Distributed photovoltaic output estimation cluster area intelligent division method

Abstract

The invention discloses a distributed photovoltaic output estimation cluster area intelligent division method, and relates to the technical field of distributed photovoltaic output estimation, and the method comprises the steps: building a distributed photovoltaic output estimation cluster division index based on the correlation degree of natural conditions; establishing a distributed photovoltaic output cluster division solving model based on the established cluster division index and a reinforcement learning algorithm, and determining a to-be-set parameter; setting parameters of the established solving model to obtain an optimized solving model; and carrying out distributed photovoltaic output estimation cluster region division strategy output by using the obtained optimization solution model. According to the method, the photovoltaic power stations with similar environmental conditions are reasonably grouped by establishing the cluster division indexes based on the natural condition correlation degree, so that data redundancy is reduced, the efficiency and accuracy of the estimation process are improved, meanwhile, parameter setting is carried out on the reinforcement learning model, and the estimation accuracy is improved. And the intelligence degree and adaptability of the cluster division model are further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed photovoltaic output estimation, and in particular to a distributed photovoltaic output estimation cluster area intelligent division method. Background Art

[0002] With the rapid development of renewable energy, especially the widespread application of distributed photovoltaic power generation, the scale of photovoltaic access to distribution networks continues to expand. The widespread distribution and dynamic changes of distributed photovoltaic systems make the estimation of photovoltaic output extremely complicated. However, accurate photovoltaic output estimation is of vital importance for the safe and stable operation of the power grid, power dispatch optimization, and system load prediction. Efficient and accurate output estimation methods can provide important decision-making basis for frequency regulation, load balancing, dispatch planning, etc. of the power grid, thereby promoting the development of smart grids and improving the adaptability of power grid systems to the volatility and intermittency of renewable energy.

[0003] At present, the output estimation of distributed photovoltaics mainly relies on the historical data and real-time meteorological information of a single photovoltaic site. However, due to the differences in the temporal and spatial distribution of photovoltaic power generation, a single estimation method is difficult to effectively adapt to the widespread and large number of distributed photovoltaic power stations. Traditional methods mostly use a one-by-one calculation method, which is time-consuming and easily reduces accuracy due to data redundancy and local volatility. In addition, with the increasing number of distributed photovoltaic systems, traditional estimation methods are difficult to cope with large-scale, high-frequency real-time output predictions, and lack sufficient intelligent support to handle the interrelationships and dependencies between data. Therefore, how to effectively group and estimate different photovoltaic systems has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to provide a distributed photovoltaic output estimation cluster area intelligent division method to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solution: a distributed photovoltaic output estimation cluster area intelligent division method, comprising the following steps:

[0006] S1. Establish distributed photovoltaic output estimation cluster division indicators based on natural condition correlation;

[0007] S2, based on the cluster division index established in S1 and the reinforcement learning algorithm, a distributed photovoltaic output cluster division solution model is established, and the parameters to be adjusted are determined;

[0008] S3, adjusting the parameters of the solution model established in S2 to obtain an optimized solution model;

[0009] S4. Use the optimization solution model obtained in S3 to output the cluster area division strategy for distributed photovoltaic output estimation.

[0010] According to the above technical solution, step S1 specifically includes the following steps:

[0011] S101, establishing a natural condition feature vector based on the natural condition correlation;

[0012] S102, calculating the similarity of natural conditions between any two photovoltaic sites;

[0013] S103: construct a similarity matrix based on the similarity between any two photovoltaic sites.

[0014] According to the above technical solution, the establishment of the natural condition feature vector is based on information data of multiple dimensions, including but not limited to meteorological factors, geographical factors, and solar irradiance;

[0015] Expressed as: X i =[x i1 ,x i2 ,...,x in ] T ;

[0016] Among them, X i represents the natural condition feature vector of the i-th distributed photovoltaic site; n represents the number of natural condition features for distributed photovoltaic output estimation, and T represents transposition;

[0017] The similarity of natural conditions between the photovoltaic sites is measured using the Euclidean distance d ij To express:

[0018]

[0019] Among them, i represents the i-th photovoltaic site, j represents the j-th photovoltaic site; k represents the number of natural condition characteristics is k;

[0020] Build a similarity matrix based on the calculated similarities:

[0021]

[0022] Among them, s ij represents the similarity between the i-th photovoltaic site and the j-th photovoltaic site; exp() represents the matrix function.

