A wind-solar power prediction and power distribution network dispatching method based on dynamic feedback of decision effectiveness

Abstract

This invention discloses a method for wind and solar power prediction and distribution network scheduling based on dynamic feedback of decision-making efficiency, belonging to the field of distribution network scheduling technology. The method includes: first, constructing a high-precision prediction method for wind and solar power at distribution network nodes based on an improved Crossformer method; second, constructing a multi-objective optimization scheduling model considering safety and economy, and efficiently solving the model using an improved second-order cone relaxation method; further, constructing a proxy dynamic efficiency deviation index adapted to the time-varying feasible region to quantify the differentiated impact of prediction errors on scheduling objectives at different times in real time; finally, using the gradient backpropagation of this dynamic index to perform closed-loop training of the prediction model. This invention achieves deep integration of prediction and scheduling by feeding back the decision-making merits at the scheduling level to the prediction level, effectively improving the efficiency of wind and solar power absorption and the safety of power grid operation.

Description

Technical Field

[0001] This invention relates to power distribution network dispatching technology, specifically to a method for predicting wind and solar power and dispatching power distribution networks based on dynamic feedback of decision-making efficiency. Background Technology

[0002] With the advancement of the global energy transition, the penetration rate of new energy sources, represented by wind power and photovoltaics, in distribution networks is continuously increasing. However, the strong randomness and volatility of wind and solar power output pose significant challenges to the safe and economical operation of active distribution networks. Current distribution network dispatch typically follows a "prediction first, optimization later" model: first, a prediction model is trained with the goal of minimizing prediction error; then, the prediction results are used as deterministic inputs to solve the dispatch problem in the optimization model. This traditional model has the following significant drawbacks:

[0003] (1) Mismatch between prediction and decision-making objectives: Traditional prediction models only focus on statistical numerical accuracy, ignoring the differentiated impact of prediction errors on scheduling objectives at different times and under different grid conditions. For example, during peak grid load or heavy line load periods, small prediction errors may lead to serious over-limit risks or high penalty costs; while during light load periods, larger errors may have little impact on scheduling decisions. High-precision prediction results often cannot be translated into optimal scheduling decisions.

[0004] (2) The physical feasible region of the distribution network changes dynamically over time. Existing prediction methods fail to fully perceive this time-varying constraint, which may result in the generated prediction curves being infeasible in subsequent scheduling.

[0005] (3) The wind and light data contain a lot of non-Gaussian noise and have complex cross-time and cross-dimensional dependencies. Traditional models have difficulty retaining key feature information while denoising, which limits the upper limit of prediction.

[0006] Therefore, there is an urgent need for a closed-loop prediction and scheduling method that can feed back the quality of decisions at the scheduling level to the prediction level and effectively handle noise and complex dependencies. Summary of the Invention

[0007] This invention aims to address the disconnect between prediction accuracy and decision quality in the existing "predict first, optimize later" model, as well as the pain points of high noise interference and difficulty in capturing spatiotemporal dependencies in wind and solar data. This invention proposes a wind and solar power prediction and distribution network scheduling method based on dynamic feedback of decision performance. By constructing a proxy dynamic performance deviation index, a gradient backpropagation path is established from the scheduling layer to the prediction layer, achieving deep integration of prediction and scheduling.

[0008] This invention employs the following technical means to achieve: a method for wind and solar power prediction and distribution network dispatch based on dynamic feedback of decision-making efficiency, comprising the following steps:

[0009] Step 1: Construct a spatiotemporal correlation prediction model. Use an improved Crossformer neural network to model the wind and solar power of each node in the distribution network, capture the cross-time and cross-dimensional dependencies between multidimensional meteorological features and power sequences, and generate power prediction sequences.

[0010] Step 2: Construct a distribution network optimization scheduling model under time-varying constraints. Input the power prediction sequence into the multi-objective optimization scheduling model. Under the premise of satisfying the time-varying physical feasible region constraint of the power grid, use the improved second-order cone relaxation method to efficiently solve the model and generate the day-ahead scheduling strategy.

