Power generation side unit carbon emission prediction method and device based on large time sequence model
By using a time-series large model-based approach and employing spatiotemporal scale alignment and feature fusion with multi-source data, the nonlinear dependency and data scarcity issues in carbon emission prediction for power generation units were resolved, achieving high-precision carbon emission prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID INFORMATION & TELECOMM BRANCH
- Filing Date
- 2025-12-25
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to capture the complex nonlinear dependencies between carbon emissions, weather, geography, and load in predicting carbon emissions from power generation units. They also have poor generalization capabilities, cannot effectively utilize spatial correlation information from multi-source data, and suffer from data scarcity and fragmentation issues.
We employ a time-series large model-based approach, aligning spatiotemporal scales using multi-source basic data, extracting spatial features using far-field and near-field encoder structures, fusing features through attention mechanisms and fully connected networks, and fine-tuning the model using a lightweight bottleneck structure adapter to achieve high-precision prediction.
By effectively extracting and fusing features from spatial data, high-precision carbon emission prediction was achieved with limited data, solving the problems of data scarcity and spatial correlation, and improving prediction accuracy.
Smart Images
Figure CN122022003A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of carbon emission prediction technology, specifically to a method and apparatus for predicting carbon emissions from power generation units based on a time-series large model. Background Technology With the enormous consumption of energy resources in the power sector, the energy system is undergoing a transformation towards cleaner, lower-carbon, safer, and more efficient systems. The main sources of carbon emissions in the power sector are concentrated in power generation units. Against this backdrop, refined prediction of carbon emissions from power generation units plays a positive role in the formulation of carbon emission strategies.
[0002] Based on a survey of the current status of carbon emission forecasting, there are two main technical approaches to traditional power generation-side carbon emission forecasting: one is the emission source quantification-statistical accounting method, which identifies emission source parameters such as fuel consumption and combustion efficiency of generator sets, calculates historical carbon emissions using IPCC accounting formulas, and then extrapolates the forecast using traditional statistical learning methods such as autoregressive integrated moving average model (ARIMA) and support vector machine (SVM); the other is the "data-driven-simple multivariate coupling" method, which attempts to combine meteorological data as covariates and use convolutional neural networks (CNN) or long short-term memory networks (LSTM) for forecasting.
[0003] However, such methods are usually limited to univariate time series fitting or linear variable coupling, making it difficult to capture the complex nonlinear dependencies between carbon emissions, weather, geography, and load, and they have poor generalization ability—requiring redesign of feature engineering and model structure for new power generation sites.
[0004] With the success of large-scale pre-trained models in Natural Language Processing (NLP) and Computer Vision, a new paradigm of "pre-training-fine-tuning" is gradually emerging in the field of time series forecasting: Time Series Foundation Models (TSFMs) learn general spatiotemporal knowledge through pre-training on massive amounts of heterogeneous data, enabling them to quickly adapt to downstream tasks (such as load forecasting and weather forecasting). For example, large-scale time series models such as Chronos (ChronosX) and TimeGPT can handle multidimensional data and complex time dependencies, but they face technical bottlenecks in the scenario of carbon emission forecasting on the power generation side: Power generation side data has the characteristics of strong spatial correlation and high data heterogeneity (grid data + point data). Existing TSFMs only use multi-source data as flat covariate inputs and cannot utilize their inherent spatial correlation information, resulting in limited prediction accuracy.
[0005] Although there have been attempts to improve the time series large model (TSFMs) field by using adapter injection, such designs still have two major limitations that prevent them from being adapted to the scenario of generating unit carbon emission prediction: 1. Adapter functionality limitations: It only processes a single type of time-series data and does not integrate multi-source spatial features. 2. Limitations of adapter structure and scenario adaptability: There is no lightweight bottleneck design, and the selection of the insertion layer lacks spatial feature adaptation logic.
[0006] Furthermore, due to the scarcity and fragmentation of carbon emission data from power generation units, achieving high-precision prediction based on limited data has become a key bottleneck in the current application of this technology. Summary of the Invention
[0007] To overcome the above-mentioned shortcomings, this invention proposes a method and device for predicting carbon emissions from power generation units based on a time-series large model.
