Artificial intelligence-based digital simulation decision analysis method and system for power distribution network
By establishing impedance loss and power flow prediction models based on artificial intelligence, and combining them with particle swarm optimization to optimize load allocation, the problems of lag and weak global optimization capabilities in distribution network optimization technology are solved. This achieves reduced network losses and load balancing, and improves the safety and stability of the distribution network.
Patent Information
- Application Number
- PCT/CN2025/141375
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-06
- Filing Date
- 2025-12-10
- Publication Date
- 2026-11-12
AI Technical Summary
Existing power distribution network optimization technologies suffer from high lag, weak global optimization capabilities, and low resource utilization, resulting in high network losses, unbalanced loads, and difficulty in coping with dynamic load changes.
An artificial intelligence-based approach is adopted to preprocess power data from the distribution network, establish an impedance loss model, combine power flow calculation and multivariate feature fusion, and use particle swarm optimization algorithm to optimize load allocation, thereby achieving accurate location of high-loss branches and dynamic load balancing.
It reduces network losses in the distribution network, improves load balancing, ensures the safety and stability of the distribution network, and can respond to load changes in real time.
Smart Images

Figure CN2025141375_12112026_PF_FP_ABST
Abstract
Description
Artificial Intelligence-Based Digital Simulation Decision Analysis Method and System for Distribution Networks
[0001] This application claims priority to Chinese Patent Application No. 202510571661.8, filed with the Chinese Patent Office on May 6, 2025, the entire contents of which are incorporated herein by reference. Technical Field
[0002] This application relates to the field of distribution network decision analysis technology, and for example to an artificial intelligence-based digital simulation decision analysis method and system for distribution networks. Background Technology
[0003] With the rapid development of smart grids and the large-scale integration of distributed energy resources, the load volatility and operational complexity of distribution networks have increased significantly, posing a severe challenge to distribution network load allocation optimization methods. Related distribution network optimization technologies mainly include offline optimization based on static power flow, rule-driven load transfer strategies, and single-model prediction methods. While these methods can all achieve load allocation, they still have certain limitations. On the one hand, offline optimization based on static power flow constructs a fixed topology model using historical data and employs heuristic algorithms such as genetic algorithms to generate load allocation schemes. However, this method has a long optimization cycle and cannot respond to load changes in real time, resulting in a lag in the scheme and making it unsuitable for distribution network optimization in dynamic scenarios. On the other hand, rule-driven load transfer strategies mostly rely on manual experience to set load rate thresholds and transfer overloaded branch loads through switching operations. This lacks a global optimization perspective and easily falls into the trap of local optimization, triggering a chain reaction of "overload – load transfer – new overload." Furthermore, single-model prediction typically separates load prediction from optimization decisions, leading to the gradual accumulation of prediction errors. It also lacks the fusion of multi-dimensional data, resulting in data silos and causing optimization schemes to easily deviate from actual operating conditions. Summary of the Invention
[0004] This application provides an artificial intelligence-based digital simulation decision analysis method and system for power distribution networks, which can solve problems such as high technological lag, weak global optimization capability, and low resource utilization, thereby achieving the technical effects of reducing network losses in power distribution networks and improving the load balance of distribution cabinets.
[0005] In a first aspect, this application provides an artificial intelligence-based digital simulation decision analysis method for power distribution networks, the method comprising:
[0006] Acquire power data from the distribution network and perform data preprocessing on the power data to generate a standardized dataset. The standardized dataset includes the branch ID, impedance value, active power, and reactive power of each branch of the distribution network.
[0007] Based on the distribution network topology and the standardized dataset, an initial loss dataset is obtained through power flow calculation. The initial loss dataset includes the loss value, voltage deviation, and power factor of each branch.
[0008] Based on the initial loss dataset and the impedance values of each branch, an impedance loss model is established, and based on the current impedance values of each branch and the impedance loss model, the loss type of each branch of the distribution network is determined.
[0009] A preset algorithm is used to identify loss patterns in the initial loss dataset and the loss types of each branch to obtain the spatial distribution pattern of loss in each branch.
[0010] The current impedance value of each branch is input into the pre-built power flow prediction model to obtain the predicted active power value of each branch.
[0011] Based on the predicted active power values and the spatial distribution pattern of losses for each branch, a load allocation optimization model is established and solved using the particle swarm optimization algorithm to obtain the optimal load allocation scheme.
[0012] In one embodiment, obtaining the initial loss dataset through power flow calculation based on the distribution network topology and the standardized dataset includes:
[0013] Based on the topology of the distribution network, construct the node-branch association matrix;
[0014] Based on the node branch association matrix and the standardized dataset, power flow calculation is performed using the improved forward-backward substitution method to obtain branch currents and node voltages.
[0015] Calculate the loss value based on the branch current and the corresponding branch resistance;
[0016] Calculate the voltage deviation based on the node voltage and the nominal voltage of the distribution network;
[0017] Calculate the power factor based on the active and reactive power of each branch;
[0018] An initial loss dataset is formed based on the branch ID of each branch, the loss value, the voltage deviation, and the power factor.
[0019] In one embodiment, the step of establishing an impedance loss model based on the initial loss dataset and the impedance values of each branch, and determining the loss type of each branch of the distribution network based on the current impedance value of each branch and the impedance loss model, includes:
[0020] A sliding time window is used to extract the timing features of the impedance values of each branch, resulting in an impedance feature matrix.
[0021] The least squares method is used to fit the impedance characteristic matrix and the loss values in the initial loss dataset to obtain the impedance loss model of each branch.
[0022] Obtain the current impedance value of each branch and input it into the corresponding impedance loss model to obtain the first predicted loss value of each branch, and extract the loss sensitivity coefficient of each branch from the corresponding impedance loss model.
[0023] If the first predicted loss value is greater than the loss threshold, or if the loss sensitivity coefficient is greater than the coefficient threshold, then the corresponding branch is marked as a high loss branch and the high loss branch mark is output.
