A new low-carbon city carbon emission collaborative prediction method and system based on digital twinning

Abstract

The application discloses a new low-carbon city carbon emission collaborative prediction method and system based on digital twinning, relates to the technical field of low-carbon city carbon emission prediction, and comprises the following steps: constructing a city multi-source carbon emission data access channel, collecting dynamic streaming data and static benchmark data to form an original data lake, generating normalized carbon emission element data set through standardization cleaning and space-time alignment, and driving a city-level digital twinning basic model to generate a twinning body real-time state sequence. Through the improved attention fusion algorithm introducing a learnable collaborative adjustment factor, an adaptive bias term and a cross-head attention interaction layer, multi-dimensional collaborative features are extracted, a carbon emission collaborative prediction model is used to output multi-time point prediction results, the results are fed back to the digital twinning model to carry out prediction scenario virtual deduction, and a prediction verification closed loop is formed. The method optimizes carbon emission feature extraction and prediction verification process, and improves the accuracy and rationality of city carbon emission collaborative prediction.

Description

Technical Field

[0001] This invention belongs to the field of carbon emission prediction technology for low-carbon cities, specifically a novel collaborative prediction method and system for carbon emissions in low-carbon cities based on digital twins. Background Technology

[0002] Existing technologies for predicting carbon emissions in low-carbon cities acquire carbon emission data from IoT sensors, urban business systems, and statistical platforms through data acquisition channels. After basic data processing, a digital twin model is used to simulate the urban system's operational status. A conventional multi-head attention mechanism is employed to extract carbon emission impact features, and a prediction model is used to calculate carbon emission values, ultimately outputting the prediction results directly. However, urban carbon emission data encompasses both dynamic streaming data and static baseline data, exhibiting differences in data type and spatiotemporal dimensions. Conventional processing methods can only achieve simple data aggregation, failing to achieve unified spatiotemporal labeling and standardization, and the level of data standardization is insufficient to meet the requirements of subsequent model calculations.

[0003] Conventional multi-head attention mechanisms can only perform single-dimensional feature extraction. The attention weight calculation process lacks adjustable parameters and adaptive correction structures, and it also fails to establish interaction links between different feature extraction units, resulting in poor synergistic fusion of multi-dimensional carbon emission impact features. Carbon emission prediction results are only used as single output data and cannot be used to influence digital twin models for virtual simulations. Furthermore, the prediction results lack corresponding verification processes, and the degree of matching between the predicted content and the actual operational status of the city cannot be verified.

[0004] The attention fusion algorithm needs structural improvements to optimize the adaptability of feature extraction and fusion. At the same time, a feedback linkage mechanism between the prediction results and the digital twin model needs to be built to make up for the problems of insufficient feature fusion accuracy and lack of prediction verification process, so as to meet the technical requirements of collaborative prediction of urban carbon emissions. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; To this end, this invention proposes a novel collaborative prediction method for carbon emissions in low-carbon cities based on digital twins, comprising: Construct a data access channel for multi-source carbon emission data in the city, receive and aggregate dynamic streaming data and static benchmark data from IoT sensors, urban business systems and statistical platforms to form a raw carbon emission data lake; The data in the raw carbon emission data lake are standardized, cleaned, and spatiotemporally aligned to generate a standardized urban carbon emission element dataset with unified spatiotemporal identification. Based on the standardized urban carbon emission element dataset, the preset urban-level digital twin basic model is driven to perform simulation evolution, generating a real-time state sequence of the twin that reflects the dynamic operating status of the urban entity system. Based on the real-time state sequence of the twin, an improved attention fusion algorithm is used to extract multi-dimensional collaborative features affecting carbon emissions. The improved attention fusion algorithm is based on a multi-head attention mechanism and introduces a learnable collaborative adjustment factor and an adaptive bias term in its attention weight calculation process, as well as adding a cross-head attention interaction layer. The multi-dimensional collaborative features are input into the carbon emission collaborative prediction model for multi-step forward calculation, and the city carbon emission collaborative prediction results for multiple future time points are output. The collaborative prediction results of urban carbon emissions are fed back to the city-level digital twin basic model, driving the city-level digital twin basic model to perform virtual simulations under the prediction scenario, forming a prediction verification closed loop.

[0006] Furthermore, the data in the raw carbon emission data lake is subjected to standardization cleaning and spatiotemporal alignment to generate a standardized urban carbon emission element dataset with unified spatiotemporal identification, including: Identify the business type and physical dimension of each data record in the raw carbon emission data lake, and convert the values ​​of different physical dimensions into standard carbon dioxide equivalent values ​​according to the preset carbon emission accounting standard. A unified timestamp coding system and a spatial grid coding system are attached to all data records in the raw carbon emission data lake. The minimum time unit of the timestamp coding system is minutes, and the spatial grid coding system adopts a predefined geographical grid division rule. The converted standard carbon dioxide equivalent values ​​are associated with the unified timestamp coding system and spatial grid coding system to form the standardized urban carbon emission element dataset, which is organized by time series and spatial grid with standardized records as units.

[0007] Furthermore, based on the standardized urban carbon emission element dataset, a pre-defined city-level digital twin basic model is driven to perform simulation evolution, generating a real-time state sequence of the twin reflecting the dynamic operating state of the urban entity system, including: In the city-level digital twin basic model, virtual model components corresponding to the city's physical subsystems are established. The virtual model components include energy supply network components, traffic flow network components, building energy consumption components, and industrial production components. The energy consumption data, traffic flow data, building operation data, and industrial output data in the standardized urban carbon emission data set are mapped and input to the energy supply network component, traffic flow network component, building energy consumption component, and industrial production component, respectively. Activate the coupling and interaction relationships between the virtual model components in the city-level digital twin basic model, advance the simulation clock of the virtual model at the rate of real time synchronization, so that each virtual model component can perform dynamic calculation and state update based on input data and interaction relationships; The internal state variables output by each virtual model component are continuously collected as the simulation clock advances. These internal state variables include at least energy flow, traffic flow velocity, building internal environmental parameters, and industrial equipment load, and are arranged in chronological order to form the real-time state sequence of the twin.

