CN117973644A - Distributed photovoltaic power virtual acquisition method considering optimization of reference power station - Google Patents

Abstract

The invention provides a distributed photovoltaic power virtual acquisition method considering reference power station optimization, which relates to the technical field of data processing and comprises the following steps: step 1: the gating circulation unit is used as a virtual collector to capture the time dependence of the distributed photovoltaic. A loss function and an attention mechanism based on Wasserstein distance measurement are introduced into the GRU to assist the virtual collector to learn complex mapping relations between different power stations; step 2: introducing a time-varying binary transfer function and a chaotic initialization strategy into a traditional badger optimization algorithm to search a reference power station set with highest virtual acquisition precision; step 3: and adopting model-free reinforcement learning depth DQN to adaptively and dynamically adjust super parameters of the virtual collector. The invention solves the problems of high acquisition cost and low acquisition reliability caused by mass distributed photovoltaic construction, and has important significance for better developing fine management and fine operation and maintenance of future distributed photovoltaic power stations.

Description

A distributed photovoltaic power virtual collection method considering reference power station optimization

Technical Field

The invention belongs to the field of electric power technology, relates to distributed photovoltaic operation data collection, and is a distributed photovoltaic power virtual collection method considering reference power station optimization.

Background Art

With the increasing shortage of energy and environmental pollution, photovoltaic systems have become one of the important means to solve energy and environmental problems. Compared with centralized photovoltaics, distributed photovoltaics have developed rapidly due to their advantages such as flexible installation and high energy utilization. However, the increase in the number of distributed photovoltaics requires a large number of sensors and communication equipment to monitor the operating status of distributed photovoltaics. Distributed photovoltaic data transmission in remote areas requires high communication costs, and the huge amount of data also requires extensive servers, databases and data monitoring platforms, resulting in many users being unwilling to use this service due to privacy or cost reasons. In addition, with the complex and changeable geographical environment and weather conditions at the installation location of distributed photovoltaics, the transmission process often has problems such as data loss, transmission congestion and equipment failure. Therefore, it is necessary to develop an economical and robust data collection method for distributed photovoltaic clusters.

Summary of the invention

The purpose of the present invention is to provide a method for virtual collection of distributed photovoltaic operation data taking into account reference power stations and hyperparameter optimization to solve the problem of high cost and low reliability of massive distributed photovoltaic operation and maintenance data collection. The main invention contents are as follows: Based on the fact that each distributed photovoltaic in the region has spatial correlation, a gated recurrent unit GRU is used as a virtual collector to capture the temporal correlation of distributed photovoltaics. The loss function and attention mechanism based on Wasserstein distance metric are introduced into GRU to mine the complex mapping relationship between different power stations. Secondly, time-varying binary transfer function and chaotic initialization strategy are introduced into the traditional honey badger optimization algorithm (HBA) to search for a set of reference power stations with the highest virtual collection accuracy. Finally, in order to improve the performance of the virtual collector under complex weather conditions, a model-free reinforcement learning deep Q network (DQN) is used to adaptively and dynamically adjust the hyperparameters of the virtual collector.

The technical solution adopted by the present invention is as follows:

A distributed photovoltaic power virtual collection method considering reference power station optimization includes the following steps:

Step S1: Design a virtual collector AL-GRU, use GRU to mine the spatiotemporal correlation characteristics between distributed photovoltaics, use the attention mechanism to dynamically adjust the weight of the reference power station at different times, and use a new loss function based on Wasserstein distance metric to improve the GRU fitting performance;

Step S2: Using the virtual collector AL-GRU as a data inference model, the reference power station in the region is optimized to collect the operating data of all distributed photovoltaic power stations in the region through the optimized reference power station, and the honey badger optimization algorithm HBA with a time-varying binary transfer function and a chaotic initialization strategy is introduced, and the optimization is performed with the highest K-fold cross-validation score as the objective function;

Step S3: Design state space, action space and reward function to guide the hyperparameter changes in the DQN learning history scenario in reinforcement learning, and adaptively adjust the hyperparameters according to the output change trend of distributed photovoltaics in the offline application stage.

Furthermore, the structure of the GRU includes an update gate and a reset gate, and the output of the reset gate Control the degree of integration between the network's current input and historical memory, and update the output of the gate Determines the proportion of the historical output power information of the reference power station to be retained. Assume that the input of GRU at time t is , combined with the reset gate and the update gate, the output at time t can be obtained , the specific calculation formula is as follows:

;

;

;

;

Where: , is the weight matrix corresponding to the reset gate; W _zy , W _zh are the weight matrices corresponding to the update gate; , Output the corresponding weight matrix for GRU; represents Hadamard operation; b represents bias vector; Represents the output of GRU at time t -1; To update the candidate value; Represents the sigmoid activation function.

