Multi-dimensional electric power carbon emission reduction path optimization method based on machine learning

Abstract

The invention discloses a multi-dimensional power carbon emission reduction path optimization method based on machine learning, and the method comprises the following steps: 1, collecting and standardizing multi-source power carbon emission data to generate a power carbon emission data set, and determining a historical low-carbon scheduling path vector; 2, constructing an electric power carbon emission influence map based on the electric power carbon emission data set; 3, constructing an attribute vector matrix and carrying out principal component analysis; 4, adopting Latin hypercube sampling to generate a candidate scheduling path vector set; step 5, calculating a Soft-DTW distance to obtain a screened candidate scheduling path vector set; 6, constructing a power flow topological graph, calculating the structure entropy value of each candidate scheduling path, and generating an optimal carbon emission reduction scheduling path; and step 7, carrying out visual display on the optimal carbon emission reduction scheduling path, and generating a power scheduling instruction set. According to the method, Latin hypercube sampling and the Soft-DTW distance are fused, and intelligent screening and scheduling of power low-carbon paths are achieved.

Description

Technical Field

[0001] This invention relates to the field of low-carbon emission reduction technology, and in particular to a multidimensional power carbon emission reduction path optimization method based on machine learning. Background Technology

[0002] With the advancement of global "dual carbon" goals, the power system, as one of the core sources of carbon emissions, is gradually transforming from a traditional dispatching model primarily focused on power supply reliability to a multi-objective coordinated dispatching model oriented towards low-carbon, green, and efficient operation. Currently, most power carbon emission reduction path optimization methods still rely on static rule setting and empirical strategy guidance, lacking in-depth modeling and dynamic response capabilities for multi-source heterogeneous power carbon emission data. This makes it difficult to accurately uncover the intrinsic characteristics of low-carbon paths within the context of the coordinated evolution of multi-dimensional dispatching variables. Especially given the complex structure of the generation side, frequent fluctuations in load demand, and the increasing proportion of renewable energy year by year, traditional dispatching path selection methods struggle to effectively measure the potential causal relationships between control variables, leading to large biases in carbon emission intensity assessment, unstable optimization results, and limited practical execution value of low-carbon dispatching instructions.

[0003] Traditional methods typically rely on static carbon emission factor models for linear programming or heuristic scheduling scheme construction, lacking the ability to model the complex relationships between scheduling variables in a graph structure and ignoring the nonlinear interactions and cross-time-series effects among control variables. Furthermore, most methods depend solely on simple distance metrics in low-dimensional feature spaces, failing to globally evaluate different scheduling strategies at the path structure level. This results in scheduling path selection results that are sensitive to initial inputs, exhibiting poor stability and generalization ability. Moreover, in the path selection stage, common methods predominantly employ hard-matching metrics, failing to consider potential nonlinear time alignment differences between samples, further weakening the learning efficiency for actual historical low-carbon experience paths.

[0004] Therefore, how to provide a multidimensional power carbon emission reduction path optimization method based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multidimensional power carbon emission reduction path optimization method based on machine learning. This invention integrates the construction of power carbon emission impact maps and principal component dimensionality reduction analysis, combined with Soft-DTW distance screening and structural entropy optimization mechanisms, to achieve intelligent generation, stability screening, and visualized deployment of multidimensional power carbon emission reduction paths, thereby improving the scientific nature and execution efficiency of power system carbon emission reduction strategies.

[0006] The multidimensional power carbon emission reduction path optimization method based on machine learning according to embodiments of the present invention includes the following steps: Step 1: Collect and standardize multi-source power carbon emission data to generate a power carbon emission dataset, and determine historical low-carbon scheduling path vectors; Step 2: Construct an impact map of electricity carbon emissions based on the electricity carbon emissions dataset; Step 3: Construct an attribute vector matrix based on the power carbon emission impact map and perform principal component analysis to build a low-dimensional scheduling feature space; Step 4: In the low-dimensional scheduling feature space, Latin hypercube sampling is used to generate a set of candidate scheduling path vectors; Step 5: Calculate the Soft-DTW distance between the candidate scheduling path vector and the historical low-carbon scheduling path vector to obtain the set of candidate scheduling path vectors after filtering; Step 6: Construct the corresponding power flow topology based on the selected candidate scheduling path vector set, calculate the structural entropy value of each candidate scheduling path, and select the candidate scheduling path with the smallest structural entropy value as the optimal carbon emission reduction scheduling path. Step 7: Visualize the optimal carbon emission reduction scheduling path and generate the corresponding power dispatch instruction set.

