Low-carbon economic dispatch method for controllable flexible load of optical storage and direct flexible system

Abstract

The application discloses a kind of controllable flexible load low-carbon economic dispatching methods of light storage direct flexible system.First, the mathematical model of controllable flexible load, photovoltaic cell and energy storage element is established;Then, the daily operation cost and carbon emission cost function of light storage direct flexible system is constructed;Finally, according to the mathematical model of photovoltaic cell and energy storage element, the total output power of photovoltaic cell and energy storage element in each period is obtained;According to the total output power of photovoltaic cell and energy storage element, considering the controllable flexible load, the daily operation cost and carbon emission cost function of light storage direct flexible system is used as fitness function, and the multi-objective particle swarm optimization algorithm is solved, and the inertia weight, cognitive acceleration coefficient and social acceleration coefficient of multi-objective particle swarm optimization algorithm are optimized based on reinforcement learning, to obtain the scheduling amount of various flexible loads in each period, to realize the minimum operation of daily operation cost and carbon emission cost of light storage direct flexible system.In the modeling of controllable flexible load, control variable is introduced, the uncertainty problem of flexible load participation response is considered, and the adaptive adjustment of particle swarm algorithm parameters is also realized, and the anti-interference ability is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of flexible load scheduling, and in particular relates to a low-carbon economic scheduling method for controllable flexible loads in a photovoltaic-storage direct-flexible system. Background Art

[0002] PV-storage direct-flexible power technology offers broad application prospects in building power systems and the energy sector due to its low-carbon nature, flexible loads, efficient power usage, and intelligent operation. Scheduling the controllable flexible loads in PV-storage direct-flexible power systems is a key approach to achieving low-carbon energy use. However, traditional linear programming methods often suffer from complex solutions and low computational accuracy when faced with long timescales and multi-objective constraints, making them difficult to meet the real-time load scheduling requirements of daily operations.

[0003] The multi-objective particle swarm optimization algorithm is the most mature algorithm for various planning problems. However, the particle swarm optimization algorithm is sensitive to parameter selection and is prone to falling into local optimal solutions in the later stages of the search. At the same time, due to the large uncertainty in the autonomous response behavior of the load, the flexible load may not be able to fully respond during the scheduling process, resulting in the final scheduling result being far from the expected target, making it impossible to achieve accurate scheduling. Summary of the Invention

[0004] In view of the shortcomings of existing flexible load scheduling, the technical problem that the present invention intends to solve is to provide a low-carbon economic scheduling method for controllable flexible loads in a photovoltaic storage direct-flexible system.

[0005] The present invention solves the technical problem by adopting the following technical solutions:

[0006] A low-carbon economic dispatch method for controllable flexible loads in a solar-storage direct-flexible system, characterized by comprising the following steps:

[0007] S1. Establish mathematical models of controllable flexible loads, photovoltaic cells, and energy storage elements;

[0008] S2. Construct the daily operation cost and carbon emission cost function of the solar-storage direct-flexible system;

[0009] The daily operation cost function is as follows:

[0010] F1=C GR +C shift +C tran +C cut (8)

[0011] Among them, C GR is the cost of electricity sales and purchase, C shift is the shiftable load dispatch cost, C tran is the transferable load dispatch cost, C cut It can reduce the load cost;

[0012] The carbon emission cost function is as follows:

[0013]

[0014] Among them, C p is the carbon emission penalty coefficient, γ grid is the carbon emission coefficient, P buy (t) is the electric power purchased by the PV-storage direct-flexible system from the grid during period t, where T represents the number of periods in a day;

[0015] S3. Based on the mathematical model of photovoltaic cells and energy storage elements, the total output power of photovoltaic cells and energy storage elements in each time period is obtained. Based on the total output power of photovoltaic cells and energy storage elements, the controllable flexible load is considered, and the daily operating cost and carbon emission cost function of the photovoltaic storage direct-flexible system are used as the fitness function. The multi-objective particle swarm optimization algorithm is used to solve the problem. The parameters of the multi-objective particle swarm optimization algorithm are optimized based on reinforcement learning to obtain the dispatch amount of various flexible loads in each time period.

