A business source-load carbon collaborative prediction method, system, device and storage medium

Abstract

The present application relates to the technical field of source-load-carbon collaborative prediction, and discloses an enterprise source-load-carbon collaborative prediction method, system, device and storage medium, which comprises the following steps: obtaining original operation data and meteorological data of an enterprise to be predicted, inputting the meteorological data into a dynamic meteorological kernel network to obtain dynamic spatial meteorological features; inputting the original operation data and the dynamic spatial meteorological features into a comparative variational mode decoupling network to obtain inertia trend features and meteorological fluctuation features; using a bidirectional cross attention and a gate mechanism to perform feature enhancement and feature fusion on the inertia trend features, the meteorological fluctuation features and the dynamic spatial meteorological features to obtain fused features; and inputting the fused features into a collaborative prediction model to obtain source-load-carbon prediction values of the enterprise to be predicted. Through deep perception of meteorological data, a feature decoupling mechanism based on physical guidance and a prediction architecture based on deep coupling of data driving and physical mechanism, the present application effectively improves the accuracy and stability of enterprise source-load-carbon collaborative prediction results.

Description

Technical Field

[0001] This invention relates to the field of source-load carbon co-prediction technology, and in particular to a method, system, device and storage medium for enterprise source-load carbon co-prediction. Background Technology

[0002] Modern corporate energy systems are becoming increasingly complex, typically integrating grid power purchases, distributed renewable energy generation, and diverse power loads. There is a profound intrinsic coupling relationship between photovoltaic (PV) output, power load, and carbon emission intensity. For example, PV output directly affects a company's net load and power purchase strategy, while the choice of power purchase time indirectly determines its carbon emission level.

[0003] Therefore, accurate and coordinated forecasting of enterprises' photovoltaic output, electricity load, and carbon emission intensity is crucial for optimizing energy costs, improving the absorption rate of new energy sources, and formulating scientific low-carbon strategies. However, existing forecasting methods face several challenges: First, traditional forecasting methods often employ a "divide and conquer" strategy, establishing independent forecasting models for load, photovoltaics, and carbon intensity. This approach completely ignores the inherent physical connections and temporal coupling between them, resulting in a lack of overall consistency in forecasting results and making it difficult to support coordinated optimization decisions. Second, load, photovoltaic, and carbon intensity data are typically highly nonlinear and non-stationary time series, influenced by various external factors such as weather and electricity prices. Traditional statistical models struggle to capture their complex dynamic patterns. Finally, although some deep learning methods have been applied to time series forecasting, they are relatively simplistic in feature engineering, typically processing raw time series data directly without fully exploring the inherent information between the data. When multiple features coexist, existing fusion mechanisms often involve simple feature splicing, lacking deep interaction and failing to effectively extract highly discriminative fusion feature representations for accurate forecasting, thus limiting the upper limit of the forecasting model's accuracy. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method, system, device, and storage medium for enterprise source-load carbon co-prediction, which can solve the problems of insufficient feature extraction, poor multi-source information fusion effect, and insufficient prediction accuracy in the processing of highly coupled source-load carbon data in existing technologies, thereby improving the accuracy and stability of source-load carbon co-prediction results.

[0005] In a first aspect, the present invention provides a method for predicting the synergistic carbon source and load of enterprises, the method comprising: The original operational data and meteorological data of the enterprise to be predicted are obtained, and the meteorological data is input into a preset dynamic meteorological kernel network to obtain dynamic spatial meteorological characteristics. The dynamic meteorological kernel network is constructed based on a convolutional neural network. The original operational data and the dynamic spatial meteorological features are input into a preset comparative variational mode decoupling network to obtain inertial trend features and meteorological fluctuation features. The comparative variational mode decoupling network is constructed based on a variational filter network with physically guided comparison anchor points. A bidirectional cross-attention and gating mechanism is used to enhance and fuse the inertial trend features, the meteorological fluctuation features, and the dynamic spatial meteorological features to obtain fused features; The fused features are input into a preset collaborative prediction model to obtain the source load carbon prediction value of the enterprise to be predicted. The collaborative prediction model adopts a multi-task prediction model architecture and is built based on a physical embedded Transformer encoder. The difference between the predicted photovoltaic output value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model is calculated in the source load carbon prediction value. The difference is input into the performance prediction model constructed based on the multilayer perceptron network to obtain the dynamic performance ratio. Based on the dynamic performance ratio and the upper limit of photovoltaic output, the optimized photovoltaic output prediction value is obtained.

[0006] Furthermore, the steps of obtaining the raw operational data and meteorological data of the enterprise to be predicted include: Obtain the raw operating data of the enterprise to be predicted in the current period. The raw operating data includes real-time source load carbon data, local meteorological data, time-derived data, physical model data, and lagged source load carbon data. The meteorological data of the enterprise to be predicted is obtained through the meteorological data service interface. The meteorological data includes dynamic meteorological data and environmental characteristic meteorological data.

[0007] Furthermore, the dynamic meteorological kernel network consists of a series of dynamic convolutional modules and a feature dimensionality reduction module; The dynamic convolution module consists of multiple cascaded dynamic convolutional layers, used to perform dynamic resampling and convolution operations on the meteorological data. The offset used by each dynamic convolutional layer during resampling is calculated by one convolutional layer from the wind speed vector in the meteorological data. The feature dimensionality reduction module consists of a global average pooling layer and a first multilayer perceptron. The global average pooling layer is used to aggregate information from the high-dimensional spatial features output by the dynamic convolution module and output the pooled features. The first multilayer perceptron is used to perform dimensionality mapping on the pooled features and output dynamic spatial meteorological features.

[0008] Furthermore, the contrastive variational mode decoupling network consists of a series of physical anchor modules and a variational filter network; The physical anchor point module consists of two parallel fully connected neural networks. The input data of the first fully connected neural network are time-derived data and physical model data in the original operating data, and the output data is behavioral pattern anchor points. The input data of the second fully connected neural network are local meteorological data and dynamic spatial meteorological features in the original operating data, and the output data is global meteorological anchor points. The variational filter network takes the real-time source load carbon data and the lagged source load carbon data in the original operating data as input data, and uses two parallel frequency domain filter channels and inverse fast Fourier transform to perform frequency domain filtering and time domain reconstruction on the input data, and outputs inertial trend characteristics and meteorological fluctuation characteristics. The loss function of the variational filter network consists of reconstruction loss, physical alignment loss and orthogonality independence loss, and the physical alignment loss consists of the similarity between the inertial trend feature and the behavioral pattern anchor point, as well as the similarity between the meteorological fluctuation feature and the global meteorological anchor point.

[0009] Furthermore, the step of using a bidirectional cross-attention and gating mechanism to enhance and fuse the inertial trend features, the meteorological fluctuation features, and the dynamic spatial meteorological features to obtain fused features includes: The inertial trend feature and the meteorological fluctuation feature are enhanced by using bidirectional cross-attention to obtain enhanced inertial trend feature and enhanced meteorological fluctuation feature; The dynamic spatial meteorological features are concatenated with the time-derived data in the original operational data to obtain a scene vector; Based on the scene vector, a gating mechanism is used to generate a dynamic weight vector. Based on the dynamic weight vector, the enhanced inertial trend feature, the enhanced meteorological fluctuation feature, and the dynamic spatial meteorological feature are fused to obtain the fused feature.

