On-line matching method for cooperative supply insurance strategy of power distribution network containing district micro-grid

Abstract

The invention discloses an on-line matching method for a cooperative supply guarantee strategy of a power distribution network containing a district micro-grid, and the method comprises the steps: 1, building an offline historical scene library of the power distribution network containing the district micro-grid, building a basic model of the power distribution network, building a historical typical scene library based on the historical operation data of the power distribution network in an offline manner, and laying a model foundation for the subsequent strategy design; calculating power flow based on a Newton-Raphson algorithm to perform first scene classification, judging optimization targets of different scenes, and establishing a historical scene library; the method comprises the following steps: 1, carrying out on-line scene matching, 2, generating an insurance supply strategy in a corresponding historical typical scene based on network reconstruction and an improved PSO algorithm so as to establish an off-line historical typical scene insurance supply strategy library, and 3, carrying out on-line scene matching based on the historical scene and the insurance supply strategy library and a deep neural network, and carrying out on-line scene matching by learning operation characteristics and load distribution of the historical scene. The historical scene most similar to the online state is recognized, the online state is matched with the optimal historical typical scene, and a real-time guarantee supply strategy is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of operation control of distribution networks, and particularly relates to an online matching method for collaborative power supply guarantee strategies of distribution networks including microgrids in substations. Background Art

[0002] Under the guidance of the "dual carbon" goal, renewable energy has developed rapidly. With the integration of a large number of distributed resources, the microgrid in the substation is promoting the transformation of the distribution network from the traditional "power grid - load" binary structure to the "power source - power grid - load - energy storage" quaternary structure. However, the power generation capacity of distributed new energy is greatly affected by weather changes, and its large-scale access may cause a large deviation between the day-ahead scheduling plan of the substation microgrid and the actual situation, thus triggering a series of problems such as frequent voltage fluctuations and increased grid losses, posing a severe challenge to the safe and economic operation of the substation microgrid and even the entire distribution network. Given the large number of substations in the microgrid and the high complexity of the power supply guarantee model of the distribution network, under the strict requirements of the intraday regulation for the speed of strategy generation, it is urgent to develop an effective method that can quickly formulate collaborative power supply guarantee strategies for distribution networks including substations in the microgrid. Summary of the Invention

[0003] The purpose of the present invention is to provide an online matching method for collaborative power supply guarantee strategies of distribution networks including substations in the microgrid to solve the problems raised in the above background art.

[0004] To achieve the above purpose, the present invention provides the following technical solution: An online matching method for collaborative power supply guarantee strategies of distribution networks including substations in the microgrid, comprising the following steps:

[0005] Step 1: Establish an offline historical scenario library of the distribution network including substations in the microgrid, establish a basic model of the distribution network, and offline construct a historical typical scenario library based on the historical operation data of the distribution network;

[0006] The basic model of the distribution network is established as follows:

[0007] The principle of the Newton - Raphson method is to gradually linearize and iterate repeatedly, and the iteration equation is:

[0008]

[0009] where X is the state variable, ΔX is the state correction amount, F(X) is the function vector, K is the constant matrix, and J is the Jacobian matrix.

[0010] The polar coordinate representation form of the node voltage is:

[0011]

[0012] Perform the first scenario classification by calculating the power flow based on the Newton-Raphson algorithm, judge the optimization objectives (voltage control, network loss optimization) of different scenarios, and establish a historical scenario library;

[0013] Compress the scenarios in the historical scenario library based on the K-means clustering algorithm, classify the scenarios with the same operating characteristics into the same typical scenarios, and obtain several typical voltage control scenarios and network loss optimization scenarios through two scenario classifications;

[0014] Step 2: Establish an off-line historical typical scenario power supply guarantee strategy library, and generate power supply guarantee strategies corresponding to historical typical scenarios based on network reconfiguration and improved PSO algorithm;

[0015] In each iteration of PSO, update the position and velocity of the particle using the individual extreme value and the population extreme value, and the update formula is as follows:

[0016]

[0017]

[0018] where k is the number of iterations; x id and v id are the position and velocity of the d-th particle respectively; ω is the inertia weight; c 1 , c 2 are the learning factors; γ 1 , γ 2 are random numbers from 0 to 1; E id is the individual extreme value; E gd is the population extreme value;

[0019] Aiming at the disadvantages of PSO being easily trapped in local optimum and having low search accuracy, improve the search efficiency by improving the inertia weight and acceleration factor;

[0020] To this end, introduce a linearly decreasing inertia weight LDW in PSO:

[0021]

[0022] where ω max and ω min are the maximum and minimum values of ω respectively, which are set to 0.9 and 0.4 in the present invention; is the number of iterations, k max is the maximum number of iterations;

[0023] Sort the distribution network topology based on the encoding method of the ordered loop network matrix;

[0024] Solve the optimal power supply guarantee strategy corresponding to the historical typical scenario based on the improved particle swarm algorithm;

[0025] Step 3: Design an online scenario matching method. Based on a deep neural network, perform online scenario matching. By learning the operating characteristics and load distribution of historical scenarios, identify the historical scenario that is most similar to the online state, and match the best historical typical scenario for the online state.

