A multi-objective pareto optimal scheduling method and system for a new power system

Abstract

The application provides a multi-objective Pareto optimal scheduling method and system for a new power system, and relates to the technical field of automatic scheduling of power systems.The method comprises the following steps: collecting operation data of the new power system, analyzing the operation data by using a pre-trained time sequence feature extraction network, determining a source and load power probability distribution scenario set and an initial feasible solution space; initializing a candidate strategy population in the initial feasible solution space, calculating mutual information entropy values between multi-dimensional objective function gradient vectors of each candidate strategy, and synthesizing a comprehensive guide potential field vector pointing to a Pareto improvement direction according to the mutual information entropy values.The application improves the global optimality and convergence stability of the new power system scheduling through multi-objective collaborative optimization and adaptive evolutionary search, realizes efficient, accurate and robust multi-objective Pareto optimal scheduling, and balances the global exploration and local optimization capabilities.

Description

Technical Field

[0001] This invention relates to the field of power system automated dispatching technology, and in particular to a multi-objective Pareto optimal dispatching method and system for new power systems. Background Technology

[0002] In new power systems, the output of new energy sources such as wind and solar power exhibits significant randomness and volatility, while load-side response behavior also displays multi-scale variation characteristics. This leads to rapid non-stationary changes in the source-load power probability distribution during intraday rolling scheduling. Existing multi-objective Pareto optimal scheduling methods (such as decomposition-based MOEA / D and non-dominated sorting genetic algorithms with elitist strategies) typically employ relatively fixed population evolution strategies. That is, during iteration, crossover, mutation, and environmental selection mechanisms rely on preset evolutionary parameters or static congestion distance metrics. When the dispersion of the source-load probability distribution scenario set changes within the scheduling period (e.g., from a stable period to a period of severe fluctuation), these methods may suffer from insufficient flexibility in coordinating exploration capabilities and convergence speed.

[0003] Specifically, some existing methods tend to explore globally in the early stages of evolution, but the search range shrinks rapidly in the later stages. If the probability distribution scenario set exhibits new high-dispersion characteristics in the later stages (such as sudden and significant fluctuations in the output of new energy sources), the shrunken population may struggle to re-cover potential new non-dominated regions, thus affecting the integrity and approximation efficiency of the Pareto front to some extent. Conversely, if a strong exploration intensity is maintained throughout, the convergence speed in stable scenarios may not be ideal.

[0004] For example, a coastal industrial park's power distribution network often experiences rapidly developing localized convective weather in the summer afternoons. On a certain scheduling day, the source-load power probability distribution is relatively stable for the first two hours (solar output is sufficient and fluctuations are mild), and the system is scheduled with economic efficiency and low carbon emissions as the main objectives. At this time, existing methods can quickly converge to a relatively optimal set of non-dominated solutions. However, convective clouds then cover the park within 15 minutes, causing solar output to plummet from 75% of rated power to 20%, while air conditioning load increases by 30% due to a sharp rise in temperature. The dispersion index of the source-load probability distribution scenario set increases significantly. At this transition moment, the population of scheduling methods using a fixed evolution strategy may have already clustered near the Pareto front of the previous period, and the variable asynchronous length and cross-distribution index fail to adapt to the new uncertainty intensity in time. This makes it difficult to effectively explore new solution regions containing combinations such as rapid energy storage discharge and gas turbine ramp-up in subsequent iterations, and some non-dominated solutions may be missed. Summary of the Invention

[0005] This invention provides a multi-objective Pareto optimal scheduling method and system for novel power systems, which solves the problem that fixed evolution strategies are difficult to balance exploration capability and convergence speed, and enhances the algorithm's adaptability to scenarios with strong source-load fluctuations.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a multi-objective Pareto optimal scheduling method for novel power systems is provided, the method comprising: Step 1: Collect operational data of the new power system, analyze the operational data using a pre-trained time series feature extraction network, and determine the source-load power probability distribution scenario set and the initial feasible solution space; Step 2: Initialize the candidate policy population in the initial feasible solution space, calculate the mutual information entropy value between the gradient vectors of the multidimensional objective function of each candidate policy, and synthesize the comprehensive guiding potential vector pointing to the Pareto improvement direction based on the mutual information entropy value. Step 3: Extract the first evolution anchor point and the second evolution anchor point from the initial feasible solution space. In each rolling scheduling cycle, regenerate the source load power probability distribution scenario set based on the latest observation data and forecast data, and calculate the dispersion index of the scenario set in real time. Based on the comparison result of the dispersion index and the preset threshold, select the corresponding evolution anchor point to construct a local convergent hyperellipsoidal domain or a global exploratory hyperellipsoidal domain. Step 4: Map the candidate strategy population to the hyperellipsoidal domain, calculate the Mahalanobis distance between each candidate strategy and the geometric center of the hyperellipsoidal domain, divide the population into subsets based on the Mahalanobis distance, and perform differential correction and search to obtain the hybrid evolutionary strategy set. Step 5: The historical iteration velocity memory term is superimposed on the integrated guiding potential field vector as an inertial component to calculate the state update vector, so as to drive the hybrid evolution strategy set to perform trajectory evolution in the initial feasible solution space and obtain the evolved non-dominated candidate strategy set. Step 6: Perform voxel distance transformation and non-dominated sorting on the evolved non-dominated candidate strategy set to obtain the Pareto scheduling strategy library. Match the current power grid operating state with the matching degree of the candidate strategies in the Pareto scheduling strategy library, select the final strategy and send it to the control terminal.

[0007] Secondly, a multi-objective Pareto optimal scheduling system for new power systems includes: The analysis module is used to collect operational data of the new power system, analyze the operational data using a pre-trained time series feature extraction network, and determine the source-load power probability distribution scenario set and the initial feasible solution space. The computation module is used to initialize the candidate policy population in the initial feasible solution space, calculate the mutual information entropy value between the gradient vectors of the multidimensional objective function of each candidate policy, and synthesize the comprehensive guiding potential vector pointing to the Pareto improvement direction based on the mutual information entropy value. The module is used to extract the first and second evolution anchor points from the initial feasible solution space. In each rolling scheduling cycle, it regenerates the source load power probability distribution scenario set based on the latest observation data and forecast data, and calculates the dispersion index of the scenario set in real time. Based on the comparison result of the dispersion index and the preset threshold, the corresponding evolution anchor point is selected to construct a local convergent hyperellipsoidal domain or a global exploratory hyperellipsoidal domain. The correction module is used to map the candidate strategy population to the hyperellipsoidal domain, calculate the Mahalanobis distance between each candidate strategy and the geometric center of the hyperellipsoidal domain, divide the population into subsets based on the Mahalanobis distance, and perform differential correction and search to obtain the hybrid evolutionary strategy set. The evolution module is used to superimpose the historical iteration velocity memory term as an inertial component with the comprehensive guiding potential field vector to calculate the state update vector, so as to drive the hybrid evolution strategy set to perform trajectory evolution in the initial feasible solution space and obtain the evolved non-dominated candidate strategy set. The control module performs voxel distance transformation and non-dominated sorting on the evolved non-dominated candidate strategy set to obtain the Pareto scheduling strategy library. It then matches the current power grid operating state with the candidate strategies in the Pareto scheduling strategy library and selects the final strategy to send to the control terminal.

[0008] The above-described solution of the present invention has at least the following beneficial effects: This invention utilizes a pre-trained temporal feature extraction network to perform spatiotemporal joint analysis of operational data from novel power systems. This enables the characterization of the uncertain distribution patterns of source-load power, reasonable definition of the initial feasible solution space, and effective elimination of ineffective strategy points, thereby improving the feasibility and security of scheduling strategies. A comprehensive guiding potential vector is constructed based on the mutual information entropy values ​​between multi-dimensional objective functions. This dynamically quantifies the coupling relationships and conflict intensity between objectives, ensuring that the population iteration direction always aligns with the Pareto improvement trend. Compared to fixed-weight guiding methods, this improves the convergence accuracy and approximation effect of non-dominated solutions. The evolution anchor point is dynamically switched according to the discreteness of the source-load power probability distribution scenario, and a locally convergent hyperellipsoidal domain or a globally exploratory hyperellipsoidal domain is adaptively constructed. This achieves flexible adaptation between rapid convergence under stable operating conditions and global search under drastic disturbance conditions, solving the problem that fixed evolution strategies cannot simultaneously balance exploration capability and convergence speed, and enhancing the algorithm's adaptability to scenarios with strong source-load fluctuations.

[0009] By employing Mahalanobis distance to partition the population into subsets and implementing differentiated search, the local development capabilities of the core region can be enhanced while maintaining population diversity in the peripheral region, avoiding premature convergence and improving the sufficiency of the solution space search. Combining historical iteration velocity memory terms with a comprehensive guiding potential vector to realize population trajectory evolution balances iterative inertia and goal-oriented advantages, optimizing the population update mechanism and further improving optimization efficiency and the quality of non-dominated solution sets. A Pareto scheduling strategy library is constructed by combining voxel distance transformation with non-dominated sorting, improving the uniformity and completeness of solution set distribution in high-dimensional target spaces and avoiding the loss of key scheduling strategies. Simultaneously, online strategy optimization based on operating state matching can quickly output scheduling instructions adapted to the current operating conditions, improving the response speed of real-time scheduling in new power systems. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a multi-objective Pareto optimal scheduling method for novel power systems provided by an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a multi-objective Pareto optimal scheduling system for a new type of power system provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown, an embodiment of the present invention proposes a multi-objective Pareto optimal scheduling method for novel power systems, the method comprising the following steps: Step 1: Collect operational data of the new power system, analyze the operational data using a pre-trained time series feature extraction network, and determine the source-load power probability distribution scenario set and the initial feasible solution space; Step 2: Initialize the candidate policy population in the initial feasible solution space, calculate the mutual information entropy value between the gradient vectors of the multidimensional objective function of each candidate policy, and synthesize the comprehensive guiding potential vector pointing to the Pareto improvement direction based on the mutual information entropy value. Step 3: Extract the first evolution anchor point and the second evolution anchor point from the initial feasible solution space. In each rolling scheduling cycle, regenerate the source load power probability distribution scenario set based on the latest observation data and forecast data, and calculate the dispersion index of the scenario set in real time. Based on the comparison result of the dispersion index and the preset threshold, select the corresponding evolution anchor point to construct a local convergent hyperellipsoidal domain or a global exploratory hyperellipsoidal domain. Step 4: Map the candidate strategy population to the hyperellipsoidal domain, calculate the Mahalanobis distance between each candidate strategy and the geometric center of the hyperellipsoidal domain, divide the population into subsets based on the Mahalanobis distance, and perform differential correction and search to obtain the hybrid evolutionary strategy set. Step 5: The historical iteration velocity memory term is superimposed on the integrated guiding potential field vector as an inertial component to calculate the state update vector, so as to drive the hybrid evolution strategy set to perform trajectory evolution in the initial feasible solution space and obtain the evolved non-dominated candidate strategy set. Step 6: Perform voxel distance transformation and non-dominated sorting on the evolved non-dominated candidate strategy set to obtain the Pareto scheduling strategy library. Match the current power grid operating state with the matching degree of the candidate strategies in the Pareto scheduling strategy library, select the final strategy and send it to the control terminal.