[0008] Firstly, a method for predicting carbon emissions from generator-side units based on a time-series large-scale model is provided, the method comprising: The time series input sequence of a pre-trained time series large model is determined based on multi-source basic data from power plants; The time-series input sequence is used as the input to a pre-trained time-series large model to obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large model. The multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data.
[0009] Preferably, the step of determining the time-series input sequence of the pre-trained time-series large model based on multi-source basic data from power plants includes: Spatiotemporal scale alignment is performed on the multi-source basic data of the power plant. Centered on the latitude and longitude of the power plant, extract pre-set far-field scale grid data and pre-set near-field scale grid data; The features of the preset far-field scale grid data and the preset near-field scale grid data are extracted by the far-field encoder structure and the near-field encoder structure, respectively, to obtain the far-field grid data features and the near-field grid data features. The far-field grid data features and the near-field grid data features are concatenated to obtain the grid data features; The attention mechanism and fully connected network are used to fuse the grid data features corresponding to carbon emission grid data and the grid data features corresponding to ERA5 meteorological data to obtain fused features. The time series sequence corresponding to the fusion feature and the time series sequence corresponding to the power plant load data are concatenated to obtain the time series input sequence.
[0010] Furthermore, the spatiotemporal scale alignment of the multi-source basic data of the power plant includes: The ERA5 meteorological data is upsampled or downsampled to make its spatial resolution consistent with that of the carbon emission grid data; Power plant load data is aggregated daily to align its time scale with carbon emission grid data and ERA5 meteorological data.
[0011] Furthermore, after upsampling the ERA5 meteorological data, the interpolation points are calculated using the following formula. Perform data interpolation:
[0012] In the above formula, Points to be interpolated interpolated data, These are the coordinates of the four vertices of the ERA5 meteorological data grid. These are the locations in the ERA5 meteorological data grid. ERA5 meteorological data.
[0013] Furthermore, the preset far-field scale grid data corresponds to a scale of 17×17, and the preset near-field scale grid data corresponds to a scale of 7×7.
[0014] Furthermore, the far-field encoder structure includes: a first convolutional layer, a first ReLU activation layer, a first max pooling layer, and a first ConvLSTM layer; the near-field encoder structure includes: a second convolutional layer, a second ReLU activation layer, a second max pooling layer, a third convolutional layer, a second ConvLSTM layer, and a global average pooling layer. The first convolutional layer comprises 32 3×3 convolutional kernels, with a stride of 1 and padding of 1. The first ConvLSTM layer has 64 hidden units and a time step of 7. The second convolutional layer comprises 32 7×7 convolutional kernels, with a stride of 1 and padding of 3. The third convolutional layer comprises 64 5×5 convolutional kernels, with a stride of 1 and padding of 2. The second ConvLSTM layer has 128 hidden units and a time step of 7.
[0015] Furthermore, the fusion features are as follows:
[0016] In the above formula, As a feature of fusion, It is the ReLU activation function. The first parameter is the first fully connected layer. As the first attention weight, The grid data features corresponding to the carbon emission grid data. As the second attention weight, The grid data features corresponding to ERA5 meteorological data. The second parameter is the first fully connected layer. The first parameter of the second fully connected layer. This is the second parameter of the second fully connected layer.
[0017] Furthermore, the first attention weight and the second attention weight are as follows:
[0018] In the above formula, The parameters of the first attention layer corresponding to the carbon emission grid data. The parameters of the second attention layer corresponding to the carbon emission grid data. These are the parameters of the first attention layer corresponding to the ERA5 meteorological data. These are the parameters of the second attention layer corresponding to the ERA5 meteorological data.
