A Quasi-Line Model Based on Improving the Consumption Capacity of New Energy and a Method for Optimizing the Configuration of Multi-Type Energy Storage Capacity

Abstract

The present invention discloses a reference line model based on improving the consumption capacity of new energy and a multi-type energy storage capacity optimization configuration method. The model adopts an AC power flow model, and linearizes all non-linear constraints. The present invention combines the node load reference line formation model based on the linearized power flow calculation method, the deep reinforcement learning algorithm and the multi-type energy storage capacity optimization configuration method, which can effectively promote the improvement of the new energy consumption level and promote the development of "new energy + energy storage".

Description

Technical Field

[0001] The present invention relates to a quasi-line model based on improving the consumption capacity of new energy and a method for optimizing the configuration of multi-type energy storage capacities. Background Art

[0002] Since new energy power stations generally lack the ability to actively support the power grid such as the inertia support and primary frequency regulation similar to conventional synchronous units, after large-scale grid connection, the operating characteristics of the power system and the active / reactive power balance mechanism will change, which will seriously deteriorate the system frequency characteristics and the safety and stability level. At the same time, the peak shaving resources are continuously decreasing, and the pressure for new energy consumption is increasing. At present, the current situation of relying solely on conventional energy for regulation urgently needs to be changed. New energy has the ability of active frequency support, and sharing the responsibilities and obligations of the safe and stable operation of the system with conventional energy is the future technological development trend. Energy storage has the characteristics of energy time-shifting, fast response, and flexible layout, and is an important technical means to improve new energy consumption and solve the problem of insufficient active support ability of new energy. As the requirements of the power grid for new energy gradually increase from "friendly grid connection" to "friendly grid connection + active support", "new energy + energy storage" has become the general trend. In recent years, scholars at home and abroad have carried out many studies on the energy storage configuration on the new energy side. The energy storage configuration technology for promoting consumption has been relatively mature and has achieved typical engineering applications. The application modes of energy storage mainly include suppressing output fluctuations, reducing the curtailment rate, compensating for power prediction errors, etc.

[0003] At present, the application scenarios of energy storage on the new energy side are still limited to promoting new energy consumption or participating in actively supporting the power grid. The single energy storage application scenario can no longer meet the diverse needs of the current increasingly complex power system for new energy. There is no literature report on the energy storage configuration method that takes into account both the consumption level and the ability to actively support the power grid, which is an urgent technical direction for research at present. Summary of the Invention

[0004] The purpose of the present invention is to provide a quasi-line model based on improving the consumption capacity of new energy and a method for optimizing the configuration of multi-type energy storage capacities, which combines the node load quasi-line formation model, the deep reinforcement learning algorithm and the multi-type energy storage capacity optimization configuration method, and can effectively promote the improvement of new energy consumption level and promote the development of "new energy + energy storage".

[0005] The technical solution of the present invention is as follows:

[0006] A quasi-line model based on improving the consumption capacity of new energy and a method for optimizing the configuration of multi-type energy storage capacities, characterized in that: this invention patent combines the node load quasi-line formation model, the deep reinforcement learning algorithm and the multi-type energy storage capacity optimization configuration method, and can effectively promote the improvement of new energy consumption level and promote the development of "new energy + energy storage".

[0007] The present invention discloses a quasi-line model and an optimized allocation method for multi-type energy storage capacity based on improving the consumption capacity of new energy. Under the goal of carbon peaking and carbon neutrality, China is accelerating the construction of a new power system with new energy as the main body. The large-scale access of new energy has led to prominent phenomena of curtailment of wind and solar power. To improve the consumption level of new energy in the power system, first, this patent proposes a node load quasi-line formation model based on the linearized power flow calculation method. This model can guide adjustable loads to adjust their power consumption periods, thereby promoting the improvement of the new energy consumption level. At the same time, this model is an AC power flow model. Its advantage is that, compared with the DC power flow model, it considers relevant constraints of the power system such as voltage constraints, and compared with other AC power flow models, this model linearizes all non-linear constraints, resulting in lower calculation costs. Second, a deep reinforcement learning algorithm is considered to solve the master-slave game between the subjects. The deep reinforcement learning algorithm uses the marginal electricity price of each node as the state space, the load quasi-line incentive price as the action space, and the cost of the regional power grid electricity seller as the feedback. By continuously training the agent, the agent can find the load quasi-line incentive price that maximizes the interests of the regional power grid electricity seller. On this basis, an optimized allocation method for multi-type energy storage capacity is further proposed considering the improvement of new energy consumption capacity, and the solution process of the complex model is optimized by introducing a typical operating condition characteristic curve extraction method, which can cover the diverse operating condition characteristics of energy storage on the new energy side and achieve the efficient solution of the complex model, thereby achieving the purpose of improving the new energy consumption capacity.

