Virtual power plant control method based on multi-objective optimization algorithm

Abstract

The invention provides a virtual power plant control method based on a multi-objective optimization algorithm. The virtual power plant control method comprises the steps of collecting real-time operation data of distributed energy equipment through a virtual power plant control center; establishing a dynamic prediction model, and predicting output fluctuation characteristics and load change characteristics of each distributed energy device in a future time period; constructing a multi-time scale collaborative optimization model; solving the multi-time-scale collaborative optimization model by adopting an improved chaotic particle swarm algorithm, and generating an equipment control parameter set; dynamically adjusting a weight coefficient matrix of the optimization model according to the deviation degree of the real-time operation data and the prediction data; the robustness of the control parameters is verified through Monte Carlo simulation, and the next step is executed when the verification passing rate is larger than or equal to 95%; and issuing a control instruction set including a time sequence control instruction and a power adjustment threshold to each distributed energy device. According to the invention, efficient scheduling and precise control of the virtual power plant can be realized, and the power tracking precision, economical efficiency and stability of distributed energy can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of virtual power plants, and more specifically, to a virtual power plant control method based on a multi-objective optimization algorithm. Background Art

[0002] With the acceleration of energy transformation, the proportion of distributed energy in the power system continues to increase. As a flexible scheduling platform that integrates distributed energy, virtual power plants can effectively coordinate power generation, energy storage, and load resources, improving energy utilization efficiency and grid operation stability. However, the intermittent and uncertain nature of distributed energy poses challenges to the control of virtual power plants. Existing virtual power plant control methods mostly use single-time scale optimization, which makes it difficult to take into account the day-ahead, intraday, and real-time scheduling requirements. The optimization algorithm also has problems such as slow convergence and easy falling into local optimality. In addition, traditional methods lack adaptability to uncertain factors, resulting in poor robustness of the control strategy.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the existing virtual power plant control method cannot effectively balance power tracking accuracy, economy and stability, and when facing distributed energy output fluctuations and load changes, the dynamic adjustment capability of the control strategy is insufficient, making it difficult to meet the optimization scheduling requirements under complex working conditions. Summary of the Invention

[0004] The present invention provides a virtual power plant control method based on a multi-objective optimization algorithm, comprising:

[0005] S1. Collect real-time operating data of distributed energy equipment through the virtual power plant control center, including power generation power, energy storage capacity, load demand, and grid dispatch instructions;

[0006] S2. Establish a dynamic prediction model to predict the output fluctuation characteristics and load change characteristics of each distributed energy device in the future T period;

[0007] S3. Construct a multi-timescale collaborative optimization model, the model including a day-ahead optimization layer, an intraday rolling layer, and a real-time adjustment layer;

[0008] S4. Using an improved chaotic particle swarm optimization algorithm to solve the multi-time scale collaborative optimization model and generate a set of device control parameters;

[0009] S5. Dynamically adjust the weight coefficient matrix of the optimization model based on the deviation between the real-time operation data and the predicted data;

[0010] S6. Verify the robustness of the control parameters through Monte Carlo simulation. When the verification pass rate is ≥95%, execute step S7;

[0011] S7. Issue a control instruction set including timing control instructions and power regulation thresholds to each distributed energy device.

[0012] Furthermore, the step S4 includes:

[0013] S41. Initialize the particle swarm parameters, set the inertia factor ω = 0.729, the acceleration constants c1 = 2.05, c2 = 2.05;

[0014] S42. Use Tent chaotic mapping to generate the initial particle position matrix X = [x1, x2, ..., x n ] T , where x i ∈[0,1] D , D is the dimension of optimization variables, n is the number of particles;

[0015] S43. Calculate the fitness function value F(X) = α·f1+β·f2+γ·f3, where α is the power tracking weight coefficient, β is the economy weight coefficient, and γ is the stability weight coefficient, and α+β+γ=1;

[0016]

[0017] represents the power tracking error, where P act (t) is the actual power at time t, P pre (t) is the predicted power at time t, and T is the total number of time periods;

[0018]

[0019] Represents the economic deviation, where Cost real (t) is the actual operating cost, Cost min is the theoretical minimum cost;

[0020]

[0021] represents the power fluctuation penalty term, where ΔP(t)=|P(t)-P(t-1)|, P max is the maximum allowable power of the equipment;

[0022] S44. Update particle velocity and position through dynamic inertia weight adjustment strategy;

[0023] S45. When max(|F(X (k) )-F(X (k-1) )|)≤ε or the number of iterations k≥K max Stop the iteration and output the optimal solution set, where ε=1e-6, K max =500, k is the current number of iterations.

[0024] Furthermore, the dynamic inertia weight adjustment strategy in step S44 satisfies:

[0025]

[0026] Among them, ω initial =0.9,ω final =0.4, λ is the nonlinear adjustment index and satisfies k is the current iteration number, K max is the maximum number of iterations.

[0027] Furthermore, the step S2 includes:

[0028] S21. Perform wavelet packet decomposition on the historical operation data to obtain n intrinsic mode components IMF1~IMF n ;

[0029] S22. Establish a deep belief network (DBN) prediction model with a hidden layer structure of [128, 64, 32] and the number of input layer nodes equal to the number of IMF components n.

[0030] S23. Use the improved Adam optimizer to train the network parameters, and the learning rate η is Decay, where η0 = 0.001, τ = 1000, and k is the number of training steps;

[0031] S24. Predict each IMF component separately and reconstruct it to obtain the final prediction result.

[0032] Furthermore, the parameter update formula of the improved Adam optimizer is:

[0033] m t =μ·m t-1 +(1-μ)·g t

[0034]

[0035] Among them, μ = 0.9 is the momentum decay rate, v = 0.999 is the second-order moment decay rate, δ = 1e-8 is the numerical stability constant, g t is the gradient vector at time t, sign() is the sign function, output is ±1, and η is the learning rate.

