A method and system for optimizing the peak-shaving of water, wind and solar energy multi-energy complementary

Abstract

The present invention discloses a method and system for optimizing peak-shaving scheduling of water-wind-solar multi-energy complementary, including: collecting output data of wind-solar power stations and preprocessing data sets, constructing a deep convolutional generative adversarial network, inputting data sets to train the deep convolutional generative adversarial network, and generating wind-solar scenario samples; collecting runoff and load data, using a martingale model to quantify and simulate the corresponding runoff or load process, and obtaining runoff and load scenario samples; based on a fast reduction algorithm for probability distance, respectively reducing and updating the wind-solar scenario samples, runoff and load scenario samples, and constructing a basin water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model, and using a particle swarm algorithm to solve it, and obtaining the optimal scheduling result. The present invention can effectively optimize the scheduling scheme by comprehensively considering the energy utilization efficiency, and ensure the safe and stable operation of the system under the peak-shaving demand of the power grid.

Description

Technical Field

[0001] The present invention belongs to the field of short-term peak-shaving optimization scheduling of a water-wind-solar multi-energy complementary system, and in particular, relates to a water-wind-solar multi-energy complementary peak-shaving optimization scheduling method and system. Background Art

[0002] Due to the difficulty in directly absorbing wind and solar resources and the large amount of abandoned light, in order to achieve efficient use of energy and maintain the real-time security of the power grid, it is usually adopted to combine wind and solar with hydropower to carry out water-wind-solar multi-energy complementary peak-shaving and optimized scheduling.

[0003] Existing peak load optimization scheduling technologies are usually one-sided when considering the uncertainty on both sides of the source and load. They are rarely able to simultaneously characterize the comprehensive uncertainty of multiple factors of water, wind and solar power, and select appropriate quantitative models based on the characteristics of energy or load. In addition, most existing scheduling methods are aimed at deterministic optimization scenarios. In fact, there are a large number of uncertain factors in the scheduling process, such as the difficulty in quantifying wind and solar energy, the linear loss of reservoirs, and the instability of load demand. If not taken into account, this uncertainty will inevitably interfere with the actual peak load regulation process of the power grid. Summary of the invention

[0004] In view of the shortcomings and defects of the above existing short-term peak-shaving optimization scheduling technologies, the present invention provides a short-term multi-energy complementary peak-shaving optimization scheduling method of water, wind and solar power that takes into account the uncertainty on both sides of the source and load. The method can fully consider the uncertainty on the source and load side in the actual grid scheduling and operation process, and is suitable for peak-shaving optimization in random scenarios. A short-term multi-energy complementary peak-shaving optimization scheduling system of water, wind and solar power that takes into account the uncertainty on both sides of the source and load is also provided.

[0005] In order to solve the above technical problems, the present invention provides a method and system for optimizing the peak-shaving of water, wind and solar power. A method for optimizing the peak-shaving of water, wind and solar power comprises:

[0006] Collect output data of wind and solar power stations and perform data set preprocessing, build a deep convolutional generative adversarial network, input the data set to train the deep convolutional generative adversarial network, and generate wind and solar scenario samples;

[0007] Collect runoff and load data, use the martingale model to quantify and simulate the corresponding runoff or load process, and obtain runoff and load scenario samples;

[0008] Based on the rapid reduction algorithm of probability distance, the wind-solar scenario samples, runoff and load scenario samples are reduced and probability updated respectively, and then the basin water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model is constructed, and the particle swarm algorithm is used to solve it to obtain the optimal scheduling result.

[0009] Preferably, the process of generating the scenery scene sample includes:

[0010] Collect hourly output data of different wind and solar power stations and reshape the data, set labels and perform maximum value normalization;

[0011] Build a deep convolutional generative adversarial network model, use the Earth Mover distance to construct the loss function, and replace the original fully connected layers in the generator and discriminator of the model with convolutional layers;

[0012] Sampling the data set in batch size and inputting it into the deep convolutional generative adversarial network model, the model updates the discriminator and generator networks in turn until both reach Nash equilibrium and the training is completed;

[0013] The noise vector extracted from the Gaussian distribution is input into the trained model, and the model generates landscape scene samples with different characteristics based on different labels.

[0014] Preferably, the deep convolutional generative adversarial network model is sampled from the data set according to the batch size, and the model updates the discriminator and the generator network in sequence until both reach Nash equilibrium. The training process includes:

[0015] Set the hyper-optimization parameters, initialize the parameters of the generator and discriminator, and then perform iterative training. In each iteration, batch sample input from historical samples and Gaussian distribution respectively;

[0016] The RMSProp optimizer is used. When the number of iterations is a multiple of the balance parameter, the generator parameters are updated. Otherwise, the discriminator parameters are updated, and the loss functions corresponding to the generator and the discriminator are calculated. When the model converges or reaches the maximum number of iterations, the training ends.