[0008] Furthermore, based on the real-time state sequence of the twin, an improved attention fusion algorithm is used to extract multi-dimensional collaborative features affecting carbon emissions, including: A fixed-time-length slice of historical state is extracted from the real-time state sequence of the twin, and the heterogeneous state variables from different virtual model components in the historical state slice are spliced ​​together to form an initial fusion tensor. The initial fusion tensor is simultaneously projected onto multiple independent feature subspaces. Within each feature subspace, the correlation weight of the state information at each time point in the historical state slice relative to the state information at all other time points is calculated. When calculating the relevance weights, an additional learnable co-regulation factor is introduced. This co-regulation factor is generated from the overall statistical features of the historical state slices and is used to dynamically adjust the balance of attention distribution in different feature subspaces. Within each feature subspace, the correlation weights adjusted by the collaborative adjustment factor are applied to the corresponding state information, and a weighted sum is performed to obtain the context-aware feature vector of each feature subspace. The context-aware feature vectors of all feature subspaces are concatenated and then dimensionality-reduced and fused through a linear transformation layer, ultimately outputting a compact multi-dimensional collaborative feature that contains the spatiotemporal dependencies of multiple source states.

[0009] Furthermore, the working principle of the improved attention fusion algorithm includes: During the attention weight calculation process, an adaptive bias term is added to the dot product similarity calculation result of the query vector, key vector, and value vector. The value of the adaptive bias term is determined by the variance and mean of the currently processed historical state slice, and is used to suppress the abnormal attention distribution caused by the difference in the stationarity of the state sequence. After performing parallel attention computation in multiple feature subspaces, a cross-head attention interaction layer is added. The cross-head attention interaction layer receives the intermediate feature maps output by all feature subspaces and learns the potential correlation between the attention focusing patterns of different feature subspaces. The cross-head attention interaction layer, through a lightweight cross-gated network, allows attention feature vectors from different feature subspaces to interact conditionally, thereby enhancing the ability of the finally fused multi-dimensional collaborative features to characterize the nonlinear coupling relationship between different carbon emission drivers.

[0010] Furthermore, the step of inputting the multi-dimensional collaborative features into the carbon emission collaborative prediction model for multi-step look-ahead calculations, and outputting the city's carbon emission collaborative prediction results for multiple future time points, includes: The carbon emission collaborative prediction model includes a temporal feature enhancement network and a multi-step prediction head; The temporal feature deepening network receives the multi-dimensional collaborative features and captures and extracts the deep temporal patterns hidden within the multi-dimensional collaborative features through stacked dilated causal convolutional layers, outputting forward-looking deepened temporal features. The multi-step prediction head contains multiple parallel output branches equal to the number of prediction steps. Each parallel output branch independently receives the deepened temporal features and is mapped to a specific scalar prediction value through a fully connected layer. The scalar prediction values ​​generated by all parallel output branches are arranged in order of their corresponding future time points to form the collaborative prediction results of urban carbon emissions at multiple future time points.

[0011] Furthermore, the temporal feature deepening network receives the multi-dimensional collaborative features and captures and extracts the deep temporal patterns implicit within the multi-dimensional collaborative features through stacked dilated causal convolutional layers, including: Multiple dilated causal convolutional layers with different dilation rates are configured, and the dilation rates are arranged in ascending order. In the order of increasing dilation rates, the dilated causal convolutional layers at the front are used to capture fine-grained, short-period temporal patterns, while the dilated causal convolutional layers at the back are used to capture coarse-grained, long-period temporal patterns. The multi-dimensional collaborative features are used as input and passed sequentially through multiple dilated causal convolutional layers with different dilation rates. In addition to being passed to the next layer, the output of each layer is also fused with the input of a subsequent specific layer through skip connections. After the last dilated causal convolutional layer, a global temporal pooling operation is introduced to aggregate the temporal features refined through multiple layers in the time dimension, extracting global temporal dependency features as the forward-looking, in-depth temporal features.

[0012] Furthermore, the collaborative prediction results of urban carbon emissions are fed back to the city-level digital twin basic model, driving the city-level digital twin basic model to perform virtual simulations under the prediction scenario, forming a prediction verification closed loop, including: The city's carbon emission collaborative prediction results are used as boundary constraints for future time periods and are input in reverse to the corresponding virtual model components in the city-level digital twin basic model. In the city-level digital twin basic model, based on the boundary constraints and combined with the model's built-in operating rules and interaction logic, virtual simulations are performed from the current simulation moment to future time periods to simulate the state evolution paths of each subsystem of the city under the carbon emission scenario defined by the city's collaborative carbon emission prediction results. Record the sequence of key intermediate variables related to carbon emissions generated during the virtual simulation process, and perform a consistency comparison analysis between the sequence of key intermediate variables and recent historical fragments of the twin real-time state sequence on which the collaborative prediction results of urban carbon emissions were based. Specifically, this includes: Extract the state variable subsequence that corresponds in physical meaning to the key intermediate variable sequence from the recent historical fragments of the twin's real-time state sequence; Calculate the degree of difference between the key intermediate variable sequence and the corresponding state variable subsequence on the same statistical features, including mean, variance and slope of change trend; A difference degree evolution curve is established, which depicts the continuous change of the difference degree from a historical moment to a future prediction moment; When the difference evolution curve exhibits a smooth transition near the prediction starting point, it is determined that the virtual inference process is consistent with the historical evolution logic.