[0010] Furthermore, the collaborative prediction model consists of a multi-layered stacked physically embedded Transformer encoder, an adaptive attention pooling layer, and a shared private decoupled prediction head; The multi-layered stacked physical embedded Transformer encoder is used to capture the long-range spatiotemporal dependencies between the fused features through a grouping mechanism of physical constraint head and global attention head, and output a long-range dependency feature sequence; wherein, the physical constraint head applies constraints through a physical constraint mask matrix, which is constructed based on the physical state similarity of the input original running data at different time steps; The adaptive attention pooling layer is used to perform adaptive attention pooling on the long-range dependent feature sequence and output an aggregate vector; The shared private decoupled prediction head is used to extract shared features from the aggregated vector, concatenate the extracted source-load carbon co-feature vector with the aggregated vector, and perform independent regression prediction on the concatenated features to obtain the source-load carbon prediction value. The source-load carbon prediction value includes the photovoltaic output prediction value, the power load prediction value, and the carbon emission intensity prediction value.

[0011] Further, the step of calculating the difference between the photovoltaic output prediction value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model, inputting the difference into the performance prediction model constructed based on a multilayer perceptron network to obtain the dynamic performance ratio, and obtaining the optimized photovoltaic output prediction value based on the dynamic performance ratio and the upper limit of photovoltaic output includes: Based on the physical parameters of the time to be predicted and the preset ideal photovoltaic output model, the upper limit of photovoltaic output is calculated. The ideal photovoltaic output model is constructed based on astronomical geometric relationships and clear-sky radiation model. The photovoltaic output prediction value is extracted from the source load carbon prediction value, and the difference between the photovoltaic output prediction value and the photovoltaic output upper limit value is calculated. The difference and the environmental parameters at the time to be predicted are input into a preset performance prediction model to obtain the dynamic performance ratio. The performance prediction model is constructed based on a multilayer perceptron network. Based on the dynamic performance ratio and the upper limit of photovoltaic output, an optimized photovoltaic output prediction value is obtained.

[0012] Secondly, the present invention provides an enterprise source-load carbon co-prediction system, the system comprising: The feature extraction module is used to acquire the original operational data and meteorological data of the enterprise to be predicted, and input the meteorological data into a preset dynamic meteorological kernel network to obtain dynamic spatial meteorological features. The dynamic meteorological kernel network is constructed based on a convolutional neural network. The original operational data and the dynamic spatial meteorological features are input into a preset comparative variational mode decoupling network to obtain inertial trend features and meteorological fluctuation features. The comparative variational mode decoupling network is constructed based on a variational filter network with physically guided comparison anchor points. The feature fusion module is used to enhance and fuse the inertial trend features, the meteorological fluctuation features, and the dynamic spatial meteorological features using a bidirectional cross-attention and gating mechanism to obtain fused features; The collaborative prediction module is used to input the fused features into a preset collaborative prediction model to obtain the source load carbon prediction value of the enterprise to be predicted. The collaborative prediction model adopts a multi-task prediction model architecture and is built based on a physically embedded Transformer encoder. The difference between the predicted photovoltaic output value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model is calculated in the source load carbon prediction value. The difference is input into the performance prediction model constructed based on the multilayer perceptron network to obtain the dynamic performance ratio. Based on the dynamic performance ratio and the upper limit of photovoltaic output, the optimized photovoltaic output prediction value is obtained.

[0013] Thirdly, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0014] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0015] This invention provides a method, system, device, and storage medium for enterprise source-load carbon co-prediction. The invention utilizes a dynamic meteorological kernel network for deep perception of meteorological data, effectively capturing directional and fluid spatial dependencies. This provides dynamic spatial context information containing physical dynamics for subsequent time-series forecasts, thereby improving the accuracy of prediction results. Through a physics-guided feature decoupling mechanism combined with a bidirectional interaction mechanism, the invention enhances the model's prediction accuracy and stability when facing sudden weather disturbances, achieving physical-level deep fusion of source-load carbon data. Furthermore, through a prediction architecture deeply coupled with data-driven and physical mechanisms, structured physical knowledge is deeply and interpretably embedded into attention computation, guiding the model to learn statistical patterns while adhering to physical laws. This enhances the model's generalization ability under extreme conditions or in data-sparse scenarios, improving the interpretability and robustness of deep learning models. In summary, this invention, through deep perception of meteorological data, a physics-guided feature decoupling mechanism, and a prediction architecture deeply coupled with data-driven and physical mechanisms, effectively improves the accuracy and stability of enterprise source-load carbon co-prediction results. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the enterprise source-load carbon co-prediction method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the enterprise source-load carbon co-prediction system in an embodiment of the present invention; Figure 3 This is an internal structural diagram of the computer device in an embodiment of the present invention.

[0017] Figure label: 10. Feature extraction module; 20. Feature fusion module; 30. Collaborative prediction module. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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] Please see Figure 1 The first embodiment of the present invention proposes a method for predicting the synergistic carbon source and load of enterprises, including steps S10 to S50: Step S10: Obtain the original operating data and meteorological data of the enterprise to be predicted, input the meteorological data into a preset dynamic meteorological kernel network to obtain dynamic spatial meteorological features. The dynamic meteorological kernel network is constructed based on a convolutional neural network. Step S20: Input the original operating data and the dynamic spatial meteorological characteristics into a preset comparative variational mode decoupling network to obtain inertial trend characteristics and meteorological fluctuation characteristics. The comparative variational mode decoupling network is constructed based on a variational filter network with physically guided comparison anchor points. Step S30: Use a bidirectional cross-attention and gating mechanism to enhance and fuse the inertial trend features, the meteorological fluctuation features, and the dynamic spatial meteorological features to obtain fused features; Step S40: Input the fused features into a preset collaborative prediction model to obtain the source load carbon prediction value of the enterprise to be predicted. The collaborative prediction model adopts a multi-task prediction model architecture and is constructed based on a Transformer encoder. Step S50: Calculate the difference between the photovoltaic power output prediction value and the upper limit of photovoltaic power output obtained based on the ideal photovoltaic power output model in the source load carbon prediction value. Input the difference into the performance prediction model constructed based on the multilayer perceptron network to obtain the dynamic performance ratio. Based on the dynamic performance ratio and the upper limit of photovoltaic power output, obtain the optimized photovoltaic power output prediction value.

[0020] In this embodiment, sensors are first used to acquire site-level raw runtime sequence data of photovoltaic output, power load, carbon emission intensity and related characteristics generated during the operation of the enterprise. In addition, through the meteorological data service interface, regional heterogeneous meteorological data of the surrounding area of ​​the enterprise's geographical location within the same time period, organized in a grid form, are acquired. Specifically, the raw runtime sequence data includes real-time source-load carbon data, local meteorological data, time-derived data, physical model data, and lagged source-load carbon data. Real-time source-load carbon data includes the current power load, photovoltaic output, and carbon emission intensity; local meteorological data includes the current solar irradiance, temperature, and humidity, as well as predicted values ​​for the next moment; time-derived data is obtained by decomposing time information into hours, days of the week, and holidays; physical model data consists of solar altitude angle, solar azimuth angle, and solar zenith angle calculated based on geographic coordinates, providing strong physical constraints for photovoltaic output prediction; lagged source-load carbon data includes the previous moment's power load, photovoltaic output, and carbon emission intensity, used to capture the system's physical inertia; additionally, enterprise cooling load data and latitude / longitude coordinates can also be collected as raw runtime data.