[0016] The particle velocity update formula is:

[0017]

[0018] r 1k =G r1 (s k-1 )(16)

[0019] r 2k =G r2 (s k-1 )(17)

[0020]

[0021] Where, v i (k+1), v i (k) are the velocities of the i-th particle in the k+1th and kth iterations, respectively, ω k is the inertia weight of the kth iteration, c 1k is the cognitive acceleration coefficient of the kth iteration, c 2k is the social acceleration coefficient of the kth iteration, x i (k) is the position of the i-th particle in the k-th iteration, r 1k and r 2k is a random number in the range of 0 to 1, is the individual best position of the i-th particle in the k-th iteration, x gbest (k) is the optimal position of the population in the kth iteration, G r1 , G r2 is the generative network, s k-1is the state of the agent in the k-1th iteration, a 0k , a 1k , a 2k are the action values in the three dimensions of reinforcement learning respectively, and 10 and c 20 are the initial values of inertia weight, cognitive acceleration coefficient and social acceleration coefficient respectively.

[0022] Compared with the prior art, the present application has the beneficial effects that:

[0023] 1. In view of the uncertainty of the response of flexible load to dispatching, control variables are introduced in the mathematical modeling of various flexible loads, and the control variables are used to represent whether the flexible load participates in the response, so that the problem of uncertainty of the response of flexible load is effectively solved.

[0024] 2. Instead of keeping the parameters unchanged or mechanically adjusting the parameters according to the number of iterations, the inertia weight, the cognitive acceleration coefficient and the social acceleration coefficient of the multi-objective particle swarm optimization algorithm are adaptively optimized by using reinforcement learning, the state of the particle swarm is taken as the state of the agent, the change of the parameters is mapped through the state of the particle swarm, the anti-interference ability of the particle swarm algorithm is enhanced, the search speed is accelerated in the early search stage, the search range is expanded in the late search stage to prevent falling into a local optimal solution, the problem of falling into a local optimal solution of the particle swarm algorithm is improved while the convergence speed is ensured, and finally the minimum operation cost of the photovoltaic storage direct flexible system and the carbon emission cost are realized. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is the overall flowchart of the present application;

[0026] Figure 2 is the flowchart of the optimization solution of the present application;

[0027] Figure 3 is the flowchart of the reinforcement learning of the present application;

[0028] Figure 4 is the flexible load distribution diagram before dispatching in the embodiment of the present application;

[0029] Figure 5 is the flexible load dispatching result diagram in the embodiment of the present application;

[0030] Figure 6 is the power distribution of each part of the photovoltaic storage direct flexible system after dispatching in the embodiment of the present application;

[0031] Figure 7 is the multi-objective Pareto solution set of different algorithms in the embodiment of the present application. DETAILED DESCRIPTION

[0032] Specific embodiments are given below with reference to the accompanying drawings, which are only used to illustrate the technical solutions of the present application in detail and do not limit the protection scope of the present application.

[0033] The present application provides a low-carbon economic dispatching method of controllable flexible load and photovoltaic energy storage system (referred to as method, see Figures 1 to 7 ), comprising the following steps:

[0034] S1, a mathematical model of controllable flexible load, photovoltaic cell and energy storage element is established;

[0035] The mathematical model of controllable flexible load reflects the dispatching rule, and the controllable flexible load includes three kinds of translatable, transferable and reducible; wherein the mathematical model of translatable load is:

[0036]

[0037] Wherein, is the translatable load in the t period after participating in response, is the translatable load in the t period before participating in response, represents the scheduling period of translatable load, represents the response control variable of translatable load, which is a 0 / 1 variable, represents that the translatable load participates in response in the period , represents that the translatable load does not participate in response.

[0038] The mathematical model of transferable load is:

[0039]

[0040] Wherein, represents the transferable load in the t period after participating in response, represents the transferable load in the t period before participating in response, represents the scheduling period of transferable load, represents the response control variable of transferable load, which is a 0 / 1 variable, represents that the transferable load participates in response in the period , represents that the transferable load does not participate in response.

[0041] The mathematical model of reducible load is:

[0042]

[0043] Wherein, is the reducible load in the t period after participating in response, is the reducible load in the t period before participating in response. Indicates the load reduction response control variable, which is a 0 / 1 variable. Indicates that the load can be reduced to participate in the response, Indicates that the load can be reduced and does not participate in the response; a cut is the load reduction factor, and its value range is [0,1].

