Power distribution network planning method and system considering distributed energy uncertainty

Abstract

The invention discloses a power distribution network planning method and system considering distributed energy uncertainty, and relates to the technical field of power grid planning, and the method comprises the steps: collecting distributed energy node data, compensating space-time migration in combination with meteorological data, and carrying out the time sequence alignment through a dynamic time warping algorithm; constructing an improved Wasserstein generative adversarial network to generate a conventional scene, and injecting Gaussian noise through potential spatial disturbance to generate an extreme scene deviating from training distribution; inputting the mixed scene set into a mixed integer nonlinear programming model, and adopting a graph neural network to establish a topology-power flow agent model to accelerate solution; updating line impedance parameters through a Kalman filter, and collecting and checking actual output; and decomposing the corrected planning scheme into cloud global optimization and edge local control, and carrying out cloud-edge collaboration. According to the method, the adaptability of a power distribution network planning scheme in a complex and uncertain environment is improved by combining spatial-temporal feature alignment, adversarial network scene enhancement, a graph neural network and cloud edge collaborative optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid planning, and particularly to a distribution network planning method and system considering the uncertainty of distributed energy sources. Background Art

[0002] With the wide access of distributed energy sources, modern distribution networks are developing towards highly intelligent and distributed coordinated control. However, due to the randomness and volatility of renewable energy sources such as wind energy and photovoltaic energy, distribution network operation faces problems such as voltage over-limit, reverse power flow, and equipment overload. In addition, traditional distribution network planning methods are usually based on static load forecasting and deterministic optimization models, and fail to fully consider the random volatility of distributed energy sources, resulting in poor adaptability of the planning scheme in actual operation. Therefore, how to effectively quantify and address the uncertainty of distributed energy sources in the distribution network planning stage has become a key problem to be solved urgently.

[0003] In dealing with the uncertainty of distributed energy sources, existing research mainly adopts methods such as Monte Carlo simulation, scenario analysis, and robust optimization. However, the Monte Carlo method has a high computational cost and is difficult to apply to large-scale distribution networks; traditional scenario generation methods cannot effectively capture extreme operating conditions, resulting in insufficient robustness of the optimization scheme; and the optimization method based on deterministic equivalence has limitations in characterizing dynamic topology adjustment and energy storage scheduling strategies. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a distribution network planning method considering the uncertainty of distributed energy sources to solve the problem that traditional scenario generation methods cannot effectively capture extreme operating conditions.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a distribution network planning method considering the uncertainty of distributed energy sources, which includes collecting distributed energy node data, compensating for spatio-temporal offsets by combining meteorological data, performing time series alignment using the dynamic time warping algorithm, constructing a spatio-temporal feature matrix, extracting local spatio-temporal patterns through a sliding window, and outputting a high-dimensional feature dataset; Constructing an improved Wasserstein generative adversarial network to generate conventional scenarios, injecting Gaussian noise through potential space perturbation to generate extreme scenarios deviating from the training distribution, and integrating them into a mixed scenario set; Inputting the mixed scenario set into a mixed integer non-linear programming model, using a graph neural network to establish a topology-power flow proxy model for accelerated solution, outputting the predicted value of the line overload probability, and performing branch and bound pruning to obtain an energy storage configuration scheme and a dynamic topology strategy library; Update the line impedance parameters through a Kalman filter, collect the actual output and conduct verification, and adopt rolling horizon optimization to adjust the energy storage configuration plan and the topological switching sequence to generate a revised planning plan; Decompose the revised planning plan into cloud - side global optimization and edge - side local control for cloud - edge collaboration.

[0007] As a preferred scheme of the distribution network planning method considering the uncertainty of distributed energy sources of the present invention, wherein: the compensation for spatio - temporal offset includes determining the cloud movement speed according to a meteorological radar, and applying a time - offset compensation based on the coordinate difference and the cloud movement speed to the output time - series data of adjacent nodes.

[0008] As a preferred scheme of the distribution network planning method considering the uncertainty of distributed energy sources of the present invention, wherein: in the improved Wasserstein generative adversarial network, the generator receives a latent noise vector and a meteorological condition label and generates an output scenario that conforms to spatio - temporal correlation; the discriminator adds a power - balance constraint loss term and calculates the power - flow violation penalty of the generated scenario to make the generated scenario satisfy Kirchhoff's laws.

[0009] As a preferred scheme of the distribution network planning method considering the uncertainty of distributed energy sources of the present invention, wherein: in the mixed - integer non - linear programming model, the decision variables include the energy storage location, capacity, line expansion plan, and binary topology matrix; the objective function is to minimize the life - cycle cost, and the constraint conditions are embedded with a linearized DisFlow power - flow equation and a radial topology verification rule.

[0010] As a preferred scheme of the distribution network planning method considering the uncertainty of distributed energy sources of the present invention, wherein: the constraint conditions of the binary topology matrix include radial constraint, connectivity constraint, and switch operation times limit.