Furthermore, the attention mechanism considers the historical output state to construct the input sliding window, thereby providing more reference information to the attention mechanism module. Assume that the input of the attention mechanism module at time t is ,in, is the number of reference power stations selected in the region, is _a time sliding window of length T1 In order to obtain the importance of each reference power station to the distributed photovoltaic power generation to be collected at this moment, A three-layer attention mechanism neural network ANN is constructed as input, and the output of ANN is normalized through the Softmax layer. The specific calculation process is as follows:

;

;

;

Where: and They are the outputs of the input layer and hidden layer of the attention mechanism module respectively; and are the activation functions of the input layer and the hidden layer respectively; and They are the weight matrix and bias term of the neural network respectively; is the weight coefficient of the i- th reference power station;

The obtained reference power station attention distribution matrix Multiply it by the current input to get the input after evaluating the importance of different reference power stations to the distributed photovoltaic power to be collected. , the virtual collector will fit the output of the power station to be collected according to the input, and the calculation process is as follows:

;

;

The daily variation curve of distributed photovoltaics to be collected is divided into K modes through the K-shape clustering algorithm. The calculation method is as follows:

;

;

;

Where: SBD is the shape-based distance; , Represents two distributed photovoltaic power time series for comparison; since the sampling time is 15 minutes, the length of the time series T = 96; Represents the cross-correlation sequence The length of ; k is the relative sliding distance between the two sequences.

Furthermore, the sample weights of the K -type curves are obtained, and the probability distribution is used to measure the difference between the K-type curves and the overall fluctuation situation. The Wasserstein distance is used to measure the probability distribution of the obtained K-type curves and the overall probability distribution. Assuming that the distributions of the two curve categories are P and Q respectively, the specific calculation formula is as follows:

;

Where: represents the set of all joint distributions of P and Q ; x and y represent the distributions from the joint distribution The samples obtained by sampling; Represents the distance between samples; in all possible joint distributions, the lower bound of the expected value of the distance between samples is the Wasserstein distance;

The weight of each type of curve is obtained through the Softmax function. During the offline training process, different weights are assigned to samples belonging to different types of curves. The specific calculation formula of the loss function of the Wasserstein distance metric guiding the virtual collector is as follows:

;

;

Where: Represents the Wasserstein distance between the i- th type curve and the overall probability distribution; Represents the number of samples of the i- th type of curve in the training set; represents the sample weight of the i -th type of curve; is the virtual acquisition result of the kth sample belonging to the i -th type of curve; is the real power.

Furthermore, the reference power station selection process in step S2 includes:

Step S201: Define the reference power plant selection problem, assuming that The set of binary state variables of distributed photovoltaics is , when the element is 1, it means that the distributed photovoltaic is selected as the reference power station, and when the element is 0, it means that the distributed photovoltaic is selected as the distributed photovoltaic to be collected;

Step S202: Introduce the K-fold cross validation score into the objective function and optimize it under the premise of satisfying the constraints. The specific description is as follows:

;

;

;

Where: Indicates the number of samples in the validation set; Indicates the number of training set samples; and They represent the ith virtual collection value and true value of the mth distributed photovoltaic to be collected; K _V is the fold of cross-validation; N _cpv represents the number of distributed photovoltaics to be collected,

The calculation formula for the number of reference power stations during the optimization process is:

;

The constraint on the number of reference power stations is:

;

Where: is the maximum number of reference power stations;

Step S203: Improve HBA by introducing a time-varying binary transfer function to adapt HBA to binary optimization tasks, introduce a chaotic initialization strategy to improve the quality of the HBA initial solution set, as described in detail below:

The time-varying binary transfer function calculation formula is:

;

Where: Control parameter , t is the current iteration number, and are the upper and lower limits of the control parameters, respectively; x represents the position of the HBA individual; e is the natural base;

The position of the HBA individual is transformed through the time-varying binary transfer function as follows:

;

Where: represents the dth dimension of the mth search agent of HBA; rand represents a random number between 0 and 1,

The chaos initialization strategy uses tent mapping to generate the initial population sequence, which is expressed as follows:

;

Step S204: Based on the objective function and constraints set in step S202, the model developed in step S1 is used as the virtual collector, and the optimization strategy proposed in step S203 is used to optimize the best reference power station set. Subsequently, the virtual collector is trained with the selected reference power station as input and other distributed photovoltaics to be collected as output to realize the collection of operating data of the entire distributed photovoltaic system in the area.

Furthermore, in the design state space in step S3, the variables affecting the control action decision are set as the state quantity of the system, and the state space is set

,in, represents the current period, W represents the weather, DR and DP represent the first-order differences of irradiance and output power in _the T2 time step, R and P represent the irradiance at the current moment and the output power of all reference power stations, and H represents the hyperparameter set of the virtual collector.

Furthermore, in the action space design in step S3, the intelligent body in DQN will adjust the hyperparameters of the virtual collector according to the current state space, and the combination of hyperparameter changes It constitutes the action space of the intelligent agent, where Indicates The change of hyperparameters, , The action space size of the hyperparameter is ; g represents the granularity vector that discretizes the continuous space.

Furthermore, the reward function in step S3 encourages the agent to take actions to improve the accuracy of virtual acquisition, and is defined as follows:

;

Where: and They represent the mean square error of virtual acquisition in the current state and after the action of the agent within 2 T ₂ steps respectively.