[0021] To obtain the time-series data of the aforementioned raw operational data, preferably, time-series data of 96 steps (i.e., 96 consecutive hours) of raw operational data are collected at 1-hour sampling intervals as input matrix A; simultaneously, regional multi-level meteorological data of the corresponding step length is obtained through the meteorological data service interface as spatial input local B. This matrix contains gridded data covering the geographical area surrounding the enterprise, aiming to provide complete spatiotemporal environmental field information, specifically including: (1) Dynamic data: Zonal wind , direction of wind And air pressure, used to characterize the flow characteristics of the atmospheric flow field; (2) Environmental characteristic data: surface downflow solar radiation, total cloud cover, air temperature and dew point temperature, used to directly characterize the physical environment that affects source load changes.

[0022] It should be noted that the models in this embodiment and subsequent embodiments are all pre-trained models, and they all adopt conventional training methods. The training steps of the models will not be described in detail here.

[0023] For the two types of data acquired—raw operational data and meteorological data—this embodiment employs different models for feature extraction. Specifically, a dynamic meteorological kernel network is used to extract the dynamic spatial features of the meteorological data, aiming to explicitly embed meteorological dynamics into the feature extraction process. The implementation includes the following three steps: (1) Construction of meteorological tensor: The spatial input matrix B, composed of meteorological data, is organized into a four-dimensional tensor. Where H and W are the spatial resolutions of the grid. The channel number is T, and the total time is T. Each channel consists of two sets of physical quantities: the dynamic driving channel, which includes dynamic meteorological data, reflecting the real-time motion state of the atmospheric flow field and serving as the physical basis for calculating spatial offset; and the environmental characteristic channel, which includes environmental characteristic data, used to directly characterize the physical environment affecting photovoltaic output and power load.

[0024] (2) Physically guided nuclear shift prediction Unlike traditional deformable convolutions that implicitly learn offsets from image texture, this embodiment employs a physics-driven offset calculation mechanism, using tensors... Grid-level wind speed vector field with spatial resolution As input, the structure employs a single lightweight layer. The convolutional layer acts as a physical mapper. This mapper leverages the local perception property of convolution to map each spatial location within the meteorological grid. Calculate the corresponding dynamic offset independently. The calculated spatial offset of the convolution kernel sampling points. Let be a tensor corresponding to the input grid size, representing the non-uniformity of airflow direction at different geographical locations. The formula is: , The input is the wind speed vector field tensor; for Convolutional layers are used to map wind speed physical quantities to an offset dimension in the feature space; It is a learnable physical scaling factor used to amplify the normalized displacement features to the actual grid pixel coordinate system. Its physical meaning represents the movement amplitude of the airflow within a unit time step (i.e., the mapping relationship between wind speed and cloud movement distance). The spatial offset tensor of the calculated convolution kernel sampling points represents the transmission path of upstream information by the airflow.

[0025] (3) Dynamic resampling and convolution Using the calculated offset For tensors All environmental feature channels and dynamic channels undergo synchronous spatial transformation and feature extraction. Specifically, the dynamic meteorological kernel network consists of five cascaded stacked dynamic convolutional layers. A single convolutional layer can only capture the meteorological influence of neighboring grids. By stacking multiple dynamic convolutional layers, the network's "effective physical receptive field" can be effectively expanded, enabling it to retrospectively track high-speed cloud clusters or cold front systems moving with the wind in the distance, thus achieving "upstream perception" of future weather changes.

[0026] In the m-th convolutional layer, the above offset is... Applied to feature maps, bilinear interpolation is used at irregular locations. Sample feature values ​​and perform convolution operations: In the formula, m represents the network layer index. Since it is a 5-layer stack, the value range is [ ]; Indicates the first Layer in position Output feature value at; This represents the total number of sampling points of the convolution kernel (if Convolution kernel, then ); This represents the weight parameter of the k-th sampling point in the m-th convolutional kernel; For the first Feature map output by layer (when) The time is the original input tensor ); This represents the k-th sampling point in the standard convolution kernel relative to the center point. Fixed grid offset; This indicates the corresponding number calculated by step (1). Dynamic physical offset of each sampling point.

[0027] After five layers of stacking, a feature map containing deep dynamic information is obtained. To facilitate feature fusion in subsequent steps, the features need to be converted into one-dimensional vectors. This embodiment employs global average pooling and a multilayer perceptron to achieve dimensionality reduction and vectorization of spatial features, expressed by the following formula: In the formula, As dynamic spatial meteorological characteristics, The high-dimensional spatial features output by the dynamic convolution module. This represents the global average pooling operation, used to reduce the spatial dimension. Information aggregation, MLP is a multilayer perceptron (containing linear layers and ReLU activation), used to perform dimensional mapping on the pooled features to obtain a one-dimensional feature vector.

[0028] The dynamic meteorological kernel network proposed in this embodiment integrates meteorological physical dynamics into the feature extraction structure. By utilizing real-time physical features such as wind speed and direction, the sampling shape and position of the convolution kernel are dynamically and asymmetrically adjusted, enabling it to adaptively focus on physically key upstream influence areas. Compared to traditional fixed convolution methods that treat meteorological data as static images, this embodiment can effectively capture spatial dependencies with directionality and fluidity (such as the evolution path of weather systems), providing dynamic spatial context information containing physical dynamics for subsequent time-series forecasting, thereby improving the accuracy of the forecast results.

[0029] For the original running data, time-frequency transformation can be used to extract time-domain and frequency-domain features. Preferably, the input matrix A is used as the time-domain feature, and a two-dimensional time-frequency feature matrix is ​​generated by continuous wavelet transform (CWT). At the same time, fast Fourier transform (FFT) is applied in parallel to generate a one-dimensional frequency-domain feature vector, thereby obtaining time-frequency and frequency-domain features.

[0030] In a preferred embodiment, to overcome the limitations of traditional time-frequency processing methods such as wavelet transform and Fourier transform, which only perform mathematical segmentation based on frequency, this invention employs a contrastive variational mode decoupling network constructed based on a variational filtering network with physically guided contrast anchors to extract features from the original running data. The contrastive variational mode decoupling network consists of a series-connected physical anchor module and a variational filtering network; the physical anchor module is composed of two parallel fully connected neural networks. Specifically, to guide the subsequent decoupling network to correctly distinguish between "system inertia" and "meteorological disturbance," this embodiment utilizes auxiliary features in the input matrix A and the extracted dynamic spatial meteorological features to construct two types of physical anchor vectors: (1) Constructing behavioral pattern anchors Time-derived data and physical model data are extracted from input matrix A and fed into a two-layer fully connected neural network to generate behavioral pattern anchors. These anchors represent deterministic inertial benchmarks in the system determined by human production and life cycles (such as weekday / weekend load patterns) and astronomical laws (such as sunrise and sunset).

[0031] (2) Constructing a global meteorological anchor point Local meteorological data is extracted from the input matrix A and concatenated with dynamic spatial meteorological features to fuse local observations and wide-area field information. The fused vector is then input into another independent fully connected neural network to generate a global meteorological anchor point. This anchor point represents the random environmental disturbance in the system caused by the combined effects of local micro-meteorological abrupt changes and the evolution of surrounding wide-area weather systems.

