Novel power system dispatching scene sample data generation method and model construction method

Abstract

The embodiment of the application provides a novel power system dispatching scene sample data generation method and a model construction method, comprising: constructing an extreme weather condition label variable and a power generation power history data sequence; performing normalization processing; inputting the normalized extreme weather condition label variable into a generator of a CGAN network to obtain a generated sample data set; splicing the extreme weather condition label variable and the power generation power history data sequence to form a matrix sequence y, and splicing the extreme weather condition label variable and the generated sample data set to form a matrix sequence h; outputting the matrix sequence y and the matrix sequence h to a discriminator of the CGAN network to output a reality judgment result; training the CGAN network through a target function; and generating a predicted power generation power sequence of each type of unit under extreme weather through the trained CGAN network, thereby improving the safety and stability of the novel power system in response to extreme weather.

Description

Technical Field

[0001] This application relates to the technical field of novel power system dispatching, specifically to methods for generating sample data and building models for novel power system dispatching scenarios. Background Technology

[0002] Building new power systems has become a key supporting technology.

[0003] The new power system is a diversified power source support system with new energy sources as the mainstay. As the proportion of new energy power generation in the power grid continues to increase, clean energy sources such as wind power, photovoltaics, and gas have been gradually incorporated into the main categories of power balance and grid stability control. However, traditional power generation forms such as thermal power and hydropower still play a dominant role in the power grid due to their significant voltage and frequency control characteristics. Therefore, the extreme weather dispatch of the new power system must not only consider the operation of traditional thermal power and hydropower units under extreme weather conditions, but also the operation of clean energy sources such as wind, solar, and gas under extreme weather conditions. In addition, unlike traditional thermal power generation, wind power and photovoltaic power generation, as the main forms of new energy power generation, are highly susceptible to the influence of external meteorological conditions, geographical environment, etc., and the power generation exhibits intermittent and fluctuating characteristics, making the new power system exhibit stronger "weather coupling" and "system vulnerability". Therefore, the power system with new energy sources as the mainstay will be more susceptible to the impact of extreme weather.

[0004] Currently, the general design standards for my country's power system are based on the traditional power supply security and investment economy of the power system. They cannot meet the development requirements of the new power system, which features "high proportion of new energy sources, high frequency of extreme weather, and high cost of power outage losses". They do not adequately consider the operating characteristics of new energy sources under extreme weather conditions, and the support and synergistic effects of traditional units are not adequately anticipated. This results in significant risks to the emergency safety operation level of the new power system. Therefore, it is crucial to study the prediction and control of dispatch scenarios for multiple power sources under extreme weather conditions in the new power system.

[0005] Currently, most existing scene generation and decision-making methods are based on probabilistic models. They typically employ methods such as Markov chain method, time series method, and scene tree generation method to establish a probabilistic model that follows the pattern of scene change based on historical data. These methods mainly generate new scenes for analysis through sampling and require a large amount of historical data as samples to support decision-making. At the same time, the biggest drawback of this method is that it is difficult to apply to the model description of multi-type feature scenes. It is only suitable for the probability prediction and decision-making of a certain feature and is not suitable for the generation and decision-making of complex and diverse feature scenes.

[0006] In recent years, deep learning models have been increasingly applied to prediction and decision-making in new energy power generation scenarios, achieving good results in addressing the randomness and intermittency of new energy output. However, prediction and decision-making in these application scenarios still rely on sufficient data samples and reasonable deep learning models. Furthermore, there are still three main problems in the prediction and control of new power systems under extreme weather conditions:

[0007] (i) The power grid operation data samples are scarce, making it difficult to meet the data-driven requirements of deep learning models;

[0008] (ii) The dispatching scenario under weather conditions not only targets new energy power generation, but also includes predictions of various forms of energy such as hydropower, thermal power, gas and other major energy sources of regional power grids, as well as energy exchange between power grids. Relying solely on deep learning networks to learn and predict multiple types of features will affect the efficiency of the training and learning process and the accuracy of the prediction.

[0009] (III) Scene control under extreme weather conditions needs to be judged based on the changes of multiple types of characteristic quantities, rather than relying on the changes of one or a few characteristic quantities. Therefore, it is necessary to select an effective identification and judgment mechanism for control. Summary of the Invention

[0010] To address one of the aforementioned technical deficiencies, this application provides a novel method for generating sample data and a model construction method for power system dispatching scenarios.

[0011] According to a first aspect of the embodiments of this application, a novel method for generating sample data for power system dispatching scenarios is provided, including:

[0012] S11, construct the extreme weather condition label variable c, and the historical power generation data sequence P of each type of unit corresponding to the extreme weather condition label variable c;

[0013] S12, normalize the extreme weather condition label variable c and the historical power generation data sequence P;

[0014] S13, input the normalized extreme weather condition label variable c* into the generator of the CGAN network to obtain the generated sample dataset p. new ;

[0015] S14, concatenate the extreme weather condition label variable c* and the historical power generation data sequence P* to form a matrix sequence y, and then combine the extreme weather condition label variable c* and the generated sample dataset p. new Concatenate the components to form a matrix sequence h;

[0016] S15, output the matrix sequence y and matrix sequence h to the discriminator of the CGAN network and output the authenticity judgment result;

[0017] S16, Train the CGAN network using the objective function to obtain the trained CGAN network;

[0018] S17, using a trained CGAN network to generate predicted power generation sequences for various types of generating units under extreme weather conditions;

[0019] The predicted power generation sequence for each type of generating unit is expressed as follows:

[0020] N(P T )=N(P PV P WD P H P WH P GH );

[0021] Where, p T P represents the transpose of the generated predicted power generation sequence; PV P WD P H P WH P GH These represent the photovoltaic power generation sequence, wind power generation sequence, thermal power generation sequence, hydropower generation sequence, and gas power generation sequence generated by the CGAN network, respectively.