[0036] Furthermore, the dynamic adjustment strategy in step S5 includes:

[0037] S51. Calculate the deviation between real-time data and predicted data:

[0038]

[0039] Among them, P real (t) is the actual power, P pre (t) is the predicted power, P rated is the rated power of the equipment;

[0040] S52. When ΔD(t)≥ΔD threshold When , press β←β+κ·(ΔD(t)-ΔD threshold ) Update the economic weight coefficient while maintaining α+β+γ=1, where κ=0.05, ΔD threshold =0.15;

[0041] S53. The adjusted weight coefficients must satisfy the constraints: α∈[0.3,0.6], β∈[0.2,0.5], γ∈[0.1,0.3].

[0042] Furthermore, the objective function of the multi-time-scale collaborative optimization model in step S3 is:

[0043]

[0044] Constraints include:

[0045] 1) Output limit:

[0046] 2) Power balance:

[0047] 3) Climbing limit: |P i (t)-P i (t-1)|≤ΔP ramp,i , where: C gen (t) is the power generation cost, C curt (t) is the cost of abandoned energy, C reg (t) is the adjustment cost, P i (t) is the output of device i in period t, I is the set of devices, P load (t) is the load power at time t, P loss (t) is the power loss at time t, ΔP ramp,i is the maximum allowed ramp rate of device i.

[0048] Furthermore, the power generation cost C gen The calculation formula for (t) is:

[0049]

[0050] Among them, a i is the quadratic cost coefficient, b i is the linear cost coefficient, c iis the fixed cost coefficient, d i and σ i is a nonlinear cost parameter, and N is the total number of distributed energy devices.

[0051] Furthermore, step S6 includes:

[0052] S61. Generate M = 1000 sets of randomly perturbed samples that follow a normal distribution N(μ,∑), where μ is the predicted mean and ∑ is the covariance matrix;

[0053] S62. Perform forward simulation on each sample to calculate the objective function value F. m ;

[0054] S63. Calculation of robustness indicators

[0055]

[0056] Among them, I() is the indicator function, when F m ≤F max 1 when it is set, otherwise 0, F max =1.2·F opt , F opt To optimize the target value;

[0057] S64. When R≥0.95, the verification is determined to be passed.

[0058] Furthermore, the control instruction set generation method in step S7 includes:

[0059] S71. Decompose the optimization results in time and space to generate the time-power curve P of each device ref (t);

[0060] S72. Add feedforward compensation

[0061]

[0062] Where, e(t) = P ref (t)-P act (t), K p =0.8 is the proportional coefficient, K i =0.2 is the integration coefficient;

[0063] S73. Smooth the instruction set to ensure that the second-order derivative constraint is satisfied:

[0064]

[0065] in, P rated is the rated power of the device.

[0066] The above-described embodiments of the present invention have at least the following beneficial effects: By constructing a multi-timescale collaborative optimization model and combining a three-layer structure of day-ahead optimization, intraday rolling, and real-time adjustment, the present invention can effectively address the operational requirements of virtual power plants at different timescales and achieve refined scheduling of distributed energy devices. Furthermore, the improved chaotic particle swarm optimization algorithm employed to solve the optimization model can rapidly converge to a global optimal solution, significantly improving optimization efficiency and accuracy and ensuring efficient and stable operation of the virtual power plant under complex operating conditions.

[0067] Furthermore, by dynamically adjusting the optimization model's weight coefficient matrix, the present invention allows flexible adjustments to the control strategy based on the deviation between real-time operating data and predicted data, further enhancing the system's adaptability and robustness. Combined with Monte Carlo simulation to verify the robustness of control parameters, this approach ensures that the optimization results meet actual operational requirements despite uncertainties, thus providing reliable guarantees for the safe, economical, and stable operation of virtual power plants. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0069] Figure 1 A flowchart of a virtual power plant control method based on a multi-objective optimization algorithm provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0070] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0071] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0072] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0073] Reference below Figure 1 , Figure 1This is a flow chart of a virtual power plant control method based on a multi-objective optimization algorithm provided by an embodiment of the present invention. Figure 1 As shown, a virtual power plant control method 100 based on a multi-objective optimization algorithm includes:

[0074] S1. Collect real-time operating data of distributed energy equipment through the virtual power plant control center, including power generation power, energy storage capacity, load demand, and grid dispatch instructions;

[0075] S2. Establish a dynamic prediction model to predict the output fluctuation characteristics and load change characteristics of each distributed energy device in the future T period;

[0076] S3. Construct a multi-timescale collaborative optimization model, the model including a day-ahead optimization layer, an intraday rolling layer, and a real-time adjustment layer;

[0077] S4. Using an improved chaotic particle swarm optimization algorithm to solve the multi-time scale collaborative optimization model and generate a set of device control parameters;

[0078] S5. Dynamically adjust the weight coefficient matrix of the optimization model based on the deviation between the real-time operation data and the predicted data;

[0079] S6. Verify the robustness of the control parameters through Monte Carlo simulation. When the verification pass rate is ≥95%, execute step S7;

[0080] S7. Issue a control instruction set including timing control instructions and power regulation thresholds to each distributed energy device.

[0081] It should be noted that the present invention proposes a virtual power plant control method based on a multi-objective optimization algorithm. This method collects real-time operating data of distributed energy equipment through the virtual power plant control center. These data include power generation power, energy storage capacity, load demand, and grid dispatch instructions. Among them, power generation power refers to the actual output power of distributed energy equipment at a certain moment; energy storage capacity refers to the remaining power or energy storage capacity of the energy storage equipment; load demand refers to the real-time electricity demand of users in the area served by the virtual power plant; and grid dispatch instructions are power adjustment requirements issued by the grid operator to the virtual power plant based on the overall grid operation status. By collecting these data, accurate input information can be provided for subsequent optimization control to ensure that the operation strategy of the virtual power plant can reflect the current operation status and external requirements in real time.

