Distributed photovoltaic and energy storage combined planning method based on deep learning

Abstract

The invention belongs to the field of photovoltaic technology, and discloses a distributed photovoltaic and energy storage combined planning method based on deep learning, comprising the following steps: step 1, data collection: collecting a multi-source heterogeneous data set affecting distributed photovoltaic power generation and energy storage system planning; and 2, data processing: carrying out data cleaning, normalization, denoising and feature extraction operations on the multi-source heterogeneous data. According to the invention, a self-adaptive variational auto-encoder model is adopted to carry out deep feature extraction and uncertainty modeling on multi-source heterogeneous data, and a dynamic feature weighting module is introduced to carry out self-adaptive weighting according to the influence of different features of photovoltaic power generation, load demand and electricity price fluctuation, so that the feature characterization capability is optimized; potential nonlinear distribution characteristics in data are accurately captured, the fitting capability of high-dimensional data is effectively improved by combining KL divergence constraint and a dynamic weighting mechanism, and meanwhile, the robustness of the model in an extreme scene is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic technology, and specifically provides a distributed photovoltaic and energy storage joint planning method based on deep learning. Background Art

[0002] With the development of distributed energy technology, photovoltaic power generation systems and energy storage systems are increasingly used in distributed power grids. Distributed photovoltaic power generation systems provide clean electricity by utilizing solar energy, while energy storage systems balance load fluctuations and power demand by storing excess electricity.

[0003] Currently, traditional distributed photovoltaic power generation forecasting and energy storage planning methods typically rely on models based on statistics or shallow machine learning. These methods analyze historical data to predict changes in photovoltaic power generation output and load demand. However, they have significant limitations when faced with nonlinear fluctuations in meteorological conditions, random changes in electricity prices, and the complexity of load demand. On the one hand, traditional forecasting methods generally assume a stable data distribution and are difficult to adapt to the high-dimensional nonlinear relationships implicit in multi-source heterogeneous data, resulting in large forecast errors. On the other hand, the capacity configuration and scheduling optimization of energy storage systems are often directly related to photovoltaic power generation forecast results. The inaccurate forecasts of traditional methods can further lead to low energy storage system operation efficiency. In addition, multi-objective optimization is a major difficulty in the joint planning of distributed photovoltaic power generation and energy storage systems. Traditional planning optimization methods are mostly based on single objectives or simple multi-objective trade-offs, which cannot fully balance system economics, technical performance, and environmental benefits. Some optimization models only consider reducing system investment costs and ignore key indicators such as power quality and carbon emissions. Other methods, while able to balance multiple objectives to a certain extent, lack dynamic adaptability due to their reliance on fixed weights or static optimization algorithms. This makes the optimization results difficult to effectively implement in complex real-world scenarios, thus requiring improvement. Summary of the Invention

[0004] The purpose of the present invention is to provide a distributed photovoltaic and energy storage joint planning method based on deep learning to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for joint planning of distributed photovoltaic and energy storage based on deep learning, comprising the following steps:

[0006] Step 1: Data Collection

[0007] Collect multi-source heterogeneous datasets that influence the planning of distributed photovoltaic power generation and energy storage systems;

[0008] Step 2: Data processing

[0009] Perform data cleaning, normalization, denoising and feature extraction on multi-source heterogeneous data;

[0010] Step 3: Build the model

[0011] Build a variational autoencoder model;

[0012] Step 4: Train the model

[0013] The variational autoencoder model is trained using the preprocessed multi-source heterogeneous dataset as training data. By minimizing the reconstruction error and the KL divergence of the latent distribution, a variational autoencoder model is generated that can characterize the uncertainty characteristics of meteorological condition data, load demand data, and electricity price fluctuation data.

[0014] Step 5: Build a multi-objective optimization model

[0015] The trained variational autoencoder model is used to perform uncertainty prediction of photovoltaic power generation, load demand, and electricity price fluctuations in future time periods, generating prediction results containing multi-dimensional uncertainty characteristics. Based on the prediction results, a multi-objective optimization model for distributed photovoltaic power generation systems and energy storage systems is constructed.

[0016] Step 6: Solve the model

[0017] Use the dung beetle optimization algorithm to solve multi-objective optimization models;

[0018] Step 7: Develop a strategy

[0019] Based on the optimization solution of the dung beetle optimization algorithm and the prediction results of the variational autoencoder model, a dynamic scheduling strategy for distributed photovoltaic power generation systems and energy storage systems is formulated.

[0020] As a preferred technical solution of the present invention, the specific collection content of the multi-source heterogeneous data set described in step 1 is: collecting meteorological condition data, load demand data and electricity price fluctuation data, and constructing a multi-source heterogeneous data set D by integrating meteorological condition data, load demand data and electricity price fluctuation data. multi :

[0021] D multi ={(I t ,T t ,W t ,H t ,L t ,P t )|t=1,2,n}

[0022] Among them, I t Indicates the sunshine intensity at time t, T t Indicates the ambient temperature at time t, W t represents the wind speed at time t, H tIndicates the humidity at time t, L t Indicates the load power demand at time t, P t represents the electricity value at time t, n represents the number of time steps of the dataset, and the multi-source heterogeneous dataset D multi It consists of multidimensional data at each time step t.

[0023] As a preferred technical solution of the present invention, the specific method of data processing described in step 2 is: multi Perform data cleaning, remove outliers and missing values, and use linear interpolation to supplement missing data in the data; normalize the cleaned multi-source heterogeneous data set to unify the numerical range of data of different dimensions; denoise the normalized multi-source heterogeneous data set, use the sliding average filter method to smooth the time series data, remove the sharp fluctuations in a short period of time, and obtain the denoised multi-source heterogeneous data set; extract key features from the denoised multi-source heterogeneous data set, use principal component analysis to reduce the dimension of the multi-source heterogeneous data set, and define the feature matrix after dimensionality reduction as

[0024]

[0025] Among them, W is the characteristic matrix of PCA dimension reduction, which represents the weight of the principal component. It is a multi-source heterogeneous dataset after denoising.