[0032] Then, physical-guided variational mode decoupling is performed. In this embodiment, the core target sequence to be predicted (real-time and lagged source-load carbon data) in the input matrix A is used as the input sequence. It is then adaptively decomposed using a parameterized variational filtering network. The specific steps include: (1) Learnable variational filtering First, the input sequence Transform to the frequency domain to obtain the frequency domain sequence Define two parallel frequency domain filter channels: trend channel. and fluctuation channel To simulate the narrowband characteristics of variational mode decomposition and maintain differentiability, the filter's transfer function adopts a learnable Lorentz function form: In the formula, Let be the frequency domain filter for the i-th channel, where trend represents the trend channel and flux represents the fluctuation channel; Let i be the center frequency parameter of the i-th channel. For frequency variables, These are the bandwidth control parameters for the i-th channel. These parameters are used as network weights and are automatically updated during training. During initialization, the center frequency parameter of the trend channel is set. To capture a low-frequency reference, the center frequency parameter of the fluctuation channel is set. It is a high-frequency band used to capture dynamic changes.

[0033] (2) Frequency domain filtering and time domain reconstruction Based on the above parallel channel filters, the frequency domain response of each channel is calculated and inverse fast Fourier transform (IFFT) is performed to restore it to the time domain signal: In the formula, This represents the trend component of the time-domain signal at time t; This represents the fluctuation component of the time-domain signal at time t; IFFT(*) represents the inverse fast Fourier transform operation. This represents the input sequence in the frequency domain, i.e., the frequency domain sequence. For frequency variables; Frequency domain filter representing the trend channel; This represents a frequency domain filter for the fluctuation channel.

[0034] (3) Physically guided contrast loss constraint To ensure that the decomposed components have real physical meaning, rather than just mathematical high- and low-frequency segmentation, this embodiment introduces a composite physical information loss function. Participating in model training: in, and The hyperparameters are used to balance the weights of the various loss terms. The specific sub-terms are defined as follows: (1) Reconstruction loss : Ensure that the summation of the decomposed components can losslessly restore the original signal: (2) Physical alignment loss (Using cosine similarity CosSim): Forced "trend component" "and behavioral pattern anchors" Similar to, forced "fluctuation component" "and "global meteorological anchor points" "resemblance: (3) Orthogonality independence loss : Force the two components to remain orthogonal in the feature space to avoid information redundancy: Will As a characteristic of inertial trend, As meteorological fluctuation characteristics, after the above processing steps, the comparative variational mode decoupling network outputs two highly decoupled feature streams with clear physical meanings. These two feature streams will replace the traditional time-frequency features and, together with the dynamic spatial meteorological features output by the dynamic meteorological kernel network, serve as the input to the collaborative prediction model.

[0035] Since inertial trend features, meteorological fluctuation features, and dynamic spatial meteorological features are multi-source heterogeneous features, feature fusion is required before inputting them into the collaborative prediction model. Furthermore, to address the "information silos" and "scene mismatch" problems that exist during the fusion process of multi-source heterogeneous features, in a preferred embodiment, this invention provides a feature fusion method based on bidirectional cross-attention and gating mechanisms to perform nonlinear mapping and deep fusion of the above features. Specific steps include: The inertial trend feature and the meteorological fluctuation feature are enhanced by using bidirectional cross-attention to obtain enhanced inertial trend feature and enhanced meteorological fluctuation feature; The dynamic spatial meteorological features are concatenated with the time-derived data in the original operational data to obtain a scene vector; Based on the scene vector, a gating mechanism is used to generate a dynamic weight vector. Based on the dynamic weight vector, the enhanced inertial trend feature, the enhanced meteorological fluctuation feature, and the dynamic spatial meteorological feature are fused to obtain the fused feature.

[0036] In this embodiment, the inertial trend characteristics are first analyzed. Characteristics of meteorological fluctuations and dynamic spatial meteorological characteristics Perform feature space alignment, because , It comes from a variational filter network, and Since features from convolutional networks have inconsistent dimensions, one-dimensional convolutional layers (Conv1D) or fully connected layers (Linear) are used to project them into a unified high-dimensional hidden space. In the formula, H trend H represents the inertial tendency characteristic of alignment. fluct H represents the aligned meteorological fluctuation characteristics. space This indicates the dynamic spatial meteorological characteristics of alignment.

[0037] To enable the "trend" to perceive potential "fluctuations" and for the "fluctuations" to reference existing "trends," this embodiment employs a bidirectional cross-attention mechanism. This mechanism facilitates information exchange and complementary correction between inertial trend features and meteorological fluctuation features before fusion. Specifically, using aligned inertial trend features as the query and aligned meteorological fluctuation features as the key and value, the inertial trend features are enhanced based on the attention mechanism to obtain enhanced inertial trend features. This pathway utilizes the current system inertia to retrieve and utilize relevant meteorological disturbance information. Simultaneously, using aligned meteorological fluctuation features as the query and aligned inertial trend features as the key and value, the meteorological fluctuation features are enhanced based on the attention mechanism to obtain enhanced meteorological fluctuation features. This pathway serves to retrospectively reference the system's basic operating baseline when facing drastic environmental changes, preventing fluctuation features from deviating from physical principles. Through this interaction, enhanced inertial trend features are generated. and enhanced meteorological fluctuation characteristics .

[0038] To address the dynamic changes in the importance of various characteristic flows under different environmental conditions (e.g., time-frequency characteristics are more important under volatile weather conditions, while time-domain trends are more important under stable weather conditions), this embodiment designs a physical scene perception gating unit, the specific steps of which include: (1) Scene vector construction: The dynamic spatial meteorological features are concatenated with the time-derived data in the input matrix A to form a scene vector that can characterize the current physical environment state. .

[0039] (2) Generation of gating weights: Scene vectors The input is fed into a lightweight multilayer perceptron (MLP) containing two linear layers and ReLU activation, and the output is a dynamic weight vector tailored to inertial trend features and meteorological fluctuation features. : In the formula, This represents the weights for inertial trend characteristics and meteorological fluctuation characteristics. This weight vector is adjusted in real time according to changes in the physical environment within the time window.

[0040] (3) Spatiotemporal adaptive injection and fusion: The generated dynamic weight vector is used to perform a weighted summation of the two interactively enhanced features, and dynamic spatial meteorological features are explicitly injected (extended to all time steps through a broadcast mechanism to make them consistent with the temporal feature dimension), generating the final fused features: In the formula, This indicates an enhanced tendency towards inertia. This indicates enhanced meteorological fluctuation characteristics. Represents the dynamic spatial meteorological characteristics of alignment. It is a learnable scaling factor used to adjust the weight of the influence of static spatial environment information on dynamic time series prediction; F fused This represents a fusion feature that retains the inertial patterns of system operation, keenly captures dynamic disturbances in the meteorological environment, and incorporates the physical environmental context. This fusion feature will serve as input to the collaborative prediction model.

[0041] This embodiment introduces two physical anchors, "behavioral patterns" and "global meteorological data," and combines them with mutual information constraints to adaptively decouple the original mixed sequence into "inertial trend characteristics" and "meteorological fluctuation characteristics." This decoupling method enables the model to distinguish between steady-state components determined by production and daily life patterns and transient components caused by sudden environmental changes. Combined with the subsequent two-way interaction mechanism, it greatly improves the model's prediction accuracy and stability when facing sudden weather disturbances, achieving physical-level deep fusion of source and load carbon data.

