Multi-objective optimal dispatching method for distribution network with distributed photovoltaic and load fluctuation characteristics

Abstract

The application discloses a kind of distributed photovoltaic and load fluctuation characteristics multi-objective power distribution network optimization scheduling method, it is related to multi-objective power distribution network optimization scheduling field, including: dynamic adjustment equipment is as edge node, its associated physical parameter is as control unit;According to the target demand of multi-objective power distribution network optimization scheduling, construct multi-objective power distribution network optimization scheduling total loss function;Multi-objective power distribution network optimization scheduling model is constructed, and the instruction scheduling value of each control unit is acquired;Based on automatic differentiation traditional power flow calculation, the correction of multi-objective power distribution network optimization scheduling model and output result is carried out;According to the correction result of multi-objective power distribution network optimization scheduling model and output result, real-time optimization scheduling is carried out to multi-objective power distribution network.The application has the advantages that: it can accurately respond to multi-objective power distribution network optimization scheduling demand, effectively coordinates distributed photovoltaic and load fluctuation characteristics by deeply fusing physical constraint conditions, realizes multi-objective collaborative optimization scheduling.

Description

Technical Field

[0001] This invention relates to the field of multi-objective distribution network optimization scheduling, specifically to a multi-objective distribution network optimization scheduling method based on distributed photovoltaic and load fluctuation characteristics. Background Technology

[0002] The power distribution network is undergoing an evolution from a traditional passive network to an active distribution network with a high proportion of distributed energy access. Distributed photovoltaic (PV) has become the main mode of access due to its clean and flexible characteristics. However, the intermittent, random, and anti-peak-shaving characteristics of its output, coupled with the increasing load fluctuations, have brought unprecedented challenges to the safe, stable, and efficient operation of the power distribution network. On the one hand, the bidirectional flow of power may lead to safety problems such as line overload and node voltage exceeding limits. On the other hand, frequent power fluctuations pose a threat to power quality, equipment lifespan, and the grid's regulation capabilities.

[0003] Against this backdrop, optimizing the dispatching of distribution networks has become an inevitable choice to ensure their safe, economical, and high-quality operation. Traditional dispatching models rely on human experience or simple rule control, which is difficult to cope with complex optimization scenarios involving multiple dimensions, changes, and objectives. Multi-objective optimization dispatching aims to coordinate multiple objectives such as economy (e.g., minimizing network losses), safety (e.g., stabilizing voltage), reliability (e.g., improving fluctuation resistance), and regulation smoothness. Its core is to find the optimal decision space in a world of massive uncertainty. In recent years, artificial intelligence technology, especially deep learning, has provided a new paradigm for solving this complex problem due to its powerful nonlinear fitting and fast reasoning capabilities. However, how to deeply integrate the strict physical operating constraints of the power grid with data-driven AI models to ensure that the AI ​​output results are not only optimal but also physically feasible and have extremely high reliability remains a technical challenge that urgently needs to be overcome in the current field, and it is also the key to realizing the intelligent dispatching upgrade of distribution networks.

[0004] Existing technologies that utilize neural networks for distribution network optimization and dispatching are purely driven by historical data distribution. They attempt to directly fit the mapping relationship between input and output through the complex black-box structure of neural networks, while ignoring the physical laws (such as power flow calculation) and safety constraints (such as voltage upper and lower limits and line power limits) that distribution network operation must strictly follow. This disconnect from physical mechanisms leads the model to pursue only the superficial correlation in the learning data, which can easily output dispatching instructions that have a low mathematical loss function value but are completely infeasible or even dangerous in actual physics. For example, in pursuit of economy, control strategies may be issued that cause system voltage to exceed limits, line overload, and even trigger cascading failures. Therefore, due to its inherent uninterpretability and unreliability, it cannot be applied to the actual dispatching of power systems with extreme requirements for safety and stability. Summary of the Invention

[0005] To address the aforementioned technical problems, this paper presents a multi-objective distribution network optimization scheduling method based on distributed photovoltaic power and load fluctuation characteristics. This technical solution solves the problem mentioned in the background technology that relies on statistical patterns in historical data while ignoring the physical laws and safety constraints that distribution network operation must strictly follow. This leads to the model only pursuing the superficial correlations in the learning data, which easily results in the output of scheduling instructions that have a very low mathematical loss function value but are completely infeasible or even dangerous in actual physics.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A multi-objective distribution network optimization scheduling method based on distributed photovoltaic power generation and load fluctuation characteristics includes:

[0008] Based on the physical structure of power distribution network optimization scheduling, the dynamically adjusting equipment is used as an edge node, and its associated physical parameters are used as a control unit.

[0009] Based on the target requirements of multi-objective distribution network optimization scheduling, and combined with the physical constraints of distributed photovoltaic and load fluctuation characteristics, a total loss function for multi-objective distribution network optimization scheduling is constructed.

[0010] Based on a deep fully connected network, a multi-objective distribution network optimization scheduling model is constructed to obtain the instruction scheduling values ​​of each control unit;

[0011] Based on the command scheduling values ​​of each control unit, and using automatic differential traditional power flow calculation, the multi-objective distribution network optimization scheduling model and its output results are corrected.

[0012] Based on the multi-objective distribution network optimization scheduling model and the correction results of the output results, the multi-objective distribution network is optimized and scheduled in real time.

[0013] Preferably, the step of constructing the total loss function for multi-objective distribution network optimization scheduling based on the target requirements of multi-objective distribution network optimization scheduling and the physical constraints of distributed photovoltaic and load fluctuation characteristics specifically includes:

[0014] In terms of economic efficiency, the goal is to minimize network loss costs. An economic loss function is constructed based on traditional power flow calculation with automatic differentiation.

[0015] In terms of safety, the goal is to minimize the risk of voltage over-limit, and the sum of squares of voltage deviations is quantified to construct a safety loss function;

[0016] In terms of fluctuation mitigation, a fluctuation mitigation loss function is constructed with the goal of maximizing reactive power reserve capacity.