[0023] According to the above technical solution, step S2 specifically includes the following steps:

[0024] S201, establishing a Markov decision model for distributed photovoltaic output estimation cluster partition solution model;

[0025] S202, using the Actor-Critic framework to perform reinforcement learning training on the Markov decision model established in S201;

[0026] S203: Generate a target Q value using the target network.

[0027] According to the above technical solution, the Markov decision model includes a state (S), an action (A) and a reward function (R);

[0028] The reward function is:

[0029]

[0030] Where C represents the set of all clusters; C k and C l They represent the kth cluster and the lth cluster respectively, and λ represents the parameter that weighs the similarity within a cluster and the difference between clusters.

[0031] According to the above technical solution, in the Actor-Critic architecture, the Actor network is used to determine the action to be performed in a given state, and the Critic network is used to evaluate the value of the action selected by the Actor network;

[0032] The Actor network outputs the current state and the corresponding action;

[0033] The parameter update of the Actor network is based on the policy gradient, and the update formula is:

[0034]

[0035] Among them, θ μ Represents the parameters of the Actor network, Q(s,a|θ Q ) represents the Q value function of the Critic network, μ(s|θ μ ) represents the policy function of the Actor network.

[0036] According to the above technical solution, the Critic network inputs the current state and the action output by the Actor and evaluates the Q value of the action;

[0037] The loss function of the Critic network is:

[0038]

[0039] Among them, y i =r i +γQ′(s i+1 ,μ′(s i+1 )|θ Q ′), γ represents the discount factor, Q′ and μ′ represent the target Critic network and the target Actor network respectively.

[0040] According to the above technical solution, in order to avoid Q value oscillation during training, the target network is used to generate the target Q value. The parameters of the target network are updated from the main network at regular intervals. The update formula of the target network is:

[0041] θ′=τθ+(1-τ)θ′;

[0042] Among them, τ∈[0,1] is the soft update parameter, and θ,θ′ represent the parameters of the main network and the target network respectively.

[0043] Through repeated training, the model can learn the optimal strategy to achieve the best clustering effect of photovoltaic sites; this strategy can dynamically adjust the clustering according to the natural conditions of the photovoltaic sites and their changes.

[0044] According to the above technical solution, step S3 specifically includes the following steps:

[0045] S301, using the parameters to be adjusted obtained in S2 as the solution space vector to be optimized of the plant cell swarm algorithm;

[0046] S302, using the plant cell group growth fitness function to evaluate the quality of the current parameters to be adjusted;

[0047] S303, calculate the fitness of each cell, select the cell with the highest fitness to enter the next generation, and eliminate the cells with lower fitness.

[0048] According to the above technical solution, the calculation model of the fitness of each cell is:

[0049]

[0050] Among them, f i : The fitness value of the i-th solution, which is used to evaluate the quality of the current parameter vector;

[0051] α, β, γ, δ: weight factors used to adjust the influence of different items in fitness calculation;

[0052] θ i : The parameter vector of the i solutions represents the parameters to be set for the current distributed photovoltaic output estimation cluster area division model;

[0053] R(θ i ): Based on the strategy θ i The cumulative reward function evaluates the performance of the strategy.

[0054] H(θ i ): Policy entropy, quantifies the exploratory nature of a policy, with high entropy indicating that the policy is more inclined to explore new behaviors.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] The present invention establishes a cluster division index based on the correlation of natural conditions to reasonably group photovoltaic power stations with similar environmental conditions, thereby reducing data redundancy and improving the efficiency and accuracy of the estimation process;

[0057] By using the method of the present invention, photovoltaic sites with similar natural conditions are divided into clusters, which can effectively reduce the amount of calculation and improve the calculation efficiency. Further optimization of the preliminary clustering results based on the reinforcement learning model can ensure that the division results achieve a good balance between accuracy and calculation efficiency; the plant cell swarm algorithm is used in the present invention to adjust the parameters of the reinforcement learning model, and has the ability of global search; the parameters in the reinforcement learning model have a great influence on the performance of the model. The use of the plant cell swarm algorithm can effectively avoid falling into the local optimum, ensure that the parameter configuration reaches the global optimum, and show stronger robustness and generalization ability under different environmental conditions;

[0058] Therefore, the proposed new method has important theoretical and practical significance for promoting the efficient realization and intelligent development of distributed photovoltaic output estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 The photovoltaic output curves corresponding to different distributed photovoltaic output estimation clusters of the present invention;