[0014] In this embodiment of the invention, a pre-trained temporal feature extraction network is used to perform spatiotemporal joint analysis on the operating data of a novel power system. This enables the characterization of the uncertainty distribution of source-load power, reasonable definition of the initial feasible solution space, effective elimination of invalid strategy points, and improvement of the feasibility and security of the scheduling strategy. A comprehensive guiding potential vector is constructed based on the mutual information entropy values ​​between multi-dimensional objective functions. This dynamically quantifies the coupling relationship and conflict intensity between objectives, ensuring that the population iteration direction always aligns with the Pareto improvement trend. Compared to fixed-weight guidance methods, this improves the convergence accuracy and approximation effect of non-dominated solutions. The evolution anchor point is dynamically switched according to the discreteness of the source-load power probability distribution scenario, and a local convergence hyperellipsoidal domain or a global exploration hyperellipsoidal domain is adaptively constructed. This achieves flexible adaptation between rapid convergence under stable operating conditions and global search under drastic disturbance conditions, solving the problem that fixed evolution strategies cannot simultaneously consider exploration capability and convergence speed, and enhancing the algorithm's adaptability to scenarios with strong source-load fluctuations.

[0015] By employing Mahalanobis distance to partition the population into subsets and implementing differentiated search, the local development capabilities of the core region can be enhanced while maintaining population diversity in the peripheral region, avoiding premature convergence and improving the sufficiency of the solution space search. Combining historical iteration velocity memory terms with a comprehensive guiding potential vector to realize population trajectory evolution balances iterative inertia and goal-oriented advantages, optimizing the population update mechanism and further improving optimization efficiency and the quality of non-dominated solution sets. A Pareto scheduling strategy library is constructed by combining voxel distance transformation with non-dominated sorting, improving the uniformity and completeness of solution set distribution in high-dimensional target spaces and avoiding the loss of key scheduling strategies. Simultaneously, online strategy optimization based on operating state matching can quickly output scheduling instructions adapted to the current operating conditions, improving the response speed of real-time scheduling in new power systems.

[0016] In a preferred embodiment of the present invention, step 1 includes: Step 100a: Collect operational data of the new power system. This operational data includes historical source-load power time-series data, meteorological environmental data, power grid topology parameters, unit ramp rate constraint data, node voltage safety boundary data, and equipment operating status data. Specifically, it includes: Historical source-load power time-series data covers the output time-series data of new energy sources such as wind power and photovoltaics, the output time-series data of conventional units, and the time-series data of various loads (industrial load, residential load, and commercial load); meteorological environmental data covers environmental parameters related to new energy output such as wind speed, light intensity, temperature, and humidity; power grid topology parameters cover the number of nodes, branch connection relationships, branch impedance, and transformer parameters; unit ramp rate constraint data covers the maximum and minimum ramp rates of each conventional unit and new energy unit, where the maximum ramp rate of conventional units is 1.5% to 3.5% of rated capacity / min, and the minimum ramp rate is -3.0% to -1.0% of rated capacity / min; new energy units... The maximum ramp rate of the generating unit is 2.0% to 5.0% of the rated capacity / min, and the minimum ramp rate is -4.0% to -1.5% of the rated capacity / min. The node voltage safety boundary data covers the upper and lower voltage limits of each grid node. The upper voltage limit of high-voltage nodes is 1.05 to 1.10 times the rated voltage, and the lower voltage limit is 0.90 to 0.95 times the rated voltage. The upper voltage limit of low-voltage nodes is 1.07 to 1.12 times the rated voltage, and the lower voltage limit is 0.88 to 0.93 times the rated voltage. The equipment operation status data covers the operation status (normal, abnormal, maintenance) and operating parameters of key equipment such as transformers, circuit breakers, and transmission lines.

[0017] The collected raw data underwent outlier removal, missing value imputation, and time series alignment. Outlier removal employed the 3σ criterion: for a given data sequence, its mean and standard deviation were calculated; if a data point satisfies |data point - mean| > 3 times the standard deviation, it was identified as an outlier and replaced using linear interpolation. Missing value imputation used adjacent-data linear interpolation; for a missing data point, the imputation result was calculated using the valid data points before and after it. The calculation formula is as follows: (in the formula) For time series indexes of missing data points, For missing data points, , (These are the valid data points before and after the data points); time series alignment uses a uniform time step (e.g., 15 minutes / step) to resample all time series data, ensuring that all types of data remain consistent in the time dimension, ultimately forming a standardized dataset under a unified time series scale.

[0018] Step 100b involves constructing a pre-trained temporal feature extraction network that includes a multi-head self-attention mechanism module, a graph convolutional neural network module, and a conditional generative adversarial network module. Specifically, this includes: A pre-trained temporal feature extraction network was constructed for temporal feature extraction. This network adopts a three-module cascaded structure: a multi-head self-attention mechanism module, a graph convolutional neural network module, and a conditional generative adversarial network module. Among them, the multi-head self-attention mechanism module is used to extract the long-period dependency features of the time dimension of historical source-load power temporal data, capturing the inherent pattern of source-load power changes over time. Its specific construction parameters and internal structure are as follows: the number of attention heads is set to 4 to 8, the input dimension of the module is completely consistent with the feature dimension of the historical source-load power temporal data, and the output dimension is fixed. Set to 64 to 128; This module contains three core layers with clearly defined functions and working together: the linear transformation layer, the attention calculation layer, and the feature aggregation layer. The linear transformation layer maps the input source-load power time-series data Xt (Xt is the time-series data matrix) to a query vector Q, a key vector K, and a value vector V through three independent linear transformation operations. The corresponding calculation formulas are Q = Xt × WQ, K = Xt × WK, and V = Xt × WV (where WQ, WK, and WV are the independent linear transformation weight matrices of the three types of vectors, and their values ​​all follow a normal distribution). The matrix is ​​randomly initialized, with dimensions equal to the input time-series data feature dimension multiplied by the dimensions of the query vector and key vector, and the dimensions of the query vector and key vector are... The multi-head self-attention mechanism module outputs a dimension (number of attention heads), and the three types of vectors, Q, K, and V, are independent of each other, corresponding to different feature dimensions of the time series data. The attention calculation layer takes the three types of vectors output by the linear transformation layer as input, calculates the similarity between the query vector and the key vector, and then uses the Softmax function to normalize the similarity to obtain the attention weight distribution between each time series point. This quantifies the degree of correlation between features at different time series points and highlights the importance of key time series features. The feature aggregation layer then performs a weighted summation operation on the value vectors based on this attention weight distribution, fuses the features of different time series points, and finally outputs a fused time series feature that can characterize the long-term periodic dependence of source-load power.

[0019] The graph convolutional neural network module is constructed to extract spatial topological correlation features between equipment operating status data and power grid topology parameters. The specific construction parameters and structure are as follows: the number of network layers is set to 3 to 5; the input dimension is consistent with the fusion feature dimension of the equipment operating status data and power grid topology parameters; the hidden layer dimension is 64 to 128; and the output dimension is consistent with the output dimension of the multi-head self-attention mechanism module. Internally, the module constructs a power grid topology adjacency matrix with power grid nodes as vertices and branch connections as edges. Adjacency matrix elements Defined as, if node With nodes In branch connections, then ,otherwise The aggregation of features between adjacent nodes is achieved through graph convolution operations. The formula for graph convolution operations is as follows: Normalized adjacency matrix The calculation formula is: (in the formula) For the input feature matrix, This is the weight matrix. For bias terms, For degree matrix, (identity matrix); degree matrix The element is defined as (in the formula) (Number of grid nodes).

[0020] The conditional generative adversarial network (GAN) module is constructed to generate a set of source-load power probability distribution scenarios containing confidence intervals under spatiotemporal fusion feature constraints. It consists of a generator and a discriminator. The generator employs a fully connected network structure, with input dimensions equal to the dimensions of the spatiotemporal fusion feature vector (i.e., the sum of the dimensions of the temporal and spatial hidden state sequences), and output dimensions equal to the feature dimensions of the source-load power scenario. The number of hidden layers is set to 4 to 6, with 128 to 256 neurons per hidden layer. The activation function used is LeakyReLU (with a slope of 0.2). The output layer uses the tanh activation function to output the source load power prediction scene; the discriminator also uses a fully connected network structure, with the input dimension being the sum of the feature dimension of the source load power scene and the dimension of the spatiotemporal fusion feature vector, and the output dimension being 1 (used to output the discrimination probability of scene authenticity); the number of hidden layers is set to 3 to 5, the number of hidden layer neurons is set to 128 to 256, the activation function is LeakyReLU (slope of 0.2), and the output layer uses the sigmoid activation function to distinguish between real historical scenes and generated prediction scenes.

[0021] Using the historical operational dataset preprocessed in step 100a, a two-stage pre-training process combining supervised and unsupervised methods was performed. In the first stage, the spatiotemporal feature extraction sub-network (multi-head self-attention mechanism module + graph convolutional neural network module) was trained. The supervised training tasks included time-series prediction of historical source-load power data and classification of power grid node features, constructing a training loss function. An Adam optimizer was used (learning rate set to 1e⁻⁴ to 1e⁻³, decay coefficients set to 0.9 and 0.999), iteratively training until the loss function converged (convergence condition: the change in loss function is less than 1e⁻⁵ in 100 consecutive iterations), completing the parameter initialization and training of the spatiotemporal feature extraction sub-network. In the second stage, the entire network was jointly trained (adding a conditional generative adversarial network module). The parameters of the spatiotemporal feature extraction sub-network trained in the first stage were fixed, and adversarial training based on scene generation was used as the core task, constructing a generator loss function. and discriminator loss function (in the formula) The number of scene samples. It is a random noise vector. For spatiotemporal fusion feature vectors, The predicted scene output by the generator. The discriminator outputs the probability. For real historical scenes, (This refers to the scene sample traversal index); the Adam optimizer (learning rate set to 1e-4, decay coefficients set to 0.9 and 0.999) is used to iteratively train the generator and discriminator until the adversarial training converges, thus completing the construction and training of the entire pre-trained temporal feature extraction network.