[0011] As a preferred scheme of the distribution network planning method considering the uncertainty of distributed energy sources of the present invention, wherein: the construction of the graph neural network includes, The node feature vector includes the photovoltaic installed capacity, load prediction value, and energy storage SOC state; The edge feature vector includes the line impedance, initial current - carrying capacity, and expansion candidate identifier; The output layer uses a sigmoid activation function to predict the line overload probability.

[0012] As a preferred scheme of the distribution network planning method considering the uncertainty of distributed energy sources of the present invention, wherein: the cloud - side global optimization refers to updating the long - term energy storage capacity plan based on the Benders decomposition method and sending the topology strategy library to each edge node through the network; the edge - side local control refers to deploying a lightweight reinforcement learning controller to collect local operation parameters in real time and execute topological switching.

[0013] In a second aspect, the present invention provides a distribution network planning system considering the uncertainty of distributed energy, including a feature extraction module, a scenario mixing module, a scheme generation module, a planning correction module, and a cloud-edge collaboration module; The feature extraction module is used to collect distributed energy node data, compensate for spatio-temporal offsets by combining meteorological data, perform time series alignment using the dynamic time warping algorithm, construct a spatio-temporal feature matrix, extract local spatio-temporal patterns through a sliding window, and output a high-dimensional feature dataset; The scenario mixing module is used to construct an improved Wasserstein generative adversarial network to generate normal scenarios, inject Gaussian noise through latent space perturbation to generate extreme scenarios of the training distribution, and integrate them into a mixed scenario set; The scheme generation module is used to input the mixed scenario set into a mixed-integer non-linear programming model, establish a topology-power flow proxy model using a graph neural network to accelerate the solution, output the predicted value of the line overload probability, perform branch and bound pruning, and obtain a energy storage configuration scheme and a dynamic topology strategy library; The planning correction module is used to update the line impedance parameters through a Kalman filter, collect the actual output and perform verification, and adopt rolling horizon optimization to adjust the energy storage configuration scheme and the topology switching sequence to generate a corrected planning scheme; The cloud-edge collaboration module is used to decompose the corrected planning scheme into global optimization in the cloud and local control at the edge for cloud-edge collaboration.

[0014] In a third aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, and: when the computer program is executed by the processor, any step of the distribution network planning method considering the uncertainty of distributed energy as described in the first aspect of the present invention is implemented.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and: when the computer program is executed by the processor, any step of the distribution network planning method considering the uncertainty of distributed energy as described in the first aspect of the present invention is implemented.

[0016] The beneficial effects of the present invention are as follows: By collecting data of distributed energy nodes and compensating for spatio-temporal offsets in combination with meteorological data, the dynamic time warping algorithm is used to accurately align time series and extract local patterns, constructing a high-dimensional feature dataset, providing a precise basis for subsequent scenario generation; An improved Wasserstein generative adversarial network is used to generate normal scenarios, and extreme scenarios are generated by injecting Gaussian noise through latent space perturbation, achieving comprehensive coverage of normal and extreme working conditions and enhancing the robustness of the method to uncertainty; The mixed scenario set is input into the mixed-integer non-linear programming model, and a graph neural network is used to construct a topology-power flow proxy model to achieve rapid prediction of line overload probability and branch-and-bound pruning, accelerating the solution of the optimal energy storage configuration and dynamic topology strategy; The line parameters are updated in real time through a Kalman filter, and the energy storage configuration and topology switching sequence are dynamically verified and corrected in combination with rolling horizon optimization; It is decomposed into cloud global optimization and edge local control, constituting a cloud-edge collaborative mechanism, thereby overall improving the economic, safe and intelligent level of distribution network planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a flowchart of a distribution network planning method considering the uncertainty of distributed energy in Embodiment 1.

[0019] Figure 2 It is a structural diagram of an improved Wasserstein generative adversarial network in Embodiment 1.

[0020] Figure 3 It is a collaborative solution model diagram of mixed-integer programming and graph neural network in Embodiment 1.

[0021] Figure 4 It is a cloud-edge collaborative optimization architecture diagram in Embodiment 1. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to make the above objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the drawings in the specification.

[0023] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways than those specifically described herein, and those skilled in the art can make similar generalizations without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0024] Secondly, as used herein, "an embodiment" or "embodiments" refer to specific features, structures, or characteristics that may be included in at least one implementation of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that are mutually exclusive of other embodiments.

[0025] Embodiment 1, referring to Figures 1 to 4 , is the first embodiment of the present invention. This embodiment provides a distribution network planning method considering the uncertainty of distributed energy, including the following steps: S1: Collect distributed energy node data, compensate for spatio-temporal offsets by combining meteorological data, perform time series alignment using the dynamic time warping algorithm, construct a spatio-temporal feature matrix, extract local spatio-temporal patterns through a sliding window, and output a high-dimensional feature dataset.