[0022] According to a second aspect of the embodiments of this application, a novel method for constructing a control decision model for power system dispatching scenarios is provided, including:

[0023] S21, Based on the CGAN network, generate the predicted power generation of various types of generating units, and construct the state-space variable s based on the predicted load and frequency of the regional new power system. t ;

[0024] The predicted power generation of each type of unit is generated using the novel power system dispatch scenario sample data generation method described in any one of claims 1 to 4.

[0025] S22, establish execution strategies based on the standby unit's operational capacity, emergency load shedding amount, and external grid interconnection transmission power; whereby each execution strategy corresponds to a specific action space a. t Based on action space a t Construct action set A;

[0026] S23, change the initial state space variable s t Inputting the DQN network into the training Q-network generates the value Q(s) corresponding to the initial state. t a t ,θ); where θ represents the neural network weights used to train the Q-network;

[0027] S24, selecting the corresponding action space a from the action set A by using the ε-greedy strategy function t ;

[0028] S25, according to the action space a t executing the control flow according to the corresponding execution policy;

[0029] S26, for the action space a t calculating the immediate reward r according to the reward function, and using Matpower to calculate the state space quantity s at the next moment by power flow calculation t+1 ;

[0030] S27, constructing an experience dataset (s t , a t , r t , s t+1 ); randomly sampling d groups of experience datasets from the experience pool, inputting them into the training Q network and target Q network of the DQN network, and calculating the value of the training Q network and the target value of the target Q network;

[0031] wherein: d < dm, dm is the data capacity of the experience pool;

[0032] S28, updating the θ value of the training Q network according to the loss function and the target value of the target Q network;

[0033] S29, determining whether a preset number of iterations is reached; if yes, generating a control decision model based on the DQN network; otherwise, performing step S24.

[0034] The embodiments of the present application have the following technical effects by adopting the above technical solutions:

[0035] 1. In the present application, the conditional generative adversarial network (CGAN network) is applied to the generation of scenario sample data for new-type power system dispatch under extreme weather conditions. In addition to considering the characteristics of new energy power generation, traditional power generation types such as thermal power and hydropower are fully considered, which can deeply explore the supporting role of traditional units in new-type power systems, so that the generated samples conform to actual situations and improve the availability of generated samples.

[0036] 2. In the present application, the traditional objective function of the conditional generative adversarial network is optimized and improved. According to the characteristics of traditional hydropower, thermal power generation and new energy power generation, a divergence objective function is proposed to constrain the prediction distribution of new energy power generation, which improves the reliability of data generation for new-type power system dispatch scenarios under extreme weather.

[0037] 3. In this application, a novel power system dispatching scenario control decision model under extreme weather conditions is established using a deep reinforcement learning network (CGAN network). It can realize feature self-extraction, nonlinear system training, and complex dispatching scenario control. Compared with traditional probabilistic analysis and decision-making methods, it can adapt to different scenario learning and decision-making.

[0038] 4. The implementation strategy in this application can scientifically and rationally formulate control strategies for new power systems under extreme weather conditions, avoid relying on a single random selection strategy that threatens the stable operation of the power grid, and improve the safety and stability of new power systems in response to extreme weather. Attached Figure Description

[0039] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0040] Figure 1 A flowchart illustrating the novel power system dispatch scenario sample data generation method provided in this application embodiment;

[0041] Figure 2 This is a schematic diagram of the data processing flow of the generator in the embodiments of this application;

[0042] Figure 3 This is a schematic diagram of the batch processing flow in the embodiments of this application;

[0043] Figure 4 This is a schematic diagram of the data processing flow of the discriminator in the embodiments of this application;

[0044] Figure 5 A flowchart illustrating the method for constructing a novel power system dispatch scenario control decision model provided in this application embodiment;

[0045] Figure 6 This is a schematic diagram illustrating the execution process of the extreme weather dispatching scenario execution strategy in the embodiments of this application. Detailed Implementation

[0046] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0047] To address the urgent needs and shortcomings of the main methods in the background technology, this application utilizes the sample data augmentation advantage of Conditional Generative Adversarial Networks (CGAN) to generate small sample data under extreme weather conditions, constructing a multi-sample dataset for various power dispatching scenarios under extreme weather conditions; and leverages the strong learning and decision-making capabilities of Deep Reinforcement Learning (DQN) to construct a control decision model for dispatching scenarios under extreme weather conditions.

[0048] Example 1

[0049] like Figure 1 As shown in the embodiments of this application, the novel power system dispatch scenario sample data generation method includes:

[0050] S11, construct the extreme weather condition label variable c, and the historical power generation data sequence P of each type of unit corresponding to the extreme weather condition label variable c;

[0051] S12, normalize the extreme weather condition label variable c and the historical power generation data sequence P;

[0052] S13, input the normalized extreme weather condition label variable c* into the generator of the CGAN network to obtain the generated sample dataset p. new ;

[0053] S14, concatenate the extreme weather condition label variable c* and the historical power generation data sequence P* to form a matrix sequence y, and then combine the extreme weather condition label variable c* and the generated sample dataset p. new Concatenate the components to form a matrix sequence h;

[0054] S15, output the matrix sequence y and matrix sequence h to the discriminator of the CGAN network and output the authenticity judgment result;

[0055] S16, Train the CGAN network using the objective function to obtain the trained CGAN network;

[0056] S17, using a trained CGAN network to generate predicted power generation sequences for various types of generating units under extreme weather conditions;

[0057] The predicted power generation sequence for each type of generating unit is expressed as follows:

[0058] N(P T )=N(P PV P WD P H P WH P GH );

[0059] Where, p TP represents the transpose of the generated predicted power generation sequence; PV P WD P H P WH P GH These represent the photovoltaic power generation sequence, wind power generation sequence, thermal power generation sequence, hydropower generation sequence, and gas power generation sequence generated by the CGAN network, respectively.