Furthermore, the adaptive adjustment of hyper parameters in step S3 includes the following steps:

Step S301: In order to improve the performance of virtual acquisition under complex weather conditions, a DQN-based virtual acquisition robustness enhancement strategy is proposed. The dynamic adjustment of hyperparameters is converted into the action selection of the agent through the DQN algorithm. The attenuated ε-greedy strategy is adopted to balance the random search and greedy behavior of the agent training process. The action is selected in this strategy. The probability of is 1- ε , the probability of random action is ε , and the action selection process is expressed as:

;

;

Where: Represents the action corresponding to the highest value function at time t ; and are the lower and upper limits of ε respectively; Represents the maximum number of iterations of reinforcement learning training. As the network training matures, the probability of random actions will gradually decrease. e is a natural base number;

Step S302: Introduce the experience replay mechanism to improve the efficiency of the interaction between the agent and the environment and reduce the correlation and dependence between samples. In this mechanism, the experience sample data obtained by the interaction between the agent and the environment at each time step is stored in the experience pool. In the subsequent training process, DQN will update the weights of the target network and the current value network based on the following loss function:

;

Where: is the output of the target value network; is the output of the current value network; Represents the number of DQN training samples; and Represent the parameters of the current value network and the target value network respectively; Represents the return discount factor; max represents the maximum value function.

Beneficial Effects

The present invention has the following beneficial effects:

1. The present invention proposes the concept of "virtual collection" to solve the data collection challenges such as insufficient data transmission equipment and reduced data transmission reliability brought about by the construction of massive distributed photovoltaics.

2. Aiming at the virtual collection problem, a virtual collector for spatiotemporal feature extraction is developed. A loss function based on Wasserstein distance metric is proposed to enhance the fitting ability of GRU under different weather fluctuation scenarios. In addition, an attention mechanism is integrated to provide different attention levels to the reference power station at different times.

3. In order to solve the problem of reference power station selection, a time-varying binary transfer function is introduced into the traditional honey badger optimization algorithm to adapt to the optimization of binary decision variables, and the quality of the initial solution of the honey badger optimization algorithm is improved through the chaos initialization strategy.

4. The present invention combines deep reinforcement learning with virtual collection technology, and dynamically adjusts the hyperparameters of the virtual collector according to changes in the distributed photovoltaic output trend, thereby enhancing the robustness of the virtual collection technology.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a distributed photovoltaic operation data virtual collection method considering reference power stations and hyperparameter optimization provided by the present invention.

FIG. 2 is a schematic diagram of a virtual collector involved in the present invention.

FIG3 is a flow chart of the reference power station selection method involved in the present invention.

FIG4 is a schematic diagram of the K-fold cross validation involved in the present invention.

FIG5 is a box plot of virtual acquisition errors of the loss function involved in the present invention and the traditional loss functions MAE and MSE.

Figure 6(a)-(d) are comparison diagrams of virtual acquisition results with and without adding the attention mechanism on Typical Day 1 to Typical Day 4, respectively.

FIG7 is a comparison diagram of the errors of virtual collection of different power stations using different virtual collectors.

Figure 8 shows the changes in virtual acquisition accuracy at different iteration times.

DETAILED DESCRIPTION

The present invention will be further described in detail below in conjunction with the accompanying drawings:

As shown in FIG1 , the present invention provides a distributed photovoltaic power virtual collection method considering reference power station optimization, the method comprising the following steps:

Step S1: Design the virtual collector AL-GRU. The gated recurrent unit (GRU) is used as the virtual collector to mine the spatiotemporal correlation characteristics between distributed photovoltaics. The attention mechanism is used to dynamically adjust the weights of the reference power station at different times, and the new loss function based on the Wasserstein distance metric is used to improve the GRU fitting performance.

Step S2: Using the virtual collector developed above as the inference model, the reference power station in the region is optimized to install sensors at the selected sites to collect the photovoltaic operation data of the entire system. In order to improve the quality of the reference power station, a honey badger optimization algorithm with a time-varying binary transfer function and a chaotic initialization strategy is proposed, and the optimization is performed with the highest K-fold cross-validation score as the objective function.

Step S3: To improve the performance of the model under different weather conditions, the state space, action space and reward function are designed to guide the hyperparameter changes in the DQN learning history scene in reinforcement learning. In the offline application stage, the hyperparameters can be adaptively adjusted according to the output change trend of distributed photovoltaics to improve the accuracy of virtual collection.

The explanation of the virtual collector in step S1 is: the data reasoning of the distributed photovoltaic virtual collection is essentially to use the collection data of the reference power station as input, and to estimate the operating data of all distributed photovoltaics in the region in real time through the regression model. Therefore, the virtual collector is a regression model developed for virtual collection tasks.

The explanation of the reference power station optimization in step S2 is:

The purpose of the present invention is to select reference power stations based on the improved honey badger optimization algorithm, and their real-time power data will be input into the intelligent computing model as a multi-dimensional feature to estimate the output of all distributed photovoltaics in the region. However, randomly selecting reference power stations in the region is unreliable. Selecting a reasonable combination of reference power stations in the power station area can complement each other, thereby effectively improving the accuracy of virtual acquisition.

Furthermore, the virtual collector AL-GRU in step S1 includes the loss function of the basic GRU, the attention mechanism module and the Wasserstein distance metric:

Step S101: The structure of GRU includes an update gate and a reset gate. The output of the reset gate Controls the degree of integration between the network's current input and the memory of historical moments. Update gate output It determines the proportion of the historical output power information of the reference power station to be retained. Assume that the input of the GRU unit at time t is , combined with the reset gate and the update gate, the output at time t can be obtained The specific calculation formula is as follows:

;

;

;

;

Where: , is the weight matrix corresponding to the reset gate; W _zy , W _zh are the weight matrices corresponding to the update gate; , Output the corresponding weight matrix for the GRU unit; represents Hadamard operation; b represents bias vector; Represents the output of the GRU unit at time t -1; To update the candidate value; Represents the sigmoid activation function.