[0026] Specifically, it includes the following steps: S1.1: Collect distributed energy node data, meteorological data, and power grid topology data.

[0027] Specifically, collect distributed energy node data of photovoltaic / wind farms in the area, with a sampling frequency of 5 minutes / time, a time stamp accuracy of ±30 ms, and a data format of [time stamp, node ID, active power, reactive power].

[0028] Collect meteorological radar images through a meteorological radar to obtain meteorological data, with a spatial resolution of irradiance (W / m²), wind speed (m / s), and cloud optical thickness of 1 km × 1 km, a time resolution of 15 minutes, and a data format of [time stamp, grid coordinates, meteorological parameters].

[0029] Import the power distribution network model in CIM / E format and extract the power grid topology data, including node coordinates (longitude and latitude), line impedance matrix, transformer turns ratio, historical curve of load nodes (kW), line cost, and unit energy storage capacity.

[0030] S1.2: Perform spatio-temporal offset compensation.

[0031] Specifically, based on three consecutive frames of meteorological radar images, use the optical flow method to calculate the cloud movement speed, and the expression is: ; Where is the cloud movement speed, is the number of matching feature clouds, is the index of the number of feature clouds, is the th displacement vector of the feature cloud between adjacent frames, is the scanning interval time of the weather radar.

[0032] For any two distributed energy nodes and , calculate the time offset between the distributed energy nodes according to their node coordinates. The expression is: ; where is the projection distance along the cloud movement direction, is the abscissa of the distributed energy node, is the ordinate of the distributed energy node, is the abscissa of another distributed energy node, is the ordinate of another distributed energy node.

[0033] The time offset is the ratio of the projection distance along the cloud movement direction to the cloud movement speed. When the distributed energy node is located downstream of the cloud movement, the output sequence needs to be compensated forward for the duration of the time offset.

[0034] S1.3: Use the dynamic time warping algorithm for time series alignment.

[0035] It should be noted that the sequence alignment process is as follows: taking the location of the power grid dispatching center as the reference node, using the original uncompensated distributed energy node data as the reference sequence, calculating the cost matrix between the compensated sequence and the reference sequence, performing backward search, recording the path that minimizes the cumulative cost, and performing non-linear stretching along the path to make its length consistent with the reference sequence.

[0036] Furthermore, the aligned distributed energy node data is spatially interpolated and matched with the meteorological data: for each distributed energy node , take the weighted average of all meteorological monitoring points within the 1km×1km grid where it is located, and the weight is the inverse square of the distance from the node coordinates.

[0037] S1.4: Construct a spatio-temporal feature matrix and extract local spatio-temporal patterns through a sliding window.

[0038] Specifically, construct a three-dimensional feature tensor, including the time dimension, the number of nodes, and the number of feature channels. The number of feature channels includes distributed energy node data, meteorological data, and power grid topology data.

[0039] Overlapping windows are used for local feature extraction. The window length is set to 72 hours and the sliding step is set to 6 hours. Data enhancement is performed by applying random time reversal and Gaussian noise to each window.

[0040] S1.5: Output high-dimensional feature dataset.

[0041] Specifically, a node output-meteorological correlation test is performed, and the Pearson correlation coefficient of each distributed energy node is calculated. When it is less than 0.7, a manual review is triggered.

[0042] The final output high-dimensional feature dataset includes the main data body and metadata. The main data body contains the spatiotemporal feature matrix stored in HDF5 format, and the metadata contains node coordinates and topological connection relationships (adjacency matrix), timestamps and meteorological scene labels (sunny / cloudy / rainstorm), and standardized parameters (mean and standard deviation of each feature). Batch loading is achieved through PyTorch's DataLoader.

[0043] Preferably, by collecting distributed energy node data, meteorological data and power grid topology data, the comprehensiveness and accuracy of the input are ensured; the optical flow method is used to calculate the cloud movement speed, and the time offset is calculated based on the node coordinates to achieve the space-time offset compensation of the distributed energy node data and improve the space-time consistency; the dynamic time warping algorithm is used to align the distributed energy node data with the power grid dispatching center as the benchmark, and the meteorological data is matched by spatial interpolation to ensure the accurate alignment of the time series and the reasonable mapping of spatial information; the three-dimensional space-time feature matrix is ​​further constructed, and the local space-time pattern is extracted through the sliding window, and the space-time robustness of the space-time feature matrix is ​​enhanced by combining random time reversal and Gaussian noise; the high-dimensional feature data set finally output not only contains the space-time feature matrix, but also comes with node topology information, timestamp labels and standardized parameters, and supports efficient loading. It improves the adaptability to the uncertainty of distributed energy in distribution network planning and provides high-quality input support for subsequent deep learning modeling and optimization.

[0044] S2: Construct an improved Wasserstein generative adversarial network to generate regular scenes, inject Gaussian noise into the latent space perturbation to generate extreme scenes that deviate from the training distribution, and integrate them into a mixed scene set.