[0024] The impedance loss model is represented by the following formula:
[0025] ;
[0026] In the formula, This represents the first predicted loss value. This represents the loss sensitivity coefficient. represents the impedance value, and b represents the correction factor.
[0027] In one embodiment, the step of using a preset algorithm to perform loss pattern recognition on the initial loss dataset and the loss types of each branch to obtain the spatial distribution pattern of the loss of each branch includes:
[0028] The PCA dimensionality reduction algorithm is used to reduce the dimensionality of the initial loss dataset to obtain the dimensionality-reduced initial loss dataset.
[0029] The K-means clustering algorithm was used to perform cluster analysis on the initial loss dataset after dimensionality reduction to obtain the load type of each branch, which includes industrial load, commercial load, residential load and mixed load.
[0030] Based on the load type and loss type of each branch, the spatial distribution pattern of loss for each branch is determined.
[0031] In one embodiment, the step of inputting the current impedance values of each branch into a pre-constructed power flow prediction model to obtain power flow distribution prediction results includes:
[0032] A sliding time window is used to extract the timing features of the current impedance value of each branch within a preset time period to obtain the impedance timing features.
[0033] Static feature extraction is performed on the initial loss dataset to obtain the baseline features of the network state;
[0034] The impedance timing characteristics and the network state baseline characteristics are input into a pre-constructed power flow prediction model to obtain the power flow distribution prediction results for each branch. The power flow prediction model is constructed using a long short-term memory neural network model.
[0035] In one embodiment, establishing a load allocation optimization model based on the predicted active power values of each branch and the spatial distribution pattern of losses includes:
[0036] Based on the predicted active power of each branch, calculate the predicted second loss value of each branch.
[0037] Determine the loss weight based on the spatial distribution pattern of loss in each branch;
[0038] Based on the loss weights, the second loss prediction values are weighted and summed to obtain the network loss of the distribution network;
[0039] Calculate the standard deviation of the load rate based on the load rate of each branch;
[0040] A load allocation optimization model is constructed with the network loss and the standard deviation of the load rate as the objective functions and voltage deviation and branch load rate as constraints.
[0041] In one embodiment, determining the loss weight based on the spatial distribution pattern of losses in each branch includes:
[0042] Set the corresponding initial weight according to the load type of each branch;
[0043] The initial weights are adjusted according to the loss type of each branch to obtain the loss weights.
[0044] In one embodiment, after obtaining the optimal load allocation scheme, the method further includes:
[0045] Acquire real-time monitoring data of the distribution network, and determine whether to trigger path transfer based on the real-time monitoring data, wherein the real-time monitoring data includes real-time branch load rate and real-time voltage;
[0046] If path transfer is triggered, all transfer paths of the branch to be transferred are determined according to the topology of the distribution network.
[0047] Power flow calculations are used to obtain the transferred load and loss reduction corresponding to each transfer path, and the transfer path with the maximum loss reduction is selected as the optimal transfer path.
[0048] The optimal load allocation scheme is modified based on the transferred load and the optimal transfer path to obtain the modified optimal load allocation scheme.
[0049] In one embodiment, determining whether to trigger path switching based on the real-time monitoring data includes:
[0050] Based on the real-time branch load rate, a sliding time window mechanism is used to calculate the load change rate between adjacent windows. If the load change rate is greater than the change rate threshold, path switching is triggered.
[0051] or,
[0052] Based on the real-time voltage obtained from each sampling, a sliding time window mechanism is used to calculate the voltage deviation within the current window, resulting in multiple voltage deviation values.
[0053] Each voltage deviation value is compared with a deviation threshold. If the number of voltage deviation values greater than the deviation threshold is greater than the number threshold, then path switching is triggered.
[0054] Secondly, this application provides an artificial intelligence-based digital simulation decision analysis system for power distribution networks, the system comprising:
[0055] The data processing module is configured to acquire power data from the distribution network, preprocess the power data to generate a standardized dataset, which includes the branch ID, impedance value, active power and reactive power of each branch of the distribution network; and obtain an initial loss dataset through power flow calculation based on the topology of the distribution network and the standardized dataset, which includes the loss value, voltage deviation and power factor of each branch.
[0056] The loss analysis module is configured to establish an impedance loss model based on the initial loss dataset and the impedance values of each branch, and determine the loss type of each branch of the distribution network based on the current impedance value of each branch and the impedance loss model; and use a preset algorithm to perform loss pattern recognition on the initial loss dataset and the loss type of each branch to obtain the spatial distribution pattern of loss of each branch.
[0057] The power flow prediction module is set to input the current impedance value of each branch into a pre-built power flow prediction model to obtain the predicted active power value of each branch.
[0058] The load allocation module is configured to establish a load allocation optimization model based on the predicted active power value and the spatial distribution pattern of losses for each branch, and to solve the model using the particle swarm optimization algorithm to obtain the optimal load allocation scheme.
[0059] This application provides an artificial intelligence-based digital simulation decision analysis method and system for distribution networks. By establishing an impedance loss model and quantifying the sensitivity of impedance to losses, this application enables precise location of high-loss branches. Through a power prediction model that fuses multivariate features, it improves power flow prediction accuracy, providing data support for forward-looking decision-making. Furthermore, through a multi-objective particle swarm optimization algorithm and a dynamic weight adjustment mechanism, it achieves optimized load allocation in the distribution network. This application effectively reduces network losses and improves load balance in the distribution network, thereby ensuring the safety and stability of distribution network operation. Attached Figure Description
[0060] Figure 1 is a flowchart illustrating the digital simulation decision analysis method for power distribution networks based on artificial intelligence in an embodiment of this application;
[0061] Figure 2 is a schematic diagram of the structure of the distribution network digital simulation decision analysis system based on artificial intelligence in an embodiment of this application. Detailed Implementation
[0062] Please refer to Figure 1. The first embodiment of this application proposes an artificial intelligence-based digital simulation decision analysis method for power distribution networks, which includes steps S10 to S60:
[0063] Step S10: Obtain power data of the distribution network and perform data preprocessing on the power data to generate a standardized dataset. The standardized dataset includes the branch ID, impedance value, active power and reactive power of each branch of the distribution network.