[0044] The mathematical model of a photovoltaic cell is:

[0045]

[0046] Among them, P pv Represents the actual output power of the photovoltaic cell, P STC is the rated output power of the photovoltaic cell under standard operating conditions; k is the derating factor of the photovoltaic cell, which is usually 0.9 to 0.95; G STC It is the irradiance intensity of photovoltaic cells under standard working conditions, and its value is 1000W / m 2 ; G c represents the irradiation intensity of the photovoltaic cell in the working environment, α is the temperature coefficient; T c is the operating temperature of the photovoltaic cell, T STC is the temperature of the photovoltaic cell under standard working conditions, the standard room temperature is 25°C; T a is the operating environment temperature; NOCT is the nominal operating temperature of the photovoltaic cell, generally 45±2℃.

[0047] The mathematical model of energy storage element charging and discharging is:

[0048] soc c (t)=(1-δ)soc c (t-1)-P c Δt c / η c E r (6)

[0049] Among them, soc c (t), soc c (t-1) represents the charge capacity of the energy storage element at the end of the tth and t-1th charging periods, δ represents the self-discharge rate of the energy storage element, and P c Indicates the charging power of the energy storage element, Δt c Represents the charging time interval of the energy storage element, η c Indicates the charging efficiency of the energy storage element, E r Indicates the rated capacity of the energy storage element;

[0050] soc d (t)=(1-δ)soc d (t-1)-P d Δt d / η d E r (7)

[0051] Among them, soc d (t), soc d (t-1) represents the charge capacity of the energy storage element at the end of the tth and t-1th discharge periods, respectively, and P d Indicates the discharge power of the energy storage element, Δt d Represents the discharge time interval of the energy storage element, η d Indicates the discharge efficiency of the energy storage element.

[0052] S2. Construct a function for the daily operating costs and carbon emission costs of a PV-storage direct-flexible system. Daily operating costs include electricity sales and purchase costs, load dispatch costs that can be shifted, load dispatch costs that can be transferred, and load dispatch costs that can be reduced.

[0053] The daily operating cost function of the PV-storage direct-flexible system is as follows:

[0054] F1=C GR +C shift +C tran +C cut (8)

[0055] Among them, C GR is the cost of electricity sales and purchase, C shift is the shiftable load dispatch cost, C tran is the transferable load dispatch cost, C cut This can reduce load costs.

[0056] The cost of electricity sales and purchase is:

[0057]

[0058] Where M(t) is the grid electricity price in period t, T represents the number of periods in a day; P Grid (t) is the interaction power between the PV-storage direct-flexible system and the grid at time t. When the system sells electricity to the grid, P Grid (t) is a negative value, when the system purchases electricity from the grid P Grid (t) is a positive value.

[0059] The cost of shiftable load scheduling is:

[0060]

[0061] Among them, C c is the unit dispatch cost of the shiftable load;

[0062] The cost of dispatching transferable load is:

[0063]

[0064] Among them, C t is the unit dispatch cost of the transferable load;

[0065] The load dispatching cost can be reduced as follows:

[0066]

[0067] Among them, C rl is the unit dispatch cost of load that can be reduced.

[0068] The carbon emission cost function of the solar-storage direct-flexible system is as follows:

[0069]

[0070] Among them, C p is the carbon emission penalty coefficient, γ grid is the carbon emission coefficient, P buy (t) is the electric power purchased by the PV-storage direct-flexible system from the grid during period t.

[0071] S3. Based on the mathematical model of the photovoltaic cells and energy storage elements, the total output power of the photovoltaic cells and energy storage elements in each time period is obtained. Based on the total output power of the photovoltaic cells and energy storage elements, the controllable flexible load is taken into account to determine the system power deficit or surplus in each time period, and then the interaction power between the solar-storage direct-flexible system and the power grid is obtained.

