Power system artificial intelligence decision enhancement method based on sample utility perception
Patent Information
- Application Number
- CN202611009865.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-08
- Publication Date
- 2026-11-03
AI Technical Summary
Existing AI decision-making methods for power systems suffer from long computation times and complex iterative processes when facing large-scale power grids and multiple operating scenarios, making it difficult to meet the needs of rapid online decision-making. Supervised learning methods rely on a large number of optimal solution labels, resulting in high data construction costs. Unsupervised learning methods fail to fully distinguish the training value of different samples, leading to insufficient utilization of sample utility and affecting the model's decision-making accuracy and constraint satisfaction capabilities.
By constructing an influence function to quantify the impact of each training sample on the unsupervised training loss and model parameter updates, high-efficiency samples are identified, and candidate samples are generated in the neighborhood of high-efficiency samples. The parameters and weights of the unsupervised deep neural network are updated synchronously through a two-layer optimization method, which strengthens the model's learning in the vicinity of constraint boundaries and nonlinear sensitive regions.
It improves the model's ability to learn from complex operating scenarios, reduces the average error and maximum error, enhances the physical feasibility of the output decision, and reduces the computational burden of traditional AC optimal power flow iteration solutions.
Smart Images
Figure CN123030946A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power dispatching, specifically to an artificial intelligence decision enhancement method for power systems based on sample utility perception. Background Technology
[0002] With the large-scale integration of new energy power generation, flexible loads, and power electronic equipment, the uncertainty and volatility of power system operation have significantly increased. To ensure the safe, economical, and stable operation of the power grid, the dispatching system needs to rapidly complete power flow analysis, optimal power flow calculation, and operational decisions under numerous operating scenarios. Among these, the optimal power flow problem is a fundamental issue in power system operation optimization, requiring the minimization of operating costs while satisfying power balance, generator output limits, node voltage limits, and line capacity constraints. Therefore, how to quickly obtain safe, feasible, and economical decision-making results under complex operating environments has become a crucial technical requirement for intelligent power system dispatching.
[0003] Existing power system optimization decision-making methods mainly include traditional optimization methods and artificial intelligence decision-making methods. Traditional optimization methods typically employ iterative algorithms such as Newton's method and interior-point method, which offer good solution accuracy. However, when facing large-scale power grids, multiple operating scenarios, and the need for rapid online decision-making, they suffer from problems such as long computation time, complex iterative processes, and insufficient real-time performance. Especially when renewable energy output and load change rapidly, the dispatching system needs to repeatedly solve a large number of optimization problems, making it difficult for traditional methods to meet the requirements of online applications.
[0004] In recent years, artificial intelligence methods such as deep neural networks have been applied to power system optimization decision-making. These methods learn the mapping relationship between operating states and optimization decision results, and can quickly output approximate decision results after training, thus significantly improving online computation speed. However, existing AI decision-making methods still have certain shortcomings. On the one hand, supervised learning methods typically require pre-generating a large number of optimal solution labels using traditional optimization solvers, resulting in high offline data construction costs. On the other hand, while unsupervised learning methods can reduce dependence on labeled data, they usually treat all training samples as equally important, failing to fully consider the differences in the contributions of different samples to model training and decision-making performance.
[0005] In reality, different samples have varying effects on training AI models. Samples located near constraint boundaries, in regions with significant nonlinear characteristics, or in areas where the model struggles to learn tend to have a greater impact on model parameter updates, loss reduction, and constraint satisfaction capabilities; while regular samples have a relatively limited effect on improving model performance. Ignoring this difference in sample utility during training can easily lead to insufficient learning of key samples, causing the model to exhibit problems such as larger decision errors, insufficient constraint satisfaction rates, and decreased generalization ability in complex operating scenarios.
[0006] Furthermore, existing sample augmentation methods mostly employ random sampling, uniform sampling, or sample selection strategies based on prediction errors. While these methods can increase the number of training samples, they struggle to accurately identify the key samples that truly impact model training performance. Random or uniform sampling lacks specificity, and error-based sample selection may only improve a few samples with large errors, failing to effectively enhance the overall model performance. Therefore, existing methods still need improvement in terms of sample utilization efficiency, key region learning ability, and adaptability to complex scenarios.
[0007] In summary, current AI-based decision-making methods for power systems suffer from the following main problems: traditional optimization methods have a heavy computational burden, making it difficult to meet the needs of rapid online decision-making; supervised learning methods rely on a large number of optimal solution labels, resulting in high data construction costs; existing unsupervised learning methods do not adequately distinguish the training value of different samples, leading to insufficient utilization of sample utility; and existing sample augmentation and weighting strategies lack accurate identification and targeted utilization of highly useful samples, affecting the model's decision-making accuracy, constraint satisfaction ability, and training stability.