[0017] The super optimization parameters include but are not limited to the number of iterations, batch size, learning rate and balance parameters.

[0018] Preferably, the deep convolutional generative adversarial network model includes a discriminator and a generator;

[0019] The discriminator includes a 4-layer convolution structure, wherein the first to third layers are 3 convolution layers, each layer is first merged with the label information, and then input into the next layer after batch normalization and activation function normalization. If necessary, a dropout layer is added to suppress the overfitting phenomenon of the model; the fourth layer first flattens the multidimensional data, and then maps the data into a scalar and normalizes it through a fully connected layer;

[0020] The generator includes a three-layer convolution structure. After merging with the label information and normalizing it, the first layer is a fully connected layer that converts the input random noise into a matrix. The second to third layers are two deconvolution layers. Except for the dropout layer, the structure is opposite to that of the discriminator.

[0021] Preferably, the process of collecting runoff and load data, quantifying and simulating the corresponding runoff or load process using a martingale model, and obtaining runoff and load scenario samples includes:

[0022] Collect runoff and load series, set standard deviation parameters for forecast improvement and correlation coefficient parameters for forecast error;

[0023] The Martingale model is used to generate forecast improvements for runoff and load series respectively, and ensemble forecasts are calculated to generate different runoff and load scenarios.

[0024] Preferably, the process of using the martingale model to generate forecast improvements of runoff and load series respectively, calculating ensemble forecasts, and generating different runoff and load scenarios includes:

[0025] Constructing a variance-covariance matrix VCV of forecast improvements, using Cholesky to decompose the variance-covariance matrix VCV to obtain an upper triangular matrix, and then using a Monte Carlo simulation method to randomly generate forecast improvements;

[0026] Combining the forecast improvements with the actual sequence information, simulation scenarios of different runoff and loads are obtained.

[0027] Preferably, the process of reducing and probabilistically updating the wind-solar scenario samples, runoff and load scenario samples respectively based on the fast reduction algorithm of probability distance includes:

[0028] Calculate the probability distance of each pair of scenarios a and a' generated in scenario set A respectively,

[0029] Select the scenario d with the smallest sum of distances from the remaining scenario probabilities, and select the scenario r with the smallest distance from the smallest scenario d in the scenario set A, add the scenario probability of the smallest scenario d to the scenario r, and perform scenario sample reduction on the smallest scenario d;

[0030] The above steps are repeated until the number of scenarios reaches the requirement, and a new scenario set is obtained.

[0031] Preferably, a basin water-wind-light short-term multi-energy complementary peak-shaving optimization scheduling model is constructed and solved using a particle swarm algorithm. The process of obtaining the optimal scheduling result includes:

[0032] The characteristic parameters of wind-solar-hydropower stations, reservoir characteristic curve information, and the generated wind-solar scenario samples, runoff and load scenario samples are input into the water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model. Each scenario is optimized separately, and the expected sequence process of each scenario result is taken as the optimization target. The model is randomly optimized and scheduled, and the PSO algorithm is introduced to solve it to obtain the optimal scheduling plan.

[0033] Preferably, before inputting the characteristic parameters of the wind-solar-hydropower station, the reservoir characteristic curve information, and the generated wind-solar scenario samples, runoff and load scenario samples into the water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model, the following steps are also included:

[0034] The decision variable of the model is set as water level, and the objective function is to minimize the sum of the residual load of the power grid in all periods under source load uncertainty;

[0035] Setting constraints of the model, wherein the constraints include reservoir water balance constraints, decision variable constraints, outflow constraints, hydropower output constraints and power grid constraints;

[0036] The scenarios that have been reduced and probability updated are weighted according to their respective probabilities and then input into the model to incorporate uncertainty factors.

[0037] The present invention also provides a water, wind and solar multi-energy complementary peak-shaving optimization scheduling system, comprising:

[0038] Data acquisition module, used to collect output data of wind and solar power stations and perform data set preprocessing;

[0039] A model building module, connected to the data acquisition module, for building a deep convolutional generative adversarial network, inputting a data set to train the deep convolutional generative adversarial network, and generating scenery scene samples;

[0040] A sample generation module, connected to the model building module, is used to collect runoff and load data, quantify and simulate corresponding runoff or load processes using a martingale model, and obtain runoff and load scenario samples;

[0041] The scheduling optimization module is connected to the sample generation module and is used for a rapid reduction algorithm based on probability distance. After reducing and probabilistically updating the wind-solar scenario samples, runoff and load scenario samples, a basin water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model is constructed, and a particle swarm algorithm is used to solve it to obtain the optimal scheduling result.

