Optimization method for participation of photovoltaic power station in day-ahead power dispatching

Abstract

The invention discloses an optimization method for a photovoltaic power station to participate in electric power day-ahead scheduling, and the method comprises the steps: carrying out the K-Means clustering of the historical photovoltaic output of a target photovoltaic power station and corresponding meteorological information, constructing a typical output scene, and depicting the dynamic characteristics of a prediction error through a deep learning model; and then, introducing a conditional value-at-risk CVaR as a regulation and control index, constructing a multi-scene random scheduling optimization model covering output uncertainty, solving the multi-scene random scheduling optimization model by using a whale optimization algorithm, and generating a day-ahead scheduling plan considering both accuracy and robustness.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power station scheduling, and more specifically, relates to an optimization method for photovoltaic power stations to participate in day-ahead power scheduling. Background Art

[0002] As the proportion of renewable energy generation continues to rise, photovoltaic power plants are gradually shifting from traditional guaranteed grid connection to a dispatch management system based on power market mechanisms. In day-ahead dispatch, the power market requires photovoltaic power plants to submit accurate next-day output plans to facilitate unified coordination and resource allocation at the system dispatch center. Therefore, developing accurate and reliable day-ahead dispatch plans has become a crucial step in the grid-connected operation of photovoltaic power plants.

[0003] However, due to the significant intermittency and uncertainty of photovoltaic power generation and its significant impact on weather, output forecasts can be subject to significant errors, increasing the risk of plan execution deviations. This not only impacts the ability of power plants to fulfill their dispatch obligations within market mechanisms but can also cause system-wide fluctuations and operational pressures. Traditional forecasting and dispatching methods often struggle to effectively address complex weather conditions, such as cloudy and rainy conditions, impacting the security and controllability of grid dispatch.

[0004] Currently, photovoltaic power stations face two core problems when participating in electricity market dispatch: first, the ability to model output forecast errors is insufficient. Conventional methods are often based on simplified statistical models and lack a dynamic characterization of uncertainty characteristics under different meteorological scenarios. Second, dispatch strategies lack sufficient flexibility and robustness to effectively respond to rapidly changing environmental conditions, thereby reducing the feasibility of dispatch plans and system adaptability.

[0005] Therefore, there is an urgent need for an intelligent dispatch optimization method that integrates advanced technologies such as artificial intelligence and machine learning to intelligently identify and model typical weather scenarios and forecast error distributions, thereby optimizing dispatch plans under multiple uncertainties. Improving the accuracy and robustness of dispatch plans will help enhance the operational coordination and system-friendliness of photovoltaic power plants within the power market mechanism, promoting their deeper and more intelligent integration into the new power system. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide an optimization method for photovoltaic power stations to participate in power day-ahead scheduling. Through the in-depth mining of multi-scenario information by intelligent algorithms and the introduction of risk control strategies, the plan executability and adaptability of photovoltaic power stations under the power market scheduling mechanism can be effectively improved, and it has strong practical application value and promotion potential.

[0007] To achieve the above-mentioned object of the invention, the present invention provides an optimization method for a photovoltaic power station to participate in day-ahead power dispatching, characterized by comprising the following steps:

[0008] S1. Combine the historical photovoltaic output of the target photovoltaic power station and the corresponding meteorological information, and cluster them using the K-Means clustering algorithm to obtain typical meteorological scenarios of different categories;

[0009] S2. Use the particle swarm optimization-gradient boosting tree prediction model to predict photovoltaic output for different categories of typical meteorological scenarios, and then calculate the prediction error;

[0010] S3. Divide the prediction errors of samples in each typical meteorological scenario into hourly time periods to form an error sample set, and then use the kernel density estimation method to construct the probability density function of the prediction error;

[0011] S4. Introducing conditional value at risk (CVaR) as a control indicator to build a multi-scenario stochastic scheduling optimization model that covers output uncertainty;

[0012] S5. The multi-scenario random scheduling optimization model is solved through the whale optimization algorithm (CWOA) to obtain the optimal day-ahead scheduling declaration strategy under different typical weather conditions and time periods.

[0013] The object of the invention of the present invention is achieved like this:

[0014] The present invention provides an optimization method for photovoltaic power stations participating in day-ahead power dispatching. The method performs K-Means clustering on the historical photovoltaic output of the target photovoltaic power station and the corresponding meteorological information to construct a typical output scenario, and characterizes the dynamic characteristics of the prediction error through a deep learning model. Then, the conditional value at risk (CVaR) is introduced as a control indicator to construct a multi-scenario random dispatching optimization model covering output uncertainty. The multi-scenario random dispatching optimization model is then solved using the whale optimization algorithm to generate a day-ahead dispatching plan that takes into account both accuracy and robustness.