[0060] Specifically, extreme weather events in different regions include high temperatures, low temperatures, sandstorms, smog, and ice storms. Among these, sandstorms and smog primarily affect solar irradiance; extremely low irradiance during extreme weather can lead to significantly reduced or even stopped solar power output. Meanwhile, high temperatures and cold waves primarily affect solar power through temperature. Therefore, the main impacts of extreme weather on solar power generation are solar irradiance and temperature. Let solar irradiance be W and temperature be T, then the extreme weather conditions corresponding to solar power generation are:

[0061] c PV ={(W1, T1), (W2, T2), ..., (W i T i ), ... (W n T n )};

[0062] Where: c PV For the description of extreme weather conditions for photovoltaic power generation, W i T i Let represent the irradiance and temperature corresponding to the i-th element of the extreme weather condition sequence; n is the number of sampling points.

[0063] Therefore, the historical data sequence of photovoltaic power generation corresponding to extreme weather is as follows:

[0064] P PV ={P PV.1 P PV.2 ..., P PV.i ..., P PV.n};

[0065] Where: P PV For historical power generation data series, P PV，i Let be the photovoltaic power generation corresponding to the i-th sampling point.

[0066] Furthermore, the factors affecting wind power output due to extreme weather are wind speed and temperature; let the wind speed be F and the temperature be T, then the extreme weather conditions corresponding to wind power are:

[0067] c WD ={(F1, T1), (F2, T2), ..., (F i Ti ), ……(F n T n )}

[0068] Where: c WD For the description of extreme weather conditions for wind power generation, F i T i Let i be the wind speed and temperature corresponding to the i-th element in the extreme weather condition sequence;

[0069] Therefore, the historical data sequence of wind power generation corresponding to extreme weather conditions is as follows:

[0070] P WD ={P WD.1 P WD.2 ..., P WD.i ..., P WD.n};

[0071] Where: P WD For historical data series of wind power generation, P WD.i Let be the wind power generation capacity corresponding to the i-th sampling point.

[0072] Furthermore, for thermal power units, sustained high temperatures can lead to significant power output reductions due to condensate circulation water temperature, while extremely cold weather can cause ice storms, disrupting the operation of coal-fired units. Therefore, the influencing factor of extreme weather on thermal power units is temperature T. Thus, the extreme weather conditions for thermal power units are:

[0073] c H ={T1, T2, ..., T i , ...T n};

[0074] Among them, c H For the description of extreme weather conditions for thermal power units, T i Let be the temperature corresponding to the i-th element in the extreme weather condition sequence;

[0075] Therefore, the historical data sequence of power generation of thermal power units corresponding to extreme weather is as follows:

[0076] P H ={P H.1 P H.2 ..., P H.i ..., P H.n};

[0077] Where: P H P is a historical data sequence of power generation from thermal power units. H，i Let be the power generation capacity of the thermal power unit corresponding to the i-th sampling point.

[0078] Furthermore, for hydropower units, both high and low temperatures can easily cause water level changes. Simultaneously, precipitation also affects the output of the hydropower unit. Therefore, the influencing factors for hydropower units are temperature T and precipitation ρ. Thus, the extreme weather conditions for hydropower units are:

[0079] c WH ={(T1, ρ1), (T2, ρ2),..., (T i , ρ i ), ..., (T n , ρ n )};

[0080] Among them, c WH For the label description of extreme weather conditions for hydropower units, T i ρ i Let i be the temperature and rainfall corresponding to the i-th element of the extreme weather condition sequence.

[0081] Therefore, the historical data sequence of power generation of hydropower units corresponding to extreme weather is as follows:

[0082] P WH ={P WH.1 P WH.2 ..., P WH.i ..., P WH.n};

[0083] Where: P WH P is a historical data sequence of power generation from thermal power units. WH，i Let be the power generation capacity of the hydropower unit corresponding to the i-th sampling point.

[0084] Furthermore, the power generation of gas-fired power units is closely related to the concentration of methane gas. Methane concentration is easily affected by high temperatures and ground pressure. Extreme weather conditions causing sudden changes in temperature and ground pressure can affect methane concentration, leading to unit shutdown. Therefore, the influencing factors for methane power generation are temperature T and pressure Q. Thus, the extreme weather conditions for methane-fired power units are:

[0085] c GH ={(T1, Q1), (T2, Q2), ..., (T i Q i ), ..., (T n Q n )};

[0086] Among them, c GH For the description of extreme weather conditions for gas generator units, T i Q i Let i be the temperature and air pressure corresponding to the i-th element of the extreme weather condition sequence.

[0087] Therefore, the historical data sequence of power generation of gas generating units under extreme weather conditions is as follows:

[0088] P GH ={P GH.1 P GH.2 ..., P GH.i ..., P GH.n};

[0089] Among them, P GH For the historical power generation data sequence of gas generating units, P GH，i Let be the power generation capacity of the gas generator unit corresponding to the i-th sampling point.

[0090] In summary, the extreme weather conditions for wind, solar, hydro, thermal, and gas power generation are five factors: temperature T, solar irradiance W, wind speed F, rainfall ρ, and air pressure Q.