[0045] Specifically, the following steps are included: S2.1: Construct an improved Wasserstein generative adversarial network.

[0046] S2.1.1: Build the generator structure.

[0047] Specifically, the output layer includes a latent noise vector with a dimension of 128, which follows a standard Gaussian distribution; the meteorological condition label encodes the meteorological types in the next 24 hours (such as sunny, cloudy, heavy rain), and is mapped to a 32-dimensional vector through an embedding layer.

[0048] The middle layer uses a 3-layer Temporal Convolutional Network (TCN). The width of the convolutional kernel in each layer is 5, and the number of channels is 256, 128, and 64 in sequence. Then a Gated Recurrent Unit (GRU) is connected to capture temporal dependencies.

[0049] The output layer outputs the spatio-temporal output scenario, representing the active power generation values of the nodes in the next 288 time steps (5-minute resolution, 24 hours). The activation function is Sigmoid (normalized to [0, 1]).

[0050] S2.1.2: Construct the discriminator structure.

[0051] Specifically, the input layer is the real / generated scenario, with the same dimension as the generator output.

[0052] The middle layer is a Graph Convolutional Network (GCN) layer: aggregating the node output and topological relationship, with an output dimension of 128. A temporal attention mechanism is used to calculate the importance weights of each time step.

[0053] The output layer is the Wasserstein distance score, representing the authenticity of the scenario; and the physical constraint loss term, calculating the degree of power balance violation.

[0054] S2.2: Construct an improved Wasserstein Generative Adversarial Network.

[0055] Specifically, the expression of the Wasserstein loss function is: ; where is the Wasserstein loss function, is the Wasserstein distance score output by the discriminator, is the real scenario, is the scenario generated by the generator, is the expected calculation.

[0056] Adding a constraint condition, the discriminator weights need to satisfy 1-Lipschitz continuity, which is achieved through a gradient penalty term (GP), and the expression is: ; where is the gradient penalty loss, is the gradient penalty coefficient (taking 10), is the linear interpolation sample of the real and generated scenarios, is the discriminator for The calculated gradient represents the rate of change of the discriminator.

[0057] Add the physical constraint loss, and the expression is: ; Where is the physical constraint departure loss, is the weight of the physical constraint loss, which determines the influence degree of this loss term on optimization (take 1, and fine-tune through experiments), is the historical predicted power of the conforming node, is the power balance operator based on the power grid topology (equivalent to Kirchhoff's current law), is the real scenario.

[0058] The optimization objective of the discriminator is: ; Where is to take the minimum value.

[0059] The optimization objective of the generator is: ; S2.3: Conduct the training of the generative adversarial network model.

[0060] Specifically, divide the high-dimensional feature dataset into a training set (80%) and a validation set (20%), and perform normalization and standardization.

[0061] Use alternating training. In each round of iteration, the discriminator is updated 5 times and the generator is updated 1 time.

[0062] The optimizer uses Adam, and the learning rate is set to The momentum parameters are set to 0.5 and 0.9 respectively.

[0063] Terminate the training when the Wasserstein loss on the validation set has not decreased for 10 consecutive rounds.

[0064] S2.4: Generate regular scenarios and extreme scenarios, and integrate them into a mixed scenario set.

[0065] Specifically, sample latent variables from the standard Gaussian distribution, and combine with meteorological labels to generate regular output scenarios that conform to the training data distribution.

[0066] Furthermore, by calculating the principal component direction (corresponding to the largest eigenvalue) of the latent space and adding perturbations, conditionally constrain the perturbed scenarios to ensure that they meet the extreme characteristics (such as continuous low processing in the whole region) to generate extreme scenarios.

[0067] Regular scenario production Samples covering typical output patterns under historical meteorological conditions. Extreme scenario generation Samples, including a sudden drop in output across the entire region (simulating cloudy / heavy rain) and severe fluctuations in local output (simulating strong winds / rapid cloud movement).

[0068] Integrate the normal scenarios and extreme scenarios, and conduct similarity statistics through the Kolmogorov-Smirnov test; and verify the extreme coverage rate, requiring that the lower limit of the output of the extreme scenarios is lower than 20% of the historical minimum value.

[0069] Preferably, construct a generator based on TCN and GRU, a discriminator combining GCN and temporal attention mechanism to achieve efficient spatio-temporal modeling. Use the Wasserstein loss to improve training stability, gradient penalty (GP) to ensure 1-Lipschitz continuity, and add physical constraint loss to ensure that the generated results satisfy the power balance constraint. The training adopts an alternating optimization strategy, and the Adam optimizer is used to update the parameters, and the convergence is monitored on the validation set. In the generation stage, first sample the standard Gaussian distribution variables, combine with meteorological labels to generate normal scenarios; then inject perturbations based on the principal component analysis (PCA) direction to construct extreme scenarios. Generate 10 4 normal samples and 500 extreme samples to ensure covering extreme cases below 20% of the historical minimum output. Evaluate the distribution consistency through the Kolmogorov-Smirnov test, so that the mixed scenario set can improve the coverage rate of extreme cases while maintaining authenticity, providing reliable data support for power grid dispatching.