[0082] Specifically, a dynamic prediction model is established to predict the output fluctuation characteristics and load variations of each distributed energy resource (DRE) device over future time periods. This model analyzes historical operating data and combines techniques such as wavelet packet decomposition and deep belief networks to provide high-precision forecasts of DRE power generation and load demand. Wavelet packet decomposition is a signal processing method that decomposes complex signals into multiple intrinsic mode components, thereby extracting signal characteristics. Deep belief networks are a prediction model based on deep learning. Using a multi-layer neural network structure to learn and fit data, they can effectively capture nonlinear relationships within the data. Furthermore, the multi-timescale collaborative optimization model comprises a day-ahead optimization layer, an intraday rolling layer, and a real-time adjustment layer, each corresponding to optimization requirements at different timescales. The day-ahead optimization layer primarily optimizes the operating plan for the next day; the intraday rolling layer makes short-term adjustments based on real-time operating conditions; and the real-time adjustment layer provides immediate control for emergencies or short-term fluctuations. This layered optimization structure enables refined scheduling of virtual power plants at different timescales.

[0083] Preferably, an improved chaotic particle swarm algorithm is used to solve the multi-time scale collaborative optimization model. When initializing the particle swarm parameters, the inertia factor is set to 0.729 and the acceleration constants are set to 2.05. The selection of these parameters is based on a large number of experimental verifications and can effectively balance the global search ability and local search ability of the algorithm. The calculation of the fitness function value comprehensively considers the power tracking error, economic deviation and power fluctuation penalty term, and balances these three goals by assigning different weight coefficients. For example, the power tracking error can be measured by calculating the difference between the actual power and the predicted power, the economic deviation is evaluated by comparing the difference between the actual operating cost and the theoretical minimum cost, and the power fluctuation penalty term is used to limit the rapid change of the equipment power to ensure the safe operation of the equipment. In addition, the dynamic inertia weight adjustment strategy can dynamically adjust the inertia factor according to the number of iterations to further improve the convergence speed and accuracy of the algorithm.

[0084] In some embodiments, step S4 includes:

[0085] S41. Initialize the particle swarm parameters, set the inertia factor ω = 0.729, the acceleration constants c1 = 2.05, c2 = 2.05;

[0086] S42. Use Tent chaotic mapping to generate the initial particle position matrix X = [x1, x2, ..., x n ] T , where x i ∈[0,1] D , D is the dimension of optimization variables, n is the number of particles;

[0087] S43. Calculate the fitness function value F(X) = α·f1+β·f2+γ·f3, where α is the power tracking weight coefficient, β is the economy weight coefficient, and γ is the stability weight coefficient, and α+β+γ=1;

[0088]

[0089] represents the power tracking error, where P act (t) is the actual power at time t, P pre (t) is the predicted power at time t, and T is the total number of time periods;

[0090]

[0091] Represents the economic deviation, where Cost real (t) is the actual operating cost, Cost min is the theoretical minimum cost;

[0092]

[0093] represents the power fluctuation penalty term, where ΔP(t)=|P(t)-P(t-1)|, P max is the maximum allowable power of the equipment;

[0094] S44. Update particle velocity and position through dynamic inertia weight adjustment strategy;

[0095] S45. When max(|F(X (k) )-F(X (k-1) )|)≤ε or the number of iterations k≥K max Stop the iteration and output the optimal solution set, where ε=1e-6, K max =500, k is the current number of iterations.

[0096] It should be noted that the implementation steps of the improved chaotic particle swarm algorithm in the present invention are one of the core links of the virtual power plant control method. The algorithm realizes efficient solution of the multi-time scale collaborative optimization model by initializing the particle swarm parameters, generating the initial particle position matrix, calculating the fitness function value, and dynamically adjusting the particle speed and position. Among them, the inertia factor and the acceleration constant are key parameters in the particle swarm algorithm, which directly affect the global search ability and convergence speed of the algorithm. Tent chaos mapping is a chaotic system used to generate the initial particle position matrix. Its randomness and ergodicity can enhance the initial population diversity of the algorithm and avoid falling into local optimality. The fitness function value is a criterion for measuring the quality of particles. It comprehensively considers the power tracking error, economic deviation and power fluctuation penalty term, and balances the relationship between different optimization objectives by allocating weight coefficients.

[0097] Specifically, the parameters of the improved chaotic particle swarm algorithm are set as follows: the inertia factor ω is set to 0.729, and the acceleration constants c1 and c2 are both set to 2.05. These parameters were selected based on extensive experimental verification and can effectively balance the algorithm's global and local search capabilities. The initial particle position matrix is generated using the Tent chaotic map. Its dimension is determined by the dimension d of the optimization variable and the number of particles n. The position value of each particle is randomly distributed in the interval [0, 1]. In the calculation of the fitness function value, the power tracking error is measured by comparing the difference between the actual power and the predicted power, and the economic deviation is evaluated by the difference between the actual operating cost and the theoretical minimum cost. The power fluctuation penalty term is used to limit rapid changes in device power and ensure safe operation of the equipment. In addition, a dynamic inertia weight adjustment strategy dynamically adjusts the inertia factor based on the current number of iterations, further improving the convergence speed and accuracy of the algorithm.

[0098] Preferably, when implementing the improved chaotic particle swarm algorithm, the following aspects can be further refined or optimized. First, for the generation of the initial particle position matrix, in addition to the Tent chaotic map, other chaotic systems, such as the Logistic chaotic map, can also be considered to enhance the diversity of the initial population. Secondly, in the weight distribution of the fitness function, the weight coefficient can be dynamically adjusted according to the actual operating needs of the virtual power plant. For example, in a scenario where economy is prioritized, the proportion of the economic weight coefficient can be appropriately increased. Finally, for the dynamic inertia weight adjustment strategy, an adaptive adjustment mechanism can be introduced to dynamically adjust the nonlinear adjustment index according to the diversity or convergence speed of the current population, thereby further improving the performance of the algorithm.