[0042] The fused features obtained through the above embodiments will be used as input data for the collaborative prediction model to perform source-load carbon collaborative prediction. The collaborative prediction model can adopt a multi-task prediction architecture with a neural network model and three independent prediction heads. The neural network model can use a Transformer encoder or a long short-term memory neural network to extract the dependencies in the fused features, and the extracted features will be predicted individually by three parallel prediction heads to obtain the source-load carbon prediction values, namely, photovoltaic power output prediction value, power load prediction value and carbon emission intensity prediction value.

[0043] In a preferred embodiment, to address the problem that traditional Transformers easily establish long-range dependencies that violate physical common sense when processing energy data (e.g., nighttime photovoltaic output incorrectly focuses on midday irradiance characteristics), and to improve the accuracy of prediction results, this invention designs a collaborative prediction model consisting of a multi-layered stacked physically embedded Transformer encoder, an adaptive attention pooling layer, and a shared private decoupled prediction head. Preferably, the multi-layered stacked physically embedded Transformer encoder (PIA-Encoder) employs a 4-layer stacked structure.

[0044] The PIA-Encoder proposed in this embodiment can be understood as an innovative extension based on the standard Transformer encoder architecture. It retains the multi-head attention mechanism and feedforward neural network structure of the Transformer encoder, but in the self-attention calculation, it deeply embeds physical knowledge through the design of a Physical Prior Embedded Attention Operator (PIA). That is, each encoder layer contains two sub-layers: one is the multi-head physical embedded attention mechanism (Multi-Head PIA), and the other is the feedforward neural network. Its specific construction and calculation process is as follows: (1) Construction of physical state vector and similarity matrix: First, based on physical laws, the "physical environment homogeneity" between any two time steps in the sequence is quantified: (a) State vector extraction: Extract the physical state vector for each time step t from the input matrix A. ,in The normalized solar elevation angle (characterizing photovoltaic potential) is given at time step t. The normalized outdoor air temperature (i.e., temperature, used to characterize the load temperature control requirements) is the temperature at time step t.

[0045] (b) Physical similarity calculation: Calculate the physical similarity between the p-th time step and the q-th time step in the sequence. To take into account the different physical characteristics of photovoltaics and loads, a composite similarity metric formula is adopted: in, All are weighting coefficients. This is the temperature sensitivity factor. The formula generates a symmetric matrix. The closer the element value is to 1, the more similar the physical environment is at two moments, and the stronger the potential feature correlation should be. , respectively, are the solar altitude angles at time steps p and q. The outdoor temperatures are at time steps p and q, respectively.

[0046] (c) Dynamic physical prior mask generation based on Matrix, construct a dynamic bias matrix for intervening in attention distribution. N represents the dimension. This matrix is ​​not merely a positional mask, but a semantic mask with physical semantics. Its generation logic comprises both hard and soft constraints: in, The learnable penalty coefficient. This indicates the threshold for determining a strong physical correlation.

[0047] Hard constraint mechanism: This mechanism is specifically designed for the 'physical constraint head' responsible for capturing the physical characteristics of photovoltaics in the subsequent multi-head attention network. For photovoltaic tasks, when time step q is at night (solar altitude angle...) When time step p is during the day, the mask value is set to negative infinity. After Softmax, the attention weight at that location is reduced to zero. This mathematically forces the physically constrained branch to strictly follow the physical law that 'photovoltaic output is only related to daytime' without affecting the global branch's capture of nighttime load inertia.

[0048] Soft constraint mechanism: For situations where the physical environment differs significantly but does not violate any laws (such as midday in summer versus midday in winter), a linear penalty term is applied. It is a learnable nonnegative scalar parameter. During backpropagation, the model automatically adjusts it. The size of the value is used to balance "data-driven relevance" with "the constraints of physical rules".

[0049] (2) Multi-head physical embedded attention computation The above dynamically generated This is injected as a bias term into the attention score calculation. To resolve the logical conflict between "PV needs to be cut off at night" and "load needs to be dependent on all time periods" in multi-task collaborative prediction, this embodiment uses the Transformer's multi-head attention mechanism for group mask control.

[0050] Specifically, assuming there are H attention heads in total, they are divided into two groups: (1) Physical constraint header: accounting for 1 / 2 of the total number of headers. In this type of header, [the following will be included] It is added directly to the attention score. For photovoltaic tasks, The hard constraints in the process cause the attention weights of the nighttime time step to be zero, thus strictly adhering to the laws of physics.

[0051] (2) Global Attention Head: In this type of head, no hard constraint containing "nighttime truncation" is injected. (or just inject) The soft constraint part allows the model to still focus on historical load and carbon emission characteristics at night, so as to preserve the time-dependent connectivity throughout the time period and capture the continuity characteristics of tasks such as load and carbon emissions.

[0052] The formula for calculating attention is: in, This is an indicator function, which is 1 when the h-th head belongs to the physical constraint group, and 0 otherwise. Let Q, K, and V represent the dimensions of the attention head. Q, K, and V represent the query, key, and value in the attention mechanism, respectively.

[0053] In this embodiment, the PIA-Encoder will input the fused feature sequence. After the above multi-layer encoding process, the output is a long-range dependency feature sequence containing physical logic. ,in N represents the time dimension. For feature dimensions.

[0054] This embodiment employs the PIA operator and its dynamic masking and grouping strategies to deeply and interpretably embed structured physical knowledge into attention computation, guiding the model to follow physical laws while learning statistical patterns in the data. This design enables the establishment of spatiotemporal dependencies to have causal logic, effectively reducing prediction biases that violate physical common sense. It achieves physical-level constraints on photovoltaic output at the microscopic level while preserving the global context integrity required for multi-task prediction at the macroscopic level. This solves the shortcomings of purely data-driven models in terms of physical consistency, while avoiding the suppression effect on non-photovoltaic tasks. This enhances the model's generalization ability under extreme conditions or data-sparse scenarios and improves the interpretability and robustness of deep learning models.

[0055] To achieve accurate and coordinated prediction of photovoltaic power, load, and carbon emissions, this embodiment designs an "adaptive attention pooling layer" and a "shared private decoupled prediction head." Specifically, traditional mean pooling tends to distribute key abrupt changes in the sequence (such as sudden drops in photovoltaic power output), while taking the last moment easily ignores historical context. Therefore, this embodiment uses adaptive attention pooling to compress the sequence and uses calculated weights to perform a weighted summation of the sequence, thereby generating a fixed-length aggregate vector. In the formula, and For learnable parameters, Indicates the first The importance weights of each time step, where N represents the total number of time steps. Indicates the first The operation, which takes a sequence of long-range dependent features at each time step (i.e., time t), allows the aggregation vector to... It can adaptively focus on key mutation points in the sequence.

[0056] Considering that photovoltaic output, electricity load, and carbon emission intensity are all affected by the same external environment (such as temperature and irradiance), there exists a "covariance" among them. To capture this implicit synergistic pattern, this embodiment constructs a cross-task sharing network. The specific calculations are as follows: Aggregate vector The input is fed into a shared multilayer perceptron to extract the common environmental response patterns shared by all three components: in, The source-load carbon synergy feature vector characterizes cross-task coupling logic such as "high temperature leads to load surge and photovoltaic efficiency decline". and is a learnable weight matrix for cross-task shared networks, used to extract covariant features among multiple tasks; GELU is the Gaussian error linear unit activation function.

