A park power distribution intelligent optimization scheduling method based on load prediction

Abstract

The present application relates to a kind of park power distribution intelligent optimization scheduling method based on load prediction, belong to power system scheduling technical field, for the inconsistency of statistical error index and scheduling economy target in the two-stage separation architecture of existing prediction and scheduling, the park power distribution network optimal power flow scheduling model is converted into convex optimization problem after second-order cone relaxation, embedded in the output end of load prediction network, form differentiable optimization layer, with the actual cost after scheduling and constraint violation penalty to construct decision perception loss function, based on KKT condition to construct implicit gradient back propagation to update prediction network parameters, and introduce graph attention network to aggregate multi-level features and correct scheduling decisions layer by layer, the present application can improve scheduling economy and reduce the risk of equipment safety margin violation.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching technology, specifically to a smart optimization dispatching method for industrial park power distribution based on load forecasting, and more particularly to a dispatching method that performs end-to-end joint training and joint inference between a load forecasting network and a power distribution network optimization dispatching model. This method can be applied to smart energy management of industrial park power distribution networks with distributed photovoltaic, electric vehicle and energy storage system access, and belongs to the technical field of combining smart park energy management, artificial intelligence application and power system dispatching. Background Technology

[0002] With the high proportion of distributed photovoltaic, electric vehicles, and energy storage systems connected to industrial parks, the operation of the park's power distribution network exhibits strong randomness and volatility. To address these operational characteristics, the industry typically employs load forecasting technology to predict electricity load and photovoltaic output over a certain period. This is combined with optimal power flow or mixed-integer linear programming (MILM) optimization models to solve for transformer output, energy storage charging and discharging power, and flexible load regulation, thereby achieving economical and safe and stable operation of the park's power distribution network. The load forecasting stage typically uses deep learning models such as Long Short-Term Memory (LSTM) networks and Transformers, updating model parameters with mean squared error as the training objective function. The scheduling optimization stage uses convex optimization solvers or commercial mixed-integer linear programming solvers to solve scheduling models that include constraints such as equipment capacity, power balance, and energy storage charge-state coupling.

[0003] The closest existing technical solution to this invention is a two-stage separation of prediction and decision-making in a park power distribution dispatching framework. In the first stage, this scheme trains a deep learning prediction model using historical data such as historical load, weather, and photovoltaic output. During the training process, the only optimization objective is to minimize the mean square error between the predicted value and the actual historical load value, thereby obtaining the predicted values ​​of load and renewable energy output for each period in the future dispatching cycle. In the second stage, the predicted values ​​output from the first stage are used as known deterministic parameters and substituted into an optimization dispatching model with the objective of minimizing economic operating costs, including constraints such as equipment capacity, power balance, and peak-valley difference. The model is then solved independently to obtain dispatching instructions such as transformer output, energy storage charging and discharging power, and flexible load adjustment, which are then issued and executed. In this scheme, the training of the prediction model and the solution of the dispatching model are two completely independent modules that are executed sequentially. The prediction model does not know or participate in the solution process and objective function of the dispatching model during the training process.

[0004] The aforementioned existing technical solutions have the following technical defects. First, the mean squared error loss function used in the prediction model training phase assigns equal weight to the prediction errors at all time points. However, the distribution network dispatch cost function has a nonlinear and uneven sensitivity to prediction errors at different time periods and nodes. For example, prediction errors during off-peak load periods have little impact on dispatch costs, while prediction errors during peak load periods or periods near equipment capacity limits may directly lead to equipment overload or trigger punitive power purchase costs. Therefore, prediction models trained with statistical error as the training objective struggle to actively identify and prioritize the prediction accuracy of dispatch-sensitive periods and nodes, ultimately resulting in good prediction statistics but poor actual dispatch economics and insufficient safety margins. Second, the dispatch optimization model includes... The scheduling model contains numerous nonlinear and non-smooth boundary constraints. Errors generated during the prediction phase are amplified by the nonlinear propagation effect of these constraints when substituted into the scheduling model. Small errors that seem acceptable at the prediction level may turn feasible solutions into infeasible solutions after propagation through the scheduling model, leading to consequences such as curtailment of solar power, wind power, or emergency load shedding. Furthermore, since the prediction model and scheduling model are designed as two completely independent, sequential, and non-differentiable modules in existing technical solutions, the training process of the prediction model cannot obtain the objective function and constraint boundary information of the scheduling model. From a technical architecture perspective, there is no channel to feed back scheduling information to the prediction model training stage, making it difficult to fundamentally solve the above-mentioned disconnect problem within the existing technical framework. Summary of the Invention