[0026] The evaluation formula for the matching effect of the best historical typical scenario is:

[0027]

[0028] where L i is the actual value of node i, that is, the node voltage, active power, and reactive power data in the online state; is the matching value of node i, that is, the historical node voltage, active power, and reactive power data of the same type of historical typical scenario;

[0029] Based on the historical power supply guarantee strategy corresponding to the historical typical scenario, that is, the matching strategy, directly serve as the real-time power supply guarantee strategy for the online state.

[0030] Preferably, in Step 1, the mathematical model of photovoltaic output is:

[0031]

[0032] where P PV is the photovoltaic output power, in kW; L is the light intensity, in kW / m 2 ; P st , L st , T st are the maximum output power, light intensity, and photovoltaic cell surface temperature (25 °C) under standard test conditions respectively; τ is the photovoltaic cell temperature coefficient. T s is the surface temperature of the photovoltaic cell, in °C, expressed as:

[0033] T s = T en + 0.0138(1 + 0.031T en )(1 - 0.042u)L

[0034] where T en is the ambient temperature; u is the wind speed, in m / s.

[0035] Preferably, in Step 1, the energy storage system model is:

[0036]

[0037]

[0038] where is the electricity stored in the battery at time t, in kWh; μ ES,lossis the self - energy loss rate of the battery; η ES,ch and η ES,dis are the charging and discharging efficiencies of the battery respectively; and are the charging and discharging powers of the battery at time t, with the unit of kW; Δt is the time interval, with the unit of h; and are the charging state and discharging state of the battery at time t, indicating that the battery is charging.

[0039] Preferably, in step two, the distribution network topology sorting process based on the ordered loop network matrix coding method is as follows:

[0040] (1) Number each branch. 1 indicates that the switch of the branch is closed, and 0 indicates that the switch is open;

[0041] (2) The number of basic loops is the same as the number of tie switches;

[0042] (3) List the loop matrix according to the basic loops. The number of rows of the loop matrix is the number of basic loops, and the number of columns is the basic loop branch with the most branches. The values at the positions where the other loops do not have branches in each column are 0.

[0043] Preferably, in step two, the linear decreasing inertia weight formula of the improved particle swarm algorithm is as follows:

[0044]

[0045] where ω max and ω min are the maximum and minimum values of ω respectively. In the present invention, they are set to 0.9 and 0.4; is the number of iterations, k max is the maximum number of iterations.

[0046] Preferably, in step two, the learning factor formula of the improved particle swarm algorithm is as follows:

[0047]

[0048] where c 1f , c 1h , c 2f , c 2h are constant constants. According to experience, they are taken as c 1f = 1.5, c 1h = 0.7, c 2f = 2.5, c 2h = 0.5.

[0049] Preferably, in step three, the improved calculation formula of the Softmax layer added after the output layer of the deep neural network is as follows:

[0050]

[0051] Among them, c is a constant, c = max(y ni ). Subtracting a constant from the exponent does not affect the final result.

[0052] Preferably, in step three, based on formula (8), the system matching deviation rate is:

[0053]

[0054] Among them, N bus is the total number of system nodes.

[0055] Preferably, in step three, the optimal historical typical scenario optimization effect evaluation formula is:

[0056]

[0057] Among them, P loss is the active power network loss of the system before optimization; is the active power network loss of the system after optimization.

[0058] Preferably, in step three, the system voltage deviation reduction rate (voltage reduction rate) is:

[0059]

[0060] Among them, ΔU is the system voltage deviation before optimization; ΔU op is the system voltage deviation after optimization.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] (1) By offline constructing a historical scenario library and a multi-objective historical strategy library, and combining an online scenario matching method based on a deep neural network, the present invention significantly improves the matching speed and accuracy of the coordinated power supply guarantee strategy for the distribution network. This method reduces the dependence on real-time complex calculations, enabling rapid response and providing accurate power supply guarantee strategies under the requirements of the generation speed of the intraday regulation countermeasure strategies.

[0063] (2) The present invention adopts network reconfiguration and an improved particle swarm optimization (PSO) algorithm to generate an optimal power supply guarantee strategy for historical typical scenarios, effectively reducing the active power network loss of the system and reducing voltage fluctuations, thereby improving the economy and security of the distribution network operation. At the same time, by comprehensively considering the network loss level and the economic and security constraint conditions for eliminating voltage over-limit, more refined control of the distribution network is achieved.