[0008] To improve the consumption level of new energy in the power system, this paper proposes a node load quasi-line formation model based on the linearized power flow calculation method. This model uses an AC power flow model, which considers relevant constraints of the power system such as voltage constraints and power flow constraints compared with the DC power flow model. At the same time, this model linearizes all non-linear constraints, resulting in lower calculation costs compared with other AC power flow models.

[0009] A deep reinforcement learning algorithm is used to solve the master-slave game between the subjects. The deep reinforcement learning algorithm uses the marginal electricity price of each node as the state space, the load quasi-line incentive price as the action space, and the cost of the regional power grid electricity seller as the feedback. By continuously training the agent, the agent can find the load quasi-line incentive price that maximizes the interests of the regional power grid electricity seller.

[0010] An optimized allocation method for multi-type energy storage capacity is further proposed considering the improvement of new energy consumption capacity, and the solution process of the complex model is optimized by introducing a typical operating condition characteristic curve extraction method, which can cover the diverse operating condition characteristics of energy storage on the new energy side and achieve the efficient solution of the complex model, thereby achieving the purpose of improving the new energy consumption capacity.

[0011] This invention patent combines the node load reference line formation model, the deep reinforcement learning algorithm, and the multi-type energy storage capacity optimization configuration method, which can effectively promote the improvement of new energy consumption level and drive the development of "new energy + energy storage".

[0012] A reference line model based on improving the new energy consumption capacity, characterized in that: an AC power flow model is adopted, and all non-linear constraints in the model are linearized.

[0013] Denote the load of a certain node i in the power grid at time t as E i,t , assuming there are T time periods in total, taking the total load of this node in T time periods as the reference value, the normalized load curve is called the load shape E * i,t

[0014]

[0015] Taking the adjustable load participating in demand response as the research object, denote the load shape of the adjustable load of a certain node as E * i,t , and then adopt a linearized AC power flow model to formulate the node load reference line; the power flow model uses the square of the voltage amplitude and the voltage phase angle as decision variables, and converts all non-convex constraints into convex constraints; the line power flow formula of the AC network of the power system is as follows:

[0016] P ij,t =g ij (v 2 i,t -v i,t v j,t cosθ ij,t )-b ij v i,t v j,t sinθi j,t (2)

[0017] Q ij,t =-b ij (v 2 i -v i,t v j,t cosθ ij,t )-g ij v i,t v j,t sinθ ij,t (3)

[0018] Among them, P ij,t is the active power of branch (i, j) at time t, Q ij,t is the reactive power of branch (i, j) at time t, g ij is the conductance of branch (i, j), bij is the susceptance of branch (i, j), and v i,t is the voltage magnitude of node i at time t, and θ ij,t is the phase angle difference between node i and node j at time t;

[0019] According to the first-order Taylor expansion of the sine function, assuming that the voltage magnitudes and phase angles in two consecutive iterative calculations do not differ much, the square of the node voltage magnitude difference can be approximately expressed as follows:

[0020]

[0021]

[0022]

[0023] To ensure that the square of the node voltage magnitude difference is always greater than 0, the following constraint is added:

[0024] v s ij,t,L ≥0 (7)

[0025] According to the definition of the node load guideline, it can be obtained that:

[0026]

[0027]

[0028] 0 ≤ E * D,i,t,k ≤ π ≤ 1 (10)

[0029] where P u,i,t is the non-adjustable load of node i at time t, and E * D,i,t,k is the load guideline of node i in the k-th iterative calculation, and E D,i,t is the adjustable load of node i at time t, and π is a constant; this can prevent the load guideline in a certain period from being too high and not conforming to the actual situation; taking the square of the node voltage magnitude as the decision variable, the voltage constraint is as follows:

[0030] v 2 t,min ≤ v 2 i,t,k ≤ v 2 t,max (11)

[0031] The branch power flow constraint is as follows:

[0032] (P ij,t,k ) 2 +(Q ij,t,k ) 2 ≤ S2 ij,max (12)