[0019] Preferably, the training process of the pre-trained time-series large model includes: The time series input sequence of the large time series model is determined based on historical multi-source basic data of power plants; Training data was constructed using the aforementioned time-series input sequence and historical carbon emission results from power plants. The training data is used to train the large-scale time series model to obtain the pre-trained large-scale time series model; In this process, bottleneck structure adapters are inserted into layers 3, 5, and 7 of the time series large model. When the pre-trained time series large model is fine-tuned for a new power plant, the training data corresponding to the new power plant is used to train the pre-trained time series large model. During the training process, the model parameters of the pre-trained time series large model are frozen, and only the parameters of the bottleneck structure adapter are adjusted.
[0020] Secondly, a carbon emission prediction device for generator-side units based on a time-series large-scale model is provided, the device comprising: The determination module is used to determine the time series input sequence of a pre-trained time series large model based on multi-source basic data from power plants; The prediction module is used to take the time-series input sequence as input to a pre-trained time-series large model and obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large model. The multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data.
[0021] The above-described technical solutions of the present invention have at least one or more of the following beneficial effects: This invention provides a method and apparatus for predicting carbon emissions from power generation units based on a time-series large-scale model, comprising: determining the time-series input sequence of a pre-trained time-series large-scale model based on multi-source basic data of power plants; using the time-series input sequence as input to the pre-trained time-series large-scale model to obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large-scale model; wherein, the multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data. The technical solution provided by this invention can effectively extract and fuse the spatial features of spatial data and achieve high-precision prediction based on limited data. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the main steps of the carbon emission prediction method for power generation units based on a time-series large model according to an embodiment of the present invention; Figure 2 This is a spatial feature extraction structure diagram according to an embodiment of the present invention. Detailed Implementation
[0023] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0024] 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.
[0025] Example 1 See appendix Figure 1 , Figure 1 This is a schematic flowchart illustrating the main steps of a generation-side unit carbon emission prediction method based on a time-series large model, according to an embodiment of the present invention. Figure 1 As shown, the carbon emission prediction method for power generation units based on a large time-series model in this embodiment of the invention mainly includes the following steps: Step S101: Determine the time series input sequence of the pre-trained time series large model based on the multi-source basic data of the power plant station; Step S102: Use the time-series input sequence as the input of the pre-trained time-series large model to obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large model; The multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data.
[0026] In one specific implementation, the multi-source basic data of the power plant is shown in Table 1: Table 1
[0027] In this embodiment, determining the time-series input sequence of the pre-trained time-series large model based on multi-source basic data from power plants includes: Spatiotemporal scale alignment is performed on the multi-source basic data of the power plant. Centered on the latitude and longitude of the power plant, extract pre-set far-field scale grid data and pre-set near-field scale grid data; The features of the preset far-field scale grid data and the preset near-field scale grid data are extracted by the far-field encoder structure and the near-field encoder structure, respectively, to obtain the far-field grid data features and the near-field grid data features. The far-field grid data features and the near-field grid data features are concatenated to obtain the grid data features; The attention mechanism and fully connected network are used to fuse the grid data features corresponding to carbon emission grid data and the grid data features corresponding to ERA5 meteorological data to obtain fused features. The time series sequence corresponding to the fusion feature and the time series sequence corresponding to the power plant load data are concatenated to obtain the time series input sequence.
[0028] In one implementation, the spatiotemporal scale alignment of the multi-source basic data of the power plant includes: The ERA5 meteorological data is upsampled or downsampled to make its spatial resolution consistent with that of the carbon emission grid data; Power plant load data is aggregated daily to align its time scale with carbon emission grid data and ERA5 meteorological data.
[0029] This step prepares input data for subsequent spatial feature extraction. Instead of analyzing the entire grid, it precisely defines the spatial range that the model needs to focus on, centered on the power plant to be predicted. Addressing the single-point limitation problem: This is the first step in realizing a multi-scale spatial coding architecture. Traditional methods only focus on single-point data, ignoring spatial correlations.
[0030] In one implementation, after upsampling the ERA5 meteorological data, the interpolation points are determined according to the following formula. Perform data interpolation:
[0031] In the above formula, Points to be interpolated interpolated data, These are the coordinates of the four vertices of the ERA5 meteorological data grid. These are the locations in the ERA5 meteorological data grid. ERA5 meteorological data.