Step S102: Introduce the attention mechanism to improve the fitting ability of GRU. The attention mechanism considers the historical output state to construct the input sliding window, thereby providing more reference information to the attention mechanism module. Assume that the input of the attention mechanism module at time t is .in, is the number of reference power stations selected in the region, is _a time sliding window of length T1 In order to obtain the importance of each reference power station to the current photovoltaic power to be collected at this moment, A three-layer attention mechanism neural network (ANN) is constructed as input, and the output of the ANN is normalized through the Softmax layer. The specific calculation process is as follows:

;

;

;

Where: and They are the outputs of the input layer and hidden layer of the attention mechanism module respectively; and are the activation functions of the input layer and the hidden layer respectively; and They are the weight matrix and bias term of the neural network respectively; is the weight coefficient of the i- th reference power station;

Step S103: Obtain the reference power station attention distribution matrix Multiply it by the current input to get the input after evaluating the importance of different reference power stations to the current photovoltaic power to be collected. The virtual collector will fit the output of the distributed photovoltaic to be collected based on the input. The calculation process is as follows:

;

;

The reasons for introducing the attention mechanism in step S102 include: each power station in the region has a unique geographical spatial location, and the importance of each reference power station to the distributed photovoltaics to be collected is not the same at different times. This motivates us to use the attention mechanism to explore the degree of spatiotemporal correlation between different photovoltaics. The attention mechanism imitates the resource allocation mechanism of human brain attention, and at a specific moment it will focus on the area that needs to be paid attention to. Therefore, the attention mechanism is introduced in the virtual collector. Combined with Figure 2, the schematic diagram of the virtual collector involved in the present invention can give different degrees of attention to different reference power stations, thereby realizing the dynamic allocation of reference power station weights.

Step S104: Divide the daily variation curve of the photovoltaic power to be collected into K modes by using the K-shape clustering algorithm. The calculation method is as follows:

;

;

;

Where: SBD is the shape-based distance; , Represents the input time series of k-shape; since the sampling time is 15 minutes, the length of the time series T = 96; Represents the cross-correlation sequence The length of ; k is the relative sliding distance between the two sequences.

Step S105: In order to obtain the sample weight of the K-type curve, the probability distribution is used to measure the difference between the K-type curve and the overall fluctuation. The advantage of this is that it can prevent a certain abnormal curve from affecting the overall weight. The Wasserstein distance does not require the length of the two sets of data to be the same, and can effectively measure their similarity even when the two probability distributions do not overlap. Therefore, the Wasserstein distance is used to measure the probability distribution of the obtained K-type curve and the overall probability distribution. The specific calculation formula is as follows:

;

Where: represents the set of all joint distributions of P and Q ; x and y represent the distributions from the joint distribution The samples obtained by sampling; Represents the distance between samples; in all possible joint distributions, the lower bound of the expected value of the distance between samples is the Wasserstein distance.

Step S105: Obtain the weight of each type of curve through the Softmax function. During the offline training process, different weights are assigned to samples belonging to different types of curves. The specific calculation formula of the loss function of the Wasserstein distance metric guiding the virtual collector is as follows:

;

;

Where: Represents the Wasserstein distance between the i- th type curve and the overall probability distribution; Represents the number of samples of the i- th type of curve in the training set; represents the sample weight of the i -th type of curve; is the virtual acquisition result of the kth sample belonging to the i -th type of curve; is the real power.

Further, in conjunction with FIG. 3 , referring to the flowchart of the reference power station selection method involved in the present invention, the reference power station selection process in step S2 includes:

Step S201: Define the reference power plant selection problem. Assume that The set of binary state variables of distributed photovoltaics is When the element is 1, it means that the distributed photovoltaic power station is selected as the reference power station, and when the element is 0, it means that the distributed photovoltaic power station is selected as the distributed photovoltaic power station to be collected;

Step S202: The optimization process needs to consider the performance of the virtual collector, and the training loss cannot reflect the model's fitting ability under unknown conditions. Therefore, in order to make full use of the existing data and reflect the comprehensive virtual collection error of different distributed photovoltaics as reference power stations during the optimization process, the K-fold cross-validation score is introduced into the objective function, and the optimization is performed under the premise of satisfying the constraints. The effect of the validation set is shown in Figure 4, which is a schematic diagram of the K-fold cross-validation involved in the present invention, and is described in detail as follows:

;

;

;

Where: Indicates the number of samples in the validation set; Indicates the number of training set samples; and They represent the ith virtual collection value and true value of the mth distributed photovoltaic to be collected respectively; K _V is the fold of cross-validation; Indicates the number of photovoltaics to be collected.

The optimization algorithm tends to increase the number of reference power stations to improve the accuracy of virtual acquisition. If sparsity is not required, the purpose of reducing acquisition costs in virtual acquisition cannot be achieved. The calculation formula for the number of reference power stations is:

;

The constraint on the number of reference power stations is:

;

Where: is the maximum number of reference power stations.