[0091] In this embodiment, the expression for the extreme weather condition label variable c is:

[0092] c = {c0, c1, ..., c} t , ..., c n};

[0093] Among them, c t This represents the condition variable element at time t, including: temperature T. t Irradiance W t Wind speed F t Rainfall ρ t and air pressure value Q t ;

[0094] c t The expression is: c t =[T t W t F t , ρ t W t ];

[0095] The expression for the historical power generation data sequence P of each type of generating unit is:

[0096] P = {P0, P1, ..., P} t ..., P n};

[0097] Among them, P t This represents a sample element representing the actual power generation at time t, including: photovoltaic power generation P. PV.t Wind power generation capacity P WD.t Thermal power generation capacity P H.t Hydropower generation capacity P WH.t and gas power generation PGH.t ;

[0098] P t The expression is: p t =[P PV.t P WD.t P H.t P WH.t P GH.t ].

[0099] After constructing the extreme weather condition label variable c and the historical power generation data sequence P of each type of generating unit, the output data samples of each type of generating unit in the new power system under extreme weather conditions can be generated through the CGAN network; including:

[0100] S12, normalize the extreme weather condition label variable c and the historical power generation data sequence P; in this embodiment, the normalization process is as follows:

[0101]

[0102]

[0103] Among them, c min c max P represents the minimum and maximum values ​​in the sequence of matrix c; min and P max The minimum and maximum values ​​in the p matrix sequence are given. Each element in the extreme weather condition label variable c and the historical power generation data sequence P is normalized to obtain normalized matrices c* and p*.

[0104] In this embodiment, step S13 involves inputting the normalized extreme weather condition label variable c* into the generator of the CGAN network to obtain the generated sample dataset p. new ;include:

[0105] S131, the normalized extreme weather condition label variable c* is concatenated with the Gaussian white noise matrix z to form the input vector matrix w;

[0106] Among them, the extreme weather condition label variable c* is an n×5 matrix;

[0107] Specifically, the expression for the Gaussian white noise matrix z is:

[0108] z = {z0, z1, ..., z} t , ..., z n};

[0109] The expression for concatenating the weather condition label variable c* with the Gaussian white noise matrix z using the concatenation function is as follows:

[0110] w = cbind(c * ,z);

[0111] The expression for the concatenated input vector matrix w is:

[0112] w = {w0, w1, ..., w} t , ..., w n};

[0113] Among them, w t =[c * t , z t ] T ;

[0114] The Gaussian white noise matrix z is a Gaussian white noise vector matrix of equal length sampled from a Gaussian distribution with a mean of 0 and a standard deviation of 1. t Let be the element at time t; w is a 2n×5 matrix, which is the input matrix obtained by vertically concatenating the conditional label and the Gaussian noise vector z. t Let be the element at time t.

[0115] S132, the input vector matrix w is input into the input layer of the CGAN network generator. After processing by the generator's deconvolution layer and output layer, the generated sample dataset p is obtained. new .

[0116] The deconvolutional layer includes deconvolution computation, batch normalization, and ReLU function activation; the output layer includes fully connected operations and Tanh function activation.

[0117] like Figure 2 As shown, in this embodiment, the number of neurons in the input layer of the CGAN network generator can be set to 5; the parameters of the deconvolution layer are set as follows: the number of deconvolution layers is set to 4, the kernel size K (which can be set to 4×4 in this embodiment), the calculation stride S (which is set to 2 in this embodiment), and the padding parameter P can be set according to the number of sampling points; then the transformed input vector after calculation by the deconvolution function is expressed as:

[0118] o a =S*(w-1)-2P+K;

[0119] Among them, o a Let be the output feature vector sequence of the a-th layer, where a = 1, 2, 3, 4.

[0120] like Figure 3 As shown, after deconvolution, the feature vector sequence o a Perform batch standardization:

[0121] Calculation of the mean of the eigenvector sequence:

[0122] Among them, o a，i o a The i-th element of the feature sequence, μ a The mean of the eigenvector sequence;

[0123] Variance calculation of the eigenvector sequence:

[0124] Among them, e a The variance of the eigenvector sequence;

[0125] standardization:

[0126] in, The sequence is the result of normalization of the feature sequence, and ε is the sequence normalization adjustment parameter, which can be set according to the sequence normalization situation, or it can be set to 0.

[0127] Train parameters γ and β, and perform batch standardization:

[0128]

[0129] Where BN is the batch normalization function, γ is the expansion parameter in batch normalization, and β is the shift parameter in batch normalization. This is the batch-standardized feature sequence.

[0130] After batch standardization, the activation function performs the activation operation. The expression for the ReLU function is:

[0131]

[0132] in, for The i-th element in the sequence.

[0133] Then, the features extracted by the convolutional layers are synthesized through a fully connected layer at the output layer, thus connecting all the features; its expression is:

[0134]

[0135] Among them, b n Let α represent the bias matrix. k Represents the weight matrix, o k This represents the extraction value of the k-th feature;

[0136] Then, the generated sample dataset is output after being processed by the Tanh activation function; the expression of the Tanh activation function is:

[0137]

[0138] In this embodiment, S14, the extreme weather condition label variable c* and the historical power generation data sequence p* are concatenated to form a matrix sequence y, and the extreme weather condition label variable c* and the generated sample dataset p* are then combined. new Concatenate the components to form a matrix sequence h; including:

[0139] y = cbind(c * p * );

[0140] y = {y0, y1, ..., y t , ..., y n};

[0141] Where cbind is the matrix row concatenation function, y t Let y be the t-th element in the matrix sequence y. t The expression is:

[0142] y t =[c t * p t * ] T ;

[0143] Output data p to the generator new Normalization is performed:

[0144]

[0145] Where, p min.new and p max.new For matrix sequence p new The minimum and maximum values ​​in p new Each element p in t.new Normalization is performed to obtain the normalized matrix p. new * .