[0011] Step 3: Construct an agent dynamic performance deviation index. Based on the agent dynamic decision performance value theory, construct a performance deviation evaluation function that adapts to the time-varying feasible region to quantify the scheduling performance gap between the scheduling strategy generated in Step 2 and the optimal strategy based on real power.

[0012] Step 4: Perform dynamic closed-loop feedback training, construct a hybrid loss function that takes into account both prediction accuracy loss and decision performance loss, and use the gradient backpropagation algorithm of this function to update the closed-loop parameters of the improved Crossformer neural network.

[0013] The beneficial effects of adopting the above technical solution in this invention are:

[0014] 1. This invention breaks through the limitations of the traditional one-way open-loop "prediction first, optimization later" approach. By introducing a proxy dynamic performance deviation index, it establishes a gradient feedback mechanism from the scheduling layer to the prediction layer. The model no longer simply pursues numerical fitting, but is guided by decision performance, automatically avoiding prediction errors at key constraint boundaries, significantly improving the safety and economy of scheduling decisions.

[0015] 2. By embedding a TRPCA layer within the Crossformer architecture to construct the ITSA structure, this invention effectively removes non-Gaussian noise using low-rank sparse decomposition theory. This ensures that cross-dimensional feature fusion is based on clean low-rank data, significantly improving the model's robustness and feature extraction accuracy under unstable data quality conditions.

[0016] 3. Addressing the characteristic of scheduling problems having multiple optimal solutions, this invention proposes an "unambiguous decision performance value." By anchoring to the most vulnerable decision boundary within the optimal solution set, it provides a clear and robust gradient guide for model training. This effectively prevents the model from degenerating and converging to all-zero trivial solutions, ensuring the convergence stability of the algorithm under uncertain environments.

[0017] 4. A multi-objective optimization model covering voltage quality, network loss, and renewable energy consumption was constructed and deeply coupled with closed-loop prediction. Under the premise of strictly meeting physical safety constraints such as voltage, the model effectively reduced system network losses and maximized the utilization rate of distributed power sources, achieving the overall optimal efficiency of active distribution network operation. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0019] Figure 1 This is a flowchart illustrating the wind and solar power prediction and distribution network scheduling method based on dynamic feedback of decision-making efficiency.

[0020] Figure 2 The active power loss of the entire IEEE 69-node network at all times.

[0021] Figure 3 The optimized voltage distribution map of the entire network at 12:00.

[0022] Figure 4 This is the photovoltaic power output curve for all time periods. Detailed Implementation

[0023] The following will refer to the appendices in the embodiments of the present invention. Figure 1-4 The technical solutions in the embodiments of the present invention will be clearly and completely described.

[0024] The aforementioned method for wind and solar power prediction and distribution network dispatching based on dynamic feedback of decision-making efficiency includes the following process:

[0025] Figure 1 This is a schematic diagram of the process architecture of a wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency.

[0026] The infrastructure computation for the improved Crossformer model described in step 1 includes:

[0027] 1) Utilize a Dimensionally Segmented Embedding Layer (DSW) to segment each dimension into segments of length 1. The data points are divided into lengths of The segments are divided, and each segment is embedded into a vector through linear projection and positional embedding:

[0028] (1);

[0029] In the formula, It is a two-dimensional vector with dimension . The Time series; For learnable projection matrix; yes The length of the dimension is The Segment vector; for Weizhongdi Learnable embedding positions of segment vectors.

[0030] 2) Capture cross-temporal and cross-dimensional dependencies using a two-stage attention layer (TSA), with its cross-dimensional stage output... The calculation is as follows:

[0031] (2);

[0032] In the formula, This is the output of the multilayer perceptron; , These are the input and output of the TSA, respectively; This represents the TSA function.

[0033] 3) Prediction is performed using a hierarchical encoder-decoder (HED), and the final prediction result is... The sum of the prediction results for each layer:

[0034] (3);

[0035] In the formula, ; It is the first The prediction results for the layer; The current moment; To predict the length of the time domain.

[0036] The Crossformer model is improved by introducing a TRPCA layer to construct the ITSA, which will output the results across time phases. Decomposed into a low-rank clean matrix and sparse noise matrix The optimization problem of the decomposition process is expressed as:

[0037] (4)

[0038] In the formula, f is the objective function of the optimization problem of the decomposition process; Indicates and For variable pairs The objective of minimizing the rank is to solve the problem. Represents the 0 norm; yes The weights.