[0070] S3: Input the mixed scenario set into the mixed-integer non-linear programming model, use the graph neural network to establish a topology-power flow surrogate model to accelerate the solution, output the predicted value of the line overload probability, perform branch and bound pruning, and obtain the energy storage configuration plan and the dynamic topology strategy library.

[0071] Specifically, it includes the following steps: S3.1: Construct a mixed-integer non-linear programming (MINLP) model.

[0072] Specifically, based on the power grid topology data, define the objective function of the MINLP model, and the expression is: ; Among them, is the set of candidate energy storage installation nodes, is the index of the candidate energy storage installation node, is the candidate branch for line capacity expansion, is the index of the candidate energy storage installation node for capacity expansion, is the unit capacity cost of the energy storage, is the line unit capacity expansion cost, is the penalty coefficient for the lack of power supply is the candidate energy storage installation node of the energy storage capacity is the line of the expansion times is the expected lack of power supply.

[0073] The decision variables include the candidate energy storage installation node of the energy storage location coordinates, capacity , the line expansion times and the binary topology matrix, where a value equal to 1 indicates that the branch is closed. The energy storage location coordinates of the candidate energy storage installation node are obtained from the power grid topology data.

[0074] The linearized DistFlow power flow equation has the expression: ; where is the active power of the branch ; is the set of downstream nodes of the branch node ; is the index of the downstream node set is the active power of the branch ; is the resistance of the branch ; is the reactive power of the branch ; is the voltage of the branch node ; is the load power of the branch node ; is the distributed power generation output of the branch node .

[0075] Ignoring the second-order terms, it is simplified to a linear equation with the expression: ; Furthermore, control the radial topology constraint, line capacity constraint, and energy storage operation constraint.

[0076] The radial topology constraint includes the connectivity constraint and the acyclic constraint. The connectivity constraint ensures that the network is connected and acyclic, which is expressed by the spanning tree constraint; the acyclic constraint ensures that for each loop, at least one branch is disconnected.

[0077] The line capacity constraint needs to limit the upper limit of the line current-carrying capacity after expansion.

[0078] The energy storage operation constraint needs to limit the dynamics and boundaries of the capacity.

[0079] S3.2: Establish a Graph Neural Network (GNN) surrogate model.

[0080] Specifically, the input layer includes node features and edge features. The node features include load power, PV capacity, and energy storage SOC, and the edge features include line impedance, initial current-carrying capacity, and expansion flag.

[0081] The graph convolutional layer aggregates neighborhood information and updates node embeddings.

[0082] The output layer performs edge-level prediction and outputs the overload probability for each edge through Sigmoid.

[0083] By randomly sampling the topological structure and load scenarios, a power flow calculator (such as OpenDSS) is called to generate label data (overload is 1, otherwise 0). The loss function uses binary cross-entropy loss.

[0084] S3.3: Accelerate the solution by the branch and bound method.

[0085] Specifically, the pruning strategy is as follows: During the branching process, the GNN is called to predict the line overload probability for the current partial solution. If there is a line whose predicted line overload probability is greater than the current upper bound, it is determined that all solutions under this subtree are infeasible, and this branch is pruned.

[0086] The solution process is as follows: Add the complete problem to the queue of nodes to be explored, and set the upper bound of the optimal solution to infinity. Use depth-first search to preferentially process nodes with lower objective values of the relaxed problem. Continuously relax the integer variables and solve the linear programming sub-problem. If the objective value of the relaxed solution exceeds the current upper bound, or the GNN predicts a high overload risk, prune. Select the non-integer variables to generate child nodes and recursively process until the queue is empty.

[0087] S3.4: Output the energy storage configuration plan and the dynamic topology strategy library.

[0088] Specifically, the energy storage configuration plan includes the optimal location and capacity, satisfying a voltage deviation of <5% in 90% of the scenarios.

[0089] The dynamic topology strategy library includes base load scenarios, closing the main trunk lines and disconnecting redundant branches to reduce losses; peak load scenarios, enabling expansion lines and adjusting the topology to balance the load distribution; fault scenarios, presetting island operation topologies to isolate the fault area.

[0090] The energy storage configuration plan and the dynamic topology strategy library are stored in JSON format.

[0091] Preferably, through the constructed MINLP model, combined with the linearized power flow equation, the energy storage configuration and dynamic topology adjustment are optimized to minimize the energy storage cost, line expansion cost, and power shortage. The GNN surrogate model is used to learn the power flow state through node features and edge features, quickly predict the line overload probability, and improve the calculation efficiency. The branch and bound method is used to accelerate the global optimization solution. An energy storage configuration plan and a dynamic topology strategy library that meet 90% of the scenario constraints are output to achieve the adaptive adjustment of the distribution network under the base load, peak load, and fault scenarios. Considering both global optimality and calculation efficiency, the calculation complexity is reduced compared with traditional optimization methods, and the safety, economy, and reliability of power grid operation are improved.