[0013] Furthermore, after forming the prediction and verification closed loop, the method further includes: Based on the results of the consistency comparison analysis, the prediction performance evaluation index of the collaborative prediction model for carbon emissions in the current round is calculated. Based on the historical multi-round prediction performance evaluation indicators, the learning rate decay strategy of key trainable parameters in the carbon emission collaborative prediction model is dynamically adjusted. Using the newly acquired standardized urban carbon emission element dataset and corresponding real carbon emission observations, the carbon emission collaborative prediction model is incrementally fine-tuned to adapt to the dynamic evolution of the urban system over time.

[0014] Furthermore, the present invention also includes a novel low-carbon city carbon emission collaborative prediction system based on digital twins. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the novel low-carbon city carbon emission collaborative prediction method based on digital twins described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: In the process of attention weight calculation, a learnable collaborative adjustment factor and an adaptive bias term are introduced. At the same time, a cross-head attention interaction layer is added. The learnable collaborative adjustment factor can dynamically match the intrinsic correlation characteristics of multi-dimensional influencing factors of carbon emissions. The adaptive bias term can reduce the interference of invalid data on feature extraction in the weight allocation stage. The cross-head attention interaction layer can realize the cross-linking of features extracted by different attention heads. The fusion fit of multi-dimensional collaborative features is improved, the correlation and integration effect of feature information of each dimension is optimized, the feature expression of carbon emission influencing factors is closer to the real state of urban system operation, the completeness and pertinence of feature extraction are improved, and the information interaction between different dimensional features is more sufficient, avoiding the information fragmentation problem caused by a single feature extraction structure.

[0016] The collaborative prediction results of urban carbon emissions are fed back to the city-level digital twin basic model, driving the model to complete virtual simulations under the predicted scenarios, forming a complete prediction verification closed loop. The digital twin model can reproduce the dynamic operation virtual scenario of the urban entity system based on the prediction results. The virtual simulation process can verify the adaptability of the prediction results in real time, improve the fit between the prediction scenario and the urban entity's operating status, enhance the credibility of the prediction results, and the state data generated by the virtual simulation can be directly applied to the parameter adaptation and adjustment of the prediction model. The dynamic optimization of the prediction process has a feasible execution path, and the prediction and simulation simulation stages form a coherent technical connection. The coherence of the entire carbon emission prediction process is optimized, the fit between the prediction results and the actual operating status of the city is further improved, and the overall stability and rationality of collaborative prediction are continuously strengthened. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the steps of a novel collaborative prediction method for low-carbon city carbon emissions based on digital twins, as described in this invention. Figure 2 A flowchart for generating a normalized dataset for standardized cleaning and spatiotemporal alignment; Figure 3 A flowchart for generating real-time state sequences of a city-level digital twin based model. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0019] See Figure 1This invention provides a novel collaborative prediction method for carbon emissions in low-carbon cities based on digital twins. The specific method includes: By constructing a data access channel for urban multi-source carbon emission-related data, dynamic streaming data and static benchmark data from IoT sensors, urban business systems, and statistical platforms are received and aggregated to form a raw carbon emission data lake. The data in this raw carbon emission data lake undergoes standardization cleaning and spatiotemporal alignment to generate a standardized urban carbon emission element dataset with unified spatiotemporal identification. Based on this standardized urban carbon emission element dataset, a pre-set city-level digital twin basic model is driven to perform simulation evolution, generating a real-time state sequence of the twin reflecting the dynamic operating state of the urban entity system. Based on this real-time state sequence of the twin, an improved... An improved attention fusion algorithm extracts multi-dimensional collaborative features affecting carbon emissions. This algorithm is based on a multi-head attention mechanism and incorporates learnable collaborative adjustment factors and adaptive bias terms in its attention weight calculation process, as well as adding a cross-head attention interaction layer. The multi-dimensional collaborative features are input into a carbon emission collaborative prediction model for multi-step look-ahead calculation, outputting collaborative prediction results of urban carbon emissions at multiple future time points. The collaborative prediction results of urban carbon emissions are fed back to the city-level digital twin basic model, driving the city-level digital twin basic model to perform virtual simulations under the prediction scenario, forming a prediction verification closed loop.

[0020] In one embodiment of the present invention, see [reference] Figure 2 The process involves identifying the business type and physical dimension of each data record in the raw carbon emission data lake, and converting the values ​​of different physical dimensions into standard carbon dioxide equivalent values ​​according to a preset carbon emission accounting standard. A unified timestamp encoding system and a spatial grid encoding system are then attached to all data records in the raw carbon emission data lake. The minimum time unit for the timestamp encoding system is minutes, and the spatial grid encoding system adopts predefined geographical grid division rules. The converted standard carbon dioxide equivalent values ​​are then associated with the unified timestamp encoding system and spatial grid encoding system to form a standardized urban carbon emission element dataset organized by time series and spatial grid, with standardized records as units.

[0021] In practice, the data records in the raw carbon emission data lake originate from different types of monitoring equipment and business platforms. For example, one record may come from a gas meter installed in an industrial park, with its business type labeled "industrial fuel consumption" and its physical unit "cubic meters." Another record may come from traffic checkpoint monitoring, with its business type labeled "motor vehicle passage" and its physical unit "vehicle trips." Identifying the business type and physical unit of each data record in the raw carbon emission data lake is a prerequisite for performing standardization conversion. In practice, each data record carries metadata tags from its source system, and the system parses these tags to determine the data's business attributes.

[0022] According to the preset carbon emission accounting standards, the values ​​of different physical dimensions are converted into standard carbon dioxide equivalent values. In specific implementation, the system maintains an emission factor mapping library, which stores the carbon emission coefficients of various activities. For the above "industrial fuel consumption" data, the system retrieves the emission factor corresponding to "natural gas combustion" in the mapping library. If the factor is... The conversion formula for kilograms of carbon dioxide equivalent per cubic meter is: in: This represents the numerical value in cubic meters from the original data. This is the calculated standard carbon dioxide equivalent value. For "motor vehicle traffic" data, the system matches the average emission factor according to the vehicle type, multiplies the "number of trips" by the emission coefficient per trip to obtain the equivalent carbon emissions. After the above processing, all data records are converted into a unified energy or mass unit, namely the standard carbon dioxide equivalent value.