[0072] The daily operating costs and carbon emission cost functions of the PV-storage direct-flexible system are used as fitness functions and solved using a multi-objective particle swarm optimization algorithm. At the same time, the inertia weight, cognitive acceleration coefficient, and social acceleration coefficient of the multi-objective particle swarm optimization algorithm are optimized based on reinforcement learning to obtain the dispatch amount of various flexible loads in each time period. Each particle position represents a set of solutions, that is, the dispatch amount of various flexible loads in a single time period.

[0073] The particle position and velocity update formulas are as follows:

[0074] x i (k+1)=x i (k)+v i (k+1)(14)

[0075]

[0076] r 1k =G r1 (s k-1 )(16)

[0077] r 2k =G r2 (s k-1 )(17)

[0078] Where x i (k+1), x i (k) are the positions of the i-th particle in the k+1th and kth iterations, respectively, and v i (k+1), v i (k) are the velocities of the i-th particle in the k+1th and kth iterations, respectively, ω k is the inertia weight of the kth iteration, c 1k is the cognitive acceleration coefficient of the kth iteration, c 2k is the social acceleration coefficient of the kth iteration, r 1k and r 2k is a random number in the range [0,1], is the individual best position of the i-th particle in the k-th iteration, x gbest (k) is the optimal position of the population in the kth iteration, G r1 , G r2 is the generative network, s k-1 is the particle swarm state in the k-1th iteration.

[0079] Reinforcement learning (RL) is a type of machine learning that aims to guide an agent to perform optimal actions in an environment to maximize cumulative rewards. In a standard RL setting, each agent interacts with the environment with the ultimate goal of maximizing the environment's reward. This interaction is formally described as a Markov decision process (MDP), which can be described by the four-tuple (S, A, R, P), where S represents the state space, A is the action space, R is the reward function, and P is the state transition probability.

[0080] The process of reinforcement learning is as follows: the agent will interact with the environment in discrete time steps. The agent's behavior is controlled by the policy π, which maps each state to an action. In each iteration, the agent receives a state s k ∈S, and perform an action a according to the policy k ∈A, and then obtain a reward value r k .

[0081] In the deep deterministic policy gradient of reinforcement learning, four neural networks are designed to obtain the best strategy: the actor network μ(s k |θ μ ), target actor network μ'(s k |θ μ′ ), action value network Q(s k ,a k |θ Q ) and target action value network Q'(s k ,a k |θ Q'), the actor network and the target actor network are used to select actions according to the current state, and the action value network and the target action value network are used to evaluate the effect of the action. During the training process, the weight parameter update formula of the target actor network and the target action value network is as follows:

[0082]

[0083] where θ μ , θ μ′ are the weight parameters of the actor network and the target actor network respectively, θ Q , θ Q′ are the weight parameters of the action value network and the target action value network respectively, and τ is the sliding average coefficient.

[0084] The weight parameter θ Q of the action value network is updated by minimizing the following loss function:

[0085] L(θ Q ) = (r k (s k , a k ) + γQ'(s k+1 , a k+1 | θ Q′ ) - Q(s k , a k | θ Q′ ) - Q(s k , a k | θ Q ) 2 (19)

[0086] where γ is a discount factor, Q(s k , a k | θ Q ) is the action value in the kth iteration, and Q'(s k+1 , a k+1 | θ Q′ ) is the target action value in the (k+1)th iteration.

[0087] The weight parameter update formula of the actor network is:

[0088]

[0089] where represents the sampled gradient, represents the gradient of the action value with respect to the action, and represents the gradient of the actor network with respect to the parameter θ μ .

[0090] The state of the reinforcement learning algorithm refers to the particle swarm state, including the iteration progress of the multi-objective particle swarm optimization algorithm, particle position diversity, and population stability. The calculation formula of the iteration progress is as follows:

[0091]

[0092] In the formula, is the iteration progress, k num is the current iteration number, k max is the maximum iteration number;

[0093] The particle position diversity is represented by the particle position diversity index, and the calculation formula is as follows:

[0094]

[0095] In the formula, Diversity(k) represents the particle position diversity index in the kth iteration, x ij (k) is the jth dimension of the ith particle in the kth iteration, is the average value of the jth dimension of all particles in the kth iteration, n s represents the dimension of the particle position space;

[0096] The population stability is represented by the population stability index, and the calculation formula is as follows:

[0097] I noim =(k num -k numlast ) / k max (23)

[0098] In the formula, I noim is the population stability index, k numlast is the iteration number when the last population best position is updated.