[0057] Based on the extraction of covariant features, independent predictions are constructed for the three targets to preserve their unique physical properties. The specific steps include: (1) Feature splicing and decoupling: For each prediction task (i.e., photovoltaic output, electricity load, and carbon emission intensity, respectively), sharing the source-load carbon synergy characteristics of all tasks. Aggregation features shared across all tasks The data is then concatenated to form the input data that serves as the independent prediction head for this task: This represents the input features of the k-th prediction head, which are content-dependent for all prediction tasks. They are exactly the same. This "jump connection" design ensures that the task can see both the original information and utilize collaborative information. What enables differentiated predictions for different tasks is not... The content is different; instead, they are independent regression prediction heads configured separately for each task k in subsequent steps, with parameters not shared. These independent prediction heads will draw from the same... In this process, information is learned and extracted based on the characteristics of each task, thereby providing task-specific prediction results.

[0058] (2) Independent regression prediction: Will Input each independent fully connected network and output preliminary predictions: In the formula, This represents the predicted electricity load. Indicates the power load forecast header. This indicates the input characteristics of the power load forecasting head. This represents the predicted photovoltaic power output. Indicates the photovoltaic power output forecast head, This indicates the input characteristics of the photovoltaic power output prediction head. This represents the predicted value of carbon emission intensity. This indicates the carbon emission intensity forecast head. This represents the input characteristics of the carbon emission intensity prediction head.

[0059] In a preferred embodiment, after obtaining the predicted source load carbon value of the enterprise to be predicted, the method further includes: Based on the physical parameters of the time to be predicted and the preset ideal photovoltaic output model, the upper limit of photovoltaic output is calculated. The ideal photovoltaic output model is constructed based on astronomical geometric relationships and clear-sky radiation model. The photovoltaic output prediction value is extracted from the source load carbon prediction value, and the difference between the photovoltaic output prediction value and the photovoltaic output upper limit value is calculated. The difference and the environmental parameters at the time to be predicted are input into a preset performance prediction model to obtain the dynamic performance ratio. The performance prediction model is constructed based on a multilayer perceptron network. Based on the dynamic performance ratio and the upper limit of photovoltaic output, an optimized photovoltaic output prediction value is obtained.

[0060] In this embodiment, the upper limit of photovoltaic output at a future time is first predicted using an ideal photovoltaic output model. This future time coincides with the time predicted by the collaborative prediction model. The ideal photovoltaic output model is constructed based on deterministic astronomical geometric relationships and a clear-sky radiation model. The calculation of the upper limit of photovoltaic output depends only on the site's geographical information and a future timestamp. The specific calculation process is as follows: (1) Obtain physical parameters: Site geographical features: including longitude ,latitude and altitude ; Photovoltaic array characteristics: including the rated installed capacity of photovoltaic modules. Installation tilt angle and installation azimuth angle ; Temporal characteristics: the future moment to be predicted (Including year, month, day, hour, and minute), used to calculate accumulated days and true solar time.

[0061] (2) Calculation of the Sun's spatial position: Based on the above input features, the Solar Position Algorithm (SPA) is used to calculate... Solar geometric parameters at time: First, calculate the solar declination angle. And the hour angle f; then calculate the solar altitude angle. With solar azimuth Among them, the solar altitude angle The calculation formula is: At the same time, calculate the solar azimuth angle. To determine the direction of light incidence: (3) Calculation of power under ideal clear sky: Using the solar geometric parameters obtained in the previous step, combined with the installation information of the photovoltaic panel itself, the angle of incidence of light is calculated. The horizontal clear-sky radiation is then projected onto the tilted surface of the photovoltaic panel to obtain the final upper limit of photovoltaic output. First, calculate the angle of incidence of the light ray. : In the formula, These are the solar altitude angle and azimuth angle calculated in the previous step; Install the tilt angle for the photovoltaic panels; The azimuth angle for installing photovoltaic panels.

[0062] Calculation of clear-sky direct radiation using the Ineichen-Perez model With scattered radiation And project it onto the photovoltaic panel plane: In the formula, for The intensity of direct sunlight under clear skies at any given moment; for The intensity of clear-sky diffuse radiation at any given moment; The angle of incidence of the light ray; Install the tilt angle for the photovoltaic panels; the latter item For the perspective factor of the isotropic sky model, This refers to the rated installation capacity of the photovoltaic modules.

[0063] After obtaining the upper limit of photovoltaic power output at future times and the predicted photovoltaic power output value predicted by the collaborative prediction model, the deviation between the two is calculated. The deviation feature can explicitly quantify the degree of deviation of the preliminary prediction from the theoretical physical upper limit. As a key clue, it keenly reflects the preliminary perception of non-ideal physical states, including but not limited to cloud cover, equipment failure, and ambient temperature, as well as the collaborative prediction model's perception of these states.

[0064] Environmental parameters, including solar irradiance, temperature, and solar altitude angle, are acquired at the time of prediction. The deviation Δ and these environmental parameters are then input into a multilayer perceptron network to predict the dynamic performance ratio of photovoltaics. To address the potential for negative values ​​or numerical divergence in traditional quadratic prediction, the output layer of this multilayer perceptron network is forced to use a sigmoid activation function, thus ensuring the output dynamic performance ratio is accurate. Strictly limited to The range.

[0065] This multilayer perceptron network employs a composite loss function. The training process is expressed by the following formula: In the formula, L mse The term L represents the mean squared error between the predicted and actual values, used to ensure numerical accuracy. bound L represents the boundary constraint term, used to penalize situations where the performance ratio exceeds the physical upper bound; smooth This represents a smooth transition constraint term used to penalize nighttime (solar altitude angle). Abnormal fluctuations in performance ratio that are not zero.

[0066] Specifically, boundary constraint terms Constructed using the ReLU activation function: In the formula, Indicates the time step (or batch size) of the training sequence; For time step index; As the activation function, this formula forces the dynamic performance ratio to remain within the [0, 1] interval.

[0067] Smooth transition constraint Defined as: in, Indicates the first The solar altitude angle at each time step; among which As an indicator function, when the solar altitude angle The function sets the value to 1 if the nighttime performance ratio is 1 and 0 otherwise, forcing the nighttime performance ratio to approach 0.

[0068] Finally, the output dynamic performance ratio is multiplied by the calculated upper limit of photovoltaic output to obtain the optimized photovoltaic output prediction value: In the formula, This represents the optimized photovoltaic power output forecast. This indicates the upper limit of photovoltaic power output. This indicates the dynamic performance ratio.

[0069] In this embodiment, the upper limit of photovoltaic output This is a theoretical upper limit calculated based on the astronomical formula for the future time to be predicted and assuming ideal clear-sky conditions (i.e., no cloud cover and no equipment degradation). It can be understood as the ideal clear-sky radiation power, serving as a physical benchmark for the future time. The deviation Δ represents the difference between the preliminary prediction value and the theoretical physical upper limit. This difference incorporates comprehensive information from multiple physical and data-driven factors: it reflects the collaborative prediction model's initial perception of the actual attenuation at the future time (such as cloud cover, ambient temperature effects, and equipment performance), and may also include the prediction uncertainty of the collaborative prediction model itself.