Step S203: Improve HBA by introducing a time-varying binary transfer function to adapt HBA to binary optimization tasks. Compared with other binary transfer functions, the time-varying transfer function can improve the algorithm's exploration ability at different search stages and avoid falling into local optimality. In addition, a chaotic initialization strategy is introduced to improve the quality of the HBA initial solution set. The specific description is as follows:

The time-varying binary transfer function calculation formula is:

;

Where: x represents the position of the HBA individual; e is the natural base; control parameter

;

t is the current iteration number, and are the upper and lower limits of the control parameters respectively.

The position of the HBA individual is transformed through the binary transfer function as follows:

;

Where: represents the dth dimension of the mth search agent of the HBA; rand represents a random number between 0 and 1.

The chaos initialization strategy is described as follows:

The traditional sparrow optimization algorithm initializes the population through random numbers. This random generation method may cause uneven distribution of generated individuals, resulting in reduced population diversity and optimization speed. Therefore, tent mapping is used to generate the initial population sequence, which is expressed as follows:

;

Step S204: Based on the objective function and constraints set in step S202, the model developed in step S1 is used as a virtual collector, and the optimization strategy proposed in step S203 is used to optimize and obtain the best reference power station set. Subsequently, the virtual collector is trained with the selected reference power station as input and other distributed photovoltaics to be collected as output, so as to realize the virtual collection of the operating data of the entire distributed photovoltaic system in the region.

Furthermore, the dynamic adjustment of the virtual collector hyperparameters in step S3 includes the following steps:

Step S301: In order to improve the performance of virtual acquisition under complex weather conditions, a DQN-based virtual acquisition robustness enhancement strategy is proposed. This method is based on the state information and acquisition errors in historical scenes, and improves the robustness of virtual acquisition by dynamically adjusting the model hyperparameters, thereby reducing the virtual acquisition error in specific scenes. The dynamic adjustment of hyperparameters is converted into the action selection of the agent through the DQN algorithm, and the attenuated ε-greedy strategy is adopted to balance the random search and greedy behavior of the agent training process. Actions are selected in this strategy. The probability of is 1- ε , and the probability of random action is ε .

;

;

Where: Represents the action corresponding to the highest value function at time t ; and are the lower and upper limits of ε respectively; Represents the maximum number of iterations of reinforcement learning training. As the network training matures, the probability of random actions will gradually decrease.

The intelligent agent in step S301 is explained as follows: the intelligent agent in reinforcement learning is composed of a neural network, and is an entity that performs hyperparameter adjustment actions during data collection by a virtual collector to complete training by minimizing virtual collection errors.

Step S302: Introduce the experience replay mechanism to improve the efficiency of the interaction between the agent and the environment and reduce the correlation and dependence between samples. In this mechanism, the experience sample data obtained by the interaction between the agent and the environment at each time step is is stored in the experience pool. In the subsequent training process, DQN will update the weights of the target network and the current value network based on the following loss function.

;

Where: is the output of the target value network; Represents the number of DQN training samples; and Represent the parameters of the current value network and the target value network respectively.

Step S303: For the virtual collection task, the description of the agent state space, action space and reward function is as follows:

State Space Design:

The variables that affect the control action decision are set as the state of the system. Our goal is to improve the performance of virtual acquisition in complex weather scenarios. The state space of the agent not only needs to reflect the output trend of the DPV, but also needs to include the hyperparameter state. Therefore, the state space is set .in, represents the current period, W represents the weather, DR and DP represent the first-order differences of irradiance and output power in _the T2 time step, R and P represent the irradiance at the current moment and the output power of all reference power stations, and H represents the hyperparameter set of the virtual collector.

Action Space Design:

The DQN agent will adjust the hyperparameters of the virtual collector based on the current state. Therefore, the combination of hyperparameter changes It constitutes the action space of the intelligent agent. Indicates The change of hyperparameters, , The action space size of the hyperparameter is ; g represents the granularity vector that discretizes the continuous space.

Reward function design:

The reward function motivates the agent to take actions that improve the accuracy of virtual acquisition and is defined as follows:

;

Where: and They represent the mean square error of virtual acquisition in the current state and after the action of the agent within 2 T ₂ steps respectively.

Best practice for specific applications:

To realize the virtual data collection of distributed photovoltaic power, it is first necessary to find a suitable distributed photovoltaic power data set and preprocess the collected data set to provide data support for subsequent case analysis. The present invention selected 29 distributed photovoltaics (DPVs) in Jiangning District, Nanjing City, Jiangsu Province, China for virtual collection testing. The data collected from February 1, 2018 to December 31, 2018 were collected in 15-minute intervals from 07:15 to 17:00. In order to facilitate the training and verification of the network, we use the data from February 1 to October 31 as the training set and verification set, and the remaining data is used to verify the effectiveness of DPV virtual collection. The accuracy of virtual collection is evaluated by the error indicators mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE):

;

;

;

Where: Indicates the number of test samples; and They represent the virtual collection value and real value of distributed photovoltaic output power respectively. When evaluating the virtual collection performance by MAPE, it is necessary to remove the sampling points where the actual power is 0.