[0007] Optionally, the multi-source power carbon emission data includes generator set power generation data, renewable energy power generation data, real-time regional power grid load data, regional power purchase contract information and power purchase ratio data, fuel type data, historical dispatch strategies and corresponding carbon emission intensity data; the standardization steps include outlier removal, missing value imputation, time alignment and normalization processing for different types of data to generate a power carbon emission dataset with a unified structure; and based on the set carbon emission intensity threshold, selecting the optimal historical low-carbon dispatch path from the historical dispatch strategies to construct a historical low-carbon dispatch path vector.

[0008] Optionally, step two specifically includes: Each scheduling control variable in the power carbon emission data set is used as a graph node. The scheduling control variables include the power generation ratio of generator sets, the power generation ratio of renewable energy, the load allocation ratio between regions, the regional electricity purchase ratio, and the fuel consumption ratio. For each pair of scheduling control variables, the corresponding edge weight is calculated according to the variable type. The calculation steps are as follows: For any two continuous scheduling control variables, extract the corresponding time series data within the set time window, calculate the Pearson correlation coefficient of the two scheduling control variables within the corresponding time window, and use it as the edge weight between the corresponding scheduling control variable pairs in the power carbon emission impact map; For any two discrete scheduling control variables, the joint distribution probability in the same time window is statistically calculated, and the mutual information value of the joint distribution probability is used as the edge weight between the corresponding scheduling control variable pairs in the power carbon emission impact map. Connect all scheduling control variables with edge weights higher than a preset threshold in the power carbon emission impact map.

[0009] Optionally, step three specifically includes: The impact map of electricity carbon emissions is represented as an attribute vector matrix. Each row of the attribute vector matrix corresponds to a scheduling control variable node, and each column is the edge weight value between pairs of scheduling control variable nodes. Perform covariance matrix calculation on the attribute vector matrix to obtain the covariance matrix; Perform eigenvalue decomposition on the covariance matrix to obtain all eigenvalues ​​and their corresponding eigenvectors; All eigenvectors are sorted in descending order of eigenvalues, and the proportion of explained variance is calculated cumulatively. The proportion of explained variance is the ratio of the sum of the first few eigenvalues ​​to the sum of all eigenvalues. Select the eigenvectors corresponding to the first few eigenvalues ​​whose cumulative explained variance ratio is greater than a preset threshold, and use them as principal component bases. The principal component bases are the low-dimensional scheduling eigenvectors after dimensionality reduction. The low-dimensional scheduling feature space is constructed based on principal component basis.

[0010] Optionally, step four specifically includes: In the low-dimensional scheduling feature space, a set of candidate scheduling path vectors is generated using Latin hypercube sampling. The Latin hypercube sampling step includes: Each principal component dimension is divided into several equally spaced sub-intervals; In each dimension, a sample point is randomly selected from each of the sub-intervals to form a one-dimensional sampling vector; For a one-dimensional sampling vector of all dimensions, an index sequence with an equal number of sampling points is generated using a Sobol sequence, wherein the index sequence is a uniformly distributed permutation of integers without repetition. The one-dimensional sampling points of the corresponding dimension are rearranged according to the index order generated by the Sobol sequence; Each dimension's one-dimensional sampling vector is treated as an independent sampling axis, and different sampling axes are orthogonally combined. A sampling point is selected sequentially from each sampling axis to form a complete multidimensional sample point, and a sampling matrix is ​​constructed. The sequence of sample points in each row of the sampling matrix corresponds to a candidate scheduling path vector. Based on the sampling matrix, construct a set of candidate scheduling path vectors.

[0011] Optionally, step five specifically includes: For each candidate scheduling path vector in the candidate scheduling path vector set, pair it with the historical low-carbon scheduling path vector and calculate the Soft-DTW distance. The specific steps for calculating the Soft-DTW distance include: Construct a two-dimensional cost matrix, where each element of the two-dimensional cost matrix represents the Euclidean distance between the corresponding dimension sample points of the candidate scheduling path vector and the historical low-carbon scheduling path vector at different time points; In the two-dimensional cost matrix, based on the Euclidean distance value, and following the minimum cumulative cost propagation rule, the cumulative cost value is updated sequentially from the first element in the upper left corner of the two-dimensional cost matrix to the lower right corner. The path with the minimum cumulative cost among all possible paths is selected as the optimal alignment path. The Euclidean distances between all paired sample points on the optimal alignment path are accumulated to obtain the Soft-DTW distance between the candidate scheduling path vector and the historical low-carbon scheduling path vector. Candidate scheduling path vectors whose Soft-DTW distance exceeds a preset threshold are removed from the candidate scheduling path vector set to obtain the filtered candidate scheduling path vector set.