[0060] Figure 2 It is a schematic diagram of the convergence process of the distributed photovoltaic output estimation cluster partitioning solution model based on the reinforcement learning algorithm of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The present invention provides a technical solution for a distributed photovoltaic output estimation cluster area intelligent division method, comprising the following steps:

[0063] S1. Establish distributed photovoltaic output estimation cluster division indicators based on natural condition correlation;

[0064] Specifically:

[0065] S101, establishing a natural condition feature vector based on the natural condition correlation;

[0066] Furthermore, the establishment of the natural condition feature vector is based on information data of multiple dimensions, including but not limited to meteorological factors, geographical factors, and solar irradiance;

[0067] Expressed as: X i =[x i1 ,x i2 ,...,x in ] T ;

[0068] Among them, X i represents the natural condition feature vector of the i-th distributed photovoltaic site; n represents the number of natural condition features for distributed photovoltaic output estimation, and T represents transposition;

[0069] S102, calculating the similarity of natural conditions between any two photovoltaic sites;

[0070] Furthermore, the similarity of natural conditions between the photovoltaic sites is calculated using the Euclidean distance d ij To express:

[0071]

[0072] Among them, i represents the i-th photovoltaic site, j represents the j-th photovoltaic site; k represents the number of natural condition characteristics is k;

[0073] S103, constructing a similarity matrix based on the similarity between any two photovoltaic sites;

[0074] Furthermore, a similarity matrix is ​​established based on the calculated similarities:

[0075]

[0076] Among them, s ij represents the similarity between the i-th photovoltaic site and the j-th photovoltaic site; exp() represents the matrix function.

[0077] S2, based on the cluster division index established in S1 and the reinforcement learning algorithm, a distributed photovoltaic output cluster division solution model is established, and the parameters to be adjusted are determined;

[0078] Specifically:

[0079] S201, establishing a Markov decision model for distributed photovoltaic output estimation cluster partition solution model;

[0080] Further, the Markov decision model includes a state (S), an action (A) and a reward function (R);

[0081] About state (S): Each state represents the current division of the PV site and its corresponding natural condition characteristics. In different states, the PV site belongs to different clusters;

[0082] Regarding actions (A): actions are defined as adjustments to the division of PV sites in the current state, such as moving a site from one cluster to another, or creating a new cluster;

[0083] About the reward function (R): The reward function is used to evaluate the quality of a certain partitioning action, with the goal of maximizing the similarity within the cluster and minimizing the difference between clusters;

[0084] The reward function is:

[0085]

[0086] Where C represents the set of all clusters; C k and C l They represent the kth cluster and the lth cluster respectively, and λ represents the parameter that weighs the similarity within a cluster and the difference between clusters.

[0087] S202, using the Actor-Critic framework to perform reinforcement learning training on the Markov decision model established in S201;

[0088] Furthermore, in the Actor-Critic architecture, the Actor network is used to decide the action to be performed in a given state, and the Critic network is used to evaluate the value of the action selected by the Actor network;

[0089] The Actor network outputs the current state and the corresponding action, i.e., the cluster division adjustment strategy of the PV site;

[0090] The parameter update of the Actor network is based on the policy gradient, and the update formula is:

[0091]

[0092] Among them, θ μ Represents the parameters of the Actor network, Q(s,a|θ Q ) represents the Q value function of the Critic network, μ(s|θ μ ) represents the strategy function of the Actor network;

[0093] The purpose of parameter update based on policy gradient is to find the parameters to update in the direction of maximum policy gradient, so that the optimal parameters can be found more quickly.

[0094] Furthermore, the Critic network inputs the current state and the action output by the Actor, and evaluates the Q value of the action, that is, the long-term benefit of performing the action in this state;

[0095] The loss function of the Critic network is:

[0096]

[0097] Among them, y i =r i +γQ′(s i+1 ,μ′(s i+1 )|θ Q′ ), γ represents the discount factor, Q′ and μ′ represent the target Critic network and the target Actor network respectively;

[0098] S203: Generate a target Q value using the target network.

[0099] Furthermore, in order to avoid Q value oscillation during training, the Q value of the next state calculated by the target network is used as the target Q value of the main network. The parameters of the target network are updated from the main network at regular intervals. The update formula of the target network is:

[0100] θ′=τθ+(1-τ)θ′;

[0101] Among them, τ∈[0,1] is the soft update parameter, and θ,θ′ represent the parameters of the main network and the target network respectively.

