Virtual power plant optimization scheduling method based on artificial intelligence

Abstract

The invention discloses a virtual power plant optimal scheduling method based on artificial intelligence, and further reduces the power operation cost through an integrated intelligent management and scheduling strategy. The system is composed of a power generation end, an energy storage end, a power utilization end, a system management platform, a power grid, a power market and other key components. The core lies in that a depth deterministic strategy gradient (DDPG) algorithm is utilized to construct a simulation environment model, and a reward function is designed to evaluate the economic benefit of the strategy. By initializing Actor and Critic neural networks, the system can generate and evaluate a scheduling strategy, and continuously iteratively updates the network through an experience playback mechanism to optimize economic benefits. The invention provides an innovative virtual power plant optimal scheduling method based on artificial intelligence, and maximization of the benefits of the virtual power plant is realized by reducing the operation cost.

Description

Technical Field

[0001] The present invention relates to the field of virtual power plant optimization scheduling, and in particular to an artificial intelligence-based virtual power plant optimization scheduling method. Background Art

[0002] As an advanced energy management system, the background technology of virtual power plant is the integration and innovation of various technologies. This system realizes real-time monitoring and management of distributed energy resources through smart grid technology, centrally manages scattered small-scale power generation resources, energy storage equipment and power load, and forms a unified dispatch energy supply platform. The automation and intelligent characteristics of smart grid enable virtual power plant to efficiently respond to changes in supply and demand of power grid and optimize the allocation of power resources.

[0003] One of the core technologies of virtual power plants is artificial intelligence, especially machine learning algorithms, which can predict electricity demand and supply and optimize power generation and storage operations.

[0004] In the prior art, technicians in this field mostly use artificial intelligence algorithms to predict and schedule virtual power plants, with energy saving or cost reduction as the optimization goal, and obtain a series of virtual power plant scheduling parameters. However, in the field of virtual power plants, a single algorithm itself lacks adaptive optimization, and scheduling is difficult to achieve the ideal optimization target level. Therefore, it is necessary to select an adaptive scheduling algorithm and improve it, optimize the production structure so that VPP scheduling can further reduce costs and improve efficiency. Summary of the invention

[0005] Purpose of the invention: To solve the problems mentioned in the background technology, the present invention discloses an optimization scheduling method for virtual power plants based on artificial intelligence. By improving the DDPG algorithm and the adaptive chaos enhanced multi-model fusion ACEMF strategy, the reasonable calculation of energy generation and distribution is further refined, the operating costs of virtual power plants are reduced, and the economic benefits and market competitiveness are improved while ensuring the stability of the power grid.

[0006] Technical solution:

[0007] The present invention provides a virtual power plant optimization scheduling method based on artificial intelligence, the method comprising the following steps:

[0008] S1 establishes a virtual power plant system, including power generation, energy storage and system management platform;

[0009] The S2 power generation terminal integrates the wind power generation model and the photovoltaic power generation model to calculate the total cost C of wind power generation w and the total cost of photovoltaic power generation C p ;

[0010] S3 energy storage terminal integrates and builds energy storage cost model to calculate the total energy storage cost CES ;

[0011] The S4 system management platform connects the power generation end and the energy storage end. Based on the improved DDPG algorithm, a VPP operating cost and scheduling model is constructed to simulate VPP operation. The adaptive chaos enhanced multi-model fusion ACEMF strategy is introduced to obtain the virtual power plant strategy target when the VPP operating cost is minimized.

[0012] Furthermore, the integrated model in the power generation end of S1 is as follows:

[0013] Wind power generation model:

[0014]

[0015] Where V(t) is the wind speed at time t, P m,w [v(t)] is the output of the mth wind turbine at the moment when the wind speed is V(t); V r is the cut-in wind speed; V R is the rated wind speed of the fan; V c is the cut-out wind speed; P m,w-N is the rated output of the mth fan;

[0016] Total cost of wind power generation:

[0017]

[0018] In the formula, C w is the total operating cost of the fan output; c w is the single operating cost of wind power generation; M is the total number of wind turbines; P k,w [V(t)] is the output of the kth wind turbine at the moment when the wind speed is V(t);

[0019] Photovoltaic power generation model:

[0020] The light radiation intensity follows the Beta distribution, as follows:

[0021]

[0022] Where Γ(.) is the gamma function; θ represents solar radiation; f(θ) is the probability distribution when the solar radiation intensity is θ; a and b are the shape parameters of the Beta function. After obtaining the expectation and variance of the solar radiation intensity, a and b are calculated.