[0092] S4: Update the line impedance parameters through the Kalman filter, collect the actual output and conduct verification, and adopt rolling horizon optimization to adjust the energy storage configuration plan and the topology switching sequence to generate a revised planning plan.

[0093] Specifically, it includes the following steps: S4.1: Update the dynamics of line impedance parameters.

[0094] Specifically, the measurement data of SCADA are accessed in real time, including the node voltage amplitude (V), current amplitude (I), active / reactive power (P / Q), with a time resolution of the order of 1 second, and high-precision synchronization is achieved through the PMU (synchronized phasor measurement unit).

[0095] State estimation is performed through the Kalman filter, and the line impedance parameters are defined as dynamic state variables, including resistance and reactance. An observation model is constructed based on the measurement data (such as node voltage difference and line power) to reflect the influence of impedance parameters on electrical quantities.

[0096] In the prediction stage, the current moment parameters are predicted based on historical impedance values, considering slow-changing factors such as equipment aging and temperature drift. In the update stage, the latest measurement data is used to correct the predicted value, and the impedance estimation value is adjusted by minimizing the prediction error covariance matrix.

[0097] The updated line impedance parameters are output for correcting the power flow calculation model.

[0098] S4.2: Verification of actual output and deviation detection.

[0099] Specifically, the actual output data (photovoltaic / wind power) is aligned in time and space with the hybrid scenario set and matched according to the same time stamp and node granularity. The mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient are calculated to evaluate the consistency between the actual output and the scenario prediction. If the MAE exceeds 20% or the correlation coefficient is lower than 0.7, it is determined as a significant deviation, and the sample is discarded.

[0100] S4.3: Rolling horizon optimization adjustment.

[0101] It should be understood that the rolling horizon optimization adjusts the nonlinear constraints (such as the DistFlow equation) by piecewise linearization, transforming them into a mixed-integer linear programming (MILP) problem. Initialize variables based on the historical optimal solution to accelerate the convergence of the solver. Solve in parallel for different time domain windows to shorten the calculation time to within 5 minutes.

[0102] S4.4: Generate and verify the revised planning scheme.

[0103] Specifically, the revised planning scheme includes the revised energy storage scheduling strategy and the topology switching sequence.

[0104] Conduct digital twin verification, simulate the operation in the next 24 hours, and check indicators such as voltage stability and line load rate.

[0105] Preferably, utilize the high-precision measurement data of SCADA, update the line impedance parameters in real time through Kalman filtering to improve the accuracy of power flow calculation. Check the actual output data of photovoltaic, wind power, etc., detect deviations and eliminate abnormal data to ensure the reliability of the optimization input. Adopt rolling horizon optimization to piecewise linearize complex nonlinear constraints, initialize the optimization variables based on the historical solution, improve the MILP solution efficiency, and shorten the calculation time to within 5 minutes. Generate a revised energy storage scheduling and topology adjustment scheme, and verify its effectiveness in aspects such as voltage stability and line load balancing through digital twin simulation of the operation in the next 24 hours.

[0106] S5: Decompose the revised planning scheme into cloud-based global optimization and edge-side local control for cloud-edge collaboration.

[0107] Specifically, it includes the following steps: S5.1: Cloud-based global optimization.

[0108] Specifically, input the revised planning scheme in the cloud, update the energy storage capacity planning and topology strategy library based on long-term historical data and global information to ensure the economy of the whole network.

[0109] Apply the Benders decomposition method. The master problem is the long-term energy storage capacity planning (integer variables), minimizing the total life cycle cost (energy storage investment + power shortage penalty). The sub-problem is to verify the feasibility of the solution of the master problem (such as power flow constraints, voltage stability), and generate Benders cuts (feasibility / optimality constraints).

[0110] The cloud solves the master problem to generate an initial solution for energy storage configuration. Send the initial solution to each edge node, call the digital twin to simulate the operation, and check for constraint violations. If there are violations (such as line overload), generate Benders cuts and feedback them to the master problem, and re-solve until convergence.

[0111] Based on the topology switching effect data (such as loss reduction ratio, switch life consumption) fed back by the edge side, dynamically adjust the weight parameters in the policy library. Send update packages (in JSON format) through the 5G network, and use digital signatures and hash checks to prevent tampering.

[0112] S5.2: Edge-side local control.

[0113] Specifically, the inputs on the edge side are local SCADA measurement data, the topology policy library issued by the cloud, and the energy storage scheduling baseline. Perform topology switching and energy storage scheduling with a millisecond-level response speed to ensure local power supply quality.

