shown in the table below: Table 1 Key Basic Parameters of the Plant

[0063] 2. Algorithm Core Parameter Configuration In this embodiment, a CNN-GRU fusion model is built based on the TensorFlow 2.10 deep learning framework. The algorithm parameters are divided into three categories: network structure parameters, training parameters, and prediction control parameters. All parameters are determined through grid search combined with validation set performance tuning. The specific configuration is shown in the table below:

[0064] 3. Algorithm Execution Process Based on the above parameter configuration, the carbon emission prediction process is divided into four core stages: "data input - model inference - result output - optimization feedback," forming a closed-loop operation mechanism that is automatically executed once per hour. The specific process and key steps are explained below: Step 1: Real-time input of multi-source data (time T0) At time T0 (e.g., 9:00 AM daily), the system automatically triggers a data collection task, synchronously acquiring all data from the previous 24 hours (from T-24 to T0) through three types of interfaces: 1. BIM Static Data Interface: Call the Revit API to read static parameters (such as thermal conductivity of the building envelope and equipment location) that have not been updated between T-24 and T0. After confirming that there are no equipment modifications, connect directly. 2. IoT Dynamic Data Interface: Real-time data such as power consumption (total power of CNC lathe 185kW) and natural gas flow (welding process 120m³) are collected from smart meters, gas meters and equipment controllers via Modbus protocol, and hourly data is generated after averaging. 3. External data interface: Obtain hourly meteorological data from T-24 to T0 from Gaode Weather API (e.g., temperature 28℃ and humidity 65% ​​at T0), and obtain workpiece output (520 pieces) and shift information from the production management system.

[0065] All data are aligned with the triple label of "timestamp-spatial grid ID-feature type" to generate 864 original data records (24 hours × 36 spatial grids).

[0066] Step 2: Data Preprocessing and Feature Matrix Generation (T0+5 minutes) The system invokes a Python preprocessing script to complete data cleaning and feature construction within 5 minutes. 1. Data cleaning: Outliers are identified using the “3σ criterion” (e.g., a welding robot with a power of 35kW > the 3σ upper limit of 32kW), marked as outliers, and replaced with the average value of the previous hour; for missing meteorological data at time T-3, linear interpolation is used to fill in the missing data (interpolating from 26℃ at time T-4 to 27℃ at time T-2). 2. Normalization: Perform normalization calculations on continuous data (e.g., the normalized value of 28℃ is 0.68, formula: (28-5) / (38-5)=0.68) to eliminate the influence of dimensions; 3. Feature Matrix Construction: Divide the space into 20×20 grids, fill the corresponding channels with 12 types of features (ambient temperature, equipment power, etc.) of each grid, and generate a 24×20×20×12 spatiotemporal feature matrix, which is used as the model input data X.

[0067] Stage 3: Model Inference Calculation (T0+10 minutes) 1. Input the feature matrix X into the CNN-GRU model and accelerate inference using TensorRT, completing spatial feature extraction and temporal inference within 10 minutes: CNN spatial feature extraction: The Conv2D-1 layer extracts the associated features of "grid 12 (west-facing workshop) - temperature - equipment power" through sliding calculation with 3×3 convolution kernels, and outputs a 32×10×10 feature map; the Conv2D-2 layer further extracts higher-order features, which are compressed into a 64×5×5 feature map by MaxPooling2D-2, and finally the Reshape layer converts it into a 24×1600 temporal vector sequence; 2. GRU time-series inference: The GRU-1 layer receives a 24×1600 vector sequence, filters key time-series information from T-6 to T0 (such as changes in equipment load rate) through a reset gate, and updates the weight of historical information controlled by the gate; the GRU-2 layer outputs a 64-dimensional hidden state, which is mapped to the carbon emission prediction value for the next 1-6 hours through a fully connected layer, where the predicted value for T0+1 hour (10:00) is 892kgCO2e, and the predicted value for T0+6 hour (15:00) is 915kgCO2e.

[0068] Step 4: Prediction Results Output and Optimization Feedback (T0+15 minutes) The prediction results processing and application are completed in T0+15 minutes (9:15). 1. Results Output: The predicted values ​​are generated into reports in the format of "time-carbon emission value-confidence level" and synchronized to the EMS system (displaying the prediction curve from T0+1 to T0+6 hours), the factory management platform (pushing to the mobile terminal of the manager), and the production scheduling system; 2. Optimization Feedback: For the high carbon emission prediction at T0+6 hours (915kgCO2e > threshold 850kgCO2e), the system automatically triggers an optimization suggestion - adjust the high-energy-consuming welding task from 13:00 to 15:00 to 0:00 to 2:00 the next day (valley time of the power grid), and pushes the suggestion to the production scheduling module; 3. Model Iteration: Record the input data, output results and confidence scores of this prediction as incremental training samples for the model. Retrain the model once a month using the full dataset to update the network parameters.