[0005] The technical problem this invention aims to solve is that in industrial park distribution networks containing distributed photovoltaic power and flexible load access, the existing two-stage technical architecture of prediction and scheduling has the problem of inconsistency between prediction evaluation indicators and scheduling decision objectives. This results in good statistical accuracy of the prediction model but poor actual scheduling economy and high risk of equipment overload. Moreover, the prediction error is easily amplified after propagation through the nonlinear constraints of the scheduling model, causing the scheduling result to be infeasible or the scheduling cost to be abnormally high. This invention aims to provide a joint optimization scheduling method that directly feeds back the physical constraints of the distribution network and the economic scheduling objectives to the load prediction model training process, so that the prediction model can actively perceive the sensitive areas on the scheduling side during the training phase, thereby solving the technical problem of the disconnect between prediction and scheduling.

[0006] To address the aforementioned technical problems, this invention provides a method for intelligent optimization scheduling of power distribution in industrial parks based on load forecasting, which includes the following technical aspects.

[0007] First, historical load data, photovoltaic output data, meteorological data, energy storage status of charge data, and electric vehicle charging time window data are collected at the building, transformer, and park levels within the park to form input features. These input features are then fed into a prediction network containing a graph attention network module and an encoder module. The graph attention network module abstracts the power consumption units at the building, transformer, and park levels as nodes in a topological graph. By calculating the attention coefficients between nodes and their neighboring nodes, the features of neighboring nodes are weighted and aggregated to obtain consistent load features across levels. The encoder module performs time-series encoding on these load features. Finally, the prediction network outputs the predicted load values ​​and photovoltaic output values ​​for each time period within the future scheduling cycle.

[0008] Secondly, the above-mentioned load forecast and photovoltaic output forecast are substituted into the differentiable optimization layer, which is a convex optimal power flow scheduling model after second-order cone relaxation transformation. The total scheduling cost function is used as the optimization objective. The solution is performed under the constraints of equipment capacity, energy storage state of charge, electric vehicle charging time window, and node power balance to obtain the scheduling decision and the corresponding Lagrange multiplier. The total scheduling cost function is composed of a weighted sum of economic operating cost, peak-valley difference penalty, and equipment safety margin penalty.

[0009] Next, based on the aforementioned scheduling decision, the actual load value for the corresponding time period, and the actual photovoltaic output value, a decision-aware loss function is calculated. This loss function is composed of a weighted average of the actual scheduling cost term obtained by substituting the scheduling decision into the real scenario and the constraint violation penalty term. This allows the loss function to directly reflect the scheduling consequences rather than simply statistical errors. To enable gradient backpropagation of the sensitivity of the scheduling decision to the predicted value, this invention constructs the Lagrangian function and KKT condition equations corresponding to the optimization problem. By inverting the KKT matrix composed of the KKT condition equations, the implicit gradients of the scheduling decision with respect to the predicted load and photovoltaic output values ​​are obtained. According to the chain rule, this implicit gradient is multiplied by the gradient of the prediction network itself with respect to the network parameters to obtain the total gradient of the decision-aware loss function with respect to the prediction network parameters. The prediction network parameters are then updated accordingly. This process is iterated repeatedly until the preset training termination condition is met, resulting in a trained prediction network.

[0010] Finally, in the inference phase, the trained prediction network is used to perform forward calculations on the real-time data of the park to obtain real-time load forecasts and real-time photovoltaic output forecasts. These are then substituted into the differentiable optimization layer to obtain real-time scheduling decisions. The building-level flexible load response actions in the scheduling decisions are aggregated layer by layer in the order of building level, transformer level, and park level. The aggregation results are checked to see if they exceed the capacity constraints of the previous level. If they do, the response actions of the corresponding buildings are proportionally corrected according to the adjustment margin of each building's flexible load, and the aggregation is checked again until the aggregation results meet the capacity constraints of all levels. Finally, the scheduling instructions that meet the multi-level consistency constraints are issued and executed.