[0099] The action space of the reinforcement learning algorithm is three-dimensional, and the action is between -1 and 1. Through linear transformation, the action output by the actor network is mapped to a reasonable interval. The three dimensions of the action space correspond to the inertia weight ω k , cognitive acceleration coefficient c 1k and social acceleration coefficient c 2k in the velocity update formula of the multi-objective particle swarm optimization algorithm, respectively. Among them, c 1k , c 2k are transformed to the interval 0.5-2.5, and ω k is transformed to the interval 0.1-0.9. The linear transformation formula of the three parameters is as follows:

[0100]

[0101] In the formula, a 0k、a 1k 、a 2k They are the action values ​​in three dimensions, ω0, c 10 and c 20 are the initial values ​​of inertia weight, cognitive acceleration coefficient and social acceleration coefficient respectively.

[0102] The reward values ​​are set as follows:

[0103]

[0104] Where G best (k), G best (k+1) are the optimal fitness values ​​of the particle swarm in the kth and k+1th iterations respectively.

[0105] Example

[0106] This embodiment takes a typical solar-storage direct-flexible system as an example to simulate the scheduling of controllable flexible loads within a day. The basic conditions of various flexible loads in the system are shown in Table 1.

[0107] Table 1 Basic conditions of various types of flexible loads

[0108]

[0109] Table 2 Time-of-use electricity prices within a day

[0110]

[0111] The distribution before flexible load scheduling is as follows Figure 4 As shown in the figure, the flexible load scheduling results are as follows: Figure 5 As shown in the figure, the power of each part of the solar-storage direct-flexible system after scheduling is as follows: Figure 6 As shown in the figure, after the scheduling, the system's daily load decreased by 102.22 kWh, and the energy self-sufficiency rate of the PV-storage direct-flexible system increased from 55.78% to 73.66%, a 17.88% improvement. After optimization, the system's minimum daily operating cost was 709,927 yuan, and the minimum carbon emission cost was 689.2 yuan. Compared to the pre-optimization system's daily operating cost of 1,062,450 yuan and carbon emission cost of 1,609 yuan, the daily operating cost decreased by 33.18% and the carbon emission cost decreased by 57.16%.

[0112] In order to prove the superiority of the method of the present invention, the multi-objective particle swarm optimization algorithm-reinforcement learning algorithm (DDPGPSO) of the present invention is compared with other existing algorithms. The multi-objective Pareto solutions of different algorithms are as follows: Figure 7As shown in the figure, the DDPGPSO algorithm achieves the lowest daily operating cost, which is 2%, 6.7%, and 4.8% lower than the daily operating costs of MOPSO (conventional multi-objective particle swarm algorithm), MOCOA (multi-objective raccoon algorithm), and NSGA2 (improved multi-objective genetic algorithm), respectively. Compared with the carbon emission cost of the MOCOA algorithm, the carbon emission cost optimization results of the other three algorithms are better, and the optimization results of the three algorithms are not much different. Comparing the MOPSO and DDPGPSO algorithms, it is clear that the DDPGPSO algorithm is more efficient in searching for solutions at the end of the search period, which can effectively solve the problem that the MOPSO algorithm is prone to falling into local optimal solutions. Adaptive optimization of the particle swarm algorithm's parameters through reinforcement learning can make the particle swarm algorithm more resistant to interference and more effective in adapting to changes in the state environment when the parameter settings are changed. When the parameter settings are unchanged, the parameters can be set more reasonably when the state environment changes.