[0023] The output power of a photovoltaic unit is related to factors such as sunshine intensity and conversion efficiency as follows:

[0024] P P,t =η p S p G t

[0025] Where P P,t Represents the output of the photovoltaic unit at time t; η p Represents the efficiency of photovoltaic power conversion; S p Represents the area of ​​the photovoltaic panel; G t represents the sunshine intensity at time t;

[0026] The total cost of photovoltaic power generation:

[0027]

[0028] In the formula, C p is the total operating cost of photovoltaic power output; c p is the single operating cost of photovoltaic power generation; M is the total number of photovoltaic power generation units; P P,t The output of the kth photovoltaic panel at time t.

[0029] Furthermore, the specific steps of S2 are:

[0030] Establishing the energy storage end model: Since the energy storage itself has the ability to charge and discharge, its output is not affected by the external environment. The distributed energy storage output model can be expressed as:

[0031] -P n,max <P n,ES(t) <P n,max

[0032] Where P n,ES(t) is the output of the nth energy storage device at time t; P n,max The maximum output of energy storage;

[0033] The total cost of energy storage is:

[0034]

[0035] In the formula, C ES is the total cost of energy storage output; c ES The output of a single energy storage device.

[0036] Furthermore, the VPP operating cost and scheduling model in S4 is constructed as follows:

[0037] S4.1 Establish the objective function:

[0038] Taking the VPP operating cost as the optimization target, the cost includes the power purchase cost or power sales revenue between the VPP and the power grid, as well as the cost required for the energy storage end to convert from the charging state to the discharging state, and establish the objective function:

[0039]

[0040] C=CES +C b +C p +C w +C v

[0041] Where, C is the VPP operating cost; C ES is the total cost of energy storage; C b The cost of purchasing electricity from the power grid; C p is the total cost of photovoltaic power generation, C w is the total cost of wind power generation, C v To sell electricity to the grid for profit, Γ B (t) is the price of electricity purchased by VPP from the large power grid, P max (t) is the exchange power between VPP and the large power grid during period t. It is positive when VPP purchases power from the large power grid and negative when VPP sells power to the large power grid. S (t) is the price of electricity sold by VPP to the large power grid, Δt is the time change;

[0042] VPP optimization scheduling must meet the power balance constraints in all time periods:

[0043] P m,w [v(t)]+P PV,t +P n,ES(t) +P max (t) = P load (t)

[0044] Where: P load (t) is the residential electricity load power in period t;

[0045] S4.2 uses the deep deterministic decision gradient optimization algorithm (DDPG) to build an environmental model that simulates VPP operation, and designs a reward function to evaluate the economic benefits of the strategy to minimize the VPP operating costs in the objective function:

[0046] S4.2.1 Initialize the neural network,

[0047] The parameters of actor and critic are expressed as:

[0048]

[0049] a=μ(s|θ μ )

[0050] Where: μ, Q are the strategy function and value function; θ μ ,θ Q are the policy network parameters and value network parameters; copy θ μ ,θ Q To the target network to get θ μ′ ,θ Q′; S, a are input state and input action;

[0051] S4.2.2 Initialize the experience replay pool R, use the experience replay buffer to store the total cost of the energy storage end, the cost of purchasing electricity from the grid, the total cost of photovoltaic power generation, the total cost of wind power generation, and the profit of selling electricity to the grid as experience samples to train the neural network, realize the objective function fitting, and minimize the VPP cost. In the VPP operation, the current state, action, reward, next state and other information are stored in the experience buffer, and the impact of different strategies on the objective function is evaluated to optimize the VPP scheduling strategy;

[0052] S4.2.3 Update the Critic network: randomly sample a batch of experience samples from the experience playback buffer to calculate the target Q value, update the weight of the Critic network through the gradient descent method to make it close to the target Q value, monitor and evaluate the effects of different strategies, including cost, benefit and system stability, identify strategies to reduce the discharge cost of the energy storage end, reduce the electricity purchase cost and increase the profit of electricity sales, and use the time difference algorithm to calculate the network parameters θ of the Critic network. Q ;