[0099] In some embodiments, the dynamic inertia weight adjustment strategy in step S44 satisfies:

[0100]

[0101] Among them, ω initial =0.9,ω final =0.4, λ is the nonlinear adjustment index and satisfies k is the current iteration number, K max is the maximum number of iterations.

[0102] It should be noted that the dynamic inertia weight adjustment strategy in the present invention is an important component of the improved chaotic particle swarm algorithm. This strategy balances the global search capability and local search capability of the particle swarm algorithm by dynamically adjusting the inertia factor during the iteration process. The inertia factor is an important parameter in the particle swarm algorithm that determines the speed at which particles move in the search space. By dynamically adjusting the inertia factor, the search strategy can be flexibly adjusted at different stages of the algorithm, thereby improving the convergence speed and accuracy of the algorithm. The initial inertia factor ωmax and the final inertia factor ωmin are set to 0.9 and 0.4, respectively. The selection of these parameters is based on a large number of experimental verifications and can effectively balance the global search capability and local search capability of the algorithm.

[0103] Specifically, the formula for the dynamic inertia weight adjustment strategy is: ω(t) = ωmax-(ωmax-ωmin)*(t / Tmax) β , where ωmax is the initial inertia factor, ωmin is the final inertia factor, t is the current iteration number, Tmax is the maximum iteration number, and β is the nonlinear adjustment exponent. The choice of the nonlinear adjustment exponent β has a significant impact on the variation curve of the inertia factor. It is usually set to 1 + 0.5 * exp(t / Tmax) to achieve a nonlinear decrease in the inertia factor. This dynamic adjustment strategy maintains a large inertia factor in the early stages of the algorithm, enhancing global search capabilities and avoiding local optima. In the later stages of the algorithm, the inertia factor is gradually reduced to enhance local search capabilities and improve solution accuracy.

[0104] Preferably, when implementing the dynamic inertia weight adjustment strategy, the following aspects can be further refined or optimized. First, the value range of the initial inertia factor and the final inertia factor can be adjusted according to the specific optimization problem and algorithm performance requirements. For example, in scenarios with high global search capability requirements, the value of the initial inertia factor can be appropriately increased. Second, an adaptive adjustment mechanism can be introduced to dynamically adjust the nonlinear adjustment index β according to the diversity or convergence speed of the current population, thereby further improving the performance of the algorithm. Finally, other optimization strategies can be combined, such as introducing a local search algorithm or a hybrid optimization algorithm, to further improve the solution efficiency and accuracy of the improved chaotic particle swarm algorithm.

[0105] In some embodiments, step S2 includes:

[0106] S21. Perform wavelet packet decomposition on the historical operation data to obtain n intrinsic mode components IMF1~IMF n ;

[0107] S22. Establish a deep belief network (DBN) prediction model with a hidden layer structure of [128, 64, 32] and the number of input layer nodes equal to the number of IMF components n.

[0108] S23. Use the improved Adam optimizer to train the network parameters, and the learning rate η is Decay, where η0 = 0.001, τ = 1000, and k is the number of training steps;

[0109] S24. Predict each IMF component separately and reconstruct it to obtain the final prediction result.

[0110] It should be noted that the establishment of a dynamic prediction model in the present invention is a key link in achieving precise control of virtual power plants. The model processes historical operating data through wavelet packet decomposition, extracts intrinsic mode components, and combines deep belief networks (DBNs) for prediction. Wavelet packet decomposition is a signal processing technology that can decompose complex signals into multiple frequency components, thereby extracting characteristic information of the signal. Deep belief networks are a prediction model based on deep learning that has powerful nonlinear fitting capabilities and can make high-precision predictions of output fluctuations and load changes of distributed energy sources. Through this combined method, the accuracy and reliability of the prediction can be effectively improved, providing a solid data foundation for subsequent optimization control.

[0111] Specifically, the implementation process of the dynamic prediction model includes the following steps: First, wavelet packet decomposition is performed on historical operating data to extract multiple intrinsic mode components (EMDs). These components can reflect the output fluctuation characteristics and load variation characteristics of distributed energy devices. Then, a deep belief network prediction model is established with a hidden layer structure of [128, 64, 32] and the number of input layer nodes equal to the number of components after wavelet packet decomposition. The network parameters are trained using an improved Adam optimizer, and the learning rate is adjusted according to the exponential decay law. The initial learning rate is set to 0.001 and the decay step size is 1000. This method effectively improves the model training efficiency and prediction accuracy. Finally, by predicting and reconstructing each intrinsic mode component separately, the final prediction results are obtained, providing accurate input data for the optimal scheduling of the virtual power plant.

[0112] Preferably, when implementing a dynamic prediction model, the following aspects can be further refined or optimized. First, the number of wavelet packet decomposition layers can be adjusted based on the complexity of the actual data to better extract the characteristic information of the signal. Second, the hidden layer structure of the deep belief network can be optimized based on the complexity of the prediction task, for example, by increasing or decreasing the number of hidden layer nodes to improve the model's fitting ability. In addition, the learning rate decay strategy of the improved Adam optimizer can also be dynamically adjusted based on the convergence status during training, for example, by introducing an adaptive learning rate adjustment mechanism to further improve the training efficiency of the model.

[0113] In some embodiments, the parameter update formula of the improved Adam optimizer is:

[0114] m t =μ·m t-1 +(1-μ)·g t

[0115]

[0116] Among them, μ = 0.9 is the momentum decay rate, v = 0.999 is the second-order moment decay rate, δ = 1e-8 is the numerical stability constant, g t is the gradient vector at time t, sign() is the sign function, output is ±1, and η is the learning rate.