[0070] This embodiment optimizes and corrects the photovoltaic (PV) output prediction based on deviations, aiming to provide a more physically interpretable PV output generation mechanism. The essential generation form of PV output is the product of a physical upper limit and a dynamic performance ratio (RPR). While the collaborative prediction model makes predictions, these predictions are black-box predictions. This embodiment uses the difference between the model's predicted value and the calculated physical upper limit as a key clue, combining it with predicted environmental parameters for future moments to more accurately estimate the RRP. In this way, the black-box prediction result is decomposed and mapped to physically interpretable components. The PV output prediction formula in this embodiment is based on the physical principle that the actual output of a PV panel at a given moment should be the theoretical maximum output power multiplied by the efficiency (i.e., the RRP) under current environmental and equipment conditions. This efficiency reflects the actual operating efficiency or performance degradation ratio of the PV panel relative to ideal conditions after being affected by factors such as cloud cover, excessively high / low temperatures, etc., in the actual environment.

[0071] The model in this embodiment predicts this by learning the relationship between Δ and environmental parameters. PR dynamic Here, Δ is an important "correction signal" used to represent the co-prediction model's initial assessment of the actual degradation. The final optimized correction formula multiplies the deterministic physical upper limit by the dynamic degradation ratio estimated by data-driven methods combined with physical cues, resulting in a final photovoltaic output prediction that is both physically reasonable and data-driven, with stronger physical consistency and interpretability. This method avoids the possibility that traditional black-box models might predict results that violate physical common sense under extreme or abnormal weather conditions, thus improving the model's robustness. This embodiment uses a physics-driven framework for refined interpretation and calibration to ensure that the final prediction possesses both data-driven accuracy and physical plausibility.

[0072] This embodiment provides a collaborative prediction method for enterprise source-load carbon data. Through a dynamic meteorological kernel network, it effectively captures spatial dependencies with directionality and mobility, providing dynamic spatial context information containing physical dynamics for subsequent time-series predictions, thereby improving the accuracy of prediction results. By comparing variational mode decoupling networks to distinguish between steady-state components determined by production and living patterns and transient components caused by environmental abrupt changes, and in conjunction with a two-way interaction mechanism, it improves the prediction accuracy and stability of the model when facing sudden weather disturbances, achieving physical-level deep fusion of source-load carbon data. Through the PIA operator and its dynamic masking and grouping strategies in the collaborative prediction model, structured physical knowledge is deeply and interpretably embedded into attention calculations, guiding the model to follow physical laws while learning data statistical patterns. This enhances the model's generalization ability under extreme conditions or data-sparse scenarios and improves the interpretability and robustness of deep learning models.

[0073] Please see Figure 2 Based on the same inventive concept, the second embodiment of this invention proposes an enterprise source-load carbon synergistic prediction system, comprising: The feature extraction module 10 is used to acquire the original operating data and meteorological data of the enterprise to be predicted, input the meteorological data into a preset dynamic meteorological kernel network, and obtain dynamic spatial meteorological features. The dynamic meteorological kernel network is constructed based on a convolutional neural network. The original operational data and the dynamic spatial meteorological features are input into a preset comparative variational mode decoupling network to obtain inertial trend features and meteorological fluctuation features. The comparative variational mode decoupling network is constructed based on a variational filter network with physically guided comparison anchor points. The feature fusion module 20 is used to perform feature enhancement and feature fusion on the inertial trend features, the meteorological fluctuation features and the dynamic spatial meteorological features using a bidirectional cross-attention and gating mechanism to obtain fused features; The collaborative prediction module 30 is used to input the fused features into a preset collaborative prediction model to obtain the source load carbon prediction value of the enterprise to be predicted. The collaborative prediction model adopts a multi-task prediction model architecture and is constructed based on a Transformer encoder. The difference between the predicted photovoltaic output value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model is calculated in the source load carbon prediction value. The difference is input into the performance prediction model constructed based on the multilayer perceptron network to obtain the dynamic performance ratio. Based on the dynamic performance ratio and the upper limit of photovoltaic output, the optimized photovoltaic output prediction value is obtained.

[0074] The technical features and effects of the enterprise source-load carbon co-prediction system proposed in this embodiment of the invention are the same as those of the method proposed in this embodiment of the invention, and will not be repeated here. Each module in the above-mentioned enterprise source-load carbon co-prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0075] Furthermore, embodiments of the present invention also propose a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0076] Please see Figure 3The diagram illustrates the internal structure of a computer device in one embodiment. This computer device can specifically be a terminal or a server. The computer device includes a processor, memory, network interface, display, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an enterprise source-load carbon co-prediction method. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0077] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0078] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0079] In summary, the present invention proposes a method, system, device, and storage medium for collaborative prediction of enterprise source-load carbon. The method acquires the original operational data and meteorological data of the enterprise to be predicted. The meteorological data is input into a preset dynamic meteorological kernel network to obtain dynamic spatial meteorological features. This dynamic meteorological kernel network is constructed based on a convolutional neural network. The original operational data and the dynamic spatial meteorological features are input into a preset contrastive variational mode decoupling network to obtain inertial trend features and meteorological fluctuation features. This contrastive variational mode decoupling network is constructed based on a variational filtering network with physically guided contrast anchors. A bidirectional cross-attention and gating mechanism is used to enhance and fuse the inertial trend features, meteorological fluctuation features, and dynamic spatial meteorological features to obtain fused features. The fused features are input into a preset collaborative prediction model to obtain the predicted source-load carbon value of the enterprise to be predicted. This collaborative prediction model adopts a multi-task prediction model architecture and is constructed based on a Transformer encoder. This invention employs a dynamic meteorological kernel network to deeply perceive meteorological data, effectively capturing spatial dependencies with directionality and fluidity. This provides dynamic spatial context information containing physical dynamics for subsequent time-series forecasts, thereby improving the accuracy of prediction results. By comparing variational mode decoupling networks to distinguish between steady-state components determined by production and lifestyle patterns and transient components caused by environmental abrupt changes, and in conjunction with a bidirectional interaction mechanism, the prediction accuracy and stability of the model are improved when facing sudden weather disturbances, achieving a physical-level deep fusion of source-load carbon data. Through the PIA operator and its dynamic masking and grouping strategies in the collaborative prediction model, structured physical knowledge is deeply and interpretably embedded into attention computation, guiding the model to follow physical laws while learning statistical patterns in the data. This enhances the model's generalization ability under extreme conditions or in data-sparse scenarios, and improves the interpretability and robustness of deep learning models. In conclusion, this invention effectively improves the accuracy and stability of enterprise source-load carbon collaborative prediction results through deep perception of meteorological data, a physical-guided feature decoupling mechanism, and a prediction architecture based on deep coupling of data-driven and physical mechanisms.