[0012] Optionally, step six specifically includes: Each candidate scheduling path vector in the candidate scheduling path vector set is treated as a candidate scheduling path composed of sample points. Based on each candidate scheduling path, a corresponding power flow topology graph is constructed. The specific construction steps are as follows: Each candidate scheduling path contains sample points as nodes in the power flow topology graph, and each node has corresponding coordinates in the low-dimensional scheduling feature space. Traverse all nodes and calculate the Euclidean distance between any two nodes in the low-dimensional scheduling feature space. If the Euclidean distance between two nodes is less than a preset connection threshold, an edge connection is established between the two corresponding nodes, and the weight of the edge is set to the Euclidean distance between the two nodes. For each node in the power flow topology graph, count the number of all outgoing edges and the sum of the weights of all outgoing edges. Divide the weight of each outgoing edge of each node by the sum of the weights of the outgoing edges of the node to obtain the information distribution probability of each outgoing edge, forming the information distribution probability set of the node. Multiply each information probability value in the information probability set of each node by the negative logarithm of the information probability value, and sum all the product results to obtain the node structure entropy; The structural entropy of all nodes in the power flow topology graph is obtained by weighting the total outgoing edge weights of each node in proportion to the total outgoing edge weights of the entire power flow topology graph. The candidate scheduling paths are sorted according to their structural entropy values, and the candidate scheduling path with the smallest structural entropy value is selected as the optimal carbon emission reduction scheduling path.

[0013] Optionally, step seven specifically includes: In the low-dimensional scheduling feature space, all sample points in the optimal carbon emission reduction scheduling path are connected sequentially in time using coordinate curves to generate a visual trajectory map of the optimal carbon emission reduction scheduling path. The scheduling control variables corresponding to each sample point in the optimal carbon emission reduction scheduling path are extracted to generate a power scheduling instruction set. Each power scheduling instruction corresponds to a combination of control parameters within a scheduling cycle. The combination of control parameters includes power allocation parameters, voltage regulation parameters, and energy consumption limit parameters. Transmit the power dispatch instruction set to the power dispatch control system and set the effective time window.

[0014] The beneficial effects of this invention are: This invention constructs a power carbon emission impact map to comprehensively characterize the potential dependencies and influence strengths among scheduling control variables in multi-source power carbon emission data, effectively overcoming the limitations of traditional models in modeling variable relationships. Combined with principal component analysis, it significantly reduces the redundancy of high-dimensional data and improves the efficiency and expressive power of scheduling path feature extraction. In the process of scheduling path generation and selection, Latin hypercube sampling and Soft-DTW path distance metric methods are introduced, which can accurately match the dynamic feature differences of historical low-carbon paths while ensuring uniform sample coverage, thus improving the accuracy and robustness of path selection. By constructing power flow topology diagrams corresponding to candidate paths and using structural entropy as an evaluation index, the information organization efficiency of each scheduling path is measured from the perspective of structural complexity, enabling the final selected path to possess better global structural synergy and carbon emission reduction potential. The generated scheduling paths are not only visually intuitive and clear but can also automatically extract scheduling control parameters to form instruction sets, effectively supporting the rapid deployment and implementation of actual scheduling systems. The overall approach has the combined advantages of high modeling accuracy, reasonable path generation, strong result stability and good operability in actual deployment, which significantly improves the intelligence level and optimization effect of low-carbon dispatching strategies in complex power system environments. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0016] Figure 1 This is an overall flowchart of the multidimensional power carbon emission reduction path optimization method based on machine learning proposed in this invention. Figure 2 This is a schematic diagram illustrating the process of generating a set of candidate scheduling path vectors using Latin hypercube sampling in the multidimensional power carbon emission reduction path optimization method based on machine learning proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1-2 The multidimensional power carbon emission reduction path optimization method based on machine learning includes the following steps: Step 1: Collect and standardize multi-source power carbon emission data to generate a power carbon emission dataset, and determine historical low-carbon scheduling path vectors; Step 2: Construct an impact map of electricity carbon emissions based on the electricity carbon emissions dataset; Step 3: Construct an attribute vector matrix based on the power carbon emission impact map and perform principal component analysis to build a low-dimensional scheduling feature space; Step 4: In the low-dimensional scheduling feature space, Latin hypercube sampling is used to generate a set of candidate scheduling path vectors; Step 5: Calculate the Soft-DTW distance between the candidate scheduling path vector and the historical low-carbon scheduling path vector to obtain the set of candidate scheduling path vectors after filtering; Step 6: Construct the corresponding power flow topology based on the selected candidate scheduling path vector set, calculate the structural entropy value of each candidate scheduling path, and select the candidate scheduling path with the smallest structural entropy value as the optimal carbon emission reduction scheduling path. Step 7: Visualize the optimal carbon emission reduction scheduling path and generate the corresponding power dispatch instruction set.