[0008] Therefore, it is necessary to propose a sample utility-aware artificial intelligence decision-making enhancement method for power systems. This method quantifies the impact of different operating samples on model training loss and parameter updates, identifies highly efficient samples, and performs local sample enhancement and weight updates around these samples. This allows the AI model to focus more on key operating regions during training. This approach improves the efficiency of training sample utilization, enhances the model's learning ability in complex operating scenarios and constraint boundary regions, thereby improving the accuracy, constraint satisfaction capability, and engineering applicability of AI decision-making in power systems. Summary of the Invention
[0009] The purpose of this invention is to provide a method for enhancing artificial intelligence decision-making in power systems based on sample utility perception, comprising the following steps:
[0010] Step 1: Obtain historical operating data of the power system or operating samples generated by simulation, construct a training sample set, and use the training sample set to train an unsupervised deep neural network model;
[0011] Step 2: Construct a function to quantify the effect of each training sample on the unsupervised training loss and model parameter updates;
[0012] The influence function is used to calculate the influence value of each training sample on the overall training loss, and one or more high-efficiency samples are selected from the training sample set according to the magnitude of the influence value.
[0013] Step 3: Generate candidate samples in the neighborhood of the efficient sample, and update the parameters and weights of the unsupervised deep neural network model synchronously through a two-layer optimization method;
[0014] Step 4: Use the updated unsupervised deep neural network model to predict the voltage amplitude and phase angle of each node in the power system.
[0015] Furthermore, in the training sample set, the active load vector and reactive load vector are used as inputs, and the node voltage magnitude and phase angle are used as outputs.
[0016] Furthermore, the unsupervised deep neural network model includes an input layer, One hidden layer and one output layer;
[0017] The outputs of each hidden layer are shown below:
[0018]
[0019]
[0020]
[0021] in, Indicates the state of the input layer. Indicates the first The output of the hidden layer, and They represent the first Layer weight matrix and bias vector, Represents a non-linear activation function. This represents the optimal power flow state prediction output by the deep neural network.
[0022] Furthermore, during the training of unsupervised deep neural network models, the goal is to minimize the loss function;
[0023] The loss function is as follows:
[0024]
[0025] in, This represents the total unsupervised training loss; This represents the operating cost item. This indicates the penalty for exceeding the limits of generator active and reactive power output. This indicates a penalty for exceeding the line capacity limit. This indicates a penalty for exceeding the limit of the node voltage amplitude. This indicates the penalty for exceeding the phase angle limit at a node; , , , and These represent the weighting coefficients of the aforementioned loss terms.
[0026] Furthermore, the penalties for exceeding operating cost limits, generator active and reactive power output limits, line capacity limits, node voltage amplitude limits, and node phase angle limits are as follows:
[0027]
[0028]
[0029]
[0030]
[0031]
[0032] in, A hinge function that penalizes only the portion exceeding the limit; variables with exponentiation, such as... , , and This represents the corresponding physical quantity calculated by combining the state prediction from a deep neural network with the AC power flow equations.
[0033] Furthermore, in step 1, the optimal model parameters after training the unsupervised deep neural network model are as follows:
[0034]
[0035] in, Indicates in the sample weight vector The optimal or near-optimal deep neural network parameters are obtained.
[0036] Furthermore, the influence function is shown below:
[0037]
[0038] in, Indicates sample The impact value, This represents the Hessian matrix of the weighted unsupervised training loss at the current optimal parameters. Describes the Hessian inverse matrix. This represents the overall training loss at the optimal model parameters. The gradient of the parameters at that point, The gradient of the parameters representing the loss of a single sample.
[0039] Furthermore, the steps for calculating the impact of each training sample on the overall training loss using the influence function include:
[0040] Step 1: Transform the influence function into a linear equation, resulting in:
[0041]
[0042] in, = Let be the intermediate vector to be solved; The gradient of loss for a single sample;
[0043] Step 2, initialize the iteration parameters of the conjugate gradient method, i.e.:
[0044]
[0045] in, Represents the initial residual. Indicates the initial conjugate search direction;
[0046] Step 3, iteratively update the parameters of the conjugate gradient method, that is:
[0047]
[0048]
[0049]
[0050] in, Indicates the first Approximate solution for the next iteration Indicates the corresponding residual. Indicates the conjugate search direction;
[0051] Step 4, if the residual norm If the number of iterations is less than a preset threshold or the maximum number of iterations is reached, the iteration is terminated, and an approximate solution is used. Calculate the impact of the current training sample on the overall training loss; otherwise, let k = k + 1 and return to step 3.
[0052] The sample impact values are shown below:
[0053]
[0054] In the formula, This represents the sample impact value.
[0055] Furthermore, efficient use of samples refers to one or more samples with the greatest impact value.
[0056] Furthermore, the steps for generating new candidate samples include:
[0057] Each high-efficiency sample is used as the central sample, and a Euclidean hypersphere neighborhood is constructed around the central sample, i.e.:
[0058]
[0059] in, Indicates For the center of the ball, The local sampling region is defined by its radius. express 3D real space, Indicates candidate samples to the center sample The 2-norm distance.