[0069] Example 3 A dynamic carbon emission prediction system for industrial parks based on the fusion of BIM and CNN-GRU includes: The model building module is used to build an integrated high-precision 3D digital model, access and integrate external dynamic data sources, and form a basic database of carbon emissions in the park. The preprocessing module is used to preprocess multi-source heterogeneous data in the basic database; normalize continuous numerical data in the basic database to eliminate the influence of dimensions; and construct a feature matrix with spatiotemporal structure from the processed data. The correlation learning module is used to input the feature matrix into the convolutional neural network, and use the convolution kernel of the convolutional neural network to perform sliding convolution operation on the feature matrix to automatically learn and extract the spatial local correlation and combination features between various influencing factors. The carbon emission change pattern learning module is used to input the time series features containing spatial information extracted by the convolutional neural network into the gated recurrent unit network. The gated recurrent unit captures the long-term dependence and dynamic evolution law of the time series features in the time dimension through the internal update gate and reset gate structure, and learns the pattern of carbon emission change over time. The prediction module stacks convolutional neural network and gated recurrent unit network layers to build an end-to-end CNN-GRU fusion prediction model. A fully connected layer is connected at the end of the model to map the hidden state output by the gated recurrent unit to the final carbon emission prediction value. The fusion prediction model is trained under supervision using historical datasets. A loss function is defined to measure the difference between the predicted value and the true value. An optimization algorithm is used to continuously adjust the network parameters to minimize the loss function. The output module is used to deploy the trained CNN-GRU fusion prediction model to a real-world application environment and output the predicted carbon emissions of the park.

[0070] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU, characterized in that, Includes the following steps: S1. Construct an integrated high-precision three-dimensional digital model, access and integrate external dynamic data sources, and form a basic database of carbon emissions in the park; S2. Perform preprocessing operations on the multi-source heterogeneous data in the basic database; The continuous numerical data in the basic database is normalized to eliminate the influence of dimensions; the processed data is then used to construct a feature matrix with a spatiotemporal structure. S3. Input the feature matrix into the convolutional neural network, and use the convolution kernel of the convolutional neural network to perform sliding convolution operation on the feature matrix to automatically learn and extract the spatial local correlation and combination features between various influencing factors. S4. Input the time series features containing spatial information extracted by the convolutional neural network into the gated recurrent unit network. The gated recurrent unit captures the long-term dependence and dynamic evolution of the time series features in the time dimension through the internal update gate and reset gate structure, and learns the pattern of carbon emissions changing over time. S5. Stack convolutional neural network and gated recurrent unit network layers to build an end-to-end CNN-GRU fusion prediction model. Connect a fully connected layer at the end of the model to map the hidden state output by the gated recurrent unit to the final carbon emission prediction value. Use historical datasets to supervise the training of the fusion prediction model. Define a loss function to measure the difference between the predicted value and the true value. Use an optimization algorithm to continuously adjust the network parameters to minimize the loss function. S6. Deploy the trained CNN-GRU fusion prediction model into the actual application environment and output the predicted carbon emissions of the park.

2. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 1, characterized in that, In step S1, an integrated high-precision three-dimensional digital model is constructed using the BIM model. BIM model data integration includes: building geometry and physical information, electromechanical system information, and dynamic operation and environmental information; The data in the high-precision 3D digital model includes the geometric structure of buildings within the park, the materials of the building envelope, the HVAC system, lighting equipment, and the patterns of human activity. External dynamic data sources include real-time weather data, energy prices, and policy factors.

3. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 1, characterized in that, In step S2, the feature matrix of the constructed spatiotemporal structure for: In the formula, For the previous time step, and These represent the height and width of the gridded spatial layout of the park, respectively, used to characterize the spatial distribution of carbon emission-related factors. This represents the number of feature channels.

4. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 1, characterized in that, In step S3, the convolutional neural network extracts spatial features through multiple convolutional and pooling operations. The calculation formula for the convolutional layer is as follows: In the formula: For the first The first layer Each output feature map at location Activation value at that location, For the first The input feature map of the layer, For the first Layer In each convolutional kernel, there is a corresponding input channel. The weights, the kernel size is For the first Layer The bias term of each convolution kernel, It is a non-linear activation function.

5. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 1, characterized in that, In step S4, the gated recurrent unit network captures time-series dependencies through its internal update and reset gates, expressed as: Reset door : Determines the hidden state at the previous moment How much information needs to be ignored: Update Gate Control the hidden state of the previous moment. and the current candidate hidden state For the current final hidden state Degree of impact: Candidate hidden state : Calculate the candidate information at the current time step, including the current input. And historical information after resetting the gate filtering: Final hidden state By combining the update gate with historical information and current candidate information through linear interpolation, the final output at the current time step is generated. In the formula: For the current time step Input; and These are the hidden states of the previous time step and the current time step, respectively. and These are the weight matrices and bias terms corresponding to each gating unit and candidate state calculation. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. This represents a vector concatenation operation. It represents the Hadamardi (or Hadama) stack.

6. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 1, characterized in that, In step S5, the network architecture of the CNN-GRU fusion model includes: Input layer: Receives the data constructed in step S2 with... Dimensional data, in which For time step; Convolutional module: Consists of one or more convolutional layers, activation function layers, and pooling layers, performing operations at each time step. Spatial features are extracted from the feature matrix of dimension 1, and the compressed feature vector sequence is output. Recurrent module: Consists of one or more stacked GRU layers, receives the feature vector sequence output by the CNN module, and performs temporal modeling; Output layer: One or more fully connected layers that map the final hidden state of the GRU layer to the carbon emission prediction for one or more future time steps.

7. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 6, characterized in that, Model training is performed by minimizing the error between the predicted and true values. The loss function and parameter update methods are as follows: Loss function: The mean squared error is used to quantify the model's predictive performance. The expression is: In the formula, The total number of training samples; For the model to the first Predicted carbon emissions for each sample; This corresponds to the actual carbon emissions; This represents the set of all parameters to be optimized in the model; Parameter optimization: An adaptive moment estimation optimization algorithm is used, which calculates the loss function with respect to the parameters. gradient The learning rate for each parameter is dynamically adjusted by combining first-order and second-order moment estimates to update the parameters: In the formula, For the first The parameters for the next iteration; The learning rate; and These are the bias correction estimates for the first and second moments of the gradient, respectively; It is a very small constant.

8. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 6, characterized in that, In step S6, the CNN-GRU fusion prediction model receives real-time updated campus data online, and after preprocessing, it is input into the CNN-GRU fusion prediction model for forward propagation calculation.

9. The method for dynamic prediction of carbon emissions in industrial parks based on the fusion of BIM and CNN-GRU as described in claim 8, characterized in that, The predicted carbon emissions of the park are serialized data representing the total carbon emissions of the park at a specified future time scale.

10. A dynamic carbon emission prediction system for industrial parks based on the fusion of BIM and CNN-GRU, characterized in that, include: The model building module is used to build an integrated high-precision 3D digital model, access and integrate external dynamic data sources, and form a basic database of carbon emissions in the park. The preprocessing module is used to perform preprocessing operations on multi-source heterogeneous data in the basic database; The continuous numerical data in the basic database is normalized to eliminate the influence of dimensions; the processed data is then used to construct a feature matrix with a spatiotemporal structure. The correlation learning module is used to input the feature matrix into the convolutional neural network, and use the convolution kernel of the convolutional neural network to perform sliding convolution operation on the feature matrix to automatically learn and extract the spatial local correlation and combination features between various influencing factors. The carbon emission change pattern learning module is used to input the time series features containing spatial information extracted by the convolutional neural network into the gated recurrent unit network. The gated recurrent unit captures the long-term dependence and dynamic evolution law of the time series features in the time dimension through the internal update gate and reset gate structure, and learns the pattern of carbon emission change over time. The prediction module stacks convolutional neural network and gated recurrent unit network layers to build an end-to-end CNN-GRU fusion prediction model. A fully connected layer is connected at the end of the model to map the hidden state output by the gated recurrent unit to the final carbon emission prediction value. The fusion prediction model is trained under supervision using historical datasets. A loss function is defined to measure the difference between the predicted value and the true value. An optimization algorithm is used to continuously adjust the network parameters to minimize the loss function. The output module is used to deploy the trained CNN-GRU fusion prediction model to a real-world application environment and output the predicted carbon emissions of the park.

Images