[0022] Step 101b: Input historical source-load power time-series data into the multi-head self-attention mechanism module. By calculating the dot product similarity of the query vector, key vector, and value vector, the long-period dependency features of the historical source-load power time-series data in the time dimension are obtained, resulting in a time-hidden state sequence. Input equipment operating status data and grid topology parameters into the graph convolutional neural network module. By aggregating the feature information of adjacent nodes, the spatial correlation features of the equipment operating status data in the spatial topology are obtained, resulting in a spatial hidden state sequence, specifically including: The preprocessed historical source-load power time-series data, equipment operating status data, and power grid topology parameters from step 100a are input into the corresponding modules of the pre-trained time-series feature extraction network to complete the extraction of temporal and spatial hidden state sequences. The temporal hidden state sequence extraction (based on a multi-head self-attention mechanism module) involves inputting the preprocessed historical source-load power time-series data into the multi-head self-attention mechanism module. Through three independent linear transformation layers, the input time-series data is mapped into query vectors, key vectors, and value vectors, respectively. The dot product similarity between the query vector and the key vector under each attention head is calculated. The dot product similarity is normalized using the Softmax function to obtain the attention weight distribution. The value vectors are then weighted and summed using the attention weights to obtain the feature output of a single attention head. The feature outputs of all attention heads are concatenated and fused to obtain the final temporal hidden state sequence.

[0023] Spatial hidden state sequence extraction (based on a graph convolutional neural network module) involves fusing preprocessed equipment operating state data with power grid topology parameters to obtain an input feature matrix, which is then fed into the graph convolutional neural network module. Based on the power grid topology connections, an adjacency matrix is ​​constructed and normalized to ensure the stability of the graph convolution operation. Through multiple layers of stacked graph convolution operations, the feature information of adjacent nodes is aggregated layer by layer. After a preset number of graph convolution operations (3 to 5 layers), an output feature matrix is ​​obtained. This output feature matrix is ​​then expanded row by row according to the order of power grid nodes to obtain a spatial hidden state sequence. Each vector in the sequence corresponds to a power grid node, and the value of each vector element reflects the degree of correlation between the node's equipment operating state and that of its neighboring nodes. A larger value indicates stronger coordination between the node's and its neighboring nodes' equipment operating states, and vice versa. This process characterizes the distribution pattern and correlation of equipment operating states in the power grid spatial topology.

[0024] Step 102b involves concatenating and fusing the temporal hidden state sequence and the spatial hidden state sequence to obtain a spatiotemporal fusion feature vector. This spatiotemporal fusion feature vector is then used as a conditional input to the conditional generative adversarial network (GAN) module. The generator within the GAN module is used to obtain the source load power prediction scene. A discriminator distinguishes the generated scene from the real historical scene. Adversarial training optimizes the generator parameters to obtain a set of source load power probability distribution scenes containing confidence intervals, specifically including: After aligning the temporal and spatial hidden state sequences by length, they are concatenated and fused along the feature dimension to obtain a spatiotemporal fused feature vector. The concatenation formula is as follows: ( For spatiotemporal fusion feature vectors, , (This is the aligned temporal and spatial hidden state sequence). The spatiotemporal fusion feature vector is used as the core constraint and input into the generator of the conditional generative adversarial network. The generator takes a normally distributed random noise vector as input and combines the long-term time dependence features of source and load and the spatial topology correlation features of the power grid contained in the spatiotemporal fusion feature vector to output source and load power prediction scenarios. Each prediction scenario is a 15-minute / step time series consistent with the preprocessing in step 100a. It fully includes the specific values ​​(unit: megawatt) of wind power output, photovoltaic power output, conventional unit output, industrial load, residential load, and commercial load at each time series point. The power values ​​of various types of source and load at each time series point meet the 95% confidence interval requirement. At the same time, it meets the physical operation constraints of the power grid such as unit ramp rate and node voltage safety boundary. There are no abnormal situations such as sudden output changes or exceeding the safety range. It accurately matches the source and load fluctuation law and actual operation status of the new power system. The above generation process is repeated 100 to 200 times to obtain 100 to 200 independent source and load power prediction scenarios covering different source and load fluctuation conditions, forming a complete scenario set to fully cover the random fluctuation range of source and load power. The generated multiple sets of source load power prediction scenarios and the real historical source load power scenarios preprocessed in step 100a (preserving the temporal fluctuation patterns of real source load power, the proportional relationships of various types of source loads, and the inherent correlation characteristics with meteorological environment and equipment operating status) are simultaneously input into the discriminator. With minimizing the generator loss and discriminator loss as the core objective, the network parameters of the generator and discriminator are iteratively updated alternately. When, during 50 consecutive iterations, the changes in both generator loss and discriminator loss are less than 1e-5, and the feature cosine similarity between the generated scenario and the real historical scenario is greater than 90%, while the discriminator's accuracy in judging the real scenario is stable above 92% and the misclassification rate for the generated scenario is less than 8%, the convergence condition is met, and the parameter updates of the generator and discriminator are immediately stopped. Statistical analysis is performed on the generated multiple sets of source load power prediction scenarios, calculating the mean and standard deviation of source load power at each time point; using the mean as a benchmark, a confidence interval is calculated at a 95% confidence level, i.e. , , They are the first The mean and standard deviation of each time series point are integrated to obtain a set of source-load power probability distribution scenarios containing confidence intervals.

[0025] Step 103b involves constructing a robust optimization uncertainty set based on the source-load power probability distribution scenario set, and combining this with grid topology parameters, unit ramp rate constraints, and node voltage safety boundaries. Invalid strategy points are eliminated through constraint projection operations to define the initial feasible solution space that satisfies the physical safety operation of the system. Specifically, this includes: Based on the source-load power probability distribution scenario set, a robust optimization uncertainty set is constructed. This uncertainty set is defined as follows: the source-load power value at any time point falls within the 95% confidence interval, ensuring that this uncertainty set can cover more than 95% of source-load power fluctuation scenarios. The power grid topology parameters, unit ramp rate constraints, and node voltage safety boundaries are used as system physical safety constraints. Specific constraints are as follows: Power grid topology constraint: the transmission power of a branch between any two nodes must not exceed the rated power of that branch (the rated power of the branch is determined according to the power grid planning and design, ranging from 100MW to 500MW), and the transmission power must not be negative. Unit ramp rate constraint: the output change rate of each unit at two adjacent time points must not exceed the unit's maximum ramp rate or be lower than its minimum ramp rate. For conventional units, the maximum ramp rate is 1.5% to 3.5% of their rated capacity / min, and the minimum ramp rate is -3.0% to -1.0% of their rated capacity / min. For new energy units, the maximum ramp rate is... The minimum ramp rate is 2.0% to 5.0% of the rated capacity per minute, and the minimum ramp rate is -4.0% to -1.5% of the rated capacity per minute. Node voltage safety constraints mean that the voltage value of each grid node at any given time point must be maintained between the upper and lower voltage limits of that node, and must not exceed the safety boundaries. Specifically, the upper voltage limit for high-voltage nodes is 1.05 to 1.10 times the rated voltage, and the lower voltage limit is 0.90 to 0.95 times the rated voltage. The upper voltage limit for low-voltage nodes is 1.0... The voltage limit is 7 to 1.12 times the rated voltage, with a lower limit of 0.88 to 0.93 times the rated voltage. A set of candidate strategy points is randomly generated, and each strategy point is substituted with all the above constraints for feasibility verification. If a strategy point does not meet the constraints, it is an invalid strategy point. It is then projected onto the constraint boundary through constraint projection. The projection method is to find the point closest to the invalid strategy point that meets all constraints as the projected strategy point. If the constraints are still not met after projection, the invalid strategy point is directly removed. All valid strategy points that have passed the verification are integrated to form an initial feasible solution space. This space is the intersection of the robust optimization uncertainty set and the system physical safety constraint set, which can ensure that all strategy points in it meet the source load power fluctuation coverage requirements and the power grid physical safety operation requirements. If the number of valid strategy points after removal is less than the preset population size (50), the random generation and projection process is repeated until a sufficient number of valid strategy points are obtained.

[0026] This embodiment comprehensively collects new power system operation data from multiple dimensions, fully covering the fluctuation characteristics of both source and load sides, grid structure constraints, and equipment operation limitations, avoiding deviations in scenario generation and feasible domain definition due to data gaps. A pre-trained temporal feature extraction network, integrating multi-head self-attention mechanisms, graph convolutional neural networks, and conditional generative adversarial networks, can extract long-period dependence features of source and load power from the time dimension and grid topology correlation features from the spatial dimension, achieving joint mining and expression of spatiotemporal features. Compared to single feature extraction methods, this approach more accurately depicts the complex fluctuation patterns of source and load power. Using spatiotemporal fusion features as conditions to drive the conditional generative adversarial network for scenario generation ensures that the generated source and load power scenarios closely resemble real-world operating patterns and possess reasonable confidence intervals, effectively improving the accuracy of source and load uncertainty modeling. A robust optimization uncertainty set is constructed based on the source and load power probability distribution scenario set, and the initial feasible solution space is defined through constraint projection operations combined with system physical constraints. This allows for the early elimination of a large number of invalid strategy points, narrowing the optimization search range and improving the overall scheduling method's operational efficiency and strategy security.

[0027] In a preferred embodiment of the present invention, step 2 includes: Step 200: Randomly initialize the candidate strategy population within the initial feasible solution space, and define the economic objective function, the low-carbon objective function, and the stability objective function, specifically including: Using the initial feasible solution space defined in step 103b as the search range, a uniform random sampling method is used to initialize the candidate strategy population, ensuring that the individuals in the population are evenly distributed within the feasible solution space. The size of the candidate strategy population is set to 50 to 100 individuals, each individual corresponding to a complete set of novel power system dispatching strategies, covering conventional unit output allocation schemes, renewable energy unit output absorption schemes, load control schemes, and branch transmission power allocation schemes at each time series point. The dimension of each individual is consistent with the dimension of the initial feasible solution space, and all individuals satisfy the system physical security constraints (grid topology constraints, unit ramp rate constraints, and node voltage security boundary constraints) defined in step 103b. After the candidate strategy population is initialized, in combination with the core requirements of novel power system dispatching, economic objective functions, low-carbon objective functions, and stability objective functions are defined respectively. All three objective functions adopt a minimization objective setting, as follows: The economic objective function takes minimizing the total cost of novel power system dispatching operation as its core objective. The total cost covers the generation cost of conventional units, the absorption cost of renewable energy units, the load control cost, and the branch loss cost. The calculation formula is as follows: ; In the formula It is an economic objective. For candidate strategy individuals, It is a time series point. This represents the total number of time series steps (on a time scale of 15 minutes per step). For the number of conventional generating units, For the first The unit power generation cost of a conventional generating unit For the first The conventional unit in the first Output at time points For the number of new energy generating units, For the first Unit consumption cost of Taiwan's new energy units For the first Taiwan's new energy units in the first Output at time points For the number of load types, For the first The unit control cost of similar loads, For the first Class load in the first The control power at the timing point, For the number of branch roads, For the first The unit loss cost of each branch road For the first The branch road in the Transmission loss power at timing points.