[0114] The state space of the lightweight reinforcement learning controller is the local load power, photovoltaic output, energy storage SOC, and voltage deviation. The action space is the combination of topology switch states and energy storage charge / discharge power commands. The main reward of the reward function is to minimize the power deviation, and the penalty terms are the number of switch actions and the over-limit of the energy storage charge / discharge depth.

[0115] The edge side fine-tunes the policy based on real-time data using the PPO (Proximal Policy Optimization) algorithm.

[0116] The execution of topology switching includes matching similar scenarios from the policy library to obtain a set of candidate topologies. Call the local GNN model to predict the overload risks of each candidate topology and screen for safe solutions. Execute switch actions and record operation logs and feedback them to the cloud.

[0117] S5.3: Cloud-edge collaborative closed-loop mechanism.

[0118] It should be noted that the collaborative optimization process is as follows: the edge side detects local anomalies (such as sudden output changes), triggers the controller to make decisions and execute actions. The abnormal data and processing results are uploaded to the cloud in real time. The cloud aggregates the data of multiple edge nodes, checks for global constraint conflicts (such as the main network being overloaded due to simultaneous expansion of multiple nodes). Update the Benders master problem based on the conflict results, re-solve it, and issue global correction instructions.

[0119] Preferably, the cloud optimizes the long-term energy storage capacity planning and topology policy library through the Benders decomposition method to ensure the economy of the entire network, and continuously adjusts and optimizes the solution in combination with the feedback from the edge side. The edge side, based on SCADA measurement data and the policies issued by the cloud, uses reinforcement learning to quickly make decisions on topology adjustment and energy storage scheduling, achieving a millisecond-level response. Dynamically fine-tune the control policy through the PPO algorithm, and use GNN to predict the safety of topology switching to ensure that the lines are not overloaded. The cloud-edge collaborative closed-loop mechanism ensures a quick response to local abnormal events, and at the same time optimizes the global policy through the aggregation of the entire network data to avoid chain reactions caused by large-scale adjustments. 5G communication and digital signature technology ensure the security and reliability of data transmission.

[0120] This embodiment also provides a distribution network planning system considering the uncertainty of distributed energy, including: a feature extraction module, a scenario mixing module, a scheme generation module, a planning correction module, and a cloud-edge collaboration module; The feature extraction module is used to collect distributed energy node data, compensate for spatio-temporal offset by combining meteorological data, perform time series alignment using the dynamic time warping algorithm, construct a spatio-temporal feature matrix, extract local spatio-temporal patterns through a sliding window, and output a high-dimensional feature dataset; The scenario mixing module is used to construct an improved Wasserstein generative adversarial network to generate conventional scenarios, inject Gaussian noise through latent space perturbation to generate extreme scenarios of the training distribution, and integrate them into a mixed scenario set; The scheme generation module is used to input the mixed scenario set into a mixed-integer non-linear programming model, establish a topology-power flow proxy model using a graph neural network to accelerate the solution, output the predicted value of the line overload probability, perform branch and bound pruning, and obtain an energy storage configuration scheme and a dynamic topology strategy library; The planning correction module is used to update the line impedance parameters through a Kalman filter, collect the actual output and conduct verification, and use rolling horizon optimization to adjust the energy storage configuration scheme and the topology switching sequence to generate a corrected planning scheme; The cloud-edge collaboration module is used to decompose the corrected planning scheme into cloud-side global optimization and edge-side local control for cloud-edge collaboration.

[0121] This embodiment also provides a computer device applicable to the case of a distribution network planning method considering the uncertainty of distributed energy, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the distribution network planning method considering the uncertainty of distributed energy as proposed in the above embodiment.

[0122] This computer device can be a terminal. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a carrier network, NFC (near field communication), or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the outer shell of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0123] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the distribution network planning method for considering the uncertainty of distributed energy as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM for short), electrically erasable programmable read-only memory (EEPROM for short), erasable programmable read-only memory (EPROM for short), programmable read-only memory (PROM for short), read-only memory (ROM for short), magnetic memory, flash memory, magnetic disk or optical disc.

