Power distribution network reactive power control method and system based on hybrid model

Abstract

The invention provides a hybrid model-based reactive power control method for a power distribution network. The method comprises the following steps of predicting a tapping position of an hour-level on-load tap-changing transformer of the power distribution network by using a mixed integer second-order cone programming algorithm; and reactive power optimization is carried out on the power distribution network by using a deep learning model, and the deep learning model is used for generating optimal reactive power output of a photovoltaic unit and a wind power unit according to the tapping position of the hour-level on-load tap-changing transformer and load, photovoltaic and wind power output data of the power distribution network, so that the sum of voltage deviation and network loss at each time point is minimum. According to the hybrid model-based reactive power control method for the power distribution network, the accuracy of a traditional optimization algorithm and the generalization ability of a deep learning algorithm can be combined to adapt to the uncertainty of DERs and loads, and more efficient and more reliable reactive power management is realized; the optimization performance is improved, the dependence on a network model is reduced, and the flexibility and adaptability of the system are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system control, and in particular relates to a hybrid model-based reactive power control method and system for a distribution network. Background Art

[0002] In modern power systems, the management and optimization of distribution networks are key to ensuring energy efficiency and power supply reliability. With the large-scale integration of distributed energy resources (DERs) such as solar photovoltaic (PV) and wind power, the operation of distribution networks has become more complex. The volatility and uncertainty of these resources have brought difficulties to the voltage control and reactive power management of the network.

[0003] Traditional reactive power control methods have difficulty adapting to rapidly changing voltage conditions. Due to the uncertainty of renewable energy output and the fluctuation of load demand, traditional model-based algorithms face the problems of increased computational complexity, slow convergence and the possibility of falling into local optimality when dealing with large-scale, high-dimensional and nonlinear problems.

[0004] To overcome these limitations, researchers have begun to explore the application of artificial intelligence technology, especially deep learning algorithms, to reactive power optimization in distribution networks in recent years. Deep learning methods can learn from historical data and predict system behavior, thereby providing faster and more flexible control strategies. However, most existing deep learning models, including traditional deep neural networks (DNNs) and convolutional neural networks (CNNs), cannot fully capture the complex relationship between global states and physical quantities when dealing with large-scale networks with high spatial correlation.

[0005] It can be seen that in the face of today's large-scale, high-dimensional distribution network structure, the existing control methods, whether they are traditional model-based algorithms or deep learning algorithms based on artificial intelligence, are difficult to simultaneously have excellent computing speed and reactive power optimization effects. Summary of the invention

[0006] One of the objects of the present invention is to solve at least one or more of the above-mentioned problems existing in the prior art. In other words, one of the objects of the present invention is to provide a distribution network reactive power control method and system based on a hybrid model that meets one or more of the above-mentioned requirements.

[0007] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a reactive power control method for a distribution network based on a hybrid model, comprising the following steps: A mixed integer second-order cone programming algorithm is used to predict the hourly on-load tap changer tap positions in distribution networks. A deep learning model is used to optimize the reactive power of the distribution network. The deep learning model is used to generate the optimal reactive power output of the photovoltaic unit and the wind power unit according to the tap position of the hourly on-load tap-changing transformer and the load, photovoltaic and wind power output data of the distribution network, so as to minimize the voltage deviation and network loss at each time point.

[0008] As a preferred implementation, the deep learning model is trained by the following method: Obtain a training data set, which includes the load, photovoltaic output, and wind power output of the distribution network; Use a deep learning model to perform reactive power optimization based on the training data set, score the results of virtual reactive power optimization, and use the score to feed back the deep learning model to optimize the virtual reactive power optimization effect of the deep learning model.

[0009] As a further preferred implementation, scoring the result of virtual reactive power optimization includes: Solve the SOCP problem using the training data set and obtain the optimal result; The results of virtual reactive power optimization are scored using the optimal result.

[0010] As a preferred implementation, reactive power optimization includes time-scale optimization of discrete power flow constraints, node voltage constraints, branch current constraints, and renewable energy reactive power output constraints.