[0017] In terms of smoothness of regulation, a smoothness of regulation loss function is constructed with the goal of minimizing the amount of regulation change in adjacent time periods;

[0018] Based on the physical constraints of photovoltaic active power fluctuations and load active power fluctuations, an active power balance penalty term is constructed.

[0019] Based on the physical constraints of voltage fluctuations in the distribution network, a voltage regulation penalty term is constructed;

[0020] Based on the penalty term and each loss function, a weighted summation method is used to construct the total loss function for multi-objective distribution network optimization scheduling.

[0021] The expression for the total loss function of the multi-objective distribution network optimal scheduling is as follows: In the formula, To optimize the total loss value of multi-objective distribution network scheduling, The number of target demands for optimizing the scheduling of multi-objective distribution networks. For the first The weights of the target demand loss values For the first Each target demand loss value, The number of penalty items. For the first The weight of each penalty item For the first The values ​​of the penalty items include the target requirements: economy, security, volatility resistance, and regulation smoothness.

[0022] Weighting coefficient These are preset hyperparameters, whose values ​​are determined by the scheduler based on actual operational needs, and are used to flexibly adjust the priority of different objectives in the overall decision-making process.

[0023] Preferably, the step of constructing a multi-objective distribution network optimization scheduling model based on a deep fully connected network and obtaining the instruction scheduling values ​​of each control unit specifically includes:

[0024] Based on the source-load fluctuation time-series prediction data, real-time power grid state variables, current state of edge nodes, target demand loss value weights, and penalty term weights, the input vector of the deep fully connected network is set.

[0025] The input vector data of the deep fully connected network is normalized and mapped to the [0,1] interval to eliminate the influence of data units;

[0026] The control unit instruction values ​​of each edge node at future scheduling moments are used as the prediction labels for the output layer of the deep fully connected network.

[0027] Based on historical data, a training sample set is constructed according to the input vector of the deep fully connected network and the predicted labels of the output layer of the deep fully connected network, and then divided into a training set and a validation set.

[0028] Based on mean squared error, the goal of training the deep fully connected network is to minimize the error between the network's predicted label and the target label, and the deep fully connected network is pre-trained.

[0029] Among them, the target labels are composed of excellent dispatching instruction cases that are recorded in historical operation, actually executed by the power grid dispatching system, and verified to be safe and effective;

[0030] Based on the pre-trained deep fully connected network, the total loss function of multi-objective distribution network optimization scheduling is used as the training objective of the deep fully connected network to fine-tune the deep fully connected network.

[0031] In the forward propagation process of the fine-tuning stage, an automatic differential traditional power flow module needs to be embedded. This module takes the output command of the neural network and the real-time state of the power grid as inputs to solve the sub-loss values ​​in the total loss function of multi-objective distribution network optimization scheduling.

[0032] Based on the training sample set, a deep fully connected network is trained. The trained deep fully connected network is defined as a multi-objective distribution network optimization scheduling model, which is used to obtain the instruction scheduling values ​​of each control unit.

[0033] Preferably, the step of correcting the multi-objective distribution network optimization scheduling model and its output results based on the command scheduling values ​​of each control unit and on the basis of automatic differential traditional power flow calculation specifically includes:

[0034] The command scheduling values ​​of each control unit output by the multi-objective distribution network optimization scheduling model are substituted into the automatic differential traditional power flow module to perform power flow calculation and obtain the predicted state of the power grid.

[0035] Constructing the constraint violation cost function It determines whether the predicted state of the power grid exceeds the limit. If so, it fine-tunes the command through the gradient descent method and outputs the corrected command scheduling value of each control unit. If not, it directly outputs the command scheduling value of each control unit.

[0036] Based on the training results of the multi-objective distribution network optimization scheduling model, its training loss value is statistically analyzed, and its predictive performance index is calculated through the validation set.

[0037] Based on historical data, the statistical distribution of training loss value and prediction performance index is calculated, and its 95th percentile is set as the boundary threshold.

[0038] Determine whether the training loss value and prediction performance metric exceed the boundary threshold. If any parameter exceeds the boundary threshold, the current model is deemed unreliable and a correction is required. Otherwise, the current model is deemed reliable.

[0039] When the current model is deemed unreliable, based on historical data, the cosine similarity between the input vector of the deep fully connected network and the historical data is calculated, and the top vectors with the highest similarity are selected from the historical data. One sample;

[0040] Based on constraint violation cost function Value, from the highest similarity From a sample, select one The instruction with the smallest value is used as the instruction scheduling value for each control unit after verification and correction;

[0041] The expression for the constraint violation cost function is:

[0042] In the formula, The first one obtained through power flow calculation Each node voltage value , These are the upper and lower limits of the allowable voltage, respectively. For the first The magnitude of the apparent power currently flowing through the line is obtained through power flow calculation. For the first The maximum apparent power capacity, i.e., the current carrying capacity, of this line is allowed to operate safely for a long period of time. , These are penalty coefficients for voltage and apparent power, respectively. These are empirical values ​​and should be adjusted and set according to the different priority requirements of voltage safety and line safety in the actual power grid.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] This invention provides a multi-objective distribution network optimization scheduling method based on distributed photovoltaic and load fluctuation characteristics. By embedding an automatic differential traditional power flow module as a differentiable component into the fine-tuning training phase of a neural network, this module calculates various sub-losses (such as network losses and voltage deviations) in the total loss function via forward propagation. This allows the physical laws of the power grid (power flow equations) and safety constraints to directly guide the model's weight updates through gradient calculation. This design enables the model to not only learn the statistical patterns of historical data but also embed the physical mechanisms of power grid operation, thereby outputting scheduling instructions that satisfy both multi-objective optimization requirements and strictly comply with physical safety constraints. This effectively avoids the black-box characteristics of purely data-driven traditional neural networks, ensuring the feasibility and security of optimization instructions. Furthermore, in online applications, the same automatic differential traditional power flow module can be used... This approach performs forward-looking power flow calculations and uses gradient descent to perform online safety verification and fine-tuning of the multi-objective distribution network optimization scheduling model output, ensuring safety immediately before command issuance. Simultaneously, this solution designs a verification and correction method based on cosine similarity retrieval of historical data. Based on the model's output, it intelligently judges the reliability of its performance and automatically and quickly selects the safest and most reliable command from historical best cases for execution. The physical compliance of the model is ensured through embedded physical constraints during the training phase, real-time verification during the online phase achieves safety confirmation before execution, and historical case retrieval for abnormal scenarios provides emergency backup solutions. The synergistic effect of these three methods effectively improves the accuracy and stability of multi-objective distribution network optimization scheduling using deep fully connected networks, ensuring that multi-objective distribution network optimization scheduling does not exceed limits, thereby preventing distribution network operation failures. Attached Figure Description