[0015] At the same time, the optimization method of a photovoltaic power station participating in power day-ahead dispatching of the present invention also has the following beneficial effects:

[0016] (1) Improving the intelligence level of scheduling plan formulation: The present invention uses a clustering-based typical meteorological scenario recognition mechanism, combined with the PSO-LGBM output prediction model and kernel density estimation technology, to construct an output prediction error model for different typical meteorological scenarios, so that the scheduling plan formulation process has stronger data-driven capabilities and scenario adaptability, significantly improving the intelligence level of photovoltaic power station scheduling decision-making.

[0017] (2) Improve the robustness of photovoltaic power station scheduling plans and enhance the ability to cope with uncertain risks: The present invention constructs a random optimization model based on conditional value at risk (CVaR), comprehensively considers the probability density function characteristics of the prediction error, effectively controls the risks brought by prediction deviations, and enhances the planning execution and strategy robustness of photovoltaic power stations under complex weather conditions.

[0018] (3) Improve the feasibility and rationality of photovoltaic scheduling plans: The day-ahead scheduling strategy generated by the present invention fully considers the uncertainty of the prediction error and its probability distribution characteristics, and can output scheduling suggestions that are more in line with the actual output capacity, effectively reducing the deviation and default risk during the strategy execution process, thereby improving the feasibility and execution stability of the photovoltaic power station scheduling plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of an optimization method for a photovoltaic power station to participate in day-ahead power dispatching according to the present invention;

[0020] Figure 2 It is a flow chart of particle swarm optimization-gradient boosting tree prediction model processing;

[0021] Figure 3 This is a comparison chart of photovoltaic predicted output and actual output under different meteorological scenarios;

[0022] Figure 4 This is a comparison chart of the optimal scheduling plan curve under sunny conditions;

[0023] Figure 5 This is a comparison chart of the optimal scheduling plan curve under cloudy conditions;

[0024] Figure 6 This is a comparison chart of the optimal scheduling plan curve under rainy day conditions; DETAILED DESCRIPTION

[0025] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when detailed descriptions of known functions and designs may dilute the main content of the present invention, such descriptions will be omitted here.

[0026] Example

[0027] In this embodiment, if Figure 1 As shown, the present invention provides an optimization method for a photovoltaic power station to participate in power day-ahead dispatching, comprising the following steps:

[0028] S1. Collect historical annual data of the target PV power station, including meteorological data and output data. Combine the historical annual PV output of the target PV power station and the corresponding meteorological information on a daily basis and cluster them using the K-Means clustering algorithm to obtain three typical meteorological scenarios: sunny, cloudy, and rainy.

[0029] S101. Obtain historical photovoltaic output of a target photovoltaic power station within one year and meteorological information at corresponding times, and perform normalization processing; wherein the meteorological information includes but is not limited to temperature, humidity, wind speed, and solar irradiance;

[0030] S102, the normalized historical photovoltaic output and meteorological information are matched according to the timestamp to form a sample vector, where the timestamp is measured in days, where the i-th group of sample vectors x i Expressed as:

[0031] x i =[r mean ,r std ,r max ,t mean ,h mean ]

[0032] Among them, r mean is the daily average of total radiation, r std is the standard deviation of the total radiation on that day, r max is the maximum value of total radiation on that day, t mean is the daily average ambient temperature, h mean is the daily average relative humidity;

[0033] S103. In this embodiment, the number of clusters k is set to 3, which are sunny, cloudy and rainy days respectively; the K-Means algorithm is used to classify all sample vectors x i Perform clustering, output the cluster center and the category to which each sample vector belongs, and obtain three cluster categories.