[0064] Step S20: Based on the topology of the distribution network and the standardized dataset, an initial loss dataset is obtained through power flow calculation. The initial loss dataset includes the loss value, voltage deviation, and power factor of each branch.
[0065] Step S30: Based on the initial loss dataset and the impedance values of each branch, establish an impedance loss model, and determine the loss type of each branch of the distribution network based on the current impedance value of each branch and the impedance loss model.
[0066] Step S40: Use a preset algorithm to perform loss pattern recognition on the initial loss dataset and the loss type of each branch to obtain the loss spatial distribution pattern of each branch.
[0067] Step S50: Input the current impedance value of each branch into the pre-built power flow prediction model to obtain the active power prediction value of each branch.
[0068] Step S60: Based on the predicted active power values and the spatial distribution pattern of losses for each branch, establish a load allocation optimization model and solve it using the particle swarm optimization algorithm to obtain the optimal load allocation scheme.
[0069] This application optimizes the load distribution scheme by digitally simulating the ring network structure and power data of the distribution network and performing power flow analysis on the power and loss of the distribution network. First, the power data of the distribution network is acquired and a standardized dataset is generated. In this embodiment, the power data can be divided into three sources: one is the impedance value obtained by the impedance monitoring device at the feeder segment point, and its data format is: (timestamp, branch ID, impedance value Ω / km); the second is the historical power circulation data from the SCADA data acquisition and monitoring control system, including branch active power, reactive power, etc.; the third is the monitoring data from smart meters, including load point ID, active power, and reactive power, etc. Then, these power data are preprocessed, including outlier filtering, missing value imputation, and time alignment. For example, Z-score is calculated for power data, outliers with absolute values > 3 are removed, impedance data exceeding the historical average by ±50% are marked as invalid, and linear interpolation is used to imput data with ≤ 3 consecutive missing time points. Since the time granularity of smart meters (15-minute granularity) may be longer than that of impedance monitoring devices and SCADA systems (5-minute granularity), the smart meter data also needs to be time aligned using cubic spline interpolation. The preprocessed data is then organized to obtain a standardized dataset. The power data in the dataset should at least include the timestamps, branch IDs, impedance values, active power, and reactive power of each branch of the distribution network.
[0070] After obtaining the standardized dataset, and combining it with the topology of the distribution network, power flow simulation analysis of the initial losses of the distribution network can be performed. The specific analysis steps include:
[0071] Based on the topology of the distribution network, construct the node-branch association matrix;
[0072] Based on the node branch association matrix and the standardized dataset, power flow calculation is performed using the improved forward-backward substitution method to obtain branch currents and node voltages.
[0073] Calculate the loss value based on the branch current and the corresponding branch resistance;
[0074] Calculate the voltage deviation based on the node voltage and the nominal voltage of the distribution network;
[0075] Calculate the power factor based on the active and reactive power of each branch;
[0076] An initial loss dataset is formed based on the branch ID of each branch, the loss value, the voltage deviation, and the power factor.
[0077] In this embodiment, the distribution network topology includes node connection relationships and branch types. Branch types include feeders, transformers, and distributed generation nodes. Considering the high resistance / reactance ratio of the distribution network, an improved forward-backward substitution method is used for power flow calculation, in conjunction with a standardized dataset. For example, based on the distribution network topology, the root node (feeder start point), end nodes, and branch hierarchy relationships are marked, and parameters are initialized. If a loop exists (e.g., Node4→Node5→Node6→Node4), the loop is temporarily broken, and the network is converted to a radial structure for iteration. Initialization parameters include setting initial voltage values, and the active and reactive power of nodes are extracted from the standardized dataset. If a node contains distributed generation (e.g., photovoltaic), it is processed as a PQ / PV node type.
[0078] When performing power flow calculations, the branch circuits are first calculated through a forward push process, i.e., pushing back from the end node to the root node to obtain the branch current. If a branch has sub-branches, the sub-branch currents are accumulated to obtain the branch current. Then, the node voltages are updated through a back-substitution process, i.e., pushing forward from the root node to the end node to obtain the node voltages. If a loop exists, loop correction is performed to restore the broken loops, calculate the loop voltage difference, and eliminate the deviation by compensating the current. At the same time, convergence conditions are set, such as the maximum voltage difference between adjacent nodes being less than a preset threshold, or reaching the maximum number of iterations, etc. Finally, the voltage deviation is obtained based on the difference between the node voltage and the nominal voltage. The branch power is calculated based on the node voltage and branch current, and the active and reactive power losses are calculated based on the difference between the power at the beginning and end of the branch. Finally, the power factor is calculated based on the active and reactive power losses. In calculating power loss, the derivation of the difference between the power at the beginning and end of a branch ultimately converts active power loss into the product of the square of the branch current and the resistance, and reactive power loss into the product of the square of the branch current and the reactance. The specific derivation process can be found in the conventional loss value calculation steps, which will not be elaborated here.
[0079] The initial loss dataset can be obtained through the power flow calculations described above. The initial loss dataset includes the branch IDs extracted from the topology data, the loss values determined based on active power loss, voltage deviation, and power factor.
[0080] This embodiment improves the convergence and efficiency of power flow calculation in distribution networks by using an improved forward-backward substitution method, combined with loop processing and distributed generation support. At the same time, it accurately outputs the initial loss dataset, providing reliable data input for subsequent decision analysis.
[0081] After obtaining the initial loss dataset, and combining it with the impedance of each branch, an impedance loss model is established to analyze the loss type of each branch. The specific steps include:
[0082] A sliding time window is used to extract the timing features of the impedance values of each branch, resulting in an impedance feature matrix.
[0083] The least squares method is used to fit the impedance characteristic matrix and the loss values in the initial loss dataset to obtain the impedance loss model of each branch.