[0028] The low-carbon objective function takes minimizing the total carbon emissions from the dispatch and operation of the new power system as its core objective. The total carbon emissions mainly originate from the power generation process of conventional generating units, while the power generation process of new energy units (wind power and photovoltaic) does not generate carbon emissions. The calculation formula is as follows: ; In the formula, It is a low-carbon goal. For the first Carbon emissions per unit of electricity generated by a conventional power plant.

[0029] The stability objective function takes minimizing voltage fluctuations in the dispatching and operation of the new power system as its core objective. Voltage fluctuations reflect the stability of the power grid operation; the smaller the fluctuations, the more stable the power grid operation. The calculation formula is as follows: ; In the formula, It is a stability objective. For the number of power grid nodes, For the first The power grid node at the ... The actual voltage at the timing point, For the first The rated voltage of each grid node.

[0030] Step 201: Calculate the gradient vector for each individual in the candidate strategy population under the economic objective function, the low-carbon objective function, and the stability objective function, respectively; based on the gradient vectors, calculate the mutual information entropy values ​​between each pair of objective functions to quantify the coupling degree and conflict intensity between objectives, specifically including: For any individual in the candidate strategy population Its gradient vectors under the three objective functions are denoted as follows: The gradient vector has the same dimension as the candidate policy individual. Each dimension corresponds to the partial derivative of a decision variable in the policy individual. The partial derivatives are calculated using the central difference method, and the formula is as follows: ; In the formula, yes, For the first The small perturbation values ​​of the decision variables, taking values ​​of 1e-5, , Decision variables Increase, decrease After that, the The values ​​of the objective function, This is the index of the decision variable. After calculating the gradient vectors of all individuals, the mutual information entropy values ​​between each pair of objective functions are calculated based on the gradient vectors. The mutual information entropy value is used to quantify the coupling degree and conflict strength between the three objective functions. The larger the mutual information entropy value, the higher the coupling degree and the stronger the conflict strength between the two objective functions; the smaller the mutual information entropy value, the lower the coupling degree and the weaker the conflict strength between the two objective functions. First, the gradient vectors of all candidate policy individuals are standardized to eliminate the difference in dimensions between the gradient vectors of different objective functions. The standardization formula is: ; In the formula, For the first The standardized result of the gradient vector of the objective function. For all candidate strategy individuals in the first The minimum value of the gradient vector under each objective function. For all candidate strategy individuals in the first The maximum value of the gradient vector under each objective function; after standardization, the range of the gradient vector is uniformly [0, 1]. Calculate the maximum value of the gradient vector under any two objective functions. and ( Mutual information entropy between ) The calculation formula is: ; In the formula, For the first The information entropy of the gradient vector of the objective function For the first The information entropy of the gradient vector of the objective function For the first The and the first The joint information entropy of the gradient vectors of the objective functions; where, information entropy The calculation formula is: ; Joint Information Entropy The calculation formula is: ; In the formula, The size of the candidate strategy population. For the first The gradient vector of each candidate policy individual in the th... The probability in the set of gradient vectors of the objective function ( =1 / M), For the first The candidate strategy individuals in the first The gradient vector of the i-th objective function and the i-th objective function The candidate strategy individuals in the first The joint probability of the gradient vectors under each objective function ( =1 / M²); mutual information entropy value The range of values ​​is Normalize it: ; Normalized mutual information entropy value The value of is in the range of [0, 1], where 0 indicates that the two objective functions are completely independent and 1 indicates that they are completely coupled; the mutual information entropy value used in subsequent steps refers to this normalized value.

[0031] Step 202: Based on the coupling degree and conflict strength, obtain the dynamic adaptive weight matrix. Use the dynamic adaptive weight matrix to perform a weighted summation of the gradient vectors to synthesize a comprehensive guiding potential vector pointing towards the Pareto improvement direction. Specifically, this includes: The weight coefficients of the three objective functions are calculated based on the normalized mutual information entropy value. The calculation of the weight coefficients follows the principle that the higher the conflict intensity, the smaller the weight coefficient; and the higher the coupling degree, the more balanced the weight coefficients. The specific calculation formula is as follows: ; ; like (i.e., the three objective functions are completely coupled pairwise), then let .

[0032] In the formula, For the first The weight coefficients of the objective function take values ​​in the range [0, 1], and satisfy = 1. When the first The larger the sum of the mutual information entropy values ​​of the first objective function and the other two objective functions (i.e., the higher the conflict intensity), then... The smaller the value, the higher the weighting coefficient. The smaller the value, the less likely the objective will dominate the optimization process; conversely, the lower the conflict intensity, the higher the weighting coefficient. The larger the value, the more fully the optimization requirements of this objective are reflected.

[0033] Constructing a dynamic adaptive weight matrix This matrix is ​​a 3×3 diagonal matrix. The elements on the diagonal are the weight coefficients of the three objective functions, and the elements off-diagonal are 0. This is the dynamic adaptive weight matrix. The dimension matches the number of objective functions, and can be dynamically updated according to changes in the coupling degree and conflict intensity between objectives, ensuring the rationality and adaptability of weight allocation. A weighted sum of the gradient vectors of the three objective functions is used with a dynamic adaptive weight matrix to synthesize a comprehensive guiding potential vector pointing towards the Pareto improvement direction. This vector can guide the candidate policy population to iterate in the direction of co-optimization of the three objective functions, avoiding the overall performance degradation caused by single-objective optimization. The specific synthesis formula is as follows: ; In the formula, To integrate the potential field vector, its dimension is consistent with the dimension of the candidate policy individuals; , , The three standardized gradient vectors of the objective function are given in step 201. During the weighted summation, the weight matrix W assigns weights to the three standardized gradient vectors using the weight coefficients of the three objective functions on its diagonal, based on the normalized mutual information entropy value calculated in step 201. Quantify the coupling degree and conflict intensity between the objective functions; define the first The sum of the mutual information entropy values ​​of the first objective function and the other two objective functions is , The range of values ​​is .based on Dynamically calculate the normalized final weights of the objective function. Ensure that the sum of the three weights is 1. According to... The size can be used to classify conflict intensity levels and give weight values, low conflict ( ≤0.6) After normalization, the weights are usually in the range of 0.4 to 0.6, ensuring that the optimization requirements of this objective are fully reflected; (0.6 < ≤1.2) After normalization, the weights are usually in the range of 0.25 to 0.4, achieving a balance between optimization needs and overall coordination; high conflict ( >1.2) After normalization, the weights are usually in the range of 0.1 to 0.25 to avoid the objective from overly dominating the optimization process.

[0034] In this embodiment, the candidate strategy population is initialized based on the previously defined initial feasible solution space, and all individuals satisfy the system's physical security constraints, avoiding invalid strategy points from participating in optimization, reducing the amount of optimization computation, and ensuring a uniform population distribution. Three objective functions—economic efficiency, low carbon emissions, and stability—are clearly defined, covering core costs, carbon emissions, and grid stability indicators in scheduling operations, ensuring the rationality and accuracy of multi-objective optimization. Gradient vector calculation accurately captures the changing trends of each objective function, and the mutual information entropy value is combined to quantify the coupling degree and conflict intensity between objectives, overcoming the limitation of not being able to accurately quantify multi-objective coupling conflicts. The dynamic adaptive weight matrix can dynamically adjust the weights according to the coupling degree and conflict intensity between objectives. Compared with fixed weight settings, this avoids a single objective excessively dominating the optimization process, improving the overall performance of the scheduling strategy. The synthesis of the comprehensive guiding potential vector is based on the standardized gradient vector and dynamic weights, pointing towards the Pareto improvement direction, further improving the optimization effect of the scheduling strategy.

[0035] In a preferred embodiment of the present invention, step 3 includes: Step 300a: Calculate the geometric centroids of all candidate strategy points in the initial feasible solution space, and use the geometric centroids as the first evolution anchor points for the corresponding system baseline load conditions; identify the boundary candidate strategy points in the initial feasible solution space with the largest Euclidean distance from the geometric centroids, and use the boundary candidate strategy points as the second evolution anchor points for the corresponding extreme disturbance boundary conditions, specifically including: Iterate through all individuals in the candidate strategy population initialized in step 200 (i.e., the current candidate strategy point), using the set of decision variables for each candidate strategy point as spatial coordinates, and calculate the geometric centroid of these candidate strategy points. First, sum the coordinates of all candidate strategy points sequentially, then divide by the total number of valid candidate strategy points. The result is the spatial coordinates corresponding to the first evolutionary anchor point. Use this geometric centroid as the first evolutionary anchor point corresponding to the baseline load condition representing normal system operation and stable load power variation.

[0036] After calculating the geometric centroid, the Euclidean distance between each individual in the current candidate strategy population and the geometric centroid is calculated one by one. The Euclidean distance values ​​of all individuals are traversed, and the individual with the largest Euclidean distance value is selected as the second evolutionary anchor point corresponding to the extreme perturbation boundary condition that represents the large fluctuation of source load and the system operation close to the safety boundary.

[0037] Step 300b: At the beginning of each rolling scheduling cycle, based on the latest observation and forecast data, a new source-load power probability distribution scenario set is generated, and the variance integral of each scenario curve in the source-load power probability distribution scenario set relative to the mean curve is calculated to obtain the dispersion index, specifically including: Taking the source-load power probability distribution scenario set regenerated in the current period (e.g., every 15 minutes) using the same method as in step 102b as the analysis object, first calculate the mean source-load power of all scenarios at the same time point to form a mean curve covering the entire time scale. After time-series calculation, a complete mean curve is formed. Then, for all scenario curves in the scenario set, calculate their variance integral relative to the mean curve over the entire time range, and use the integral result as a dispersion index characterizing the overall dispersion of source-load fluctuations. The formula for calculating the dispersion index is: ; In the formula, This is the start time of the time series; This is the time when the sequence ends; This represents the total number of scenarios included in the source-load power probability distribution scenario set; For the first The scene in The source load power values ​​at each time point; For all scenarios in the first The average source load power at each time point.