[0053] S4.2.4 Update the Actor network: Generate the action of the current state through the Actor network, use the Critic network to estimate its value function, evaluate the Critic network to generate a better scheduling strategy, update the Actor network weights through the gradient ascent method to generate a better action strategy, learn how to more effectively manage the discharge of the energy storage end, optimize the power purchase and sales strategy to minimize the operating cost of the VPP, and use the deterministic policy gradient algorithm to calculate the network parameters θ of the Actor. μ ;

[0054] S4.2.5 Use a sliding average method to soft-update the objective function to ensure that the strategy continuously optimizes the operating costs of the VPP;

[0055] S4.2.6 Repeat steps S4.2.2 to S4.2.5 to iteratively train the VPP operating cost and scheduling model.

[0056] Furthermore, the reward function is detailed as follows:

[0057] The expression of the reward function is:

[0058]

[0059] In the formula, R(s,a) is the immediate reward after taking action a in state s, R t is the immediate reward at time t, and γ is the discount factor.

[0060] Furthermore, the improved DDPG algorithm described in S4 introduces the adaptive chaos enhanced multi-model fusion ACEMF strategy process as follows:

[0061] The ACEMF strategy aims to integrate the randomness of chaos theory and the robustness of multi-model fusion. The core is to build multiple expert models, each of which focuses on different aspects of VPP scheduling: power generation efficiency, cost minimization and environmental impact. The expert models are all based on the DDPG algorithm and capture their respective optimization goals through different parameter initialization and reward functions. i represents the i-th expert model, where i = 1, 2, ..., n, n is the total number of expert models. The strategy update process of each expert model introduces chaotic perturbations, and the intensity of chaotic perturbations is dynamically adjusted through an adaptive mechanism. t For the corresponding chaotic disturbance intensity, the strategy parameter update formula is defined as follows:

[0062]

[0063] in, represents the strategy parameter of the i-th expert model at the t-th iteration, is the performance measure, i.e., the expected return of the strategy under the current parameters, is the disturbance term generated according to the chaotic map;

[0064] The multi-model fusion rule combines the decisions of n expert models to form the final scheduling strategy. is the weight of the i-th expert model at the t-th iteration, satisfying The final strategy can be expressed as:

[0065]

[0066] Among them, P t i is the performance evaluation of the i-th expert model at the t-th iteration, f(P t ) is a function that adjusts the intensity of chaotic perturbations according to the overall performance evaluation.

[0067] Beneficial effects:

[0068] 1. The present invention is based on the deep deterministic policy gradient (DDPG) algorithm, assisted by the improvement of the adaptive chaos enhanced multi-model fusion ACEMF strategy, which further refines the VPP scheduling scheme with the minimum operating cost as the goal. At the same time, based on the joint action of the DDPG algorithm and the ACEMF strategy, the multi-expert model fusion optimization decision is improved to improve the effectiveness of the algorithm in processing VPP scheduling under different weather conditions, making it more suitable for different VPP scheduling occasions, and the adaptability of the method is further improved.

[0069] 2. The VPP of the present invention is mainly based on two new energy reserve scheduling methods: photovoltaic and wind power. Through the DDPG algorithm and the establishment of a power generation model, its multi-objective parallel optimization characteristics are used to accurately make scheduling decisions for new energy power generation under unstable power fluctuations, further reducing energy waste and cost consumption.

[0070] 3. The present invention sets a reward function in accordance with the reduction of electricity operating costs, so that the algorithm can converge to the optimal strategy more quickly, further enhancing the response speed and robustness of the method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a schematic diagram of the virtual power plant structure;

[0072] Figure 2 It is the algorithm flow chart of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0074] like Figure 1 and Figure 2 As shown, the embodiment of the present invention discloses a virtual power plant optimization scheduling method based on artificial intelligence, and the method steps are as follows:

[0075] S1 establishes a virtual power plant system, including power generation, energy storage, power consumption, power grid, power market and system management platform;

[0076] On the power generation side, it is characterized by its ability to aggregate scattered and diverse small-scale power generation resources into a unified dispatchable and highly controllable energy supply platform through intelligent management and market participation capabilities, so as to achieve optimized operation and maximize economic benefits.