[0117] It should be noted that the dynamic adjustment strategy in the present invention is an important part of the virtual power plant control method, which is used to dynamically adjust the weight coefficient matrix of the optimization model according to the deviation between real-time operating data and predicted data. This adjustment mechanism can ensure the flexibility and adaptability of the virtual power plant control strategy to cope with the uncertainty of distributed energy output and the dynamic changes in load demand. Among them, the deviation degree is a key indicator to measure the difference between actual operating data and predicted data. By calculating the deviation degree and adjusting the weight coefficient accordingly, the control target of the virtual power plant can be optimized to achieve a better balance between power tracking, economy and stability.

[0118] Specifically, the implementation process of the dynamic adjustment strategy is as follows:

[0119] First, calculate the deviation between real-time data and predicted data. The formula is:

[0120] ΔP(t)=||P actual (t)-P predicted (t)||2 / P rated

[0121] Among them, P actual (t) is the actual power, P predicted (t) is the predicted power, P rated is the rated power of the equipment. When the deviation exceeds the preset threshold (for example, 0.15), the economic weight coefficient is adjusted according to certain rules. For example, if the deviation is large, it means that there is a significant difference between the actual operation and the prediction. At this time, the economic weight coefficient needs to be increased to reduce the operating cost while keeping the sum of the weight coefficients at 1. The adjusted weight coefficients must also meet the constraints, such as the power tracking weight coefficient α∈[0.3,0.6], the economic weight coefficient β∈[0.2,0.5], and the stability weight coefficient γ∈[0.1,0.3]. Through this dynamic adjustment mechanism, the control strategy of the virtual power plant can be flexibly adjusted according to the actual operating conditions, so as to better adapt to complex working conditions.

[0122] Preferably, when implementing the dynamic adjustment strategy, the following aspects can be further refined or optimized. First, the calculation of the deviation degree can introduce data in more dimensions, such as considering factors such as power fluctuation rate or cost deviation, to more comprehensively reflect the difference between actual operation and prediction. Secondly, the adjustment rules of the weight coefficient can be customized according to the specific operating objectives of the virtual power plant. For example, in scenarios with high stability requirements, the adjustment range of the stability weight coefficient can be appropriately increased. In addition, an adaptive adjustment mechanism can be introduced to dynamically adjust the adjustment step size of the threshold and weight coefficient according to historical data and real-time operating conditions, thereby further improving the adaptability and flexibility of the control strategy.

[0123] In some embodiments, the dynamic adjustment strategy in step S5 includes:

[0124] S51. Calculate the deviation between real-time data and predicted data:

[0125]

[0126] Among them, P real (t) is the actual power, P pre (t) is the predicted power, P rated is the rated power of the equipment;

[0127] S52. When ΔD(t)≥ΔD threshold When , press β←β+κ·(ΔD(t)-ΔD threshold ) Update the economic weight coefficient while maintaining α+β+γ=1, where κ=0.05, ΔD threshold =0.15;

[0128] S53. The adjusted weight coefficients must satisfy the constraints: α∈[0.3,0.6], β∈[0.2,0.5], γ∈[0.1,0.3].

[0129] It should be noted that the dynamic adjustment strategy in the present invention is used to dynamically adjust the weight coefficient matrix of the optimization model based on the deviation between real-time operating data and predicted data. Deviation is a key indicator to measure the difference between actual operating data and predicted data. By calculating the deviation and adjusting the weight coefficient accordingly, the control objectives of the virtual power plant can be optimized to achieve a better balance between power tracking, economy and stability. Specifically, the calculation formula for deviation is:

[0130] ΔP(t)=||P actual (t)-P predicted (t)||2 / P rated

[0131] Among them, P actual (t) is the actual power, Ppredicted (t) is the predicted power, P rated is the rated power of the device.

[0132] Specifically, when the deviation ΔP(t) exceeds a preset threshold (e.g., 0.15), the economic weight coefficient β is adjusted according to certain rules, while keeping the sum of the power tracking weight coefficient α, the economic weight coefficient β, and the stability weight coefficient γ at 1. The adjusted weight coefficients must meet the constraints: α∈[0.3,0.6], β∈[0.2,0.5], γ∈[0.1,0.3]. For example, when ΔP(t)≥0.15, the economic weight coefficient β is updated according to β←β+k*(ΔP(t)-0.15), where k is the adjustment step size, usually set to 0.05. Through this dynamic adjustment mechanism, the control strategy of the virtual power plant can be flexibly adjusted according to the actual operating conditions, thereby better adapting to complex operating conditions.

[0133] Preferably, when implementing the dynamic adjustment strategy, the following aspects can be further refined or optimized. First, the calculation of the deviation degree can introduce data of more dimensions, such as considering factors such as power fluctuation rate or cost deviation, so as to more comprehensively reflect the difference between actual operation and prediction. Secondly, the adjustment rules of the weight coefficient can be customized according to the specific operating objectives of the virtual power plant. For example, in scenarios with high stability requirements, the adjustment range of the stability weight coefficient γ can be appropriately increased. In addition, an adaptive adjustment mechanism can be introduced to dynamically adjust the adjustment step of the threshold and weight coefficient according to historical data and real-time operating conditions, thereby further improving the adaptability and flexibility of the control strategy.

[0134] In some embodiments, the objective function of the multi-time-scale collaborative optimization model in step S3 is:

[0135]

[0136] Constraints include:

[0137] 1) Output limit:

[0138] 2) Power balance:

[0139] 3) Climbing limit: |P i (t)-P i (t-1)|≤ΔP ramp,i , where: C gen (t) is the power generation cost, C curt (t) is the cost of abandoned energy, C reg (t) is the adjustment cost, P i (t) is the output of device i in period t, I is the set of devices, Pload (t) is the load power at time t, P loss (t) is the power loss at time t, ΔP ramp,i is the maximum allowed ramp rate of device i.