[0032] In one implementation, the preset far-field scale grid data corresponds to a scale of 17×17, and the preset near-field scale grid data corresponds to a scale of 7×7. Near-field scale: used to capture local spatial correlations around the power plant, such as neighboring power plants or local weather conditions. Far-field scale: used to capture large-scale background influences, such as the impact of large-scale temperature changes on overall regional carbon emissions.
[0033] In one implementation, the far-field encoder structure and the near-field encoder structure use a CNN+ConvLSTM hybrid model to extract spatial features of carbon emissions and meteorological data, respectively. Specifically: The far-field encoder structure includes: a first convolutional layer, a first ReLU activation layer, a first max pooling layer, and a first ConvLSTM layer; the near-field encoder structure includes: a second convolutional layer, a second ReLU activation layer, a second max pooling layer, a third convolutional layer, a second ConvLSTM layer, and a global average pooling layer. The first convolutional layer comprises 32 3×3 convolutional kernels, with a stride of 1 and padding of 1. The first ConvLSTM layer has 64 hidden units and a time step of 7. The second convolutional layer comprises 32 7×7 convolutional kernels, with a stride of 1 and padding of 3. The third convolutional layer comprises 64 5×5 convolutional kernels, with a stride of 1 and padding of 2. The second ConvLSTM layer has 128 hidden units and a time step of 7.
[0034] like Figure 2 As shown, the far-field encoder structure outputs far-field grid data features V_far_carbon corresponding to 128-dimensional carbon emission grid data and far-field grid data features V_far_T corresponding to 128-dimensional ERA5 meteorological data. The near-field encoder structure outputs far-field grid data features V_near_carbon corresponding to 64-dimensional carbon emission grid data and far-field grid data features V_near_T corresponding to 64-dimensional ERA5 meteorological data. The far-field grid data features and near-field grid data features are concatenated to obtain grid data features V_carbon corresponding to 192-dimensional carbon emission grid data and grid data features V_T corresponding to 192-dimensional ERA5 meteorological data.
[0035] This process addresses the loss of spatial information: traditional methods use multi-source data as flat covariate inputs, failing to utilize their inherent spatial correlation information. CNNs and ConvLSTs are standard deep learning methods for processing gridded and spatial data, automatically capturing implicit complex spatial patterns and correlations. Multi-scale encoding is achieved: the design of V_near (64-dimensional) and V_far (128-dimensional) corresponds to the surrounding area truncation step, using vectors of different dimensions to carry spatial information at different scales, realizing a structured representation of spatial features.
[0036] In one implementation, the two independent feature vectors obtained in the previous step (192-dimensional V_carbon and 192-dimensional V_T) are nonlinearly combined to form a unified feature vector V_fusion containing multi-source information. The fused features are as follows:
[0037] In the above formula, As a feature of fusion, It is the ReLU activation function. The first parameter is the first fully connected layer. As the first attention weight, The grid data features corresponding to the carbon emission grid data. As the second attention weight, The grid data features corresponding to ERA5 meteorological data. The second parameter is the first fully connected layer. The first parameter of the second fully connected layer. This is the second parameter of the second fully connected layer.
[0038] In one implementation, the first attention weight and the second attention weight are as follows:
[0039] In the above formula, (192×1) represents the parameters of the first attention layer corresponding to the carbon emission grid data. (1×1) represents the parameters of the second attention layer corresponding to the carbon emission grid data. (192×1) represents the first attention layer parameter corresponding to the ERA5 meteorological data. (1×1) represents the second attention layer parameter corresponding to the ERA5 meteorological data.
[0040] This process addresses the problem of simple coupling: traditional methods struggle to capture the complex nonlinear dependencies between carbon emissions and weather. Dynamic weighting: This is the attention-weighted fusion mechanism of this invention. It outperforms traditional simple concatenation or summation. Simulating realistic dependencies: The attention mechanism allows the model to dynamically allocate weights. For example, during hot weather, the weight of meteorological features may be automatically increased by the model, while it decreases during normal weather. This more accurately captures the chain dependency relationship of "weather → load → carbon emissions".