[0011] As a preferred implementation, the goal of setting the hourly OLTC tapping position of the distribution network is to reduce the total amount of voltage variation, network loss and OLTC decomposed operation cost.

[0012] As a preferred implementation, the structure of the deep learning model includes an encoder and a decoder; The hidden layer size of the encoder and decoder is 256 and the number of layers is 4; The deep learning model includes an attention mechanism, the number of attention heads is 2, and the activation function is the ReLU function.

[0013] On the other hand, the present invention also provides a distribution network reactive power control system based on a hybrid model, comprising: a day-ahead control prediction module, a deep learning model; The day-ahead prediction module is used to predict the hourly on-load tap changer tap positions of the distribution network using a mixed integer second-order cone programming algorithm. The deep learning model is used to optimize the reactive power of the distribution network. According to the hourly on-load tap-changing transformer tap position and the load, photovoltaic and wind power output data of the distribution network, the optimal reactive power output of the photovoltaic unit and the wind power unit is generated to minimize the voltage deviation and network loss at each time point.

[0014] As a preferred implementation, the deep learning model is trained by the following method: Obtain a training data set, which includes the load, photovoltaic output, and wind power output of the distribution network; Use a deep learning model to perform reactive power optimization based on the training data set, score the results of virtual reactive power optimization, and use the score to feed back the deep learning model to optimize the virtual reactive power optimization effect of the deep learning model.

[0015] As a preferred implementation, the day-ahead control prediction module uses a mixed integer second-order cone programming algorithm to predict the hourly on-load tap changer tap positions of the distribution network, with the goal of reducing the total amount of voltage variation, network loss and OLTC decomposed operation cost.

[0016] As a preferred implementation, the structure of the deep learning model includes an encoder and a decoder; The hidden layer size of the encoder and decoder is 256 and the number of layers is 4; The deep learning model includes an attention mechanism, the number of attention heads is 2, and the activation function is the ReLU function.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The reactive power control method and system of the distribution network based on the hybrid model of the present invention can combine the accuracy of the traditional optimization algorithm and the generalization ability of the deep learning algorithm to adapt to the uncertainty of DERs and loads, and realize more efficient and reliable reactive power management. The traditional algorithm ensures the globality of the optimization problem and the satisfaction of physical constraints, while the deep learning algorithm provides a fast response to real-time data and a deep understanding of complex relationships. This hybrid method not only improves the optimization performance, but also reduces the dependence on the network model and improves the flexibility and adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of a reactive power control method for a distribution network based on a hybrid model according to an embodiment of the present invention; Figure 2 It is a structural diagram of a test power saving system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] The following description provides examples and does not limit the scope, applicability or examples set forth in the claims. Changes may be made to the functions and arrangements of the elements described without departing from the scope of the present invention. Various processes or components may be appropriately omitted, substituted or added to each example. For example, the described method may be performed in an order different from the described order, and various steps may be added, omitted or combined. In addition, the features described in some examples may be combined in other examples.

[0021] The present invention provides a reactive power control method for a distribution network based on a hybrid model, the process of which is as follows: Figure 1 As shown, the following steps are included: Flowchart as Figure 1 As shown, the steps are as follows: S1. Use mixed integer second-order cone programming algorithm to predict the hourly on-load tap changer tap positions of distribution networks.

[0022] The mixed integer second-order cone programming algorithm described above is capable of handling nonlinear power flow constraints. By linearizing these constraints, the algorithm can accurately simulate and optimize the relationship between physical quantities in the distribution network, such as the relationship between node voltage and line current.

[0023] With the above characteristics, the mixed integer second-order cone programming algorithm is able to optimize the entire distribution network, taking into account all relevant physical and operational constraints. This means that it can capture and optimize the operating status of the network at a global level, including the topology of the distribution network, line parameters, load forecasts, and renewable energy output forecasts to determine the hourly on-load tap changer transformer tap position.