[0140] It should be noted that the objective function and constraints of the multi-time-scale collaborative optimization model in the present invention are the core components of the virtual power plant control method. The model realizes the refined scheduling of virtual power plants at different time scales by comprehensively considering multi-objective optimization problems such as power tracking, economy and stability. The construction of the objective function integrates the power generation cost, energy abandonment cost and regulation cost, aiming to achieve the economy and stability of the virtual power plant operation by optimizing the equipment output. The constraints include output limit, power balance and climbing limit. These constraints ensure that the equipment operates within a safe range while meeting the power demand of the power grid.

[0141] Specifically, the objective function of the multi-time-scale collaborative optimization model is:

[0142] ∑[α·C g eneration(t)+β·C c urtailment(t)+γ·C r egulation(t)]

[0143] Among them, α, β, and γ are the weight coefficients of power tracking, economy, and stability respectively, and C g Eneration(t) is the cost of power generation, C c Curtailment(t) is the cost of energy abandonment, C r Regulation(t) is the regulation cost. The calculation formula of power generation cost is: C g eneration(t)=∑[a i ·P i (t) 2 +b i ·P i (t)+c i +d i ·exp(e i ·P i (t))]

[0144] Among them, a i 、b i 、c i are quadratic, linear and fixed cost coefficients, d i and e i is the nonlinear cost parameter, P i (t) is the output of device i in period t. Constraints include: output limit P m in,i≤Pi8 (t)≤P m ax,i, power balance ∑P i8 (t) = P l oad(t)+P l oss(t), climbing limit|P i8 (t)-P i (t-1)|≤R m ax,i,These constraints ensure that the device output is within the allowable range while satisfying the power balance and device ramp rate limits.

[0145] Preferably, when implementing the multi-time scale collaborative optimization model, the following aspects can be further refined or optimized. First, the weight coefficient in the objective function can be dynamically adjusted according to the actual operating needs of the virtual power plant. For example, in a scenario where economy is prioritized, the proportion of the economic weight coefficient β can be appropriately increased. Secondly, the calculation of power generation costs can introduce more factors, such as equipment maintenance costs or environmental costs, to more comprehensively reflect the operating costs. In addition, the ramp limits in the constraints can be adjusted according to the actual characteristics of the equipment, such as introducing nonlinear ramp limits or considering the dynamic response characteristics of the equipment, thereby further improving the adaptability and flexibility of the model.

[0146] In some embodiments, the power generation cost C gen The calculation formula for (t) is:

[0147]

[0148] Among them, a i is the quadratic cost coefficient, b i is the linear cost coefficient, c i is the fixed cost coefficient, d i and σ i is a nonlinear cost parameter, and N is the total number of distributed energy devices.

[0149] It should be noted that the calculation formula for power generation cost in the present invention is an important component of the multi-time scale collaborative optimization model, which is used to accurately evaluate the operating cost of distributed energy equipment in virtual power plants. The calculation of power generation cost not only takes into account the power output of the equipment, but also combines the quadratic cost coefficient, linear cost coefficient, fixed cost coefficient and nonlinear cost parameter, which can fully reflect the economic efficiency of the equipment under different operating conditions. Through this refined cost calculation, the optimal scheduling of virtual power plants can more scientifically balance the economy and operating efficiency, providing strong support for the economic operation of virtual power plants.

[0150] Specifically, the calculation formula for power generation cost is:

[0151] C generation(t)=∑[a i ·P i (t) 2 +b i ·P i (t)+c i +d i ·exp(e i ·P i (t))]

[0152] Among them: a i 、b i 、c i are quadratic, linear, and fixed cost coefficients, respectively, used to describe the cost variation of the equipment at different power outputs;

[0153] d i and e i is a nonlinear cost parameter, which is used to consider the nonlinear cost increase that may occur when the equipment outputs high power; P i (t) is the output of device i in period t; n is the total number of distributed energy devices.

[0154] These parameters can be set according to the actual characteristics of the equipment, for example, the quadratic cost coefficient a i Usually related to the efficiency curve of the equipment, the linear cost coefficient b i Related to the operation and maintenance costs of the equipment, the fixed cost coefficient c j This segmented cost calculation can more accurately reflect the economic efficiency of the equipment under different operating conditions and provide a scientific basis for optimized scheduling.

[0155] Preferably, when implementing the power generation cost calculation, the following aspects can be further refined or optimized. First, for the nonlinear cost parameter d i and d i , a fitting can be performed based on the actual operating data of the equipment to more accurately reflect cost changes at high power output. Secondly, dynamic cost parameters for the equipment can be introduced, such as considering startup and shutdown costs or aging costs, to further improve the cost calculation model. Furthermore, for virtual power plants with a large number of distributed energy devices, group calculation can be used to classify the devices by type or power level and calculate their power generation costs separately, thereby improving computational efficiency and simplifying the optimization model.

[0156] In some embodiments, step S6 includes:

[0157] S61. Generate M = 1000 sets of randomly perturbed samples that follow a normal distribution N(μ,∑), where μ is the predicted mean and ∑ is the covariance matrix;

[0158] S62. Perform forward simulation on each sample to calculate the objective function value F. m ;

[0159] S63. Calculation of robustness indicators

[0160]

[0161] Among them, I() is the indicator function, when F m ≤F max 1 when it is set, otherwise 0, F max =1.2·F opt , F opt To optimize the target value;

[0162] S64. When R≥0.95, the verification is determined to be passed.

[0163] It should be noted that the robustness verification step in the present invention is a key link to ensure the reliable operation of the virtual power plant control strategy under uncertain conditions. Random disturbance samples are generated through Monte Carlo simulation, and these samples are forward simulated to calculate the objective function value, thereby evaluating the robustness of the control strategy. Monte Carlo simulation is a calculation method based on random sampling. It can simulate the uncertainty of the system through a large number of samples, thereby providing a quantitative evaluation of the reliability of the control strategy. The calculation of the robustness index is based on the tolerance range of the optimization target value to ensure that the control strategy can still achieve the expected operating effect under the influence of uncertain factors.