[0080] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0081] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for predicting the synergistic carbon source and load of enterprises, characterized in that, include: The original operational data and meteorological data of the enterprise to be predicted are obtained, and the meteorological data is input into a preset dynamic meteorological kernel network to obtain dynamic spatial meteorological characteristics. The dynamic meteorological kernel network is constructed based on a convolutional neural network. The original operational data and the dynamic spatial meteorological features are input into a preset comparative variational mode decoupling network to obtain inertial trend features and meteorological fluctuation features. The comparative variational mode decoupling network is constructed based on a variational filter network with physically guided comparison anchor points. A bidirectional cross-attention and gating mechanism is used to enhance and fuse the inertial trend features, the meteorological fluctuation features, and the dynamic spatial meteorological features to obtain fused features; The fused features are input into a preset collaborative prediction model to obtain the source load carbon prediction value of the enterprise to be predicted. The collaborative prediction model adopts a multi-task prediction model architecture and is built based on a physical embedded Transformer encoder. The difference between the predicted photovoltaic output value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model is calculated in the source load carbon prediction value. The difference is input into the performance prediction model constructed based on the multilayer perceptron network to obtain the dynamic performance ratio. Based on the dynamic performance ratio and the upper limit of photovoltaic output, the optimized photovoltaic output prediction value is obtained.

2. The enterprise source-load carbon synergistic prediction method according to claim 1, characterized in that, The steps for obtaining the raw operational data and meteorological data of the enterprise to be predicted include: Obtain the raw operating data of the enterprise to be predicted in the current period. The raw operating data includes real-time source load carbon data, local meteorological data, time-derived data, physical model data, and lagged source load carbon data. The meteorological data of the enterprise to be predicted is obtained through the meteorological data service interface. The meteorological data includes dynamic meteorological data and environmental characteristic meteorological data.

3. The enterprise source-load carbon synergistic prediction method according to claim 2, characterized in that, The dynamic meteorological kernel network consists of a series of dynamic convolutional modules and a feature dimensionality reduction module; The dynamic convolution module consists of multiple cascaded dynamic convolutional layers, used to perform dynamic resampling and convolution operations on the meteorological data. The offset used by each dynamic convolutional layer during resampling is calculated by one convolutional layer from the wind speed vector in the meteorological data. The feature dimensionality reduction module consists of a global average pooling layer and a first multilayer perceptron. The global average pooling layer is used to aggregate information from the high-dimensional spatial features output by the dynamic convolution module and output the pooled features. The first multilayer perceptron is used to perform dimensionality mapping on the pooled features and output dynamic spatial meteorological features.

4. The enterprise source-load carbon synergistic prediction method according to claim 2, characterized in that, The contrastive variational mode decoupling network consists of a series of physical anchor modules and a variational filter network; The physical anchor point module consists of two parallel fully connected neural networks. The input data of the first fully connected neural network are time-derived data and physical model data in the original operating data, and the output data is behavioral pattern anchor points. The input data of the second fully connected neural network are local meteorological data and dynamic spatial meteorological features in the original operating data, and the output data is global meteorological anchor points. The variational filter network takes the real-time source load carbon data and the lagged source load carbon data in the original operating data as input data, and uses two parallel frequency domain filter channels and inverse fast Fourier transform to perform frequency domain filtering and time domain reconstruction on the input data, and outputs inertial trend characteristics and meteorological fluctuation characteristics. The loss function of the variational filter network consists of reconstruction loss, physical alignment loss and orthogonality independence loss, and the physical alignment loss consists of the similarity between the inertial trend feature and the behavioral pattern anchor point, as well as the similarity between the meteorological fluctuation feature and the global meteorological anchor point.

5. The enterprise source-load carbon synergistic prediction method according to claim 2, characterized in that, The steps of using a bidirectional cross-attention and gating mechanism to enhance and fuse the inertial trend features, meteorological fluctuation features, and dynamic spatial meteorological features to obtain fused features include: The inertial trend feature and the meteorological fluctuation feature are enhanced by using bidirectional cross-attention to obtain enhanced inertial trend feature and enhanced meteorological fluctuation feature; The dynamic spatial meteorological features are concatenated with the time-derived data in the original operational data to obtain a scene vector; Based on the scene vector, a gating mechanism is used to generate a dynamic weight vector. Based on the dynamic weight vector, the enhanced inertial trend feature, the enhanced meteorological fluctuation feature, and the dynamic spatial meteorological feature are fused to obtain the fused feature.

6. The enterprise source-load carbon synergistic prediction method according to claim 1, characterized in that, The collaborative prediction model consists of a multi-layered stacked physically embedded Transformer encoder, an adaptive attention pooling layer, and a shared private decoupled prediction head; The multi-layered stacked physical embedded Transformer encoder is used to capture the long-range spatiotemporal dependencies between the fused features through a grouping mechanism of physical constraint head and global attention head, and output a long-range dependency feature sequence; wherein, the physical constraint head applies constraints through a physical constraint mask matrix, which is constructed based on the physical state similarity of the input original running data at different time steps; The adaptive attention pooling layer is used to perform adaptive attention pooling on the long-range dependent feature sequence and output an aggregate vector; The shared private decoupled prediction head is used to extract shared features from the aggregated vector, concatenate the extracted source-load carbon co-feature vector with the aggregated vector, and perform independent regression prediction on the concatenated features to obtain the source-load carbon prediction value. The source-load carbon prediction value includes the photovoltaic output prediction value, the power load prediction value, and the carbon emission intensity prediction value.

7. The enterprise source-load carbon synergistic prediction method according to claim 1, characterized in that, The steps of calculating the difference between the predicted photovoltaic output value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model, inputting the difference into the performance prediction model constructed based on a multilayer perceptron network to obtain the dynamic performance ratio, and obtaining the optimized predicted photovoltaic output value based on the dynamic performance ratio and the upper limit of photovoltaic output include: Based on the physical parameters of the time to be predicted and the preset ideal photovoltaic output model, the upper limit of photovoltaic output is calculated. The ideal photovoltaic output model is constructed based on astronomical geometric relationships and clear-sky radiation model. The photovoltaic output prediction value is extracted from the source load carbon prediction value, and the difference between the photovoltaic output prediction value and the photovoltaic output upper limit value is calculated. The difference and the environmental parameters at the time to be predicted are input into a preset performance prediction model to obtain the dynamic performance ratio. The performance prediction model is constructed based on a multilayer perceptron network. Based on the dynamic performance ratio and the upper limit of photovoltaic output, an optimized photovoltaic output prediction value is obtained.

8. A corporate source-load carbon co-prediction system, characterized in that, include: The feature extraction module is used to acquire the original operational data and meteorological data of the enterprise to be predicted, and input the meteorological data into a preset dynamic meteorological kernel network to obtain dynamic spatial meteorological features. The dynamic meteorological kernel network is constructed based on a convolutional neural network. The original operational data and the dynamic spatial meteorological features are input into a preset comparative variational mode decoupling network to obtain inertial trend features and meteorological fluctuation features. The comparative variational mode decoupling network is constructed based on a variational filter network with physically guided comparison anchor points. The feature fusion module is used to enhance and fuse the inertial trend features, the meteorological fluctuation features, and the dynamic spatial meteorological features using a bidirectional cross-attention and gating mechanism to obtain fused features; The collaborative prediction module is used to input the fused features into a preset collaborative prediction model to obtain the source load carbon prediction value of the enterprise to be predicted. The collaborative prediction model adopts a multi-task prediction model architecture and is built based on a physically embedded Transformer encoder. The difference between the predicted photovoltaic output value and the upper limit of photovoltaic output obtained based on the ideal photovoltaic output model is calculated in the source load carbon prediction value. The difference is input into the performance prediction model constructed based on the multilayer perceptron network to obtain the dynamic performance ratio. Based on the dynamic performance ratio and the upper limit of photovoltaic output, the optimized photovoltaic output prediction value is obtained.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images