[0146] The normalized generated data matrix p new * By concatenating the columns of the extreme weather condition label variable c*, we obtain the matrix sequence h:

[0147] h = cbind(c * p new * );

[0148] h = {h0, h1, ..., h} t , ..., h n};

[0149] Among them, ht Let h be the t-th element in the matrix sequence h. t The expression is:

[0150] h t =[h t * p t.new * ] T .

[0151] In this embodiment, the discriminator of the CGAN network is used to analyze h. t and y t The process involves making a judgment: determining the authenticity of data from generated (fake) and real (true) data. Based on game theory principles, the system trains the system to judge the authenticity of scenario data. The stronger the discriminator, the easier it is for the generator's data to be judged as fake. Therefore, the generator needs to improve its ability to generate data that more closely approximates the probability distribution of real data. As the generator's imitation ability increases, the generated data becomes harder for the discriminator to distinguish between real and fake. Thus, the discriminator will improve its judgment ability through training, eventually reaching an equilibrium state where the generated samples perfectly match reality.

[0152] like Figure 4 As shown, in step S15, the matrix sequence y and the matrix sequence h are output to the discriminator of the CGAN network to output the authenticity judgment result; wherein: the input sequence of the discriminator is the matrix sequence y and the matrix sequence h, which are processed by the input layer, the convolutional layer (convolution, batch normalization, LeakyReLU activation function), and the output layer (convolution, LeakyReLU activation function) to obtain the discrimination result.

[0153] In this embodiment, the parameter settings of the discriminator's convolutional layer are similar to those of the generator's deconvolutional layer: the number of layers can be set to 4, the convolutional kernel size K (set to 4×4 in this embodiment), the calculation stride S (set to 2 in this embodiment), and the padding parameter P can be set according to the number of sampling points; then the input vector transformation after the convolution is calculated by the convolution calculation function is expressed as:

[0154] m a = (h + 2P - K) / S + 1;

[0155] Where, m a Let be the output feature vector sequence of the a-th layer, where a = 1, 2, 3, 4.

[0156] Then, batch standardization is performed using the training parameters γ and β:

[0157]

[0158] Where BN is the batch normalization function, γ is the expansion parameter in batch normalization, and β is the shift parameter in batch normalization. The feature sequence after batch standardization; This is the sequence after the feature sequence has been standardized.

[0159] Then, it is activated using the LeakyReLU activation function, the expression of which is:

[0160]

[0161] in, for The i-th element in the sequence.

[0162] In this embodiment, the output layer of the discriminator does not have a fully connected layer. Instead, a sigmoid function is added after the convolutional layer to construct a binary classification network, which determines the probability of a true classification result.

[0163]

[0164] Where m represents the final feature extraction result of the convolution calculation.

[0165] In this embodiment, S16, the CGAN network is trained using an objective function to obtain a trained CGAN network, wherein the expression of the objective function is:

[0166]

[0167] in:

[0168]

[0169] p s =p PV +p WH +p GH ;

[0170]

[0171]

[0172] In the above formula, E represents the expected value, D(p) represents the output data of the discriminator, G(z) represents the generated data sequence of the generator; p ~ Date(p) represents the probability distribution of the learned true power sequence; z ~ GN(z) represents the probability distribution of Gaussian noise data input into the generator;

[0173] N(.) represents the generative output model of the adversarial generative network, and p ~ N(p) represents the probability distribution of the generated power sequence of the adversarial generative network model; N(p PVN(p) represents the photovoltaic power generation sequence in the CGAN network production sequence; s This represents the power sequence of the sum of the generating capacities of thermal power, hydropower, and gas power in the CGAN network production sequence; The difference in probability distribution between the power generation sequences of photovoltaic power and those of thermal power, hydropower, and gas power is represented by the divergence; p ~ Date(p) represents the probability distribution of the learned true power sequence. The power generation sequence p represents the sum of the actual photovoltaic power generation sequence PPV and the power generation sequence of thermal power, hydropower, and gas power. s The divergence.

[0174] Since traditional CGAN networks only consider the adversarial and balanced training objectives of the generator and discriminator, this application considers the power prediction of various types of power generation in new power systems. Therefore, the spatial distribution probability of the generated sample sequences should meet the actual needs of the regional power grid.

[0175] Among them, thermal power generation is only affected by high temperatures under extreme weather conditions, and its power generation is relatively stable. Wind, solar, hydro, gas and other factors are also affected by other important factors. Therefore, the power generation sequence of thermal power units is relatively stable, and thermal power units play a leading role in the power grid and are irreplaceable in terms of grid voltage and frequency regulation.

[0176] In addition, hydropower and gas-fired power units can play a certain role in voltage and frequency control under extreme weather conditions to compensate for the insufficient power generation of thermal power units and maintain the safe and stable operation of the power grid. This type of power generation is generally a traditional synchronous generator unit, whose energy can be stored in large quantities, and whose power generation is controllable and relatively stable, and has certain stabilizing factors under extreme weather conditions. Therefore, the distribution training guidance is based on the power generation sequences of thermal power units, hydropower units, and gas-fired power units, as well as the KL divergence of wind power and photovoltaic power, which are two types of power generation with strong randomness and volatility. That is, the spatial correlation distribution of wind power and photovoltaic power output and thermal power generation should conform to the distribution characteristics of the true sequence.