[0024] The specific method of predicting the hourly on-load tap changer tap position of the distribution network by using the mixed integer second-order cone programming algorithm is as follows: According to the forecast values ​​of photovoltaic, wind power and load, the tapping position of the on-load tap-changing transformer is arranged every hour one day in advance. In the process of arranging the tapping position of the on-load tap-changing transformer, the objective function is to reduce the total amount of voltage change, network loss and on-load tap-changing transformer operation cost in the distribution network as shown in the following formula (1): (1) in, , and represents the cost factor; and Respectively represent photovoltaic and wind power at nodes Time Reactive power output; Indicates branch The resistance of and represents the square of node voltage and line current; The tap of the on-load tap-changing transformer is The state below.

[0025] Specifically, the constraints on physical quantities in the distribution network are set as follows: (2) (3) (4) (5) (6) (7) (8) (9) (10) Among the above constraints, and Respectively represent photovoltaic and wind power at nodes Active power output at; and Respectively represent nodes Active and reactive loads at Indicates that wind power is at the node The capacity of the place; Indicates single-tap adjustable voltage; Indicates the maximum control time of OLTC.

[0026] Among them, formulas (1) and (2) are the power flow constraints of the distribution network, representing the power constraint and voltage constraint respectively. Formulas (3) and (4) are the convex relaxation constraints of the nonlinear power flow using second-order cone programming. Formulas (5) and (6) are the upper and lower bounds of the node voltage and branch current respectively. Formulas (7) and (8) are the reactive power output constraints of the photovoltaic generator set and the wind turbine set respectively. Formula (9) is the voltage constraint of the root node of the distribution network. Formula (10) is the limit on the number of adjustments of the on-load tap changer.

[0027] Among the above constraints, equations (9) and (10) are highly nonlinear and need to be linearized before they can be solved. The constraints of this linearization process are as follows: (11) (12) in, represents the tap state variable; Indicates the maximum adjustable step of the tap; Indicates the number of adjustment steps; Indicates the change of tap position; Indicates the control time of the on-load tap-changing transformer.

[0028] Through the above steps, the nonlinear on-load tap changer tap control problem can be transformed into a mixed integer second-order cone programming problem, so that the commercial solver can be used to obtain the optimal on-load tap changer control information of the on-load tap changer one day in advance.

[0029] Step S2: Use a deep learning model to optimize the reactive power of the distribution network. The deep learning model includes three models, which are used to optimize the reactive power of wind turbines, photovoltaic generators during the day, and photovoltaic generators at night.

[0030] Specifically, the deep learning model fixes the root node voltage information by importing the hourly on-load tap changer tap positions from step S1, and the root node voltage information comes from the constraints of the above formula (8).

[0031] After importing the hourly on-load tap-changing transformer tap position, the deep learning model generates the optimal reactive power output of the photovoltaic unit and the wind power unit based on the hourly on-load tap-changing transformer tap position and the load, photovoltaic and wind power output data of the distribution network, so as to minimize the voltage deviation and network loss at each time point.

[0032] For the three models, their inputs and outputs for reactive power optimization are as follows: Wind turbine model: enter

[0033] Output

[0034] Daytime photovoltaic unit model: enter

[0035] Output

[0036] Nighttime photovoltaic unit model: enter

[0037] Output

[0038] In order to minimize the sum of voltage deviation and network loss at each time point, the objective functions of reactive power optimization of the three deep learning models are set as follows:

[0039] This embodiment also provides a training method for the above deep learning model: First, a training data set is obtained. The training data set includes the load, photovoltaic output, and wind power output of the distribution network.

[0040] The training data set of this embodiment selects the load, photovoltaic and wind power output data of the UK National Grid within 25 days. Specifically, the training data set collects the load, photovoltaic and wind power output data of the UK National Grid with a resolution of 7200 data points every 5 minutes within the 25 days, and performs normalization processing.

[0041] The test dataset is one day of data randomly selected from the same dataset.