[0041] In this embodiment, the training process of the pre-trained time-series large model includes: The time series input sequence of the large time series model is determined based on historical multi-source basic data of power plants; Training data was constructed using the aforementioned time-series input sequence and historical carbon emission results from power plants. The training data is used to train the large-scale time series model to obtain the pre-trained large-scale time series model; In this process, bottleneck structure adapters are inserted into layers 3, 5, and 7 of the time series large model. When the pre-trained time series large model is fine-tuned for a new power plant, the training data corresponding to the new power plant is used to train the pre-trained time series large model. During the training process, the model parameters of the pre-trained time series large model are frozen, and only the parameters of the bottleneck structure adapter are adjusted.
[0042] Leveraging the powerful time series forecasting capabilities of the Time Series Large Model (TSFM), this approach addresses the issues of TSFM's inability to directly process specific spatial information and the scarcity of carbon data.
[0043] This process addresses the problem of data scarcity: carbon emission data from the power generation side is scarce and fragmented. Training a large model from scratch is not feasible. It leverages pre-trained knowledge: TSFM has already learned general spatiotemporal knowledge through pre-training on massive amounts of data. Lightweight transfer learning: adapter design is key. It eliminates the need to fine-tune all the parameters of TSFM, but only these newly inserted adapters with a small number of parameters. This significantly reduces computation and effectively avoids overfitting on scarce data, which is the core of solving the model adaptation problem in data-scarce scenarios.
[0044] The bottleneck structure design of this invention can retain the core information of spatial features. The bottleneck structure (W_down→GELU→W_up) of this adapter has been verified in the fine-tuning adaptation of BERT in NLP, which has greatly reduced the number of parameters and solved the overfitting problem in the case of scarce data. Among them, W_down (384×128) is the dimensionality reduction weight, W_up (128×384) is the dimensionality increase weight, and GELU is the activation function. The insertion layer is selected from the 3rd, 5th and 7th layers (the middle layer of TSFMs), which is the optimal fusion node based on theoretical verification. The shallow layer (the 1st and 2nd layers of TSFMs) only extracts the basic time series trend, the deep layer (the 8th and 10th layers) focuses on the global time series dependence, and the middle layer (the 3rd to 7th layers) has the feature interaction capability. At this time, the spatial features (V_fusion) are injected to maximize the capture of the chain dependence of meteorological spatial distribution → load fluctuation → carbon emission change.
[0045] In one specific implementation, the present invention uses the mean absolute ratio error to evaluate the accuracy of the model and uses the weighted quantile loss to evaluate the reliability of the predicted distribution.
[0046] Example 2 Based on the same inventive concept, this invention also provides a carbon emission prediction device for generator-side units based on a time-series large-scale model, the carbon emission prediction device for generator-side units based on a time-series large-scale model comprising: The determination module is used to determine the time series input sequence of a pre-trained time series large model based on multi-source basic data from power plants; The prediction module is used to take the time-series input sequence as input to a pre-trained time-series large model and obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large model. The multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data.
[0047] Preferably, the step of determining the time-series input sequence of the pre-trained time-series large model based on multi-source basic data from power plants includes: Spatiotemporal scale alignment is performed on the multi-source basic data of the power plant. Centered on the latitude and longitude of the power plant, extract pre-set far-field scale grid data and pre-set near-field scale grid data; The features of the preset far-field scale grid data and the preset near-field scale grid data are extracted by the far-field encoder structure and the near-field encoder structure, respectively, to obtain the far-field grid data features and the near-field grid data features. The far-field grid data features and the near-field grid data features are concatenated to obtain the grid data features; The attention mechanism and fully connected network are used to fuse the grid data features corresponding to carbon emission grid data and the grid data features corresponding to ERA5 meteorological data to obtain fused features. The time series sequence corresponding to the fusion feature and the time series sequence corresponding to the power plant load data are concatenated to obtain the time series input sequence.