[0177] like Figure 5 As shown in the embodiments of this application, a novel method for constructing a control decision model for power system dispatching scenarios is also provided, including:

[0178] S21, Based on the CGAN network, generate the predicted power generation of various types of generating units, and construct the state-space variable s based on the predicted load and frequency of the regional new power system. t ;

[0179] The predicted power generation of each type of unit is generated using the novel power system dispatch scenario sample data generation method described in any one of claims 1 to 4.

[0180] S22, establishing an execution strategy according to the input capacity of standby units, the emergency load shedding amount and the transmission power of external grid tie-lines; wherein each execution strategy corresponds to a corresponding action space a t ; based on the action space a t construct an action set A;

[0181] S23, inputting the initial state space variable s t into the training Q-network of the DQN network, and generating the value Q(s t , a t , θ) corresponding to the initial state; wherein θ represents the neural network weight of the training Q-network;

[0182] S24, selecting the corresponding action space a from the action set A by using an ε-greedy strategy function t ;

[0183] S25, executing the control process according to the execution strategy corresponding to the action space a t ;

[0184] S26, the action space a t calculates the immediate reward r according to the reward function, and calculates the next-time state space variable s by using Matpower power flow calculation t+1 ;

[0185] S27, constructing an experience data set (s t , a t , r t , s t+1 ); randomly sampling d groups of experience data sets from the experience pool, inputting them into the training Q-network and the target Q-network of the DQN network, and calculating the value of the training Q-network and the target value of the target Q-network;

[0186] wherein: d<dm, dm is the data capacity of the experience pool;

[0187] S28, updating the θ value of the training Q-network according to the loss function and the target value of the target Q-network;

[0188] S29, judging whether a preset number of iterations is reached, if yes, generating a control decision model based on the DQN network; otherwise, executing step S24.

[0189] Specifically, the expression of the state space variable s t is:

[0190] s t ={P HN.t , P WHN.t , P GHN.t , P WDN.t , P PVN.t , PL.t P B.t P D1.t ..., P Dk.t , Δf t};

[0191] In the formula, P HN.t P WHN.t P GHN.t P WDN.t P PVN.t Δt represents the power generation of thermal power, hydropower, gas power, wind power, and photovoltaic power at time t, respectively; Δf represents the frequency deviation of the regional power grid at time t.

[0192] When the output of generating units is limited under extreme weather conditions in the regional power grid, the region relies on standby generating units and external grid power support for regional power supply; P L.t P represents the predicted load of the regional power grid at time t. B.t P represents the standby generating capacity of the region at time t; D1.t ... P Dk.t These represent the power flow at important cross-sections of the connecting line, where k is the number of power flows at important cross-sections.

[0193] Based on the above state-space variables s t Information is used to implement safe and stable control of a new type of power system via a DQN network. Under extreme weather conditions, the control of this new power system is a short-term transient control, and the control cycle is divided into n periods.

[0194] The action space a t The expression is:

[0195] a t =(ΔP) B.t ΔP L.t ΔP D1.t , ..., ΔP Dk.t );

[0196] Where: ΔP B.t ΔP represents the reserve capacity deployed after the system makes a decision at time t. L.t ΔP represents the control variable for the load after the system decision at time t. D1.t , ..., ΔP Dk.t This indicates that the power of the k tie-line sections increases after the system makes a decision.

[0197] In the established implementation strategy, the reserve capacity should meet the energy balance requirements of the regional new power system:

[0198] ΔP L.t+1 =ΔP L.t +K S (Δf t+1 -Δf t );

[0199] ΔP B.t+1 =ΔP B.t +ΔP L.t+1 ;

[0200] K S =K R *K G +K L ;

[0201] First, if increasing the reserve capacity at time t cannot meet the increased load supply, additional reserve capacity will be issued at time t+1; where ΔP L.t+1 K represents the load control quantity at time t+1. S K is the unit regulating power of the system. G K represents the unit regulating power of the system's generator set. L ΔP is the unit regulating power of the system load. B.t+1 K represents the reserve capacity control value at time t+1. R The system reserve factor represents the ratio of the system's installed capacity to its load.

[0202] Where: load control variable ΔP L.t The calculation expression is:

[0203] ΔP L.t =P L.t -P L.t-1 ;

[0204] P L.t and P L.t-1 These are the points at time t-1 and time t on the system load power prediction curve.

[0205] Secondly, the allocation of reserve capacity should meet the limitations of the regional system's reserve capacity:

[0206]

[0207] Within n cycles, when the reserve capacity reaches its upper limit P B.max At that time, consider using tie lines to support power transmission from external regional power grids:

[0208]

[0209] Where: ΔP W To support switching power for external networks, K W For external network unit adjustment power, ΔP O This is the power control quantity for the external network.

[0210] When ΔP W When the value is greater than 0, the external power grid can provide power support to the system and participate in frequency regulation;

[0211] When ΔP W =-ΔP L.t At that time, the external power grid was just enough to provide power support for the system;

[0212] In this formula: the sum of the transmission power of each section is the external grid-supported transmission power;

[0213] Transport flow restrictions at each cross-section: P Dk.min ≤P Dk.t ≤P Dk.max In this formula: P Dk.t P represents the power flow at the k-th connection section at time t; Dk.min and P Dk.max This is the limit value for the power flow at the k-th connecting line section.

[0214] When ΔP W When the value is less than 0, or when the external power grid support power cannot meet the system power change due to the power transmission capacity limitation of the cross-section, the load reduction of the system should be considered to balance the power and stabilize the grid.

[0215] ΔP L.t+1 =K L (Δf t+1 -Δf t );

[0216] At this point, grid stability is maintained solely through system load control; therefore, only the load regulation power KL is considered to reflect the load control quantity. Simultaneously, system load control must be executed according to the emergency control load sequence table established by the system grid.