[0077] On the energy storage side, its characteristics are reflected in its ability to efficiently integrate and manage various energy storage systems, such as battery energy storage, pumped storage, compressed air energy storage, etc., to smooth the intermittency of renewable energy, provide grid regulation services, and enhance the flexibility and reliability of the power system.

[0078] The electricity consumption end is characterized by the use of intelligent demand response technology and load management strategies to achieve precise control and optimized scheduling of electricity loads, so as to adapt to changes in power grid supply and demand, improve energy utilization efficiency, and participate in market transactions to obtain economic benefits.

[0079] System management platform: Virtual power plants require a set of technologies and systems to intelligently aggregate and manage the above resources, including signal transmission and reception, data analysis, aggregate routing, and operation strategies.

[0080] Power grid, virtual power plants provide a new management and operation mode for the power grid, which helps the power grid to achieve more efficient, environmentally friendly and economical operation.

[0081] In the electricity market, virtual power plants need to interact with the electricity market, participate in electricity trading and ancillary service markets, and achieve optimal resource allocation and economic benefits.

[0082] The S2 power generation terminal integrates the wind power generation model and the photovoltaic power generation model to calculate the total cost C of wind power generation w and the total cost of photovoltaic power generation C p ;

[0083] Wind power generation model: Wind power generation mainly depends on the input wind speed and light intensity, so the wind power generation model can be expressed as:

[0084]

[0085] Where V(t) is the wind speed at time t, P m,w [v(t)] is the output of the mth wind turbine at the moment when the wind speed is V(t); V r is the cut-in wind speed; V R is the rated wind speed of the fan; V c is the cut-out wind speed; P m,w-N is the rated output of the mth fan;

[0086] The total cost of wind power generation is:

[0087]

[0088] In the formula, C w is the total operating cost of the fan output; c w is the single operating cost of wind power generation; M is the total number of wind turbines; P k,w [V(t)] is the output of the kth wind turbine when the wind speed is V(t).

[0089] Photovoltaic power generation model: There is a certain relationship between the output power of photovoltaic power generation and the intensity of solar radiation. According to research, the intensity of solar radiation follows the Beta distribution, as follows:

[0090]

[0091] Where Γ(.) is the gamma function; θ represents solar radiation; f(θ) is the probability distribution when the solar radiation intensity is θ; a and b are the shape parameters of the Beta function. After obtaining the expectation and variance of the solar radiation intensity, a and b can be calculated.

[0092] The output power of a photovoltaic unit is related to factors such as sunshine intensity and conversion efficiency as follows: P,t =η p S p G t

[0093] Where P P,t Represents the output of the photovoltaic unit at time t; η p Represents the efficiency of photovoltaic power conversion; S p Represents the area of ​​the photovoltaic panel; G t represents the sunshine intensity at time t.

[0094] The total cost of photovoltaic power generation is:

[0095]

[0096] In the formula, C p is the total operating cost of photovoltaic power output; c p is the single operating cost of photovoltaic power generation; M is the total number of photovoltaic power generation units; P P,t The output of the kth photovoltaic panel at time t.

[0097] S3 energy storage terminal integrates and builds energy storage cost model to calculate the total energy storage cost C ES ;

[0098] Since the energy storage itself has the ability to charge and discharge, its output is not affected by the external environment. The distributed energy storage output model can be expressed as:

[0099] -P n,max <P n,ES(t) <P n,max

[0100] Where P n,ES(t) is the output of the nth energy storage device at time t; P n,max The maximum output of energy storage.

[0101] The total cost of energy storage is:

[0102]

[0103] In the formula, C ES is the total cost of energy storage output; c ES The output of a single energy storage device.

[0104] The S4 system management platform connects the power generation end and the energy storage end. Based on the improved DDPG algorithm, a VPP operating cost and scheduling model is constructed to simulate VPP operation. The adaptive chaos enhanced multi-model fusion ACEMF strategy is introduced to obtain the virtual power plant strategy target when the VPP operating cost is minimized. The steps are as follows:

[0105] S1: Establish the objective function:

[0106] In order to reflect the economic efficiency of VPP optimization scheduling, the VPP operating cost is considered as the optimization target. The cost includes the power purchase cost or power sales revenue between VPP and the power grid, as well as the cost required for the energy storage end to convert from the charging state to the discharging state. The objective function is established as follows:

[0107] C=C ES +C b +C p +C w +C v

[0108] Where, C is the VPP operating cost; C ES is the total cost of energy storage; C b The cost of purchasing electricity from the power grid; C p is the total cost of photovoltaic power generation, C w is the total cost of wind power generation, C v To make profits from selling electricity to the grid.