[0045] Figure 1 This is a flowchart of a multi-objective distribution network optimization scheduling method based on distributed photovoltaic power generation and load fluctuation characteristics, according to the present invention.

[0046] Figure 2 The flowchart for constructing the total loss function of multi-objective distribution network optimization scheduling is shown below.

[0047] Figure 3 The flowchart for obtaining the instruction scheduling values ​​of each control unit in constructing a multi-objective distribution network optimization scheduling model according to the present invention is shown below.

[0048] Figure 4 This is a structural diagram of the electronic device proposed in this invention;

[0049] Figure 5 This is a schematic diagram of the structure of the computer-readable storage medium proposed in this invention. Detailed Implementation

[0050] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0051] Reference Figure 1 As shown, a multi-objective distribution network optimization scheduling method based on distributed photovoltaic power generation and load fluctuation characteristics includes:

[0052] Based on the physical structure of power distribution network optimization scheduling, the dynamically adjusting equipment is used as an edge node, and its associated physical parameters are used as a control unit.

[0053] Based on the target requirements of multi-objective distribution network optimization scheduling, and combined with the physical constraints of distributed photovoltaic and load fluctuation characteristics, a total loss function for multi-objective distribution network optimization scheduling is constructed.

[0054] Based on a deep fully connected network, a multi-objective distribution network optimization scheduling model is constructed to obtain the instruction scheduling values ​​of each control unit;

[0055] Based on the command scheduling values ​​of each control unit, and using automatic differential traditional power flow calculation, the multi-objective distribution network optimization scheduling model and its output results are corrected.

[0056] Based on the multi-objective distribution network optimization scheduling model and the correction results of the output results, the multi-objective distribution network is optimized and scheduled in real time.

[0057] It can be explained that traditional power flow calculation is one of the most fundamental and core calculations in power systems. Given the input conditions of the power grid, such as generator output, load consumption, network topology, and control commands, it calculates the operating state of the power grid by solving a set of nonlinear equations describing the steady-state characteristics of the power system. This results in the voltage magnitude and phase angle of each node, and thus the power of all branches. It serves as a simulator of the physical world and is a core tool for verifying whether the optimal scheduling scheme for the distribution network conforms to physical constraints and is feasible. It runs through the entire process of model training, online verification, and anomaly handling. The automatic differentiation traditional power flow module refers to converting the algorithmic process of distribution network power flow calculation (such as the forward-backward substitution method or the Newton-Raphson method) into a differentiable computational graph model using automatic differentiation (AD) technology. This is done by using a programming language that supports automatic differentiation (such as...) to perform the core iterative process of power flow calculation. This approach reimplements the algorithm in Python or encapsulates existing power flow calculation code using the automatic differentiation engine of a deep learning framework. This makes the entire power flow calculation process behave like a neural network layer, capable of forward computation of state variables and backward propagation of gradient information. The input to this module is the command scheduling value of each control unit, and the output includes not only system state variables but also the gradient information of these state variables with respect to the input command. This provides the necessary mathematical foundation for subsequent gradient descent-based safety correction algorithms. Therefore, this solution is based on the traditional power flow module of automatic differentiation and provides emergency backup solutions for ensuring the physical compliance of the model by embedding physical constraints during the training phase, implementing pre-execution safety confirmation through real-time verification during the online phase, and retrieving historical cases of abnormal scenarios. This ensures that the multi-objective distribution network optimization scheduling does not deviate from physical constraints and that the command scheduling value of each control unit is within the range allowed by physical constraints, thereby avoiding the multi-objective distribution network optimization scheduling exceeding limits and causing distribution network operation failures.

[0058] The physical structure based on power distribution network optimization scheduling, which uses dynamically adjusting equipment as edge nodes and their associated physical parameters as control units, specifically includes:

[0059] Identify and obtain all instances of dynamically adjusting devices participating in scheduling in the existing distribution network, define each independent device as an edge node, and assign a unique digital identifier to each edge node;

[0060] Based on the type and physical characteristics of each edge node, obtain its remotely controlled physical parameters and define these parameters as the control unit of that edge node;

[0061] Based on the device design specifications of each edge node, the basic physical constraints of each control unit are obtained, including the upper and lower limits of the control unit values, accuracy, and frequency of action.

[0062] The edge node refers to a physical device in the distribution network that has remote or automatic control capabilities, and is the final execution unit and data acquisition unit for optimized scheduling instructions.