[0064] (3) The method of the present invention can adapt to the uncertainty and volatility of the output of distributed new energy in the distribution network. Especially when the weather change has a great impact on the distributed new energy generation capacity, by online matching the best historical typical scenarios and real-time adjusting the power supply guarantee strategy, the adaptability and robustness of the system to abnormal situations are enhanced. This dynamic adjustment ability helps to maintain the stable operation of the power grid and can maintain efficient power supply even in the face of emergencies or unpredictable load changes. Description of the Drawings

[0065] Figure 1 is the flowchart of the present invention;

[0066] Figure 2 is the schematic diagram of the orderly loop network matrix coding method based on the IEEE33 example of the present invention;

[0067] Figure 3 is the flowchart of the improved particle swarm optimization algorithm of the present invention;

[0068] Figure 4 is the schematic diagram of the six-layer neural network structure of the DNN adopted by the present invention. Detailed Embodiment

[0069] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0070] The present invention provides a method for online matching of a cooperative power supply guarantee strategy for a distribution network with a substation area microgrid as shown in Figures 1-4 ;

[0071] 1. It includes the following steps:

[0072] Step 1: Establish an offline historical scenario library for the distribution network with a substation area microgrid;

[0073] Step 2: Establish an offline historical typical scenario power supply guarantee strategy library;

[0074] Step 3: Design an online scenario matching method.

[0075] 2. Establishing an offline historical scenario library for the distribution network with a substation area microgrid includes:

[0076] 2.1 Establish a basic model of the distribution network, and based on the Newton-Raphson algorithm, calculate the power flow for the first scenario classification, judge the optimization objectives (voltage control, network loss optimization) of different scenarios, and establish a historical scenario library to lay a foundation for the following content:

[0077] The basic model of the distribution network is established as follows:

[0078] (1) The principle of the Newton-Raphson method is to linearize step by step and iterate repeatedly. The iteration equation is:

[0079]

[0080] where X is the state variable, ΔX is the state correction amount, F(X) is the function vector; J is the Jacobian matrix.

[0081] The polar coordinate representation form of the node voltage is:

[0082]

[0083] Except for the slack node, the injection power of the node is:

[0084]

[0085] where P i and Q i are the active injection power and reactive injection power of node i; V i and V j are the voltage amplitudes of node i and node j respectively; G ij and B ij are the conductance and susceptance of line ij respectively, and θij is the phase angle difference of line ij.

[0086] The deviation of the node injection power is:

[0087]

[0088] where is the Jacobian matrix J, and its off-diagonal elements (i≠j) are:

[0089]

[0090] The diagonal elements of the Jacobian matrix (i = j) are:

[0091]

[0092] where i∈j is the set of nodes connected to node i.

[0093] (2) Distributed photovoltaic model

[0094] Photovoltaic power generation applies the "photovoltaic effect" of semiconductors. Through the semiconductor materials embedded in the photovoltaic panels, solar energy is converted into electrical energy. The influencing factors of solar photovoltaic power generation units mainly include light intensity, cell junction temperature, and ambient temperature. The mathematical model of PV output is as follows:

[0095]

[0096] Among them, P PV is the photovoltaic output power, with the unit of kW; L is the light intensity, with the unit of kW / m 2 ; T s is the surface temperature of the photovoltaic cell, with the unit of °C; P st , L st , T st are the maximum output power, light intensity, and surface temperature of the photovoltaic cell (25 °C) under standard test conditions respectively; τ is the temperature coefficient of the photovoltaic cell.

[0097] The surface temperature of the photovoltaic cell is:

[0098] T s = T en + 0.0138(1 + 0.031T en )(1 - 0.042u)L (8)

[0099] Among them, T en is the ambient temperature; u is the wind speed, with the unit of m / s.

[0100] PVs in the distribution network have both active power regulation ability and reactive power regulation ability. To ensure full consumption of photovoltaic power generation, the present invention considers not reducing the active power output of PVs and only uses the reactive power regulation ability of PVs for optimal control. The reactive power regulation constraint of PV is:

[0101]

[0102] Among them, P PV,i and Q PV,i are the active and reactive powers of the i-th PV node respectively. The expression of P PV,i is shown in Equation (18); cosθ i is the minimum power factor of the i-th PV node; Ω PV is the set of all PV nodes in the distribution network.

[0103] (3) Electric vehicle charging station model

[0104] In the ADN, the controllable loads mainly include three categories: residential loads, commercial loads, and industrial loads. The controllable residential loads are further divided into air conditioners, electric water heaters, electric vehicles, etc. Considering the wide access of EVs to the power system, the controllable load in this invention is studied taking the EV charging station as an example. The charging model of the EV charging station is as follows:

[0105]

[0106] Among them, P EV (t) is the actual power of EV charging at time t; is the rated power of EV charging; s EV (t) is the EV charging state function, where 0 means fully charged and 1 means not fully charged.

[0107] (4) Energy storage system model

[0108] In the optimization and control of ADN, the high proportion of access of DG and EVs, as well as the user's electricity demand, bring risks such as uncertainty and volatility to the distribution network. ES is regarded as an important means to improve the power supply quality, operation safety and economy, and reduce the power supply cost of the system due to its ability of cross-period power energy scheduling. The battery in ES has been widely used due to its advantages of large capacity and low cost. The charge and discharge model of the battery is as follows:

[0109]

[0110]

[0111] Among them, is the electric quantity stored in the battery at time t, with the unit of kWh; μ ES,loss is the self-energy loss rate of the battery; η ES,ch and η ES,dis are the charging and discharging efficiencies of the battery respectively; and are the charging and discharging powers of the battery at time t respectively, with the unit of kW; Δt is the time interval, with the unit of h; and are the charging state and discharging state of the battery at time t, indicating that the battery is charging.