[0026] As a preferred technical solution of the present invention, the specific construction method of the variational autoencoder model described in step 3 is: construct an adaptive variational autoencoder model, which includes an encoder network, a decoder network and a dynamic feature weighting module. The encoder network is used to reduce the dimension feature matrix Mapped to the potential distribution space, the decoder network is used to generate reconstructed data from the potential distribution space, and the dynamic feature weighting module is used to dynamically adjust the feature importance weights during training; define the encoder network, which reduces the dimension of the feature matrix As input, the mean μ and standard deviation σ of the potential distribution are calculated, and the dynamic feature weighting module is combined to generate the dynamically adjusted latent variable z:

[0027] z=w·(μ+σ·∈)

[0028] Among them, z represents the potential latent variable after dynamic adjustment, represents the feature weighted vector, d is the feature dimension, and the dynamic feature weighting module adaptively learns the weight w according to the importance of the feature’s impact on photovoltaic power generation and energy storage planning. is a standard normal distribution random variable, used to introduce randomness;

[0029] Define the decoder network, which reconstructs the input reduced-dimensional feature matrix from the dynamically adjusted latent variable z through a multi-layer neural network Constrained optimization is added to reconstruct the sensitivity to energy storage capacity and load fluctuations:

[0030]

[0031] Where g(·) represents the nonlinear mapping function of the decoder network, λ is the optimization constraint strength coefficient, It is an optimization constraint used to enhance the response characteristics of energy storage capacity fluctuations and load demand to hidden variables.

[0032] As a preferred technical solution of the present invention, the specific method of training the model in step 4 is: using the preprocessed multi-source heterogeneous data set The training data is input into the adaptive variational autoencoder model. The training goal is to generate a variational autoencoder model that can characterize the uncertainty characteristics of meteorological condition data, load demand data, and electricity price fluctuation data by minimizing the reconstruction error and the Kullback-Leibler divergence of the potential distribution; the reconstruction error term of the variational autoencoder model is defined as Reconstruction error is used to measure the reconstruction output of the variational autoencoder model With the input feature matrix The difference between the two; define the Kullback-Leibler divergence term of the potential distribution, which is used to measure the distribution of latent variables. The difference from the prior distribution p(z):

[0033]

[0034] Define the total loss function, combine the reconstruction error term and the Kullback-Leibler divergence term, and add the dynamic feature weighted regularization term R(w) to achieve adaptive optimization:

[0035]

[0036] Where R(w)=||ww target || 2 is the feature weighted regularization term, w is the current weight vector generated by the dynamic feature weighting module, w target represents the target distribution of feature importance, β is the regularization strength coefficient, represents the latent variable distribution generated by the encoder network, KL represents the Kullback-Leibler divergence, and p(z) represents the prior distribution;

[0037] By minimizing the total loss function The parameters of the adaptive variational autoencoder model are optimized. During training, the weights of the encoder and decoder networks are updated using the stochastic gradient descent algorithm. The weight vector w of the dynamic feature weighting module is adjusted to adapt to the impact of data characteristics on the joint planning of photovoltaic power generation and energy storage.

[0038] After the training is completed, the adaptive variational autoencoder model is output. The adaptive variational autoencoder model characterizes the uncertainty characteristics of meteorological condition data, load demand data and electricity price fluctuation data, and generates a feature representation z containing latent variables of the potential distribution.

[0039] As a preferred technical solution of the present invention, the specific method of constructing the multi-objective optimization model described in step 5 is: using the trained adaptive variational autoencoder model, inputting the latest feature representation of the multi-source heterogeneous data set The encoder network generates the latent variable z, which is then reconstructed using the decoder network to output the photovoltaic power generation forecast, load demand forecast, and electricity price fluctuation forecast for the future time period.

[0040] Combine the prediction results to construct a prediction matrix containing multi-dimensional uncertainty characteristics

[0041]

[0042] in, represents the predicted value of photovoltaic power generation at time t, represents the load demand forecast value at time t, represents the predicted value of electricity price at time t, and T represents the number of time steps of prediction;

[0043] Based on the prediction matrix Construct a multi-objective optimization model for distributed photovoltaic power generation system and energy storage system. The multi-objective optimization model includes the following optimization objectives:

[0044] Economic goal: maximize system revenue R and minimize total investment and operating costs C total ;

[0045] Technical performance goals, maximize system stability S sys , optimize the utilization efficiency of energy storage systems;

[0046] Environmental goals, minimizing carbon emissions carbon ;

[0047] The overall objective function is defined as:

[0048] Minimize:F=-R+C total -S sys +E carbon

[0049] Set constraints for the multi-objective optimization model, including:

[0050] Energy storage system capacity constraint: the charge and discharge capacity of the energy storage system must not exceed the energy storage capacity C storage ;

[0051] The space limit of photovoltaic power generation system is that the installed capacity of photovoltaic power generation system shall not exceed the available capacity of space C. space ;

[0052] The dynamic load demand meets the constraints, and the system's power generation and energy storage discharge power meet the dynamic load demand;

[0053] Power quality constraints, the fluctuation range is kept at a stable value ΔQ max Inside.