[0038] Step 301b: Determine whether the dispersion index is less than a preset threshold. When the dispersion index is less than the preset threshold, it indicates that the source load fluctuation is in a stable state. Using the first evolution anchor point as the center and the preset first covariance matrix as the shape parameter, construct a locally convergent hyperellipsoidal domain, specifically including: Set the preset threshold for the dispersion index as follows: The dispersion index calculated in step 300b is compared with the preset threshold. When the dispersion index is less than the preset threshold, it is determined that the current source load power is in a stable fluctuation state, and the system operating characteristics are close to the baseline load condition. Using the first evolution anchor point constructed in step 300a as the spatial center and the preset first covariance matrix as the shape parameter, a locally convergent hyperellipsoidal domain is constructed for the local convergence search of the optimization algorithm. This hyperellipsoidal domain is used to limit the optimization search range under stable conditions, ensuring that the search process focuses on the vicinity of the baseline condition and avoiding invalid searches. The first covariance matrix is ​​a diagonal matrix, with diagonal elements taking values ​​of 1.0, 1.0, 0.8, and 0.8 respectively. The matrix dimension is completely consistent with the spatial dimension of the decision variables obtained in step 200. The constraints of the hyperellipsoidal domain are expressed as follows: In the formula For any candidate strategy point within the hyperellipsoidal domain, As the first evolutionary anchor point, Let be the first covariance matrix. It is the inverse of the first covariance matrix. The symbol is a transpose. This constraint, by quantifying the spatial relationship between candidate strategy points and the first evolution anchor point, limits the optimization search range to the vicinity of the baseline condition, effectively improving the convergence speed of the optimization algorithm. At the same time, it ensures that all candidate strategy points within the search range meet the system's physical security constraints. The dimension of the first covariance matrix is ​​equal to the spatial dimension of the decision variables defined in step 200. If the dimension of the decision variables is greater than 4, the remaining diagonal elements are uniformly set to 1.0.

[0039] Step 302b: When the dispersion index is greater than or equal to a preset threshold, it indicates that the source load fluctuation is in a state of severe disturbance. Using the second evolution anchor point as the center and the preset second covariance matrix as the shape parameter, a global exploration hyperellipsoidal domain is constructed, specifically including: When the dispersion index calculated in step 300b is greater than or equal to When the current source load power is determined to be in a state of severe disturbance, the source load power fluctuates greatly and the system operating characteristics are close to the extreme disturbance boundary conditions defined in step 300a. The optimization requirements are mainly to fully cover the feasible region and avoid missing the optimal scheduling strategy, so it is necessary to expand the search range to achieve global exploration.

[0040] Based on the above operating condition determination results, using the second evolution anchor point constructed in step 300a as the spatial center and the preset second covariance matrix as the shape parameter, a global exploration hyperellipsoidal domain is constructed for the global range search of the optimization algorithm. This hyperellipsoidal domain is used to expand the optimization search range under drastically disturbed operating conditions, ensuring coverage of all feasible regions corresponding to extreme operating conditions. The second covariance matrix is ​​also a diagonal matrix, with diagonal elements taking values ​​of 3.0, 3.0, 2.5, and 2.5 respectively. The matrix dimension is completely consistent with the spatial dimension of the decision variables obtained in step 200. The constraints of the hyperellipsoidal domain are expressed as follows: In the formula As the second evolutionary anchor point, The second covariance matrix, The second covariance matrix is ​​the inverse of the second covariance matrix. This constraint expands the optimization search range by quantifying the spatial relationship between candidate strategy points and the second evolution anchor point, fully covering the feasible region corresponding to extreme conditions, effectively avoiding the omission of the optimal scheduling strategy due to drastic fluctuations in source load, and ensuring the integrity and reliability of the optimization results. The dimension of the second covariance matrix is ​​equal to the spatial dimension of the decision variables defined in step 200; if the dimension of the decision variables is greater than 4, the remaining diagonal elements are uniformly taken as 3.0.

[0041] This embodiment constructs dual evolution anchor points using the geometric centroid and the farthest boundary point, corresponding to the system's baseline and extreme disturbance conditions respectively, thus matching optimization requirements under different source load fluctuation characteristics. It employs variance integral quantification of the source load fluctuation dispersion index, with a complete calculation formula and standardized dimensions, objectively reflecting the severity of source load power fluctuations. Based on the comparison between the dispersion index and a fixed threshold, it adaptively selects either a local convergence domain or a global exploration domain, balancing the need for rapid convergence under stable conditions with the global search requirement under severe disturbance conditions, thereby improving optimization efficiency and solution accuracy. A covariance matrix is ​​used to control the shape of the hyperellipsoidal domain, while ensuring clear search domain boundaries and well-defined constraints.

[0042] In a preferred embodiment of the present invention, step 4 includes: Step 400: Project the candidate strategy population onto the hyperellipsoid domain, and calculate the Mahalanobis distance between each candidate strategy point and the geometric center of the hyperellipsoid domain after mapping; divide candidate strategy points with Mahalanobis distance less than the Mahalanobis distance threshold into a core dense subset, and divide candidate strategy points with Mahalanobis distance greater than or equal to the Mahalanobis distance threshold into an edge sparse subset, specifically including: The candidate strategy population initialized and optimized in step 200 is projected onto the hyperellipsoidal domain constructed in step 301b or 302b. The core purpose of this projection mapping is to ensure that all candidate strategy points are within the optimization search range corresponding to the current working condition, avoiding invalid iterations caused by exceeding the search domain. The projection mapping rule is as follows: if a candidate strategy point is already within the hyperellipsoidal domain, its position remains unchanged; if a candidate strategy point exceeds the boundary of the hyperellipsoidal domain, it is projected onto the boundary of the hyperellipsoidal domain. The projection method is to find the intersection point of the line connecting the candidate strategy point to the geometric center of the hyperellipsoidal domain and the boundary of the hyperellipsoidal domain, and use this intersection point as the projected candidate strategy point.

[0043] After projection, the Mahalanobis distance between each mapped candidate strategy point and the geometric center of the hyperellipsoid is calculated. The Mahalanobis distance quantifies the spatial correlation between the candidate strategy point and the geometric center. A Mahalanobis distance threshold of 0.6 is set (this value is less than the Mahalanobis distance of 1 corresponding to the hyperellipsoid boundary, ensuring that points near the center are distinguished from those near the boundary). The Mahalanobis distance of each candidate strategy point is compared with this threshold, resulting in two subsets: candidate strategy points with a Mahalanobis distance less than 0.6 are classified as a core dense subset (concentrated near the center of the hyperellipsoid, with high optimization potential); candidate strategy points with a Mahalanobis distance greater than or equal to 0.6 and less than or equal to 1 are classified as an edge sparse subset (distributed at the edge of the hyperellipsoid, used to maintain population diversity). If the edge sparse subset is empty, some points from the core dense subset are randomly selected and moved to the edge sparse subset to ensure population diversity.

[0044] Step 401: For the core dense subset, calculate the second-order partial derivative approximation matrices of the economic objective function, the low-carbon objective function, and the stability objective function. Use the second-order partial derivative approximation matrices to perform curvature correction on the gradient direction and update the positions of each candidate strategy point in the core dense subset to accelerate local convergence. For the marginal sparse subset, determine a random step-size vector that follows a heavy-tailed distribution. Use the random step-size vector to drive candidate strategy points to perform long-distance position jumps and update the positions of each candidate strategy point in the marginal sparse subset to maintain population diversity. Merge the updated core dense subset and the updated marginal sparse subset to obtain a hybrid evolutionary strategy set, specifically including: Differentiated update strategies are adopted for the partitioned core dense subset and marginal sparse subset to achieve accelerated local convergence and maintenance of population diversity, respectively. The specific update process is as follows: For each candidate strategy point in the core dense subset, the second-order partial derivative approximation matrix of the economic objective function, the low-carbon objective function, and the stability objective function is calculated. This matrix is ​​a finite difference approximation of the Hessian matrix and is used to characterize the curvature characteristics of each objective function at the current candidate strategy point. The elements on the diagonal of the matrix correspond to the second-order partial derivatives of each objective function with respect to a single decision variable. The sign of their values ​​determines the curvature direction; a positive value indicates that the objective function at that decision variable... The decision variables exhibit convexity in one dimension, while negative values ​​indicate concavity. The magnitude of the absolute value determines the steepness of the curvature; the larger the absolute value, the steeper the curvature, and the faster the objective function changes in that dimension. The off-diagonal elements of the matrix correspond to the second-order mixed partial derivatives of each objective function with respect to two different decision variables. Their magnitude and sign represent the coupling curvature influence between different decision variables; the larger the absolute value, the stronger the joint influence of the two decision variables on the objective function, indicating higher coupling. This matrix, through the above dimensional representations, provides a precise basis for gradient direction correction, avoiding oscillations during gradient descent and accelerating local convergence. The second-order partial derivative approximation matrix is ​​calculated using the finite difference approximation method of the Hessian matrix. Taking the economic objective function as an example, its Hessian matrix (second-order partial derivative approximation matrix) is... The calculation formula is: ; In the formula, For the economic objective function, Candidate policy points of One decision variable, For the economic objective function, the first The and the first The second-order mixed partial derivatives of the decision variables are approximated using the central difference method, and the calculation formula is as follows: ; In the formula, The first The, the The small perturbation values ​​of each decision variable are all 1e-5 (matching the magnitude of the decision variables). , It is the index of the decision variables; the second-order partial derivative approximation matrix of the low-carbon objective function and the stability objective function. The same method as described above is used for calculation.

[0045] The synthesized guided potential vector synthesized in step 202 is standardized using the second-order partial derivative approximation matrices of the three objective functions before curvature correction. First, the synthesized guided potential vector... To standardize the values ​​of all components to a range of [0, 1], the standardization formula is as follows: ; Then, the curvature correction of the standardized integrated guided potential vector is performed using the second-order partial derivative approximation matrices of the three objective functions. The correction formula is as follows: ; in, This is the corrected combined gradient vector. The standardized composite guiding potential field vector (obtained by standardizing the composite guiding potential field vector synthesized in step 202, with each component taking values ​​in the range of [0, 1]). It is the minimum value of the comprehensive guiding potential vector at each candidate strategy point. This represents the maximum value of the integrated guiding potential vector at each candidate policy point. The average of the three second-order partial derivative approximation matrices is used to balance the curvature effects of the three objective functions, ensuring the rationality of the correction direction. Based on the corrected integrated gradient vector, the positions of each candidate policy point in the core dense subset are updated using the following formula: ; In the formula, The positions of candidate strategy points after the core dense subset is updated. This is the position before the update. The learning rate is set to 0.01 (balancing convergence speed and stability). The negative sign indicates updating along the gradient descent direction, achieving faster local convergence. For each candidate policy point in the marginal sparse subset, a random step size vector following a heavy-tailed distribution is determined. The Cauchy distribution is chosen as the heavy-tailed distribution because it has a long tail and can generate a large random step size, driving candidate policy points to make long-distance position jumps, avoiding the population from getting trapped in local optima and maintaining population diversity. Random step size vector The formula for generating it is: ; In the formula, The scaling parameter of the Cauchy distribution is set to 0.5 (to control the overall magnitude of the random step size and avoid exceeding the feasible region due to an excessively large step size). The dimension of the random step size vector is consistent with the dimension of the decision variable space, and the components of each dimension follow the aforementioned Cauchy distribution. Using this random step size vector, the candidate policy points in the marginal sparse subset are driven to perform long-distance position jumps. The update formula is as follows: ; In the formula, The updated candidate policy point positions are for the edge sparse subset. The addition operation is used to achieve long-distance jumps between candidate policy points based on their original positions, thus maintaining population diversity. The updated core dense subset and the updated edge sparse subset are merged to remove duplicate candidate policy points, resulting in a new candidate policy set, which is the hybrid evolutionary policy set. This set retains the optimization potential of the core region while maintaining population diversity.