[0109] in,

[0110]

[0111] Where: is Γ B (t) is the price of electricity purchased by VPP from the large power grid, P max (t) is the exchange power between VPP and the large power grid during period t. It is positive when VPP purchases power from the large power grid and negative when VPP sells power to the large power grid. S (t) is the price of electricity sold by VPP to the large power grid, and Δt is the time change.

[0112] VPP optimization scheduling must meet the power balance constraints in all time periods:

[0113] P m,w [v(t)]+P PV,t +P n,ES(t) +P max (t) = P load (t)

[0114] Where: P load (t) is the residential electricity load power during period t.

[0115] S2: Using the Deep Deterministic Decision Gradient Optimization (DDPG) algorithm, we build an environment model that simulates VPP operations and design a reward function to evaluate the economic benefits of the strategy to minimize the VPP operating costs in the objective function:

[0116] The expression of the reward function is:

[0117]

[0118] In the formula, R(s,a) is the immediate reward after taking action a in state s, R t is the immediate reward at time t, and γ is the discount factor.

[0119] The reward function is designed to incentivize strategies that can reduce the discharge cost of energy storage, reduce the cost of purchasing electricity from the grid, and increase the profit of selling electricity to the grid. Next, initialize the Actor and Critic neural networks, the Actor network generates scheduling strategies, and the Critic network evaluates the value of these strategies. The operating experience is stored through the experience replay mechanism, and the experience is used to iteratively update the two networks. During the training process, the strategy is gradually adjusted to optimize the economic benefits, and finally the trained strategy is applied to the actual scheduling of VPP to maximize cost savings and market benefits.

[0120] S2.1: Initialize Neural Networks: Initialize two neural networks, an Actor network for strategies and a Critic network for evaluating value, so that you can start testing different strategy combinations in a simulation environment to find the best strategy to minimize the operating costs of the VPP. These two networks can be neural network structures of any depth. The initialization of these two networks is to be able to start generating and evaluating scheduling strategies, that is, strategies that can minimize the total cost by optimizing the discharge, purchase and sale of electricity at the energy storage end. This is the basis for achieving the optimization of VPP operating costs mentioned in the objective function.

[0121] The parameters of actor and critic are expressed as:

[0122]

[0123] a=μ(s|θ μ )

[0124] Where: μ, Q are the strategy function and value function; θ μ ,θ Q are the policy network parameters and value network parameters; copy θ μ ,θ Q To the target network to get θ μ' ,θ Q' ; S and a are input status and input action.

[0125] S2.2: Experience replay: Initialize the experience replay pool R, and use the experience replay buffer to store data such as the total cost of the energy storage end, the cost of purchasing electricity from the grid, the total cost of photovoltaic power generation, the total cost of wind power generation, and the profit of selling electricity to the grid as experience samples. After collecting the data, these data will be used to train the neural network to better predict and optimize the scheduling strategy of the VPP, so as to achieve the cost minimization mentioned in the objective function. In the VPP operation, the current state, action, reward, next state and other information are stored in the experience buffer, and the impact of different strategies on the objective function is evaluated to optimize the scheduling strategy of the VPP.

[0126] S2.3: Update the Critic network: Randomly sample a batch of experience samples from the experience playback buffer to calculate the target Q-value as the optimization target of the Critic network. Update the weight of the Critic network through the gradient descent method to make it close to the target Q-value, so as to monitor and evaluate the effects of different strategies, including cost, benefit and system stability, and further help identify those strategies that can reduce the discharge cost of the energy storage end, reduce the cost of purchasing electricity and increase the profit of selling electricity, which has a positive impact on the objective function. The network parameter θ of the Critic network Q The update adopts the temporal difference (TD) algorithm, in the form of TDerror, and the loss function L is to minimize the mean square error:

[0127]

[0128] Where: y i Calculate the state action pair (s i+1 ,a i+1 ) is the target value; γ is the target value loss factor.