[0063] It can be explained that, in multi-objective distribution network optimization scheduling, clearly knowing which equipment in the distribution network and which physical parameters of that equipment can be scheduled is an essential step. Secondly, clarifying the basic physical constraints of these physical parameters is necessary to prevent adjustments from exceeding the acceptable range. Therefore, this scheme uses dynamically adjustable equipment as edge nodes and their associated physical parameters as control units, clearly indicating that adjustments are made to control units during multi-objective distribution network optimization scheduling. For ease of understanding, specific examples are given, but are not limited to the following:

[0064] Define dynamically adjustable devices as edge nodes, including: distributed photovoltaic inverters, energy storage converters, static var generators, on-load tap-changing transformers, and parallel capacitor banks;

[0065] The control unit of the distributed photovoltaic inverter includes active power setpoint and reactive power setpoint. Its basic physical constraints are the upper and lower limits of active power setpoint (0kW - rated power), the upper and lower limits of reactive power setpoint, and the upper limit of active power change rate. The control unit of the energy storage converter includes active power setpoint (charging / discharging) and state of charge (SOC). Its basic physical constraints are the upper and lower limits of active power setpoint (maximum charging power - maximum discharging power), the upper and lower limits of SOC operation, and the charging / discharging switching time interval. The control unit of the static var generator is the reactive power setpoint. Its basic physical constraints are the upper and lower limits of reactive power setpoint. The control unit of the on-load tap changer is the tap position. Its basic physical constraints are the upper and lower limits of the tap position, the maximum number of daily operations, and the time interval between adjacent operations. The control unit of the parallel capacitor bank is the number of switching groups. Its basic physical constraints are the upper and lower limits of the number of switching groups, the maximum number of daily operations, and the time interval between adjacent operations.

[0066] Based on the above definitions and examples, clear and operable control objects and their safety boundaries are provided for the optimal scheduling of multi-objective distribution networks, ensuring the feasibility of optimization commands and equipment safety.

[0067] Reference Figure 2 As shown, the construction of the total loss function for multi-objective distribution network optimization scheduling specifically includes:

[0068] In terms of economic efficiency, the goal is to minimize network loss costs. An economic loss function is constructed based on traditional power flow calculation with automatic differentiation.

[0069] In terms of safety, the goal is to minimize the risk of voltage over-limit, and the sum of squares of voltage deviations is quantified to construct a safety loss function;

[0070] In terms of fluctuation mitigation, a fluctuation mitigation loss function is constructed with the goal of maximizing reactive power reserve capacity.

[0071] In terms of smoothness of regulation, a smoothness of regulation loss function is constructed with the goal of minimizing the amount of regulation change in adjacent time periods;

[0072] Based on the physical constraints of photovoltaic active power fluctuations and load active power fluctuations, an active power balance penalty term is constructed.

[0073] Based on the physical constraints of voltage fluctuations in the distribution network, a voltage regulation penalty term is constructed;

[0074] Based on the penalty term and each loss function, a weighted summation method is used to construct the total loss function for multi-objective distribution network optimization scheduling.

[0075] The expression for the total loss function of the multi-objective distribution network optimal scheduling is as follows: In the formula, To optimize the total loss value of multi-objective distribution network scheduling, The number of target demands for optimizing the scheduling of multi-objective distribution networks. For the first The weights of the target demand loss values For the first Each target demand loss value, The number of penalty items. For the first The weight of each penalty item For the first The values ​​of the penalty items include the target requirements: economy, security, volatility resistance, and regulation smoothness.

[0076] Weighting coefficient , These are preset hyperparameters, whose values ​​are determined by the scheduler based on actual operational needs, and are used to flexibly adjust the priority of different objectives in the overall decision-making process.

[0077] This can be explained by the fact that when performing multi-objective distribution network optimization scheduling, it is not enough to focus solely on either economic or safety requirements; multiple aspects need to be considered. Based on the actual multi-objective distribution network optimization scheduling needs, the weights of each requirement are adjusted to clarify the priority of the objectives in the overall decision-making process. This approach is more suitable for the flexible scheduling characteristics of multi-objective distribution network optimization. This scheme constructs a mechanism that transforms the complex physical rules and operational requirements (economy, safety, stability, smoothness) of multi-objective distribution network optimization scheduling into differentiable mathematical expressions. This mathematical expression is then embedded as prior physical knowledge into the total loss function of the neural network training process. This guides the scheduling instructions generated by the neural network to not only pursue the optimality of a single objective but also automatically and synchronously satisfy the multiple constraints and comprehensive optimality of power grid operation.

[0078] The expression for the economic loss function is: In the formula, This represents the economic loss value. This is the electricity price factor, which converts the active power lost through network losses into economic costs. The active power loss value of a single branch obtained through power flow calculation;

[0079] For a specific transmission line, power flow calculations can yield the active power P, reactive power Q, and line resistance R parameters transmitted along the line, according to the formula... This allows us to calculate the active power loss of the line. ,in, Line current;

[0080] The security loss function is expressed as follows: In the formula, This represents the security loss value. The first one obtained through power flow calculation Each node voltage value , These are the upper and lower limits of the allowable voltage, respectively;

[0081] The expression for the volatility mitigation loss function is: In the formula, To mitigate losses due to fluctuations, For the first The maximum capacity of a single reactive power source device For the first The current actual capacity of each reactive power source device It is the absolute value symbol.

[0082] Indicates the first The total remaining reactive capacity of the system is obtained by summing the remaining reactive capacity of each reactive power source device. Since the loss function is to be minimized, a negative sign is added in front of it, so that maximizing the system's reactive power reserve capacity is equivalent to minimizing it.

[0083] The expression for the smoothness loss function is: In the formula, To adjust the smoothness loss value, for A vector composed of all control unit command values ​​at any given time. for A vector composed of all control unit command values ​​at any given time. It is an L2 norm;

[0084] This value comprehensively reflects the total amplitude of action of all control devices in adjacent time periods, and when expanded, it is... , of which each Indicates the first Each control unit Time relative to The square of the motion amplitude at each moment, summed up by adding this value from all control units, yields the quantized value of the total motion amplitude of the entire system at adjacent moments. The larger the motion amplitude, the greater the quantization value. The larger the value, the smoother the motion, and the smaller the value. Number of control units;

[0085] The expression for the active power balance penalty term is: In the formula, The value of the active power balance penalty term. This is the penalty coefficient for meritorious service. This refers to the system's net active power fluctuation. The active power unit instructions for energy storage devices are output by the neural network. The distributed power source active unit instructions are output by the neural network, where... , This represents the load active power fluctuation, which is the deviation between the actual active power and the predicted active power at the current moment. Photovoltaic active power fluctuation is the deviation between the actual active power and the predicted active power of photovoltaic power at the current moment.