[0046] In this embodiment, the candidate strategy population is projected into the hyperellipsoidal domain to ensure that all iterative individuals are within the search range corresponding to the current working condition. The core and peripheral subsets are divided by combining Mahalanobis distance, which not only focuses on the local optimization of the core region, but also takes into account the diversity maintenance of the peripheral region, thereby improving the targeting and comprehensiveness of the optimization.

[0047] The core dense subset uses a second-order partial derivative approximation matrix for gradient curvature correction, which effectively avoids oscillations during gradient descent and accelerates local convergence. The marginal sparse subset uses a heavy-tailed distributed random step-size vector to achieve long-distance jumps, which effectively maintains population diversity, avoids getting trapped in local optima, and balances convergence speed and optimization accuracy.

[0048] In a preferred embodiment of the present invention, step 5 includes: Step 500: Obtain the velocity vector of the hybrid evolution strategy set in the previous iteration cycle, and use the velocity vector as the historical iteration velocity memory term; combine the historical iteration velocity memory term with the preset inertia weight coefficient to obtain the inertial component; superimpose the inertial component with the comprehensive guiding potential field vector to obtain the state update vector, specifically including: For the first iteration (generation 1), there is no previous iteration cycle. At this time, the historical iteration velocity memory term is initialized to a zero vector, with its dimension consistent with the dimension of the decision variables of the candidate policy individuals. Starting from generation 2, the velocity vectors of each candidate policy point within the mixed evolutionary policy set in the previous iteration cycle are obtained. This velocity vector is obtained during the trajectory evolution of the candidate policy points in the previous iteration cycle (i.e., the value of the state update vector in the previous cycle), and is used to characterize the position update magnitude and direction of the candidate policy points in the previous iteration cycle. This velocity vector is used as the historical iteration velocity memory term. .

[0049] Set the preset inertia weight coefficient. The value is set to 0.7 to balance the influence of the historical iteration velocity memory term. The historical iteration velocity memory term is multiplied by a preset inertia weight coefficient to obtain the inertia component. The inertial component is superimposed with the synthesized guiding potential vector obtained in step 202 to obtain the state update vector, calculated using the following formula: ; In the formula, The guiding strength coefficient is set to 0.05 (to control the influence of the integrated guiding potential field on the state update and avoid excessive oscillations). During the superposition process, the dimensions and units of the two vectors are ensured to be consistent (both have undergone standardization). This state update vector integrates historical iteration inertia information and the current Pareto improvement direction, providing clear guidance for the trajectory evolution of candidate policy points; in the first iteration... The zero vector is the updated state update vector, which is stored in the historical iteration speed memory for use in the next cycle.

[0050] A preset inertia weight coefficient is set to balance the influence of the historical iteration velocity memory term. Considering the convergence requirements of the optimization process, the inertia weight coefficient is set to 0.7 (within a range of 0.5 to 0.9, balancing inertial influence and convergence speed). The historical iteration velocity memory term is multiplied by the preset inertia weight coefficient to obtain the inertia component. This inertia component is then vector-superimposed with the synthesized guiding potential field vector from step 202, ensuring that the two vectors have the same dimension and uniformity (both are standardized) to obtain the state update vector. This vector integrates historical iteration information and current comprehensive guiding information, providing a clear direction for the trajectory evolution of candidate strategy points.

[0051] Step 501: Update the position coordinates of each candidate strategy point in the hybrid evolution strategy set according to the state update vector to realize trajectory evolution; perform boundary reflection operation on the evolved candidate strategy points to reflect candidate strategy points that exceed the boundary of the initial feasible solution space back into the initial feasible solution space; perform non-dominated screening on the reflected candidate strategy set to remove dominated individuals, and obtain the evolved non-dominated candidate strategy set, specifically including: Based on the state update vector synthesized in step 500, update the position coordinates of each candidate policy point in the hybrid evolution policy set to realize the trajectory evolution of the candidate policy points. The position update formula is: The position of the candidate policy point after evolution in the current iteration cycle = the position in the previous iteration cycle + the state update vector of the current iteration cycle. After the trajectory evolution is completed, a boundary reflection operation is performed on all evolved candidate policy points. The core purpose is to reflect candidate policy points that exceed the initial feasible solution space boundary defined in step 103b back into the initial feasible solution space, ensuring that all candidate policy points satisfy the system's physical security constraints. The boundary reflection operation rule is that if the value of a certain dimension decision variable of a candidate policy point exceeds the feasible boundary (upper or lower limit) of that dimension, then the value of that dimension is reflected back into the boundary. The reflection formula is: when hour, ; when hour, ; In the formula, For the first generation after evolution The values ​​that the decision variable can take. , The first The feasible upper and lower limits of the decision variables are determined based on the initial feasible solution space in step 103b. For decision variables related to the output of conventional units, the feasible upper limit is 100% of the rated capacity of the unit, and the feasible lower limit is 10% of the rated capacity of the unit. For decision variables related to the output of new energy units, the feasible upper limit is 100% of the rated capacity of the unit, and the feasible lower limit is 0MW. For decision variables related to load regulation, the feasible upper limit is 10% of the rated power of the corresponding load, and the feasible lower limit is -10% of the rated power of the corresponding load. For decision variables related to branch transmission power, the feasible upper limit is the rated power of the corresponding branch (100MW to 500MW), and the feasible lower limit is 0MW. For the first reflection After reflection, all decision variables are within the feasible range, taking values ​​from the given decision variables.

[0052] The candidate strategy set after boundary reflection is subjected to non-dominated selection using a non-dominated ranking method. Dominated individuals are removed, and non-dominated individuals are retained, resulting in the evolved non-dominated candidate strategy set. The non-dominated selection rule is as follows: for any two individuals in the candidate strategy set... and ,like The values ​​obtained in the three objective functions of economy, low carbon emissions, and stability are all no worse than those of [previous standard]. And at least one objective function has a value better than Then determine Individuals that are dominated are eliminated; if two individuals do not dominate each other, they are both retained. The resulting set of non-dominated candidate strategies after screening represents the optimal candidate strategy points within the current iteration period.

[0053] In this embodiment, the state update vector integrates the historical iteration velocity memory term and the comprehensive guiding potential field vector, and combines the inertia weight coefficient to balance historical information with current guiding information, ensuring that the trajectory evolution of candidate strategy points is smooth and orderly, avoiding iterative abrupt changes, and improving the stability of the optimization process. After trajectory evolution, a boundary reflection operation is performed to ensure that all candidate strategy points are within the initial feasible solution space, satisfying the system's physical safety constraints, avoiding invalid individuals from participating in subsequent optimization, and reducing computational burden; non-dominated screening removes dominated individuals and retains the optimal candidate strategy point, improving the quality of the optimization results.

[0054] In a preferred embodiment of the present invention, step 6 includes: Step 600: Perform fast non-dominated sorting on the evolved non-dominated candidate policy set to divide the frontier surfaces into different dominance levels; for all candidate policy points within the same frontier surface, map the corresponding multidimensional objective function values ​​to the target space, and establish a three-dimensional voxel mesh covering all candidate policy points of the corresponding frontier surface in the target space, with each voxel unit being a cube of equal size; traverse all voxel units, marking voxel units containing at least one candidate policy point as occupied voxels, and the remaining ones as empty voxels, specifically including: A fast non-dominated candidate policy set obtained in step 501 is subjected to a sorting operation. The purpose is to divide the frontier into different levels according to the dominance relationship of the candidate policy points. The first frontier is the set of the best non-dominated candidate policy points in the current iteration period. The dominance level of the subsequent frontiers decreases in turn. The sorting process follows the non-dominated screening rules to ensure that the candidate policy points in the same frontier do not dominate each other and that the candidate policy points in different frontiers have a clear dominance relationship.

[0055] The specific process of fast non-dominated sorting is as follows: First, traverse all candidate strategy points in the evolved non-dominated candidate strategy set, count the number of times each candidate strategy point is dominated by other candidate strategy points, and divide the candidate strategy points with a domination count of 0 into the first frontier; then remove all candidate strategy points in the first frontier, count the domination count of the remaining candidate strategy points again, and divide the candidate strategy points with a domination count of 0 into the second frontier; repeat the above process until all candidate strategy points are divided into frontiers of the corresponding level, thus completing fast non-dominated sorting.

[0056] For all candidate strategy points within each frontier, the economic objective function value, low-carbon objective function value, and stability objective function value corresponding to each candidate strategy point are used as three-dimensional coordinates and mapped to a three-dimensional target space. The three coordinate axes of the target space correspond to the three objective functions, and the range of values ​​of the coordinate axes is consistent with the range of values ​​of each objective function. The visualization distribution of candidate strategy points in the target space is realized through coordinate mapping.

[0057] A 3D voxel mesh covering all candidate policy points on the corresponding front surface is constructed within the target space. The mesh construction range completely encompasses the coordinate range of all candidate policy points within the front surface, ensuring that all candidate policy points are covered by the mesh. Each voxel element is a cube of uniform size. Before constructing the voxel mesh, the three objective function values ​​of all candidate policy points within the same front surface are first standardized using the following formula: ; in It is the first The result of standardizing the objective function values ​​(within the range of [0, 1]). Candidate strategy points In the The original values ​​of each objective function (economic, low-carbon, or stability objective function values). It is the current frontier surface containing all candidate strategy points at the th... Minimum value on each objective function It is the current frontier surface containing all candidate strategy points at the th... The maximum value of each objective function It is the index of the objective function ( =1 corresponds to the economic objective. =2 corresponds to the low-carbon target. =3 corresponds to the stability objective), mapping the standardized objective function value to... Cube target space. The side length of the voxel unit is 0.5 (dimensionless). This ensures the mesh uniformly covers the normalized target space. The number of voxel units is determined based on the construction range and side length, using the following formula: ; In the formula, This represents the total number of voxel units in a 3D voxel mesh. , These represent the maximum and minimum values ​​of the X-axis of the objective space (the economic objective function); , These represent the maximum and minimum values ​​of the target space Y-axis (low-carbon objective function); , These represent the maximum and minimum values ​​of the Z-axis (stability objective function) in the objective space, respectively. The value is 0.5, representing the side length of the voxel cell, ensuring that the mesh completely covers all candidate policy points.