[0019] This step effectively reduces computational complexity by constructing a power carbon emission impact map and compressing the scheduling dimension based on principal component analysis. It combines Latin hypercube sampling and Soft-DTW distance to achieve accurate matching between historical experience and candidate paths. Then, it evaluates the stability of power scheduling paths through structural entropy to ensure the low-carbon nature and controllability of path selection. Finally, it forms a visualized scheduling instruction set, which improves the scientific nature, visibility, and execution efficiency of carbon emission reduction strategies.

[0020] In this embodiment, the multi-source power carbon emission data includes generator set power generation data, renewable energy power generation data, real-time regional power grid load data, regional power purchase contract information and power purchase ratio data, fuel type data, historical dispatch strategies and corresponding carbon emission intensity data; the standardization step includes outlier removal, missing value imputation, time alignment and normalization processing for different types of data to generate a power carbon emission dataset with a unified structure; and based on the set carbon emission intensity threshold, the optimal historical low-carbon dispatch path is selected from the historical dispatch strategies to construct a historical low-carbon dispatch path vector.

[0021] This step integrates multi-source heterogeneous power data and performs standardization processing to ensure data quality and spatiotemporal consistency, thereby improving the fundamental reliability of data-driven scheduling optimization. At the same time, it introduces a carbon emission intensity threshold, selects representative low-carbon paths based on historical scheduling strategies, and establishes a historical low-carbon scheduling path vector. This provides a real and reliable reference benchmark for subsequent path matching and optimization, significantly improving the effectiveness and accuracy of carbon emission reduction path optimization.

[0022] In this embodiment, step two specifically includes: Each scheduling control variable in the power carbon emission data set is used as a graph node. The scheduling control variables include the power generation ratio of generator sets, the power generation ratio of renewable energy, the load allocation ratio between regions, the regional electricity purchase ratio, and the fuel consumption ratio. For each pair of scheduling control variables, the corresponding edge weight is calculated according to the variable type. The calculation steps are as follows: For any two continuous scheduling control variables, extract the corresponding time series data within the set time window, calculate the Pearson correlation coefficient of the two scheduling control variables within the corresponding time window, and use it as the edge weight between the corresponding scheduling control variable pairs in the power carbon emission impact map; For any two discrete scheduling control variables, the joint distribution probability in the same time window is statistically calculated, and the mutual information value of the joint distribution probability is used as the edge weight between the corresponding scheduling control variable pairs in the power carbon emission impact map. Connect all scheduling control variables with edge weights higher than a preset threshold in the power carbon emission impact map.

[0023] This step constructs a power carbon emission impact map, expressing multidimensional scheduling control variables in a graph structure. It uses Pearson correlation coefficients and mutual information values ​​to accurately characterize the correlation strength between variables, achieving unified modeling of continuous and discrete variable relationships. This effectively enhances the expressive ability of carbon emission relationships in complex scheduling systems, providing a structured foundation for subsequent feature extraction and path optimization, and helping to achieve more scientific and interpretable carbon emission reduction path design.

[0024] In this embodiment, step three specifically includes: The impact map of electricity carbon emissions is represented as an attribute vector matrix. Each row of the attribute vector matrix corresponds to a scheduling control variable node, and each column is the edge weight value between pairs of scheduling control variable nodes. Perform covariance matrix calculation on the attribute vector matrix to obtain the covariance matrix; Perform eigenvalue decomposition on the covariance matrix to obtain all eigenvalues ​​and their corresponding eigenvectors; All eigenvectors are sorted in descending order of eigenvalues, and the proportion of explained variance is calculated cumulatively. The proportion of explained variance is the ratio of the sum of the first few eigenvalues ​​to the sum of all eigenvalues. Select the eigenvectors corresponding to the first few eigenvalues ​​whose cumulative explained variance ratio is greater than a preset threshold, and use them as principal component bases. The principal component bases are the low-dimensional scheduling eigenvectors after dimensionality reduction. The low-dimensional scheduling feature space is constructed based on principal component basis.