[0112] Based on the above basic model of the distribution network, according to the Newton-Raphson power flow calculation algorithm, the first scenario classification is carried out based on the optimization objectives. According to the objectives of voltage control and network loss optimization, historical scenarios can be divided into three categories: excessive network loss, voltage exceeding the upper limit, and voltage exceeding the lower limit. Among them, the phenomenon of voltage exceeding the upper limit generally occurs at noon when the PV output is high, which is a light load scenario; the situation of voltage exceeding the lower limit is a heavy load scenario when the load is too large, such as a large number of EVs accessing the distribution network for charging. The three scenario types are shown in the following table.

[0113] Table 1 Three Scenario Types

[0114]

[0115] In the present invention, the normal operating range of the node voltage is set to [0.95, 1.05] p.u., where V i is the voltage amplitude of node i, and V low is the lower limit value of the node voltage, and V up is the upper limit value of the node voltage; according to the empirical value, the threshold of the total network loss rate for the 10 kV distribution network is specified as 10%, where is the system network loss rate:

[0116]

[0117] where P loss is the active network loss of the system; and P sup is the total supplied electrical energy of the system.

[0118] 2.2 Compress the scenarios in the historical scenario library based on the K-means clustering algorithm, classify the scenarios with the same operating characteristics into the same typical scenarios, and obtain several typical voltage control scenarios and network loss optimization scenarios through two scenario classifications.

[0119] On the basis of the content in 2.1, for the three types of scenarios of excessive network loss, voltage exceeding the upper limit, and voltage exceeding the lower limit, a second scenario classification is carried out respectively based on the K-means clustering algorithm, and the scenarios with the same operating characteristics are classified into the same typical scenarios.

[0120] The K-means clustering algorithm is a typical clustering method with the characteristics of simple idea, fast convergence speed, and easy implementation, and can achieve efficient clustering of data. Its basic process is as follows:

[0121] Step1: Input the sample data, and set the number of clusters to N clu = 5;

[0122] Step2: Randomly select N clu sample points as the initial cluster centers;

[0123] Step 3: Calculate the Euclidean distance:

[0124]

[0125] Among them, D ij is the Euclidean distance between the i-th sample point and the j-th cluster center in the n-dimensional space; is the value of the i-th sample point in the k-th dimension, and is the value of the j-th cluster center in the k-th dimension;

[0126] Step 4: Based on the results of Step 3, classify the sample data according to the principle of the nearest distance;

[0127] Step 5: According to the results of Step 4, update the cluster centers by using the arithmetic mean of the sample data in each cluster category:

[0128]

[0129] Among them, u′ jk is the value of the j-th new cluster center in the k-th dimension; a′ ik is the value of the i-th sample point in the k-th dimension; Ω j is the set of sample points in the j-th class, and the cluster category of a′ ik is the j-th class; N j is the number of sample points in the j-th class in the k-th dimension;

[0130] Step 6: Repeat Steps 3 to 5 until the cluster centers remain unchanged;

[0131] Step 7: Output the clustering results.

[0132] 3. Establish an off-line historical typical scenario supply guarantee strategy library, and its method is as follows:

[0133] 3.1 Sort the distribution network topology based on the ordered loop network matrix coding method;

[0134] When the intelligent optimization algorithm processes the multi-peak function problem, by reordering the feasible solution space, reducing the peaks or transforming it into a single-peak problem, the search efficiency and the quality of the solution are improved, and the number of iterations is reduced. The present invention sorts the distribution network topology based on the ordered loop network matrix coding method:

[0135] 1) Number each branch, where 1 indicates that the switch of the branch is closed and 0 indicates that the switch is open;

[0136] 2) The number of basic loops is the same as the number of tie switches;

[0137] 3) List the loop network matrix according to the basic loops. The number of rows of the loop network matrix is the number of basic loops, and the number of columns is the basic loop branch with the most branches. The values at the positions where the other loops do not have branches in each column are 0.

[0138] Taking the IEEE 33-node distribution network as an example, as Figure 2 shown, branches B33, B34, B35, B36, and B37 are tie-line branches, and sectional switches are installed on the remaining branches. Therefore, there are 5 basic loops, and the ordered loop matrix of the IEEE 33-node distribution network is as follows:

[0139]

[0140] 3.2 Solving the optimal power supply guarantee strategy for the corresponding historical typical scenarios based on the improved particle swarm optimization algorithm;

[0141] The principle of the Particle Swarm Optimization (PSO) algorithm is simple, easy to implement, and has a fast convergence speed. It has good parallelism, robustness, and anti-interference ability. However, it has limitations such as being sensitive to initial values, being prone to falling into local optima, and having low search accuracy. The three characteristics of particles in PSO are position, velocity, and fitness. Among them, the position of the particle is the solution to the extreme value optimization problem; the velocity is a vector, including the magnitude of the particle's speed and the moving direction; the fitness function is the objective function of the extreme value optimization problem, and the fitness value is the solution to this objective function. In each iteration of PSO, the position and velocity of the particle are updated using the individual extreme value and the population extreme value. The update formulas are as follows:

[0142]

[0143]

[0144] Among them, k is the number of iterations; x id and v id are the position and velocity of the d-th particle respectively; ω is the inertia weight; c 1 , c 2 are the learning factors; γ 1 , γ 2 are random numbers in the range of 0 to 1; E id is the individual extreme value; E gd is the population extreme value.