[0060] Randomly generate the sampling direction and radius scaling factor, i.e.:
[0061]
[0062]
[0063] in, Represents the unit sampling direction vector. Let represent a random vector that follows a standard normal distribution. The mean is The covariance matrix is The standard normal distribution;
[0064] Based on the sampling direction and radius scaling factor, candidate samples are uniformly generated within the Euclidean hypersphere neighborhood of each central sample.
[0065] By quantifying sample utility through influence functions, this invention can identify the most critical running samples for training deep neural networks under unsupervised conditions. Through efficient neighborhood sampling and two-layer sample weight updates, this invention can enhance the model's learning near constraint boundaries and nonlinear sensitive regions, thereby reducing average and maximum errors, improving the physical feasibility of output decisions, and reducing the computational burden of traditional AC optimal power flow iteration in multi-scenario online applications. Attached Figure Description
[0066] Figure 1 This is a flowchart illustrating the artificial intelligence decision enhancement method for power systems based on sample utility perception, as described in this invention.
[0067] Figure 2 Efficiently utilize sample influence value evolution curves for the 118-node system;
[0068] Figure 3 The distribution of sample influence values before and after training of the 4-node system;
[0069] Figure 4 The distribution of sample influence values before and after training of the 588-node system;
[0070] Figure 5 shows the convergence curve of loss versus training rounds for the 4-node system.
[0071] Figure 6 shows the convergence curve of loss versus training rounds for the 118-node system.
[0072] Figure 7 shows the convergence curve of loss versus training rounds for the 588-node system. Detailed Implementation
[0073] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.
[0074] Example 1:
[0075] A sample utility-aware AI-powered decision enhancement method for power systems includes the following steps:
[0076] Step 1: Obtain historical operating data of the power system or operating samples generated by simulation, construct a training sample set, and use the training sample set to train an unsupervised deep neural network model;
[0077] Step 2: Construct a function to quantify the effect of each training sample on the unsupervised training loss and model parameter updates;
[0078] The influence function is used to calculate the influence value of each training sample on the overall training loss, and one or more high-efficiency samples are selected from the training sample set according to the magnitude of the influence value.
[0079] Step 3: Generate candidate samples in the neighborhood of the efficient sample, and update the parameters and weights of the unsupervised deep neural network model synchronously through a two-layer optimization method;
[0080] Step 4: Use the updated unsupervised deep neural network model to predict the voltage amplitude and phase angle of each node in the power system.
[0081] Example 2:
[0082] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as in Example 1. In addition, the active load vector and reactive load vector are used as inputs in the training sample set, and the node voltage magnitude and phase angle are used as outputs.
[0083] Example 3:
[0084] A method for enhancing AI-based decision-making in power systems based on sample utility perception, with technical content identical to any one of embodiments 1-2, further comprising an unsupervised deep neural network model including an input layer, One hidden layer and one output layer;
[0085] The outputs of each hidden layer are shown below:
[0086]
[0087]
[0088]
[0089] in, Indicates the state of the input layer. Indicates the first The output of the hidden layer, and They represent the first Layer weight matrix and bias vector, Represents a non-linear activation function. This represents the optimal power flow state prediction output by the deep neural network.
[0090] Example 4:
[0091] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as any one of Examples 1-3. Furthermore, when training the unsupervised deep neural network model, the goal is to minimize the loss function.
[0092] The loss function is as follows:
[0093]
[0094] in, This represents the total unsupervised training loss; This represents the operating cost item. This indicates the penalty for exceeding the limits of generator active and reactive power output. This indicates a penalty for exceeding the line capacity limit. This indicates a penalty for exceeding the limit of the node voltage amplitude. This indicates the penalty for exceeding the phase angle limit at a node; , , , and These represent the weighting coefficients of the aforementioned loss terms.
[0095] Example 5:
[0096] The method for enhancing AI-based decision-making in power systems based on sample utility perception has the same technical content as any one of Examples 1-4. Further, the operating cost item, generator active and reactive power output exceeding limit penalty item, line capacity exceeding limit penalty item, node voltage amplitude exceeding limit penalty item, and node phase angle exceeding limit penalty item are respectively as follows:
[0097]
[0098]
[0099]
[0100]
[0101]
[0102] in, A hinge function that penalizes only the portion exceeding the limit; variables with exponentiation, such as... , , and This represents the corresponding physical quantity calculated by combining the state prediction from a deep neural network with the AC power flow equations.
[0103] Example 6:
[0104] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as any one of Examples 1-5. Further, in step 1, the optimal model parameters after training the unsupervised deep neural network model are as follows:
[0105]
[0106] in, Indicates in the sample weight vector The optimal or near-optimal deep neural network parameters are obtained.
[0107] Example 7:
[0108] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as any one of Examples 1-6, and the influence function is as follows:
[0109]
[0110] in, Indicates sample The impact value, This represents the Hessian matrix of the weighted unsupervised training loss at the current optimal parameters. Describes the Hessian inverse matrix. This represents the overall training loss at the optimal model parameters. The gradient of the parameters at that point, The gradient of the parameters representing the loss of a single sample.