[0164] Specifically, the robustness verification process includes the following steps: First, 1,000 sets of normally distributed random perturbation samples are generated, with the mean being the predicted value and the covariance matrix being obtained based on historical data. These samples are used to simulate the uncertainty of distributed energy output and load demand. Second, a forward simulation is performed on each set of samples to calculate the objective function value, which reflects the operating cost or performance index of the system under different perturbations. Finally, the robustness index is calculated using the formula:

[0165] R=(1 / N)·∑[I(f i ≤ft t hreshold)]

[0166] Among them, N is the total number of samples, I is the indicator function, when the objective function value f i Less than or equal to the threshold f t The value is 1 when hreshold is set, otherwise it is 0. t hreshold is usually set to 1.2 times the optimization target value to evaluate whether the system's performance meets the requirements under uncertain conditions. When the robustness index R ≥ 0.95, the verification is considered passed, indicating that the control strategy has sufficient robustness.

[0167] Preferably, when implementing robustness verification, the following aspects can be further refined or optimized. First, the number of random perturbation samples can be adjusted according to actual needs. For example, in scenarios with high robustness requirements, the number of samples can be increased to improve verification accuracy. Second, the calculation of the objective function value can be combined with more actual operating constraints, such as the operating state restrictions of the equipment or the stability requirements of the power grid, to more comprehensively reflect the operating performance of the system. In addition, an adaptive adjustment mechanism can be introduced to dynamically adjust the parameters or control strategy of the optimization model based on the verification results, thereby further improving the robustness of the system.

[0168] In some embodiments, the control instruction set generation method in step S7 includes:

[0169] S71. Decompose the optimization results in time and space to generate the time-power curve P of each device ref (t);

[0170] S72. Add feedforward compensation

[0171]

[0172] Where, e(t) = P ref (t)-P act (t), K p =0.8 is the proportional coefficient, K j =0.2 is the integration coefficient;

[0173] S73. Smooth the instruction set to ensure that the second-order derivative constraint is satisfied:

[0174]

[0175] in, P rated is the rated power of the device.

[0176] It should be noted that the control instruction set generation method in this invention is the final step in the virtual power plant control method. It is used to convert the optimization results into specific control instructions and issue them to each distributed energy device. The control instruction set includes timing control instructions and power adjustment thresholds to ensure that each device operates according to the optimization results. Timing control instructions refer to the device operation plan generated based on the time series, while power adjustment thresholds are the power output range of the device in different time periods. By generating these instructions, precise control of distributed energy devices can be achieved, ensuring the safe, economical, and stable operation of the virtual power plant.

[0177] Specifically, the generation process of the control instruction set includes the following steps: First, the optimization results are decomposed into time and space to generate the time-power curve P of each device. sTime-space decomposition is to decompose the optimization results according to time and equipment to obtain the power output plan of each equipment in each time period. Secondly, add the feedforward compensation term P ff (t), the formula is:

[0178] P ff (t) = K p ·e(t)+K i ·∫e(t)dt

[0179] Where, e(t) is the power error, K p and K i are the proportional coefficient and the integral coefficient respectively. The feedforward compensation term is used to correct the power output of the device to make it closer to the optimized result. Finally, the instruction set is smoothed to ensure that the second-order derivative constraint |d 2 P / dt 2 |≤P m ax a cc, where P m ax a cc is the maximum acceleration of the device. Smoothing can avoid drastic changes in the device's power output and ensure safe operation of the device.

[0180] Preferably, when implementing the control instruction set generation method, the following aspects can be further refined or optimized. First, the granularity of time-space decomposition can be adjusted according to the actual operating characteristics of the device. For example, for a device with a faster response speed, a finer time granularity can be used for decomposition. Second, the proportional coefficient K of the feedforward compensation term is p and the integral coefficient K i Optimization can be performed based on actual operating data to improve the compensation effect. In addition, the constraints of the smoothing process can be adjusted according to the dynamic characteristics of the equipment, such as introducing a nonlinear smoothing function or considering the dynamic response characteristics of the equipment, thereby further improving the accuracy and reliability of the control instructions.

[0181] The above-mentioned embodiments of the present invention have the following beneficial effects: By constructing a multi-time-scale collaborative optimization model and combining the three-layer structure of day-ahead optimization, intraday rolling and real-time adjustment, the present invention can effectively take into account the operating requirements of the virtual power plant at different time scales and realize the refined scheduling of distributed energy equipment. The improved chaotic particle swarm algorithm is used to solve the optimization model, which can quickly converge to the global optimal solution, significantly improve the optimization efficiency and accuracy, and ensure that the virtual power plant maintains efficient and stable operation under complex operating conditions. By dynamically adjusting the weight coefficient matrix of the optimization model, the control strategy can be flexibly adjusted according to the deviation between the real-time operating data and the predicted data, further enhancing the adaptability and robustness of the system.

[0182] In addition, the present invention verifies the robustness of the control parameters through Monte Carlo simulation, which can ensure that the optimization results can still meet the actual operation requirements under the influence of uncertain factors. Combined with wavelet packet decomposition and deep belief network prediction model, the prediction accuracy of output fluctuations and load changes of distributed energy equipment can be improved. The use of the improved Adam optimizer can accelerate the training process of network parameters and improve the convergence speed and accuracy of the prediction model. By generating a control instruction set containing timing control instructions and power adjustment thresholds, precise control of distributed energy equipment can be achieved, ensuring the safe, economical and stable operation of the virtual power plant.

[0183] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, a server, a mobile phone, or a tablet.