[0011] The technical effects that this invention can achieve include: 1. By directly incorporating the scheduling consequences into the training loss of the prediction network, the prediction model can proactively perceive the economic costs and safety constraints of the scheduling side during the training phase. Under similar prediction statistical error levels, this can reduce the actual total scheduling cost and the probability of equipment safety margin violations. 2. By using an implicit gradient backpropagation mechanism based on KKT conditions, the scheduling model achieves analytical transmission of gradient information under non-smooth constraints, resulting in a more stable training process compared to heuristic patching methods. 3. By using graph attention networks to achieve cross-level aggregation of building-level, transformer-level, and park-level features, and with the help of layer-by-layer aggregation verification and proportional correction mechanisms, the aggregated response actions of the bottom flexible loads meet the capacity constraints of the upper layers, which is conducive to improving the level of local consumption of new energy. Attached Figure Description

[0012] Figure 1 This is a flowchart of the training process for the decision perception and prediction network of this invention. Figure 2 This is a flowchart of the differentiable optimization layer scheduling decision-solving process of the present invention; Figure 3 This is a flowchart of the multi-level consistency correction process of the present invention. Detailed Implementation

[0013] The following is in conjunction with the appendix Figure 1-3The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. The intelligent optimization scheduling method for industrial park power distribution provided by the present invention is applicable to industrial park power distribution networks with rooftop photovoltaic, electric vehicle charging facilities and energy storage systems connected. The overall architecture of the method includes four functional links that cooperate with each other: data perception, decision perception and prediction, differentiable optimization scheduling and multi-level consistency correction. The four links form a closed-loop joint optimization structure through forward data flow and reverse gradient flow.

[0014] I. Data Perception and Multi-level Feature Aggregation In one specific implementation, the park's power distribution network collects building-level electricity consumption data through advanced metering infrastructure devices, and collects transformer-level and park-level operational data through a monitoring and data acquisition system. The collected data includes historical load sequences, photovoltaic output sequences, meteorological characteristics, energy storage state of charge, and electric vehicle charging time window information at the building, transformer, and park levels. The above data constitutes the input feature tensor, denoted as the feature matrix.

[0015] To achieve consistent aggregation of features across building, transformer, and park levels, this invention abstracts the park power distribution network as a topology graph. The node set represents the power consumption units at the building, transformer, and park levels, and the edge set represents the electrical connections between units at each level. For any node and its set of neighboring nodes, the feature aggregation process of the graph attention network is implemented in the following way: ; ; in, Represents a node The original feature vector contains the node's historical load, photovoltaic output, and flexible load adjustment margin. Represents a node The set of neighboring nodes, This is a learnable linear transformation weight matrix used to map the original features to a unified feature space. For a learnable attention weight vector, the symbol is... This represents a vector concatenation operation. Represents a node Its neighboring nodes Attention coefficient, used to measure the node For nodes The importance of feature aggregation is determined by normalizing the attention scores of all candidate nodes within the neighbor node set. Represents a non-linear activation function. Represents a node The updated feature vector obtained after aggregating neighbor information, together with the graph attention network module and the encoder module for temporal coding, constitutes the prediction network of this invention. After time-series encoding by the encoder module, the prediction network outputs the load forecast and photovoltaic output forecast values ​​for each time period within the future scheduling cycle, which are denoted as the forecast values. .

[0016] II. Construction of Differentiable Optimizable Layers This invention abandons the traditional approach of using only statistical error as the training objective in prediction models, and transforms the industrial park power distribution network dispatching model into a component of the loss function, assuming the dispatching decision variable is... Including transformer output, energy storage charging and discharging power, and flexible load regulation, the optimal power flow scheduling model can be expressed as the following parameterized optimization problem: ; ; in, The total scheduling cost function is composed of a weighted average of the economic operating cost term, the peak-to-valley difference penalty term, and the equipment safety margin penalty term. This represents the set of inequality constraints, including equipment capacity constraints, upper and lower limits of energy storage state of charge constraints, and electric vehicle charging time window constraints. This represents the set of equality constraints, including nodal power balance constraints. Indicates that at a given predicted value The scheduling decision is obtained by solving the scheduling model under the given conditions.

[0017] To make the optimization problem differentiable during the training phase, this invention employs second-order cone relaxation to transform the non-convex power flow equations into convex constraint forms, making the scheduling model a convex optimization problem overall. This ensures that the corresponding KKT condition equations have a unique solution and are differentiable with respect to the input variables. Based on this, this invention constructs the following decision-aware loss function: ; in, This represents the actual load and the true value of photovoltaic output. This represents the actual scheduling cost incurred after substituting the scheduling decisions obtained from the predicted values ​​into the real-world scenario. This represents the constraint violation penalty function, used to quantify the degree of violation of inequality constraints. Specifically, it can be implemented by summing the squares of the positive violation portions of the inequality constraints. The penalty coefficient is used to balance the weight between the actual scheduling cost and the constraint violation penalty. The decision perception loss function mentioned above is fundamentally different from the traditional statistical error loss function. Its value is directly determined by the scheduling consequences, rather than by the numerical difference between the predicted value and the actual value.