In order to prove the effectiveness of the IHBA developed by the present invention, the honey badger optimization algorithm (HBA), coyote algorithm (COA), gray wolf algorithm (GWO) and Harris Hawk optimization algorithm (HHO) are selected to optimize the reference power station. The objective function values of each optimization algorithm under different numbers of reference power stations are shown in Table 1. It can be seen that the reference power station obtained by the IHBA optimization algorithm used in the present invention has the best objective function value.

Table 1 Objective function values of each optimization algorithm under different reference power station numbers (W ² )

In order to verify the effect of the virtual collector, the total number of reference power stations is selected as 14, and the reference power stations are numbered as follows: 2, 4, 7, 10, 11, 12, 14, 15, 16, 18, 20, 21, 24, 27 through the IHBA algorithm for optimization. First, the superiority of the loss function proposed in the present invention is demonstrated, and the accuracy of virtual acquisition under the classic loss function MAE, MSE and the proposed loss function is compared. Without setting the random number seed, the experiment is repeated 20 times to obtain the error shown in Figure 5, the virtual acquisition error box plot of the loss function involved in the invention and the traditional loss function MAE, MSE. It can be seen that the virtual acquisition error of the proposed loss function is the lowest and has a lower standard deviation. In addition, compared with MAE, MSE can better reflect the sampling points with larger deviations. The loss function proposed in the present invention can further reflect the virtual acquisition error under different fluctuating weather conditions on the basis of the MSE loss function, and is more competitive in the virtual acquisition problem than other loss functions. For the convenience of representation, we call the GRU with attention mechanism A-GRU and the GRU with improved loss function L-GRU in the following text.

We selected four typical days with different output trends in the test set, and used Power Station No. 1 as a representative to verify the effect of adding the attention mechanism. The virtual collection results are shown in Figure 6, which is a comparison chart of the virtual collection results with and without the attention mechanism on four typical days. Overall, the three lines are relatively close, indicating that the reference power station selected by the present invention can complete virtual collection with higher accuracy. In addition, the A-GRU with the added attention mechanism performs better than the GRU. This is because the attention mechanism dynamically assigns weights to the reference power station according to the data of the sliding window during this period of time, which increases the weight of the reference power station that is closer to the distributed photovoltaic output trend to be collected. In short, the proposed method has good virtual collection performance under both stable and fluctuating weather conditions.

The virtual acquisition accuracy under different improvement schemes is shown in Table 2. L-GRU has a more obvious effect on reducing MAE and RMSE errors. This shows that the loss function developed by the present invention can improve the fitting performance when the actual output is large, while MAPE is often affected by sampling points with smaller actual output. The MAE and RMSE reduction effects of A-GRU are not as good as those of L-GRU, but the MAPE reduction effect is better. In addition, the attention mechanism and GRU share the loss function. The proposed loss function can assist the attention mechanism in giving each power station a more discriminative weight under different weather conditions. Therefore, the combination of the two has the best virtual acquisition accuracy.

Table 2 Virtual acquisition error under different improvement schemes

Furthermore, in order to reflect the superiority of the proposed virtual collector, the performance of different virtual collectors is compared, as shown in the error comparison diagram of virtual collection of different power stations using different virtual collectors in Figure 7. From the perspective of the model, the MAE and RMSE of the BP neural network are both the highest, and it is difficult to complete the virtual collection task with high accuracy. Ensemble learning has good fitting performance. Among them, the RMSE of XGBoost is higher than that of LightGBM, but the MAPE is lower than that of LightGBM. This is because most of the abnormal points of XGBoost appear at sampling points with large actual power, while LightGBM is the opposite. In the future, it can be considered to fuse the two models to improve the collection accuracy. The AL-GRU proposed in the present invention has the best virtual collection accuracy. In addition, the GRU with time series memory function is also better than the BP neural network and the RF in the ensemble learning. Therefore, the time series analysis function is very important for the task of virtual collection, and the subsequent research should focus on this feature. From the perspective of the power station, the virtual collection accuracy of Power Station No. 1 and Power Station No. 13 is lower than that of other power stations. If it is used as a reference power station, it may affect the virtual acquisition accuracy of all power stations. Therefore, IHBA regards it as a non-reference power station, which indirectly reflects the rationality of IHBA's selection of a reference power station.

The output trend of distributed photovoltaics is not always as standardized as the β distribution, and the trend changes significantly under different weather conditions. In order to verify the effectiveness of the DQN dynamic hyperparameter adjustment method proposed in the present invention, we conducted simulation verification based on 6 virtual collectors. In order to clearly show the changes in model performance under different iteration numbers, we selected AL-GRU as a typical example for display. As shown in Figure 8, the changes in virtual acquisition accuracy under different iteration numbers, in order to make the display clearer, only the changes in RMSE are shown. It can be seen that the performance of the AL-GRU model has seriously declined due to insufficient DQN experience in the early stage. This shows that the hyperparameter setting of the model seriously affects the performance of virtual acquisition. With the increase in the number of iterations, DQN effectively improves the ability to dynamically adjust hyperparameters by learning more interactive information. When the maximum number of iterations exceeds 1000, the virtual acquisition performance of ALGRU with DQN-assisted hyperparameter adjustment tends to be stable. Therefore, we selected Max_iter=1200 for the experiment, repeated the experiment 15 times, and observed the virtual acquisition errors of different VCMs assisted by DQN.