[0042] Compared with the prior art, the present invention has the following advantages and technical effects:

[0043] The optimization scheduling method of the present invention can provide a basis for the scheduling plan of real-time tracking of the power grid load curve under multiple uncertain factors. On the one hand, combining mathematical statistics and generative deep learning theory, the uncertainty of the source-load side is randomly quantified and simulated: according to the characteristics of the uncertainty factors, a deep convolutional generative adversarial network is used to generate wind and solar scenarios, and a martingale model is introduced to simulate runoff and load. On the other hand, the uncertainty factors on the source-load side are input, and a basin water-wind-solar short-term complementary optimization scheduling model with the goal of real-time tracking of the load curve is constructed, and the PSO algorithm is used to solve it.

[0044] The optimization scheduling method of the present invention is an effective means to solve the problem of wind and solar grid connection and alleviate the pressure of uncertainty. By comprehensively considering the energy utilization efficiency, it can effectively optimize the scheduling plan and ensure the safe and stable operation of the system under the peak load demand of the power grid. At the same time, by combining mathematical statistics models and generative adversarial networks, this method can quickly respond to uncertainty disturbances, provide strong support for the formulation of short-term scheduling plans for wind, solar and hydropower systems in complex river basins, and promote the development of clean energy and sustainable development of energy systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0046] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0047] Figure 2 Schematic diagram of the process of step S1 of an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of the process of step S2 of an embodiment of the present invention;

[0049] Figure 4 Schematic diagram of the process of step S3 of an embodiment of the present invention;

[0050] Figure 5 Schematic diagram of the process of step S4 of an embodiment of the present invention;

[0051] Figure 6 Schematic diagram of the DCGAN model of an embodiment of the present invention. DETAILED DESCRIPTION

[0052] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0053] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0054] like Figure 1-6 As shown, this embodiment provides a method for optimizing the peak-shaving of water, wind and solar energy by complementary energy, including the following steps:

[0055] Step S1, collecting output data of different wind and solar power stations and preprocessing the data set, constructing a deep convolutional generative adversarial network (DCGAN), inputting the data set to train the DCGAN, and obtaining different wind and solar scenario samples generated by it;

[0056] Step S2, collecting runoff and load data, using a martingale model to quantify and simulate the corresponding runoff or load process, and obtaining runoff and load scenario samples;

[0057] Step S3: Use a fast reduction algorithm based on probability distance to reduce and update the probability of wind and solar, runoff and load scenarios respectively;

[0058] Step S4, input the above uncertain factors, build a basin water-wind-light short-term multi-energy complementary peak-shaving optimization scheduling model, and use particle swarm optimization (PSO) to solve it.

[0059] Further, in step S11, after collecting the hourly output data of different wind and solar power stations, the data set is reshaped, labels are set, and maximum value normalization is performed;

[0060] Step S12: construct a DCGAN, use the Earth Mover distance to construct the loss function, and replace the original fully connected layers in the generator and discriminator of the model with convolutional layers;

[0061] Step S13: Sample the data set according to the batch size and input it into DCGAN. The model updates the discriminator and generator networks in turn until both reach Nash equilibrium and the training is completed.

[0062] Step S14: input the noise vector extracted from the Gaussian distribution into the trained model, and the model generates landscape scene samples with different characteristics according to different labels.

[0063] Renewable energy generation has strong uncertainty and spatiotemporal correlation. The explicit expression of the joint probability distribution of its output is usually difficult to obtain accurately. The traditional scenario generation method that requires known distribution is not completely applicable. In recent years, the continuous development of deep learning has provided new ideas for scenario generation. Generative adversarial networks are a type of deep learning model. Since its proposal, it has attracted widespread attention due to its strong generalization ability. DCGAN is one of the most representative models. DCGAN is usually composed of a generator (G) and a discriminator (D). It transforms a set of noise vectors that follow a Gaussian distribution into a The generator is input and tries to generate fake samples close to real data. The discriminator receives real samples and generated samples at the same time and distinguishes true from false. In this process, the two-way confrontation is repeated, and the weights of G and D are optimized through continuous iteration. and , until both parties reach Nash equilibrium, thus forming a minimax game model, whose objective function is:

[0064]

[0065] In the formula, is the expectation of a given variable; , are the real data and noise vectors respectively; , are the distributions of real data and noise vectors, respectively.