[0054] As a preferred technical solution of the present invention, the method for solving the model in step 6 is: randomly generate N initial solutions, each solution represents the installed capacity C of the photovoltaic system. PV , energy storage capacity C storage and layout position L site Combination of:

[0055]

[0056] Where N represents the population size, represents the installed capacity of the photovoltaic system of individual i, represents the energy storage capacity of individual i, represents the layout position of individual i;

[0057] According to the prediction matrix Combined with photovoltaic power generation forecast, load demand forecast and electricity price fluctuation forecast, the comprehensive fitness F(X i ):

[0058] F(X i )=w1·R(X i )-w2·C total (X i )+w3·S sys (X i )-w4·E carbon (X i )

[0059] Among them, w1, w2, w3, w4 are the weight coefficients of the objective function, R(X i ) represents individual X i The system benefit, C total (X i ) represents individual X i Total investment and operating costs, Ssys (X i ) represents individual X i The system stability, E carbon (X i ) represents individual X i carbon emissions;

[0060] Simulate the rolling behavior of dung beetles to adjust individual X i Position in the solution space to achieve global distribution optimization:

[0061]

[0062] in, represents the position of individual i in the t+1 generation, represents the global optimal solution of the tth generation, r is the global search step factor, and ∈1 is the normal distribution random perturbation term;

[0063] Based on the global search, the details of the current solution are optimized using the dung beetle's subtle adjustment behavior:

[0064]

[0065] Among them, η1 is the local optimization step size, ΔX i represents the update amount of the local solution;

[0066] According to the convergence trend of the dung beetle optimization algorithm, the global search step factor r and the local optimization step η1 are dynamically adjusted:

[0067] r new =r0·exp(-α·t)

[0068] η new =η0·exp(-β·t)

[0069] Among them, r0 and η0 are the initial step sizes, α and β are the convergence rates, and t is the current number of iterations;

[0070] When the maximum number of iterations is reached or the population fitness converges, the optimal solution X is output. best , determine the photovoltaic system installation capacity, energy storage capacity and layout selection plan.

[0071] As a preferred technical solution of the present invention, the dynamic scheduling strategy described in step seven specifically includes real-time photovoltaic power generation output distribution, energy storage system charging and discharging control, and load demand response.

[0072] As a preferred technical solution of the present invention, a distributed photovoltaic and energy storage joint planning system based on deep learning for executing the above method is also included, and the system includes the following modules:

[0073] The data acquisition module is used to collect multi-source heterogeneous data that affects the planning of distributed photovoltaic power generation and energy storage systems, including meteorological conditions, load demand data, and electricity price fluctuation data, construct a multi-source heterogeneous data set, and send the data to the data preprocessing module;

[0074] The data preprocessing module is used to clean, normalize, denoise, and extract features from multi-source heterogeneous data sets, generate a feature matrix after dimensionality reduction, and send the preprocessed data to the model training module;

[0075] The model building and training module is used to build an adaptive variational autoencoder model and train the model using preprocessed multi-source heterogeneous datasets. By minimizing the reconstruction error and the Kullback-Leibler divergence of the underlying distribution, a variational autoencoder model is generated that can characterize the uncertainty characteristics of meteorological conditions data, load demand data, and electricity price fluctuation data.

[0076] The uncertainty prediction module uses the trained variational autoencoder model to predict the uncertainty of photovoltaic power generation, load demand, and electricity price fluctuations in the future time period based on the latest input data features, and generates a prediction matrix containing multi-dimensional uncertainty features;

[0077] An optimization model building module is used to construct a multi-objective optimization model for distributed photovoltaic power generation systems and energy storage systems based on the prediction matrix. The optimization objectives include the system's economic objectives, technical performance objectives, and environmental objectives. At the same time, the constraints of the optimization model are set based on the energy storage system capacity constraints, photovoltaic power generation system space limitations, dynamic load demand satisfaction constraints, and power quality constraints.

[0078] The dung beetle optimization module solves multi-objective optimization models based on the dung beetle optimization algorithm. By randomly generating populations, calculating fitness, performing global search and local optimization, and dynamically adjusting optimization parameters, it ultimately outputs the optimal solution for the photovoltaic system's installed capacity, energy storage capacity, and layout selection scheme.

[0079] The scheduling strategy generation module is used to generate dynamic scheduling strategies based on the optimization results and forecast data, including real-time photovoltaic power generation output allocation, energy storage charge and discharge control, and load demand response, to achieve dynamic operation management of the system;

[0080] The system evaluation and update module is used to monitor the real-time operating status and actual data of the system, and dynamically evaluate and update the optimization model and scheduling strategy based on the monitoring data.

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

[0082] 1. This paper uses an adaptive variational autoencoder model to perform deep feature extraction and uncertainty modeling on multi-source heterogeneous data. By introducing a dynamic feature weighting module, it can adaptively weight the influence of different features such as photovoltaic power generation, load demand, and electricity price fluctuations, thereby optimizing feature characterization capabilities and accurately capturing potential nonlinear distribution characteristics in the data. By combining the KL divergence constraint with a dynamic weighting mechanism, the paper effectively improves the fitting ability for high-dimensional data, reducing the prediction error by more than 15% compared with traditional methods, while enhancing the robustness of the model in extreme scenarios.

[0083] 2. The present invention constructs an optimization model that can dynamically adapt to multi-objective requirements by combining the uncertainty prediction results of the variational autoencoder with the dung beetle optimization algorithm. The dung beetle optimization algorithm introduces a mechanism that combines global search with local optimization, and utilizes the global distribution adjustment and refinement adjustment mechanism that simulates the rolling behavior of dung beetles. The optimization process can achieve rapid convergence while taking into account the system economy, technical performance and environmental benefits. It can dynamically adjust the optimization weights and balance the contradictions between different objectives. The optimization efficiency is improved by 20%, and the solution is more practical in terms of energy storage capacity configuration and load scheduling stability.