The virtual acquisition accuracy of each model is shown in Table 3. It can be seen that the virtual acquisition accuracy of each VCM has been improved with the assistance of DQN. Since LightGBM and XGBoost in ensemble learning have more hyperparameters and have a larger hyperparameter adjustment space, the accuracy improvement is more obvious. In addition, the standard deviation of the ensemble learning model is larger because more iterations are required to make the model more stable.

Table 3 Virtual collection errors of different virtual collectors after adaptively adjusting hyperparameters

Furthermore, we observe the virtual acquisition effect of DQN dynamically adjusting hyperparameters under different training set ratios. The average virtual acquisition error when the training set ratio increases from 40% to 80% is shown in Table 4. It can be seen that the parameter improvement effect will improve with the increase of training sets, which shows that reinforcement learning requires more historical state scenarios for learning. Therefore, the subsequent improvement of this technology should focus on the collection of diversified output scenarios.

Table 4 Virtual acquisition error at different training set ratios

An electronic device includes a memory, wherein the processors are communicatively connected to each other, the memory stores computer instructions, and the processor executes any of the above-mentioned methods for virtual collection of distributed photovoltaic operation data taking into account reference power stations and hyperparameter optimization by executing the computer instructions.

A computer storage medium having computer instructions stored thereon, wherein the computer readable storage medium is used to enable a computer to execute any of the above-mentioned methods for virtual collection of distributed photovoltaic operation data taking into account reference power stations and hyperparameter optimization.

It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The solutions in the embodiments of the present invention may be implemented in various computer languages, for example, object-oriented programming language Java and interpreted scripting language JavaScript, etc.

The present invention is described with reference to the flowcharts and/or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and/or box in the flowchart and/or block diagram, as well as the combination of the processes and/or boxes in the flowchart and/or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

The applicant has made a detailed illustration and description of the embodiments of the present invention in conjunction with the drawings in the specification. However, this is for the purpose of helping readers better understand the spirit of the present invention rather than limiting the scope of protection of the present invention. On the contrary, any improvement or modification based on the spirit of the present invention should fall within the scope of protection of the present invention.

Claims (9)