[0102] Through repeated training, the model can learn the optimal strategy to achieve the best clustering effect of photovoltaic sites; this strategy can dynamically adjust the clustering according to the natural conditions of the photovoltaic sites and their changes.

[0103] S3, adjusting the parameters of the solution model established in S2 to obtain an optimized solution model;

[0104] Specific:

[0105] S301, using the parameters to be adjusted obtained in S2 as the solution space vector to be optimized of the plant cell swarm algorithm, the parameters to be adjusted include but are not limited to learning rate, discount factor, etc.;

[0106] S302, using the plant cell group growth fitness function to evaluate the quality of the current parameters to be adjusted;

[0107] Specifically:

[0108]

[0109] Among them, T is the number of training rounds, R t is the cumulative reward of round t;

[0110] S303, calculate the fitness of each cell, select the cell with the highest fitness to enter the next generation, and eliminate the cells with lower fitness.

[0111] The parameter configuration of the distributed photovoltaic output estimation cluster partitioning solution model based on the reinforcement learning algorithm after tuning by the plant cell swarm algorithm can significantly improve the model training effect and cluster partitioning performance. The optimized model can converge to the optimal strategy faster and show stronger robustness and generalization ability under different environmental conditions.

[0112] Furthermore, the calculation model of the fitness of each cell is:

[0113]

[0114] Among them, f i : The fitness value of the i-th solution, which is used to evaluate the quality of the current parameter vector;

[0115] α, β, γ, δ: weight factors used to adjust the influence of different items in fitness calculation;

[0116] θ i : The parameter vector of the i solutions represents the parameters to be set for the current distributed photovoltaic output estimation cluster area division model;

[0117] R(θ i ): Based on the strategy θ i The cumulative reward function evaluates the performance of the strategy.

[0118] H(θ i ): Policy entropy, quantifies the exploratory nature of a policy, with high entropy indicating that the policy is more inclined to explore new behaviors.

[0119] S4. Use the optimization solution model obtained in S3 to output the cluster area division strategy for distributed photovoltaic output estimation.

[0120] In this embodiment, in order to verify the effectiveness of the present invention, we selected 100 distributed photovoltaic sites connected to a distribution network as research objects; these photovoltaic sites are distributed in different geographical areas and have different natural conditions, such as: meteorological factors, geographical locations, irradiance, etc.; the data used include meteorological data and geographical information of each photovoltaic site, meteorological data such as: temperature, humidity, wind speed, irradiance, etc.; geographical information such as latitude, longitude, altitude, etc.;

[0121] like Figure 1As shown, the photovoltaic output curves corresponding to different distributed photovoltaic output estimation clusters are shown, such as clusters C1, C2, and C3; it can be seen that there are obvious differences in the photovoltaic outputs of different distributed photovoltaic output estimation clusters, which means that the method can effectively distinguish the distributed photovoltaic outputs under different natural conditions, and gather the distributed photovoltaic outputs of the same natural condition type together for output estimation, thereby increasing the efficiency of distributed photovoltaic output;

[0122] like Figure 2 As shown in the figure, the convergence process of the distributed photovoltaic output estimation cluster partitioning solution model based on the reinforcement learning algorithm is demonstrated. It can be seen that due to the use of the plant cell swarm algorithm, the model can quickly converge to the optimal value within 100 iterations, thereby improving the overall optimization solution efficiency of the distributed photovoltaic output estimation cluster partitioning solution model.

[0123] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A distributed photovoltaic output estimation cluster area intelligent division method, characterized in that: The following steps are involved: S1. Establish distributed photovoltaic output estimation cluster division indicators based on natural condition correlation; S2, based on the cluster division index established in S1 and the reinforcement learning algorithm, a distributed photovoltaic output cluster division solution model is established, and the parameters to be adjusted are determined; S3, adjusting the parameters of the solution model established in S2 to obtain an optimized solution model; S4. Use the optimization solution model obtained in S3 to output the cluster area division strategy for distributed photovoltaic output estimation.

2. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 1, characterized in that: Step S1 specifically includes the following steps: S101, establishing a natural condition feature vector based on the natural condition correlation; S102, calculating the similarity of natural conditions between any two photovoltaic sites; S103: construct a similarity matrix based on the similarity between any two photovoltaic sites.

3. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 2, characterized in that: The establishment of the natural condition feature vector is based on information data of multiple dimensions; Expressed as: X i =[x i1 ,x i2 ,...,x in ] T ; Among them, X i represents the natural condition feature vector of the i-th distributed photovoltaic site; n represents the number of natural condition features for distributed photovoltaic output estimation, and T represents transposition; The similarity of natural conditions between the photovoltaic sites is measured using the Euclidean distance d ij To express: Among them, i represents the i-th photovoltaic site, j represents the j-th photovoltaic site; k represents the number of natural condition characteristics is k; Build a similarity matrix based on the calculated similarities: Among them, s ij represents the similarity between the i-th PV site and the j-th PV site; exp() represents the function of finding the exponential of the variables in the brackets; σ is the standard deviation of the characteristic values ​​of each PV site.

4. The method for intelligently dividing distributed photovoltaic output estimation cluster areas according to claim 1, characterized in that: Step S2 specifically includes the following steps: S201, establishing a Markov decision model for distributed photovoltaic output estimation cluster partition solution model; S202, using the Actor-Critic framework to perform reinforcement learning training on the Markov decision model established in S201; S203: Generate a target Q value using the target network.

5. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 4, characterized in that: The Markov decision model includes a state (S), an action (A) and a reward function (R); The reward function is: Where C represents the set of all clusters; C k and C l They represent the kth cluster and the lth cluster respectively, and λ represents the parameter that weighs the similarity within a cluster and the difference between clusters.

6. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 5, characterized in that: In the Actor-Critic architecture, the Actor network is used to decide the action to be performed in a given state, and the Critic network is used to evaluate the value of the action selected by the Actor network; The Actor network outputs the current action, that is, the current distributed photovoltaic output estimation cluster division strategy, including the number of divided clusters and the affiliation of each distributed photovoltaic cluster. The parameter update of the Actor network is based on the policy gradient, and the update formula is: Among them, θ μ Represents the parameters of the Actor network, Q(s,a|θ Q ) represents the Q value function of the Critic network, μ(s|θ μ ) represents the policy function of the Actor network.

7. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 6, characterized in that: The Critic network inputs the current state and the action output by the Actor and evaluates the Q value of the action; The loss function of the Critic network is: In the formula, y i =r i +γQ′(s i+1 ,μ′(s i+1 )|θ Q′ ), γ represents the discount factor, Q′ and μ′ represent the target Critic network and the target Actor network respectively. L(θ Q ) means that when the strategy is θ Q ; i represents the agent i used for iteration in the Critic network; N represents the total number of agents used for iteration in the Critic network; s represents the agent state used for iteration in the Critic network; a represents the agent action used for iteration in the Critic network, that is, the judgment of the cluster division strategy of the current distributed photovoltaic output estimation output by the Actor network; Q represents the current Critic network function.

8. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 7, characterized in that: The target network is used to generate the target Q value. The parameters of the target network are updated from the main network at regular intervals. The update formula of the target network is: θ′=τθ+(1-τ)θ′; Among them, τ∈[0,1] is the soft update parameter, and θ,θ′ represent the parameters of the main network and the target network respectively.

9. The method for intelligently dividing distributed photovoltaic output estimation cluster areas according to claim 1, characterized in that: Step S3 specifically includes the following steps: S301, using the parameters to be adjusted obtained in S2 as the solution space vector to be optimized of the plant cell swarm algorithm; S302, using the plant cell group growth fitness function to evaluate the quality of the current parameters to be adjusted; S303, calculate the fitness of each cell, select the cell with the highest fitness to enter the next generation, and eliminate the cells with lower fitness.

10. A distributed photovoltaic output estimation cluster area intelligent division method according to claim 9, characterized in that: The calculation model of the fitness of each cell is: Among them, f i : The fitness value of the i-th solution, which is used to evaluate the quality of the current parameter vector; α, β, γ, δ: weight factors used to adjust the influence of different items in fitness calculation; θ i : The parameter vector of the i solutions represents the parameters to be set for the current distributed photovoltaic output estimation cluster regional division model; R(θ i ): Based on the strategy θ i The cumulative reward function evaluates the performance of the strategy. H(θ i ): Policy entropy, quantifies the exploratory nature of a policy, with high entropy indicating that the policy is more inclined to explore new behaviors.

Images