A virtual power plant optimization scheduling method based on machine learning

Abstract

The present invention belongs to the field of power system optimization and scheduling, and in particular relates to a virtual power plant optimization and scheduling method based on machine learning. First, real-time data such as power generation and load demand of distributed power sources, energy storage equipment, and load management systems in the virtual power plant are collected, cleaned, and standardized. Then, a comprehensive objective function is constructed, covering multiple factors such as economic benefits, environmental benefits, and price fluctuations in the electricity market. Then, an improved NSGA‑II algorithm is used for optimization. Its dual selection mechanism combines local and global optimality to accelerate convergence, improves congestion metrics to accurately evaluate individuals, and adaptive crossover and mutation operations are adjusted according to population dynamics. Through this method, a balance between economic and environmental benefits can be maximized, equipment operation constraints can be met, the operating efficiency and stability of virtual power plants can be improved, and an innovative solution can be provided for power system optimization.

Description

Technical Field

[0001] The present invention belongs to the field of power system optimization and scheduling, and in particular relates to a virtual power plant optimization and scheduling method based on machine learning. Background Art

[0002] Against the backdrop of accelerating energy transition and electricity marketization, virtual power plants (VPPs) have emerged as an emerging model for integrating and optimizing power resources. Power systems face numerous challenges. The large-scale integration of distributed energy resources (DREs) reduces the stability and controllability of power supply. The intermittent and volatile nature of their generation can easily lead to grid frequency and voltage fluctuations, threatening their safe operation. Furthermore, with increasingly stringent environmental protection requirements, reducing carbon emissions has become a key task for the power industry. The traditional power generation structure, which is primarily based on thermal power, urgently needs to be adjusted, and the proportion of clean energy in the power supply needs to be increased. However, clean energy generation is constrained by natural conditions, further complicating power supply and demand balancing and scheduling. In the electricity market, with frequent price fluctuations, existing scheduling methods struggle to balance economic and environmental benefits, and are unable to efficiently allocate resources based on market dynamics and equipment performance. Furthermore, existing VPP scheduling technologies, when considering the coordinated operation of multiple energy devices, often overlook comprehensive factors such as the charging and discharging efficiency of energy storage devices, the performance constraints of various devices, and power generation volatility. This lacks a systematic optimization solution and makes it difficult to meet the demands of modern power systems for efficient, stable, and green operation. Summary of the Invention

[0003] In response to the technical problems existing in the above-mentioned background technology, the present invention proposes a virtual power plant optimization scheduling method based on machine learning.

[0004] In order to achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0005] S1. First, collect real-time data from each distributed power source, energy storage device, and load management system in the virtual power plant;

[0006] S2. Clean and standardize the collected data;

[0007] S3. Construct an optimized objective function that not only considers economic and environmental benefits, but also introduces price fluctuations in the electricity market, grid load forecasts, performance constraints of various energy devices, charging and discharging efficiency of energy storage devices, and power generation volatility of virtual power plants. The comprehensive objective function is: F(x) = α·C(x)-β·E(x)-γ·R(x), where C(x) represents economic benefits, E(x) represents environmental benefits, R(x) is a constraint function, and α, β, and γ are weight coefficients.

[0008] S4. Finally, the improved non-dominated sorting genetic algorithm NSGA-II is used to optimize the objective function for multiple objectives, which improves the optimization efficiency;

[0009] The improved non-dominated sorting genetic algorithm NSGA-II is implemented as follows:

[0010] S41, first initialize the population and initialize individual parameters;

[0011] S42. Use double selection mechanism to calculate the non-dominated order of individuals and improve the crowding measure;

[0012] S43, performing adaptive crossover and mutation operations;

[0013] S44. Generate a new population, terminate the algorithm when the maximum set number of iterations of 30 is reached, and output the results.

[0014] Preferably, the specific form of the economic benefit C(x) in step S3 is: Among them, P sell (t) represents the amount of electricity sold at time t, P buy (t) is the amount of electricity purchased at time t, λ(t), μ(t) are the sales and purchase prices in the electricity market, respectively, and C fuel,i (t), C op,i (t) represents the combustion cost and operating cost of equipment of type i.