[0033] The branch power flow constraint is a quadratic constraint. Obviously, the mathematical stability of the linear constraint model is much greater than that of the quadratic constraint. Therefore, it is necessary to use the piecewise linearization method to convert the quadratic constraint into a linear constraint;

[0034] Generator active power output constraint:

[0035] P g,min ≤P g,t,k ≤P g,max (13)

[0036] where P g,t,k is the active power output of generator t at the k-th iteration, P g,max is the maximum active power output of the generator, and P g,min is the minimum active power output of the generator;

[0037] Generator ramp rate constraint

[0038] -ΔP g D ≤P g,t,k -P g,t-1,k ≤ΔP g U (14)

[0039] where ΔP g D and ΔP g U represent the upward and downward ramp rates respectively;

[0040] The formulation of the node load reference line is to promote the consumption of new energy and reduce the power generation cost. Therefore, the goal of the power grid dispatching center is to maximize social welfare, that is, to minimize the power generation cost and the cost of abandoning wind and light of new energy; the objective function expression is as follows:

[0041]

[0042] where a g is the quadratic cost coefficient of the generator, b g is the linear cost coefficient of the generator, c g is the constant term, and k e is the wind abandonment;

[0043] The node formulation method model based on the linearized power flow calculation method is as follows:

[0044]

[0045] The deep reinforcement learning algorithm is used to solve the master-slave game between the agents; specifically, the deep deterministic policy gradient (DDPG) algorithm is used for model solving, and the policy network delay update strategy of the double-delayed deep deterministic policy gradient algorithm is introduced;

[0046] DDPG consists of a total of 4 neural networks, which are used to approximate the Q-value function and the policy function; among them, the loss function during the update of the evaluation training network is:

[0047]

[0048] where N is the size of the batch data samples used for training and updating, y i is the cumulative return, and Q w (s i , a i ) is the Q-value function of the evaluation training network outputting the state-action at the current moment;

[0049] Combined with the Q-value function of the evaluation training network, the policy gradient of the Actor during parameter update can be obtained:

[0050]

[0051] where J is the expectation of the Q-value function, π θ is the policy of the policy training network, and θ is the parameter of the policy training network;

[0052] For the update of the evaluation target network parameter w' and the action target network parameter θ, DDPG ensures that the parameters can be updated slowly through a soft update mechanism, thereby improving the stability of learning:

[0053] w'←ξw+(1-ξ)w'

[0054] θ'←ξθ+(1-ξ)θ' (19)

[0055] where ξ is the soft update coefficient.

[0056] Taking the maximum annual on-grid electricity and the maximum annual net income of multiple types of energy storage systems as the goals:

[0057] maxF=[f1,f2] (20)

[0058] In the formula: f 1 is the annual on-grid electricity of the new energy power station; f 2 is the annual net income of multiple types of energy storage systems;

[0059] After configuring multiple types of energy storage systems, multiple benefits can be obtained. The annual net income of multiple types of energy storage systems:

[0060] f 2 =I on+I Acc +I rel -C LCC C LCC = C inv +C om +C rep +C rec (21)

[0061] Where: I on is the income obtained by storing part or all of the curtailed power during the power rationing period in the energy storage system and then feeding it into the power grid during the non-power rationing period to increase the annual power generation for grid connection; I Acc is the annual mid- and short-term power prediction accuracy assessment cost of the new energy power station reduced by compensating for the mid- and short-term power prediction errors in the new energy power station; I rel is the income obtained by increasing the annual power generation for grid connection due to the exemption of active power reserve; C LCC is the annual value of the cost of multiple types of energy storage systems during the planning period;

[0062] Taking the 25-year operation cycle of the new energy power station as the planning cycle, calculate the annual value of the cost C of multiple types of energy storage systems LCC :

[0063] C LCC = C inv +C om +C rep +C rec (22)

[0064] Where: C inv is the annual equipment purchase cost; C om is the annual operation and maintenance cost, including the costs of labor, operation inspection, maintenance, etc. invested to ensure the normal operation of the energy storage power station; C rep is the annual equipment replacement cost. During the planning period, when the capacity retention rate of the battery energy storage system decays to 80%, new equipment needs to be replaced. The service life of the flywheel energy storage reaches 25 years and does not need to be replaced; C rec is the annual equipment residual value and recovery cost;

[0065]