[0039] The improved Crossformer model introduces a closed-loop feedback mechanism, utilizing cross-dimensional stage outputs. Output across time stages Perform adaptive calibration, and correct the features. The calculation formula is:

[0040] (5);

[0041] In the formula, For activation functions; This is the feedback intensity coefficient; Represents element-wise multiplication; A learnable mapping matrix; corrected features It is substituted into the TRPCA optimization model to replace the original input.

[0042] Step 2, which involves constructing and solving a proactive distribution network safety and economic optimization scheduling model, includes:

[0043] (1) Establishment of an active distribution network safety and economic optimization scheduling model

[0044] To ensure the safe, reliable, and economical operation of active distribution networks, this invention constructs a distribution network optimization scheduling model considering multi-objective collaboration. The specific objective function comprises the following three parts: 1) minimizing system network losses; 2) minimizing node voltage deviations; and 3) maximizing distributed generation output. Furthermore, the analytic hierarchy process (AHP) is used to transform the multi-objective optimization scheduling problem into a single-objective optimization scheduling problem, providing theoretical support for the generation of subsequent scheduling schemes.

[0045] 1) Minimize system active power loss:

[0046] (6);

[0047] In the formula, T represents the total number of observation periods; G ij V is the conductance of line ij; i and V j θ represents the voltage amplitude at nodes i and j; Δt represents the time interval between the two observation points; θ ij Let be the phase angle difference between nodes i and j.

[0048] 2) Minimize the absolute value of voltage offset:

[0049] (7);

[0050] The nodes at the end of the line and those connected to state variables are defined as weak nodes in the system. The objective function for optimizing other regular nodes is shown above. In the formula, N is the number of regular nodes; V i,t V i,ref Let be the voltage amplitude and the expected voltage value of the i-th node at time t.

[0051] To address the system's weak points, penalties are applied to the parts that exceed limits. The objective function at this point is...

[0052] (8);

[0053] In the formula, N weak denoted as the number of weak nodes; k is the penalty coefficient.

[0054] 3) Maximize DG output:

[0055] (9);

[0056] In the formula, N G The number of distributed power sources in the system; Let be the output of DG at node i at time t.

[0057] It is worth noting that the three indicators mentioned above do not belong to the same dimension and need to be normalized before being combined and weighted using the analytic hierarchy process to obtain the final objective function.

[0058] (10);

[0059] In the formula, α, β, and γ are the corresponding weight coefficients, determined by the analytic hierarchy process, where α + β + γ = 1; 10 with f 20 These represent the system loss and voltage deviation when the system is not optimized, f. 30 This represents the total installed capacity of DG.

[0060] (2) Solving the active distribution network safety and economic optimization scheduling model

[0061] The optimization objective problem can be described as follows:

[0062] (11);

[0063] In the formula, It is a decision vector; Let p be the target value; the inequalities Ap + Bu ≤ g(t) and Fv ≤ k are collectively referred to as the feasible region S; p is the decision vector, representing the core independent variable to be optimized; A is the first coefficient matrix corresponding to the decision vector and p; u is the input vector or external influence parameter vector; B is the second coefficient matrix corresponding to the input vector and u; g(t) is the constraint threshold vector (or time-varying boundary vector) that changes with time t; F is the constraint coefficient matrix of the relevant variable v; v is the system state variable vector or auxiliary decision vector; k is the constant constraint vector, representing the fixed boundary or limit threshold.

[0064] To address the problem of safe and economical optimal scheduling in active distribution networks, this invention establishes a mathematical model based on second-order cone programming (SOCP). While second-order cone relaxation can transform non-convex power flow constraints into convex constraints to improve solution efficiency, it often exhibits non-zero duality gaps when facing the highly dynamic characteristics of the network and complex operating conditions. This makes it difficult for the calculated scheduling strategy to strictly satisfy Kirchhoff's laws at the physical level, thus affecting the safety of grid operation. Therefore, this invention further introduces an iterative recovery algorithm based on a relaxation error penalty function, building upon the second-order cone solution. This method uses the second-order cone relaxation solution as the initial value, constructs an error term representing the deviation between the relaxation variables and the physical truth value, and introduces it as a penalty function into the objective function. This ensures both scheduling economy and the physical feasibility of the solution, as well as the strict safety of grid operation.