[0084] Obtain the current impedance value of each branch and input it into the corresponding impedance loss model to obtain the first predicted loss value of each branch, and extract the loss sensitivity coefficient of each branch from the corresponding impedance loss model.
[0085] If the first predicted loss value is greater than the loss threshold, or if the loss sensitivity coefficient is greater than the coefficient threshold, then the corresponding branch is marked as a high-loss branch, and the high-loss branch mark is output.
[0086] In this embodiment, the data used to establish the impedance loss model of each branch includes impedance data from the standardized dataset and loss data from the initial loss dataset. First, the impedance values are processed based on a sliding time window. The window length is set and features are extracted, including the impedance mean, variance, maximum value, and minimum value for each window. Then, based on the impedance and loss values, the least squares method is used for modeling.
[0087] ;
[0088] In the formula, This represents the first predicted loss value. R represents the loss sensitivity coefficient, R represents the impedance value, and b represents the correction coefficient.
[0089] In the initial model, R is chosen as the mean impedance. Initial loss values were selected, and the loss sensitivity coefficient and correction coefficient were determined by data fitting with a goodness of fit greater than 0.9 and a residual standard deviation of less than 5% as verification conditions, thus obtaining the impedance loss model.
[0090] The current impedance value of each branch is input into the corresponding impedance loss model to obtain the first predicted loss value for each branch. Then, based on the impedance loss model and the predicted loss value, the loss type of each branch is determined. Here, the loss type includes high loss, medium loss, and low loss. In this embodiment, there are two methods for determining the loss type. One is based on the predicted loss value, with different loss thresholds preset. If the predicted loss value is greater than the high loss threshold, the branch is considered a high loss branch; if it is less than the low loss threshold, the branch is considered a low loss branch; otherwise, it is a medium loss branch.
[0091] Another method is based on the loss sensitivity coefficient of the impedance loss model. According to the model formula, impedance and loss are positively correlated. However, the sensitivity of different branches to loss varies significantly. This sensitivity is represented by the loss sensitivity coefficient in the formula. For example, for branches in industrial areas, due to high load rates and large currents, a small increase in impedance can lead to a significant increase in loss. The loss sensitivity coefficient in their impedance loss model is relatively large. Conversely, for branches in residential areas, the load rate and current are significantly lower than in industrial areas, resulting in a smaller loss sensitivity coefficient in their model. Therefore, the loss sensitivity coefficient can be used to determine whether a branch is a high-loss branch. Different coefficient thresholds are set. When the loss sensitivity coefficient of a branch's model is greater than the high threshold, it indicates that the branch is a high-loss branch and is marked. If it is less than the low threshold, it is considered a low-loss branch; otherwise, it is a medium-loss branch. In this embodiment, the key is marking high-loss branches because they have a relatively large impact on the load.
[0092] In this embodiment, the loss value of the branch is accurately predicted by the impedance loss model, and the impact of impedance on loss is measured by the loss sensitivity coefficient. The key branch that contributes the most to the loss of the whole network is identified, providing clear optimization targets and priorities for subsequent algorithms to achieve efficient resource allocation.
[0093] Based on the initial loss dataset and loss types of each branch obtained from the above steps, loss pattern recognition is used to determine the spatial distribution pattern of loss in each branch. Specific steps include:
[0094] The PCA dimensionality reduction algorithm is used to reduce the dimensionality of the initial loss dataset to obtain the dimensionality-reduced initial loss dataset.
[0095] The K-means clustering algorithm was used to perform cluster analysis on the initial loss dataset after dimensionality reduction to obtain the load type of each branch, which includes industrial load, commercial load, residential load and mixed load.
[0096] Based on the load type and loss type of each branch, the spatial distribution pattern of loss for each branch is determined.
[0097] In this embodiment, cluster analysis is performed on the initial loss dataset to determine the load type of each branch. The load type is then combined with the loss type to determine the spatial distribution pattern of the branch's losses. The load types include industrial loads, commercial loads, residential loads, and mixed loads. Different load types have different loss characteristics. For industrial loads, due to the high proportion of motors, transformers, and other equipment (resulting in a low power factor and requiring significant reactive power compensation), the current is large and continuous, leading to high losses and high voltage deviations. Industrial loads are characterized by high continuous load, low power factor, and small but high average loss fluctuations. For commercial loads, due to the concentrated operation of air conditioning and lighting equipment during business hours, the power factor is high, but short-term load concentration leads to medium to high losses. Therefore, they are characterized by bi-peak loads, high power factor, and moderate loss fluctuations. For residential loads, due to the dispersed nature of household appliances and concentrated usage time, the power factor is close to 1. However, due to long lines and low voltage at the ends, losses are low, but voltage deviations are large. Therefore, they are characterized by nighttime peaks, dispersed and volatile loads, and high power factor. For mixed loads, which consist of multiple types of loads superimposed, the losses and voltage deviations fall between those of industrial and residential loads.
[0098] Based on the differences in losses of different loads, this embodiment uses the PCA dimensionality reduction algorithm to reduce the dimensionality of the initial loss dataset. For example, the 20-dimensional features (loss, voltage, power factor, etc.) are reduced to 5-8 principal components (retaining 85% variance). By extracting time-series statistics such as mean, variance, and peak value as features, the periodicity of the load is captured, and the core pattern of load behavior is extracted. Then, the K-means clustering algorithm is used to perform cluster analysis on the dimensionality-reduced principal component features, thereby clustering each branch into different load types.
[0099] After obtaining the load type of each branch, and combining it with the loss type of the branch determined by the impedance loss model, the spatial distribution pattern of loss based on load type and loss type of each branch is determined. For example, branch A is industrial load - high loss, branch B is residential load - low loss, and branch C is commercial load - medium loss, etc.
[0100] In this application, in addition to predicting the loss value, the impedance value of each branch is also related to the active power of the branch. In this embodiment, the active power of each branch is predicted by a power flow prediction model, and the specific steps include:
[0101] A sliding time window is used to extract the timing features of the current impedance value of each branch within a preset time period to obtain the impedance timing features.