[0042] After obtaining the training data set, a deep learning model is used to perform reactive power optimization based on the training data set, and the results of virtual reactive power optimization are scored. The score is used to feed back the deep learning model, thereby training the deep learning model so that the virtual reactive power optimization results of the deep learning model gradually approach the optimal results, thereby achieving the optimal virtual reactive power optimization effect of the deep learning model after training.

[0043] Specifically, the above-mentioned scoring of the results of virtual reactive power optimization includes: Solve the SOCP problem using the training data set and obtain the optimal result.

[0044] The results of virtual reactive power optimization are scored using the optimal result.

[0045] Although the traditional reactive power optimization algorithm based on SOCP has a large amount of calculation and takes a long time, it has excellent accuracy. The method of this embodiment uses the reactive power optimization result of SOCP as the optimal result, and thus as the scoring standard for the results of virtual reactive power optimization. That is, the reactive power optimization ability of the deep learning model can be gradually optimized in the scoring-feedback training process, so that its reactive power optimization results gradually approach the results of the traditional reactive power optimization algorithm.

[0046] This embodiment also provides a specific implementation of the above-mentioned deep learning model, wherein the structure of the model includes an encoder, a decoder and an attention mechanism The hidden layer size of encoder and decoder is 256, and the number of layers of encoder and decoder is 4.

[0047] The number of attention heads of the model is 2, the dropout rate is 0.3, and the activation function is ReLU.

[0048] The model consists of an encoder, a decoder, and an attention mechanism. It uses a global attention mechanism to discover the deep coupling relationship between multiple energy loads and their related data, and then uses this mechanism to provide this information to multiple task-specific decoders.

[0049] The encoder consists of multiple encoding layers, each with a feedforward neural network layer and a multi-head attention layer. The feedforward neural network enhances the nonlinear fitting ability of the model so that it can be applied consistently to each position. For this network, it is two convolutions with a kernel size of 1, which consists of two linear transformations divided by a ReLU activation function.

[0050]

[0051] in, , denote the weights and biases of the encoder FFN respectively.

[0052] In neural networks, the goal of normalization methods is to transform input data into normalized data with a mean of 0 and a variance of 1. In order to prevent the input data from being in the saturation region of the activation function, normalization is usually performed before the activation function. In theory, normalization speeds up the convergence of the model. The proposed model uses layer normalization for both the encoder and the decoder. Unlike batch normalization, layer normalization is independent of the batch size, ensuring that the amount of data used in the normalization calculation is not affected by the batch size.

[0053] For samples , in its flattened form, , and calculate the average accordingly and variance .

[0054]

[0055]

[0056] Each sample is then normalized to achieve zero mean and unit variance, resulting in a normalized sample . This is to ensure numerical stability and prevent the denominator from accidentally becoming zero.

[0057]

[0058] Finally, the output is processed through a feed-forward network layer, where and represents learnable parameters.

[0059]

[0060] The multiple decoding layers that make up the encoder structure are mirrored in the decoder. Each decoding layer is able to utilize the full range of encoded information throughout the decoding process. Unlike traditional dynamic decoding methods used in natural language processing, this decoder directly produces output in a single iteration.

[0061] The attention mechanism mimics the way humans allocate attention. Instead of distributing attention evenly across the entire painting, it emphasizes certain parts. By assigning different weight values ​​to each input element, the attention mechanism allows the model to suppress irrelevant data and prioritize important information.

[0062] Additive attention and dot product (multiplication) attention are two typical attention functions. Dot product attention is consistent with our algorithm and is more efficient in terms of speed and space because it leverages highly optimized matrix multiplication code.

[0063] Dot product attention represents a query and a set of key-value pairs mapped to output as matrices. The calculation method is as follows.

[0064]

[0065]

[0066] in, , and represents the linear transformation matrix. Q, K, and V are matrices generated by the network input.

[0067] There are two attention modules in the decoder. The first module functions similarly to the attention module of the encoder, where Q is obtained by linearly transforming the input, while K and V are derived similarly. Instead, the output of the first multi-head attention module in the encoder and the current decoding layer is the source of the hidden information matrix, which is used for the second attention module.