[0018] III. Implicit Gradient Backpropagation Mechanism Based on KKT Conditions because The implicit function is obtained by solving an optimization problem and cannot be directly obtained through the backpropagation chain of a conventional neural network. To calculate the gradient, this invention proposes a gradient calculation method based on KKT conditional implicit differentiation, and constructs the Lagrangian function for this optimization problem: ; in, Let Lagrange multiplier vectors be the vectors corresponding to the inequality constraints. Let Lagrange multiplier vectors be the equation constraints, used in scheduling decisions. At this point, the problem satisfies the following KKT conditions: ; ; ; The above conditions are applied to scheduling decisions. Regarding Taking the total differential, we obtain the following system of linear equations using the KKT matrix, where the coefficient matrix is ​​denoted as the KKT matrix. : ; ; in, Represents the Lagrange function with respect to decision variables. The second derivative matrix, , These represent inequality constraints and equality constraints, respectively. The first derivative matrix, Represent the Lagrange function pair and The mixed second derivative matrix, , These represent the constraint pairs. The first derivative matrix, obtained by applying the matrix By performing the inverse operation, the scheduling decision can be obtained. For predicted values Implicit gradient: ; in, Representation matrix Zhongyu The corresponding row block in the matrix In implementations with large dimensions and where direct inversion is time-consuming, the conjugate gradient method can be used to iteratively solve the above linear equations to obtain an approximate value of the implicit gradient, thus replacing the matrix... Explicit inversion is performed to improve computational efficiency while maintaining gradient accuracy. This implicit gradient is then combined with the gradient of the prediction network itself along the backpropagation chain of the neural network. Multiply, where By representing the predicted network parameters, we can obtain the total gradient of the decision-aware loss function with respect to the predicted network parameters: ; A gradient descent optimizer is used to update the prediction network parameters based on the total gradient mentioned above. The process of forward prediction, scheduling solution, loss calculation, implicit gradient solution and parameter update is repeatedly executed until the decision-aware loss function meets the preset training termination condition, and the trained decision-aware prediction model is obtained.

[0019] IV. Multi-level Consistency Correction Mechanism To ensure that the aggregated responses of the building-level flexible loads meet the capacity constraints of the upper-level transformer and park levels, this invention sets up a multi-level consistency correction stage after the graph attention feature aggregation module and scheduling decision output of the prediction network. During the inference phase, the obtained building-level flexible load responses are aggregated layer by layer in the order of building level, transformer level, and park level, and the aggregation result is checked to see if it exceeds the capacity constraint of the upper level. If the aggregation result exceeds the upper level limit, the response of the corresponding building is proportionally corrected according to the adjustment margin of each building's flexible load, and the aggregation check is re-executed until the constraints of all levels are met. Finally, a scheduling instruction that meets the multi-level consistency constraints is output and sent to the park microgrid edge controller for execution.

[0020] V. Implementation methods in specific scenarios To further illustrate the specific implementation process of the present invention, the following description is based on a campus power distribution network scenario containing three buildings. The campus power distribution network includes Building A, Building B, and Building C. Building A and Building B are connected to the campus medium-voltage busbar through Transformer 1, and Building C is connected to the campus medium-voltage busbar through Transformer 2. Each building is equipped with rooftop photovoltaics and three types of flexible loads: movable, reduceable, and transferable. The campus is also equipped with a centralized energy storage system.

