A power distribution network operation management method fusing learning-optimization

Abstract

The application discloses a kind of fusion learning-optimal power distribution network operation management method, comprising the following steps: S1, constructing power distribution network operation model and distributed resource cluster response model;S2, propose constraint learning algorithm for distributed resource cluster, to learn the response characteristics of distributed resource cluster, obtain the data-driven response model based on neural network;S3, the data-driven response model is equivalently converted into neural network response constraint model, and is fused to power distribution network dispatching model, realizes fusion scheduling;S4, optimizes power distribution network decision.

Description

TECHNICAL FIELD

[0001] The present application relates to power distribution network operation management technology, and in particular to a power distribution network operation management method combining learning and optimization. BACKGROUND

[0002] To realize low-carbon energy transformation of the power system, a large number of renewable distributed resources represented by distributed photovoltaic, wind power and energy storage are connected to the power grid, and the diversity and complexity of the operation characteristics of the renewable distributed resources bring great challenges to the operation of the power distribution network. The traditional centralized control cannot adapt to the grid connection of a large number of distributed resources, and there are problems such as system reliability, massive communication and information privacy. A large number of distributed resources often form clusters by aggregation, and participate in the management of the power distribution network in the form of clusters. However, the response characteristics of the distributed resource clusters are complex, and how to effectively motivate and manage the distributed resources is a key problem in the operation and management of the power distribution network power system. In addition, the internal control strategy of the distributed resources is complex and involves user privacy, and the operator of the power distribution network cannot uniformly model and control. SUMMARY

[0003] A large number of distributed resources often form clusters by aggregation, and participate in the management of the power distribution network in the form of clusters. However, the response characteristics of the distributed resource clusters are complex, and how to model the response characteristics of the distributed resources, how to effectively motivate and manage the distributed resources, and how to effectively consider the response characteristics of the distributed resources in operation and dispatch are all key problems in the operation and management of the power distribution network power system, which need to be strengthened. In view of this, the present application proposes a power distribution network operation management method combining learning and optimization to solve the above problems.

[0004] According to one aspect of the present application, a power distribution network operation management method combining learning and optimization is provided, comprising the following steps: S1, constructing a power distribution network operation model and a distributed resource cluster response model; S2, proposing a constraint learning algorithm for distributed resource clusters to learn the response characteristics of the distributed resource clusters and obtain a data-driven response model based on a neural network; S3, equivalently converting the data-driven response model into a neural network response constraint model and integrating it into a power distribution network dispatching model to realize integrated dispatching; and S4, optimizing power distribution network decision-making.

[0005] Further, in step S1, when constructing the power distribution network operation model and the distributed resource cluster response model, the operator of the power distribution network takes minimizing the operation cost of the operation system as the target, considers the power flow constraints, line flow capacity constraints and node voltage constraints of the power distribution network, and motivates the distributed resource clusters to participate in the operation and dispatching of the power system.

[0006] Further, in step S1, based on the power distribution network operation model, a distributed resource cluster response model is constructed: the distributed resource cluster accepts the power distribution network price, takes maximizing cluster revenue as the goal, considers photovoltaic, energy storage, gas turbine standby operation constraints and load demand constraints in the distributed resource, and constructs the distributed resource cluster response model.

[0007] Further, step S2 includes: first analyzing the response characteristics of the distributed resource, constructing a distributed resource data set through data sampling, and then constructing and training the neural network-based data-driven response model.

[0008] Further, the input of the neural network is the incentive price, the hidden layer is the neuron based on the ReLU activation function, and the output is the distributed resource cluster response; by training the neural network, the neural network-based data-driven response model is constructed, the inference of the response characteristics of the distributed resource is realized, and the trained network model does not depend on the operation parameters of the distributed resource, realizing the protection of user privacy.

[0009] Further, step S3 equivalently converts the data-driven response model into a neural network response constraint model, including: the nonlinear part of the neural network is the ReLU activation function of the hidden layer neuron, the core of the ReLU activation function is to set the negative input value to 0 output, and the positive value remains unchanged output, therefore, the ReLU is equivalently converted into a mixed integer linear constraint by using the large M method, and the neural network response constraint model of the distributed resource cluster is constructed by using the mixed integer linear constraint.

[0010] Further, step S3 includes: based on the neural network response constraint model, a power distribution network operation framework FLO4DNO integrating machine learning and optimization is proposed, and carbon emission constraints are considered to realize carbon emission management of the power distribution network, the Mccomick method is used to process nonlinear terms, and the FLO4DNO low-carbon operation model is proposed by combining the power distribution network operation model, realizing the management of distributed resources and carbon emission management of the power distribution network.