[0023] A unified timestamp encoding system and spatial grid encoding system are attached to all data records in the raw carbon emission data lake. In specific implementation, the timestamp encoding system adopts Coordinated Universal Time (UTC) format, with the smallest time unit accurate to the minute. For example, a data record generated on March 15, 2025 at 08:30 has a timestamp encoding of "202503150830". The spatial grid encoding system adopts predefined geographic grid division rules, such as dividing urban areas into square grids with a side length of 500 meters. Each grid is assigned a unique code, such as "GRID_12345". The geographic coordinates of the data record are matched to the corresponding grid code through spatial indexing.

[0024] The converted standard carbon dioxide equivalent values ​​are associated with the unified timestamp coding system and spatial grid coding system. In specific implementation, each standardized record contains three core fields: standard carbon dioxide equivalent value, timestamp code, and spatial grid code. These records are organized according to time series and spatial grid. For example, multiple records within the same timestamp and the same grid will be aggregated to form a basic unit of the standardized urban carbon emission element dataset. The standardized urban carbon emission element dataset is ultimately presented as a multidimensional data array indexed by time and space.

[0025] In one embodiment of the present invention, see [reference] Figure 3 Virtual model components corresponding to urban entity subsystems are established within the city-level digital twin foundation model. These virtual model components include an energy supply network component, a traffic flow network component, a building energy consumption component, and an industrial production component. Energy consumption data, traffic flow data, building operation data, and industrial output data from the standardized urban carbon emission element dataset are mapped and input to the energy supply network component, traffic flow network component, building energy consumption component, and industrial production component, respectively. The coupling and interaction relationships between the virtual model components in the city-level digital twin foundation model are activated, and the simulation clock of the virtual model is advanced at a rate synchronized with real time, enabling each virtual model component to perform dynamic calculations and state updates based on the input data and interaction relationships. Internal state variables output by each virtual model component are continuously collected under the advancement of the simulation clock. These internal state variables include at least energy flow, traffic velocity, building internal environmental parameters, and industrial equipment load, arranged in chronological order to form the real-time state sequence of the twin.

[0026] In practical implementation, the city-level digital twin basic model is a complex system model coupled across multiple domains. Within it, virtual model components are established, each corresponding one-to-one with a specific subsystem of the city. These virtual model components include an energy supply network component, a traffic flow network component, a building energy consumption component, and an industrial production component. The energy supply network component models the energy transmission and node balance relationships of the power grid and gas pipeline network; the traffic flow network component models the road network topology and vehicle dynamics; the building energy consumption component models the thermal characteristics of the building envelope and the HVAC system; and the industrial production component models typical technological processes and material energy balance.

[0027] The energy consumption data, traffic flow data, building operation data, and industrial output data in the standardized urban carbon emission data set are mapped and input to the corresponding virtual model components. In specific implementation, energy consumption data, such as substation outgoing power readings, are sent to the corresponding electrical nodes of the energy supply network component; traffic flow data, such as road segment detector traffic density, are sent to the corresponding road segment agent of the traffic flow network component; building operation data, such as building sub-item electricity consumption and indoor temperature, are sent to the building instance object of the building energy consumption component; and industrial output data, such as production line output and raw material input, are sent to the process unit module of the industrial production component.

[0028] The coupling and interaction relationships between the virtual model components in the city-level digital twin basic model are activated, and the simulation clock of the virtual model is advanced at a rate synchronized with real time. In some embodiments, the vehicle movement of the traffic flow network component dynamically loads the road lighting and traffic light loads, thereby changing the local power demand of the energy supply network component. The electricity price fluctuations of the energy supply network component, in turn, affect the scheduling plans of the industrial production component. This bidirectional or multidirectional effect constitutes a coupling and interaction relationship. The simulation clock advances in steps of seconds or minutes in sync with the real world. Each virtual model component performs dynamic calculations and status updates based on the current input data and interaction relationships. For example, the energy supply network component solves the node power balance equations. in: Indicates time Power injected into the node, Indicates time Affected by temperature The sum of the power of various loads affected, This represents the instantaneous change in energy storage at this node of the power grid.

[0029] The system continuously collects the internal state variables output by each virtual model component under the advancement of the simulation clock, and arranges them in chronological order to form the real-time state sequence of the twin. It can be understood that the internal state variables output by the energy supply network component include active and reactive power flows for each line; the internal state variables output by the traffic flow network component include the average vehicle speed and queue length for each road segment; the internal state variables output by the building energy consumption component include the temperature, humidity, and instantaneous power of the air conditioning system for each room; and the internal state variables output by the industrial production component include the temperature, pressure, and carbon emission intensity per unit time for each reactor. Optionally, the system records the values ​​of these variables at fixed sampling intervals, forming a time-indexed multidimensional array, which constitutes the real-time state sequence of the twin reflecting the dynamic operating state of the urban physical system.