[0184] The above descriptions are merely some preferred embodiments of the present invention and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A virtual power plant control method based on a multi-objective optimization algorithm, characterized in that: The following steps are involved: S1. Collect real-time operating data of distributed energy equipment through the virtual power plant control center, including power generation power, energy storage capacity, load demand, and grid dispatch instructions; S2. Establish a dynamic prediction model to predict the output fluctuation characteristics and load change characteristics of each distributed energy device in the future T period; S3. Construct a multi-timescale collaborative optimization model, the model including a day-ahead optimization layer, an intraday rolling layer, and a real-time adjustment layer; S4. Using an improved chaotic particle swarm optimization algorithm to solve the multi-time scale collaborative optimization model and generate a set of device control parameters; S5. Dynamically adjust the weight coefficient matrix of the optimization model based on the deviation between the real-time operation data and the predicted data; S6. Verify the robustness of the control parameters through Monte Carlo simulation. When the verification pass rate is ≥95%, execute step S7; S7. Issue a control instruction set including timing control instructions and power regulation thresholds to each distributed energy device.

2. The method according to claim 1, characterized in that The step S4 comprises: S41. Initialize the particle swarm parameters, set the inertia factor ω = 0.729, the acceleration constants c1 = 2.05, c2 = 2.05; S42. Use Tent chaotic mapping to generate the initial particle position matrix X = [x1, x2, ..., x n ] T , where x i ∈[0,1] D , D is the dimension of optimization variables, n is the number of particles; S43. Calculate the fitness function value F(X) = α·f1+β·f2+γ·f3, where α is the power tracking weight coefficient, β is the economy weight coefficient, and γ is the stability weight coefficient, and α+β+γ=1; represents the power tracking error, where P act (t) is the actual power at time t, P pre (t) is the predicted power at time t, and T is the total number of time periods; Represents the economic deviation, where Cost real (t) is the actual operating cost, Cost min is the theoretical minimum cost; represents the power fluctuation penalty term, where ΔP(t)=|P(t)-P(t-1)|, P max is the maximum allowable power of the equipment; S44. Update particle velocity and position through dynamic inertia weight adjustment strategy; S45. When max(|F(X (k) )-F(X (k-1) )|)≤ε or the number of iterations k≥K max Stop the iteration and output the optimal solution set, where ε=1e-6, K max =500, k is the current number of iterations.

3. The method according to claim 2, characterized in that The dynamic inertia weight adjustment strategy in step S44 satisfies: Among them, ω initial =0.9,ω final =0.4, λ is the nonlinear adjustment index and satisfies k is the current iteration number, K max is the maximum number of iterations.

4. The method according to claim 1, wherein The step S2 comprises: S21. Perform wavelet packet decomposition on the historical operation data to obtain n intrinsic mode components IMF1~IMF n ; S22. Establish a deep belief network (DBN) prediction model with a hidden layer structure of [128, 64, 32] and the number of input layer nodes equal to the number of IMF components n. S23. Use the improved Adam optimizer to train the network parameters, and the learning rate η is Decay, where η0 = 0.001, τ = 1000, and k is the number of training steps; S24. Predict each IMF component separately and reconstruct it to obtain the final prediction result.

5. The method according to claim 4, characterized in that The parameter update formula of the improved Adam optimizer is: m t =μ·m t-1 +(1-μ)·g t Among them, μ = 0.9 is the momentum decay rate, v = 0.999 is the second-order moment decay rate, δ = 1e-8 is the numerical stability constant, g t is the gradient vector at time t, sign() is the sign function, output is ±1, and η is the learning rate.

6. The method according to claim 1, characterized in that The dynamic adjustment strategy in step S5 includes: S51. Calculate the deviation between real-time data and predicted data: Among them, P real (t) is the actual power, P pre (t) is the predicted power, P rated is the rated power of the equipment; S52. When ΔD(t)≥ΔD threshold When , press β←β+κ·(ΔD(t)-ΔD threshold ) Update the economic weight coefficient while maintaining α+β+γ=1, where κ=0.05, ΔD threshold =0.15; S53. The adjusted weight coefficients must satisfy the constraints: α∈[0.3,0.6], β∈[0.2,0.5], γ∈[0.1,0.3].

7. The method according to claim 1, characterized in that The objective function of the multi-time-scale collaborative optimization model in step S3 is: Constraints include: 1) Output limit: P min,i ≤P i (t)≤P max,i , 2) Power balance: 3) Climbing limit: |P i (t)-P i (t-1)|≤ΔP ramp,i , where: C gen (t) is the power generation cost, C curt (t) is the cost of abandoned energy, C reg (t) is the adjustment cost, P i (t) is the output of device i in period t, I is the set of devices, P load (t) is the load power at time t, P loss (t) is the power loss at time t, ΔP ramp,i is the maximum allowed ramp rate of device i.

8. The method according to claim 7, characterized in that The power generation cost C gen The calculation formula for (t) is: Among them, a i is the quadratic cost coefficient, b i is the linear cost coefficient, c i is the fixed cost coefficient, d i and σ i is a nonlinear cost parameter, and N is the total number of distributed energy devices.

9. The method according to claim 1, characterized in that The step S6 comprises: S61. Generate M = 1000 sets of randomly perturbed samples that follow a normal distribution N(μ,∑), where μ is the predicted mean and ∑ is the covariance matrix; S62. Perform forward simulation on each sample to calculate the objective function value F. m ; S63. Calculation of robustness indicators Among them, I() is the indicator function, when F m ≤F max 1 when it is set, otherwise 0, F max =1.2·F opt , F opt To optimize the target value; S64. When R≥0.95, the verification is determined to be passed.

10. The method according to claim 1, characterized in that The control instruction set generation method in step S7 includes: S71. Decompose the optimization results in time and space to generate the time-power curve P of each device ref (t); S72. Add feedforward compensation Where, e(t) = P ref (t)-P act (t), K p =0.8 is the proportional coefficient, K i =0.2 is the integration coefficient; S73. Smooth the instruction set to ensure that the second-order derivative constraint is satisfied: in, P rated is the rated power of the device.

Images