[0111] Example 8:
[0112] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as any one of Examples 1-7. Further, the step of calculating the impact value of each training sample on the overall training loss using an influence function includes:
[0113] Step 1: Transform the influence function into a linear equation, resulting in:
[0114]
[0115] in, = Let be the intermediate vector to be solved; The gradient of loss for a single sample;
[0116] Step 2, initialize the iteration parameters of the conjugate gradient method, i.e.:
[0117]
[0118] in, Represents the initial residual. Indicates the initial conjugate search direction;
[0119] Step 3, iteratively update the parameters of the conjugate gradient method, that is:
[0120]
[0121]
[0122]
[0123] in, Indicates the first Approximate solution for the next iteration Indicates the corresponding residual. Indicates the conjugate search direction;
[0124] Step 4, if the residual norm If the number of iterations is less than a preset threshold or the maximum number of iterations is reached, the iteration is terminated, and an approximate solution is used. Calculate the impact of the current training sample on the overall training loss; otherwise, let k = k + 1 and return to step 3.
[0125] The sample impact values are shown below:
[0126]
[0127] In the formula, This represents the sample impact value.
[0128] Example 9:
[0129] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as any one of Examples 1-8, but further, the high-utility sample is one or more samples with the largest influence value.
[0130] Example 10:
[0131] The method for enhancing AI decision-making in power systems based on sample utility perception has the same technical content as any one of embodiments 1-9. Further, the step of generating new candidate samples includes:
[0132] Each high-efficiency sample is used as the central sample, and a Euclidean hypersphere neighborhood is constructed around the central sample, i.e.:
[0133]
[0134] in, Indicates For the center of the ball, The local sampling region is defined by its radius. express 3D real space, Indicates candidate samples to the center sample The 2-norm distance.
[0135] Randomly generate the sampling direction and radius scaling factor, i.e.:
[0136]
[0137]
[0138] in, Represents the unit sampling direction vector. Let represent a random vector that follows a standard normal distribution. The mean is The covariance matrix is The standard normal distribution;
[0139] Based on the sampling direction and radius scaling factor, candidate samples are uniformly generated within the Euclidean hypersphere neighborhood of each central sample.
[0140] Example 11:
[0141] This invention provides an AI-enhanced decision-making method for power systems based on sample utility perception. This method targets rapid decision-making scenarios for optimal AC power flow. It constructs an unlabeled training sample set based on historical power system operating data or simulated load operation samples; uses a deep neural network to learn the mapping relationship between operating states and optimal power flow decision variables; quantifies the effect of each training sample on unsupervised training loss and model parameter updates using an influence function; and generates candidate samples in a targeted manner within the neighborhood of highly efficient samples. Furthermore, it synchronously updates the deep neural network parameters and sample weights through two-layer optimization, thereby ensuring that the model training process continuously focuses on key operating areas that significantly impact economic efficiency, feasibility, and constraint fulfillment.
[0142] like Figure 1 As shown, the artificial intelligence decision enhancement method for power systems based on sample utility perception according to an embodiment of the present invention includes the following steps:
[0143] Step 201: Obtain historical operating data of the power system or operating samples generated by simulation, construct a training sample set, and establish an unsupervised deep neural network model.
[0144] Specifically, power grid operation data such as active and reactive loads, generator output limits, node voltage limits, node phase angle limits, line admittance parameters, and line capacity limits are acquired for each node. Operational samples are then generated based on historical operating conditions, uniform sampling, or preset load fluctuation ranges. For each operational sample, the active and reactive load vectors are used as inputs to a deep neural network, while the node voltage amplitude and phase angle are used as the decision states to be predicted. Generator active and reactive outputs, as well as line power flow, are calculated from the predicted states combined with AC power flow equations.
[0145] In this step, the AC optimal power flow model is used to provide the physical basis for the training loss of this invention, and its goal is to minimize the generation cost while satisfying the grid operation security constraints.
[0146]
[0147] in, Represents a set of generators. Indicates the index of the generator or node. Indicates the first The active power output of the generator, Indicates the first The reactive power output of the generator Represents a node voltage amplitude, Represents a node phase angle, Indicates the first The operating cost function of the generator with respect to its active power output; the objective function for... belong The total generator costs are summed, with the lower limit of the summation placed below the summation sign. The net injected power at each node is determined by the following formula:
[0148]
[0149]
[0150] in, Represents a set of nodes. and Representing nodes respectively The net active power injection and net reactive power injection; and Representing the nodes respectively The active and reactive power outputs of the connected generators; and Representing nodes respectively Active and reactive load requirements; superscript Corresponding to the generation-side variables, superscript Corresponding load-side variables, superscript Corresponding net injected variables. The AC active and reactive power flow balance constraints are as follows:
[0151]
[0152]
[0153] in, and These represent the nodes in the nodal admittance matrix. With nodes The conductivity and susceptivity of the elements, Represents a node voltage amplitude, Represents a node phase angle, This represents the phase angle difference between the nodes at both ends of the line. belong This represents the summation of power flow over all adjacent or related nodes. The constraints for generators, voltage, phase angle, and line capacity are:
[0154]
[0155]
[0156]
[0157]
[0158]
[0159] in, and They represent the first The generator has a lower and upper limit for active power output. and They represent the first Lower and upper limits of reactive power output of a generator; and These represent the lower and upper limits of the node voltage amplitude, respectively. and These represent the lower and upper limits of the node phase angle, respectively. and They represent the lines respectively. The trend of meritorious and ineffective actions This indicates the thermal stability capacity or power flow limit of the line. This represents a set of routes.