[0025] This step transforms the power carbon emission impact map into an attribute vector matrix and performs principal component analysis. This significantly reduces the feature dimensionality while preserving the correlation characteristics between key scheduling variables, constructing a compact and discriminative low-dimensional scheduling feature space. This effectively alleviates redundancy and noise interference present in high-dimensional feature spaces, improves the efficiency of subsequent path sampling and matching, and provides a clearer feature representation basis for carbon emission reduction scheduling strategies.

[0026] In this embodiment, step four specifically includes: In the low-dimensional scheduling feature space, a set of candidate scheduling path vectors is generated using Latin hypercube sampling. The Latin hypercube sampling step includes: Each principal component dimension is divided into several equally spaced sub-intervals; In each dimension, a sample point is randomly selected from each of the sub-intervals to form a one-dimensional sampling vector; For a one-dimensional sampling vector of all dimensions, an index sequence with an equal number of sampling points is generated using a Sobol sequence, wherein the index sequence is a uniformly distributed permutation of integers without repetition. The one-dimensional sampling points of the corresponding dimension are rearranged according to the index order generated by the Sobol sequence; Each dimension's one-dimensional sampling vector is treated as an independent sampling axis, and different sampling axes are orthogonally combined. A sampling point is selected sequentially from each sampling axis to form a complete multidimensional sample point, and a sampling matrix is ​​constructed. The sequence of sample points in each row of the sampling matrix corresponds to a candidate scheduling path vector. Based on the sampling matrix, construct a set of candidate scheduling path vectors.

[0027] This step effectively ensures the uniformity of candidate scheduling path vector coverage and sample diversity across all principal component dimensions by introducing a strategy combining Latin hypercube sampling and Sobol sequences into the low-dimensional scheduling feature space. This enhances the representativeness of path sampling and spatial exploration capabilities. This method avoids the sample clustering or gapping issues that may occur with traditional random sampling, thereby improving the stability and optimization space of subsequent path matching and selection processes, and significantly improving the global optimality of carbon emission reduction scheduling path search.

[0028] In this embodiment, step five specifically includes: For each candidate scheduling path vector in the candidate scheduling path vector set, pair it with the historical low-carbon scheduling path vector and calculate the Soft-DTW distance. The specific steps for calculating the Soft-DTW distance include: Construct a two-dimensional cost matrix, where each element of the two-dimensional cost matrix represents the Euclidean distance between the corresponding dimension sample points of the candidate scheduling path vector and the historical low-carbon scheduling path vector at different time points; In the two-dimensional cost matrix, based on the Euclidean distance value, and following the minimum cumulative cost propagation rule, the cumulative cost value is updated sequentially from the first element in the upper left corner of the two-dimensional cost matrix to the lower right corner. The path with the minimum cumulative cost among all possible paths is selected as the optimal alignment path. The Euclidean distances between all paired sample points on the optimal alignment path are accumulated to obtain the Soft-DTW distance between the candidate scheduling path vector and the historical low-carbon scheduling path vector. Candidate scheduling path vectors whose Soft-DTW distance exceeds a preset threshold are removed from the candidate scheduling path vector set to obtain the filtered candidate scheduling path vector set.

[0029] This step introduces the Soft-DTW distance metric and combines it with minimum cumulative cost propagation to achieve flexible alignment between candidate scheduling paths and historical low-carbon paths, effectively addressing the issue of nonlinear differences in scheduling path time length and sample distribution. Compared to traditional Euclidean distance matching methods, Soft-DTW can more accurately capture path change trends and overall structural similarity, improving the robustness and reliability of path selection, thereby ensuring that paths entering the next step of structural entropy analysis have high representativeness and carbon reduction potential.

[0030] In this embodiment, step six specifically includes: Each candidate scheduling path vector in the candidate scheduling path vector set is treated as a candidate scheduling path composed of sample points. Based on each candidate scheduling path, a corresponding power flow topology graph is constructed. The specific construction steps are as follows: Each candidate scheduling path contains sample points as nodes in the power flow topology graph, and each node has corresponding coordinates in the low-dimensional scheduling feature space. Traverse all nodes and calculate the Euclidean distance between any two nodes in the low-dimensional scheduling feature space. If the Euclidean distance between two nodes is less than a preset connection threshold, an edge connection is established between the two corresponding nodes, and the weight of the edge is set to the Euclidean distance between the two nodes. For each node in the power flow topology graph, count the number of all outgoing edges and the sum of the weights of all outgoing edges. Divide the weight of each outgoing edge of each node by the sum of the weights of the outgoing edges of the node to obtain the information distribution probability of each outgoing edge, forming the information distribution probability set of the node. Multiply each information probability value in the information probability set of each node by the negative logarithm of the information probability value, and sum all the product results to obtain the node structure entropy; The structural entropy of all nodes in the power flow topology graph is obtained by weighting the total outgoing edge weights of each node in proportion to the total outgoing edge weights of the entire power flow topology graph. The candidate scheduling paths are sorted according to their structural entropy values, and the candidate scheduling path with the smallest structural entropy value is selected as the optimal carbon emission reduction scheduling path.