[0048] Furthermore, the spatiotemporal scale alignment of the multi-source basic data of the power plant includes: The ERA5 meteorological data is upsampled or downsampled to make its spatial resolution consistent with that of the carbon emission grid data; Power plant load data is aggregated daily to align its time scale with carbon emission grid data and ERA5 meteorological data.
[0049] Furthermore, after upsampling the ERA5 meteorological data, the interpolation points are calculated using the following formula. Perform data interpolation:
[0050] In the above formula, Points to be interpolated interpolated data, These are the coordinates of the four vertices of the ERA5 meteorological data grid. These are the locations in the ERA5 meteorological data grid. ERA5 meteorological data.
[0051] Furthermore, the preset far-field scale grid data corresponds to a scale of 17×17, and the preset near-field scale grid data corresponds to a scale of 7×7.
[0052] Furthermore, the far-field encoder structure includes: a first convolutional layer, a first ReLU activation layer, a first max pooling layer, and a first ConvLSTM layer; the near-field encoder structure includes: a second convolutional layer, a second ReLU activation layer, a second max pooling layer, a third convolutional layer, a second ConvLSTM layer, and a global average pooling layer. The first convolutional layer comprises 32 3×3 convolutional kernels, with a stride of 1 and padding of 1. The first ConvLSTM layer has 64 hidden units and a time step of 7. The second convolutional layer comprises 32 7×7 convolutional kernels, with a stride of 1 and padding of 3. The third convolutional layer comprises 64 5×5 convolutional kernels, with a stride of 1 and padding of 2. The second ConvLSTM layer has 128 hidden units and a time step of 7.
[0053] Furthermore, the fusion features are as follows:
[0054] In the above formula, As a feature of fusion, It is the ReLU activation function. The first parameter is the first fully connected layer. As the first attention weight, The grid data features corresponding to the carbon emission grid data. As the second attention weight, The grid data features corresponding to ERA5 meteorological data. The second parameter is the first fully connected layer. The first parameter of the second fully connected layer. This is the second parameter of the second fully connected layer.
[0055] Furthermore, the first attention weight and the second attention weight are as follows:
[0056] In the above formula, The parameters of the first attention layer corresponding to the carbon emission grid data. The parameters of the second attention layer corresponding to the carbon emission grid data. These are the parameters of the first attention layer corresponding to the ERA5 meteorological data. These are the parameters of the second attention layer corresponding to the ERA5 meteorological data.
[0057] Preferably, the training process of the pre-trained time-series large model includes: The time series input sequence of the large time series model is determined based on historical multi-source basic data of power plants; Training data was constructed using the aforementioned time-series input sequence and historical carbon emission results from power plants. The training data is used to train the large-scale time series model to obtain the pre-trained large-scale time series model; In this process, bottleneck structure adapters are inserted into layers 3, 5, and 7 of the time series large model. When the pre-trained time series large model is fine-tuned for a new power plant, the training data corresponding to the new power plant is used to train the pre-trained time series large model. During the training process, the model parameters of the pre-trained time series large model are frozen, and only the parameters of the bottleneck structure adapter are adjusted.
[0058] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0059] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting carbon emissions from generator-side units based on a large time-series model, characterized in that, The method includes: The time series input sequence of a pre-trained time series large model is determined based on multi-source basic data from power plants; The time-series input sequence is used as the input to a pre-trained time-series large model to obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large model. The multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data.
2. The method as described in claim 1, characterized in that, The determination of the time series input sequence of the pre-trained time series large model based on multi-source basic data from power plants includes: Spatiotemporal scale alignment is performed on the multi-source basic data of the power plant. Centered on the latitude and longitude of the power plant, extract pre-set far-field scale grid data and pre-set near-field scale grid data; The features of the preset far-field scale grid data and the preset near-field scale grid data are extracted by the far-field encoder structure and the near-field encoder structure, respectively, to obtain the far-field grid data features and the near-field grid data features. The far-field grid data features and the near-field grid data features are concatenated to obtain the grid data features; The attention mechanism and fully connected network are used to fuse the grid data features corresponding to carbon emission grid data and the grid data features corresponding to ERA5 meteorological data to obtain fused features. The time series sequence corresponding to the fusion feature and the time series sequence corresponding to the power plant load data are concatenated to obtain the time series input sequence.