[0021] During the offline training phase, the load sequence, photovoltaic output sequence, meteorological characteristics, energy storage state of charge, and electric vehicle charging time window information for the three buildings and two transformers within the historical scheduling cycle are first retrieved from the historical database to form an input feature tensor. This feature tensor is then input into a graph attention network module for feature aggregation from the building level to the transformer level and from the transformer level to the park level. It is then fed into the encoder module for forward computation to obtain the predicted load and photovoltaic output values ​​for the three buildings and two transformers at each time period within the future scheduling cycle. These predicted values ​​are then substituted into a differentiable optimization layer to construct an optimal power flow scheduling model with economic operating cost, peak-valley difference penalty, and equipment safety margin penalty as objectives. A second-order cone relaxation is then used to transform this model. As this is a convex optimization problem, a convex optimization solver is invoked to solve the scheduling model, obtaining the scheduling decisions corresponding to transformer output, energy storage charging and discharging power, and flexible load adjustment of each building, as well as the corresponding Lagrange multipliers. Based on the scheduling decisions and the actual load values ​​corresponding to the batch of samples, the decision-aware loss function is calculated, and the KKT condition matrix of the optimization problem is constructed. By inverting the matrix or iteratively solving the conjugate gradient, the implicit gradient of the scheduling decision with respect to the predicted value is obtained. The total gradient of the decision-aware loss function with respect to the prediction network parameters is calculated according to the chain rule, and the prediction network parameters are updated using a gradient descent optimizer. The above process is repeated until the decision-aware loss function meets the preset training termination condition, and the trained decision-aware prediction model is output.

[0022] During the online inference phase, real-time load, photovoltaic output, and equipment status data of Building A, Building B, Building C, Transformer 1, Transformer 2, and the park busbar are collected. A multi-level topology graph is constructed, and this graph is input into a pre-trained graph attention network module for feature aggregation to obtain a load feature representation with cross-level consistency. This feature representation is then input into a pre-trained decision-aware prediction network to perform forward calculation and output the predicted load and photovoltaic output values ​​of the three buildings and two transformers within the future scheduling cycle. The predicted values ​​are then substituted into a differentiable optimization layer to satisfy transformer capacity constraints, energy storage state-of-charge coupling constraints, and electric vehicle charging time window constraints. Under the given conditions, the scheduling decision is solved. If the aggregated flexible load response actions of Building A and Building B exceed the capacity constraint of Transformer 1, the response actions of Building A and Building B are proportionally corrected according to their respective adjustment margins, so that the corrected aggregated result does not exceed the capacity constraint of Transformer 1. The aggregation verification from building level to transformer level and from transformer level to park level is re-executed until the constraints of all levels are met. Finally, a scheduling instruction that satisfies the multi-level consistency constraints is output and sent to the park microgrid edge controller. The edge controller controls the corresponding equipment to perform energy storage charging and discharging, flexible load response, and transformer output adjustment actions.

[0023] VI. Other alternative implementation methods In an alternative implementation, for scenarios where the KKT matrix has a large dimension and direct inversion is inconvenient, a logarithmic barrier function can be used to soften the inequality constraints in the differentiable optimization layer into differentiable penalty terms, constructing an approximate softened scheduling model. This makes the overall mapping relationship between scheduling decisions and predicted values ​​differentiable. Then, an automatic differentiation tool can be used to directly backpropagate the softened scheduling model to obtain an approximate value of the implicit gradient, replacing the need for precise inversion of the KKT matrix. This implementation has certain advantages in terms of computational efficiency.

[0024] In another alternative implementation, the prediction network and the scheduling decision-making process can be modeled as a policy network. The operating status of the park's distribution network is taken as the environmental observation, the scheduling action is taken as the policy output, and the economic cost and safety margin after scheduling are taken as the reward signal. A policy optimization algorithm with a value evaluation structure is used to train the policy network containing the load prediction module. This implementation does not require explicit solution of convex optimization problems and KKT conditions, and can be used as an alternative implementation of the present invention in scenarios where the distribution network operating status exhibits strong non-convex characteristics.