[0034] S2. Use the particle swarm optimization-gradient boosting tree prediction model to predict photovoltaic output for different categories of typical meteorological scenarios, and then calculate the prediction error;

[0035] S201, traverse the sample vectors in each typical meteorological scene throughout the year, input the meteorological information corresponding to each sample vector into the particle swarm optimization-gradient boosting tree prediction model, and output the corresponding predicted photovoltaic output; in this embodiment, the particle swarm optimization-gradient boosting tree prediction model is an existing model, and its processing flow is as follows Figure 2 As shown;

[0036] S202, calculate the difference between the predicted photovoltaic output and the corresponding historical photovoltaic output to obtain the prediction error e, and select a representative weather day prediction result in three typical meteorological scenarios as follows: Figure 3 As shown, Figure 3 (a) is the comparison between the predicted photovoltaic output and the actual processing under sunny conditions. Figure 3 (b) is the comparison between the predicted photovoltaic output and the actual processing under cloudy conditions. Figure 3 (c) is the comparison between the predicted photovoltaic output and the actual processing under rainy conditions. Figure 3 In (a), we can observe that the predicted output under sunny weather is very close to the actual output, indicating that the overall uncertainty of the error is small. Figure 3 (b) and Figure 3 (c) In cloudy and rainy weather, the predicted output deviates greatly from the actual output, and the uncertainty of the prediction error is large, indicating that the uncertainty of power prediction increases significantly under non-ideal meteorological conditions, and large prediction errors are more likely to occur, reflecting the necessity of clustering typical meteorological scenarios in the method of the present invention.

[0037] S3. Divide the prediction errors of samples in three typical meteorological scenarios into hourly time periods to form an error sample set, and then use the kernel density estimation method to construct the probability density function of the prediction error;

[0038] S301. Divide the prediction errors of samples in three typical meteorological scenarios into hourly time periods, where the prediction error of the i-th group of samples under the c-th category at time t is expressed as The categories are divided into three categories c∈{1,2,3}, and the time is divided into 13 time periods t∈{1,2,…,13};

[0039] S302: The prediction errors of all samples in the cth typical meteorological scene at time t constitute an error sample set n c is the number of samples in the cth typical meteorological scene;

[0040] S303, constructing an error sample set matrix E;

[0041] Finally, all conditional error subsets ε c,t Form a two-dimensional error sample subset matrix:

[0042]

[0043] S304, for each sample subset ε c,t , and use the kernel density estimation method to fit its probability density function:

[0044]

[0045] Among them, hc is the kernel function bandwidth. In this embodiment, the kernel function bandwidth is selected as σ c is the standard deviation, e is the function variable, K() is the kernel function, and in this embodiment, the kernel function selects the standard Gaussian kernel function:

[0046]

[0047] S4. Introducing conditional value at risk (CVaR) as a control indicator to build a multi-scenario stochastic scheduling optimization model that covers output uncertainty;

[0048] S401, set the prediction error interval to [-50, 50], indicating the upper and lower limits of the prediction error; uniformly sample M = 100 error grid points within the prediction error interval [-50, 50], where the jth error grid point is represented by e j ,j=1,2,…,100;

[0049] S402, calculate each e j The approximate cumulative distribution function value F corresponding to each typical meteorological scenario c,t (e j );

[0050]

[0051] Wherein, Δe represents the grid step size, and in this embodiment, Δe=0.25;

[0052] S403. Construct an approximate inverse function of the cumulative distribution function using linear interpolation:

[0053]

[0054] Among them, u is the function variable;

[0055] S404, approximating the inverse function by Latin hypercube sampling Sampling is performed to obtain the error scenario under the cth typical meteorological scenario N represents the total number of all error scenarios. In this embodiment, N is 1000.

[0056] Finally, all the error scenarios Form an error scenario matrix under a typical meteorological scenario of type C:

[0057]

[0058] S405, record the photovoltaic output predicted by the particle swarm optimization-gradient boosting tree prediction model under the c-th typical meteorological scenario as represents the predicted photovoltaic output at time t under the cth typical meteorological scenario;

[0059] The error scenario ζ c Superimposed with the predicted photovoltaic output Get the photovoltaic output scenario matrix:

[0060]

[0061] Each of these lines Corresponding to a complete photovoltaic output scenario, each column represents the output of different scenarios under a specific period of time t. A total of 1000 photovoltaic output data under different error scenarios were obtained.