[0011] Further, the FLO4DNO low-carbon operation model includes two parts of optimization decision and machine learning, the optimization decision provides a data basis for machine learning, and the machine learning provides a data-driven constraint model for the optimization decision, and the two parts are integrated through the data generation and constraint learning process.

[0012] Further, the optimization decision includes: a distributed resource cluster response, specifically including: constructing the distributed resource cluster response model, calculating the response amount under different incentive prices, and generating a data set for training; a power distribution network decision model, specifically including: based on the power distribution network scheduling model, constructing an economic optimization as the power distribution network optimization scheduling model; a neural network constraint, specifically including: the data-driven response model based on the neural network is equivalent to the neural network response constraint model, which is used for joint optimization scheduling; power distribution network decision, specifically including: solving the joint model of the power distribution network and the neural network response constraint model to obtain the optimal scheduling scheme of the power distribution network.

[0013] Further, the machine learning includes: data sampling, specifically including: using the distributed resource cluster response model to collect the response under different incentive prices to form an offline training data set; model training, specifically including: based on the sampling data set, constructing a neural network model, initializing related parameters, and optimizing model parameters to minimize inference error to construct the data-driven response model based on the neural network; model equivalence, specifically including: based on the equivalent conversion strategy of the large M method, the trained data-driven response model is equivalent to a set of constraints to construct the neural network response constraint model; learning evaluation, specifically including: through establishing related index analysis data model modeling accuracy and effectiveness.

[0014] The learning-optimization integrated power distribution network operation management method provided by the present application uses a machine learning model to mine and represent distributed resource response characteristic power system knowledge from distributed resource response data, uses optimization theory to model the response characteristics, and integrates the two to support the operation management of the power distribution network by using the advantages of machine learning and optimization theory, and realizes the safe and economic operation and carbon emission management of the power distribution network. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of the learning-optimization integrated power distribution network operation management method provided by the embodiment of the present application.

[0016] Figure 2 is a neural network structure diagram of the embodiment of the present application.

[0017] Figure 3 is a power distribution network operation framework integrated with machine learning and optimization of the embodiment of the present application. DETAILED DESCRIPTION

[0018] The present application will be further described below in combination with the drawings and specific implementation manners.

[0019] The embodiment of the application provides a power distribution network operation management method fusing learning and optimization, which uses a machine learning model to mine and represent unknown power system knowledge from data, uses optimization theory to model power system existing knowledge, fuses the two, and uses advantages of machine learning and optimization theory to support operation management of the power distribution network. Figure 1 The power distribution network operation management method fusing learning and optimization of the embodiment of the application specifically comprises the following steps:

[0020] S1, constructing a power distribution network operation model and a distributed resource cluster response model;

[0021] The method of the embodiment of the application first constructs a power distribution network operation model and a distributed resource cluster response model. A power distribution network operator takes minimizing operation cost of an operation system as a target, considers power distribution network power flow constraints, line power flow capacity constraints and node voltage constraints, and encourages a distributed resource cluster to participate in power system operation scheduling. Based on this, the following power distribution network operation model is constructed:

[0022]

[0023]

[0024]

[0025]

[0026] ||2P ij,t 2Q ij,t l ij,t -v i,t ||≤l ij,t +v i,t

[0027]

[0028]

[0029] Wherein, t represents a scheduling time, T represents a planning time length, λ S,t represents a main grid price, P t S represents power purchased from the main grid, DER represents a distributed resource set, λ a,t represents a price of the distributed resource cluster a at the time t, represents power purchased from the cluster a, E represents a line set, P jk,t represents active power flowing through the line jk at the time t, P ij,t represents power flowing through the line ij at the time t, r ij represents resistance of the line ij, l ij,tdenotes the square of the current of line ij at time t, denotes the photovoltaic active power output of node j at time t, denotes the load demand of node j at time t, Q jk,t denotes the reactive power flowing through line jk at time t, Q ij,t denotes the reactive power flowing through line ij at time t, x ij denotes the reactance of line ij, denotes the reactive power output of node j at time t, denotes the reactive power output of node j at time t, v j,t denotes the square of the voltage of node j at time t, v i,t denotes the square of the voltage of node i at time t, denotes the minimum and maximum values of the voltage of node i at time t, N B denotes the set of system nodes, denotes the minimum and maximum values of the current of line ij at time t.