[0129] S2.4: Update the Actor network: Generate the action of the current state through the Actor network, and use the Critic network to estimate its value function, and evaluate the Critic network to generate a better scheduling strategy. Update the weight of the Actor network through the gradient ascent method to generate a better action strategy. This means that the Actor network will learn how to more effectively manage the discharge of the energy storage end, optimize the power purchase and sales strategy, and minimize the operating costs of the VPP. The network parameter θ of the Actor μ The update of adopts the deterministic policy gradient algorithm, which is expressed as:

[0130]

[0131] Where: is a Jacobian matrix, and the i-th column of the matrix represents the gradient of the i-th dimension of the strategy relative to the strategy parameter θ

[0132] S2.5: Update the objective function: Regularly update the target Critic network and the target Actor network, and gradually update the weights of the current network to the target network through soft updates or exponential moving averages. This soft update or exponential moving average method helps stabilize the learning process and ensures that our strategy can continuously optimize the operating costs of VPP. In this way, we can ensure the long-term effectiveness of the strategy. The parameters in the target network are soft updated using a sliding average method:

[0133]

[0134] Where: τ is the network parameter loss factor.

[0135] S2.6: Iterative training: Repeat steps S2.2 to S2.5 to continuously update and optimize the strategy based on new data and experience. This iterative training allows the models of each section of the VPP to learn from new data and adapt to changes in grid conditions, helping to ensure that the strategy can continuously minimize the discharge costs of the energy storage end, optimize the purchase and sale of electricity, and thus minimize the operating costs of the VPP.

[0136] The DDPG algorithm may have calculation errors in the process of calculating the VPP operating costs, making the results not optimal, so the algorithm needs to be optimized. The present invention adopts an innovative adaptive chaos enhanced multi-model fusion (ACEMF) strategy, which aims to significantly improve the performance and adaptability of the algorithm by integrating the randomness of chaos theory and the robustness of multi-model fusion. The core of this strategy is to build multiple expert models, each of which focuses on different aspects of VPP scheduling, such as power generation efficiency, cost minimization and environmental impact. These expert models are all based on the DDPG algorithm, but capture their respective optimization objectives through different parameter initialization and reward functions. i represents the i-th expert model, where i=1, 2, ..., n, and n is the total number of expert models.

[0137] The strategy update process of each expert model introduces chaotic perturbations to enhance its exploration ability. The intensity of chaotic perturbations is dynamically adjusted through an adaptive mechanism based on model performance feedback. t For the corresponding chaotic disturbance intensity, we define the strategy parameter update formula as follows:

[0138]

[0139] in, represents the strategy parameter of the i-th expert model at the t-th iteration, is the performance measure, i.e., the expected return of the strategy under the current parameters, is the disturbance term generated according to the chaotic map.

[0140] The multi-model fusion rule combines the decisions of each expert model to form the final scheduling strategy. is the weight of the i-th expert model at the t-th iteration, satisfying The final strategy can be expressed as:

[0141]

[0142] Weight Dynamically adjust the model based on its performance on historical data and real-time data to optimize the performance of the overall scheduling strategy.

[0143] The performance feedback and model update mechanism ensures that the algorithm can be adjusted according to real-time data and performance feedback. t is the performance evaluation index for the tth iteration, and the adjustment formula of the weight and chaotic perturbation intensity can be expressed as:

[0144]

[0145] α t+1 =α t ·f(P t )

[0146] Among them, P t i is the performance evaluation of the i-th expert model at the t-th iteration, f(P t ) is a function that adjusts the intensity of chaotic perturbations according to the overall performance evaluation.

[0147] In terms of expected effects, the ACEMF strategy significantly improves the efficiency and accuracy of VPP optimal scheduling through enhanced exploration capabilities and robustness. This strategy not only improves the global search capability of the algorithm, but also enhances its adaptability to changes in grid conditions. The ACEMF strategy provides a new improvement idea for the application of the DDPG algorithm in VPP optimal scheduling. By introducing the characteristics of chaos theory and multi-model fusion mechanism, this strategy is expected to reduce calculation errors, improve economic benefits, and enhance the robustness and global search capabilities of the algorithm. Future work will focus on further optimization of the ACEMF strategy and verification in actual VPP systems.