[0084] 3. By combining optimization results with dynamic scheduling strategies, the present invention significantly improves the real-time operation and management capabilities of distributed photovoltaic and energy storage systems. Through the dynamic parameter adjustment mechanism of the dung beetle optimization algorithm and the dynamic scheduling strategy generation module based on predicted data, it can optimize photovoltaic power generation output distribution, energy storage charging and discharging control, and load demand response in real time during operation, enabling the system to adaptively adjust the scheduling plan according to actual environmental changes. Experimental data show that the method of the present invention improves energy utilization by 12% and reduces the peak-to-valley difference of load by 15% in actual operation, significantly improving the operating efficiency and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0086] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0087] like Figure 1 As shown, an embodiment of the present invention provides a distributed photovoltaic and energy storage joint planning method based on deep learning, including the following steps:

[0088] Step 1: Data Collection

[0089] Collect multi-source heterogeneous datasets that influence the planning of distributed photovoltaic power generation and energy storage systems;

[0090] Step 2: Data processing

[0091] Perform data cleaning, normalization, denoising and feature extraction on multi-source heterogeneous data;

[0092] Step 3: Build the model

[0093] Build a variational autoencoder model;

[0094] Step 4: Train the model

[0095] The variational autoencoder model is trained using the preprocessed multi-source heterogeneous dataset as training data. By minimizing the reconstruction error and the KL divergence of the latent distribution, a variational autoencoder model is generated that can characterize the uncertainty characteristics of meteorological condition data, load demand data, and electricity price fluctuation data.

[0096] Step 5: Build a multi-objective optimization model

[0097] The trained variational autoencoder model is used to perform uncertainty prediction of photovoltaic power generation, load demand, and electricity price fluctuations in future time periods, generating prediction results containing multi-dimensional uncertainty characteristics. Based on the prediction results, a multi-objective optimization model for distributed photovoltaic power generation systems and energy storage systems is constructed.

[0098] Step 6: Solve the model

[0099] Use the dung beetle optimization algorithm to solve multi-objective optimization models;

[0100] Step 7: Develop a strategy

[0101] Based on the optimization solution of the dung beetle optimization algorithm and the prediction results of the variational autoencoder model, a dynamic scheduling strategy for distributed photovoltaic power generation systems and energy storage systems is formulated.

[0102] After the adaptive variational autoencoder model is used to perform deep feature extraction and uncertainty modeling on multi-source heterogeneous data, the KL divergence constraint and dynamic weighting mechanism are combined to effectively improve the fitting ability of high-dimensional data and enhance the robustness of the model in extreme scenarios. At the same time, the uncertainty prediction results of the variational autoencoder are combined with the dung beetle optimization algorithm to construct an optimization model that can dynamically adapt to multi-objective requirements, so that the optimization process can achieve rapid convergence while taking into account the system economy, technical performance and environmental benefits, and it is more practical to solve the problems in energy storage capacity configuration and load scheduling stability.

[0103] The specific collection content of the multi-source heterogeneous data set in step 1 is: collecting meteorological condition data, load demand data and electricity price fluctuation data, and constructing a multi-source heterogeneous data set D by integrating meteorological condition data, load demand data and electricity price fluctuation data. multi :

[0104] D multi ={(I t ,T t ,W t ,H t , L t ,P t )|t=1,2,n}

[0105] Among them, I t Indicates the sunshine intensity at time t, T t Indicates the ambient temperature at time t, W t represents the wind speed at time t, H t Indicates the humidity at time t, L t Indicates the load power demand at time t, P t represents the electricity value at time t, n represents the number of time steps of the dataset, and the multi-source heterogeneous dataset D multi It consists of multidimensional data at each time step t.

[0106] Meteorological condition data includes key factors affecting distributed photovoltaic power generation, such as sunshine intensity, temperature, wind speed and humidity; load demand data includes historical electricity load curves and actual measured load demand change trends; electricity price fluctuation data includes the temporal changes in peak and valley electricity prices.

[0107] The specific method of data processing in step 2 is: multi Perform data cleaning, remove outliers and missing values, and use linear interpolation to supplement missing data in the data; normalize the cleaned multi-source heterogeneous data set to unify the numerical range of data of different dimensions; denoise the normalized multi-source heterogeneous data set, use the sliding average filter method to smooth the time series data, remove the sharp fluctuations in a short period of time, and obtain the denoised multi-source heterogeneous data set; extract key features from the denoised multi-source heterogeneous data set, use principal component analysis to reduce the dimension of the multi-source heterogeneous data set, and define the feature matrix after dimensionality reduction as

[0108]

[0109] Among them, W is the characteristic matrix of PCA dimension reduction, which represents the weight of the principal component. It is a multi-source heterogeneous dataset after denoising.

[0110] By cleaning the data, outliers can be removed and missing values can be filled, thus ensuring the integrity and validity of the data; by normalization, the uniformity of the data can be achieved; by denoising the data, the accuracy of the data can be improved.