[0031] This step constructs a power flow topology graph from candidate scheduling path vectors and introduces a structural entropy index to quantify the connection complexity and information uncertainty between nodes in the path, effectively measuring the stability and controllability of the path within the system. Compared to traditional methods that select paths based on a single indicator such as energy consumption or carbon emissions, the structural entropy index can take into account both the spatial continuity of the scheduling path and the efficiency of information transmission, thereby selecting the optimal carbon emission reduction scheduling path under system constraints and improving the overall system performance and interpretability of path planning.

[0032] In this embodiment, step seven specifically includes: In the low-dimensional scheduling feature space, all sample points in the optimal carbon emission reduction scheduling path are connected sequentially in time using coordinate curves to generate a visual trajectory map of the optimal carbon emission reduction scheduling path. The scheduling control variables corresponding to each sample point in the optimal carbon emission reduction scheduling path are extracted to generate a power scheduling instruction set. Each power scheduling instruction corresponds to a combination of control parameters within a scheduling cycle. The combination of control parameters includes power allocation parameters, voltage regulation parameters, and energy consumption limit parameters. Transmit the power dispatch instruction set to the power dispatch control system and set the effective time window.

[0033] This step visualizes the optimal carbon emission reduction scheduling path in a low-dimensional scheduling feature space using coordinates, intuitively showcasing the evolution trend and characteristics of the carbon emission reduction path, thus improving the interpretability of the scheduling strategy. Simultaneously, each sample point in the path is transformed into a specific, executable set of power dispatching instructions, covering key parameters such as power allocation, voltage regulation, and energy consumption control. This ensures that the scheduling strategy can be seamlessly deployed to the dispatching control system and executed on time, achieving a closed-loop process from optimization modeling to actual deployment, thereby improving the implementation capability and execution efficiency of carbon emission reduction scheduling.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to a dispatch optimization project for a regional power grid. This regional power grid comprises four main substations, twelve medium-sized thermal power units, three wind farms, and one photovoltaic base, exhibiting a typical multi-source heterogeneous power system structure. Traditional power carbon emission management primarily relies on the experience of dispatchers for static strategy configuration, lacking in-depth modeling of the dynamic impact mechanisms of carbon emissions. This makes it difficult to sustainably reduce carbon emission intensity, especially as the proportion of renewable energy integration increases, leading to a rapid rise in the complexity of dispatch paths.

[0035] In this invention, the method first collects 180 consecutive days of generator operation data, wind and solar power output data, regional electricity purchase ratio, load curve, fuel consumption data, and historical dispatch strategies within the region, with a uniform time granularity of one sampling point every 15 minutes. After data cleaning and standardization, a power carbon emission dataset containing 172,800 records is constructed, and a carbon emission intensity threshold of 0.45 kg / kWh is set. From this dataset, 312 optimal historical low-carbon dispatch paths are selected as control samples. Subsequently, a power carbon emission impact map is constructed, which includes 28 dispatch control variable nodes. The edge weight calculation adopts a hybrid modeling method using Pearson correlation coefficient and mutual information value, identifying 143 effective edges.

[0036] Based on the power flow topology map, a 28-dimensional attribute vector matrix is ​​generated, and principal component analysis is performed. The top 5 principal components retained when the cumulative variance reaches 93.7% are used as the basis for the low-dimensional scheduling feature space. In this space, a total of 1000 candidate scheduling path vectors are generated through Latin hypercube sampling and Sobol sequences. Soft-DTW distance is calculated between each vector and historical paths, and 327 candidate paths with high similarity are retained based on a distance threshold (set to 20). Then, a power flow topology map of these 327 paths is constructed, and the path with the lowest structural entropy (0.134) is selected as the final scheduling path after calculating the structural entropy value.

[0037] Ultimately, this path was converted into a power dispatch instruction set and put into operation in a simulation environment for 30 days for A / B testing. Compared with traditional dispatch strategies, the method of this invention significantly reduces carbon emission intensity and carbon emissions per unit of electricity while maintaining system load balance and voltage stability.