[0065] Node power balance constraints:

[0066] (12);

[0067] In the formula, k is the index of the iteration number of the iterative algorithm; This is the set of downstream child nodes connected to node j; , These represent the active and reactive power flowing through branch (i,j) during time period t, respectively. Let be the square of the current amplitude flowing through branch (i,j) during time period t; , These represent the active and reactive power generated by the distributed power source at node i during time period t, respectively. , These represent the active and reactive power of the load at node j during time period t; X ij These are the resistance and reactance values ​​of branch (i,j), respectively.

[0068] Node voltage drop constraints:

[0069] (13);

[0070] In the formula, , These are the squares of the voltage amplitudes at nodes i and j, respectively.

[0071] Second-order cone relaxation constraint (SOCR):

[0072] (14);

[0073] In the formula, the symbol It is the Euclidean norm.

[0074] To ensure the physical feasibility of the solution, an iterative recovery algorithm based on penalty functions is used for iterative solution. The specific steps are as follows:

[0075] (1) Set the iteration number k=0 and initialize the penalty factor. .

[0076] (2) The second-order cone relaxation method is used to solve the optimization scheduling model and obtain the relaxed solution.

[0077] (3) Calculate the maximum relaxation error of each branch. , where the superscript k represents the number of iterations.

[0078] (4) If ( The preset convergence accuracy is 10. -4 If the solution is deemed physically feasible, the optimal scheduling strategy is output, and the algorithm terminates.

[0079] (5) If convergence fails, dynamically adjust the penalty factor based on the error magnitude, increasing the penalty weight for branches with larger errors: , Let be the step size coefficient. Return to step 2.

[0080] The construction of the agent decision-making effectiveness deviation index mentioned in step 3 includes:

[0081] (1) Decision effectiveness value Construction

[0082] Decision efficiency value The optimal target deviation is defined as the error between the prediction and the actual decision. The formula is as follows:

[0083] (14);

[0084] In the formula, Representatives contribute to the forecasting of scenic spots The optimal decision objective is as follows; Representatives contribute to the actual scenery The optimal decision objective is determined by the given conditions.

[0085] (2) Construction of unambiguous decision performance value

[0086] In the process of making decisions based on predictions, for the same predicted wind and solar power value There are often multiple different scheduling decision schemes. All solutions satisfy the constraints and achieve the same theoretically optimal objective value, forming a non-unique optimal solution set. This leads to the prediction model degenerating incorrectly and converging to trivial solutions of all zeros during training due to the lack of explicit gradient guidance, thus losing its ability to regulate the dynamics of the real power grid. Therefore, this invention constructs an unambiguous decision performance value. By selecting the most vulnerable decision in the optimal solution set, the model is forced to learn the upper bound of the decision's risk, thereby improving the system's robustness.

[0087] (15);

[0088] In the formula, This indicates the decision with the weakest resistance to disturbances; The target corresponding to the decision with the weakest resistance to disturbances .

[0089] (3) Construction of easily manageable agency decision-making performance deviation indicators

[0090] The unambiguous decision performance values ​​described above may be non-convex and discontinuous, which is detrimental to the application of machine learning models because their backpropagation is achieved through gradient computation. By employing duality theory and approximate transformations, a tractable dynamic form of surrogate decision performance value is developed.

[0091] The unambiguous decision effectiveness value can be derived in the following form:

[0092] (16);

[0093] In the formula, This represents the surrogate target value after dual transformation. This represents the predicted power output of wind and solar power at time t. The optimal decision objective after dual transformation; Represents the actual wind and light output at time t. The optimal decision objective after dual transformation; This is expressed as finding the minimum upper bound for reducing the relaxation duality gap; It is an introduced proportional parameter.