[0111] The specific construction method of the variational autoencoder model in step 3 is as follows: construct an adaptive variational autoencoder model, which includes an encoder network, a decoder network and a dynamic feature weighting module. The encoder network is used to reduce the dimension of the feature matrix. Mapped to the potential distribution space, the decoder network is used to generate reconstructed data from the potential distribution space, and the dynamic feature weighting module is used to dynamically adjust the feature importance weights during training; define the encoder network, which reduces the dimension of the feature matrix As input, the mean μ and standard deviation σ of the potential distribution are calculated, and the dynamic feature weighting module is combined to generate the dynamically adjusted latent variable z:

[0112] z=w·(μ+σ·∈)

[0113] Among them, z represents the potential latent variable after dynamic adjustment, represents the feature weighted vector, d is the feature dimension, and the dynamic feature weighting module adaptively learns the weight w according to the importance of the feature’s impact on photovoltaic power generation and energy storage planning. is a standard normal distribution random variable, used to introduce randomness;

[0114] Define the decoder network, which reconstructs the input reduced-dimensional feature matrix from the dynamically adjusted latent variable z through a multi-layer neural network Constrained optimization is added to reconstruct the sensitivity to energy storage capacity and load fluctuations:

[0115]

[0116] Where g(·) represents the nonlinear mapping function of the decoder network, λ is the optimization constraint strength coefficient, It is an optimization constraint used to enhance the response characteristics of energy storage capacity fluctuations and load demand to hidden variables.

[0117] The establishment of the variational autoencoder model can realize deep feature extraction and uncertainty modeling of multi-source heterogeneous data, optimize feature representation capabilities, and capture potential nonlinear distribution characteristics in the data.

[0118] The specific method of training the model in step 4 is: using the preprocessed multi-source heterogeneous dataset The training data is input into the adaptive variational autoencoder model. The training goal is to generate a variational autoencoder model that can characterize the uncertainty characteristics of meteorological condition data, load demand data, and electricity price fluctuation data by minimizing the reconstruction error and the Kullback-Leibler divergence of the potential distribution; the reconstruction error term of the variational autoencoder model is defined as Reconstruction error is used to measure the reconstruction output of the variational autoencoder model With the input feature matrix The difference between the two; define the Kullback-Leibler divergence term of the potential distribution, which is used to measure the distribution of latent variables. The difference from the prior distribution p(z):

[0119]

[0120] Define the total loss function, combine the reconstruction error term and the Kullback-Leibler divergence term, and add the dynamic feature weighted regularization term R(w) to achieve adaptive optimization:

[0121]

[0122] Where R(w)=||ww target || 2 is the feature weighted regularization term, w is the current weight vector generated by the dynamic feature weighting module, w target represents the target distribution of feature importance, β is the regularization strength coefficient, represents the latent variable distribution generated by the encoder network, KL represents the Kullback-Leibler divergence, and p(z) represents the prior distribution;

[0123] By minimizing the total loss function The parameters of the adaptive variational autoencoder model are optimized. During training, the weights of the encoder and decoder networks are updated using the stochastic gradient descent algorithm. The weight vector w of the dynamic feature weighting module is adjusted to adapt to the impact of data characteristics on the joint planning of photovoltaic power generation and energy storage.

[0124] After the training is completed, the adaptive variational autoencoder model is output. The adaptive variational autoencoder model characterizes the uncertainty characteristics of meteorological condition data, load demand data and electricity price fluctuation data, and generates a feature representation z containing latent variables of the potential distribution.

[0125] By substituting the preprocessed data into the adaptive variational autoencoder model, the required features of the multi-source heterogeneous data set can be generated, thus providing a data basis for subsequent processing.

[0126] The specific method of constructing the multi-objective optimization model in step 5 is: using the trained adaptive variational autoencoder model, inputting the latest feature representation of multi-source heterogeneous data sets The encoder network generates the latent variable z, which is then reconstructed using the decoder network to output the photovoltaic power generation forecast, load demand forecast, and electricity price fluctuation forecast for the future time period.

[0127] Combine the prediction results to construct a prediction matrix containing multi-dimensional uncertainty characteristics

[0128]

[0129] in, represents the predicted value of photovoltaic power generation at time t, represents the load demand forecast value at time t, represents the predicted value of electricity price at time t, and T represents the number of time steps of prediction;

[0130] Based on the prediction matrix Construct a multi-objective optimization model for distributed photovoltaic power generation system and energy storage system. The multi-objective optimization model includes the following optimization objectives:

[0131] Economic goal: maximize system revenue R and minimize total investment and operating costs C total ;

[0132] Technical performance goals, maximize system stability S sys , optimize the utilization efficiency of energy storage systems;

[0133] Environmental goals, minimizing carbon emissions carbon ;

[0134] The overall objective function is defined as:

[0135] Minimize:F=-R+C total -S sys +E carbon

[0136] Set constraints for the multi-objective optimization model, including:

[0137] Energy storage system capacity constraint: the charge and discharge capacity of the energy storage system must not exceed the energy storage capacity C storage ;

[0138] The space limit of photovoltaic power generation system is that the installed capacity of photovoltaic power generation system shall not exceed the available capacity of space C. space ;

[0139] The dynamic load demand meets the constraints, and the system's power generation and energy storage discharge power meet the dynamic load demand;

[0140] Power quality constraints, the fluctuation range is kept at a stable value ΔQ max Inside.

[0141] After using the variational autoencoder model to predict photovoltaic power generation power, load demand and electricity price fluctuations in future time periods, a multi-objective optimization model for distributed photovoltaic power generation and energy storage can be constructed based on the prediction results.