[0030] In one embodiment of the present invention, a fixed-length historical state slice is extracted from the real-time state sequence of the twin. Heterogeneous state variables from different virtual model components in the historical state slice are concatenated to form an initial fusion tensor. This initial fusion tensor is simultaneously projected onto multiple independent feature subspaces. Within each feature subspace, the correlation weight of the state information at each time point in the historical state slice relative to the state information at all other time points is calculated. When calculating the correlation weight, a learnable co-regulation factor is introduced. This co-regulation factor is generated from the overall statistical features of the historical state slice and is used to dynamically adjust the balance of attention distribution within different feature subspaces. The correlation weights adjusted by the co-regulation factor within each feature subspace are applied to the corresponding state information, and a weighted sum is obtained to obtain the context-aware feature vector for each feature subspace. The context-aware feature vectors of all feature subspaces are concatenated and fused through a linear transformation layer, ultimately outputting a compact multi-dimensional co-regulation feature that contains multi-source state cross-temporal dependencies. During the attention weight calculation process, an adaptive bias term is added to the dot product similarity calculation results of the query vector, key vector, and value vector. The value of the adaptive bias term is jointly determined by the variance and mean of the currently processed historical state slice, and is used to suppress abnormal attention distribution caused by differences in the stationarity of the state sequence. After parallel attention calculation in multiple feature subspaces, a cross-head attention interaction layer is added. This cross-head attention interaction layer receives intermediate feature maps output from all feature subspaces and learns the potential correlations between attention focusing patterns in different feature subspaces. The cross-head attention interaction layer allows conditional interaction of attention feature vectors from different feature subspaces through a lightweight cross-gated network, thereby enhancing the ability of the finally fused multi-dimensional collaborative features to represent the nonlinear coupling relationship between different carbon emission driving factors.

[0031] In specific implementation, a fixed-time historical state slice is extracted from the real-time state sequence of the twin, with the fixed time length set to the past 24 hours, sampling at one minute, totaling 1440 time steps. Heterogeneous state variables from different virtual model components in the historical state slice are concatenated to form an initial fusion tensor. These heterogeneous state variables include line power from the energy supply network component, road segment speed from the traffic flow network component, indoor temperature from the building energy consumption component, and equipment load from the industrial production component; these variables differ in both numerical range and physical meaning. The initial fusion tensor has a dimension of 1440×K, where K is the total number of selected state variables, each row corresponds to a multi-source state vector at a given time point, and each column corresponds to a sequence of a specific state variable changing over time.

[0032] The initial fusion tensor is simultaneously projected onto multiple independent feature subspaces. In some embodiments, eight independent linear transformation layers are used to project the initial fusion tensor onto eight different 64-dimensional feature subspaces. Within each feature subspace, the correlation weight of the state information at each time point in the historical state slice relative to the state information at all other time points is calculated. This correlation weight reflects the degree to which the state at different times contributes to the feature expression at the current time. When calculating the correlation weight, a learnable co-regulation factor is introduced. This co-regulation factor is generated from the overall statistical features of the historical state slice, for example, by transforming the mean and standard deviation of the historical state slice using a small neural network. This factor is used to dynamically adjust the balance of attention distribution within different feature subspaces, preventing some subspaces from over-focusing on a few abnormal time points. The correlation weights adjusted by the co-regulation factor within each feature subspace are applied to the corresponding state information and weighted summed to obtain the context-aware feature vector for each feature subspace. This context-aware feature vector integrates information from the entire historical period and can represent long-term dependencies.

[0033] The context-aware feature vectors from all feature subspaces are concatenated and then dimensionality-reduced and fused through a linear transformation layer. The final output is a compact, multi-dimensional collaborative feature that encapsulates the spatiotemporal dependencies of multiple source states. The dimensionality of this multi-dimensional collaborative feature is much smaller than the total number of variables in the original concatenation, achieving information compression and abstraction. During the attention weight calculation, an adaptive bias term is added to the dot product similarity calculation result of the query vector, key vector, and value vector. The value of this adaptive bias term is determined by the variance and mean of the currently processed historical state slice, and the specific calculation formula is as follows: in: Indicates the adaptive bias term. This represents the total variance of all variables in the historical state slice. This represents the overall mean of all variables. and These are learnable parameters. This is a small constant added for numerical stability. This adaptive bias term is used to suppress anomalies in the attention distribution caused by differences in the stationarity of the state sequence, and automatically adjusts the weight distribution when the sequence fluctuates drastically.

[0034] After parallel attention computation across multiple feature subspaces, a cross-head attention interaction layer is added. This layer receives intermediate feature maps from all feature subspaces and learns the potential correlations between attention focusing patterns in different feature subspaces. The cross-head attention interaction layer, through a lightweight cross-gating network, allows conditional interaction between attention feature vectors from different feature subspaces, thereby enhancing the ability of the final fused multi-dimensional collaborative features to represent the nonlinear coupling relationships between different carbon emission drivers. Optionally, the cross-gating network uses the attention weights of one subspace as a gating signal to adjust the transmission strength of the feature vectors from another subspace, enabling the fused features to capture the complex linkage effects between states in multiple domains such as energy and transportation. Referring to Table 1, the table shows the state variable values ​​from different virtual model components at some time points in the historical state slices. Table 1: Numerical Table of Multi-Source State Variables in Historical State Slices In one embodiment of the present invention, the carbon emission collaborative prediction model includes a temporal feature enhancement network and a multi-step prediction head. The temporal feature enhancement network receives the multi-dimensional collaborative features and captures and refines the deep temporal patterns implicit within the multi-dimensional collaborative features through stacked dilated causal convolutional layers, outputting forward-looking enhanced temporal features. The multi-step prediction head includes multiple parallel output branches equal to the number of prediction steps. Each parallel output branch independently receives the enhanced temporal features and maps them to a specific scalar prediction value through a fully connected layer. The scalar prediction values ​​generated by all parallel output branches are arranged according to their corresponding future time points to form the collaborative prediction results of urban carbon emissions at multiple future time points. Multiple dilated causal convolutional layers with different dilation rates are set, and the dilation rates are arranged in ascending order. In the order of increasing dilation rates, the dilated causal convolutional layers at the front are used to capture fine-grained, short-period temporal patterns, and the dilated causal convolutional layers at the back are used to capture coarse-grained, long-period temporal patterns. The multi-dimensional collaborative features are used as input and sequentially passed through multiple dilated causal convolutional layers with different dilation rates. The output of each layer is not only passed to the next layer but also fused with the input of a subsequent specific layer via skip connections. After the last dilated causal convolutional layer, a global temporal pooling operation is introduced to aggregate the multi-layered refined temporal features along the time dimension, extracting global temporal dependency features as the forward-looking, deepened temporal features.