[0148] The above description of the embodiments enables professionals and technicians in the field to implement or use the present invention. Various modifications to the embodiments will be apparent to professionals and technicians. The general principles of the present invention can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should cover the widest range consistent with the principles and novel features disclosed in the present invention.

Claims

1. A virtual power plant optimization scheduling method based on artificial intelligence, characterized in that: The method comprises the following steps: S1 establishes a virtual power plant system, including power generation, energy storage and system management platform; The S2 power generation terminal integrates the wind power generation model and the photovoltaic power generation model to calculate the total cost C of wind power generation w and the total cost of photovoltaic power generation C p ; S3 energy storage terminal integrates and builds energy storage cost model to calculate the total energy storage cost C ES ; The S4 system management platform connects the power generation end and the energy storage end. Based on the improved DDPG algorithm, a VPP operating cost and scheduling model is constructed to simulate VPP operation. The adaptive chaos enhanced multi-model fusion ACEMF strategy is introduced to obtain the virtual power plant strategy target when the VPP operating cost is minimized.

2. The method for optimizing and dispatching a virtual power plant based on artificial intelligence according to claim 1 is characterized in that: The integrated model in the power generation end described in S1 is as follows: Wind power generation model: Where V(t) is the wind speed at time t, P m,w [v(t)] is the output of the mth wind turbine at the moment when the wind speed is V(t); V r is the cut-in wind speed; V R is the rated wind speed of the fan; V c is the cut-out wind speed; P m,w-N is the rated output of the mth fan; Total cost of wind power generation: In the formula, C w is the total operating cost of the fan output; c w is the single operating cost of wind power generation; M is the total number of wind turbines; P k,w [V(t)] is the output of the kth wind turbine at the moment when the wind speed is V(t); Photovoltaic power generation model: The light radiation intensity follows the Beta distribution, as follows: Where Γ(.) is the gamma function; θ represents solar radiation; f(θ) is the probability distribution when the solar radiation intensity is θ; a and b are the shape parameters of the Beta function. After obtaining the expectation and variance of the solar radiation intensity, a and b are calculated. The output power of a photovoltaic unit is related to factors such as sunshine intensity and conversion efficiency as follows: P P,t =the p S p G t Where P P,t Represents the output of the photovoltaic unit at time t; η p Represents the efficiency of photovoltaic power conversion; S p Represents the area of ​​the photovoltaic panel; G t represents the sunshine intensity at time t; The total cost of photovoltaic power generation: In the formula, C p is the total operating cost of photovoltaic power output; c p is the single operating cost of photovoltaic power generation; M is the total number of photovoltaic power generation units; P P,t The output of the kth photovoltaic panel at time t.

3. The method for optimizing and dispatching a virtual power plant based on artificial intelligence according to claim 1 is characterized in that: The specific steps of S2 are: Establishing the energy storage end model: Since the energy storage itself has the ability to charge and discharge, its output is not affected by the external environment. The distributed energy storage output model can be expressed as: -P n,max <P n,ES(t) <P n,max Where P n,ES(t) is the output of the nth energy storage device at time t; P n,max The maximum output of energy storage; The total cost of energy storage is: In the formula, C ES is the total cost of energy storage output; c ES The output of a single energy storage device.