[0068] Compared to single attention with the same number of parameters, multi-head attention is used in this embodiment to enhance feature extraction capabilities. Multi-head attention applies different linear transformations to Q, K, and V before concatenating the results of multiple single attention actions. Compared to using a single attention head, this approach guarantees more comprehensive information collection by allowing the model to focus on data from many representation subspaces.

[0069]

[0070]

[0071] The deep learning model trained by the above method can, after importing the hourly on-load tap-changing transformer tap position, generate the optimal reactive power output of the photovoltaic unit and the wind power unit according to the hourly on-load tap-changing transformer tap position and the load, photovoltaic and wind power output data of the distribution network, so as to minimize the voltage deviation and network loss at each time point.

[0072] like Figure 2The figure shows a standard IEEE33 power saving system used to verify the effect of the method of this embodiment. The root node of the system is equipped with an OLTC unit, and four photovoltaic units (PV nodes in the figure) and four wind turbines (WD nodes in the figure) are installed along the feeder. The method of this embodiment is applied to it, and the results obtained are shown in Table 1 and Table 2, where Table 1 is a comparison of the total network loss between the method of this embodiment and other methods, and the other methods include traditional pure model-driven algorithms and non-control methods.

[0073] Table 1

[0074] Table 2 is a voltage control effect curve diagram of the method of this embodiment compared with other methods, where the other methods include traditional pure model-driven algorithms and non-control methods.

[0075] Table 2

[0076] Based on the results of the above two tables, it can be seen that the reactive power control method for distribution network based on the hybrid model provided in this embodiment has an excellent control speed and an optimization performance close to that of the traditional algorithm.

[0077] The above-mentioned reactive power control method of the distribution network based on the hybrid model of this embodiment first uses the traditional algorithm for preliminary calculation to ensure the globality of the optimization problem and the satisfaction of physical constraints, and then uses the deep learning algorithm to continue the optimization, providing a rapid response to real-time data and an in-depth understanding of complex relationships. Through the two-stage optimization strategy, it is possible to more effectively cope with the uncertainty of distributed energy resources and loads, achieve stable control of the distribution network voltage, not only improve the optimization performance, but also reduce the dependence on the network model, and improve the flexibility and adaptability of the system.

[0078] An embodiment of the present invention further provides a distribution network reactive power control system based on a hybrid model, comprising: a day-ahead control prediction module, a deep learning model; The day-ahead prediction module is used to predict the hourly on-load tap changer tap positions of the distribution network using a mixed integer second-order cone programming algorithm. The deep learning model is used to optimize the reactive power of the distribution network. According to the hourly on-load tap-changing transformer tap position and the load, photovoltaic and wind power output data of the distribution network, the optimal reactive power output of the photovoltaic unit and the wind power unit is generated to minimize the voltage deviation and network loss at each time point.

[0079] Specifically, in the above system, the deep learning model is trained by the following method: Obtain a training data set, which includes the load, photovoltaic output, and wind power output of the distribution network; Use a deep learning model to perform reactive power optimization based on the training data set, score the results of virtual reactive power optimization, and use the score to feed back the deep learning model to optimize the virtual reactive power optimization effect of the deep learning model.

[0080] As a preferred implementation, the day-ahead control prediction module uses a mixed integer second-order cone programming algorithm to predict the hourly on-load tap changer tap positions of the distribution network, with the goal of reducing the total amount of voltage variation, network loss and OLTC decomposed operation cost.

[0081] Specifically, the structure of the deep learning model includes an encoder and a decoder; The hidden layer size of the encoder and decoder is 256 and the number of layers is 4; The deep learning model includes an attention mechanism, the number of attention heads is 2, and the activation function is the ReLU function.