[0035] In its implementation, the carbon emission collaborative prediction model is a deep neural network architecture specifically designed for multi-step look-ahead computation, comprising a temporal feature enhancement network and a multi-step prediction head. The temporal feature enhancement network receives the multi-dimensional collaborative features, which are two-dimensional tensors with a shape of (time steps, feature dimension), for example, a 256-dimensional feature tensor extracted from sampling points per minute over the past 24 hours. The temporal feature enhancement network captures and refines the deep temporal patterns implicit within the multi-dimensional collaborative features through stacked dilated causal convolutional layers, outputting forward-looking enhanced temporal features.

[0036] Multiple dilated causal convolutional layers with different dilation rates are configured, arranged in ascending order of layer size (e.g., four layers with dilation rates of 1, 2, 4, and 8). In this ascending order, the earlier layers capture fine-grained, short-cycle temporal patterns, such as hourly energy fluctuations; the later layers capture coarse-grained, long-cycle temporal patterns, such as daily or weekday trends. The multi-dimensional collaborative features are used as input and sequentially passed through these layers. The output of each layer is not only passed to the next layer but also fused with the input of a specific subsequent layer via skip connections. For example, the output of layer 1 is not only input to layer 2 but also directly added to the input of layer 3 via skip connections to preserve features at different scales.

[0037] Following the final dilated causal convolutional layer, a global temporal pooling operation is introduced to aggregate the multi-layered refined temporal features along the time dimension, extracting global temporal dependency features as the aforementioned forward-looking, deepened temporal features. It can be understood that the global temporal pooling operation can be max pooling or average pooling, mapping the temporal features into a fixed-length vector that summarizes the global information of the entire historical window.

[0038] The multi-step prediction head contains multiple parallel output branches equal to the number of prediction steps. For example, if carbon emissions need to be predicted for 12 future time points (steps = 12), then 12 completely independent parallel output branches are set. Each parallel output branch independently receives the enhanced temporal features and maps them to a specific scalar prediction value through a fully connected layer. Each fully connected layer has independent weight parameters and is specifically learned to map the enhanced temporal features, which share input, to the target value at a specific future time point. The scalar prediction values ​​generated by all parallel output branches are arranged in order of their corresponding future time points to form the collaborative prediction result of urban carbon emissions for the multiple future time points. Refer to Table 2, which shows the dilation rate corresponding to different dilated causal convolutional layers and the range of the main temporal patterns they capture: Table 2: Configuration and Capture Mode of Dilated Causal Convolutional Layers In some embodiments, the operation of the dilated causal convolutional layer involves sampling and weighted summing of input features at specific locations. For a given layer, the output features at each time point are... From input features The calculation shows that: in: Indicates time The output feature vector, Indicates the first The weight vector at each convolutional kernel position. Indicates the input features at time 10:00 eigenvectors, This is the expansion rate of the current layer. This refers to the size of the convolution kernel. Optionally, this can be adjusted by changing the dilation rate. Without increasing the number of parameters, the receptive field of the convolution operation is significantly expanded, thereby effectively capturing the long-term periodic patterns and slowly drifting trend components in the multi-dimensional collaborative features.

[0039] In one embodiment of the present invention, the collaborative prediction results of urban carbon emissions are used as boundary constraints for future time periods and are input in reverse to the corresponding virtual model component in the city-level digital twin basic model. In the city-level digital twin basic model, based on the boundary constraints and combined with the model's built-in operating rules and interaction logic, virtual simulations are performed from the current simulation moment towards future time periods, simulating the state evolution paths of various urban subsystems under the carbon emission scenario defined by the collaborative prediction results of urban carbon emissions. The process involves recording the sequence of key intermediate variables related to carbon emissions generated during the virtual simulation. A consistency comparison analysis is then performed between this sequence and recent historical segments of the twin's real-time state sequence used to generate the collaborative prediction results for urban carbon emissions. Specifically, this includes: extracting a subsequence of state variables that physically corresponds to the sequence of key intermediate variables from the recent historical segments of the twin's real-time state sequence; calculating the degree of difference between the sequence of key intermediate variables and the corresponding subsequence of state variables on the same statistical characteristics, including mean, variance, and slope of the trend; establishing a degree of difference evolution curve, which depicts the continuous change of the degree of difference from a historical moment to a future prediction moment; and determining that the virtual simulation process is consistent with the historical evolution logic when the degree of difference evolution curve exhibits a smooth transition near the prediction starting point. Based on the results of the consistency comparison analysis, the prediction performance evaluation index for the current round of carbon emission collaborative prediction model is calculated; based on the prediction performance evaluation index of historical multiple rounds, the learning rate decay strategy of key trainable parameters in the carbon emission collaborative prediction model is dynamically adjusted; using the latest acquired standardized urban carbon emission element dataset and the corresponding real carbon emission observation values, the carbon emission collaborative prediction model is incrementally fine-tuned to adapt to the dynamic evolution of the urban system over time.

[0040] In practical implementation, the city's carbon emission collaborative prediction results are used as boundary constraints for future time periods. These constraints are then input back into the corresponding virtual model components within the city-level digital twin basic model. This means that if the prediction indicates an increase in industrial carbon emissions in a future period, this total constraint is decomposed into the production cap of industrial production components or the fossil fuel quota of the energy supply network components. Within the city-level digital twin basic model, based on these boundary constraints and combined with the model's built-in operating rules and interaction logic, a virtual simulation is performed from the current simulation moment towards future time periods. This simulates the state evolution paths of various urban subsystems under the carbon emission scenario defined by the city's carbon emission collaborative prediction results. For example, under energy constraints, the model simulates adjustments to generator scheduling strategies, optimized responses to traffic signal timing, and fluctuations in building temperature control settings, thereby deduce the system's operational trajectory to meet the predicted emission targets.