1. A distributed photovoltaic power virtual collection method considering reference power station optimization, characterized in that it includes the following steps: Step S1: Design a virtual collector AL-GRU, use GRU to mine the spatiotemporal correlation characteristics between distributed photovoltaics, use the attention mechanism to dynamically adjust the weight of the reference power station at different times, and use a new loss function based on Wasserstein distance metric to improve the GRU fitting performance; Step S2: Using the virtual collector AL-GRU as a data inference model, the reference power station in the region is optimized to collect the operating data of all distributed photovoltaic power stations in the region through the optimized reference power station, and the honey badger optimization algorithm HBA with a time-varying binary transfer function and a chaotic initialization strategy is introduced, and the optimization is performed with the highest K-fold cross-validation score as the objective function; Step S3: Design state space, action space and reward function to guide the hyperparameter changes in the DQN learning history scenario in reinforcement learning, and adaptively adjust the hyperparameters according to the output change trend of distributed photovoltaics in the offline application stage. 2. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1 is characterized in that: the structure of the GRU includes an update gate and a reset gate, and the output of the reset gate Control the degree of integration between the network's current input and historical memory, and update the output of the gate Determines the proportion of the historical output power information of the reference power station to be retained. Assume that the input of GRU at time t is , combined with the reset gate and the update gate, the output at time t can be obtained , the specific calculation formula is as follows: ; ; ; ; Where: , is the weight matrix corresponding to the reset gate; W _zy , W _zh are the weight matrices corresponding to the update gate; , Output the corresponding weight matrix for GRU; represents Hadamard operation; b represents bias vector; Represents the output of GRU at time t -1; To update the candidate value; Represents the sigmoid activation function. 3. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1 is characterized in that: the attention mechanism considers the historical output state to construct an input sliding window, thereby providing more reference information to the attention mechanism module. Assume that the input of the attention mechanism module at time t is ,in, is the number of reference power stations selected in the region, is _a time sliding window of length T1 In order to obtain the importance of each reference power station to the distributed photovoltaic power generation to be collected at this moment, A three-layer attention mechanism neural network ANN is constructed as input, and the output of ANN is normalized through the Softmax layer. The specific calculation process is as follows: ; ; ; Where: and They are the outputs of the input layer and hidden layer of the attention mechanism module respectively; and are the activation functions of the input layer and the hidden layer respectively; and They are the weight matrix and bias term of the neural network respectively; is the weight coefficient of the i- th reference power station; The obtained reference power station attention distribution matrix Multiply it by the current input to get the input after evaluating the importance of different reference power stations to the distributed photovoltaic power to be collected. , the virtual collector will fit the output of the power station to be collected according to the input, and the calculation process is as follows: ; ; The daily variation curve of distributed photovoltaics to be collected is divided into K modes through the K-shape clustering algorithm. The calculation method is as follows: ; ; ; Where: SBD is the shape-based distance; , Represents two distributed photovoltaic power time series for comparison; since the sampling time is 15 minutes, the length of the time series T = 96; Represents the cross-correlation sequence The length of ; k is the relative sliding distance between the two sequences. 4. The distributed photovoltaic power virtual acquisition method considering reference power station optimization according to claim 3 is characterized by: obtaining the sample weight of the K -type curve, using probability distribution to measure the difference between the K-type curve and the overall fluctuation situation, using Wasserstein distance to measure the probability distribution of the obtained K-type curve and the overall probability distribution, assuming that the distributions of the two curve categories are P and Q respectively, the specific calculation formula is as follows: ; Where: represents the set of all joint distributions of P and Q ; x and y represent the distributions from the joint distribution The samples obtained by sampling; Represents the distance between samples; in all possible joint distributions, the lower bound of the expected value of the distance between samples is the Wasserstein distance; The weight of each type of curve is obtained through the Softmax function. During the offline training process, different weights are assigned to samples belonging to different types of curves. The specific calculation formula of the loss function of the Wasserstein distance metric guiding the virtual collector is as follows: ; ; Where: Represents the Wasserstein distance between the i- th type curve and the overall probability distribution; Represents the number of samples of the i- th type of curve in the training set; represents the sample weight of the i -th type of curve; is the virtual acquisition result of the kth sample belonging to the i -th type of curve; is the real power. 5. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1, characterized in that: the reference power station selection process in step S2 includes: Step S201: Define the reference power plant selection problem, assuming that The set of binary state variables of distributed photovoltaics is , when the element is 1, it means that the distributed photovoltaic is selected as the reference power station, and when the element is 0, it means that the distributed photovoltaic is selected as the distributed photovoltaic to be collected; Step S202: Introduce the K-fold cross validation score into the objective function and optimize it under the premise of satisfying the constraints. The specific description is as follows: ; ; ; Where: Indicates the number of samples in the validation set; Indicates the number of training set samples; and They represent the ith virtual collection value and true value of the mth distributed photovoltaic to be collected; K _V is the fold of cross-validation; N _cpv represents the number of distributed photovoltaics to be collected, The calculation formula for the number of reference power stations during the optimization process is: ; The constraint on the number of reference power stations is: ; Where: is the maximum number of reference power stations; Step S203: Improve HBA by introducing a time-varying binary transfer function to adapt HBA to binary optimization tasks, introduce a chaotic initialization strategy to improve the quality of the HBA initial solution set, as described in detail below: The time-varying binary transfer function calculation formula is: ; Where: Control parameter , t is the current iteration number, and are the upper and lower limits of the control parameters, respectively; x represents the position of the HBA individual; e is the natural base; The position of the HBA individual is transformed through the time-varying binary transfer function as follows: ; Where: represents the dth dimension of the mth search agent of HBA; rand represents a random number from 0 to 1. The chaos initialization strategy uses tent mapping to generate the initial population sequence, which is expressed as follows: ; Step S204: Based on the objective function and constraints set in step S202, the model developed in step S1 is used as the virtual collector, and the optimization strategy proposed in step S203 is used to optimize the optimal reference power station set. Subsequently, the virtual collector is trained with the selected reference power station as input and other distributed photovoltaics to be collected as output to realize the collection of operating data of the entire distributed photovoltaic system in the area. 6. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1 is characterized in that: in the design state space in step S3, the variables affecting the control action decision are set as the state quantity of the system, and the state space is set ,in, represents the current period, W represents the weather, DR and DP represent the first-order differences of irradiance and output power in _the T2 time step, R and P represent the irradiance at the current moment and the output power of all reference power stations, and H represents the hyperparameter set of the virtual collector. 7. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1 is characterized in that: in the action space design in step S3, the intelligent experience in the DQN will adjust the hyperparameters of the virtual collector according to the current state space, and the combination of hyperparameter changes It constitutes the action space of the intelligent agent, where Indicates The change of hyperparameters, , The action space size of the hyperparameter is ; g represents the granularity vector that discretizes the continuous space. 8. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1 is characterized in that: the reward function in step S3, which encourages the agent to take actions to improve the virtual collection accuracy, is defined as follows: ; Where: and They represent the mean square error of virtual acquisition in the current state and after the action of the agent within 2 T ₂ steps respectively. 9. The distributed photovoltaic power virtual collection method considering reference power station optimization according to claim 1, characterized in that: the adaptive adjustment of hyperparameters in step S3 comprises the following steps: Step S301: In order to improve the performance of virtual acquisition under complex weather conditions, a DQN-based virtual acquisition robustness enhancement strategy is proposed. The dynamic adjustment of hyperparameters is converted into the action selection of the agent through the DQN algorithm. The attenuated ε-greedy strategy is adopted to balance the random search and greedy behavior of the agent training process. The action is selected in this strategy. The probability of is 1- ε , the probability of random action is ε , and the action selection process is expressed as: ; ; Where: Represents the action corresponding to the highest value function at time t ; and are the lower and upper limits of ε respectively; Represents the maximum number of iterations of reinforcement learning training. As the network training matures, the probability of random actions will gradually decrease. e is a natural base number; Step S302: Introduce the experience replay mechanism to improve the efficiency of the interaction between the agent and the environment and reduce the correlation and dependence between samples. In this mechanism, the experience sample data obtained by the interaction between the agent and the environment at each time step is stored in the experience pool. In the subsequent training process, DQN will update the weights of the target network and the current value network based on the following loss function: ; Where: is the output of the target value network; is the output of the current value network; Represents the number of DQN training samples; and Represent the parameters of the current value network and the target value network respectively; Represents the return discount factor; max represents the maximum value function.