[0217]

[0218] Among them, P SL.i The power supply sequence of the load users in the load control sequence table is m users in total;

[0219] The process of executing a strategy, such as Figure 6 As shown.

[0220] In this embodiment, the deep reinforcement learning network (DQN network) can be configured as follows:

[0221] The input layer consists of state-space variables s t The number of layers is set to s t The number of elements in the middle is 7+k layers;

[0222] The output layer is the value Q(s) of each action in this state. t a i ,θ), where i is each action of the output layer, i = 2 + k;

[0223] The middle layer is the hidden layer, and the number of layers can be set according to the specific training sample data.

[0224] The DQN network contains two identical neural networks, with the parameters of the input layer, hidden layer, and output layer set according to the parameters mentioned above; one neural network is the training Q-network, and the other is the current Q-network.

[0225] In this embodiment, the main objective of dispatching scenario control under extreme weather conditions is system stability and to meet load supply requirements as much as possible. System stability is mainly reflected in frequency stability deviation control, namely:

[0226]

[0227] When the system frequency is within the stable frequency range, no penalty term is added; when the system frequency exceeds the limit, a penalty term d is added.

[0228] When controlling and supplying power to the system load, the importance of the load and the cost of control should be considered:

[0229]

[0230] Controlling costs involves controlling both lost electricity and the cost of generating electricity from the generating units.

[0231]

[0232] In the formula, r1 represents the reward function based on frequency stability deviation, and d represents the penalty term when the frequency exceeds the limit;

[0233] r2 represents the reward function based on load importance, m1 represents primary load, m2 represents secondary load, m3 represents tertiary load, and k1, k2, and k3 represent load weight parameters, where k1 > k2 > k3. The more important the lost load, the greater the penalty. SL.i This indicates the amount of data used by the user to control the load.

[0234] r3 represents the reward function based on cost control; β represents the penalty coefficient of the cost control item; Δt represents a single control period; c1 represents the unit electricity cost of the load; c2 represents the converted cost of the regional reserve capacity per unit of power generation; c3 represents the converted cost of the tie line per unit of power supply; P B.t P represents the reserve capacity put into operation in the t-th cycle, for a total of n cycles; Dj.t This represents the power generation of the j-th tie line in the t-th cycle of the system, with a total of k tie lines.

[0235] Since the reward function should minimize the loss cost when making intelligent decisions, the expression of the reward function in this application is: r = -(r1 + r2 + r3).

[0236] In step S28, the expression for the loss function is:

[0237]

[0238] In the formula: j represents the current iteration number; y j Let θ represent the target value calculated by the Q-network for the j-th iteration. j θ represents the parameter θ for the j-th iteration; T represents the preset number of iterations.

[0239] In this embodiment, the expression for calculating the target value of the target Q-network is:

[0240]

[0241] In the formula: r j Let γ represent the immediate reward during the j-th iteration, and let γ represent the discount factor.

[0242] In this embodiment, step S24 involves using an ε-greedy policy function to select the corresponding action space a from the action set A. t In the above, the expression for the ε-greedy policy function is:

[0243]

[0244] Where π(a|s) is the ε-greedy policy function, representing the probability of choosing action a when it is the optimal choice. Otherwise, the probability of choosing is

[0245] ε is a very small positive number, and ε < 1, representing the probability of random selection;

[0246] arg max Q(a, s) represents the maximum Q value corresponding to the optimal action a; |A| represents the action set;

[0247] This strategy states that when action a is the optimal choice, its selection probability is 1 - ε - ε / |A|, otherwise the selection probability is ε / |A|.

[0248] This application also provides an electronic device, including:

[0249] Memory; processor; and computer programs;

[0250] The computer program is stored in the memory and configured to be executed by the processor to implement the method described above.

[0251] This application also provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement the method described above.

[0252] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0253] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0254] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0255] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0256] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0257] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for generating sample data for a novel power system dispatching scenario, characterized in that, Comprising: S11, constructing an extreme weather condition label variable c, and a historical power generation data sequence P of each type of unit corresponding to the extreme weather condition label variable c; S12, performing normalization processing on the extreme weather condition label variable c and the historical power generation data sequence P; S13, the normalized extreme weather condition label variables The data is input into the generator of the CGAN network to obtain the generated sample dataset. ; S14, labeling extreme weather conditions variables and normalized historical power generation data sequence The data is concatenated to form a matrix sequence y, which includes the label variables for extreme weather conditions. and generate sample datasets Concatenate the components to form a matrix sequence h; S15, inputting the matrix sequence y and the matrix sequence h into a discriminator of a CGAN network, and outputting an authenticity judgment result; S16, training the CGAN network through an objective function to obtain a trained CGAN network; S17, generating a predicted power generation sequence of each type of unit under extreme weather through the trained CGAN network; The expression of the predicted power generation sequence of each type of unit is: ； in, This represents the transpose of the generated predicted power generation sequence; , , , , These represent the photovoltaic power generation sequence, wind power generation sequence, thermal power generation sequence, hydropower generation sequence, and gas power generation sequence generated by the CGAN network, respectively. The expression of the extreme weather condition label variable c is: ； in, This represents the condition variable elements at time t, including: temperature. Irradiance Wind speed Rainfall and air pressure value ; The expression is: ; The expression of the historical power generation data sequence P of each type of unit is: ； in, This represents a sample element representing the actual power generation at time t, including: photovoltaic power generation. Wind power generation capacity Thermal power generation capacity Hydropower generation capacity and gas power generation capacity ; The expression is: ; In said S16, the expression of the objective function is: ； Wherein: ； ； ； ； In the above formula, E represents the expected value. This represents the output data of the discriminator. This represents the sequence of data generated by the generator; This represents the probability distribution for learning the true power sequence; This represents the probability distribution in the Gaussian noise data input generator; This represents the generative output model of a generative adversarial network. This represents the probability distribution of the power sequence generated by the adversarial generative network model; This represents the power generation sequence of photovoltaics in the CGAN network production sequence; This represents the power sequence representing the sum of the generating capacities of thermal power, hydropower, and gas power in the CGAN network production sequence. The difference in probability distribution of the power generation sequence of photovoltaic power generation compared to that of thermal power, hydropower, and gas power generation is called divergence. This represents the probability distribution for learning the true power sequence. Represents the actual photovoltaic power generation sequence The power generation sequence of thermal power, hydropower, and gas power generation. The divergence.