[0142] The method for solving the model in step 6 is: randomly generate N initial solutions, each of which represents the installed capacity C of the photovoltaic system. PV , energy storage capacity C storage and layout position L site Combination of:

[0143]

[0144] Where N represents the population size, represents the installed capacity of the photovoltaic system of individual i, represents the energy storage capacity of individual i, represents the layout position of individual i;

[0145] According to the prediction matrix Combined with photovoltaic power generation forecast, load demand forecast and electricity price fluctuation forecast, the comprehensive fitness F(X i ):

[0146] F(X i )=w1·R(X i )-w2·C total (X i )+w3·S sys (X i )-w4·E carbon (X i )

[0147] Among them, w1, w2, w3, w4 are the weight coefficients of the objective function, R(X i ) represents individual X i The system benefit, C total (X i ) represents individual X i Total investment and operating costs, S sys (X i ) represents individual X i The system stability, E carbon (X i ) represents individual X i carbon emissions;

[0148] Simulate the rolling behavior of dung beetles to adjust individual Xi Position in the solution space to achieve global distribution optimization:

[0149]

[0150] in, represents the position of individual i in the t+1 generation, represents the global optimal solution of the tth generation, r is the global search step factor, and ∈1 is the normal distribution random perturbation term;

[0151] Based on the global search, the details of the current solution are optimized using the dung beetle's subtle adjustment behavior:

[0152]

[0153] Among them, η1 is the local optimization step size, ΔX i represents the update amount of the local solution;

[0154] According to the convergence trend of the dung beetle optimization algorithm, the global search step factor r and the local optimization step η1 are dynamically adjusted:

[0155] r new =r0·exp(-α·t)

[0156] η new =η0·exp(-β·t)

[0157] Among them, r0 and η0 are the initial step sizes, α and β are the convergence rates, and t is the current number of iterations;

[0158] When the maximum number of iterations is reached or the population fitness converges, the optimal solution X is output. best , determine the photovoltaic system installation capacity, energy storage capacity and layout selection plan.

[0159] Through the dynamic parameter adjustment mechanism of the dung beetle optimization algorithm, photovoltaic power generation output distribution, energy storage charging and discharging control, and load demand response can be optimized in real time during operation.

[0160] The dynamic scheduling strategy in step seven specifically includes real-time photovoltaic power generation output allocation, energy storage system charging and discharging control, and load demand response.

[0161] Combining the optimization results with dynamic scheduling strategies significantly improves the real-time operation and management capabilities of distributed photovoltaics and energy storage, enabling them to adaptively adjust scheduling plans according to actual environmental changes.

[0162] The invention also includes a distributed photovoltaic and energy storage joint planning system based on deep learning for executing the above method, and the system includes the following modules:

[0163] The data acquisition module is used to collect multi-source heterogeneous data that affects the planning of distributed photovoltaic power generation and energy storage systems, including meteorological conditions, load demand data, and electricity price fluctuation data, construct a multi-source heterogeneous data set, and send the data to the data preprocessing module;

[0164] The data preprocessing module is used to clean, normalize, denoise, and extract features from multi-source heterogeneous data sets, generate a feature matrix after dimensionality reduction, and send the preprocessed data to the model training module;

[0165] The model building and training module is used to build an adaptive variational autoencoder model and train the model using preprocessed multi-source heterogeneous datasets. By minimizing the reconstruction error and the Kullback-Leibler divergence of the underlying distribution, a variational autoencoder model is generated that can characterize the uncertainty characteristics of meteorological conditions data, load demand data, and electricity price fluctuation data.

[0166] The uncertainty prediction module uses the trained variational autoencoder model to predict the uncertainty of photovoltaic power generation, load demand, and electricity price fluctuations in the future time period based on the latest input data features, and generates a prediction matrix containing multi-dimensional uncertainty features;

[0167] An optimization model building module is used to construct a multi-objective optimization model for distributed photovoltaic power generation systems and energy storage systems based on the prediction matrix. The optimization objectives include the system's economic objectives, technical performance objectives, and environmental objectives. At the same time, the constraints of the optimization model are set based on the energy storage system capacity constraints, photovoltaic power generation system space limitations, dynamic load demand satisfaction constraints, and power quality constraints.

[0168] The dung beetle optimization module solves multi-objective optimization models based on the dung beetle optimization algorithm. By randomly generating populations, calculating fitness, performing global search and local optimization, and dynamically adjusting optimization parameters, it ultimately outputs the optimal solution for the photovoltaic system's installed capacity, energy storage capacity, and layout selection scheme.

[0169] The scheduling strategy generation module is used to generate dynamic scheduling strategies based on the optimization results and forecast data, including real-time photovoltaic power generation output allocation, energy storage charge and discharge control, and load demand response, to achieve dynamic operation management of the system;

[0170] The system evaluation and update module is used to monitor the real-time operating status and actual data of the system, and dynamically evaluate and update the optimization model and scheduling strategy based on the monitoring data.

[0171] The distributed photovoltaic and energy storage joint planning system based on deep learning ensures the accurate execution of the distributed photovoltaic and energy storage joint planning method based on deep learning.

[0172] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0173] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A distributed photovoltaic and energy storage joint planning method based on deep learning, characterized in that: The following steps are involved: Step 1: Data Collection Collect multi-source heterogeneous datasets that influence the planning of distributed photovoltaic power generation and energy storage systems; Step 2: Data processing Perform data cleaning, normalization, denoising and feature extraction on multi-source heterogeneous data; Step 3: Build the model Build a variational autoencoder model; Step 4: Train the model The variational autoencoder model is trained using the preprocessed multi-source heterogeneous dataset as training data. By minimizing the reconstruction error and the KL divergence of the latent distribution, a variational autoencoder model is generated that can characterize the uncertainty characteristics of meteorological condition data, load demand data, and electricity price fluctuation data. Step 5: Build a multi-objective optimization model The trained variational autoencoder model is used to perform uncertainty prediction of photovoltaic power generation, load demand, and electricity price fluctuations in future time periods, generating prediction results containing multi-dimensional uncertainty characteristics. Based on the prediction results, a multi-objective optimization model for distributed photovoltaic power generation systems and energy storage systems is constructed. Step 6: Solve the model Use the dung beetle optimization algorithm to solve multi-objective optimization models; Step 7: Develop a strategy Based on the optimization solution of the dung beetle optimization algorithm and the prediction results of the variational autoencoder model, a dynamic scheduling strategy for distributed photovoltaic power generation systems and energy storage systems is formulated.