[0160] Therefore, the training samples not only include the load operating state, but also implicitly contain the physical feasible domain of the power system, enabling deep neural networks to be trained through physical loss even under unlabeled conditions.
[0161] Step 202: Based on the generation cost item and power balance, generator output, node voltage and line power flow constraints, perform initial training on the deep neural network.
[0162] The output variables are stacked as follows:
[0163]
[0164] Deep neural networks are used to learn the mapping relationship from the operating state to the optimal power flow decision variables:
[0165]
[0166] in, The input vector of the deep neural network is represented by the active load vector. and reactive load vector Obtained by stacking; The output vector of a deep neural network is represented by the node voltage magnitude vector. Phase angle vector of nodes Obtained by stacking; This represents all trainable parameters of a deep neural network. Indicates by parameters A defined nonlinear mapping function. The deep neural network can be a fully connected feedforward network, including... There are several hidden layers; the calculation process for each layer is as follows:
[0167]
[0168]
[0169]
[0170] in, Indicates the state of the input layer. Indicates the first The output of the hidden layer, and They represent the first Layer weight matrix and bias vector, Represents a non-linear activation function. This represents the optimal power flow state prediction value output by the deep neural network. After calculating the generator output and line power flow based on the network output, an unsupervised training loss is constructed:
[0171]
[0172] in, This represents the total unsupervised training loss; This represents the operating cost item. This indicates the penalty for exceeding the limits of generator active and reactive power output. This indicates a penalty for exceeding the line capacity limit. This indicates a penalty for exceeding the limit of the node voltage amplitude. This indicates the penalty for exceeding the phase angle limit at a node; , , , and These represent the weighting coefficients of the aforementioned loss terms, used to adjust the relative importance between economic objectives and various physical constraint penalties. Each term can be specifically written as:
[0173]
[0174]
[0175]
[0176]
[0177]
[0178] in, A hinge function that penalizes only the portion exceeding the limit; variables with exponentiation, such as... , , and This represents the corresponding physical quantity calculated by combining the state prediction from a deep neural network with the AC power flow equations.
[0179] By using the loss function described above, it is not necessary to pre-solve a large number of optimal power flow labels, and the model can simultaneously focus on the economic efficiency and the degree of satisfaction of physical constraints during the training process.
[0180] Step 203: Calculate the impact value of each training sample on the training loss based on the current model, and quantify the sample utility using the impact value.
[0181] Specifically, learnable or updatable sample weights are assigned to each training sample to construct a weighted unsupervised training objective:
[0182]
[0183] in, Indicates the first training samples, Indicates the first The weights of each sample, Indicates the number of training samples. This represents a weight vector composed of the weights of all samples. Indicates sample In model parameters Unsupervised physical loss under given sample weight vectors. Then, the optimal model parameters at training convergence are:
[0184]
[0185] in, Indicates in the sample weight vector The optimal or near-optimal deep neural network parameters obtained below, This represents the parameter value that minimizes the weighted training loss. To analyze the importance of individual samples, the th... Apply a small perturbation to the weights of each sample:
[0186]
[0187] in, This represents the perturbed sample weight vector. Indicates the application to the first A small perturbation on the weight of each sample Indicates the first There are standard basis vectors. The optimal parameters after perturbation satisfy the first-order optimality condition:
[0188]
[0189] in, This indicates the parameters of a deep neural network. Find the gradient; the first-order optimality condition indicates that the perturbation weights... Under these conditions, the gradient of the training loss with respect to the parameters is zero. (The perturbation amount is missing from the original text.) Performing a first-order Taylor expansion at zero, we obtain:
[0190]
[0191] in, This represents the change in model parameters before and after the perturbation. Indicates the parameter The second derivative of , i.e., the Hessian correlation term. Therefore, the optimal model parameters for the . Sample weights The sensitivity is approximately:
[0192]
[0193] The above formula shows that the sample weights For optimal parameters The influence is determined by both the Hessian inverse matrix and the single-sample gradient. Further, according to the chain rule, the th... training samples The impact value on the overall training loss is defined as:
[0194]
[0195] in, Indicates sample The impact value, This represents the Hessian matrix of the weighted unsupervised training loss at the current optimal parameters. Describes the Hessian inverse matrix. Indicates the overall training loss at The gradient of the parameters at that point, This represents the gradient of the parameters in the loss of a single sample. A larger gradient indicates a more significant impact of the weight changes on the model parameters and training loss, and thus a higher training utility for that sample. To avoid explicit construction and large-scale inversion... In this invention, the calculation of influence values is transformed into solving linear equations:
[0196]
[0197] in, Let be the intermediate vector to be solved, theoretically corresponding to ; The gradient of the single-sample loss is, i.e. The conjugate gradient method is used for iterative approximation, with the initialization as follows:
[0198]
[0199] in, Represents the initial residual. Indicates the initial conjugate search direction. In the... In the next iteration, the step size residual Search direction and direction update coefficient Update according to the following formula:
[0200]
[0201]
[0202]
[0203] in, Indicates the first Approximate solution for the next iteration Indicates the corresponding residual. Indicates the conjugate search direction. Hessian vector product. Calculated using second-order automatic differentiation:
[0204]
[0205] When the residual norm The iteration terminates when the value falls below a preset threshold or the maximum number of iterations is reached, and an approximate solution is used. The sample influence value was obtained:
[0206]
[0207] Step 204: Select a high-efficiency sample as the center sample, and sample within its local neighborhood to generate new candidate samples.