[0145] Aiming at the shortcomings of PSO being prone to falling into local optima and having low search accuracy, the search efficiency is improved by improving the inertia weight and the acceleration factor.

[0146] 1) Improving the inertia weight

[0147] The parameter inertia weight ω in PSO can maintain the inertia of particle movement, enabling the particle to have the ability to develop new search ranges, that is, the ability to balance global and local searches. A relatively large inertia weight is introduced in the initial stage to enhance the search and analysis ability. As the number of iterations increases, the weight factor gradually decreases to enhance the analysis result ability, thereby improving the search efficiency. Therefore, a linearly decreasing inertia weight LDW is introduced in PSO:

[0148]

[0149] Among them, ω max and ω min are the maximum and minimum values of ω respectively, which are set to 0.9 and 0.4 in the present invention; is the number of iterations, k max is the maximum number of iterations.

[0150] 2) Improved learning factor

[0151] According to experimental analysis, it is known that the learning factor that changes with time has a better effect, that is, at the initial stage, c 1 is larger and c 2 is smaller, which is beneficial to enhancing the global search ability and not easily falling into the local optimum. As the number of iterations increases, c 1 gradually decreases and c 2 gradually increases, which will enhance the local search ability and make the particles in the middle and late stages more likely to converge to the global optimum. The time-varying expression of the learning factor is:

[0152]

[0153] Among them, c 1f , c 1h , c 2f , c 2h are constant constants, and are taken as c 1f =1.5, c 1h =0.7, c 2f =2.5, c 2h =0.5 according to experience.

[0154] The improved PSO algorithm flow is as Figure 3 shown, and the specific steps are as follows:

[0155] Step1: Initialize the parameters, set the inertia weight, acceleration factor, particle population size, particle dimension, extreme values of particle velocity and position, maximum number of iterations, etc.;

[0156] Step2: Randomly initialize the positions and velocities of the particles within the feasible solution space;

[0157] Step3: Calculate the initial fitness value of the particles according to the set objective function;

[0158] Step4: Calculate the initial individual minimum value and the population minimum value;

[0159] Step5: Calculate the dynamic inertia weight and learning factor according to Equations (19) and (20);

[0160] Step6: Update the position and velocity of the particle according to Equation (17);

[0161] Step7: Update the fitness value of the particle;

[0162] Step8: Compare the fitness value of the particle with the individual extreme value and update the individual extreme value;

[0163] Step9: Compare the fitness value of the particle with the population extreme value and update the population extreme value;

[0164] Step10: Determine whether the iteration condition is satisfied (reaching the maximum number of iterations or meeting the convergence accuracy). If the convergence condition is satisfied, terminate the operation and output the global optimal value; otherwise, return to Step6.

[0165] 3) Objective function and constraint conditions

[0166] The calculation formula for the active power network loss of the system is as follows:

[0167]

[0168]

[0169] Among them, P loss is the active power network loss of the distribution network; Ω br is the set of all branches of the distribution network; I ij and r ij are the current and resistance of branch ij respectively; P ij and Q ij are the active power and reactive power at the head end of branch ij respectively; V i is the voltage amplitude of the head end node.

[0170] Considering the switch state on the branch for network reconfiguration, based on Equation (21), the active power network loss of the system is:

[0171]

[0172] Among them, P' loss is the active power network loss of the distribution network after network reconfiguration; s ij is the switch state on branch ij, 1 indicates closed, and 0 indicates open.

[0173] In summary, for the scenarios with excessive network losses in the historical scenario library of the distribution network, the present invention sorts the network topology by using the ordered ring network matrix coding method, performs network reconstruction based on the improved PSO algorithm, takes the tie switches and sectionalizing switches as the control objects, and aims to reduce the active network losses of the system to improve the economic efficiency of the system operation. The objective function of the network reconstruction is as follows:

[0174] F 1 = min P′ loss (24)

[0175] s.t. P′ loss <0.2 (25)

[0176] 4) Multi-objective historical strategy considering network reconstruction

[0177] Based on the results of the network reconstruction, using the improved PSO algorithm, aiming at minimizing the active network losses and voltage fluctuations of the system, comprehensively considering the network loss level and the economic and security constraint conditions for eliminating voltage violations, offline generate historical optimization strategies corresponding to historical scenarios one by one, obtain a multi-objective historical strategy library, and store the key control objects and optimization strategies in the historical strategy library. The objective function of the multi-objective historical optimization strategy is as follows:

[0178] F 2 = min(αP′ loss + βΔU) (26)

[0179]

[0180] where α and β are weight coefficients, and in the present invention, α = 0.1 and β = 1 are set; V i is the voltage amplitude of node i, V up and V low are the upper and lower limits of the node voltage respectively. In the present invention, the normal range of the node voltage is set to [0.95, 1.05] p.u.; ΔU is the system voltage deviation:

[0181]

[0182] where N bus is the total number of nodes in the distribution network; V is the rated voltage of node i, and in per-unit value, it is expressed as

[0183] In addition to satisfying the constraint formula (27), the objective function (26) of the multi-objective historical strategy also needs to satisfy the control variable constraints, including the PV constraint as shown in formula (9) and the ES constraint as shown in formula (12). At the same time, it also needs to satisfy the system safe operation constraints:

[0184]

[0185] wherein, P ij and Q ij are the active and reactive powers of branch ij respectively; S ij,max is the maximum allowable capacity of branch ij.