[0025] The above specific embodiments are merely illustrative examples of the technical solutions of the present invention. The scope of protection of the present invention is not limited to the number of buildings, the number of transformers, and the specific parameter values ​​involved in the above specific scenarios. Any campus power distribution network scheduling method implemented by adopting any one or a combination of the decision perception loss function construction method, the implicit gradient backpropagation mechanism based on KKT conditions, and the multi-level consistency correction mechanism described in the present invention falls within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization scheduling of power distribution in industrial parks based on load forecasting, characterized in that, include: Historical load data, photovoltaic output data, meteorological data, energy storage status of charge data, and electric vehicle charging time window data are collected at the building level, transformer level, and park level within the park to form input features; The input features are input into a prediction network, which includes a graph attention network module and an encoder module. The graph attention network module abstracts the power consumption units at the building level, transformer level and park level into topology graph nodes and performs cross-level feature aggregation to obtain consistent load features across levels. The encoder module performs time-series encoding on the load features to obtain the load prediction value and photovoltaic output prediction value for each time period in the future scheduling cycle. Substitute the predicted load value and the predicted photovoltaic output value into the differentiable optimization layer. The differentiable optimization layer is a convex optimal power flow scheduling model after second-order cone relaxation transformation. The total scheduling cost function is used as the optimization objective. The solution is obtained under the constraints of equipment capacity, energy storage state of charge, electric vehicle charging time window, and node power balance, so as to obtain the scheduling decision and the corresponding Lagrange multiplier. Based on the scheduling decision, the actual load value for the corresponding time period, and the actual photovoltaic output value, a decision perception loss function is calculated. The decision perception loss function is composed of a weighted average of the actual scheduling cost term obtained by substituting the scheduling decision into the real scenario and the constraint violation penalty term. Construct the KKT condition equation set for the optimization problem corresponding to the differentiable optimization layer, and obtain the implicit gradient of the scheduling decision with respect to the load prediction value and the photovoltaic output prediction value by performing the inversion operation on the KKT matrix composed of the KKT condition equation set. According to the chain rule, the implicit gradient is multiplied by the gradient of the prediction network with respect to the network parameters to obtain the total gradient of the decision-aware loss function with respect to the prediction network parameters, and the parameters of the prediction network are updated according to the total gradient. Repeat the process of obtaining predicted values, solving scheduling decisions, calculating the decision-aware loss function, solving implicit gradients, and updating network parameters until the preset training termination condition is met, and a trained prediction network is obtained. The trained prediction network is used to perform forward calculations on the real-time data of the park to obtain real-time load predictions and real-time photovoltaic output predictions. These real-time load predictions and real-time photovoltaic output predictions are then substituted into the differentiable optimization layer to obtain real-time scheduling decisions, which are then issued and executed.

2. The method according to claim 1, characterized in that, The graph attention network module performs weighted aggregation of neighbor node features by calculating the attention coefficients between a node and its neighbor nodes. The encoder module performs time-series encoding on the load features to obtain time-series encoded features. The prediction network outputs the load prediction value and the photovoltaic output prediction value based on the time-series encoded features.

3. The method according to claim 2, characterized in that, Also includes: The building-level flexible load response actions in the scheduling decision are aggregated layer by layer in the order of building level, transformer level and park level, and the aggregation result is checked to see if it exceeds the capacity constraint of the previous level. If the aggregation result exceeds the capacity constraint of the previous level, the response action of the corresponding building is proportionally corrected according to the adjustment margin of each building's flexible load, and the aggregation is checked again until the aggregation result meets the capacity constraint conditions of all levels.

4. The method according to claim 1, characterized in that, The total scheduling cost function is composed of a weighted average of the economic operating cost term, the peak-valley difference penalty term, and the equipment safety margin penalty term.

5. The method according to claim 1, characterized in that, The constraint violation penalty term is obtained by quantifying the degree of violation of the inequality constraint, and the decision perception loss function is the sum of the product of the actual scheduling cost term, the penalty coefficient, and the constraint violation penalty term.

6. The method according to claim 1, characterized in that, The differentiable optimization layer transforms the non-convex power flow equations into convex constraint forms through second-order cone relaxation, ensuring that the KKT condition equations of the optimization problem corresponding to the differentiable optimization layer have a unique solution.

7. The method according to claim 1, characterized in that, The KKT condition equations are obtained by taking the total differential of the Lagrange function at the scheduling decision point with respect to the load forecast and the photovoltaic output forecast. The Lagrange function is composed of the total scheduling cost function, the product of the inequality constraint term and the corresponding Lagrange multiplier, and the product of the equality constraint term and the corresponding Lagrange multiplier.

8. The method according to claim 1, characterized in that, The conjugate gradient method is used to iteratively solve the linear equation system formed by the KKT matrix to obtain an approximate value of the implicit gradient, which replaces the inversion operation of the KKT matrix.

9. The method according to claim 2, characterized in that, The process of using the trained prediction network to perform forward computation on real-time data of the park includes: collecting real-time load data, photovoltaic output data, and equipment status data at the building level, transformer level, and park level at the current moment; constructing a multi-level topology graph; inputting the multi-level topology graph into the trained graph attention network module for feature aggregation to obtain real-time load features; inputting the real-time load features into the trained encoder module for time-series encoding to obtain real-time time-series encoded features; and inputting the real-time time-series encoded features into the prediction network for forward computation to obtain the real-time load prediction value and the real-time photovoltaic output prediction value.

Images