[0066] Further, step S13a, setting the number of iterations, batch size, learning rate, balance parameters and other super-optimization parameters, and initializing the parameters of the generator and the discriminator;

[0067] Step S13b: In each iteration, batch samples are taken from the historical samples and the Gaussian distribution and input into the model;

[0068] Step S13c: Use the RMSProp optimizer to update the generator parameters when the number of iterations is a multiple of the balance parameter, otherwise update the discriminator parameters, and calculate the loss functions corresponding to the generator and the discriminator.

[0069] Step S13d: When the model converges or reaches the maximum number of iterations, the training ends.

[0070] In the network structure of DCGAN, the discriminator has a total of 4 convolutional structures. The first to third layers are 3 convolutional layers. Each layer is first merged with the label information, and then input into the next layer after batch normalization and activation function normalization. If necessary, a dropout layer is added to suppress the overfitting of the model. The fourth layer first flattens the multidimensional data, and then maps the data to a scalar and normalizes it through a fully connected layer. The generator has a total of 3 convolutional structures. After merging with the label information and normalization, the first layer is a fully connected layer that converts the input random noise into a matrix. The second to third layers are 2 deconvolution layers. Except for the dropout layer, the structure is opposite to that of the discriminator.

[0071] The model is trained using the RMSProp optimizer. All initialized weights in the network structure obey the N(0,1) Gaussian distribution. Context labels are added to each layer except the input layer and batch normalized. The generator uses the ReLU activation function except the output layer, and the discriminator uses the Leaky-ReLU activation function for activation.

[0072] Further, in step S21, the runoff and load series are collected, and the standard deviation parameter of the forecast improvement and the correlation coefficient parameter of the forecast error are set;

[0073] Step S22: using the martingale model to generate forecast improvements for the runoff and load series respectively, calculating the ensemble forecast, and obtaining different runoff and load scenarios generated thereby.

[0074] Further, step S22a, constructing a forecast-improved variance-covariance matrix VCV;

[0075] Step S22b, using Cholesky to decompose the VCV matrix to obtain its upper triangular matrix, and then using the Monte Carlo simulation method to randomly generate forecast improvements;

[0076] Step S22c: Combining the forecast improvement with the actual sequence information, a certain number of simulation scenarios are obtained.

[0077] In the short-term forecast of the water-wind-solar hybrid system, an hourly rolling scheduling strategy is generally used to dynamically update the future sequence information in real time. The Martingale model was first applied to the real-time forecast of product demand in supply chain management, and was later often used for sequence forecasting. As a simplified theory in probability theory and mathematical statistics, the Martingale model assumes that the future runoff forecast changes are unbiased, and uses this as a basis to measure the error distribution between historical and predicted samples. It assumes that the forecast improvement at each moment during the scheduling period is represented by the same variance-covariance matrix VCV:

[0078]

[0079]

[0080] In the formula, is the forecast improvement from time series w to w+H-1; VCV is The variance-covariance matrix of and They are The variance and covariance of .

[0081] Finally, the actual sequence information and forecast improvement are combined to obtain the ensemble forecast of the sequence:

[0082]

[0083] In the formula, is the predicted value of each time series; for The measured value at the time.

[0084] Further, step S31, respectively calculating the probability distance of each pair of scenarios a and a' generated in the scenario set A;

[0085] Step S32, selecting the scenario d with the smallest sum of distances to the remaining scenario probabilities, and selecting the scenario r with the smallest distance to d in the scenario set A, and adding the scenario probability of d to r, so as to achieve the effect of reducing d;

[0086] Step S33, loop the above steps until the number of scenarios reaches the requirement, and obtain a new scenario set.

[0087] The scenarios generated by scenario generation are usually numerous. If they are directly input into the subsequent model, not only will the amount of calculation and the complexity of calculation increase dramatically, but they may also cause problems such as the curse of dimensionality when high-dimensional calculations are involved. In order to comprehensively balance the scheduling effect of the model and the burden of computer solving, it is usually adopted to select representative scenarios from the original scenarios to replace the original scenarios through scenario reduction and probability update, while retaining the characteristics of the original time series scenarios as much as possible.

[0088] The probability distance-based rapid reduction algorithm is a commonly used scenario forward reduction method. The algorithm evaluates the similarity between scenarios based on the probability distance between them and performs rapid reduction based on the similarity to reduce the number of scenarios and retain diversity. At the same time, the probability information corresponding to the reduced scenarios can be obtained.