[0086] when or When the active unit instruction output by the neural network violates the prior physical rules of active power balance, a corresponding penalty is imposed. The magnitude of the penalty is proportional to the degree of violation. or At that time, no punishment shall be imposed;

[0087] The expression for the voltage regulation penalty term is: In the formula, To adjust the value of the penalty term, This is the reactive power penalty coefficient. This is the node voltage deviation, which is the difference between the actual voltage and the reference voltage. The reactive power equipment command output by the neural network is the static var generator reactive power command.

[0088] when When the reactive power equipment command output by the neural network violates the physical prior rules of voltage regulation, a penalty is imposed. The magnitude of the penalty is proportional to the degree of violation. At that time, no punishment shall be imposed;

[0089] because The function is differentiable, which makes it possible for the function to be differentiable in the following ways. The subgradient at the point can satisfy the backpropagation requirements, ensuring that the entire penalty term function is differentiable, thus enabling it to be seamlessly integrated into the training process of the neural network. During the backpropagation process, the network parameters are updated by calculating the gradient of the penalty term with respect to the neural network parameters, so that the output of the neural network gradually conforms to the physical prior rules.

[0090] Among them, the voltage regulation penalty term, as a physical prior constraint embedded in the training process of the neural network, forces its output command to conform to the voltage-reactive power sensitivity law, thereby ensuring the rationality of behavior in the decision-making process. The safety loss function, as a direct quantification of the objective, drives the safety of the system's operating state by penalizing the final result of voltage exceeding the limit. The two work together from the two dimensions of decision-making process and operating result to jointly ensure the physical feasibility and operational reliability of the optimized scheduling scheme.

[0091] Reference Figure 3 As shown, the specific steps for constructing a multi-objective distribution network optimization scheduling model to obtain the command scheduling values ​​of each control unit include:

[0092] Based on the source-load fluctuation time-series prediction data, real-time power grid state variables, current state of edge nodes, target demand loss value weights, and penalty term weights, the input vector of the deep fully connected network is set.

[0093] The input vector data of the deep fully connected network is normalized and mapped to the [0,1] interval to eliminate the influence of data units;

[0094] The control unit instruction values ​​of each edge node at future scheduling moments are used as the prediction labels for the output layer of the deep fully connected network.

[0095] Based on historical data, a training sample set is constructed according to the input vector of the deep fully connected network and the predicted labels of the output layer of the deep fully connected network, and then divided into a training set and a validation set.

[0096] Based on mean squared error, the goal of training the deep fully connected network is to minimize the error between the network's predicted label and the target label, and the deep fully connected network is pre-trained.

[0097] Among them, the target labels are composed of excellent dispatching instruction cases that are recorded in historical operation, actually executed by the power grid dispatching system, and verified to be safe and effective;

[0098] Based on the pre-trained deep fully connected network, the total loss function of multi-objective distribution network optimization scheduling is used as the training objective of the deep fully connected network to fine-tune the deep fully connected network.

[0099] In the forward propagation process of the fine-tuning stage, an automatic differential traditional power flow module needs to be embedded. This module takes the output command of the neural network and the real-time state of the power grid as inputs to solve the sub-loss values ​​in the total loss function of multi-objective distribution network optimization scheduling.

[0100] Based on the training sample set, a deep fully connected network is trained. The trained deep fully connected network is defined as a multi-objective distribution network optimization scheduling model, which is used to obtain the instruction scheduling values ​​of each control unit.

[0101] This can be explained by the fact that Deep Fully-Connected Networks (MLPs), also known as Multilayer Perceptrons (MLPs), are a fundamental and widely used artificial neural network structure. In distribution network optimization and scheduling, the relationships between source load fluctuations, grid status, edge node status, and control unit commands are characterized by strong nonlinearity and multivariate coupling. Deep Fully-Connected Networks, through nonlinear transformations of multiple hidden layers (such as the ReLU activation function), can autonomously learn these complex mapping relationships without the need for manual derivation of physical formulas, making them particularly suitable for handling distribution network optimization and scheduling. The implicit correlation between power grid fluctuations and regulation is explored. Source-load fluctuation time-series forecast data refers to forecast data for the next scheduling cycle, specifically including: predicted active and reactive power output of all distributed photovoltaic nodes, and predicted active and reactive load of all load nodes. This data is acquired based on existing photovoltaic power forecasting and load forecasting systems, providing the network with information on the sources and characteristics of future disturbances and fluctuations. It serves as a crucial basis for the network to make forward-looking decisions. Real-time power grid status quantities refer to the monitored quantities of the current power grid status, composed of continuous and discrete data, specifically including: actual photovoltaic output, actual load value, and so on. The initial state of the power grid is provided by real-time measurement data from the SCADA system and Advanced Measurement Infrastructure (AMI), including the voltage amplitude of nodes, voltage phase angles of all nodes, power flow of each branch, and the current state of charge of the energy storage system. All optimization decisions must be based on the current state. The current state of edge nodes is the control unit state of all edge nodes at the current moment, consisting of continuous and discrete data. Specifically, it includes: the active power setpoint and reactive power setpoint of the distributed photovoltaic inverter, the active power setpoint (charging / discharging) and state of charge (SOC) of the energy storage converter, the reactive power setpoint of the static var generator, and the on-load tap changer. The tap position of the transformer and the number of parallel capacitor banks are based on the real-time status reported by the SCADA system or equipment controller to provide the current base point of the control equipment. The optimization instructions are usually the changes relative to this base point. At the same time, they serve as the benchmark for calculating the prior penalty term and are used to determine whether the instructions output by the neural network violate physical prior rules such as active power balance and voltage regulation. Among them, for discrete variables in the input and output, the deep fully connected network can directly quantify them. During the training phase, the output value remains continuous to maintain differentiability. During the deployment and application phase, the output value is rounded to generate the final executable discrete instructions.