3. The method as described in claim 2, characterized in that, The spatiotemporal scale alignment of the multi-source basic data of the power plant includes: The ERA5 meteorological data is upsampled or downsampled to make its spatial resolution consistent with that of the carbon emission grid data; Power plant load data is aggregated daily to align its time scale with carbon emission grid data and ERA5 meteorological data.
4. The method as described in claim 3, characterized in that, After upsampling the ERA5 meteorological data, data interpolation is performed on the points to be interpolated using the following formula: In the above formula, Points to be interpolated interpolated data, These are the coordinates of the four vertices of the ERA5 meteorological data grid. These are the locations in the ERA5 meteorological data grid. ERA5 meteorological data.
5. The method as described in claim 2, characterized in that, The preset far-field scale grid data corresponds to a scale of 17×17, and the preset near-field scale grid data corresponds to a scale of 7×7.
6. The method as described in claim 2, characterized in that, The far-field encoder structure includes: a first convolutional layer, a first ReLU activation layer, a first max pooling layer, and a first ConvLSTM layer; the near-field encoder structure includes: a second convolutional layer, a second ReLU activation layer, a second max pooling layer, a third convolutional layer, a second ConvLSTM layer, and a global average pooling layer. The first convolutional layer comprises 32 3×3 convolutional kernels, with a stride of 1 and padding of 1. The first ConvLSTM layer has 64 hidden units and a time step of 7. The second convolutional layer comprises 32 7×7 convolutional kernels, with a stride of 1 and padding of 3. The third convolutional layer comprises 64 5×5 convolutional kernels, with a stride of 1 and padding of 2. The second ConvLSTM layer has 128 hidden units and a time step of 7.
7. The method as described in claim 2, characterized in that, The fusion features are as follows: In the above formula, As a feature of fusion, It is the ReLU activation function. The first parameter is the first fully connected layer. As the first attention weight, The grid data features corresponding to the carbon emission grid data. As the second attention weight, The grid data features corresponding to ERA5 meteorological data. The second parameter is the first fully connected layer. The first parameter of the second fully connected layer. This is the second parameter of the second fully connected layer.
8. The method as described in claim 7, characterized in that, The first attention weight and the second attention weight are as follows: In the above formula, The parameters of the first attention layer corresponding to the carbon emission grid data. The parameters of the second attention layer corresponding to the carbon emission grid data. These are the parameters of the first attention layer corresponding to the ERA5 meteorological data. These are the parameters of the second attention layer corresponding to the ERA5 meteorological data.
9. The method as described in claim 1, characterized in that, The training process of the pre-trained time-series large model includes: The time series input sequence of the large time series model is determined based on historical multi-source basic data from power plants; Training data was constructed using the aforementioned time-series input sequence and historical carbon emission results from power plants. The training data is used to train the large-scale time series model to obtain the pre-trained large-scale time series model; In this process, bottleneck structure adapters are inserted into layers 3, 5, and 7 of the time series large model. When the pre-trained time series large model is fine-tuned for a new power plant, the training data corresponding to the new power plant is used to train the pre-trained time series large model. During the training process, the model parameters of the pre-trained time series large model are frozen, and only the parameters of the bottleneck structure adapter are adjusted.
10. A carbon emission prediction device for generator-side units based on a large time-series model, characterized in that, The device includes: The determination module is used to determine the time series input sequence of a pre-trained time series large model based on multi-source basic data from power plants; The prediction module is used to take the time-series input sequence as input to a pre-trained time-series large model and obtain the carbon emission prediction results of the power plant output by the pre-trained time-series large model. The multi-source basic data includes: power plant load data, carbon emission grid data, and ERA5 meteorological data.