4. The method for optimizing and dispatching a virtual power plant based on artificial intelligence according to any one of claims 1 to 3, characterized in that: The VPP operating cost and scheduling model in S4 is constructed as follows: S4.1 Establish the objective function: Taking the VPP operating cost as the optimization target, the cost includes the power purchase cost or power sales revenue between the VPP and the power grid, as well as the cost required for the energy storage end to convert from the charging state to the discharging state, and establish the objective function: C=C ES +C b +C p +C w +C v Where, C is the VPP operating cost; C ES is the total cost of energy storage; C b The cost of purchasing electricity from the power grid; C p is the total cost of photovoltaic power generation, C w is the total cost of wind power generation, C v To sell electricity to the grid for profit, Γ B (t) is the price of electricity purchased by VPP from the large power grid, P max (t) is the exchange power between VPP and the large power grid during period t. It is positive when VPP purchases power from the large power grid and negative when VPP sells power to the large power grid. S (t) is the price of electricity sold by VPP to the large power grid, Δt is the time change; VPP optimization scheduling must meet the power balance constraints in all time periods: P m,w [v(t)]+P PV,t +P n,ES(t) +P max (t)=P load (t) Where: P load (t) is the residential electricity load power in period t; S4.2 uses the deep deterministic decision gradient optimization algorithm (DDPG) to build an environmental model that simulates VPP operation, and designs a reward function to evaluate the economic benefits of the strategy to minimize the VPP operating costs in the objective function: S4.2.1 Initialize the neural network, The parameters of actor and critic are expressed as: a=μ(s|θ μ ) Where: μ, Q are the strategy function and value function; θ μ ,θ Q are the policy network parameters and value network parameters; copy θ μ ,θ Q To the target network to get θ μ′ ,θ Q′ ; S, a are input state and input action; S4.2.2 Initialize the experience replay pool R, use the experience replay buffer to store the total cost of the energy storage end, the cost of purchasing electricity from the grid, the total cost of photovoltaic power generation, the total cost of wind power generation, and the profit of selling electricity to the grid as experience samples to train the neural network, realize the objective function fitting, and minimize the VPP cost. In the VPP operation, the current state, action, reward, next state and other information are stored in the experience buffer, and the impact of different strategies on the objective function is evaluated to optimize the VPP scheduling strategy; S4.2.3 Update the Critic network: randomly sample a batch of experience samples from the experience playback buffer to calculate the target Q value, update the weight of the Critic network through the gradient descent method to make it close to the target Q value, monitor and evaluate the effects of different strategies, including cost, benefit and system stability, identify strategies to reduce the discharge cost of the energy storage end, reduce the electricity purchase cost and increase the profit of electricity sales, and use the time difference algorithm to calculate the network parameters θ of the Critic network. Q ; S4.2.4 Update the Actor network: Generate the action of the current state through the Actor network, use the Critic network to estimate its value function, evaluate the Critic network to generate a better scheduling strategy, update the Actor network weights through the gradient ascent method to generate a better action strategy, learn how to more effectively manage the discharge of the energy storage end, optimize the power purchase and sales strategy to minimize the operating cost of the VPP, and use the deterministic policy gradient algorithm to calculate the network parameters θ of the Actor. μ ; S4.2.5 Use a sliding average method to soft-update the objective function to ensure that the strategy continuously optimizes the operating costs of the VPP; S4.2.6 Repeat steps S4.2.2 to S4.2.5 to iteratively train the VPP operating cost and scheduling model.

5. The method for optimizing and dispatching a virtual power plant based on artificial intelligence according to claim 4 is characterized in that: The reward function is as follows: The expression of the reward function is: In the formula, R(s,a) is the immediate reward after taking action a in state s, R t is the immediate reward at time t, and γ is the discount factor.

6. The method for optimizing and dispatching a virtual power plant based on artificial intelligence according to claim 4 is characterized in that: The improved DDPG algorithm described in S4 introduces the adaptive chaos enhanced multi-model fusion ACEMF strategy process as follows: The ACEMF strategy aims to integrate the randomness of chaos theory and the robustness of multi-model fusion. The core is to build multiple expert models, each of which focuses on different aspects of VPP scheduling: power generation efficiency, cost minimization and environmental impact. The expert models are all based on the DDPG algorithm and capture their respective optimization goals through different parameter initialization and reward functions. i represents the i-th expert model, where i = 1, 2, ..., n, n is the total number of expert models. The strategy update process of each expert model introduces chaotic perturbations, and the intensity of chaotic perturbations is dynamically adjusted through an adaptive mechanism. t For the corresponding chaotic disturbance intensity, the strategy parameter update formula is defined as follows: Among them, θ t i represents the strategy parameter of the i-th expert model at the t-th iteration, is the performance measure, i.e., the expected return of the strategy under the current parameters, is the disturbance term generated according to the chaotic map; The multi-model fusion rule combines the decisions of n expert models to form the final scheduling strategy. is the weight of the i-th expert model at the t-th iteration, satisfying The final strategy can be expressed as: Among them, P t i is the performance evaluation of the i-th expert model at the t-th iteration, f(P t ) is a function that adjusts the intensity of chaotic perturbations according to the overall performance evaluation.

Images