[0089] Further, in step S41, characteristic parameters of wind-solar-hydropower stations, reservoir characteristic curve information, and wind-solar-runoff and load scenarios generated by different methods are input into a constructed water-wind-solar multi-energy complementary peak-shaving optimization scheduling model;

[0090] Step S42: optimize each scenario separately, and then take the expected sequence process of each scenario result as the optimization target, perform random optimization scheduling of the model, and introduce the PSO algorithm to solve it, so as to obtain the optimal scheduling solution.

[0091] In the field of mathematics, stochastic programming theory is often used to deal with some uncertain decision-making problems. As a framework for modeling optimization problems, stochastic programming introduces random variables and known or estimated probability models on the basis of traditional deterministic planning to consider uncertainty factors. It is an effective means to solve uncertainty problems and has been widely used in the dispatch of power systems. In order to handle and quantify multiple uncertain factors on both sides of the simulated source and load, this embodiment establishes a stochastic peak-shaving optimization dispatching model for a cascaded hydro-wind-solar hybrid system based on stochastic programming theory, and uses the PSO algorithm to solve it.

[0092] The standard particle swarm optimization (PSO) is a swarm intelligence optimization algorithm inspired by social behaviors such as bird foraging. Its core idea is to solve the optimization problem by simulating the collaborative behavior of birds in a flock and generating an evolutionary process from disorder to order in the solution space. At present, the PSO algorithm is commonly used for nonlinear, high-dimensional complex constrained optimization problems due to its advantages such as easy implementation and fast convergence speed. In the PSO algorithm, the solution to the problem is represented as an individual in a particle swarm. Each individual has two parameters, a position vector and a velocity vector, and is dynamically adjusted according to the current best individual position and the global best position. In this way, the position and velocity of the particles are continuously updated in an iterative manner to obtain the optimal solution. The following evolution equation is usually used during the update:

[0093]

[0094]

[0095] In the formula, represents the velocity vector for the next iteration; represents the position vector of the next iteration; , They are cognitive factor and social learning factor; , are random numbers between [0,1]; is the individual best historical position of the particle; is the global optimal position of the particle swarm; For particles The inertia weight for the current iteration, assumed to change linearly from start to end.

[0096] Further, step S41a, setting the decision variable of the model to the water level, and the objective function to minimize the sum of the residual load of the power grid in all time periods under the uncertainty of source and load;

[0097] The most ideal peak-shaving state in the dispatching model is to completely eliminate the peak-to-valley difference, so that the load curve becomes a smooth straight line. Then thermal power generation can be connected, thereby greatly reducing the start-up and shutdown time and operating costs of thermal power units. Therefore, the objective function is to minimize the sum of the residual load of the power grid in all periods under source-load uncertainty. The specific expression is:

[0098]

[0099]

[0100] In the formula, is the power load at the current moment; For the current moment, water, wind and light work together; , , They are the wind turbine output, photovoltaic output and hydropower output at the current moment respectively; For the scheduling period.

[0101] Step S41b, setting the constraints of the model including reservoir water balance constraints, decision variable constraints, outflow constraints, hydropower output constraints and power grid constraints;

[0102] In the iterative process, only by ensuring the execution of each constraint condition can feasible solutions be screened out efficiently and accurately, and infeasible solutions be eliminated, thereby improving the efficiency of model operation. Among them, equality constraints can be expressed using equations, while non-equality constraints are often nonlinear and highly coupled, making them difficult to model directly. Therefore, consider using a dynamic penalty term , the original non-equality constraints are converted into equality penalty terms, and the original constrained optimization problem can be converted into an unconstrained problem, reducing the difficulty of solving. The specific conversion process is shown in the following formula:

[0103]

[0104]

[0105]

[0106] From the above formula, we can see that only the penalty term of the feasible solution that satisfies all constraints is 0, and the rest will be subject to the maximum penalty term, which is used to implement the non-equality constraint. The expression of the dynamic penalty function is as follows:

[0107]

[0108] In the formula, , and is a constant; is the number of current iterations; is the number of non-equality constraints.

[0109] Step S41c: The scenarios that have been reduced and probability updated are weighted according to their respective probabilities and then input into the model to incorporate uncertainty factors.

[0110] A random optimization model is constructed based on random scenarios and corresponding probabilities. The specific expression is:

[0111]

[0112]

[0113] In the formula, For the scenario The corresponding probability; For the scenario The power load at the current moment; For the scenario At this moment, water, wind and light work together; , , Separate scenarios Wind turbine output, photovoltaic output and hydropower output at the current moment.