[0208] Specifically, the influence value on each sample in the training sample set. Sort the samples and select one or more samples with the largest influence values as efficient center samples. Let the center samples be... The sample dimension is The neighborhood radius is Construct a Euclidean hypersphere neighborhood around the central sample:
[0209]
[0210] in, Indicates For the center of the ball, The local sampling region is defined by its radius. express 3D real space, Indicates candidate samples to the center sample The L2 norm distance. To generate candidate samples uniformly within this local neighborhood, sampling directions are first randomly generated:
[0211]
[0212] in, Represents the unit sampling direction vector. Let represent a random vector that follows a standard normal distribution. The mean is The covariance matrix is The standard normal distribution. Regenerated radius scaling factor:
[0213]
[0214] Example 12:
[0215] To verify the technical effectiveness of the method of the present invention under different power system scales, a 4-node test system, an IEEE 118-node test system, and a real 588-node power grid system were used for computational verification. The 4-node and 118-node systems were used to verify the mechanistic effectiveness and main performance of the method of the present invention, while the 588-node system was used to verify the applicability and scalability of the method of the present invention in large-scale power grids. Based on the available load sample set, load operation scenarios were generated through uniform random sampling within a preset load fluctuation range and used as unsupervised training data. A total of 15,000 load samples were generated, of which 5,000 samples were used for model training, and the remaining 10,000 samples were used as an independent test set. During the testing phase, the IPOPT solver was used to obtain the reference optimal power flow solution, which was only used for performance evaluation and did not participate in the training of the deep neural network.
[0216] For comparative verification, four methods were set up: M1 is the baseline unsupervised method, which is to perform unlabeled training using a physical information loss function composed of power generation cost and physical constraint penalty terms; M2 is the unsupervised method based on sample reweighting, which dynamically updates sample weights through two-layer optimization under the physical loss objective; M3 is the unsupervised method based on random sample augmentation, which randomly generates additional load samples to augment the training set in each training round; M4 is the unsupervised deep neural network optimal power flow decision augmentation method based on sample utility awareness proposed in this invention.
[0217] Evaluation metrics include: mean squared error The value is used to measure the deviation between the predicted solution and the reference solution of the deep neural network, and the average and maximum values are calculated respectively. The constraint satisfaction rate is used to evaluate the proportion of the predicted solution that satisfies the physical operation constraints, where the constraint satisfaction rates of generator active power, generator reactive power, line power flow, node voltage magnitude, and node phase angle are denoted as follows: , , , and Optimality error This is used to measure the deviation of the predicted solution's operating cost from the reference solution's operating cost; feasibility error. The metric is used to measure the degree to which the predicted solution violates the constraints of the system operation; the training convergence curve is used to compare the loss reduction process and the final convergence level under different training strategies.
[0218] The deep neural network employs a fully connected feedforward structure, with LeakyReLU as the activation function in the hidden layers, Xavier uniform initialization for network weights, and zero initialization for biases. The optimizer is Adam. The number of hidden layer neurons in the three systems is set to 32, 64, and 128, respectively, with a batch size of 32 for all systems. The learning rates are set to 1×10^-3, 1×10^-3, and 5×10^-4, respectively, and the initial loss weights are set to k. g0 =k l0 =k v0 =k δ0 =1; the number of training rounds was set to 40. All comparison methods used the same network structure and operating environment, and were trained on a workstation with a multi-core CPU processor and 32 GB of memory; to ensure the reliability of the results, each experiment was repeated three times and the average value was taken.
[0219] Table 1. Comparison of mean square errors of the four methods under different system sizes.