2. The method for joint planning of distributed photovoltaic and energy storage systems based on deep learning according to claim 1, characterized in that: The specific collection content of the multi-source heterogeneous data set described in step 1 is: collecting meteorological condition data, load demand data and electricity price fluctuation data, and constructing a multi-source heterogeneous data set D by integrating meteorological condition data, load demand data and electricity price fluctuation data. multi : D multi ={(I t ,T t ,W t ,H t ,L t ,P t )∣t=1,2,n} Among them, I t Indicates the sunshine intensity at time t, T t Indicates the ambient temperature at time t, W t represents the wind speed at time t, H t Indicates the humidity at time t, L t Indicates the load power demand at time t, P t represents the electricity value at time t, n represents the number of time steps of the dataset, and the multi-source heterogeneous dataset D multi It consists of multidimensional data at each time step t.

3. The method for joint planning of distributed photovoltaic and energy storage systems based on deep learning according to claim 1, characterized in that: The specific method of data processing described in step 2 is: multi Perform data cleaning, remove outliers and missing values, and use linear interpolation to supplement missing data in the data; normalize the cleaned multi-source heterogeneous data set to unify the numerical range of data of different dimensions; denoise the normalized multi-source heterogeneous data set, use the sliding average filter method to smooth the time series data, remove the sharp fluctuations in a short period of time, and obtain the denoised multi-source heterogeneous data set; extract key features from the denoised multi-source heterogeneous data set, use principal component analysis to reduce the dimension of the multi-source heterogeneous data set, and define the feature matrix after dimensionality reduction as Among them, W is the characteristic matrix of PCA dimension reduction, which represents the weight of the principal component. It is a multi-source heterogeneous dataset after denoising.

4. The method for joint planning of distributed photovoltaic and energy storage systems based on deep learning according to claim 1, characterized in that: The specific construction method of the variational autoencoder model described in step 3 is: construct an adaptive variational autoencoder model, which includes an encoder network, a decoder network and a dynamic feature weighting module. The encoder network is used to reduce the dimension feature matrix Mapping to the potential distribution space, the decoder network is used to generate reconstructed data from the potential distribution space, and the dynamic feature weighting module is used to dynamically adjust the feature importance weights during training; Define the encoder network, which reduces the dimension of the feature matrix As input, the mean μ and standard deviation σ of the potential distribution are calculated, and the dynamic feature weighting module is combined to generate the dynamically adjusted latent variable z: z=w·(μ+σ·∈) Among them, z represents the potential latent variable after dynamic adjustment, represents the feature weighted vector, d is the feature dimension, and the dynamic feature weighting module adaptively learns the weight w according to the importance of the feature’s impact on photovoltaic power generation and energy storage planning. is a standard normal distribution random variable, used to introduce randomness; Define the decoder network, which reconstructs the input reduced-dimensional feature matrix from the dynamically adjusted latent variable z through a multi-layer neural network Constrained optimization is added to reconstruct the sensitivity to energy storage capacity and load fluctuations: Where g(·) represents the nonlinear mapping function of the decoder network, λ is the optimization constraint strength coefficient, It is an optimization constraint used to enhance the response characteristics of energy storage capacity fluctuations and load demand to hidden variables.

5. The method for joint planning of distributed photovoltaic and energy storage based on deep learning according to claim 1, characterized in that: The specific method of training the model described in step 4 is: using the preprocessed multi-source heterogeneous dataset The training data is input into the adaptive variational autoencoder model. The training goal is to generate a variational autoencoder model that can characterize the uncertainty characteristics of meteorological condition data, load demand data, and electricity price fluctuation data by minimizing the reconstruction error and the Kullback-Leibler divergence of the potential distribution; the reconstruction error term of the variational autoencoder model is defined as Reconstruction error is used to measure the reconstruction output of the variational autoencoder model With the input feature matrix The difference between the two; define the Kullback-Leibler divergence term of the potential distribution, which is used to measure the distribution of latent variables. The difference from the prior distribution p(z): Define the total loss function, combine the reconstruction error term and the Kullback-Leibler divergence term, and add the dynamic feature weighted regularization term R(w) to achieve adaptive optimization: Where R(w)=||ww target || 2 is the feature weighted regularization term, w is the current weight vector generated by the dynamic feature weighting module, w target represents the target distribution of feature importance, β is the regularization strength coefficient, represents the latent variable distribution generated by the encoder network, KL represents the Kullback-Leibler divergence, and p(z) represents the prior distribution; By minimizing the total loss function The parameters of the adaptive variational autoencoder model are optimized. During training, the weights of the encoder and decoder networks are updated using the stochastic gradient descent algorithm. The weight vector w of the dynamic feature weighting module is adjusted to adapt to the impact of data characteristics on the joint planning of photovoltaic power generation and energy storage. After the training is completed, the adaptive variational autoencoder model is output. The adaptive variational autoencoder model characterizes the uncertainty characteristics of meteorological condition data, load demand data and electricity price fluctuation data, and generates a feature representation z containing latent variables of the potential distribution.