[0114] The following is a study of a water-wind-solar multi-energy complementary system in the upper reaches of the Yellow River as the research object, and an exemplary description is given of the water-wind-solar multi-energy complementary peak-shaving optimization scheduling method that takes into account the uncertainties on both the source and load sides provided in this embodiment.

[0115] The construction details of the DCGAN model are as follows:

[0116] The input dataset is reshaped with a data dimension of 12×12×2. The generator uses a 2×2 deconvolution layer to batch process the 100×1 noise vector z, and finally generates samples of the specified data dimension; the discriminator uses a 2×2 convolution layer to batch process the 12×12×2 input samples, and finally outputs a scalar. The maximum number of iterations is 7000, the batch size is 28, the learning rate is 0.002, and the balance parameter is 4.

[0117] The PSO algorithm parameter settings are as follows:

[0118] The relevant parameters of the PSO algorithm are set as follows: the population size is selected as 300, the maximum number of iterations is selected as 4000, the minimum inertia weight is selected as 0.3, the maximum value is selected as 0.9, and the self-learning factor and the group learning factor are both selected as 2. The optimal fitness function value is obtained through continuous iteration of the PSO algorithm.

[0119] In order to verify the superiority of the proposed method, five dispatch modes are set to analyze their load-balancing effects and the benefits of joint dispatch. The specific modes are: Mode 1: subtract wind power output from the original hybrid system; Mode 2: subtract photovoltaic output from the original hybrid system; Mode 3: subtract wind power and photovoltaic power from the original hybrid system; Mode 4: cascade hydropower does not undertake the multi-energy complementary peak-shaving task, but independently optimizes the dispatch with the goal of maximizing the cascade's own power generation; Mode 5: water-wind-light combined peak-shaving optimization dispatch. The three indicators of load rate, standard deviation and peak-valley difference are used to evaluate the performance of the dispatch mode. The performance evaluation results under different dispatch modes are shown in Table 1.

[0120] Table 1

[0121]

[0122] Comparing the results of modes 1, 2, and 3, it can be seen that when wind power or photovoltaic power is connected to the model alone, the water-wind and water-light scheduling modes have a certain degree of complementarity with pure hydropower scheduling. The peak-to-valley difference improvement rate of mode 3 is -6.65%, which is even worse than the peak-to-valley difference of the original load curve, while the corresponding improvement rates of modes 1 and 2 are 23.7% and 30.3%. The standard deviation indicators in the three modes have also been improved, but the performance in the maximum load rate is not good, indicating that although the complementarity of the two energies can stabilize the peak-to-valley difference of the load, it cannot completely eliminate the instability and volatility caused by the integration of wind and solar power into the power grid.

[0123] Compared with modes 1, 2, and 3, the performance of mode 5 (hydro-wind-solar joint peak-shaving optimization) has been basically improved, with the peak-to-valley difference, standard deviation, and maximum load rate increased by 35.8%, 39.4%, and 12.0%, respectively, indicating that hydropower fully exerts its dispatchability in this mode, reflecting the complementarity between hydropower and wind power and photovoltaic power, which not only greatly alleviates the peak-shaving pressure, but also solves the problem of the irregularity of wind and solar power generation. The above analysis can prove that the effect of hydro-wind-solar joint dispatching to stabilize the load curve is the best.

[0124] In addition to the goal of stabilizing the remaining load, mode 4 considers the maximum power generation efficiency from the perspective of optimizing the joint power generation of cascade reservoirs and analyzing the impact of the complementary mechanism on the power generation efficiency of cascade hydropower stations. By comparing and analyzing modes 4 and 5, the total power generation under mode 4 is 100,540.4 MWh, an increase of 210.8 MWh compared with mode 5, an increase of 0.2%. However, the cost of maximizing power generation is that the optimized remaining load curve fluctuates violently, which may increase the subsequent start-up and shutdown and operation costs of thermal power generating units, which is not conducive to the overall benefits. Therefore, it is more appropriate to sacrifice only 0.2% of hydropower generation in exchange for a smooth load operation process.

[0125] In addition, this embodiment also provides a water-wind-solar multi-energy complementary peak-shaving optimization scheduling system, including:

[0126] Data acquisition module, used to collect output data of wind and solar power stations and perform data set preprocessing;

[0127] A model building module, connected to the data acquisition module, for building a deep convolutional generative adversarial network, inputting a data set to train the deep convolutional generative adversarial network, and generating scenery scene samples;

[0128] A sample generation module, connected to the model building module, is used to collect runoff and load data, quantify and simulate corresponding runoff or load processes using a martingale model, and obtain runoff and load scenario samples;

[0129] The scheduling optimization module is connected to the sample generation module and is used for a rapid reduction algorithm based on probability distance. After reducing and probabilistically updating the wind-solar scenario samples, runoff and load scenario samples, a basin water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model is constructed, and a particle swarm algorithm is used to solve it to obtain the optimal scheduling result.