[0062] S406. Construct a dispatch objective function that reflects the deviation penalty and dispatchability constraint. The dispatch objective function under the τ-th photovoltaic output scenario is:

[0063]

[0064] Among them, q t is the dispatch declaration quantity in period t, p t is the electricity price in the tth period, α is the unit deviation penalty coefficient; in this embodiment, p t Set to 0.5; α is set to 1.5;

[0065] S407. Introduce the conditional value at risk indicator CVaR into the construction of the scheduling objective function to construct an optimization objective function that takes into account both expected performance and risk constraints:

[0066]

[0067] Among them, [η-R (τ) (q t )] + =max(0,η-R (τ) (q t )), indicating that only η-R (τ) (q t ) is greater than 0, and the sum is taken when η-R is less than 0. (τ) (q t ) takes the value of 0;

[0068] The conditional risk value indicator is introduced to control the adverse effects of extreme deviations. Furthermore, in order to facilitate the solution in the optimization model, a non-negative auxiliary variable z is introduced. (τ) To linearize the objective function CVaR β (x), then the linearized objective function is:

[0069]

[0070] The constraints are:

[0071]

[0072] Among them, η represents the risk threshold; δ is the scheduling deviation tolerance parameter; β is the confidence level, which is an adjustable parameter used to characterize the scheduling scheme's tolerance to extreme situations, and its value range is between 0 and 1.

[0073] S5. The multi-scenario random scheduling optimization model is solved through the whale optimization algorithm (CWOA) to obtain the optimal day-ahead scheduling declaration strategy under different typical weather conditions and time periods.

[0074] Below we describe the solution process of the whale optimization algorithm:

[0075] S501. Initialize the parameters of the whale optimization algorithm. In this embodiment, the population size is set to 300, the maximum number of iterations is 50, the convergence threshold is 0.1, and the initial population is generated by the chaotic mapping method to enhance the diversity of the population.

[0076] S502. Set the optimization objective function as the expected performance maximization model with CVaR risk constraints. Calculate the expected scheduling performance and tail risk index under multiple scenarios for each individual according to their corresponding scheduling application plan as the basis for fitness evaluation.

[0077] S503. Iterate the positions of the individuals in the population according to the update mechanism of the whale optimization algorithm, including the shrinking and encircling mechanism, the spiral update position, and the chaotic perturbation mechanism.

[0078] S504: Determine whether the termination condition (such as reaching the maximum number of iterations or the convergence threshold) is met. If so, output the current optimal individual as the final scheduling declaration strategy; otherwise, return to S502 and continue the iterative optimization process;

[0079] In this embodiment, three representative sunny days, cloudy days, and rainy days are selected to finally obtain the final dispatch declaration strategy, among which, Figure 4 Figure 2 shows the optimal scheduling curve generated by the optimization model under typical sunny conditions. The results show that when sunlight resources are relatively stable, the generated scheduling strategy is smoother, better fits the predicted output characteristics, and has good traceability and plan execution.

[0080] Figure 5 This is the optimal dispatch plan curve for typical cloudy weather conditions. As can be seen, compared to sunny days, the dispatch curve exhibits more pronounced adjustment fluctuations and deviates somewhat from the predicted output curve, demonstrating the model's ability to respond to increased output uncertainty. The optimization results more fully account for forecast error characteristics, improving the robustness of the strategy.

[0081] Figure 6 The optimal dispatch plan curve generated under typical rainy weather conditions is shown. Under these conditions, output fluctuates dramatically. The optimization model significantly reduces the risk exposure range through the CVaR mechanism, resulting in a more conservative and robust dispatch plan. This effectively reduces potential deviation risks and demonstrates the model's risk mitigation capabilities under extreme weather conditions.

[0082] Although the above describes the illustrative specific embodiments of the present invention to facilitate understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concepts of the present invention are protected.

Claims

1. A method for optimizing photovoltaic power plants participating in day-ahead power dispatching, characterized in that: The following steps are involved: S1. Combine the historical photovoltaic output of the target photovoltaic power station and the corresponding meteorological information, and cluster them using the K-Means clustering algorithm to obtain typical meteorological scenarios of different categories; S2. Use the particle swarm optimization-gradient boosting tree prediction model to predict photovoltaic output for different categories of typical meteorological scenarios, and then calculate the prediction error; S3. Divide the prediction errors of samples in each typical meteorological scenario into hourly time periods to form an error sample set, and then use the kernel density estimation method to construct the probability density function of the prediction error; S4. Introducing conditional value at risk (CVaR) as a control indicator to build a multi-scenario stochastic scheduling optimization model that covers output uncertainty; S5. The multi-scenario random scheduling optimization model is solved through the improved whale optimization algorithm CWOA to obtain the optimal day-ahead scheduling declaration strategy under different typical weather conditions and time periods.