6. The method for joint planning of distributed photovoltaic and energy storage based on deep learning according to claim 1, characterized in that: The specific method of constructing the multi-objective optimization model described in step 5 is: using the trained adaptive variational autoencoder model, inputting the latest feature representation of multi-source heterogeneous data sets The encoder network generates the latent variable z, which is then reconstructed using the decoder network to output the photovoltaic power generation forecast, load demand forecast, and electricity price fluctuation forecast for the future time period. Combine the prediction results to construct a prediction matrix containing multi-dimensional uncertainty characteristics in, represents the predicted value of photovoltaic power generation at time t, represents the load demand forecast value at time t, represents the predicted value of electricity price at time t, and T represents the number of time steps of prediction; Based on the prediction matrix Construct a multi-objective optimization model for distributed photovoltaic power generation system and energy storage system. The multi-objective optimization model includes the following optimization objectives: Economic goal: maximize system revenue R and minimize total investment and operating costs C total ; Technical performance goals, maximize system stability S sys , optimize the utilization efficiency of energy storage systems; Environmental goals, minimizing carbon emissions carbon ; The overall objective function is defined as: Minimize:F=-R+C total -S sys +E carbon Set constraints for the multi-objective optimization model, including: Energy storage system capacity constraint: the charge and discharge capacity of the energy storage system must not exceed the energy storage capacity C storage ; The space limit of photovoltaic power generation system is that the installed capacity of photovoltaic power generation system shall not exceed the available capacity of space C. space ; The dynamic load demand meets the constraints, and the system's power generation and energy storage discharge power meet the dynamic load demand; Power quality constraints, the fluctuation range is kept at a stable value ΔQ max Inside.

7. The method for joint planning of distributed photovoltaic and energy storage systems based on deep learning according to claim 1, characterized in that: The method for solving the model described in step 6 is: randomly generate N initial solutions, each solution represents the installed capacity C of the photovoltaic system PV , energy storage capacity C storage and layout position L site Combination of: Where N represents the population size, represents the installed capacity of the photovoltaic system of individual i, represents the energy storage capacity of individual i, represents the layout position of individual i; According to the prediction matrix Combined with photovoltaic power generation forecast, load demand forecast and electricity price fluctuation forecast, the comprehensive fitness F(X i ): F(X i )=w1·R(X i )-w2·C total (X i )+w3·S sys (X i )-w4·E carbon (X i ) Among them, w1, w2, w3, w4 are the weight coefficients of the objective function, R(X i ) represents individual X i The system benefit, C total (X i ) represents individual X i Total investment and operating costs, S sys (X i ) represents individual X i The system stability, E carbon (X i ) represents individual X i carbon emissions; Simulate the rolling behavior of dung beetles to adjust individual X i Position in the solution space to achieve global distribution optimization: in, represents the position of individual i in the t+1 generation, represents the global optimal solution of the tth generation, r is the global search step factor, and ∈1 is the normal distribution random perturbation term; Based on the global search, the details of the current solution are optimized using the dung beetle's subtle adjustment behavior: Among them, η1 is the local optimization step size, ΔX i represents the update amount of the local solution; According to the convergence trend of the dung beetle optimization algorithm, the global search step factor r and the local optimization step η1 are dynamically adjusted: r new =r0·exp(-α·t) or new =η0·exp(-β·t) Among them, r0 and η0 are the initial step sizes, α and β are the convergence rates, and t is the current number of iterations; When the maximum number of iterations is reached or the population fitness converges, the optimal solution X is output. best , determine the photovoltaic system installation capacity, energy storage capacity and layout selection plan.

8. The method for joint planning of distributed photovoltaic and energy storage systems based on deep learning according to claim 1, characterized in that: The dynamic scheduling strategy described in step seven specifically includes real-time photovoltaic power generation output allocation, energy storage system charging and discharging control, and load demand response.

9. A method for joint planning of distributed photovoltaic and energy storage systems based on deep learning according to any one of claims 1 to 8, characterized in that: Also included is a distributed photovoltaic and energy storage joint planning system based on deep learning for executing the method, the system comprising the following modules: The data acquisition module is used to collect multi-source heterogeneous data that affects the planning of distributed photovoltaic power generation and energy storage systems, including meteorological conditions, load demand data, and electricity price fluctuation data, construct a multi-source heterogeneous data set, and send the data to the data preprocessing module; The data preprocessing module is used to clean, normalize, denoise, and extract features from multi-source heterogeneous data sets, generate a feature matrix after dimensionality reduction, and send the preprocessed data to the model training module; The model building and training module is used to build an adaptive variational autoencoder model and train the model using preprocessed multi-source heterogeneous datasets. By minimizing the reconstruction error and the Kullback-Leibler divergence of the underlying distribution, a variational autoencoder model is generated that can characterize the uncertainty characteristics of meteorological conditions data, load demand data, and electricity price fluctuation data. The uncertainty prediction module uses the trained variational autoencoder model to predict the uncertainty of photovoltaic power generation, load demand, and electricity price fluctuations in the future time period based on the latest input data features, and generates a prediction matrix containing multi-dimensional uncertainty features; An optimization model building module is used to construct a multi-objective optimization model for distributed photovoltaic power generation systems and energy storage systems based on the prediction matrix. The optimization objectives include the system's economic objectives, technical performance objectives, and environmental objectives. At the same time, the constraints of the optimization model are set based on the energy storage system capacity constraints, photovoltaic power generation system space limitations, dynamic load demand satisfaction constraints, and power quality constraints. The dung beetle optimization module solves multi-objective optimization models based on the dung beetle optimization algorithm. By randomly generating populations, calculating fitness, performing global search and local optimization, and dynamically adjusting optimization parameters, it ultimately outputs the optimal solution for the photovoltaic system's installed capacity, energy storage capacity, and layout selection scheme. The scheduling strategy generation module is used to generate dynamic scheduling strategies based on the optimization results and forecast data, including real-time photovoltaic power generation output allocation, energy storage charge and discharge control, and load demand response, to achieve dynamic operation management of the system; The system evaluation and update module is used to monitor the real-time operating status and actual data of the system, and dynamically evaluate and update the optimization model and scheduling strategy based on the monitoring data.

Images