[0102] Static feature extraction is performed on the initial loss dataset to obtain the baseline features of the network state;
[0103] The impedance timing characteristics and the network state baseline characteristics are input into a pre-constructed power flow prediction model to obtain the power flow distribution prediction results for each branch. The power flow prediction model is constructed using a long short-term memory neural network model.
[0104] In this embodiment, a sliding time window is used to extract features of the current impedance values of each branch over a period of time. Assuming the sliding time window is set to a 24-hour window with a 15-minute step size, the extracted features include a 96-dimensional time-series mean, i.e., the impedance mean for every 15-minute interval, reflecting intraday periodic changes; and 4-dimensional statistics, i.e., mean, variance, maximum, and minimum, describing the impedance fluctuation characteristics within the window, thus obtaining the impedance time-series feature matrix. This embodiment uses a Long Short-Term Memory (LSTM) neural network model to construct the power flow prediction model. LSTM, as a model suitable for processing time-series data, can utilize these time-series features to learn and predict future impedance change trends, thereby influencing power flow distribution. The impedance time-series feature matrix is input into the trained power flow prediction model, outputting the predicted active power values of the branches.
[0105] To improve prediction accuracy, in an optional embodiment, this application adds the static state of branches to the input data of the power flow prediction model. This static state is provided by the initial loss dataset. Since the impedance time-series feature matrix provides dynamic time-varying information, and the initial loss dataset provides the current network loss and electrical state, the combination of these two allows the model to consider both historical trends and the current actual state, improving prediction accuracy. In this embodiment, the loss values, voltage deviations, and power factors in the initial loss dataset provide a baseline value for the current network state, reflecting the actual loss and electrical state of the current branch. Using both the impedance time-series features and the network state baseline features as input data for the model yields accurate prediction results.
[0106] In this embodiment, the future impedance trend is predicted by dynamic time-series changes, which drives the power flow distribution change. The prediction results are consistent with the current loss distribution by constraining the static network state, avoiding violation of physical laws. The fusion of dynamic and static data integrates historical trends and current state, which effectively improves the prediction robustness of the model.
[0107] Through the above steps, the predicted active power and spatial distribution pattern of losses for each branch are determined. Then, based on the predicted active power and spatial distribution pattern of losses, a load allocation optimization model is established to achieve optimized load allocation in the distribution network. The specific steps include:
[0108] Based on the predicted active power of each branch, calculate the predicted second loss value of each branch.
[0109] Determine the loss weight based on the spatial distribution pattern of loss in each branch;
[0110] Based on the loss weights, the second loss prediction values are weighted and summed to obtain the network loss of the distribution network;
[0111] Calculate the standard deviation of the load rate based on the load rate of each branch;
[0112] A load allocation optimization model is constructed with the network loss and the standard deviation of the load rate as the objective functions and voltage deviation and branch load rate as constraints.
[0113] In this embodiment, the second loss prediction value for each branch is first calculated based on the predicted active power value of each branch. Since the branch current is determined by the active power, that is, the branch current is equal to the square root of the sum of the squares of the active power and reactive power, divided by the voltage value. The voltage value can be obtained through power flow calculation. Since the power factor of branches with different load types is relatively fixed, the reactive power can be calculated from the active power and the power factor. Finally, the branch current is calculated based on the predicted active power value. The square of the branch current is proportional to the branch loss value, that is, the active power loss. Therefore, the predicted loss value of the branch can be obtained by multiplying the square of the branch current by the branch resistance.
[0114] The above embodiments provide an impedance loss model, which can calculate the loss value through impedance. In this embodiment, the reason why this model is not used for prediction is that the linear model of the fitting formula is a simplified approximation of the nonlinear relationship (the loss value is positively correlated with the square of the current). The model assumes that the impedance is stable in the short term and reflects the loss pattern under the current load distribution by fitting historical data. The coefficients of the model implicitly contain the average current effect under the current load level and do not directly reflect the relationship with the current. That is, the calculation efficiency is improved by sacrificing accuracy. Since the model is mainly used to classify the loss of the branch, there is no accuracy requirement for its predicted value.
[0115] The load allocation optimization model in this embodiment is used for refined prediction and optimization. It needs to accurately reflect the impact of current changes on losses. Therefore, the model used for loss classification mentioned above is not used. Instead, the model is used to calculate the instantaneous loss value of each branch based on the current fluctuation caused by changes in active power.
[0116] Then, the loss weight of each branch is determined based on the spatial distribution pattern of losses in each branch. For example, a basic weight is set according to the load type of the branch. In this embodiment, the basic weight of industrial load is the highest, for example, set to 0.8. The weights of other loads are lower than those of industrial loads, for example, set to 0.7. Then, the basic weight is adjusted according to the loss type of the branch. If it is a high-loss branch, the weight is increased by the correction system, for example, the weight is multiplied by 1.2 to strengthen the penalty for its loss. Correspondingly, if it is a low-loss branch, the weight is reduced by the correction coefficient, for example, the weight is multiplied by 0.8. Medium-loss branches remain unchanged, thereby determining the loss weight of each branch. Finally, the predicted loss value and loss weight of each branch are weighted and summed to obtain the network loss of the distribution network. Then, the loss weight of the network loss is set, for example, 0.7. This weight value is set relatively high because during the operation of the distribution network, energy loss not only means energy waste, but may also lead to problems such as line overheating and equipment aging, affecting the economy and reliability of the distribution network.
[0117] The load factor of each branch is obtained. The load factor is the ratio of the actual load to the rated capacity, reflecting the proportion of the current load on the branch relative to its maximum rated capacity. Then, the standard deviation of the load factor for all branches is calculated. The standard deviation is a statistic used to measure the dispersion of a set of data. Here, it represents the degree of dispersion of the load factor of each branch relative to the average load factor, that is, it reflects the uniformity of the load distribution of each branch. In the optimization objective, making the load distribution of each branch uniform is also of certain importance. If the branch load distribution is uneven, it may lead to some branches being overloaded while others are lightly loaded. This will not only affect the overall operating efficiency of the distribution network, but may also cause safety hazards, such as equipment damage in overloaded branches. Therefore, a load balancing term is added to the objective function.