[0015] Preferably, the specific form of the environmental benefit E(x) in step S3 is: where α j , β eco is the environmental protection factor, CF j is the carbon emission factor of the jth type of power generation, P gen,j (t) is the amount of electricity generated by the jth type of equipment at time t, P gen,coal (t) is the amount of electricity generated by coal-fired power plants.

[0016] Preferably, the specific form of the constraint function R(x) in step S3 is: Where SOC(t) represents the state of charge of the device, and this constraint function ensures that all devices in the virtual power plant meet their operating constraints during the scheduling process.

[0017] Preferably, the steps for implementing the double selection mechanism in step S42 to calculate the non-dominated order of individuals and improve the congestion metric are as follows:

[0018] S421, first calculate the non-dominated relationship;

[0019] S422. The improved crowding metric considers the solution density in both the target space and the decision space. The specific approach is: where f k (x i ) is individual x i The function value on the kth target, f k (x i ′) is x i Adjacent individual x i ′The function value on the kth target, x i (d) is individual x i The value of the d-th decision variable in the decision space;

[0020] S423. Filter the local optimal solution and sort the solutions in each non-dominated level from high to low according to the congestion degree, and preferably select the solution with high congestion degree: priority local (x i )=-D crowd (x i ), select N from the current non-dominated level local local optimal solutions, forming a local solution set;

[0021] S424, combining non-dominated level and congestion degree, calculate global priority: priority global (x i )=F(x i )+D crowd (x i ), where F(x i ) is the non-dominated sorting level, sorted from low to high according to the global priority, and select N from the population global A global optimal solution;

[0022] S425. Finally, the local solution set and the global solution set are merged to obtain a candidate set.

[0023] Preferably, the adaptive crossover and mutation operations in step S43 are specifically implemented as follows: the adaptive crossover implementation formula is: Diversity is a measure of population diversity; the adaptive mutation operation is: Where s(x i ) is individual x i The fitness, s max is the individual with the best fitness in the population.

[0024] Compared with the existing technology, the advantages and positive effects of the present invention are that, in the construction of the objective function, it comprehensively integrates economic and environmental benefits, and innovatively introduces factors such as electricity market price fluctuations, grid load forecasts, energy equipment performance constraints, energy storage equipment charging and discharging efficiency, and power generation volatility, breaking through the limitations of single considerations of traditional methods. The improved NSGA-II algorithm, with its dual selection mechanism that integrates local and global optimality, can quickly locate local optimal solutions and guide global optimality in virtual power plant equipment scheduling, accelerating convergence. The improved congestion metric takes into account the density of multi-space solutions and accurately evaluates individuals. The adaptive genetic operation flexibly adjusts the crossover mutation probability and amplitude according to population dynamics, balances global and local searches, effectively improves the optimization effect, and enhances overall adaptability. DETAILED DESCRIPTION

[0025] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described below in conjunction with the embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0026] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0027] Embodiment: In the current energy landscape, with the large-scale access of distributed energy, the stability of power supply faces huge challenges. For example, in the power system of a certain region, the power generation of solar and wind power stations fluctuates violently in different seasons and weather conditions. During sunny and stable wind periods, power generation may exceed local demand, while in rainy or windless weather, power generation drops sharply, seriously affecting the supply and demand balance of the power grid. At the same time, environmental protection pressure has prompted the power industry to reduce carbon emissions, and the power generation structure dominated by traditional thermal power is in urgent need of adjustment. However, clean energy power generation is constrained by natural conditions, which further increases the complexity of power dispatching. In addition, electricity market prices fluctuate frequently, and existing dispatching methods are difficult to balance economic and environmental benefits, and cannot achieve efficient allocation of resources. Therefore, the present invention proposes a virtual power plant optimization dispatching method based on machine learning.

[0028] First, in the data collection link, the present invention is committed to extensively collecting rich real-time data from various distributed power sources, energy storage devices, and load management systems within the virtual power plant. These data types are diverse, including power generation, load demand, energy storage status, and market electricity price information. Due to various quality issues with the collected data, data cleaning and standardization are required. During the data cleaning process, advanced filtering algorithms, such as the Kalman filter algorithm, are used for noisy data to effectively remove high-frequency interference and measurement errors and restore the true appearance of the data. For missing values, reasonable interpolation methods are used to fill them according to the time series characteristics and correlation of the data.