[0030] On the basis of the above-mentioned power distribution network operation model, a distributed resource cluster response model is constructed. The distributed resource cluster accepts the price of the power distribution network, takes maximizing the cluster benefit as the target, considers the photovoltaic, energy storage, gas turbine standby operation constraints and load demand constraints in the distributed resource, and constructs the following distributed resource cluster response model:

[0031]

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] wherein U(·) denotes the benefit function of the load, denotes the load demand of user u at time t, denotes the price of the power distribution network at time t, denotes the active power and reactive power input of the power distribution network, C DG (·) denotes the cost function of the distributed power supply in the cluster, denotes the active power and reactive power output of the distributed power supply at time t, denotes the active power and reactive power output of the renewable energy at time t, Let represent the active power demand and reactive power demand of node i at time t, respectively; Let represent the minimum and maximum output of the distributed power source at time t, respectively. This indicates the maximum ramp power of the distributed power source. This indicates the maximum active power output of renewable energy.

[0038] S2. A constraint learning algorithm for distributed resource clusters is proposed to learn the response characteristics of distributed resource clusters and obtain a data-driven response model based on neural networks.

[0039] Based on the above problem modeling, this invention proposes a constraint learning algorithm for distributed resource clusters. First, the response characteristics of distributed resources are analyzed, a distributed resource dataset is constructed through data sampling, and then a data-driven response model based on neural networks is constructed and learned.

[0040] The constrained learning algorithm includes two processes: the neural network model training process and the equivalent transformation of the network model into a constrained model. The electricity generation and carbon emissions of cluster a are determined by the incentive price decision λi of the upper-level distribution network, based on which a cluster price response model is constructed.

[0041] To accurately model complex cluster response relationships from data, this invention utilizes the powerful fitting performance of Neural Networks (NNs) in machine learning—specifically, constraint learning—to construct a data-driven response model, namely, the neural network-based data-driven response model. Considering the output... Since the variables are continuous, learning response characteristics is essentially a regression problem. Therefore, to construct a neural network based on the rectified linear unit (ReLU) to learn response characteristics, please refer to [reference needed]. Figure 2 The neural network takes the incentive price as input, its hidden layers consist of neurons based on the ReLU activation function, and its output is the distributed resource cluster response, including electricity consumption and carbon emissions. By training this neural network, a data-driven response model based on the neural network is constructed, enabling inference about the response characteristics of distributed resources. The trained network model does not depend on the operating parameters of the distributed resources, thus protecting user privacy. Therefore, the following data-driven response model based on the neural network is constructed:

[0042]

[0043]

[0044] Among them, v l i ReLU(x) represents the i-th neuron in the l-th layer of the network, where ReLU(x) := max(0,x). N represents the bias of the i-th neuron in the l-th layer of the network. l-1 This represents the set of neurons in the (l-1)th layer of the network. This represents the weights between neurons i and j in the l-th layer of the network. Let L represent the j-th neuron in the (l-1)-th layer of the network, and L represent the number of layers in the neural network. v represents the predicted response of the neural network output. L This represents the set of neurons in the output layer. N represents the bias of a neuron. L-1 This represents the set of neurons in the (L-1)th layer of the neural network. This represents the weight of neuron j in the Lth layer of the neural network. Let j represent the neuron j in the (L-1)th layer of the neural network, v0 represent the input neuron, and λ represent the incentive price.

[0045] Based on the neural network structure, a loss function is defined to quantify the output value (predicted value) of the neural network. and the true value y der The distance between them is calculated by optimizing the neural network parameters θ to make the predicted value closer to the true value. Addressing the nature of the regression problem, the loss function is constructed using Mean Square Error (MSE):

[0046]

[0047] Where N represents the number of samples.

[0048] Based on neural network structure and loss function, to minimize loss function L mse With the goal of improving the accuracy of inference values, we use the Stochastic Gradient Descent (SGD) algorithm to optimize the neural network parameters θ.

[0049] S3. The neural network-based data-driven response model obtained after training in step S2 is equivalently converted into a neural network response constraint model and integrated into the distribution network dispatching model to achieve integrated dispatching.

[0050] Due to the strong nonlinearity of neural networks, data-driven response models are difficult to directly participate in the operation and scheduling of distribution networks. To address this issue, this invention first analyzes the internal structure of neural networks and the connection characteristics of their hidden layers. Then, based on the characteristics of data-driven response models, a neural network response constraint model is constructed. This response constraint model is a mixed integer constraint, suitable for the joint scheduling of distributed resources and distribution networks.

[0051] The nonlinear part of the neural network model is the ReLU activation function of the hidden layer neurons. The core of the ReLU function is to set the negative input value to 0 output, and the positive value remains unchanged output. Therefore, the ReLU equivalent is converted into a mixed integer linear constraint by using the large M method as follows:

[0052] v≥x,v≥0

[0053] v≤x-M L (1-w)

[0054] v≤M U w,w∈{0,1}

[0055] Where M L is the minimum value, M U is the maximum value, and w is an auxiliary 0-1 variable.