[0118] Then, the objective function is constructed by minimizing the network loss and the standard deviation of the load rate:
[0119] ;
[0120] In the formula, Indicates the loss weight. This represents the load factor weight, and P represents the network loss of the distribution network. This represents the standard deviation of the load rate.
[0121] At the same time, constraints are added, including voltage deviation constraints and load rate constraints. The voltage deviation constraint means that the absolute difference between the node voltage value and the nominal voltage is less than a threshold, such as limiting the voltage deviation after optimization to no more than 5%. The load rate constraint means that the load rate of the branch is less than the load rate constraint threshold. The load rate constraint is mainly for high-loss branches to avoid excessive load rate.
[0122] In this embodiment, the objective function F aims to minimize the value of F by comprehensively considering the total loss of the distribution network and the uniformity of the load rate of each branch, in order to find the optimal load distribution scheme. The loss term emphasizes reducing the total loss to improve the economy of the distribution network and reduce energy waste, while the load balancing term focuses on making the load distribution of each branch uniform to ensure the reliability and overall operating efficiency of the distribution network. The optimal load distribution scheme is output to achieve the high efficiency, economy and reliability of the distribution network operation.
[0123] For the load allocation optimization model, a particle swarm optimization (PSO) algorithm is used for solution. First, the particle swarm is initialized, with the load allocation ratio of each branch as the particle position vector. The initial value is based on the historical load rate, and the initial load rate of high-loss branches is set to a lower value (e.g., 0.75). Then, fitness is calculated based on the input data (loss prediction, classification label, voltage deviation). When constraints are violated, the fitness value is multiplied by a penalty coefficient (e.g., 1.5). Then, dynamic weight adjustment is performed. The weight of industrial branches gradually increases with the number of iterations to strengthen their optimization, while the inertia weight decreases linearly. Initially, a global search is performed, and later, local refinement is achieved. Finally, the global optimal solution is output, which is the optimal load allocation ratio of each branch. This ensures that under this allocation scheme, the distribution network loss is minimized and the load is balanced, thereby achieving efficient and economical operation of the distribution network.
[0124] In an optional embodiment, after obtaining the optimal load allocation scheme, this application also provides a method for dynamically adjusting the allocation scheme, the specific steps of which include:
[0125] Acquire real-time monitoring data of the distribution network, and determine whether to trigger path transfer based on the real-time monitoring data, wherein the real-time monitoring data includes real-time branch load rate and real-time voltage;
[0126] If path transfer is triggered, all transfer paths of the branch to be transferred are determined according to the topology of the distribution network.
[0127] Power flow calculations are used to obtain the transferred load and loss reduction corresponding to each transfer path, and the transfer path with the maximum loss reduction is selected as the optimal transfer path.
[0128] The optimal load allocation scheme is modified based on the transferred load and the optimal transfer path to obtain the modified optimal load allocation scheme.
[0129] In this embodiment, real-time monitoring data of the distribution network is acquired, including branch load rate and voltage deviation. This data is directly collected in real time by monitoring devices (such as smart meters, voltage transformers, etc.) installed on each branch of the distribution network. The monitoring devices measure and calculate the operating parameters of the branch at certain time intervals (e.g., every 1 minute) to obtain real-time monitoring data. Then, based on the branch load rate or voltage deviation, it is determined whether to trigger path transfer.
[0130] For example, when using load rate for judgment, a sliding time window is set. At each time point, the average load rate of the branch within the current window is calculated, and the average load rate of the branch within the previous time window is recorded. Based on the average load rate of adjacent windows, the load rate change is calculated. If the load rate change is greater than a threshold, such as greater than 10%, and the duration is consistent with the time window length, such as 5 minutes, then a sudden change in load rate is judged, and load switching is required. When using voltage deviation for judgment, a sliding time window is also set, and the window length is determined according to the number of consecutive samples. After each voltage value is collected, its deviation percentage from the nominal voltage is calculated. If, in consecutive sampling, the calculated voltage deviation value is greater than the offset threshold each time, such as greater than 5%, then voltage fluctuation is considered to have occurred, and load switching is required.
[0131] When load transfer is triggered, all possible transfer paths for the current branch are determined based on the distribution network topology. These paths consist of a series of connected branches that can transfer load from the current branch to other branches to achieve load redistribution. Each path must meet voltage constraints and branch capacity limits. For example, for branch A, there may be multiple candidate transfer paths such as path A-branch B-branch C, path A-branch D-branch E, etc.
[0132] The transferable load of a current branch is limited by its load rate and the remaining capacity of adjacent branches. The difference between the current branch's load rate and the minimum allowable load rate is calculated, along with the difference between the rated capacity of adjacent branches and their current load rates. The minimum of these two differences is taken as the upper limit for load transfer. When calculating the optimal load transfer amount, minimizing the total distribution network loss is the objective, with voltage and branch capacity constraints as conditions. Using particle swarm optimization or a greedy algorithm, the load transfer amount is treated as the optimization variable, and the optimal transfer ratio satisfying the objective function is obtained through iterative search.
[0133] After determining the load transfer amount for each candidate transfer path, for each candidate transfer path, a power flow calculation model (such as the improved forward-backward substitution method), combined with current real-time monitoring data (including branch load rate, voltage, etc.) and the optimized load allocation scheme, is used to calculate the change in the total loss of the entire distribution network after transferring part of the load to that path. Assuming the initial loss of the distribution network before the transfer is P1, and the new loss of the distribution network after transferring the load to a candidate transfer path is P2, then the loss reduction of that transfer path is (P1-P2) / P1. The loss reduction of all candidate transfer paths is compared, and the path with the largest loss reduction is selected as the priority transfer path. For example, if the loss reduction of path A-branch B-branch C is 3.8%, while the loss reductions of other paths are 2.5%, 3.2%, etc., then path A-branch B-branch C is selected as the transfer path.