[0038] Table 1. Comparative Analysis of Carbon Reduction Effect and Scheduling Performance of Different Path Optimization Methods

[0039] As shown in Table 1, the method of this invention is comprehensively superior to traditional dispatching methods in terms of carbon emission indicators, reducing the average carbon emission intensity by about 13% and the carbon emission per unit of electricity by about 61 t / GWh. Simultaneously, it maintains voltage fluctuations within ±5V, and the system load deviation rate and dispatch command response delay are comparable to or even better than traditional methods. Particularly in terms of renewable energy absorption capacity, this method achieves a utilization rate exceeding 84%, significantly improving the utilization efficiency of wind and solar power output, demonstrating a good dispatching optimization capability that balances environmental protection and economic efficiency. This embodiment shows that the present invention can effectively reduce carbon emissions, improve the utilization rate of new energy sources, and enhance system stability in actual power systems, and has extremely strong practical application value.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for multi-dimensional power carbon abatement pathway optimization based on machine learning, characterized in that, The method comprises the following steps: Step 1: collecting and standardizing multi-source power carbon emission data to generate a power carbon emission dataset, and determining a historical low-carbon scheduling path vector; Step 2: constructing a power carbon emission influence graph based on the power carbon emission dataset; Step 3: constructing an attribute vector matrix based on the power carbon emission influence graph, and performing principal component analysis to construct a low-dimensional scheduling feature space; Step 4: in the low-dimensional scheduling feature space, Latin hypercube sampling is used to generate a candidate scheduling path vector set; Step 5: calculating the Soft-DTW distance between the candidate scheduling path vector and the historical low-carbon scheduling path vector to obtain a screened candidate scheduling path vector set; Step 6: constructing a corresponding power flow topology graph according to the screened candidate scheduling path vector set, calculating the structural entropy value of each candidate scheduling path, and selecting the candidate scheduling path with the minimum structural entropy value as the optimal carbon emission reduction scheduling path; Step 7: visualizing the optimal carbon emission reduction scheduling path and generating a corresponding power scheduling instruction set. 2.The method of claim 1, wherein, The multi-source power carbon emission data includes power generator data, renewable energy power generation data, regional power grid load real-time data, regional power purchase contract information and power purchase proportion data, fuel type data, historical scheduling strategy and corresponding carbon emission intensity data; the standardization step includes outlier rejection, missing value filling, time alignment and normalization processing for different types of data to generate a unified structure power carbon emission dataset; and according to the set carbon emission intensity threshold, the optimal historical low-carbon scheduling path is selected from the historical scheduling strategy to construct the historical low-carbon scheduling path vector. 3.The method of claim 1, wherein, The step 2 is specifically: Each scheduling control variable in the power carbon emission dataset is taken as a graph node, and the scheduling control variable includes a power generator proportion, a renewable energy power generation proportion, a regional load distribution proportion, a regional power purchase proportion and a fuel consumption proportion; For each pair of scheduling control variables, the corresponding edge weight is calculated according to the variable type, and the calculation step is: For any two continuous scheduling control variables, the corresponding time series data in the set time window is extracted, the Pearson correlation coefficient of the two scheduling control variables in the corresponding time window is calculated, and the Pearson correlation coefficient is taken as the edge weight between the corresponding scheduling control variable pair in the power carbon emission influence graph; For any two discrete scheduling control variables, the joint distribution probability in the same time window is respectively calculated, and the mutual information value of the joint distribution probability is calculated as the edge weight between the corresponding scheduling control variable pair in the power carbon emission influence graph; The scheduling control variable pairs with all edge weights higher than the preset threshold are connected in the power carbon emission influence graph.