[0130] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for optimizing peak-shaving by complementing water, wind and solar energy, characterized in that: include: Collect output data of wind and solar power stations and perform data set preprocessing, build a deep convolutional generative adversarial network, input the data set to train the deep convolutional generative adversarial network, and generate wind and solar scenario samples; Collect runoff and load data, use the martingale model to quantify and simulate the corresponding runoff or load process, and obtain runoff and load scenario samples; Based on the rapid reduction algorithm of probability distance, the wind-solar scenario samples, runoff and load scenario samples are reduced and probability updated respectively, and then the basin water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model is constructed, and the particle swarm algorithm is used to solve it to obtain the optimal scheduling result; The process of collecting runoff and load data, quantifying and simulating the corresponding runoff or load process using the martingale model, and obtaining runoff and load scenario samples includes: Collect runoff and load series, set standard deviation parameters for forecast improvement and correlation coefficient parameters for forecast error; The Martingale model is used to generate forecast improvements for runoff and load series respectively, and ensemble forecasts are calculated to generate different runoff and load scenarios. The process of using the Martingale model to generate forecast improvements for runoff and load series respectively, calculating ensemble forecasts, and generating different runoff and load scenarios includes: Constructing a variance-covariance matrix VCV of forecast improvements, using Cholesky to decompose the variance-covariance matrix VCV to obtain an upper triangular matrix, and then using a Monte Carlo simulation method to randomly generate forecast improvements; Combining forecast improvements with actual sequence information, simulation scenarios of different runoff and loads are obtained; The short-term multi-energy complementary peak-shaving optimization scheduling model of water, wind and light in the basin is constructed and solved by particle swarm algorithm. The process of obtaining the optimal scheduling result includes: The characteristic parameters of wind-solar-hydropower stations, reservoir characteristic curve information, and generated wind-solar scenario samples, runoff and load scenario samples are input into the water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model, and each scenario is optimized separately. Then, the expected sequence process of each scenario result is taken as the optimization target, and the model is randomly optimized and scheduled. The PSO algorithm is introduced to solve it and obtain the optimal scheduling plan. Before inputting the characteristic parameters of wind-solar-hydropower stations, reservoir characteristic curve information, and generated wind-solar scenario samples, runoff and load scenario samples into the water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model, it also includes: The decision variable of the model is set as water level, and the objective function is to minimize the sum of the residual load of the power grid in all periods under source load uncertainty; Setting constraints of the model, wherein the constraints include reservoir water balance constraints, decision variable constraints, outflow constraints, hydropower output constraints and power grid constraints; The scenarios that have been reduced and probability updated are weighted according to their respective probabilities and then input into the model to incorporate uncertainty factors.

2. The method for optimizing peak-shaving by water, wind and solar energy multi-energy complementary peak-shaving according to claim 1 is characterized in that: The process of generating the scenery scene sample includes: Collect hourly output data of different wind and solar power stations and reshape the data, set labels and perform maximum value normalization; Build a deep convolutional generative adversarial network model, use the Earth Mover distance to construct the loss function, and replace the original fully connected layers in the generator and discriminator of the model with convolutional layers; Sampling the data set in batch size and inputting it into the deep convolutional generative adversarial network model, the model updates the discriminator and generator networks in turn until both reach Nash equilibrium and the training is completed; The noise vector extracted from the Gaussian distribution is input into the trained model, and the model generates landscape scene samples with different characteristics based on different labels.

3. The method for optimizing peak-shaving by water, wind and solar energy multi-energy complementary peak-shaving according to claim 2 is characterized in that: The deep convolutional generative adversarial network model is sampled from the data set according to the batch size. The model updates the discriminator and generator networks in turn until both reach Nash equilibrium. The training process includes: Set the hyper-optimization parameters, initialize the parameters of the generator and discriminator, and then perform iterative training. In each iteration, batch sample input from historical samples and Gaussian distribution are respectively input into the model; The RMSProp optimizer is used. When the number of iterations is a multiple of the balance parameter, the generator parameters are updated. Otherwise, the discriminator parameters are updated, and the loss functions corresponding to the generator and the discriminator are calculated. When the model converges or reaches the maximum number of iterations, the training ends. The super optimization parameters include but are not limited to the number of iterations, batch size, learning rate and balance parameters.