[0134] Based on the selected load transfer path, determine the branch circuit switches that need to be operated. For example, for path A-branch B-branch C, it may be necessary to close the connecting switch between branch A and branch B, and simultaneously disconnect some load switches on branch A to transfer the load from branch A to branch B and branch C. Then, based on the actual load transfer situation, the optimal load allocation scheme is modified to obtain the optimized load allocation ratio for each branch.
[0135] This embodiment provides an artificial intelligence-based digital simulation decision analysis method for distribution networks. By establishing an impedance loss model and quantifying the sensitivity of impedance to loss, this application achieves accurate location of high-loss branches. Through a power prediction model that integrates multivariate features, it improves the accuracy of power flow prediction, providing data support for forward-looking decision-making. Furthermore, through a multi-objective particle swarm optimization algorithm and a dynamic weight adjustment mechanism, it achieves optimized allocation of distribution network loads. This application can effectively reduce network losses in distribution networks, improve load balance, and thus ensure the safety and stability of distribution network operation.
[0136] Please refer to Figure 2. A digital simulation decision analysis system for power distribution networks based on artificial intelligence, as proposed in the second embodiment of this application, includes:
[0137] The data processing module 10 is used to acquire power data of the distribution network and perform data preprocessing on the power data to generate a standardized dataset. The standardized dataset includes the branch ID, impedance value, active power and reactive power of each branch of the distribution network.
[0138] Based on the distribution network topology and the standardized dataset, an initial loss dataset is obtained through power flow calculation. The initial loss dataset includes the loss value, voltage deviation, and power factor of each branch.
[0139] The loss analysis module 20 is used to establish an impedance loss model based on the initial loss dataset and the impedance values of each branch, and to determine the loss type of each branch of the distribution network based on the current impedance value of each branch and the impedance loss model.
[0140] A preset algorithm is used to identify loss patterns in the initial loss dataset and the loss types of each branch to obtain the spatial distribution pattern of loss in each branch.
[0141] The power flow prediction module 30 is used to input the current impedance value of each branch into the pre-built power flow prediction model to obtain the active power prediction value of each branch.
[0142] The load allocation module 40 is used to establish a load allocation optimization model based on the predicted active power value of each branch and the spatial distribution pattern of losses, and to solve the model using the particle swarm optimization algorithm to obtain the optimal load allocation scheme.
[0143] The technical features and effects of the AI-based digital simulation decision analysis system for distribution networks proposed in this application are the same as the methods proposed in this application, and will not be repeated here. Each module in the aforementioned AI-based digital simulation decision analysis system for distribution networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0144] In summary, the embodiments of this application propose an artificial intelligence-based digital simulation decision analysis method and system for distribution networks. The method acquires power data from the distribution network and preprocesses the data to generate a standardized dataset. This standardized dataset includes the branch ID, impedance value, active power, and reactive power of each branch in the distribution network. Based on the distribution network topology and the standardized dataset, an initial loss dataset is obtained through power flow calculation. This initial loss dataset includes the loss value, voltage deviation, and power factor of each branch. Based on the initial loss dataset and the impedance of each branch... This application establishes an impedance loss model and determines the loss type of each branch in the distribution network based on the current impedance value of each branch and the impedance loss model. A preset algorithm is used to perform loss pattern recognition on the initial loss dataset and the loss types of each branch to obtain the spatial distribution pattern of losses for each branch. The current impedance value of each branch is input into a pre-constructed power flow prediction model to obtain the predicted active power value of each branch. Based on the predicted active power value of each branch and the spatial distribution pattern of losses, a load allocation optimization model is established and solved using a particle swarm optimization algorithm to obtain the optimal load allocation scheme. This application achieves accurate location of high-loss branches by establishing an impedance loss model and quantifying the sensitivity of impedance to losses. Through a power prediction model that integrates multi-variable features, it improves the accuracy of power flow prediction, providing data support for forward-looking decision-making. Furthermore, through a multi-objective particle swarm optimization algorithm and a dynamic weight adjustment mechanism, it achieves optimized load allocation in the distribution network. This application can effectively reduce network losses in the distribution network, improve load balance, and thus ensure the safety and stability of the distribution network operation.
[0145] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple; refer to the description of the method embodiments for relevant parts. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as these combinations of technical features do not contradict each other, they should be considered within the scope of this specification.
Claims
1. An artificial intelligence-based digital simulation decision analysis method for power distribution networks, comprising: Acquire power data from the distribution network and perform data preprocessing on the power data to generate a standardized dataset. The standardized dataset includes the branch ID, impedance value, active power, and reactive power of each branch of the distribution network. Based on the distribution network topology and the standardized dataset, an initial loss dataset is obtained through power flow calculation. The initial loss dataset includes the loss value, voltage deviation, and power factor of each branch. Based on the initial loss dataset and the impedance values of each branch, an impedance loss model is established, and based on the current impedance values of each branch and the impedance loss model, the loss type of each branch of the distribution network is determined. A preset algorithm is used to identify loss patterns in the initial loss dataset and the loss types of each branch to obtain the spatial distribution pattern of loss in each branch. The current impedance value of each branch is input into the pre-built power flow prediction model to obtain the predicted active power value of each branch. Based on the predicted active power values and the spatial distribution pattern of losses for each branch, a load allocation optimization model is established and solved using the particle swarm optimization algorithm to obtain the optimal load allocation scheme.
2. The artificial intelligence-based digital simulation decision analysis method for power distribution networks according to claim 1, wherein, The initial loss dataset is obtained through power flow calculation based on the distribution network topology and the standardized dataset, including: Based on the topology of the distribution network, construct the node-branch association matrix; Based on the node branch association matrix and the standardized dataset, power flow calculation is performed using the improved forward-backward substitution method to obtain branch currents and node voltages. Calculate the loss value based on the branch current and the corresponding branch resistance; Calculate the voltage deviation based on the node voltage and the nominal voltage of the distribution network; Calculate the power factor based on the active and reactive power of each branch; An initial loss dataset is formed based on the branch ID of each branch, the loss value, the voltage deviation, and the power factor.