[0029] Next, constructing the optimized objective function is one of the core links of the present invention. The objective function constructed by the present invention is highly comprehensive. It not only comprehensively considers the two key aspects of economic and environmental benefits, but also innovatively introduces multiple important factors, including price fluctuations in the electricity market, grid load forecasts, performance constraints of various energy devices, charging and discharging efficiency of energy storage devices, and power generation volatility factors of virtual power plants. The specific form of economic benefits C(x) is: Among them, P sell (t) represents the amount of electricity sold at time t, P buy (t) is the amount of electricity purchased at time t, λ(t), μ(t) are the sales and purchase prices in the electricity market, respectively, and C fuel,i (t), C op,i (t) represents the combustion cost and operating cost of the i-th type of equipment. Through this formula, the economic benefits of the virtual power plant under different market electricity prices and equipment operating costs can be accurately calculated. The specific form of environmental benefits E(x) is: where α j , β eco is the environmental protection factor, CF j is the carbon emission factor of the jth type of power generation, P gen,j (t) is the amount of electricity generated by the jth type of equipment at time t, P gen,coal (t) is the amount of electricity generated by coal-fired power plants. Through this formula, during the optimization process, clean energy power generation equipment can be prioritized, reducing dependence on coal-fired power plants, thereby effectively reducing carbon emissions. The specific form of the constraint function R(x) is: Where SOC(t) represents the state of charge of the equipment. This constraint function ensures that all equipment in the virtual power plant meets its operating constraints during scheduling. Therefore, the comprehensive objective function is: F(x) = α·C(x) - β·E(x) - γ·R(x), where C(x) represents economic benefits, E(x) represents environmental benefits, R(x) is the constraint function, and α, β, and γ are weight coefficients. The weight coefficients are determined using the Analytic Hierarchy Process (AHP), and the calculated values ​​are α = 0.5, β = 0.35, and γ = 0.15.

[0030] Taking into account the many shortcomings of the traditional NSGA-II algorithm, such as the lack of selection pressure, the easy omission of excellent individuals, and the slow evolution of the population; it is difficult to maintain population diversity, and it is easy to converge to the local optimum too early, missing the global optimal solution; the computational efficiency is extremely low when dealing with large-scale problems. The improved NSGA-II algorithm in the present invention integrates local and global optimality in a dual selection mechanism. In the virtual power plant, it can not only quickly explore the optimization scheme locally, but also guide to the optimal solution from the global perspective to accelerate convergence. The improved congestion metric comprehensively considers the target and the density of the decision space solution. The adaptive genetic operation is flexibly adjusted according to the state of the population. When the diversity is high, the crossover probability is increased to stimulate new solutions. The variation range of individuals with different fitness is reasonably changed to balance the global and local searches, effectively improve the optimization efficiency and quality, and enhance the adaptability of the algorithm. The implementation of the improved non-dominated sorting genetic algorithm NSGA-II is: first initialize the population and initialize the individual parameters; calculate the non-dominated relationship, for any two individuals x in the population i and x j , according to the objective function value f k (x) determines the dominant relationship between them, if x i Not inferior to x in all objectives j , and is better than x in at least one objective j , then x i dominatex j , then the improved crowding metric considers the solution density in both the target space and the decision space at the same time, specifically: where f k (x i ) is individual x i The function value on the kth target, f k (x i ′) is x i Adjacent individual x i ′The function value on the kth target, x i (d) is individual x i The value of the d-th decision variable in the decision space; then screen the local optimal solution, sort it from high to low according to the congestion in each non-dominated level, and preferably choose the solution with high congestion: priority local (x i )=-D crowd (x i ), select N from the current non-dominated level local local optimal solutions to form a local solution set; combining the non-dominated level and congestion degree to calculate the global priority: priority global (x i )=F(x i )+D crowd (x i), where F(x i ) is the non-dominated sorting level, sorted from low to high according to the global priority, and select N from the population global The global optimal solution is obtained by merging the local solution set with the global solution set to obtain the candidate set. Then, adaptive crossover and mutation operations are performed; when the population diversity is high, the crossover probability is increased, and when the population tends to converge, the crossover probability is reduced. The implementation formula is: Diversity is a measure of population diversity. Individuals with poor fitness increase their variation to try to escape the current solution area. Individuals with good fitness reduce their variation to explore the area near the current solution in detail. Where s(x i ) is individual x i The fitness, s max The individual with the best fitness in the population is generated. Finally, a new population is generated. The algorithm terminates when the maximum set number of iterations, 30, is reached, and the results are output. The optimal scheduling plan for the virtual power plant at different time periods is obtained, maximizing the balance between economic and environmental benefits while ensuring that all operational constraints are met.