[0066] Where: is the operation and maintenance cost per unit charge and discharge of the energy storage system; is the discharge amount of multiple types of energy storage systems at time t;

[0067]

[0068] Where: k is the number of equipment replacement times of the battery energy storage system; α is the average annual decline ratio of the energy storage system cost;

[0069]

[0070] Where: C rc and C rv are the recovery cost and the equipment salvage value generated when the battery energy storage device needs to be scrapped, respectively;

[0071] During the process of optimizing the capacity configuration of multi-type energy storage systems, constraints such as the safe and stable operation of the energy storage system, the minimum configured energy storage capacity for participating in primary frequency modulation, and the curtailment rate of new energy power stations need to be considered;

[0072] The constraints for the safe and stable operation of multi-type energy storage systems include power constraints and SOC constraints;

[0073]

[0074]

[0075]

[0076] Where: is the charging and discharging power of the battery energy storage system at time t; is the charging and discharging power of the flywheel energy storage system at time t; is the charging and discharging power of the multi-type energy storage system at time t;

[0077]

[0078]

[0079] are the state of charge of the battery and the flywheel energy storage system at time t, respectively; are the upper and lower limit values of the SOC operating range of the battery energy storage system; are the upper and lower limit values of the SOC operating range of the flywheel energy storage system;

[0080] Then, based on the chronological power demand of the energy storage, the total power demand of the energy storage considering multiple scenarios is calculated.

[0081] The beneficial effects of the present invention are as follows:

[0082] An alignment model based on improving the new energy consumption capacity and an optimization configuration of multi-type energy storage capacities are invented. To improve the new energy consumption level and solve complex models more efficiently, a node load alignment formation model based on a linearized power flow calculation method is proposed. This model uses an AC power flow model. Compared with the DC power flow model, it considers power system-related constraints such as voltage constraints and power flow constraints. At the same time, this model linearizes all non-linear constraints, and has a lower calculation cost compared with other AC power flow models.

[0083] An alignment model based on improving the consumption capacity of new energy and an optimized allocation of multi-type energy storage capacities are invented. In order to solve complex models more efficiently, a deep reinforcement learning algorithm is considered to solve the master-slave game between the main bodies. The deep reinforcement learning algorithm takes the marginal electricity price of each node as the state space, the load alignment incentive price as the action space, and the cost of the regional power grid electricity seller as the feedback. By continuously training the agent, the agent can find the load alignment incentive price that maximizes the interests of the regional power grid electricity seller.

[0084] On this basis, a method for optimizing the allocation of multi-type energy storage capacities is further proposed considering the improvement of the new energy consumption capacity. By introducing a method for extracting the characteristic curves of typical working conditions, the solution process of the complex model is optimized, which can cover the diverse working condition characteristics of the energy storage on the new energy side and achieve the efficient solution of the complex model, so as to achieve the purpose of improving the new energy consumption capacity. Brief Description of the Drawings

[0085] The present invention will be further described below in conjunction with the drawings and embodiments.

[0086] Figure 1 It is the framework diagram of the DDPG algorithm.

[0087] Figure 2 It is the flowchart for calculating the energy storage power considering multiple scenarios. Detailed Embodiments

[0088] An alignment model based on improving the consumption capacity of new energy and an optimized allocation of multi-type energy storage capacities. To improve the consumption level of new energy and solve complex models more efficiently, a node load alignment formation model based on the linearized power flow calculation method is proposed. This model uses the AC power flow model. Compared with the DC power flow model, it considers power system related constraints such as voltage constraints and power flow constraints. At the same time, this model linearizes all non-linear constraints, and has a lower calculation cost compared with other AC power flow models.

[0089] Denote the load of a certain node i in the power grid at time t as E i,t , assuming there are a total of T time periods (in this paper, the load alignment is one hour as a scheduling period, a total of 24 time periods). Taking the total load of this node in T time periods as the reference value, the normalized load curve is called the load shape E * i,t

[0090]

[0091] For this node in the power grid, the ideal load profile is called the load reference line of this node. The load reference line is only a per-unit value target, and the adjustable load aggregator independently decides whether to target the load reference line to obtain economic incentives from the power grid. This patent takes the adjustable load participating in demand response as the research object, and records the load profile of the adjustable load at a certain node as E * i,t Next, a linearized AC power flow model will be adopted to formulate the node load reference line. This power flow model uses the square of the voltage magnitude and the voltage phase angle as decision variables, and converts all non-convex constraints into convex constraints, greatly accelerating the calculation time of the optimal power flow. The line power flow formula for the AC network of the power system is as follows:

[0092] P ij,t =g ij (v 2 i,t -v i,t v j,t cosθ ij,t )-b ij v i,t v j,t sinθi j,t (2)

[0093] Q ij,t =-b ij (v 2 i -v i,t v j,t cosθ ij,t )-g ij v i,t v j,t sinθ ij,t (3)

[0094] Among them, P ij,t is the active power of branch (i, j) at time t, Q ij,t is the reactive power of branch (i, j) at time t, g ij is the conductance of branch (i, j), b ij is the susceptance of branch (i, j), v i,t is the voltage magnitude of node i at time t, θ ij,t is the phase angle difference between node i and node j at time t.

[0095] According to the first-order Taylor expansion of the sine function, assuming that the voltage magnitudes and phase angles in the previous and subsequent iterative calculations do not differ much, the square of the node voltage magnitude difference can be approximately expressed as follows:

[0096]

[0097]

[0098]

[0099] To ensure that the square of the node voltage amplitude difference is always greater than 0, the following constraints are added:

[0100] v s ij,t,L ≥0 (7)

[0101] According to the definition of the node load guideline, it can be obtained that:

[0102]

[0103]

[0104] 0 ≤ E * D,i,t,k ≤ π ≤ 1 (10)

[0105] Among them, P u,i,t is the non-adjustable load of node i at time t, E * D,i,t,k is the load guideline of node i in the k-th iteration calculation, E D,i,t is the adjustable load of node i at time t, and π is a constant. This can prevent the load guideline of a certain time period from being too high and not conforming to the actual situation. Taking the square of the node voltage amplitude as the decision variable, the voltage constraints are as follows:

[0106] v 2 t,min ≤ v 2 i,t,k ≤ v 2 t,max (11)

[0107] The branch power flow constraints are as follows:

[0108] (P ij,t,k ) 2 +(Q ij,t,k ) 2 ≤ S 2 ij,max (12)

[0109] The branch power flow constraint is a quadratic constraint. Obviously, the mathematical stability of the linear constraint model is much greater than that of the quadratic constraint. Therefore, it is necessary to use the piecewise linearization method to transform the quadratic constraint into a linear constraint.

[0110] Generator active power output constraint:

[0111] P g,min ≤ P g,t,k ≤ P g,max (13)

[0112] Among them, P g,t,k is the active power output of generator t during the kth iteration calculation, P g,max is the maximum active power output of the generator, P g,min is the minimum active power output of the generator.

[0113] Generator ramp rate constraint

[0114] -ΔP g D ≤P g,t,k -P g,t-1,k ≤ΔP g U (14)

[0115] Among them, ΔP g D and ΔP g U represent the upward and downward ramp rates respectively.

[0116] The formulation of the nodal load reference is to promote the consumption of new energy and reduce the power generation cost. Therefore, the goal of the power grid dispatching center is to maximize social welfare, that is, to minimize the power generation cost and the cost of curtailment of wind and solar power. The objective function expression is as follows:

[0117]

[0118] Among them, a g is the quadratic cost coefficient of the generator, b g is the linear cost coefficient of the generator, c g is the constant term, k e is the curtailment of wind

[0119] To sum up, the nodal formulation method model based on the linearized power flow calculation method is as follows:

[0120]

[0121] s.t. (4)-(14)

[0122] The deep reinforcement learning algorithm is used to solve the master-slave game between the agents. The deep reinforcement learning algorithm takes the marginal electricity price of each node as the state space, the incentive price of the load reference as the action space, and the cost of the regional power grid seller as the feedback. By continuously training the agent, the agent can find the incentive price of the load reference that maximizes the interests of the regional power grid seller.

[0123] In this paper, the Deep Deterministic Policy Gradient (DDPG) algorithm is used to solve the model, and at the same time, the strategy network delayed update strategy of the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm is introduced. On the one hand, compared with the Deep Q-network (DQN) algorithm, the DDPG algorithm can solve problems in the multi-dimensional continuous incentive price and node marginal price state space. On the other hand, compared with the DDPG algorithm, the DDPG algorithm with the strategy network delayed update strategy has a more stable training process and is easier to converge. Reinforcement learning means that an agent learns in a "trial and error" manner, and guides its behavior through the rewards obtained by interacting with the environment, with the goal of maximizing the rewards obtained by the agent. Deep learning refers to learning the internal laws and representation levels of data. Deep reinforcement learning combines the perception ability of deep learning and the decision-making ability of reinforcement learning, providing a solution idea for the perception and decision-making problems of complex systems.