3. The artificial intelligence-based digital simulation decision analysis method for power distribution networks according to claim 1, wherein, The step of establishing an impedance loss model based on the initial loss dataset and the impedance values of each branch, and determining the loss type of each branch of the distribution network based on the current impedance value of each branch and the impedance loss model, includes: A sliding time window is used to extract the timing features of the impedance values of each branch, resulting in an impedance feature matrix. The least squares method is used to fit the impedance characteristic matrix and the loss values in the initial loss dataset to obtain the impedance loss model of each branch. Obtain the current impedance value of each branch and input it into the corresponding impedance loss model to obtain the first predicted loss value of each branch, and extract the loss sensitivity coefficient of each branch from the corresponding impedance loss model. If the first predicted loss value is greater than the loss threshold, or if the loss sensitivity coefficient is greater than the coefficient threshold, then the corresponding branch is marked as a high loss branch and the high loss branch mark is output. The impedance loss model is represented by the following formula: In the formula, This represents the first predicted loss value. R represents the loss sensitivity coefficient, R represents the impedance value, and b represents the correction coefficient.
4. The artificial intelligence-based digital simulation decision analysis method for power distribution networks according to claim 1, wherein, The step of using a preset algorithm to perform loss pattern recognition on the initial loss dataset and the loss types of each branch to obtain the spatial distribution pattern of loss for each branch includes: The PCA dimensionality reduction algorithm is used to reduce the dimensionality of the initial loss dataset to obtain the dimensionality-reduced initial loss dataset. The K-means clustering algorithm was used to perform cluster analysis on the initial loss dataset after dimensionality reduction to obtain the load type of each branch, which includes industrial load, commercial load, residential load and mixed load. Based on the load type and loss type of each branch, the spatial distribution pattern of loss for each branch is determined.
5. The artificial intelligence-based digital simulation decision analysis method for power distribution networks according to claim 1, wherein, The step of inputting the current impedance value of each branch into a pre-constructed power flow prediction model to obtain the power flow distribution prediction result includes: A sliding time window is used to extract the timing features of the current impedance value of each branch within a preset time period to obtain the impedance timing features. Static feature extraction is performed on the initial loss dataset to obtain the baseline features of the network state; The impedance timing characteristics and the network state baseline characteristics are input into a pre-constructed power flow prediction model to obtain the power flow distribution prediction results for each branch. The power flow prediction model is constructed using a long short-term memory neural network model.
6. The method for digital simulation and decision analysis of power distribution networks based on artificial intelligence according to claim 1, wherein, The step of establishing a load allocation optimization model based on the predicted active power values and the spatial distribution pattern of losses for each branch includes: Based on the predicted active power of each branch, calculate the predicted second loss value of each branch. Determine the loss weight based on the spatial distribution pattern of loss in each branch; Based on the loss weights, the second loss prediction values are weighted and summed to obtain the network loss of the distribution network; Calculate the standard deviation of the load rate based on the load rate of each branch; A load allocation optimization model is constructed with the network loss and the standard deviation of the load rate as the objective functions and voltage deviation and branch load rate as constraints.
7. The artificial intelligence-based digital simulation decision analysis method for power distribution networks according to claim 6, wherein, The step of determining the loss weight based on the spatial distribution pattern of loss in each branch includes: Set the corresponding initial weight according to the load type of each branch; The initial weights are adjusted according to the loss type of each branch to obtain the loss weights.
8. The distribution network digital simulation decision analysis method based on artificial intelligence according to claim 1, further comprising, after obtaining the optimal load allocation scheme: Acquire real-time monitoring data of the distribution network, and determine whether to trigger path transfer based on the real-time monitoring data, wherein the real-time monitoring data includes real-time branch load rate and real-time voltage; If path transfer is triggered, all transfer paths of the branch to be transferred are determined according to the topology of the distribution network. Power flow calculations are used to obtain the transferred load and loss reduction corresponding to each transfer path, and the transfer path with the maximum loss reduction is selected as the optimal transfer path. The optimal load allocation scheme is modified based on the transferred load and the optimal transfer path to obtain the modified optimal load allocation scheme.
9. The artificial intelligence-based digital simulation decision analysis method for power distribution networks according to claim 8, wherein, The step of determining whether to trigger path switching based on the real-time monitoring data includes: Based on the real-time branch load rate, a sliding time window mechanism is used to calculate the load change rate between adjacent windows. If the load change rate is greater than the change rate threshold, path switching is triggered. or, Based on the real-time voltage obtained from each sampling, a sliding time window mechanism is used to calculate the voltage deviation within the current window, resulting in multiple voltage deviation values. Each voltage deviation value is compared with a deviation threshold. If the number of voltage deviation values greater than the deviation threshold is greater than the number threshold, then path switching is triggered.
10. An artificial intelligence-based digital simulation decision analysis system for power distribution networks, comprising: The data processing module is configured to acquire power data from the distribution network, preprocess the power data to generate a standardized dataset, which includes the branch ID, impedance value, active power and reactive power of each branch of the distribution network; and obtain an initial loss dataset through power flow calculation based on the topology of the distribution network and the standardized dataset, which includes the loss value, voltage deviation and power factor of each branch. The loss analysis module is configured to establish an impedance loss model based on the initial loss dataset and the impedance values of each branch, and determine the loss type of each branch of the distribution network based on the current impedance value of each branch and the impedance loss model; and use a preset algorithm to perform loss pattern recognition on the initial loss dataset and the loss type of each branch to obtain the spatial distribution pattern of loss of each branch. The power flow prediction module is set to input the current impedance value of each branch into a pre-built power flow prediction model to obtain the predicted active power value of each branch. The load allocation module is configured to establish a load allocation optimization model based on the predicted active power value and the spatial distribution pattern of losses for each branch, and to solve the model using the particle swarm optimization algorithm to obtain the optimal load allocation scheme.