[0124] In summary, the present invention: collects distributed energy node data, compensates for spatio-temporal offset by combining meteorological data, uses the dynamic time warping algorithm to precisely align time series and extract local patterns, constructs a high-dimensional feature dataset, providing an accurate basis for subsequent scenario generation; uses an improved Wasserstein generative adversarial network to generate normal scenarios and injects Gaussian noise through latent space perturbation to generate extreme scenarios, achieving comprehensive coverage of normal and extreme working conditions and enhancing the robustness of the method to uncertainty; inputs the mixed scenario set into a mixed-integer non-linear programming model, constructs a topology-power flow surrogate model using a graph neural network, realizes rapid prediction of line overload probability and branch-and-bound pruning, and accelerates the solution of optimal energy storage configuration and dynamic topology strategy; updates line parameters in real time through a Kalman filter, and dynamically checks and corrects the energy storage configuration and topology switching sequence in combination with rolling horizon optimization; decomposes into cloud global optimization and edge local control to form a cloud-edge collaborative mechanism, thereby overall improving the economic efficiency, security and intelligent level of distribution network planning.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A distribution network planning method considering the uncertainty of distributed energy, characterized in that: including Collecting distributed energy node data, compensating for spatio-temporal offset by combining meteorological data, using the dynamic time warping algorithm for time series alignment, constructing a spatio-temporal feature matrix, extracting local spatio-temporal patterns through a sliding window, and outputting a high-dimensional feature dataset; Constructing an improved Wasserstein generative adversarial network to generate normal scenarios, injecting Gaussian noise through latent space perturbation to generate extreme scenarios deviating from the training distribution, and integrating them into a mixed scenario set; Inputting the mixed scenario set into a mixed-integer non-linear programming model, using a graph neural network to establish a topology-power flow surrogate model for accelerated solution, outputting the predicted value of line overload probability, and pruning using the branch and bound method to obtain an energy storage configuration plan and a dynamic topology strategy library; Updating the line impedance parameters through a Kalman filter, collecting the actual output and checking it, and using rolling horizon optimization to adjust the energy storage configuration plan and topology switching sequence to generate a corrected planning plan; Decomposing the corrected planning plan into cloud-side global optimization and edge-side local control for cloud-edge collaboration.

2. The distribution network planning method considering the uncertainty of distributed energy as described in claim 1, characterized in that: The compensating for spatio-temporal offset includes determining the cloud movement speed according to a meteorological radar, and applying a time offset compensation based on the coordinate difference and the cloud movement speed to the output time series data of adjacent nodes.

3. The distribution network planning method considering the uncertainty of distributed energy as described in claim 1, characterized in that: In the improved Wasserstein generative adversarial network, the generator receives a latent noise vector and a meteorological condition label, and generates an output scenario that conforms to spatio-temporal correlation; the discriminator adds a power balance constraint loss term and calculates the power flow violation penalty of the generated scenario to make the generated scenario satisfy Kirchhoff's law.

4. The distribution network planning method considering the uncertainty of distributed energy according to claim 1, characterized in that: In the mixed-integer non-linear programming model, the decision variables include the energy storage location, capacity, line expansion plan, and binary topology matrix; the objective function is to minimize the life cycle cost, and the constraint conditions are embedded with a linearized DisFlow power flow equation and a radial topology verification rule.

5. The distribution network planning method considering the uncertainty of distributed energy as described in claim 4, characterized in that: The constraint conditions of the binary topology matrix include radial constraint, connectivity constraint, and switch action times limit.

6. The distribution network planning method considering the uncertainty of distributed energy as described in claim 1, characterized in that: The construction of the graph neural network includes The node feature vector includes the photovoltaic installed capacity, load prediction value, and energy storage SOC state; The edge feature vector includes the line impedance, initial current-carrying capacity, and expansion candidate identifier; The output layer uses a sigmoid activation function to predict the line overload probability.

7. The distribution network planning method considering the uncertainty of distributed energy as described in claim 1, characterized in that: The cloud-side global optimization refers to updating the long-term energy storage capacity plan based on the Benders decomposition method and sending the topology strategy library to each edge node through the network; the edge-side local control refers to deploying a lightweight reinforcement learning controller to collect local operation parameters in real time and perform topology switching.

8. A distribution network planning system considering the uncertainty of distributed energy, based on the distribution network planning method considering the uncertainty of distributed energy according to any one of claims 1 to 7, characterized in that: including a feature extraction module, a scenario mixing module, a plan generation module, a planning correction module, and a cloud-edge collaboration module; The feature extraction module is used to collect distributed energy node data, compensate for spatio-temporal offset by combining meteorological data, use the dynamic time warping algorithm for time series alignment, construct a spatio-temporal feature matrix, extract local spatio-temporal patterns through a sliding window, and output a high-dimensional feature dataset; The scenario mixing module is used to construct an improved Wasserstein generative adversarial network to generate normal scenarios, inject Gaussian noise through latent space perturbation to generate extreme scenarios of the training distribution, and integrate them into a mixed scenario set; The scheme generation module is used to input the hybrid scenario set into the mixed-integer non-linear programming model, accelerate the solution by establishing a topology-power flow surrogate model using a graph neural network, output the predicted value of line overload probability, perform branch and bound pruning, and obtain the energy storage configuration scheme and the dynamic topology strategy library; The planning correction module is used to update the line impedance parameters through a Kalman filter, collect the actual output and conduct verification, and adopt rolling horizon optimization to adjust the energy storage configuration scheme and the topology switching sequence to generate a corrected planning scheme; The cloud-edge collaboration module is used to decompose the corrected planning scheme into global optimization in the cloud and local control at the edge to perform cloud-edge collaboration.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the distribution network planning method considering the uncertainty of distributed energy according to any one of claims 1 to 7.

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

Images