[0124] The DDPG algorithm is a classic method based on the actor-critic structure. The DDPG method can better handle problems with a continuous and high-dimensional action space and has achieved good results in multiple application scenarios. In the DDPG algorithm architecture, a dual neural network architecture is used, and dual neural network models (i.e., training network and target network) are used for both the policy function and the value function, making the learning process of the algorithm more stable and accelerating the convergence speed. At the same time, the algorithm introduces an experience replay mechanism. The experience data samples generated by the actor interacting with the environment are stored in the experience pool, and a batch of data samples are extracted for training, which is similar to the experience replay mechanism of DQN, removing the correlation and dependence of the samples and making the algorithm easier to converge.

[0125] DDPG contains a total of 4 neural networks, which are used to approximate the Q-value function and the policy function. The loss function during the update of the evaluation training network is as follows:

[0126]

[0127] where N is the size of the batch data samples used for training and update, y i is the cumulative return, and Q w (s i , a i ) is the Q-value function of the evaluation training network outputting the state-action at the current moment.

[0128] Combined with the Q-value function of the evaluation training network, the policy gradient of the actor during parameter update can be obtained:

[0129]

[0130] where J is the expectation of the Q-value function, and π θ is the policy of the policy training network, and θ is the parameter of the policy training network.

[0131] For the update of the evaluation target network parameter w' and the action target network parameter θ, DDPG ensures the slow update of the parameters through a soft update mechanism (updating part of the parameters each time learn is called), thereby improving the stability of learning:

[0132] w'←ξw+(1-ξ)w'

[0133] θ'←ξθ+(1-ξ)θ' (19)

[0134] where ξ is the soft update coefficient.

[0135] Furthermore, a multi-type energy storage capacity optimization configuration method is proposed, and the solution process of the complex model is optimized by introducing a typical working condition characteristic curve extraction method, which can cover the diverse working condition characteristics of the energy storage on the new energy side and achieve the efficient solution of the complex model, thereby achieving the purpose of improving the new energy consumption capacity.

[0136] With the goal of maximizing the annual grid-connected power and the annual net income of the multi-type energy storage system:

[0137] maxF=[f1,f2] (20)

[0138] In the formula: f 1 is the annual grid-connected power of the new energy power station; f 2 is the annual net income of the multi-type energy storage system.

[0139] After configuring the multi-type energy storage system, multiple benefits can be obtained. The annual net income of the multi-type energy storage system:

[0140] f 2 =I on +I Acc +I rel -C LCC C LCC =C inv +C om +C rep +C rec (21)

[0141] In the formula: I on is the income obtained by storing part or all of the curtailed power during the curtailment period in the energy storage system and then feeding it into the grid during the non-curtailment period to increase the annual grid-connected power; I AccTo reduce the annual mid - short - term power prediction accuracy assessment cost of a new - energy power station due to compensating for the reduction of mid - short - term power prediction error; I rel The revenue obtained from increasing the annual on - grid electricity due to exempting the active power reserve; C LCC The annual value of the cost of a multi - type energy storage system during the planning period.

[0142] Taking the 25 - year operation cycle of the new - energy power station as the planning period, calculate the annual value of the cost C of the multi - type energy storage system LCC :

[0143] C LCC =C inv +C om +C rep +C rec (22)

[0144] In the formula: C inv Is the annual equipment purchase cost; C om Is the annual operation and maintenance cost, including the costs of labor, operation inspection, maintenance, etc. invested to ensure the normal operation of the energy storage power station; C rep Is the annual equipment replacement cost. During the planning period, when the capacity retention rate of the battery energy storage system decays to 80%, new equipment needs to be replaced. The service life of the flywheel energy storage can reach 25 years and does not need to be replaced; C rec Is the annual equipment salvage value and recovery cost.

[0145]

[0146] In the formula: Is the operation and maintenance cost per unit charge - discharge amount of the energy storage system; Is the discharge amount of the multi - type energy storage system at time t.

[0147]

[0148] In the formula: k is the number of equipment replacement times of the battery energy storage system; α is the average annual decline ratio of the energy storage system cost.