[0041] The key intermediate variable sequences related to carbon emissions generated during the virtual simulation are recorded. These sequences are then compared with recent historical segments of the twin real-time state sequence used to generate the collaborative prediction results for urban carbon emissions. From these recent historical segments, state variable subsequences corresponding in physical meaning to the key intermediate variable sequences are extracted. For example, average vehicle speed on main roads, total building power in commercial areas, and carbon intensity in industrial parks are selected as comparison objects. The degree of difference between the key intermediate variable sequences and their corresponding state variable subsequences on the same statistical characteristics is calculated. These statistical characteristics include mean, variance, and slope of the trend. A difference evolution curve is established, depicting the continuous change of the difference from a historical moment to a future prediction moment. When the difference evolution curve exhibits a smooth transition near the prediction starting point, the virtual simulation process is considered consistent with the historical evolution logic.

[0042] Based on the results of the consistency comparison analysis, the prediction performance evaluation index for the collaborative prediction model of carbon emissions in the current cycle is calculated. In some embodiments, the prediction performance evaluation index... The calculation is based on a combination of the smoothness and offset of the difference curve: in: Indicates the predicted start time. Indicates the end time of the prediction. Indicates time Normalized variance The attenuation coefficient is used to amplify the effect of near-end differences. A smaller value indicates a closer connection between the deduction path and historical logic, resulting in higher reliability of the prediction results. Based on the prediction performance evaluation indicators from multiple historical rounds, the learning rate decay strategy for key trainable parameters in the carbon emission collaborative prediction model is dynamically adjusted. For example, when multiple consecutive rounds... As the value increases, the learning rate decay rate is slowed down to enhance model training. Using the latest acquired normalized urban carbon emission element dataset and corresponding real carbon emission observations, the carbon emission collaborative prediction model is incrementally fine-tuned to adapt to the dynamic evolution of the urban system over time. Optionally, only the most recent week's data is used for each small-batch gradient update to avoid forgetting long-term patterns.

[0043] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A novel collaborative prediction method for carbon emissions in low-carbon cities based on digital twins, characterized in that, include: Construct a data access channel for multi-source carbon emission data in the city, receive and aggregate dynamic streaming data and static benchmark data from IoT sensors, urban business systems and statistical platforms to form a raw carbon emission data lake; The data in the raw carbon emission data lake are standardized, cleaned, and spatiotemporally aligned to generate a standardized urban carbon emission element dataset with unified spatiotemporal identification. Based on the standardized urban carbon emission element dataset, the preset urban-level digital twin basic model is driven to perform simulation evolution, generating a real-time state sequence of the twin that reflects the dynamic operating status of the urban entity system. Based on the real-time state sequence of the twin, an improved attention fusion algorithm is used to extract multi-dimensional collaborative features affecting carbon emissions. The improved attention fusion algorithm is based on a multi-head attention mechanism and introduces a learnable collaborative adjustment factor and an adaptive bias term in its attention weight calculation process, as well as adding a cross-head attention interaction layer. Based on the real-time state sequence of the twin, an improved attention fusion algorithm is used to extract multi-dimensional collaborative features affecting carbon emissions, including: A fixed-time-length slice of historical state is extracted from the real-time state sequence of the twin, and the heterogeneous state variables from different virtual model components in the historical state slice are spliced ​​together to form an initial fusion tensor. The initial fusion tensor is simultaneously projected onto multiple independent feature subspaces. Within each feature subspace, the correlation weight of the state information at each time point in the historical state slice relative to the state information at all other time points is calculated. When calculating the relevance weights, an additional learnable co-regulation factor is introduced. This co-regulation factor is generated from the overall statistical features of the historical state slices and is used to dynamically adjust the balance of attention distribution in different feature subspaces. Within each feature subspace, the correlation weights adjusted by the collaborative adjustment factor are applied to the corresponding state information, and a weighted sum is performed to obtain the context-aware feature vector of each feature subspace. The context-aware feature vectors of all feature subspaces are concatenated and then dimensionality-reduced and fused through a linear transformation layer, ultimately outputting a compact multi-dimensional collaborative feature that contains the spatiotemporal dependencies of multi-source states. The working principle of the improved attention fusion algorithm includes: During the attention weight calculation process, an adaptive bias term is added to the dot product similarity calculation result of the query vector, key vector, and value vector. The value of the adaptive bias term is determined by the variance and mean of the currently processed historical state slice, and is used to suppress the abnormal attention distribution caused by the difference in the stationarity of the state sequence. After performing parallel attention computation in multiple feature subspaces, a cross-head attention interaction layer is added. The cross-head attention interaction layer receives the intermediate feature maps output by all feature subspaces and learns the potential correlation between the attention focusing patterns of different feature subspaces. The cross-head attention interaction layer allows attention feature vectors from different feature subspaces to interact conditionally through a lightweight cross-gated network, thereby enhancing the ability of the finally fused multi-dimensional collaborative features to characterize the nonlinear coupling relationship between different carbon emission drivers. The multi-dimensional collaborative features are input into the carbon emission collaborative prediction model for multi-step look-ahead calculations, outputting the collaborative prediction results of urban carbon emissions at multiple future time points, including: The carbon emission collaborative prediction model includes a temporal feature enhancement network and a multi-step prediction head; The temporal feature deepening network receives the multi-dimensional collaborative features and captures and extracts the deep temporal patterns hidden within the multi-dimensional collaborative features through stacked dilated causal convolutional layers, outputting forward-looking deepened temporal features. The multi-step prediction head contains multiple parallel output branches equal to the number of prediction steps. Each parallel output branch independently receives the deepened temporal features and is mapped to a specific scalar prediction value through a fully connected layer. Arrange the scalar prediction values ​​generated by all parallel output branches in the order of their corresponding future time points to form the collaborative prediction results of urban carbon emissions at multiple future time points. The collaborative prediction results of urban carbon emissions are fed back to the city-level digital twin basic model, driving the city-level digital twin basic model to perform virtual simulations under the prediction scenario, forming a prediction verification closed loop.