[0058] After the mesh is constructed, all voxel cells are traversed to determine whether each voxel cell contains at least one candidate policy point. If a voxel cell contains at least one candidate policy point, it is marked as an occupied voxel; if a voxel cell does not contain any candidate policy points, it is marked as an empty voxel. The classification of voxel cells is achieved through marking.

[0059] Step 601: Identify the occupied voxels corresponding to the extreme points located on the boundary of the corresponding front surface from all candidate policy points within the corresponding front surface, and use the occupied voxels as the source point set for the distance transformation; using a multi-source parallel propagation method, calculate the shortest three-dimensional Manhattan distance from each occupied voxel to the source point set of the distance transformation layer by layer outward to obtain the distance transformation field corresponding to the corresponding front surface; for the case where the same occupied voxel contains multiple candidate policy points, take the distance value of the distance transformation field at the center of the corresponding voxel as the common voxel distance value of all candidate policy points within the corresponding voxel, specifically including: For each frontier, from all candidate policy points within the frontier, the voxels corresponding to the extreme points located on the boundary of the corresponding frontier are identified. The extreme points are identified by selecting the candidate policy points with the largest and smallest values ​​in the three objective function dimensions in the target space. These candidate policy points are the boundary extreme points of the frontier, and their corresponding voxels are the source point set of the distance transformation. The voxels in the source point set are used as the reference points for distance calculation.

[0060] A multi-source parallel propagation approach is adopted, starting with all occupied voxels within the source point set and propagating outward layer by layer. The shortest 3D Manhattan distance from each occupied voxel to the source point set of distance transformation is calculated. The 3D Manhattan distance accurately quantifies the spatial distance between two voxel units and is computationally efficient, adapting to the requirements of multi-source parallel propagation. Through calculation, the distance transformation field corresponding to the leading edge is obtained. This distance transformation field characterizes the spatial correlation between each occupied voxel and the boundary extremum point of the leading edge. A larger distance value indicates that the occupied voxel is closer to the center of the leading edge, and the corresponding candidate strategy point has a more balanced overall performance.

[0061] When multiple candidate policy points are contained within the same voxel, in order to ensure the uniformity of distance values, the distance value of the distance transformation field at the center of the corresponding voxel is taken as the common voxel distance value of all candidate policy points within that voxel.

[0062] Step 602: Within the same frontal plane, sort the candidate strategy points according to their voxel distance values ​​from largest to smallest, and select the top preset proportion of candidate strategy points in the sorted sequence to obtain the Pareto scheduling strategy library; collect real-time snapshots of the current power grid operating status, calculate the corresponding three objective function reference values ​​based on the current power grid operating status snapshots, and construct a three-dimensional reference vector of the current power grid operating status; calculate the cosine similarity between the three-dimensional reference vector and the three-dimensional objective function vectors of each candidate strategy point in the Pareto scheduling strategy library as the matching degree; select the candidate strategy point with the highest matching degree and the lowest overall cost as the final scheduling strategy; decompose the final scheduling strategy into active power output command sequences and reactive power output command sequences for each control unit, and send them to the control terminal for execution, specifically including: Within the same frontal plane, candidate strategy points are sorted in descending order of their voxel distance values ​​(i.e., the three-dimensional Manhattan distance calculated in step 601). A larger voxel distance value indicates that the occupied voxel of the corresponding candidate strategy point is closer to the center of the frontal plane, and its comprehensive performance in terms of economic efficiency, low carbon emissions, and stability is more balanced, making it more suitable as a candidate strategy for real-time scheduling. A preset proportion is set, taking into account the practicality and diversity requirements of scheduling strategies. The top 20% of candidate strategy points in the sorted sequence are selected and integrated to form a Pareto scheduling strategy library. This library contains the candidate scheduling strategies with the best and most balanced comprehensive performance within the current iteration cycle, providing alternative solutions for real-time scheduling. Real-time snapshots of the current power grid operation status are collected. The collected content is consistent with the system operation data collected in step 100a, including real-time values ​​of wind power output, photovoltaic power output, conventional unit output, industrial load, residential load, and commercial load, as well as real-time operating parameters such as grid node voltage and branch transmission power. The collection frequency is consistent with the 15-minute / step time series scale.

[0063] After real-time acquisition of a snapshot of the current power grid operation status, three corresponding objective function reference values ​​are calculated based on the snapshot: economic reference value, low-carbon reference value, and stability reference value. The calculation method is consistent with the objective function definition in step 200. The measured data of unit output, load, and voltage at the current moment are substituted into the objective function formulas for economic, low-carbon, and stability to obtain the three-dimensional reference vector under the current state. The feature vector of the candidate strategy point remains its three-dimensional objective function value vector Vc = After normalizing the three-dimensional reference vector and Vc in the current state, the cosine similarity between the two is calculated as the matching degree. The similarity value ranges from [0, 1]. The closer the value is to 1, the stronger the adaptability between the current power grid operation state and the alternative scheduling strategy; the closer the value is to 0, the weaker the adaptability.

[0064] Using the calculated cosine similarity as the matching score, all candidate scheduling policies in the Pareto scheduling policy library are traversed, and the candidate policy with the highest matching score is selected. If multiple candidate policy points have the same matching score, the comprehensive cost of these candidate policy points is further calculated. The comprehensive cost is the weighted sum of the three objective functions, and the calculation formula is as follows: ; In the formula, For the first The combined cost of each alternative scheduling strategy; , , The weight coefficients in the dynamic adaptive weight matrix obtained in step 202 satisfy the following conditions: =1; , , The first The objective function values ​​of the economy, low carbon emissions, and stability of the alternative scheduling strategies.

[0065] The candidate strategy point with the highest matching degree and the lowest overall cost is selected as the final scheduling strategy. This strategy is suitable for the current real-time operation of the power grid and achieves optimal synergy among the three objectives. The final scheduling strategy is decomposed according to the type of control unit, specifically into active power output command sequences for each conventional unit (thermal power unit, hydropower unit, etc.), active power output command sequences for each new energy unit (wind power, photovoltaic, etc.), and reactive power output command sequences for each control unit (unit, reactive power compensation device, etc.). The timing scale of the command sequences is consistent with 15 minutes / step. The decomposed output command sequences are sent to the control terminals of each control unit. The control terminals execute control operations according to the command sequences, completing real-time scheduling and forming a complete scheduling closed loop.

[0066] This embodiment uses rapid non-dominated sorting to divide the frontier into different levels, focuses on the optimal frontier to construct a voxel grid, and uses voxel labeling to achieve spatial classification of candidate strategy points, improving the targeting of strategy selection. Multi-source parallel propagation is used to calculate the three-dimensional Manhattan distance, generating a distance transformation field that can characterize the comprehensive performance balance of candidate strategy points, avoiding bias caused by subjective selection. A Pareto scheduling strategy library is constructed by selecting the optimal proportion of candidate strategy points based on voxel distance values, preserving both strategy diversity and ensuring comprehensive strategy performance. Cosine similarity is used to achieve accurate matching between the current power grid operating state and candidate strategies. Combined with comprehensive cost selection, the final scheduling strategy is ensured to be suitable for real-time operating conditions and achieve optimal synergy between economy, low carbon emissions, and stability, improving the accuracy and adaptability of scheduling. The final scheduling strategy is decomposed into standardized active and reactive power output command sequences, which are issued to the control terminal for execution, forming a complete closed loop of optimization, selection, matching, and execution, improving the practicality and operability of scheduling.

[0067] like Figure 2 As shown, embodiments of the present invention also provide a multi-objective Pareto optimal scheduling system for novel power systems, comprising: The analysis module is used to collect operational data of the new power system, analyze the operational data using a pre-trained time series feature extraction network, and determine the source-load power probability distribution scenario set and the initial feasible solution space. The computation module is used to initialize the candidate policy population in the initial feasible solution space, calculate the mutual information entropy value between the gradient vectors of the multidimensional objective function of each candidate policy, and synthesize the comprehensive guiding potential vector pointing to the Pareto improvement direction based on the mutual information entropy value. The module is used to extract the first and second evolution anchor points from the initial feasible solution space. In each rolling scheduling cycle, it regenerates the source load power probability distribution scenario set based on the latest observation data and forecast data, and calculates the dispersion index of the scenario set in real time. Based on the comparison result of the dispersion index and the preset threshold, the corresponding evolution anchor point is selected to construct a local convergent hyperellipsoidal domain or a global exploratory hyperellipsoidal domain. The correction module is used to map the candidate strategy population to the hyperellipsoidal domain, calculate the Mahalanobis distance between each candidate strategy and the geometric center of the hyperellipsoidal domain, divide the population into subsets based on the Mahalanobis distance, and perform differential correction and search to obtain the hybrid evolutionary strategy set. The evolution module is used to superimpose the historical iteration velocity memory term as an inertial component with the comprehensive guiding potential field vector to calculate the state update vector, so as to drive the hybrid evolution strategy set to perform trajectory evolution in the initial feasible solution space and obtain the evolved non-dominated candidate strategy set. The control module performs voxel distance transformation and non-dominated sorting on the evolved non-dominated candidate strategy set to obtain the Pareto scheduling strategy library. It then matches the current power grid operating state with the matching degree of the candidate strategies in the Pareto scheduling strategy library and selects the final strategy to send to the control terminal.