[0102] The deep fully connected network adopts a multi-layer fully connected structure. The number of neurons in the input layer is determined by the dimension of the input vector, the number of neurons in the hidden layer is 256, 128, and 64 respectively, and the number of neurons in the output layer is determined by the number of control units. The activation function is ReLU to avoid gradient vanishing. The optimizer is Adam, and the learning rate is initially set to 0.001. Preliminary pre-training is used to quickly converge to the usable solution space, avoiding gradient oscillations when directly training the total loss function of multi-objective distribution network optimization scheduling. This significantly shortens the overall training time. In other words, preliminary pre-training lays the foundation for the model to be accurate in accordance with the basic scheduling logic, while fine-tuning strengthens the optimization adaptation to multi-objective requirements on this basis. The two form a progressive relationship from basic feasibility to objective optimization.

[0103] The step of correcting the multi-objective distribution network optimization scheduling model and its output results based on the command scheduling values ​​of each control unit and on the basis of automatic differential traditional power flow calculation specifically includes:

[0104] The command scheduling values ​​of each control unit output by the multi-objective distribution network optimization scheduling model are substituted into the automatic differential traditional power flow module to perform power flow calculation and obtain the predicted state of the power grid.

[0105] Constructing the constraint violation cost function It determines whether the predicted state of the power grid exceeds the limit. If so, it fine-tunes the command through the gradient descent method and outputs the corrected command scheduling value of each control unit. If not, it directly outputs the command scheduling value of each control unit.

[0106] Based on the training results of the multi-objective distribution network optimization scheduling model, its training loss value is statistically analyzed, and its predictive performance index is calculated through the validation set.

[0107] Based on historical data, the statistical distribution of training loss value and prediction performance index is calculated, and its 95th percentile is set as the boundary threshold.

[0108] Determine whether the training loss value and prediction performance metric exceed the boundary threshold. If any parameter exceeds the boundary threshold, the current model is deemed unreliable and a correction is required. Otherwise, the current model is deemed reliable.

[0109] When the current model is deemed unreliable, based on historical data, the cosine similarity between the input vector of the deep fully connected network and the historical data is calculated, and the top vectors with the highest similarity are selected from the historical data. One sample;

[0110] Based on constraint violation cost function Value, from the highest similarity From a sample, select one The instruction with the smallest value is used as the instruction scheduling value for each control unit after verification and correction;

[0111] The expression for the constraint violation cost function is:

[0112] In the formula, The first one obtained through power flow calculation Each node voltage value , These are the upper and lower limits of the allowable voltage, respectively. For the first The magnitude of the apparent power currently flowing through the line is obtained through power flow calculation. For the first The maximum apparent power capacity, i.e., the current carrying capacity, of this line is allowed to operate safely for a long period of time. , These are penalty coefficients for voltage and apparent power, respectively. These are empirical values ​​and should be adjusted and set according to the different priority requirements of voltage safety and line safety in the actual power grid.

[0113] This can be explained by the fact that, due to the importance of stable distribution network operation, it is impossible to directly delegate the power of distribution network optimization scheduling entirely to a multi-objective distribution network optimization scheduling model. This is because the model's output is not entirely accurate and may be affected by external factors or model performance degradation, leading to deviations in the output. For example, directly using incorrect control unit command scheduling values ​​could cause the distribution network optimization scheduling to deviate from expectations, or even cause system paralysis. Therefore, it is necessary to add a verification and correction step to the output of the multi-objective distribution network optimization scheduling model, i.e., the control unit command scheduling values, to monitor and optimize the control unit command scheduling values. Therefore, this scheme constructs a constraint violation cost function based on automatic differential traditional power flow calculation. By constraining the feedback of cost function violations, the command scheduling values ​​of the control unit are verified and corrected. Secondly, based on the statistical distribution of model training loss values ​​and prediction performance indicators, the reliability of the model is determined. If the model is unreliable, cosine similarity is used to select a suitable model from historical data. The instruction with the smallest value is used as the instruction scheduling value for each control unit after verification and correction. This allows for timely and effective optimization scheduling measures without sacrificing timeliness when using a multi-objective distribution network optimization scheduling model to optimize the distribution network. This improves the efficiency and timeliness of distribution network optimization scheduling and ensures its stability and reliability. Specifically, fine-tuning the instruction using gradient descent refers to calculating the constraint violation cost function. Command scheduling value gradient , and according to The correction is performed in this way, and the corrected command values ​​are trimmed to the upper and lower limits allowed by the basic physical constraints of each control unit, so that... The value decreases, and this process is repeated until... Output the corrected version , among which, if If the predicted state of the power grid exceeds the limit, the command is fine-tuned using the gradient descent method. The instruction is then output directly, where, The step size parameter for gradient descent controls the magnitude of each correction instruction. A positive decimal close to zero indicates an acceptable constraint violation tolerance.

[0114] Node voltage values ​​in the safety loss function The node voltage values ​​in the constraint violation cost function Both results were obtained through the same automatic differential conventional power flow module, and their physical essence is the same; their voltage values ​​are identical. During the training phase, it serves as an intermediate variable for calculating the total loss, guiding the multi-objective distribution network optimization scheduling model to learn the physical laws of the power grid. During the correction phase, it serves as a safety verification benchmark before issuing instructions, ensuring the absolute reliability of the instruction scheduling values ​​of each control unit. This design guarantees the inherent consistency between the objectives of the multi-objective distribution network optimization scheduling model and the final safety criteria.