[0094] Step 4 describes the execution of the dynamic closed-loop feedback training process, which includes:

[0095] (1) Constructing a hybrid loss function

[0096] Construct a hybrid loss function that includes prediction accuracy loss and decision performance loss. To minimize scheduling decision efficiency while ensuring the accuracy of wind and solar power prediction; the hybrid loss function The calculation formula is as follows:

[0097] (17);

[0098] In the formula, These are hyperparameter weight coefficients used to balance the contributions of loss terms with different dimensions to the gradient, ensuring that the model is not biased towards a single metric with a large numerical magnitude during training. To predict accuracy loss, mean square error is used for calculation, as shown in the following formula:

[0099] (18);

[0100] In the formula, N is the number of training samples; This represents the actual wind and solar power output at the distribution network nodes. To improve the predicted values ​​generated by the Crossformer model.

[0101] (2) Perform dynamic closed-loop feedback training

[0102] Based on the constructed hybrid loss function The gradient descent algorithm is used to perform end-to-end closed-loop training on the improved Crossformer neural network in step 1.

[0103] The specific process is as follows: Calculate the hybrid loss function. To improve the gradient of the internal parameters of the Crossformer model, the gradient information is transmitted from the scheduling decision layer back to the prediction model layer using the backpropagation algorithm. The learnable projection matrix, attention weights and fully connected layer parameters in the Crossformer are iteratively updated until the hybrid loss function converges or reaches the preset number of training rounds, thereby obtaining the optimal prediction model for wind and solar power that takes into account both physical constraints and scheduling objectives.

[0104] To verify the applicability of the proposed algorithm in large-scale systems, this invention uses an IEEE 69-node power distribution system for simulation experiments. The parameter configurations of each decision variable in the system are as follows: For distributed power sources, two photovoltaic cells are connected to nodes 26 and 35; two wind turbines are connected to nodes 56 and 69, with active power output ranging from 0 to 0.5 MW; one micro gas turbine is connected to node 17, with both active and reactive power output ranging from 0 to 0.3 MW; and two battery energy storage devices are connected to nodes 18 and 27, with both active and reactive power output ranging from 0 to 0.5 MW. For reactive power compensation devices, two reactive power compensators (SVCs) are connected to nodes 10 and 52, with a reactive power output ranging from 0 to 1.0 MVar; and ten capacitor banks (CBs) are distributed and connected to nodes 10, 17, 27, 35, 39, 54, and 69, with a reactive power output ranging from 0 to 0.5 MVar.

[0105] To verify the effectiveness of this invention, three comparative methods were constructed: Method 1 is a general open-loop scheduling method that combines traditional prediction algorithms with conventional optimization, serving as a basic comparison; Method 2 is an open-loop scheduling method that introduces an improved Crossformer algorithm to achieve high-precision prediction but is still in a one-way transmission mode, used to verify the local effect of improving prediction accuracy; Method 3 is the scheme proposed in this invention, namely, a closed-loop optimization method that introduces a dynamic feedback mechanism for decision-making efficiency and achieves deep integration of prediction and scheduling through gradient backpropagation.

[0106] Figure 2 Analysis of the active network loss of the entire IEEE 69-node network at various time periods shows that the network loss of the method of this invention is the lowest in all time periods, especially during the system heavy load periods around 11 am and 8 pm, where its improvement effect is particularly obvious. Figure 3 To optimize the voltage distribution map of the entire network at 12:00, the analysis shows that although Method 2 can eliminate the voltage overshoot phenomenon that occurred in the traditional Method 1, the voltage fluctuation curve obtained by the restored solution model is smoother, indicating that it has a better control effect on the voltage stability of the entire network.

[0107] Figure 4 The full-time photovoltaic (PV) output curve shows that analyzing the 24-hour PV output at node 26 reveals that Method 2's scheduling strategy is relatively conservative, resulting in the lowest PV output among the compared methods. In contrast, the PV output obtained by the recovery solution model falls between the basic predicted output and Method 1. This result demonstrates that the recovery solution model effectively overcomes the conservatism of Method 2 while ensuring safe system operation, significantly improving the system's utilization rate of distributed PV power resources.