[0186] 4. Propose an online scenario matching method

[0187] 4.1 Perform online state recognition based on a deep neural network to match the best historical typical scenarios

[0188] The most basic form of deep learning is the artificial neural network. The BP neural network is the most widely used and typical non - linear artificial intelligence algorithm in artificial neural networks. It is a multi - layer feed - forward neural network with a fully - connected form between layers. Based on the perceptron model, a hidden layer is added. The relationship between the input layer and the output layer is represented by the way of learning and training, and it can be used to predict the output result when the quantitative relationship between the unknown output and input is not known. The DNN is a deepening of the artificial neural network, with multiple hidden layers, also known as a multi - layer perceptron. The DNN can represent more complex functional relationships, has numerous adjustable parameters and training algorithms, and has strong operability. It is widely used in aspects such as fault detection, load forecasting, and transient stability assessment in power systems. The DNN adopted in this invention is a six - layer neural network structure, as Figure 4 shown, including an input layer, hidden layers, and an output layer, and the number of hidden layers is three.

[0189] For the training set, use the historical power and voltage data of the same - type historical typical scenarios as the input:

[0190]

[0191] wherein, V, P, Q are matrices of M×N bus , which are the node voltages, injected active powers, and reactive powers of each historical typical scenario in sequence. M is the number of historical typical scenarios, and N bus is the number of distribution network nodes.

[0192] The output of the training set is the matching degrees of M historical typical scenarios:

[0193]

[0194] wherein, the value of each row is the matching degree of the scenario corresponding to the row number among the M historical typical scenarios, and the column where the maximum value of each row is located is the number of the historical typical scenario matched by the scenario corresponding to the row number.

[0195] Define the input of the test set as the real - time measurement data of the online state to be optimized:

[0196]

[0197] Among them, V′, P′, and Q′ are matrices of M′×N ob , where M′ is the number of online states to be optimized, and N ob is the number of real-time measurement nodes.

[0198] The output of the test set is defined as:

[0199]

[0200] Among them, Y′ is the matching result of M′ online states to be optimized in M historical typical scenarios.

[0201] The hidden layer abstracts the features of the input data into other dimensional spaces, presenting more abstract features of the input data for better linear division. The neurons in the DNN test set are consistent with those in the training set. In the present invention, the number of neurons in the four hidden layers is h, l, k, and f in sequence, so the four hidden layers are M×h, M×l, M×k, and M×f matrices in sequence.

[0202] The functional relationships between the layers of the DNN are as follows:

[0203] B = ωA + b (34)

[0204] Among them, A is the input of a certain layer of the neural network, and B is the output of this layer of the neural network; ω and b are the weight and threshold respectively, and the initial parameters are often random values. In the present invention, the dimensions of the weights and thresholds from the input layer to the first hidden layer are N bus ×h and 1×h, from the first hidden layer to the second hidden layer are h×l and 1×l, from the second hidden layer to the third hidden layer are l×k and 1×k, from the third hidden layer to the fourth hidden layer are k×f and 1×f, and from the fourth hidden layer to the output layer are f×M and 1×M.

[0205] The functions of the activation layer include restricting the variable range, preventing gradient disappearance, etc. The commonly used activation functions are the step function, the Sigmod function, and the ReLU function. Since the ReLU function has high accuracy, the activation layer in the present invention adopts the ReLU function:

[0206] R(r) = max(0, r) (35)

[0207] To make the output result more intuitive, a Softmax layer needs to be added after the output layer to normalize the output result and convert it into a probability value for easier finding of the classification result with the maximum probability. The calculation formula of the Softmax layer is as follows:

[0208]

[0209]

[0210] Among them, n, i, j ∈ [1, M], and e = 2.7182…. In the scenario matching problem of the present invention, since i > 1000, in computer programming, e 1000 will be recognized as an infinite value, resulting in the inability to continue subsequent calculations. Now, Equation (36) is improved as follows:

[0211]

[0212] Among them, c is a constant, c = max(y ni ), and subtracting a constant from the exponent does not affect the final result.

[0213] The quality of the result calculated by Softmax using cross-entropy loss is evaluated as follows:

[0214]

[0215] Among them, S′ n is the optimal output value of the nth row of the Softmax layer, that is, the maximum probability value of this row. If S′ n → 1, then error → 0.