4. The method for optimizing peak-shaving by water, wind and solar energy multi-energy complementary peak-shaving according to claim 3 is characterized in that: The deep convolutional generative adversarial network model includes a discriminator and a generator; The discriminator includes a 4-layer convolution structure, wherein the first to third layers are 3 convolution layers, each layer is first merged with the label information, and then input into the next layer after batch normalization and activation function normalization, and the overfitting phenomenon of the model is suppressed by adding a dropout layer; the fourth layer first flattens the multidimensional data, and then maps the data into a scalar and normalizes it through a fully connected layer; The generator includes a three-layer convolution structure. After merging with the label information and normalizing it, the first layer is a fully connected layer that converts the input random noise into a matrix. The second to third layers are two deconvolution layers. Except for the dropout layer, the structure is opposite to that of the discriminator.

5. The method for optimizing peak-shaving by water, wind and solar energy multi-energy complementary peak-shaving according to claim 1 is characterized in that: Based on the fast reduction algorithm of probability distance, the process of reducing and probabilistically updating the wind and solar scenario samples, runoff and load scenario samples respectively includes: Calculate the probability distance of each pair of scenarios a and a' generated in scenario set A respectively, Select the scenario d with the smallest sum of distances from the remaining scenario probabilities, and select the scenario r with the smallest distance from the smallest scenario d in the scenario set A, add the scenario probability of the smallest scenario d to the scenario r, and perform scenario sample reduction on the smallest scenario d; The above steps are repeated until the number of scenarios reaches the requirement, and a new scenario set is obtained.

6. A water, wind and solar multi-energy complementary peak-shaving optimization scheduling system, characterized in that: include: Data acquisition module, used to collect output data of wind and solar power stations and perform data set preprocessing; A model building module, connected to the data acquisition module, for building a deep convolutional generative adversarial network, inputting a data set to train the deep convolutional generative adversarial network, and generating scenery scene samples; A sample generation module, connected to the model building module, is used to collect runoff and load data, quantify and simulate corresponding runoff or load processes using a martingale model, and obtain runoff and load scenario samples; The scheduling optimization module is connected to the sample generation module and is used for a fast reduction algorithm based on probability distance. After reducing and updating the probability of the wind-solar scenario samples, runoff and load scenario samples, a watershed water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model is constructed, and a particle swarm algorithm is used to solve the problem to obtain the optimal scheduling result. The process of collecting runoff and load data, quantifying and simulating the corresponding runoff or load process using the martingale model, and obtaining runoff and load scenario samples includes: Collect runoff and load series, set standard deviation parameters for forecast improvement and correlation coefficient parameters for forecast error; The Martingale model is used to generate forecast improvements for runoff and load series respectively, and ensemble forecasts are calculated to generate different runoff and load scenarios. The process of using the Martingale model to generate forecast improvements for runoff and load series respectively, calculating ensemble forecasts, and generating different runoff and load scenarios includes: Constructing a variance-covariance matrix VCV of forecast improvements, using Cholesky to decompose the variance-covariance matrix VCV to obtain an upper triangular matrix, and then using a Monte Carlo simulation method to randomly generate forecast improvements; Combining forecast improvements with actual sequence information, simulation scenarios of different runoff and loads are obtained; The short-term multi-energy complementary peak-shaving optimization scheduling model of water, wind and light in the basin is constructed and solved by particle swarm algorithm. The process of obtaining the optimal scheduling result includes: The characteristic parameters of wind-solar-hydropower stations, reservoir characteristic curve information, and generated wind-solar scenario samples, runoff and load scenario samples are input into the water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model, and each scenario is optimized separately. Then, the expected sequence process of each scenario result is taken as the optimization target, and the model is randomly optimized and scheduled. The PSO algorithm is introduced to solve it and obtain the optimal scheduling plan. Before inputting the characteristic parameters of wind-solar-hydropower stations, reservoir characteristic curve information, and generated wind-solar scenario samples, runoff and load scenario samples into the water-wind-solar short-term multi-energy complementary peak-shaving optimization scheduling model, it also includes: The decision variable of the model is set as water level, and the objective function is to minimize the sum of the residual load of the power grid in all periods under source load uncertainty; Setting constraints of the model, wherein the constraints include reservoir water balance constraints, decision variable constraints, outflow constraints, hydropower output constraints and power grid constraints; The scenarios that have been reduced and probability updated are weighted according to their respective probabilities and then input into the model to incorporate uncertainty factors.

Images