[0056] Equivalence analysis: when the input x<0, w takes 0, and the output v is 0; when the input x>0, w takes 1, and the output v is x, verifying the equivalence of the constraints.

[0057] The constraint equivalent conversion relationship of the ReLU applying the optimization relaxation strategy large M method is applied to the ReLU activation function inside the neural network. The data-driven response model after training is further equivalent to a set of mixed integer linear constraints, and a neural network response constraint model of distributed resource clusters is constructed:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Where, represents the i-th neuron in the first layer network, represents the bias of the i-th neuron in the first layer network, X represents the input layer neuron set, x j represents the input variable of the j-th neuron, N1 represents the neuron set of the first layer network, represents the neuron unactivated output of the l-th layer network, N l represents the neuron set of the l-th layer network, represents the minimum value of the i-th neuron of the l-th layer network, u l,i represents an auxiliary 0-1 variable, maxi, l represents the maximum value of neuron i in the lth layer network, biasi, l represents the bias of the ith neuron in the Lth layer network, N L-1 represents a set of neurons of the L-1th layer network, wi, j, l represents the weight between neuron i and neuron j in the Lth layer network, represents the jth neuron of the L-1th layer network, N L represents a set of neurons of the Lth layer network.

[0065] The above response characteristic model learning and equivalent conversion to the response constraint model are the core of constraint learning, and the final construction of the response constraint model is the basis for realizing the low-carbon operation of the distribution network by fusing machine learning and optimization.

[0066] Based on the neural network response constraint model obtained as described above, a distribution network operation framework (Fusing Learning and Optimization for Distribution Network Operation, FLO4DNO) is proposed, which realizes carbon emission management of the distribution network by considering carbon emission constraints, processes nonlinear terms by using the Mccomick method, and proposes a FLO4DNO low-carbon operation model by combining a distribution network operation model, so as to realize distributed resource management and carbon emission management of the distribution network.

[0067] The FLO4DNO low-carbon operation framework is as shown in Figure 3 , which includes an optimization decision and a machine learning, the optimization decision provides a data basis for the machine learning, the machine learning provides a data-driven constraint model for the optimization decision, and the two parts are fused through a data generation and constraint learning process.

[0068] As shown in Figure 3 , the optimization decision part includes:

[0069] Distributed resource cluster response, specifically including: constructing the distributed resource cluster response model, calculating the response amount under different incentive prices, and generating a data set for training;

[0070] Distribution network decision model, specifically including: based on the distribution network scheduling model, constructing an optimal economic model as the distribution network optimization scheduling model;

[0071] Neural network constraint, specifically including: equivalent conversion of the data-driven response model based on the neural network into the neural network response constraint model, for joint optimization scheduling;

[0072] Distribution network decision, specifically including: solving the joint model of the distribution network and the neural network response constraint model to obtain the optimal scheduling scheme of the distribution network.

[0073] The constraint learning method for the distributed resource cluster is proposed for the distributed resource operation problem of the power distribution network.

[0074] With reference to the foregoing Figure 3 The machine learning part includes:

[0075] Data sampling, specifically including: collecting the response under different incentive prices by using the distributed resource cluster response model to form an offline training data set;

[0076] Model training, specifically including: based on the sampling data set, constructing a neural network model, initializing the related parameters, optimizing the model parameters with the minimum inference error as the target, and constructing the neural network-based data-driven response model;

[0077] Model equivalence, specifically including: based on the equivalent conversion strategy of the large M method, the trained data-driven response model is equivalent to a set of constraints, and the neural network response constraint model is constructed;

[0078] Learning evaluation, specifically including: analyzing the accuracy and effectiveness of the data model modeling by establishing relevant indicators.

[0079] Compared with the optimization scheduling method based on optimization or machine learning only, the constraint learning method for the distributed resource cluster is proposed based on machine learning and optimization decision, and the low-carbon energy transformation of the power system is faced. In the scenario of a large number of renewable distributed resources such as distributed photovoltaic, wind power, and energy storage accessing the power grid, a data-driven response model based on a neural network is constructed. Then, a basic framework integrating machine learning and optimization is constructed based on constraint learning. Considering the constraint model, carbon emission constraint, power distribution network operation model, and bilinear relaxation strategy, the power distribution network operation model integrating machine learning and optimization is constructed to realize the safe and economic operation of the power distribution network and carbon emission management.

[0080] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent substitutions or obvious variations can be made, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present application.