[0082] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0083] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0084] The above is only an exemplary embodiment of the present disclosure, and the scope of the present disclosure cannot be limited thereto. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the specification and practicing the disclosure here, those skilled in the art will easily think of the implementation scheme of the present disclosure. The present invention is intended to cover any variation, use or adaptive change of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the technical field not recorded in the present disclosure. The description and examples are regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A reactive power control method for distribution network based on hybrid model, characterized in that: The steps include: A mixed integer second-order cone programming algorithm is used to predict the hourly on-load tap changer tap positions in distribution networks. A deep learning model is used to perform reactive power optimization on the distribution network. The deep learning model is used to generate the optimal reactive power output of the photovoltaic unit and the wind power unit according to the tap position of the hourly on-load tap-changing transformer and the load, photovoltaic and wind power output data of the distribution network, so as to minimize the voltage deviation and network loss at each time point.

2. A distribution network reactive power control method based on a hybrid model as claimed in claim 1, characterized in that: The deep learning model is trained by the following method: Acquire a training data set, wherein the training data set includes load, photovoltaic output, and wind power output of the distribution network; The deep learning model is used to perform reactive power optimization according to the training data set, the result of the virtual reactive power optimization is scored, and the score is used to feed back the deep learning model so that the virtual reactive power optimization effect of the deep learning model is optimized.

3. A reactive power control method for distribution network based on hybrid model as claimed in claim 2, characterized in that: The scoring of the result of the virtual reactive power optimization comprises: Solve the SOCP problem using the training data set to obtain an optimal result; The optimal result is used to score the result of the virtual reactive power optimization.

4. A reactive power control method for a distribution network based on a hybrid model as claimed in claim 1, characterized in that: The reactive power optimization includes time scale optimization of discrete power flow constraints, node voltage constraints, branch current constraints, and renewable energy reactive power output constraints.

5. A reactive power control method for distribution network based on hybrid model as claimed in claim 1, characterized in that: The goal of predicting the hourly on-load tap changer tap positions of the distribution network is to reduce the total amount of voltage variation, network losses and on-load tap changer decomposed operating costs.

6. A reactive power control method for distribution network based on hybrid model as claimed in claim 1, characterized in that: The structure of the deep learning model includes an encoder and a decoder; The hidden layer size of the encoder and the decoder is 256, and the number of layers is 4; The deep learning model includes an attention mechanism, the number of attention heads is 2, and the activation function is a ReLU function.

7. A reactive power control system for a distribution network based on a hybrid model, characterized in that: include: Day-ahead control prediction module, deep learning model; The day-ahead prediction module is used to predict the hourly on-load tap changer position of the distribution network using a mixed integer second-order cone programming algorithm; The deep learning model is used to optimize the reactive power of the distribution network, generate the optimal reactive power output of the photovoltaic unit and the wind power unit according to the tap position of the hourly on-load tap-changing transformer and the load, photovoltaic and wind power output data of the distribution network, so as to minimize the voltage deviation and network loss at each time point.

8. A distribution network reactive power control system based on a hybrid model as claimed in claim 7, characterized in that: The deep learning model is trained by the following method: Acquire a training data set, wherein the training data set includes load, photovoltaic output, and wind power output of the distribution network; The deep learning model is used to perform reactive power optimization according to the training data set, the result of the virtual reactive power optimization is scored, and the score is used to feed back the deep learning model so that the virtual reactive power optimization effect of the deep learning model is optimized.

9. A distribution network reactive power control system based on a hybrid model as claimed in claim 7, characterized in that: The day-ahead control prediction module uses a mixed integer second-order cone programming algorithm to predict the hourly on-load tap changer tap positions of the distribution network, with the goal of reducing the total amount of voltage variation, network loss and OLTC decomposed operation cost.

10. A reactive power control system for a distribution network based on a hybrid model as claimed in claim 7, characterized in that: The structure of the deep learning model includes an encoder and a decoder; The hidden layer size of the encoder and the decoder is 256, and the number of layers is 4; The deep learning model includes an attention mechanism, the number of attention heads is 2, and the activation function is a ReLU function.

Images