[0113] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A low-carbon economic dispatch method for controllable flexible loads in a solar-storage direct-flexible system, characterized by: The following steps are involved: S1. Establish mathematical models of controllable flexible loads, photovoltaic cells, and energy storage elements; S2. Construct the daily operation cost and carbon emission cost function of the solar-storage direct-flexible system; The daily operation cost function is as follows: F1=C GR +C shift +C tran +C cut (8) Among them, C GR is the cost of electricity sales and purchase, C shift is the shiftable load dispatch cost, C tran is the transferable load dispatch cost, C cut It can reduce the load cost; The carbon emission cost function is as follows: Among them, C p is the carbon emission penalty coefficient, γ grid is the carbon emission coefficient, P buy (t) is the electric power purchased by the PV-storage direct-flexible system from the grid during period t, where T represents the number of periods in a day; S3. Based on the mathematical model of photovoltaic cells and energy storage elements, the total output power of photovoltaic cells and energy storage elements in each time period is obtained. Based on the total output power of photovoltaic cells and energy storage elements, the controllable flexible load is considered, and the daily operating cost and carbon emission cost function of the photovoltaic storage direct-flexible system are used as the fitness function. The multi-objective particle swarm optimization algorithm is used to solve the problem. The parameters of the multi-objective particle swarm optimization algorithm are optimized based on reinforcement learning to obtain the dispatch amount of various flexible loads in each time period. The particle velocity update formula is: r 1k =G r1 (s k-1 )(16) r 2k =G r2 (s k-1 )(17) Where, v i (k+1), v i (k) are the velocities of the i-th particle in the k+1th and kth iterations, respectively, ω k is the inertia weight of the kth iteration, c 1k is the cognitive acceleration coefficient of the kth iteration, c 2k is the social acceleration coefficient of the kth iteration, x i (k) is the position of the i-th particle in the k-th iteration, r 1k and r 2k is a random number in the range of 0 to 1, is the individual best position of the i-th particle in the k-th iteration, x gbest (k) is the optimal position of the population in the kth iteration, G r1 , G r2 is the generative network, s k-1 is the state of the agent in the k-1th iteration, a 0k 、a 1k 、a 2k They are the action values ​​in three dimensions in reinforcement learning, ω0, c 10 and c 20 are the initial values ​​of inertia weight, cognitive acceleration coefficient and social acceleration coefficient respectively.

2. The low-carbon economic dispatch method for controllable flexible loads of a PV-storage direct-flexible system according to claim 1 is characterized in that: The state of the agent refers to the particle swarm state, which includes iteration progress, particle position diversity, and population stability. The calculation formula for iteration progress is as follows: Where, is the iteration progress, k num is the current iteration number, k max is the maximum number of iterations; The particle position diversity index is used to characterize the particle position diversity, and the calculation formula is: Where Diversity(k) represents the particle position diversity index of the kth iteration, x ij (k) is the jth dimension of the i-th particle in the k-th iteration, is the average value of the jth dimension of all particles in the kth iteration, n s Represents the dimension of the particle position space; The population stability index is used to characterize population stability, and the calculation formula is: I noim =(k num -k numlast ) / k max (23) Where, I noim is the population stability index, k numlast The number of iterations when the best position of the population was last updated.

3. The low-carbon economic dispatching method for controllable flexible loads of a PV-storage direct-flexible system according to claim 1 or 2, characterized in that: The controllable flexible load includes loads that can be translated, transferred, and reduced. The mathematical model of the load that can be translated is: Where, is the translatable load during the period t after participating in the response, is the translatable load in the period t before participating in the response, Indicates the load scheduling period that can be shifted. represents the translatable load response control variable, Indicates that the load can be translated in the time period Internal participation response, Indicates that the translatable load does not participate in the response; The mathematical model of the transferable load is: Where, represents the transferable load during the t period after participating in the response, represents the transferable load for the period t before participating in the response, represents the transferable load scheduling period, represents the transferable load response control variable, Indicates the transferable load in the time period Internal participation response, Indicates that the transferable load does not participate in the response; The mathematical model of load reduction is: in, is the load that can be reduced in period t after participating in the response, is the load that can be reduced in period t before participating in the response; represents the load response control variable that can be reduced, Indicates that the load can be reduced to participate in the response, Indicates that the load can be reduced and does not participate in the response; a cut is the load reduction factor, and its value range is [0,1].

4. The low-carbon economic dispatch method for controllable flexible loads of a PV-storage direct-flexible system according to claim 3 is characterized in that: The reward value in reinforcement learning satisfies the following formula: Where r k is the reward value of the kth iteration, G best (k), G best (k+1) are the optimal fitness values ​​of the particle swarm in the kth and k+1th iterations respectively.

Images