[0068] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0069] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multi-objective Pareto optimal scheduling method for novel power systems, characterized in that, The method includes: Step 1: Collect operational data of the new power system, analyze the operational data using a pre-trained time series feature extraction network, and determine the source-load power probability distribution scenario set and the initial feasible solution space; Step 2: Initialize the candidate policy population in the initial feasible solution space, calculate the mutual information entropy value between the gradient vectors of the multidimensional objective function of each candidate policy, and synthesize the comprehensive guiding potential vector pointing to the Pareto improvement direction based on the mutual information entropy value. Step 3: Extract the first evolution anchor point and the second evolution anchor point from the initial feasible solution space. In each rolling scheduling cycle, regenerate the source load power probability distribution scenario set based on the latest observation data and forecast data, and calculate the dispersion index of the scenario set in real time. Based on the comparison result of the dispersion index and the preset threshold, select the corresponding evolution anchor point to construct a local convergent hyperellipsoidal domain or a global exploratory hyperellipsoidal domain. Step 4: Map the candidate strategy population to the hyperellipsoidal domain, calculate the Mahalanobis distance between each candidate strategy and the geometric center of the hyperellipsoidal domain, divide the population into subsets based on the Mahalanobis distance, and perform differential correction and search to obtain the hybrid evolutionary strategy set. Step 5: The historical iteration velocity memory term is superimposed on the integrated guiding potential field vector as an inertial component to calculate the state update vector, so as to drive the hybrid evolution strategy set to perform trajectory evolution in the initial feasible solution space and obtain the evolved non-dominated candidate strategy set. Step 6: Perform voxel distance transformation and non-dominated sorting on the evolved non-dominated candidate strategy set to obtain the Pareto scheduling strategy library. Match the current power grid operating state with the matching degree of the candidate strategies in the Pareto scheduling strategy library, select the final strategy and send it to the control terminal.

2. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 1, characterized in that, The operational data of the new power system includes historical source-load power time-series data, meteorological environment data, power grid topology parameters, unit ramp rate constraint data, node voltage safety boundary data, and equipment operating status data.

3. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 2, characterized in that, By analyzing the runtime data using a pre-trained temporal feature extraction network, the scenario set of source-load power probability distribution and the initial feasible solution space are determined, including: Construct a pre-trained temporal feature extraction network that includes a multi-head self-attention mechanism module, a graph convolutional neural network module, and a conditional generative adversarial network module; Historical source-load power time series data are input into the multi-head self-attention mechanism module. By calculating the dot product similarity of the query vector, key vector, and value vector, the long-period dependency features of the historical source-load power time series data in the time dimension are obtained, resulting in a time hidden state sequence. Equipment operating status data and power grid topology parameters are input into the graph convolutional neural network module. By aggregating the feature information of adjacent nodes, the spatial correlation features of the equipment operating status data in the spatial topology are obtained, resulting in a spatial hidden state sequence. The temporal hidden state sequence and the spatial hidden state sequence are concatenated and fused to obtain a spatiotemporal fusion feature vector. The spatiotemporal fusion feature vector is then used as a conditional input to the conditional generative adversarial network module. The generator in the conditional generative adversarial network module is used to obtain the source load power prediction scene. The discriminator is used to distinguish the generated scene from the real historical scene. The generator parameters are optimized through adversarial training to obtain a set of source load power probability distribution scenes containing confidence intervals. Based on the source-load power probability distribution scenario set, a robust optimization uncertainty set is constructed. Combined with the power grid topology parameters, unit ramp rate constraints, and node voltage safety boundaries, invalid strategy points are eliminated through constraint projection operations to define the initial feasible solution space that satisfies the physical safety operation of the system.

4. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 3, characterized in that, Step 2 includes: Randomly initialize the candidate strategy population within the initial feasible solution space, and define the economic objective function, the low-carbon objective function, and the stability objective function; The gradient vector of each individual in the candidate strategy population is calculated under the economic objective function, the low-carbon objective function, and the stability objective function respectively; based on the gradient vector, the mutual information entropy between each pair of objective functions is calculated to quantify the coupling degree and conflict intensity between objectives. Based on the coupling degree and conflict intensity, a dynamic adaptive weight matrix is ​​obtained. The gradient vector is then weighted and summed using the dynamic adaptive weight matrix to synthesize a comprehensive guiding potential vector pointing towards the Pareto improvement direction.

5. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 4, characterized in that, Extracting the first and second evolutionary anchor points from the initial feasible solution space, including: Calculate the geometric centroid of all candidate strategy points in the initial feasible solution space, and use the geometric centroid as the first evolution anchor point for the corresponding system baseline load condition; identify the boundary candidate strategy point with the largest Euclidean distance from the geometric centroid in the initial feasible solution space, and use the boundary candidate strategy point as the second evolution anchor point for the corresponding extreme disturbance boundary condition.

6. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 5, characterized in that, In each rolling scheduling cycle, a new source-load power probability distribution scenario set is generated based on the latest observation and forecast data. The dispersion index of this scenario set is calculated in real time. Based on the comparison between the dispersion index and a preset threshold, corresponding evolution anchor points are selected to construct a locally convergent hyperellipsoidal domain or a globally exploratory hyperellipsoidal domain, including: At the beginning of each rolling scheduling cycle, based on the latest observation data and forecast data, a new source load power probability distribution scenario set is generated, and the variance integral of each scenario curve in the source load power probability distribution scenario set relative to the mean curve is calculated to obtain the dispersion index. Determine whether the dispersion index is less than a preset threshold. When the dispersion index is less than the preset threshold, it indicates that the source load fluctuation is in a stable state. With the first evolution anchor point as the center and the preset first covariance matrix as the shape parameter, construct a locally convergent hyperellipsoidal domain. When the dispersion index is greater than or equal to the preset threshold, it indicates that the source load fluctuation is in a state of severe disturbance. With the second evolution anchor point as the center and the preset second covariance matrix as the shape parameter, a global exploration hyperellipsoidal domain is constructed.

7. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 6, characterized in that, Step 4 includes: The candidate strategy population is projected onto the hyperellipsoid domain, and the Mahalanobis distance between each candidate strategy point and the geometric center of the hyperellipsoid domain is calculated. Candidate strategy points with Mahalanobis distance less than the Mahalanobis distance threshold are divided into core dense subsets, and candidate strategy points with Mahalanobis distance greater than or equal to the Mahalanobis distance threshold are divided into marginal sparse subsets. For the core dense subset, the second-order partial derivative approximation matrices of the economic objective function, the low-carbon objective function, and the stability objective function are calculated. The curvature correction of the gradient direction is performed using the second-order partial derivative approximation matrices, and the positions of each candidate strategy point in the core dense subset are updated to accelerate local convergence. For the marginal sparse subset, a random step size vector following a heavy-tailed distribution is determined. The random step size vector is used to drive the candidate strategy points to perform long-distance position jumps, and the positions of each candidate strategy point in the marginal sparse subset are updated to maintain population diversity. The updated core dense subset and the updated marginal sparse subset are merged to obtain the hybrid evolutionary strategy set.

8. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 7, characterized in that, Step 5 includes: Obtain the velocity vector of the hybrid evolution strategy set in the previous iteration cycle, and use the velocity vector as the historical iteration velocity memory term; combine the historical iteration velocity memory term with the preset inertia weight coefficient to obtain the inertia component; and perform vector superposition of the inertia component with the comprehensive guiding potential field vector to obtain the state update vector. The position coordinates of each candidate strategy point in the hybrid evolution strategy set are updated according to the state update vector to realize trajectory evolution; the boundary reflection operation is performed on the evolved candidate strategy points to reflect the candidate strategy points that exceed the boundary of the initial feasible solution space back into the initial feasible solution space; the non-dominated candidate strategy set is subjected to non-dominated screening to remove dominated individuals, and the evolved non-dominated candidate strategy set is obtained.

9. The multi-objective Pareto optimal scheduling method for novel power systems according to claim 8, characterized in that, Step 6 includes: Fast non-dominated sorting is performed on the evolved non-dominated candidate policy set to divide the frontiers into different dominance levels. For all candidate policy points within the same frontier, the corresponding multidimensional objective function values ​​are mapped to the target space, and a three-dimensional voxel grid covering all candidate policy points of the corresponding frontier is established in the target space, with each voxel being a cube of equal size. All voxel units are traversed, and voxel units containing at least one candidate policy point are marked as occupied voxels, and the rest are marked as empty voxels. From all candidate policy points within the corresponding front surface, identify the occupied voxels corresponding to the extreme points located on the boundary of the corresponding front surface, and use the occupied voxels as the source point set of the distance transformation. Using a multi-source parallel propagation method, calculate the shortest three-dimensional Manhattan distance from each occupied voxel to the source point set of the distance transformation layer by layer to obtain the distance transformation field corresponding to the corresponding front surface. For the case where the same occupied voxel contains multiple candidate policy points, take the distance value of the distance transformation field at the center of the corresponding voxel as the common voxel distance value of all candidate policy points within the corresponding voxel. Within the same frontal plane, candidate strategy points are sorted in descending order of voxel distance values. A predetermined proportion of candidate strategy points are selected from the sorted sequence to obtain the Pareto scheduling strategy library. Real-time snapshots of the current power grid operating status are collected, and three corresponding objective function reference values ​​are calculated based on these snapshots to form a three-dimensional reference vector of the current power grid operating status. The cosine similarity between the three-dimensional reference vector and the three-dimensional objective function vectors of each candidate strategy point in the Pareto scheduling strategy library is calculated as the matching degree. The candidate strategy point with the highest matching degree and the lowest overall cost is selected as the final scheduling strategy. The final scheduling strategy is decomposed into active power output command sequences and reactive power output command sequences for each control unit and sent to the control terminal for execution.

10. A multi-objective Pareto optimal scheduling system for novel power systems, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The analysis module is used to collect operational data of the new power system, analyze the operational data using a pre-trained time series feature extraction network, and determine the source-load power probability distribution scenario set and the initial feasible solution space. The computation module is used to initialize the candidate policy population in the initial feasible solution space, calculate the mutual information entropy value between the gradient vectors of the multidimensional objective function of each candidate policy, and synthesize the comprehensive guiding potential vector pointing to the Pareto improvement direction based on the mutual information entropy value. The module is used to extract the first and second evolution anchor points from the initial feasible solution space. In each rolling scheduling cycle, it regenerates the source load power probability distribution scenario set based on the latest observation data and forecast data, and calculates the dispersion index of the scenario set in real time. Based on the comparison result of the dispersion index and the preset threshold, the corresponding evolution anchor point is selected to construct a local convergent hyperellipsoidal domain or a global exploratory hyperellipsoidal domain. The correction module is used to map the candidate strategy population to the hyperellipsoidal domain, calculate the Mahalanobis distance between each candidate strategy and the geometric center of the hyperellipsoidal domain, divide the population into subsets based on the Mahalanobis distance, and perform differential correction and search to obtain the hybrid evolutionary strategy set. The evolution module is used to superimpose the historical iteration velocity memory term as an inertial component with the comprehensive guiding potential field vector to calculate the state update vector, so as to drive the hybrid evolution strategy set to perform trajectory evolution in the initial feasible solution space and obtain the evolved non-dominated candidate strategy set. The control module performs voxel distance transformation and non-dominated sorting on the evolved non-dominated candidate strategy set to obtain the Pareto scheduling strategy library. It then matches the current power grid operating state with the candidate strategies in the Pareto scheduling strategy library and selects the final strategy to send to the control terminal.

Images