4. The machine learning based multi-dimensional electricity carbon emission reduction path optimization method according to claim 1, characterized in that, The step 3 is specifically: The power carbon emission influence graph is represented as an attribute vector matrix, each row of the attribute vector matrix corresponds to a scheduling control variable node, and each column is an edge weight value between a pair of scheduling control variable nodes; The covariance matrix calculation is performed on the attribute vector matrix to obtain a covariance matrix; The eigenvalue decomposition is performed on the covariance matrix to obtain all eigenvalues and corresponding eigenvectors; Sort all eigenvectors in descending order of eigenvalues, and cumulatively calculate an explained variance ratio, which is a ratio of a sum of the first several eigenvalues to a sum of all eigenvalues; Select eigenvectors corresponding to the first several eigenvalues with a cumulative explained variance ratio greater than a preset threshold as principal component bases, which are low-dimensional scheduling eigenvectors after dimension reduction; Construct the low-dimensional scheduling eigenspace based on the principal component bases. 5.The method of claim 1, wherein, The fourth step specifically includes: In the low-dimensional scheduling eigenspace, generate a candidate scheduling path vector set by Latin hypercube sampling, which includes the following steps: Divide each principal component dimension into several equally spaced subintervals; In each dimension, randomly select a sample point from each subinterval to form a one-dimensional sampling vector; Generate an index sequence with equal sample point numbers by using a Sobol sequence on one-dimensional sampling vectors of all dimensions, which is a non-repeated integer arrangement with uniform distribution properties; According to the index sequence generated by the Sobol sequence, rearrange one-dimensional sampling points of the corresponding dimension; Take each one-dimensional sampling vector of each dimension as a sampling axis of an independent dimension, and orthogonally combine different sampling axes; Select a sample point from each sampling axis in turn to form a complete multi-dimensional sample point, and construct a sampling matrix, in which a sample point sequence of each row corresponds to a candidate scheduling path vector; Construct the candidate scheduling path vector set according to the sampling matrix.

6. The machine learning based multi-dimensional electricity carbon abatement pathway optimization method according to claim 1, wherein, The fifth step specifically includes: Pair each candidate scheduling path vector in the candidate scheduling path vector set with a historical low-carbon scheduling path vector and calculate a Soft-DTW distance, which includes the following steps: Construct a two-dimensional cost matrix, each element of which represents the Euclidean distance between the corresponding dimension sample points of the candidate scheduling path vector and the historical low-carbon scheduling path vector at different time points; In the two-dimensional cost matrix, based on the Euclidean distance value, update the cumulative cost value from the first element in the top left corner of the two-dimensional cost matrix as the starting point to the lower right corner in turn according to the minimum cumulative cost propagation rule; Select the path with the minimum cumulative cost value from all possible paths as the optimal alignment path; Add up the Euclidean distances between all paired sample points on the optimal alignment path to obtain the Soft-DTW distance between the candidate scheduling path vector and the historical low-carbon scheduling path vector; Remove the candidate scheduling path vector with a Soft-DTW distance exceeding a preset threshold from the candidate scheduling path vector set to obtain a screened candidate scheduling path vector set.

7. The machine learning based multi-dimensional electricity carbon abatement pathway optimization method according to claim 1, characterized in that, The sixth step specifically includes: Take each candidate scheduling path vector in the candidate scheduling path vector set as a candidate scheduling path composed of sample points; Construct a corresponding power flow topology graph based on each candidate scheduling path, which includes the following steps: Take the sample points contained in each candidate scheduling path as nodes of the power flow topology graph, and each node has a corresponding coordinate in the low-dimensional scheduling eigenspace; Traverse all nodes, calculate the Euclidean distance between any two nodes in the low-dimensional scheduling feature space; If the Euclidean distance between the two nodes is less than the preset connection threshold, an edge connection is established between the corresponding two nodes, and the weight of the edge is set to the Euclidean distance between the two nodes; For each node in the power flow topology graph, count the number of all outgoing edges and the sum of the weights of all outgoing edges; Divide the edge weight of each outgoing edge of each node by the total weight of the node outgoing edges to obtain the information distribution probability of each outgoing edge, forming a set of information distribution probabilities of the node; Multiply each information distribution probability value in the information distribution probability set of each node by the negative logarithm of the information distribution probability value, and sum all the products to obtain the node structure entropy; For all node structure entropies in the power flow topology graph, weighted average according to the proportion of the total weight of the outgoing edges of each node in the total weight of the outgoing edges of the entire power flow topology graph to obtain the structure entropy value of the entire power flow topology graph; Sort different candidate scheduling paths according to the structure entropy value, and select the candidate scheduling path with the smallest structure entropy value as the optimal carbon emission reduction scheduling path.

8. The machine learning based multi-dimensional electricity carbon abatement pathway optimization method of claim 1, wherein, The step seven is specifically: In the low-dimensional scheduling feature space, all sample points in the optimal carbon emission reduction scheduling path are connected in time sequence in the form of coordinate curve to generate a visual trajectory graph of the optimal carbon emission reduction scheduling path; Extract the scheduling control variables corresponding to each sample point in the optimal carbon emission reduction scheduling path to generate a power scheduling instruction set, each power scheduling instruction corresponding to a control parameter combination in a scheduling period, the control parameter combination including power allocation parameters, voltage adjustment parameters and energy consumption limit parameters; The power scheduling instruction set is transmitted to the power scheduling control system and the effective time window is set.

Images