[0108] The proposed solution recovery method achieves multi-dimensional performance improvements. Although Method 2 outperforms Method 1 in terms of voltage deviation, active power loss, and solution time, it sacrifices distributed generation utilization, which is only 65.23%. Method 3 reduces the average voltage deviation to 0.412 and active power loss to 2847.4 kW·h, both of which are better than Method 2. It also increases distributed generation utilization to a maximum of 85.19%. While its solution time of 25.63 seconds is slightly longer than Method 2, it is still much shorter than the 49.040 seconds of Method 1, indicating that this method achieves the optimal overall scheduling effect without significantly increasing the computational burden.

Claims

1. A method for wind and solar power prediction and distribution network dispatching based on dynamic feedback of decision-making efficiency, characterized in that, Includes the following steps: Step 1: Construct a spatiotemporal correlation prediction model. Use an improved Crossformer neural network to model the wind and solar power of each node in the distribution network to obtain the improved Crossformer model. Capture the cross-time and cross-dimensional dependencies between multidimensional meteorological features and power sequences to generate power prediction sequences. Step 2: Construct a multi-objective optimization scheduling model for the active distribution network under time-varying constraints. Input the power prediction sequence into the multi-objective optimization scheduling model. Under the premise of satisfying the time-varying physical feasible region constraint of the power grid, use the improved second-order cone relaxation method to efficiently solve the model and generate the day-ahead scheduling strategy. Step 3: Construct an agent dynamic performance deviation index. Based on the agent dynamic decision performance value theory, construct a performance deviation evaluation function that adapts to the time-varying feasible region to quantify the scheduling performance gap between the scheduling strategy generated in Step 2 and the optimal strategy based on real power. Step 4: Perform dynamic closed-loop feedback training, construct a hybrid loss function that takes into account both prediction accuracy loss and decision performance loss, and use the gradient backpropagation algorithm of this function to update the closed-loop parameters of the improved Crossformer neural network.

2. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: The infrastructure computation for the improved Crossformer model described in step 1 includes: 1) Utilize a dimension-segmented embedding layer to segment each dimension into segments of length 1. The data points are divided into lengths of The segments are divided, and each segment is embedded into a vector through linear projection and positional embedding: (1)； In the formula, It is a two-dimensional vector with dimension . The Time series; For learnable projection matrix; yes The length of the dimension is The Segment vector; for Weizhongdi Learnable embedding positions of segment vectors; 2) Utilize a two-stage attention layer TSA to capture cross-time and cross-dimensional dependencies, with its cross-dimensional stage output... The calculation is as follows: (2)； In the formula, This is the output of the multilayer perceptron; , These are the input and output of the TSA, respectively; Represents the TSA function; 3) Prediction is performed using a hierarchical encoder-decoder, and the final prediction result is obtained. The sum of the prediction results for each layer: (3)； In the formula, ; It is the first The prediction results for the layer; The current moment; To predict the length of the time domain.

3. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: The improved Crossformer model described in step 1 introduces a total variation regularized robust principal component analysis (TRPCA) layer into the Crossformer model and constructs an improved two-stage attention layer, ITSA, to process the output across time stages. Decomposed into a low-rank clean matrix and sparse noise matrix The optimization problem of the decomposition process is expressed as: (4)； In the formula, f is the objective function of the optimization problem of the decomposition process; Indicates and For variable pairs The objective of minimizing the rank is to solve the problem. Represents the 0 norm; yes The weights.

4. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: The improved Crossformer model described in step 1 introduces a closed-loop feedback mechanism, utilizing cross-dimensional stage outputs. Output across time stages Perform adaptive calibration, and correct the features. The calculation formula is: (5)； In the formula, For activation functions; This is the feedback intensity coefficient; Represents element-wise multiplication; A learnable mapping matrix; corrected features It is substituted into the TRPCA optimization model to replace the original input.

5. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: Step 2 involves constructing an active distribution network multi-objective optimization scheduling model; To ensure the safe, reliable and economical operation of the active distribution network, a distribution network optimization scheduling model considering multi-objective collaboration was constructed. The specific objective function includes the following three parts: (1) minimizing node voltage deviation; (2) minimizing system network loss; and (3) maximizing the output of distributed power sources. The multi-objective optimization scheduling problem was transformed into a single-objective optimization scheduling problem using the analytic hierarchy process, providing theoretical support for the generation of subsequent scheduling schemes. The optimization objective problem is described as follows: (6)； In the formula, It is a decision vector; Let p be the target value; the inequalities Ap + Bu ≤ g(t) and Fv ≤ k are collectively referred to as the feasible region S; p is the decision vector, representing the core independent variable to be optimized; A is the first coefficient matrix corresponding to the decision vector and p; u is the input vector or external influence parameter vector; B is the second coefficient matrix corresponding to the input vector and u; g(t) is the constraint threshold vector (or time-varying boundary vector) that changes with time t; F is the constraint coefficient matrix of the relevant variable v; v is the system state variable vector or auxiliary decision vector; k is the constant constraint vector, representing the fixed boundary or limit threshold.

6. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: In step 2, the multi-objective optimization scheduling model for the active distribution network is solved efficiently. To address the problem of safe and economical optimal scheduling of active distribution networks, a mathematical model based on second-order cone programming is established. Furthermore, based on the solution of the second-order cone programming, an iterative recovery algorithm based on a relaxation error penalty function is introduced. This method uses the relaxed solution of the second-order cone programming as the initial value, constructs an error term representing the deviation between the relaxation variables and the physical true value, and introduces it as a penalty function into the objective function. This ensures both the economic efficiency of scheduling and the physical feasibility of the solution, as well as the strict safety of power grid operation.

7. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: In step 3, the performance deviation evaluation function for adapting to the time-varying feasible region is derived using an unambiguous decision performance value oriented towards gradient optimization. The unambiguous decision performance value in the above form is non-convex and discontinuous, which is not conducive to the application of machine learning models because its backpropagation is achieved through gradient calculation. Through dual theory and approximate transformation, a dynamic form of proxy decision performance value that is easy to handle is developed. The unambiguous decision effectiveness value is derived in the following form: （9）； In the formula, Representatives contribute to the forecasting of scenic spots The optimal decision objective is as follows; Representatives contribute to the actual scenery The optimal decision objective is as follows: This indicates the decision with the weakest resistance to disturbances; The target corresponding to the decision with the weakest resistance to disturbances , This represents the surrogate target value after dual transformation. This represents the predicted power output of wind and solar power at time t. The optimal decision objective after dual transformation; Represents the actual wind and light output at time t. The optimal decision objective after dual transformation; This is expressed as finding the minimum upper bound for reducing the relaxation duality gap; It is an introduced proportional parameter.

8. The wind and solar power prediction and distribution network dispatching method based on dynamic feedback of decision-making efficiency according to claim 1, characterized in that: In step 4, The aforementioned dynamic closed-loop feedback training specifically includes: Based on the constructed hybrid loss function The gradient descent algorithm is used to perform end-to-end closed-loop training on the improved Crossformer neural network in step 1. The specific process is as follows: Calculate the hybrid loss function. To improve the gradient of the internal parameters of the Crossformer model, the gradient information is transmitted from the scheduling decision layer back to the prediction model layer using the backpropagation algorithm. The learnable projection matrix, attention weights and fully connected layer parameters in the Crossformer are iteratively updated until the hybrid loss function converges or reaches the preset number of training rounds, thereby obtaining the optimal prediction model for wind and solar power that takes into account both physical constraints and scheduling objectives. The construction of the hybrid loss function specifically includes: Construct a hybrid loss function that includes prediction accuracy loss and decision performance loss. To minimize scheduling decision efficiency while ensuring the accuracy of wind and solar power prediction; the hybrid loss function The calculation formula is as follows: （10）； In the formula, i These are hyperparameter weight coefficients used to balance the contributions of loss terms with different dimensions to the gradient, ensuring that the model is not biased towards a single metric with a large numerical magnitude during training. To predict accuracy loss, the mean squared error is used for calculation, as shown in the following formula: (11)； In the formula, N is the number of training samples; This represents the actual wind and solar power output at the distribution network nodes. To improve the predicted values ​​generated by the Crossformer model.

Images