Claims

1. A distribution network operation management method integrating learning and optimization, characterized in that, Includes the following steps: S1. Constructing the distribution network operation model and the distributed resource cluster response model; When constructing the distribution network operation model and the distributed resource cluster response model, the distribution network operator aims to minimize the operating cost of the operating system, considering distribution network power flow constraints, line power flow capacity constraints and node voltage constraints, and incentivizing the distributed resource cluster to participate in power system operation and scheduling. In step S1, based on the distribution network operation model, a distributed resource cluster response model is constructed: the distributed resource cluster accepts the distribution network price, with the goal of maximizing cluster revenue, and considers the operational constraints of photovoltaic, energy storage, and gas turbine within the distributed resources, as well as load demand constraints, to construct the distributed resource cluster response model. S2. A constraint learning algorithm for distributed resource clusters is proposed to learn the response characteristics of distributed resource clusters and obtain a data-driven response model based on neural networks. S3. The neural network-based data-driven response model is equivalently converted into a neural network response constraint model and integrated into the distribution network scheduling model to achieve integrated scheduling. The nonlinear part of the neural network is the ReLU activation function of the hidden layer neurons. The core of the ReLU activation function is to set the negative input value to 0 and the positive value to remain unchanged. Therefore, the Big M method is used to equivalently convert ReLU into mixed integer linear constraints. Then, the mixed integer linear constraints are used to construct the neural network response constraint model for the distributed resource cluster. Based on the neural network response constraint model, a distribution network operation framework integrating machine learning and optimization is proposed. Carbon emission constraints are considered to achieve distribution network carbon emission management. Then, nonlinear terms are processed, and a low-carbon operation model is proposed in conjunction with the distribution network operation model to achieve distributed resource management and carbon emission management of the distribution network. S4. Optimize distribution network decisions.

2. The fusion learning-optimization distribution network operation management method as described in claim 1, characterized in that, Step S2 includes: firstly analyzing the response characteristics of distributed resources, constructing a distributed resource dataset through data sampling, and then constructing and training the data-driven response model based on neural networks.

3. The fusion learning-optimization distribution network operation management method as described in claim 2, characterized in that, The neural network takes the incentive price as input, has neurons in its hidden layer based on the ReLU activation function, and outputs the distributed resource cluster response. By training the neural network, a data-driven response model based on the neural network is constructed, enabling reasoning about the response characteristics of distributed resources. The trained network model does not depend on the operating parameters of the distributed resources, thus protecting user privacy.

4. The fusion learning-optimization distribution network operation management method as described in claim 1, characterized in that, In step S3, a distribution network operation framework FLO4DNO integrating machine learning and optimization is proposed. Carbon emission constraints are considered to realize carbon emission management of the distribution network. The McConick method is used to handle nonlinear terms. Combined with the distribution network operation model, the FLO4DNO low-carbon operation model is proposed to realize distributed resource management and carbon emission management of the distribution network.

5. The fusion learning-optimization distribution network operation management method as described in claim 4, characterized in that, The FLO4DNO low-carbon operation model comprises two parts: optimization decision-making and machine learning. Optimization decision-making provides the data foundation for machine learning, while machine learning provides the data-driven constraint model for optimization decision-making. The two parts are integrated through data generation and constraint learning processes.

6. The fusion learning-optimization distribution network operation management method as described in claim 5, characterized in that, The optimization decision includes: The distributed resource cluster response specifically includes: constructing the distributed resource cluster response model, calculating the response volume under different incentive prices, and generating a dataset for training. The distribution network decision-making model specifically includes: based on the distribution network scheduling model, constructing an optimal distribution network scheduling model with economic optimization as the goal; Neural network constraints specifically include: the equivalent transformation of a data-driven response model based on neural networks into a neural network response constraint model for joint optimization scheduling; Distribution network decision-making specifically includes: solving the joint model of the distribution network and the neural network response constraint model to obtain the optimal dispatch scheme for the distribution network.

7. The fusion learning-optimization distribution network operation management method as described in claim 5, characterized in that, The machine learning includes: Data sampling specifically includes: using the distributed resource cluster response model to collect response data under different incentive prices to form an offline training dataset; Model training specifically includes: constructing a neural network model based on the sampled dataset, initializing relevant parameters, optimizing the model parameters with the goal of minimizing inference error, and constructing the data-driven response model based on the neural network. Model equivalence specifically includes: an equivalence transformation strategy based on the Big M method, which equivalences the trained data-driven response model to a set of constraints, and constructs the neural network response constraint model; Learning assessment specifically includes: analyzing the accuracy and effectiveness of data modeling by establishing relevant indicators.

Images