[0220]
[0221] As shown in Table 1, the method M4 of this invention achieves the lowest average MSE and maximum MSE in 4-node, 118-node, and 588-node systems. Specifically, in the 4-node system, the average MSE and maximum MSE of M4 are 0.86 and 1.53, respectively, lower than M1, M2, and M3; in the 118-node system, the average MSE and maximum MSE of M4 are 0.67 and 0.944, respectively; and in the 588-node system, the average MSE and maximum MSE of M4 are 1.52 and 4.46, respectively. These results demonstrate that this invention, by identifying key samples that significantly impact training loss through sample utility and performing local enhancement and weight updates around these key samples, enables deep neural networks to more fully learn operating regions that are difficult to fit, thereby reducing the prediction error of optimal power flow decisions.
[0222] Table 2. Comparison of constraint satisfaction rates of the four methods under different system scales.
[0223] (a) 4-node system
[0224]
[0225] (b) 118-node system
[0226] (c) 588-node system
[0227]
[0228] As shown in Table 2, the method M4 of this invention achieves a higher constraint satisfaction rate across the three systems, indicating that its output decision has better physical feasibility. Compared to M1, M2, and M3, the method of this invention can identify key samples close to the constraint boundary that have a significant impact on the training loss, and strengthen these samples and their neighborhoods during training, making the model pay more attention to the operating regions where constraint exceedances are likely to occur. For example, in the 588-node system, the generator active power constraint satisfaction rate increased from 88.3%, 91.6%, and 88.4% for M1, M2, and M3 to 93.4% for M4; the generator reactive power constraint satisfaction rate increased to 96.6%. The node phase angle constraint satisfaction rate remained at 100.0% in all systems, demonstrating that this invention improves the accuracy of economic prediction without sacrificing the feasibility of basic physical constraints.
[0229] Table 3 Comparison of the effects of error sample selection and influence value sample selection based on M4
[0230] (a) 4-node system
[0231]
[0232] (c) 588-node system
[0233]
[0234] As shown in Table 3, under the M4 framework, compared to neighborhood sampling around the maximum error sample, sample selection based on influence values achieved lower average optimality error and average feasibility error in all three systems, indicating that influence value sample selection is more conducive to improving overall operational economy and overall feasibility. Although maximum error sample selection is lower in some maximum error metrics in the 118-node and 588-node systems because it directly targets the individual sample with the largest current error, making it easier to reduce the error of that extreme sample, influence value sample selection can identify samples that are more critical to the parameter update and overall loss reduction of the deep neural network. It not only corrects single-point errors but also improves the overall training quality of the model, and is therefore more suitable for sample utility enhancement training in this invention.
[0235] like Figure 2As shown, in the 118-node system, the influence value of a highly useful sample decreases rapidly in the early stages of training and then gradually stabilizes. This phenomenon indicates that the method of this invention can effectively reduce the impact of key samples on the overall training loss. In the early stages of training, the model has not yet fully learned the operational characteristics near the high-influence sample, therefore this sample has a strong effect on parameter updates and loss changes; as training progresses, the model's fitting ability in this key region gradually improves, and the sample's impact on the overall training loss decreases accordingly. In the later stages of training, the influence value does not drop to zero, indicating that this sample still has certain training value and can continuously guide the model to learn the key operational regions that are difficult to fit.
[0236] like Figure 3 and Figure 4 As shown, the distribution of influence values for 10,000 test samples before and after training is compared in 4-node and 588-node systems. Green dots represent influence values before training, and red dots represent influence values after training. Before training, the influence value distribution in the 4-node system is relatively dispersed, with high-influence samples widely distributed across the sample space. In the 588-node system, the influence values are more concentrated before training, indicating that key samples in large-scale systems are more likely to be concentrated in sensitive operating regions. After sample utility-aware training according to this invention, the high-influence regions in both systems are significantly reduced, indicating that this invention can effectively identify samples that have a significant impact on training loss and perform local augmentation around these samples. Simultaneously, the two-layer optimization framework further adjusts the sample weights and deep neural network parameters, allowing the model to be trained more fully in key regions.
[0237] like Figures 5 to 7 As shown, the training convergence curves of different methods in 4-node, 118-node, and 588-node systems are compared. In the 4-node and 118-node systems, M1 to M3 lack targeted training for key sample regions, ultimately converging to a higher loss level. M4 introduces highly useful samples during training and uses these samples to update the model, causing the deep neural network to refocus on highly useful sample regions. Although there may be periodic fluctuations during training, it ultimately achieves a lower convergence loss. In the 588-node system, due to the higher complexity of large-scale optimal power flow learning, the initial loss is larger and the curve fluctuations are more obvious; however, M4 still maintains an overall downward trend and eventually reaches the lowest convergence loss. This indicates that the method of this invention generates new samples and updates sample weights through sample utility, enabling the training process to focus on more valuable samples and exhibiting good scalability for large-scale power grids.