[0031] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A virtual power plant optimization scheduling method based on machine learning, characterized in that: The following steps are involved: S1. First, collect real-time data from each distributed power source, energy storage device, and load management system in the virtual power plant; S2. Clean and standardize the collected data; S3. Construct an optimized objective function that not only considers economic and environmental benefits, but also introduces price fluctuations in the electricity market, grid load forecasts, performance constraints of various energy devices, charging and discharging efficiency of energy storage devices, and power generation volatility of virtual power plants. The comprehensive objective function is: F(x) = α·C(x)-β·E(x)-γ·R(x), where C(x) represents economic benefits, E(x) represents environmental benefits, R(x) is a constraint function, and α, β, and γ are weight coefficients. S4. Finally, the improved non-dominated sorting genetic algorithm NSGA-II is used to optimize the objective function for multiple objectives, which improves the optimization efficiency; The improved non-dominated sorting genetic algorithm NSGA-II is implemented as follows: S41, first initialize the population and initialize individual parameters; S42. Use double selection mechanism to calculate the non-dominated order of individuals and improve the crowding measure; S43, performing adaptive crossover and mutation operations; S44, generate a new population, terminate the algorithm when the maximum set number of iterations of 30 generations is reached, and output the result; The steps for implementing the double selection mechanism in step S42 to calculate the non-dominated order of individuals and improve the congestion metric are as follows: S421, first calculate the non-dominated relationship; S422. The improved crowding metric considers the solution density in both the target space and the decision space. The specific approach is: where f k (x i ) is individual x i The function value on the kth target, f k (x i ′) is x i Adjacent individual x i ′The function value on the kth target, x i (d) is individual x i The value of the d-th decision variable in the decision space; S423. Filter the local optimal solution and sort the solutions in each non-dominated level from high to low according to the congestion degree, and preferably select the solution with high congestion degree: priority local (x i )=-D crowd (x i ), select N from the current non-dominated level local local optimal solutions, forming a local solution set; S424, combining non-dominated level and congestion degree, calculate global priority: priority global (x i )=F(x i )+D crowd (x i ), where F(x i ) is the non-dominated sorting level, sorted from low to high according to the global priority, and select N from the population global A global optimal solution; S425. Finally, the local solution set and the global solution set are merged to obtain a candidate set.

2. The virtual power plant optimization scheduling method based on machine learning according to claim 1 is characterized in that: The specific form of the economic benefit C(x) in step S3 is: Among them, P sell (t) represents the amount of electricity sold at time t, P buy (t) is the amount of electricity purchased at time t, λ(t), μ(t) are the sales and purchase prices in the electricity market, respectively, and C fuel,i (t), C op,i (t) represents the combustion cost and operating cost of equipment of type i.

3. The virtual power plant optimization scheduling method based on machine learning according to claim 1 is characterized in that: The specific form of the environmental benefit E(x) in step S3 is: where α j , β eco is the environmental protection factor, CF j is the carbon emission factor of the jth type of power generation, P gen,j (t) is the amount of electricity generated by the jth type of equipment at time t, P gen,coal (t) is the amount of electricity generated by coal-fired power plants.

4. The virtual power plant optimization scheduling method based on machine learning according to claim 1 is characterized in that: The specific form of the constraint function R(x) in step S3 is: Where SOC(t) represents the state of charge of the device, and this constraint function ensures that all devices in the virtual power plant meet their operating constraints during the scheduling process.

5. The virtual power plant optimization scheduling method based on machine learning according to claim 1 is characterized in that: The specific implementation of the adaptive crossover and mutation operation in step S43 is: the adaptive crossover implementation formula is: Diversity is a measure of population diversity; the adaptive mutation operation is: Where s(x i ) is individual x i The fitness, s max is the individual with the best fitness in the population.

Images