[0216] The core principle of DNN is to perform repeated iterations through training and learning, compare the output result with the actual result, continuously update the weights and thresholds of each neuron in the network, and continuously reduce the cross-entropy loss. The backpropagation of the present invention uses the gradient descent algorithm, takes the output of the Softmax layer as the starting point, calculates the weight correction amount Δω and threshold correction amount Δb of each layer. At the same time, in order to avoid the centralization of the weight result, a regularization penalty term is introduced:

[0217] dω = dω + reg·ω (40)

[0218] Among them, reg is the regularization penalty term parameter.

[0219] Finally, the weights and thresholds are updated using the learning rate ε:

[0220]

[0221] 4.2 Generate an online real-time optimization and control scheme based on the best matching historical typical scenario.

[0222] In order to evaluate the effect of scenario matching and the optimization effect of the matching strategy, corresponding evaluation indicators are proposed for quantitative evaluation.

[0223] (1) Evaluation of matching effect

[0224] For the evaluation of the matching effect in the online scenario, two evaluation indexes, namely the node matching deviation rate and the system matching deviation rate, are proposed. The node matching deviation rate is as follows:

[0225]

[0226] where L i is the actual value of node i, that is, the node voltage, active power, and reactive power data in the online state; is the matching value of node i, that is, the historical node voltage, active power, and reactive power data of the same type of historical typical scenario.

[0227] Based on Equation (42), the system matching deviation rate is as follows:

[0228]

[0229] where N bus is the total number of system nodes.

[0230] (2) Optimization effect evaluation

[0231] For the evaluation of the optimization effect of the matching strategy, two evaluation indexes, namely the system loss reduction rate and the system voltage reduction rate, are proposed. The active power network loss reduction rate (loss reduction rate) of the system is as follows:

[0232]

[0233] where P loss is the active power network loss of the system before optimization, and the calculation formula is shown in Equation (32); is the active power network loss of the system after optimization.

[0234] The reduction rate of the system voltage deviation (voltage reduction rate) is as follows:

[0235]

[0236] where ΔU is the system voltage deviation before optimization, and the calculation formula is shown in Equation (28); ΔU op is the system voltage deviation after optimization.

[0237] To sum up, the present invention proposes a method for generating a distribution network matching strategy based on online scenario matching, including an offline stage and an online stage. The flow chart is as Figure 1 shown, and the specific operation process is as follows:

[0238] step1: Offline construction of a historical scenario library. Considering the operation scenarios and characteristics of various resources of the source network, load, and energy storage, based on a data-driven approach, using the historical operation data of the distribution network, that is, historical node voltage, node injected active power, and reactive power data, etc., construct a historical scenario library offline;

[0239] Step 2: Offline generation of the historical typical scenario library. First, based on power flow calculation, the optimization objectives are judged, including three types: excessive network loss, voltage lower limit violation, and voltage upper limit violation. Then, based on the K-means clustering algorithm (the number of clusters is set to 5), the second scenario classification is carried out, and those with similar operating characteristics and physical features are classified into the same typical scenario to improve the speed and accuracy of scenario matching.

[0240] Step 3: Offline generation of the multi-objective historical strategy library. Considering the voltage control and network loss optimization objectives, a multi-objective historical strategy library is generated based on network reconfiguration and the improved PSO algorithm, and the key control objects and optimization control strategies are stored in the historical strategy library to lay a foundation for generating matching strategies.

[0241] Step 4: Preprocess the online state to be optimized. For the online state to be optimized, two scenario classifications are carried out according to Step 2.

[0242] Step 5: Online scenario matching. Using the historical power and voltage data of the same type of historical typical scenarios of the online state to be optimized as the training set, and the real-time voltage and power measurement information of the online state as the training set, online scenario matching is carried out based on the deep neural network. By learning the operating characteristics and load distribution of historical scenarios, the most similar historical scenario and power flow information are matched for the online state.

[0243] Step 6: Evaluation of the matching effect. After matching, two matching effect evaluation indicators, namely the node matching deviation rate and the system matching deviation rate, are calculated.

[0244] Step 7: Generation of the matching strategy. The historical strategy corresponding to the matched historical scenario, that is, the matching strategy, is directly used as the real-time optimization control scheme for the online state.

[0245] Step 8: Evaluation of the optimization effect. Calculate the optimization effect evaluation indicators, the system power loss reduction rate and the power loss reduction rate.