[0238] In summary, the method of this invention demonstrates significant technical advantages in prediction accuracy, constraint satisfaction capability, sample selection effectiveness, key region learning ability, training convergence, and applicability to large-scale systems. By quantifying sample utility through influence functions, this invention can identify the most critical running samples for training deep neural networks under unsupervised conditions. Through efficient neighborhood sampling and two-layer sample weight updates, this invention can enhance the model's learning near constraint boundaries and nonlinearly sensitive regions, thereby reducing average and maximum errors, improving the physical feasibility of output decisions, and reducing the computational burden of traditional AC optimal power flow iterative solutions in multi-scenario online applications.
Claims
1. A method for enhancing AI decision-making in power systems based on sample utility perception, characterized in that, Including the following steps: Step 1: Obtain historical operating data of the power system or operating samples generated by simulation, construct a training sample set, and use the training sample set to train an unsupervised deep neural network model; Step 2: Construct a function to quantify the effect of each training sample on the unsupervised training loss and model parameter updates; The influence function is used to calculate the influence value of each training sample on the overall training loss, and one or more high-efficiency samples are selected from the training sample set according to the magnitude of the influence value. Step 3: Generate candidate samples in the neighborhood of the efficient sample, and update the parameters and weights of the unsupervised deep neural network model synchronously through a two-layer optimization method; Step 4: Use the updated unsupervised deep neural network model to predict the voltage amplitude and phase angle of each node in the power system.
2. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 1, characterized in that, The training sample set uses active load vectors and reactive load vectors as inputs and node voltage magnitudes and phase angles as outputs.
3. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 1, characterized in that, The unsupervised deep neural network model includes an input layer, One hidden layer and one output layer; The outputs of each hidden layer are shown below: in, Indicates the state of the input layer. Indicates the first The output of the hidden layer, and They represent the first Layer weight matrix and bias vector, Represents a non-linear activation function. This represents the optimal power flow state prediction output by the deep neural network.
4. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 1, characterized in that, When training an unsupervised deep neural network model, the goal is to minimize the loss function. The loss function is as follows: in, This represents the total unsupervised training loss; This represents the operating cost item. This indicates the penalty for exceeding the limits of generator active and reactive power output. This indicates a penalty for exceeding the line capacity limit. This indicates a penalty for exceeding the limit of the node voltage amplitude. This indicates the penalty for exceeding the phase angle limit at a node; , , , and These represent the weighting coefficients of the aforementioned loss terms.
5. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 4, characterized in that, The penalties for exceeding operating cost limits, generator active and reactive power output limits, line capacity limits, node voltage amplitude limits, and node phase angle limits are as follows: in, A hinge function that penalizes only the portion exceeding the limit; variables with exponentiation, such as... , , and This represents the corresponding physical quantity calculated by combining the state prediction from a deep neural network with the AC power flow equations.
6. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 1, characterized in that, In step 1, the optimal model parameters after training the unsupervised deep neural network model are shown below: in, Indicates in the sample weight vector The optimal or near-optimal deep neural network parameters are obtained.
7. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 6, characterized in that, The influence function is shown below: in, Indicates sample The impact value, This represents the Hessian matrix of the weighted unsupervised training loss at the current optimal parameters. Describes the Hessian inverse matrix. This represents the overall training loss at the optimal model parameters. The gradient of the parameters at that point, The gradient of the parameters representing the loss of a single sample.
8. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 7, characterized in that, The steps for calculating the impact of each training sample on the overall training loss using the influence function include: Step 1: Transform the influence function into a linear equation, resulting in: in, = Let be the intermediate vector to be solved; The gradient of loss for a single sample; Step 2, initialize the iteration parameters of the conjugate gradient method, i.e.: in, Represents the initial residual. Indicates the initial conjugate search direction; Step 3, iteratively update the parameters of the conjugate gradient method, that is: in, Indicates the first Approximate solution for the next iteration Indicates the corresponding residual. Indicates the conjugate search direction; Step 4, if the residual norm If the number of iterations is less than a preset threshold or the maximum number of iterations is reached, the iteration is terminated, and an approximate solution is used. Calculate the impact of the current training sample on the overall training loss; otherwise, let k = k + 1 and return to step 3. The sample impact values are shown below: In the formula, This represents the sample influence value.
9. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 1, characterized in that, The most efficient sample is one or more samples with the greatest impact.
10. The method for enhancing artificial intelligence decision-making in power systems based on sample utility perception according to claim 1, characterized in that, The steps to generate new candidate samples include: Each high-efficiency sample is used as the central sample, and a Euclidean hypersphere neighborhood is constructed around the central sample, i.e.: in, Indicated by For the center of the ball, The local sampling region is defined by its radius. express 3D real space, Indicates candidate samples to the center sample The 2-norm distance; Randomly generate the sampling direction and radius scaling factor, i.e.: in, Represents the unit sampling direction vector. Let represent a random vector that follows a standard normal distribution. The mean is The covariance matrix is The standard normal distribution; Based on the sampling direction and radius scaling factor, candidate samples are uniformly generated within the Euclidean hypersphere neighborhood of each central sample.