[0115] The real-time optimization scheduling of the multi-objective distribution network based on the multi-objective distribution network optimization scheduling model and the correction results of the output results specifically includes:

[0116] Deploy a monitoring device to collect the data required for the input vectors of the deep fully connected network, and acquire the monitoring data of the input vectors of the deep fully connected network in real time;

[0117] The multi-objective distribution network optimization scheduling model is executed according to a preset rolling cycle to obtain the instruction scheduling values ​​of each control unit;

[0118] Power flow calculation is performed based on the automatic differential traditional power flow module, the command scheduling values ​​of each control unit are verified and corrected, and the verified and corrected command scheduling values ​​of each control unit are obtained.

[0119] The command scheduling values ​​of each control unit after verification and correction are issued to optimize and control the control units of each edge node;

[0120] Furthermore, by utilizing IoT technology, the monitoring data of the input vector of the deep fully connected network and the feedback data after optimization and control are packaged together and uploaded to the data processing center.

[0121] Construct multi-objective distribution network optimization scheduling data to classify, store, and record data received by the data processing center, facilitating the updating and optimization of the multi-objective distribution network optimization scheduling model.

[0122] This solution addresses the optimization and scheduling of multi-objective distribution networks. It constructs a comprehensive multi-objective distribution network optimization and scheduling process, encompassing steps such as data acquisition, optimization scheduling instruction acquisition, instruction verification and correction, instruction issuance, data feedback and uploading, and database storage. This ensures that the multi-objective distribution network optimization and scheduling can be executed stably and accurately, effectively coordinating distributed photovoltaic power generation with load fluctuation characteristics, and achieving multi-objective collaborative optimization and scheduling.

[0123] Furthermore, the method according to the embodiments of this application can also be achieved by means of... Figure 4 The architecture of the electronic device shown is used to implement this. For example... Figure 4 As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, a read-only memory (ROM) 503, a random access memory (RAM) 504, a communication port 505 connected to a network, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as the ROM 503 or the hard disk 507, may store a multi-objective distribution network optimization scheduling method based on distributed photovoltaic and load fluctuation characteristics provided in this application. The electronic device 500 may also include a user interface 508. Of course, Figure 4 The architecture shown is merely exemplary and can be omitted as needed when implementing different devices. Figure 4 One or more components in the illustrated electronic device.

[0124] Figure 5 This is a schematic diagram of a computer-readable storage medium structure provided in one embodiment of this application. Figure 5 The diagram illustrates a computer-readable storage medium 600 according to one embodiment of this application. The computer-readable storage medium 600 stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a multi-objective distribution network optimization scheduling method based on distributed photovoltaic power and load fluctuation characteristics, as described above with reference to the accompanying drawings, according to an embodiment of this application. The storage medium 600 includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0125] In summary, the advantages of this invention are: it can accurately address the multi-objective power distribution network optimization and scheduling needs, and through deep integration of physical constraints, it can effectively coordinate distributed photovoltaic power generation with load fluctuation characteristics to achieve multi-objective collaborative optimization scheduling.

[0126] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-objective distribution network optimization scheduling method based on distributed photovoltaic power generation and load fluctuation characteristics, characterized in that, include: Based on the physical structure of power distribution network optimization scheduling, the dynamically adjusting equipment is used as an edge node, and its associated physical parameters are used as a control unit. Based on the target requirements of multi-objective distribution network optimization scheduling, and combined with the physical constraints of distributed photovoltaic and load fluctuation characteristics, a total loss function for multi-objective distribution network optimization scheduling is constructed. Based on a deep fully connected network, a multi-objective distribution network optimization scheduling model is constructed to obtain the instruction scheduling values ​​of each control unit; Based on the command scheduling values ​​of each control unit, and using automatic differential traditional power flow calculation, the multi-objective distribution network optimization scheduling model and its output results are corrected. Based on the multi-objective distribution network optimization scheduling model and the correction results of the output results, the multi-objective distribution network is optimized and scheduled in real time. The physical structure based on power distribution network optimization scheduling, which uses dynamically adjusting equipment as edge nodes and their associated physical parameters as control units, specifically includes: Identify and obtain all instances of dynamically adjusting devices participating in scheduling in the existing distribution network, define each independent device as an edge node, and assign a unique digital identifier to each edge node; Based on the type and physical characteristics of each edge node, obtain its remotely controlled physical parameters and define these parameters as the control unit of that edge node; Based on the device design specifications of each edge node, the basic physical constraints of each control unit are obtained, including the upper and lower limits of the control unit values, accuracy, and frequency of action. The edge node refers to a physical device in the distribution network that has remote or automatic control capabilities, and is the final execution unit and data acquisition unit for optimized scheduling instructions; The step of correcting the multi-objective distribution network optimization scheduling model and its output results based on the command scheduling values ​​of each control unit and on the basis of automatic differential traditional power flow calculation specifically includes: The command scheduling values ​​of each control unit output by the multi-objective distribution network optimization scheduling model are substituted into the automatic differential traditional power flow module to perform power flow calculation and obtain the predicted state of the power grid. Constructing the constraint violation cost function It determines whether the predicted state of the power grid exceeds the limit. If so, it fine-tunes the command through the gradient descent method and outputs the corrected command scheduling value of each control unit. If not, it directly outputs the command scheduling value of each control unit. Based on the training results of the multi-objective distribution network optimization scheduling model, its training loss value is statistically analyzed, and its predictive performance index is calculated through the validation set. Based on historical data, the statistical distribution of training loss value and prediction performance index is calculated, and its 95th percentile is set as the boundary threshold. Determine whether the training loss value and prediction performance metric exceed the boundary threshold. If any parameter exceeds the boundary threshold, the current model is deemed unreliable and a correction is required. Otherwise, the current model is deemed reliable. When the current model is deemed unreliable, based on historical data, the cosine similarity between the input vector of the deep fully connected network and the historical data is calculated, and the top vectors with the highest similarity are selected from the historical data. One sample; Based on constraint violation cost function Value, from the highest similarity From a sample, select one The instruction with the smallest value is used as the instruction scheduling value for each control unit after verification and correction; The expression for the constraint violation cost function is: In the formula, The first one obtained through power flow calculation Each node voltage value , These are the upper and lower limits of the allowable voltage, respectively. For the first The magnitude of the apparent power currently flowing through the line is obtained through power flow calculation. For the first The maximum apparent power capacity, i.e., the current carrying capacity, of this line is allowed to operate safely for a long period of time. , These are penalty coefficients for voltage and apparent power, respectively. These are empirical values ​​and should be adjusted and set according to the different priority requirements of voltage safety and line safety in the actual power grid.