[0246] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An online matching method for a coordinated supply guarantee strategy of a distribution network containing a microgrid in a substation, characterized in that: The following steps are involved: Step 1: Establish an offline historical scenario library of the distribution network including the substation microgrid, establish a basic model of the distribution network, and build an offline historical typical scenario library based on the historical operation data of the distribution network; The basic model of the distribution network is established as follows: The principle of the Newton-Raphson method is to linearize step by step and iterate repeatedly. The iterative equation is: Among them, X is the state variable, ΔX is the state correction, F(X) is the function vector, K is the constant matrix, and J is the Jacobian matrix. The polar coordinate representation of the node voltage is: The power flow is calculated based on the Newton-Raphson algorithm to perform the first scenario classification, determine the optimization goals of different scenarios, and establish a historical scenario library; Based on the K-means clustering algorithm, the scenarios in the historical scenario library are compressed, and the scenarios with the same operating characteristics are classified into the same typical scenarios. Several typical voltage control scenarios and network loss optimization scenarios are obtained through two scenario classifications. Step 2: Establish an offline historical typical scenario supply guarantee strategy library, and generate supply guarantee strategies for corresponding historical typical scenarios based on network reconstruction and improved PSO algorithm; In each iteration, PSO uses individual extreme values ​​and group extreme values ​​to update the position and velocity of particles. The update formula is as follows: Where k is the number of iterations; x id and v id are the position and velocity of the dth particle respectively; ω is the inertia weight; c1 and c2 are learning factors; γ1 and γ2 are random numbers with values ​​of 0 to 1; E id is the individual extreme value; E gd is the population extreme value; Introducing linear decreasing inertia weight LDW in PSO: Among them, ω max and ω min are the maximum and minimum values ​​of ω, respectively, and k max is the maximum number of iterations; Sorting the distribution network topology based on the ordered ring network matrix coding method; Based on the improved particle swarm algorithm, the optimal supply guarantee strategy corresponding to the historical typical scenario is solved; Step 3: Design an online scene matching method, perform online scene matching based on a deep neural network, identify the historical scene that is most similar to the online state by learning the operating characteristics and load distribution of historical scenes, and match the online state with the best historical typical scene; The evaluation formula for the best historical typical scene matching effect is: Among them, L i is the actual value of node i, is the matching value of node i; Based on the historical supply guarantee strategy corresponding to historical typical scenarios, it serves as the real-time supply guarantee strategy for the online state.

2. According to claim 1, a method for online matching of a coordinated supply guarantee strategy of a distribution network containing a substation microgrid is characterized by: In step 1, the mathematical model of photovoltaic output is: Among them, P PV is the photovoltaic output power, in kW; L is the light intensity, in kW / m 2 ;P st , L st , T st are the maximum output power, light intensity, and photovoltaic cell surface temperature under standard test conditions; τ is the photovoltaic cell temperature coefficient, T s is the surface temperature of the photovoltaic cell, in °C, expressed as: T s =T en +0.0138(1+0.031T en )(1-0.042u)L Among them, T en is the ambient temperature; u is the wind speed, in m / s.

3. According to claim 1, a method for online matching of a coordinated supply guarantee strategy of a distribution network containing a substation microgrid is characterized by: In step 1, the energy storage system model is: in, is the amount of electricity stored in the battery at time t, in kWh; μ ES,loss is the energy loss rate of the battery itself; η ES,ch and η ES,dis are the battery charging and discharging efficiencies respectively; and are the battery charging and discharging power at time t, in kW; Δt is the time interval, in h; and is the charge state and discharge state of the battery at time t, Indicates that the battery is charging.

4. According to claim 1, a method for online matching of a coordinated supply guarantee strategy of a distribution network containing a substation microgrid is characterized by: In step 2, the distribution network topology sorting process based on the ordered ring network matrix encoding method is as follows: (1) Number each branch, 1 means the switch of the branch is closed, and 0 means the switch is open; (2) The number of basic loops is the same as the number of tie switches; (3) List the ring network matrix based on the basic loops. The number of rows in the ring network matrix is ​​the number of basic loops, and the number of columns is the number of basic ring branches with the largest number of branches. The value of the position where there is no branch in the remaining loops in each column is 0.

5. According to claim 1, a method for online matching of a coordinated supply guarantee strategy of a distribution network containing a substation microgrid is characterized by: In step 2, the linear decreasing inertia weight formula of the improved particle swarm algorithm is as follows: Among them, ω max and ω min are the maximum and minimum values ​​of ω respectively.

6. The online matching method for the coordinated supply guarantee strategy of a distribution network containing a substation microgrid according to claim 1 is characterized by: In step 2, the learning factor formula of the improved particle swarm algorithm is as follows: Among them, c 1f 、c 1h 、c 2f 、c 2h is a constant.

7. The online matching method for the coordinated supply guarantee strategy of a distribution network containing a substation microgrid according to claim 1 is characterized by: In step 3, the improved calculation formula of the Softmax layer added after the output layer of the deep neural network is as follows: Where c is a constant, c = max(y ni ).

8. The online matching method for the coordinated supply guarantee strategy of a distribution network including a substation microgrid according to claim 1 is characterized by: In step 3, in the formula Based on , the system matching deviation rate is: Among them, N bus is the total number of system nodes.

9. The online matching method for the coordinated supply guarantee strategy of a distribution network containing a substation microgrid according to claim 1 is characterized by: In step 3, the best historical typical scenario optimization effect evaluation formula is: Among them, P loss is the system active network loss before optimization; is the active network loss of the optimized system.

10. The online matching method for the coordinated supply guarantee strategy of a distribution network including a substation microgrid according to claim 1 is characterized in that: In step 3, the system voltage deviation reduction rate is: Among them, ΔU is the system voltage deviation before optimization; ΔU op is the optimized system voltage deviation.

Images