[0149]

[0150] In the formula: C rc 、C rv Are the recovery cost and equipment salvage value generated when the battery energy storage equipment needs to be scrapped, respectively.

[0151] During the process of optimizing the configuration of the multi - type energy storage system capacity, it is necessary to consider the constraints such as the safe and stable operation of the energy storage system, the minimum configuration of the energy storage capacity participating in primary frequency modulation, and the curtailment rate of the new - energy power station.

[0152] The safe and stable operation constraints of the multi - type energy storage system include power constraints and SOC constraints.

[0153]

[0154]

[0155]

[0156] wherein: is the charging and discharging power of the battery energy storage system at time t; is the charging and discharging power of the flywheel energy storage system at time t; is the charging and discharging power of the multi-type energy storage system at time t.

[0157]

[0158]

[0159] are the state of charge of the battery and the flywheel energy storage system at time t, respectively; are the upper and lower limit values of the SOC operating range of the battery energy storage system; are the upper and lower limit values of the SOC operating range of the flywheel energy storage system.

[0160] Then, based on the time-series power demand of the energy storage, calculate the total power demand of the energy storage that takes into account multiple scenarios, as shown in Figure 2 shown.

Claims

1. A quasi-line model based on improving the capacity of absorbing new energy, characterized by: The AC power flow model is used, and all nonlinear constraints are linearized in the model; The load of a node i in the power grid during period t is E i,t Assuming that there are T time periods, the total load of the node in T time periods is taken as the reference value, and the normalized load curve is called load shape E * i,t The adjustable load participating in demand response is taken as the research object, and the load shape of the adjustable load at a certain node is denoted as E * i,t , then a linearized AC power flow model is used to formulate the node load criterion; the power flow model uses the square of the voltage amplitude and the voltage phase angle as decision variables, and converts all non-convex constraints into convex constraints; the line power flow formula of the power system AC network is as follows: P ij,t =g ij (v 2 i,t -v i,t v j,t cosθ ij,t )-b ij v i,t v j,t sinθi j,t (2) Q ij,t =-b ij (v 2 i -v i,t v j,t cosθ ij,t )-g ij v i,t v j,t sinθ ij,t (3) Among them, P ij,t is the active power of branch (i, j) in period t, Q ij,t is the reactive power of branch (i, j) in period t, g ij is the conductance of branch (i, j), b ij is the susceptance of branch (i, j), v i,t is the voltage amplitude of node i at time period t, θ ij,t is the phase angle difference between node i and node j during period t; According to the first-order Taylor expansion of the sine function, assuming that the voltage amplitude and phase angle before and after the iterative calculation are not much different, the square of the node voltage amplitude difference can be approximately expressed as follows: To ensure that the square of the node voltage amplitude difference is always greater than 0, add the following constraints: v s ij,t,L ≥0 (7) According to the definition of the node load criterion, we can get: 0≤E * D,i,t,k ≤π≤1 (10) Among them, P u,i,t is the non-adjustable load of node i in period t, E * D,i,t,k is the load directrix of node i in the kth iteration calculation, E D,i,t is the adjustable load of node i in time period t, and π is a constant; it can prevent the load criterion in a certain period from being too high and not in line with the actual situation; the square of the node voltage amplitude is used as the decision variable, and the voltage constraint is as follows: in 2 t,min ≤in 2 i,t,k ≤in 2 t,max (11) The branch power flow constraints are as follows: (P ij,t,k ) 2 +(Q ij,t,k ) 2 ≤S 2 ij,max (12) The branch power flow constraint is a quadratic constraint. Obviously, the mathematical stability of the linear constraint model is much greater than that of the quadratic constraint. Therefore, it is necessary to use the piecewise linearization method to transform the quadratic constraint into a linear constraint. Generator active output constraints: P g,min ≤P g,t,k ≤P g,max (13) Among them, P g,t,k is the active output of the generator in period t in the kth iteration calculation, P g,max is the maximum active output of the generator, P g,min It is the minimum active output of the generator; Generator ramp constraint -ΔP g D ≤P g,t,k -P g,t-1,k ≤ΔP g U (14) Among them, ΔP g D and ΔP g U Represent the up and down climbing rates respectively; The formulation of the node load standard is to promote the consumption of new energy and reduce the cost of power generation. Therefore, the goal of the power grid dispatching center is to maximize social welfare, that is, to minimize the cost of power generation and the cost of new energy abandonment. The objective function expression is as follows: Among them, a g is the generator quadratic cost coefficient, b g is the primary cost coefficient of the generator, c g is a constant term, k e For abandoning the wind; The node formulation method model based on the linear power flow calculation method is as follows:

2. The criterion model based on improving the new energy consumption capacity according to claim 1 is characterized in that: A deep reinforcement learning algorithm is used to solve the master-slave game between the subjects. Specifically, the deep deterministic policy gradient DDPG algorithm is used to solve the model, and the policy network delayed update strategy of the double-delayed deep deterministic policy gradient algorithm is introduced. DDPG contains a total of 4 neural networks, which are used to approximate the Q-value function and the policy function; the loss function for evaluating the update of the training network is: Where N is the batch data sample size used for training updates, y i is the cumulative return, Q w (s i ,a i ) is the Q value function of the state-action output at the current moment for evaluating the training network; Combined with the Q-value function of the evaluation training network, we can get the policy gradient of the Actor during parameter update: Where J is the expectation of the Q-value function, π θ is the strategy of the strategy training network, θ is the parameter of the strategy training network; For the update of the evaluation target network parameters w' and the action target network parameters θ, DDPG uses a soft update mechanism to ensure that the parameters can be updated slowly, thereby improving the stability of learning: w'←ξw+(1-ξ)w' θ'←ξθ+(1-ξ)θ' (19) Where ξ is the soft update coefficient.

3. A multi-type energy storage capacity optimization configuration method based on a quasi-line model for improving the new energy consumption capacity according to claim 1, characterized in that: The goal is to maximize the annual grid-connected electricity and the annual net profit of multiple types of energy storage systems: max F=[f1,f2] (20) Where: f1 is the annual grid-connected electricity of the new energy power station; f2 is the annual net income of the multi-type energy storage system; After configuring multiple types of energy storage systems, you can obtain multiple benefits. The annual net benefits of multiple types of energy storage systems are: f2=I on +I Acc +I rel -C LCC C LCC =C inv +C om +C rep +C rec (21) Where: I on To store part or all of the power wasted during the power-limited period in the energy storage system, and then feed it into the grid during the non-power-limited period to increase the annual grid-connected power revenue; Acc To reduce the annual short- and medium-term power forecast accuracy assessment costs of new energy power stations by compensating for the short- and medium-term power forecast errors of new energy power stations; rel The income from increasing the annual grid-connected power due to the exemption of active power reserve reserve; C LCC The annual cost value of various types of energy storage systems during the planning period; Taking the operation period of new energy power station as 25 years as the planning period, calculate the annual cost value C of various types of energy storage systems LCC : C LCC =C inv +C om +C rep +C rec (22) Where: C inv is the annual equipment purchase cost; C om is the annual operation and maintenance cost, including the labor, operation and inspection, and maintenance costs invested to ensure the normal operation of the energy storage power station; C rep is the annual equipment replacement cost. In the planned period, when the capacity retention rate of the battery energy storage system decays to 80%, new equipment needs to be replaced. The service life of the flywheel energy storage is 25 years and does not need to be replaced. rec is the annual equipment residual value and recovery cost; Where: The operation and maintenance cost per unit charge and discharge capacity of the energy storage system; is the discharge amount of multi-type energy storage systems in period t; Where: k is the number of times the battery energy storage system equipment is replaced; α is the average annual reduction ratio of the energy storage system cost; Where: C rc , C rv They are the recycling cost and residual value of the battery energy storage equipment when it needs to be scrapped; In the process of optimizing the capacity of multi-type energy storage systems, it is necessary to consider constraints such as the safe and stable operation of the energy storage system, the minimum configuration of energy storage capacity participating in primary frequency regulation, and the power abandonment rate of new energy power stations; The safe and stable operation constraints of multi-type energy storage systems include power constraints and SOC constraints; Where: is the charging and discharging power of the battery energy storage system during period t; is the charge and discharge power of the flywheel energy storage system during period t; is the charging and discharging power of the multi-type energy storage system during period t; are the charge states of the battery and flywheel energy storage system during period t respectively; The upper and lower limits of the SOC operating range of the battery energy storage system; The upper and lower limits of the SOC operating range of the flywheel energy storage system; Then, based on the energy storage time-series power demand, the total energy storage power demand taking into account multiple scenarios is calculated.

Images