2. The multi-objective distribution network optimization scheduling method based on distributed photovoltaic power and load fluctuation characteristics according to claim 1, characterized in that, The construction of the total loss function for multi-objective distribution network optimization scheduling, based on the target requirements of multi-objective distribution network optimization scheduling and the physical constraints of distributed photovoltaic power generation and load fluctuation characteristics, specifically includes: In terms of economic efficiency, the goal is to minimize network loss costs. An economic loss function is constructed based on traditional power flow calculation with automatic differentiation. In terms of safety, the goal is to minimize the risk of voltage over-limit, and the sum of squares of voltage deviations is quantified to construct a safety loss function; In terms of fluctuation mitigation, a fluctuation mitigation loss function is constructed with the goal of maximizing reactive power reserve capacity. In terms of smoothness of regulation, a smoothness of regulation loss function is constructed with the goal of minimizing the amount of regulation change in adjacent time periods; Based on the physical constraints of photovoltaic active power fluctuations and load active power fluctuations, an active power balance penalty term is constructed. Based on the physical constraints of voltage fluctuations in the distribution network, a voltage regulation penalty term is constructed; Based on the penalty term and each loss function, a weighted summation method is used to construct the total loss function for multi-objective distribution network optimization scheduling. The expression for the total loss function of the multi-objective distribution network optimal scheduling is as follows: In the formula, To optimize the total loss value of multi-objective distribution network scheduling, The number of target demands for optimizing the scheduling of multi-objective distribution networks. For the first The weights of the target demand loss values For the first Each target demand loss value, The number of penalty items. For the first The weight of each penalty item For the first The values ​​of the penalty items include the target requirements: economy, security, volatility resistance, and regulation smoothness. Weighting coefficient , These are preset hyperparameters, whose values ​​are determined by the scheduler based on actual operational needs, and are used to flexibly adjust the priority of different objectives in the overall decision-making process.

3. The multi-objective distribution network optimization scheduling method based on distributed photovoltaic power and load fluctuation characteristics according to claim 2, characterized in that, The process of constructing a multi-objective distribution network optimization scheduling model based on a deep fully connected network and obtaining the instruction scheduling values ​​of each control unit specifically includes: Based on the source-load fluctuation time-series prediction data, real-time power grid state variables, current state of edge nodes, target demand loss value weights, and penalty term weights, the input vector of the deep fully connected network is set. The input vector data of the deep fully connected network is normalized and mapped to the [0,1] interval to eliminate the influence of data units; The control unit instruction values ​​of each edge node at future scheduling moments are used as the prediction labels for the output layer of the deep fully connected network. Based on historical data, a training sample set is constructed according to the input vector of the deep fully connected network and the predicted labels of the output layer of the deep fully connected network, and then divided into a training set and a validation set. Based on mean squared error, the goal of training the deep fully connected network is to minimize the error between the network's predicted label and the target label, and the deep fully connected network is pre-trained. Among them, the target labels are composed of excellent dispatching instruction cases that are recorded in historical operation, actually executed by the power grid dispatching system, and verified to be safe and effective; Based on the pre-trained deep fully connected network, the total loss function of multi-objective distribution network optimization scheduling is used as the training objective of the deep fully connected network to fine-tune the deep fully connected network. In the forward propagation process of the fine-tuning stage, an automatic differential traditional power flow module needs to be embedded. This module takes the output command of the neural network and the real-time state of the power grid as inputs to solve the sub-loss values ​​in the total loss function of multi-objective distribution network optimization scheduling. Based on the training sample set, a deep fully connected network is trained. The trained deep fully connected network is defined as a multi-objective distribution network optimization scheduling model, which is used to obtain the instruction scheduling values ​​of each control unit.

4. The multi-objective distribution network optimization scheduling method based on distributed photovoltaic power and load fluctuation characteristics according to claim 3, characterized in that, The real-time optimization scheduling of the multi-objective distribution network based on the multi-objective distribution network optimization scheduling model and the correction results of the output results specifically includes: Deploy a monitoring device to collect the data required for the input vectors of the deep fully connected network, and acquire the monitoring data of the input vectors of the deep fully connected network in real time; The multi-objective distribution network optimization scheduling model is executed according to a preset rolling cycle to obtain the instruction scheduling values ​​of each control unit; Power flow calculation is performed based on the automatic differential traditional power flow module, the command scheduling values ​​of each control unit are verified and corrected, and the verified and corrected command scheduling values ​​of each control unit are obtained. The command scheduling values ​​of each control unit after verification and correction are issued to optimize and control the control units of each edge node; Furthermore, by utilizing IoT technology, the monitoring data of the input vector of the deep fully connected network and the feedback data after optimization and control are packaged together and uploaded to the data processing center. Construct multi-objective distribution network optimization scheduling data to classify, store, and record data received by the data processing center, facilitating the updating and optimization of the multi-objective distribution network optimization scheduling model.

5. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a multi-objective distribution network optimization scheduling method for distributed photovoltaic and load fluctuation characteristics as described in any one of claims 1-4.

6. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a multi-objective distribution network optimization scheduling method for distributed photovoltaic power and load fluctuation characteristics as described in any one of claims 1-4.

Images