2. The optimization method for photovoltaic power station participation in day-ahead power dispatching according to claim 1, characterized in that: The step S1 specifically includes: S101. Obtain historical photovoltaic output of a target photovoltaic power station within a preset period and meteorological information at the corresponding time, and perform normalization processing; wherein the meteorological information includes but is not limited to temperature, humidity, wind speed, and solar irradiance; S102, the normalized historical photovoltaic output and meteorological information are matched according to the timestamp to form a sample vector, where the timestamp is measured in days, where the i-th group of sample vectors x i Expressed as: x i =[r mean ,r std ,r max ,t mean ,h mean ] Among them, r mean is the daily average of total radiation, r std is the standard deviation of the total radiation on that day, r max is the maximum value of total radiation on that day, t mean is the daily average ambient temperature, h mean is the daily average relative humidity; S103, set the number of clusters k, and use the K-Means algorithm to classify all sample vectors x i Perform clustering and output the cluster center and the category to which each sample vector belongs. Each category corresponds to a typical meteorological scene.

3. The optimization method for photovoltaic power station participation in day-ahead power dispatching according to claim 1, characterized in that: The step S2 specifically includes: S201, traversing the sample vectors in each typical meteorological scenario, inputting the meteorological information corresponding to each sample vector into the particle swarm optimization-gradient boosting tree prediction model, and outputting the corresponding predicted photovoltaic output; S202: Calculate the difference between the predicted photovoltaic output and the corresponding historical photovoltaic output to obtain a prediction error e.

4. The optimization method for photovoltaic power station participation in day-ahead power dispatching according to claim 1, characterized in that: The step S3 specifically includes: S301. Divide the prediction error of samples in each typical meteorological scenario into hourly time periods, where the prediction error of the i-th group of samples under the c-th category at time t is expressed as c∈{1,2,…,k}, t∈{1,2,…,T}, T is the number of divided time periods; S302: The prediction errors of all samples in the cth typical meteorological scene at time t constitute an error sample set n c is the number of samples in the cth typical meteorological scene; S303, constructing an error sample set matrix E; Finally, all conditional error subsets ε c,t Form a two-dimensional error sample subset matrix: S304, for each sample subset ε c,t , and use the kernel density estimation method to fit its probability density function: Among them, h c is the kernel function bandwidth, e is the function variable, and K() is the kernel function.

5. The optimization method for photovoltaic power station participation in day-ahead power dispatching according to claim 1, characterized in that: The step S4 specifically includes: S401, set the prediction error interval [e min ,e max ], e min ,e max Indicates the upper and lower limits of the forecast error; in the forecast error interval [e min ,e max ] uniformly sample M error grid points within the grid, where the jth error grid point is represented by e j ,j=1,2,…,M; S402, calculate each e j The approximate cumulative distribution function value F corresponding to each typical meteorological scenario c,t (e j ); Where Δe represents the grid step size; S403. Construct an approximate inverse function of the cumulative distribution function using linear interpolation: Among them, u is the function variable; S404, approximating the inverse function by Latin hypercube sampling Sampling is performed to obtain the error scenario under the cth typical meteorological scenario τ=1,2,…,N, where N represents the total number of all error scenarios; Finally, all the error scenarios Form an error scenario matrix under a typical meteorological scenario of type C: S405, record the photovoltaic output predicted by the particle swarm optimization-gradient boosting tree prediction model under the c-th typical meteorological scenario as represents the predicted photovoltaic output at time t under the cth typical meteorological scenario; The error scenario ζ c Superimposed with the predicted photovoltaic output Get the photovoltaic output scenario matrix: Each row It is a complete photovoltaic output scenario, and each column represents the output of different scenarios in a period t; S406. Construct a dispatch objective function that reflects the deviation penalty and dispatchability constraint. The dispatch objective function under the τ-th photovoltaic output scenario is: Among them, q t is the dispatch declaration quantity in period t, p t is the electricity price in period t, α is the unit deviation penalty coefficient; S407. Introduce the conditional value at risk indicator CVaR into the construction of the scheduling objective function to construct an optimization objective function that takes into account both expected performance and risk constraints: Among them, [η-R (τ) (q t )] + =max(0,η-R (τ) (q t )), indicating that only η-R (τ) (q t ) is greater than 0, and the sum is taken when η-R is less than 0. (τ) (q t ) takes the value of 0; Furthermore, in order to facilitate the solution in the optimization model, a non-negative auxiliary variable z is introduced (τ) To linearize the objective function CVaR β (x), then the linearized objective function is: The constraints are: Where η represents the risk threshold; δ is the scheduling deviation tolerance parameter, and β is the confidence level.

Images