2. The novel power system dispatch scenario sample data generation method according to claim 1, characterized in that, S13 will normalize the extreme weather condition label variables. The data is input into the generator of the CGAN network to obtain the generated sample dataset. ;include: S131, normalized extreme weather condition label variables The input vector matrix is ​​concatenated with the Gaussian white noise matrix z to form the input vector matrix w. Among them, extreme weather condition label variables It is an n×5 matrix; S132, the input vector matrix w is input into the input layer of the CGAN network generator. After processing by the generator's deconvolution layer and output layer, the generated sample dataset is obtained. .

3. A novel method for constructing a control decision model for power system dispatching scenarios, characterized in that, Comprising: S21. Based on the CGAN network, generate the predicted power generation of various types of generating units, and construct state-space variables based on the predicted load and frequency of the regional new power system. ; Wherein, the predicted power generation of each type of unit is generated by the method for generating new power system dispatching scenario data according to any one of claims 1 to 2; S22, establish execution strategies based on the standby unit capacity, emergency load shedding amount, and external grid interconnection transmission power; each execution strategy corresponds to a specific action space. Based on action space Construct action set A; S23, initial state space variables Inputting the value into the training Q-network of the DQN network generates the value corresponding to the initial state. Where θ represents the neural network weights used to train the Q-network; S24, the ε-greedy policy function is used to select the corresponding action space from the action set A. ; S25, based on motion space The corresponding execution strategy execution control process; S26, Motion Space Calculate the immediate reward based on the reward function. The state space quantities at the next time step are calculated using Matpower power flow. ; S27, Constructing an empirical dataset ; Randomly sample d sets of experience datasets from the experience pool and input them into the training Q network and the target Q network of the DQN network, and calculate the value of the training Q network and the target value of the target Q network; Wherein: d<dm, dm is the capacity of experience pool data; S28, updating the θ value of the training Q network according to the loss function and the target value of the target Q network; S29, judging whether a preset number of iterations is reached, if yes, generating a control decision model based on a DQN network; otherwise, executing step S24.

4. The method for constructing a novel power system dispatch scenario control decision model according to claim 3, characterized in that, In said S28, the expression of the loss function is: ； In the formula: j represents the current iteration number; This represents the target value calculated by the Q-network for the j-th iteration. θ represents the parameter θ for the j-th iteration; T represents the preset number of iterations.

5. The method for constructing a novel power system dispatch scenario control decision model according to claim 4, characterized in that, The calculation expression of the target value of the target Q network is: ； In the formula: This represents the immediate reward during the j-th iteration. This represents the discount factor.

6. The method for constructing a novel power system dispatching scenario control decision model according to claim 4, characterized in that, The expression of the reward function is: ； ； ； ； In the formula, Let d represent the reward function based on frequency stability deviation, and d represent the penalty term when the frequency exceeds the limit; This represents the reward function based on the importance of the load. Indicates a level 1 load. Indicates a secondary load. Indicates a level three load. , , This represents the load weighting parameter, and The more critical the loss load, the greater the penalty. This indicates the amount of data used by the user to control the load. This represents a reward function based on cost control. Indicates the penalty coefficient for cost control items; Indicates a single control cycle; This represents the cost per unit of electricity consumed by the load. This indicates the cost per unit of electricity generated to represent the region's reserve capacity. This indicates the cost per unit of power supplied by the tie line. This indicates that reserve capacity is deployed in the t-th cycle, with a total of n cycles; This represents the power generation of the j-th tie line in the t-th cycle of the system, with a total of k tie lines.

7. The method for constructing a novel power system dispatch scenario control decision model according to claim 4, characterized in that, The state space variables The expression is: In the formula, These represent the power generation of thermal power, hydropower, gas power, wind power, and photovoltaic power at time t, respectively. This represents the predicted load of the regional power grid at time t. This represents the available generating capacity of the standby generator units in the region at time t; These represent the power flow at important cross-sections of the connecting line, where k is the number of power flows at important cross-sections. This represents the frequency deviation of the regional power grid at time t; The action space The expression is: In the formula: This indicates that the system will allocate reserve capacity after making a decision at time t. The control variable representing the load after the system decision at time t is... This indicates that the power of the k tie-line sections increases after the system makes a decision.

8. The method for constructing a novel power system dispatch scenario control decision model according to claim 4, characterized in that, In step S24, the ε-greedy policy function is used to select the corresponding action space from the action set A. In the above, the expression for the ε-greedy policy function is: in, Let ε-greedy be the policy function, representing the action. When it is the optimal choice, its selection probability is: Otherwise, the probability of choosing is ; ε is a very small positive number, and , representing the probability of random selection; Indicates the optimal action The corresponding maximum Q value; |A| represents the set of actions; This strategy states: when the action When it is the optimal choice, its selection probability is: Otherwise, the probability of choosing is: .

Images