2. The novel low-carbon city carbon emission collaborative prediction method based on digital twins according to claim 1, characterized in that, The data in the raw carbon emission data lake are standardized, cleaned, and spatiotemporally aligned to generate a standardized urban carbon emission element dataset with unified spatiotemporal identification, including: Identify the business type and physical dimension of each data record in the raw carbon emission data lake, and convert the values ​​of different physical dimensions into standard carbon dioxide equivalent values ​​according to the preset carbon emission accounting standard. A unified timestamp coding system and a spatial grid coding system are attached to all data records in the raw carbon emission data lake. The minimum time unit of the timestamp coding system is minutes, and the spatial grid coding system adopts a predefined geographical grid division rule. The converted standard carbon dioxide equivalent values ​​are associated with the unified timestamp coding system and spatial grid coding system to form the standardized urban carbon emission element dataset, which is organized by time series and spatial grid with standardized records as units.

3. The novel collaborative prediction method for low-carbon city carbon emissions based on digital twins according to claim 1, characterized in that, Based on the standardized urban carbon emission element dataset, a pre-defined city-level digital twin model is driven to perform simulation evolution, generating a real-time state sequence of the twin reflecting the dynamic operating state of the urban entity system, including: In the city-level digital twin basic model, virtual model components corresponding to the city's physical subsystems are established. The virtual model components include energy supply network components, traffic flow network components, building energy consumption components, and industrial production components. The energy consumption data, traffic flow data, building operation data, and industrial output data in the standardized urban carbon emission data set are mapped and input to the energy supply network component, traffic flow network component, building energy consumption component, and industrial production component, respectively. Activate the coupling and interaction relationships between the virtual model components in the city-level digital twin basic model, advance the simulation clock of the virtual model at the rate of real time synchronization, so that each virtual model component can perform dynamic calculation and state update based on input data and interaction relationships; The internal state variables output by each virtual model component are continuously collected as the simulation clock advances. These internal state variables include at least energy flow, traffic flow velocity, building internal environmental parameters, and industrial equipment load, and are arranged in chronological order to form the real-time state sequence of the twin.

4. The novel collaborative prediction method for low-carbon city carbon emissions based on digital twins according to claim 1, characterized in that, The temporal feature deepening network receives the multi-dimensional collaborative features and captures and extracts the deep temporal patterns implicit within the multi-dimensional collaborative features through stacked dilated causal convolutional layers, including: Multiple dilated causal convolutional layers with different dilation rates are configured, and the dilation rates are arranged in ascending order. In the order of increasing dilation rates, the dilated causal convolutional layers at the front are used to capture fine-grained, short-period temporal patterns, while the dilated causal convolutional layers at the back are used to capture coarse-grained, long-period temporal patterns. The multi-dimensional collaborative features are used as input and passed sequentially through multiple dilated causal convolutional layers with different dilation rates. In addition to being passed to the next layer, the output of each layer is also fused with the input of a subsequent specific layer through skip connections. After the last dilated causal convolutional layer, a global temporal pooling operation is introduced to aggregate the temporal features refined through multiple layers in the time dimension, extracting global temporal dependency features as the forward-looking, in-depth temporal features.

5. The novel collaborative prediction method for low-carbon city carbon emissions based on digital twins according to claim 1, characterized in that, The collaborative prediction results of urban carbon emissions are fed back to the city-level digital twin basic model, driving the city-level digital twin basic model to perform virtual simulations under the prediction scenario, forming a prediction verification closed loop, including: The city carbon emission collaborative prediction results are used as boundary constraints for future time periods and are input in reverse to the corresponding virtual model components in the city-level digital twin basic model. In the city-level digital twin basic model, based on the boundary constraints and combined with the model's built-in operating rules and interaction logic, virtual simulations are performed from the current simulation moment to future time periods to simulate the state evolution paths of each subsystem of the city under the carbon emission scenario defined by the city's collaborative carbon emission prediction results. Record the sequence of key intermediate variables related to carbon emissions generated during the virtual simulation process, and perform a consistency comparison analysis between the sequence of key intermediate variables and recent historical fragments of the twin real-time state sequence on which the collaborative prediction results of urban carbon emissions were based. Specifically, this includes: Extract the state variable subsequence that corresponds in physical meaning to the key intermediate variable sequence from the recent historical fragments of the twin's real-time state sequence; Calculate the degree of difference between the key intermediate variable sequence and the corresponding state variable subsequence on the same statistical features, including mean, variance and slope of change trend; A difference degree evolution curve is established, which depicts the continuous change of the difference degree from a historical moment to a future prediction moment; When the difference evolution curve exhibits a smooth transition near the prediction starting point, it is determined that the virtual inference process is consistent with the historical evolution logic.

6. The novel collaborative prediction method for low-carbon city carbon emissions based on digital twins according to claim 1, characterized in that, After forming the prediction verification closed loop, the method further includes: Based on the results of the consistency comparison analysis, the prediction performance evaluation index of the collaborative prediction model for carbon emissions in the current round is calculated. Based on the historical multi-round prediction performance evaluation indicators, the learning rate decay strategy of key trainable parameters in the carbon emission collaborative prediction model is dynamically adjusted. Using the newly acquired standardized urban carbon emission element dataset and corresponding real carbon emission observations, the carbon emission collaborative prediction model is incrementally fine-tuned to adapt to the dynamic evolution of the urban system over time.

7. A novel low-carbon city carbon emission collaborative prediction system based on digital twins, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the novel low-carbon city carbon emission collaborative prediction method based on digital twins as described in any one of claims 1 to 6.

Images